blob: d394d9545c4eb7ea184409698546d98d7bda8589 [file] [log] [blame]
Thomas Gleixner457c8992019-05-19 13:08:55 +01001// SPDX-License-Identifier: GPL-2.0-only
Linus Torvalds1da177e2005-04-16 15:20:36 -07002/*
3 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01004 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07005 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006#include <linux/list.h>
7#include <linux/init.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07008#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -070016#include <linux/compiler.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080017#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080018#include <linux/mutex.h>
Mike Rapoport97ad1082018-10-30 15:09:44 -070019#include <linux/memblock.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070020#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090021#include <linux/slab.h>
Joonsoo Kimbbe88752020-08-11 18:37:38 -070022#include <linux/sched/mm.h>
Mike Kravetz63489f82018-03-22 16:17:13 -070023#include <linux/mmdebug.h>
Ingo Molnar174cd4b2017-02-02 19:15:33 +010024#include <linux/sched/signal.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090025#include <linux/rmap.h>
Matthew Wilcoxc6247f72017-07-10 15:48:56 -070026#include <linux/string_helpers.h>
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +090027#include <linux/swap.h>
28#include <linux/swapops.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070029#include <linux/jhash.h>
Anshuman Khandual98fa15f2019-03-05 15:42:58 -080030#include <linux/numa.h>
Waiman Longc77c0a82020-01-04 13:00:15 -080031#include <linux/llist.h>
Roman Gushchincf11e852020-04-10 14:32:45 -070032#include <linux/cma.h>
Mina Almasry8cc5fcb2021-06-30 18:48:19 -070033#include <linux/migrate.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070034
David Gibson63551ae2005-06-21 17:14:44 -070035#include <asm/page.h>
Mike Rapoportca15ca42020-08-06 23:22:28 -070036#include <asm/pgalloc.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070037#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070038
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070039#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070040#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070041#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a3052302009-12-14 17:58:25 -080042#include <linux/node.h>
Michal Hockoab5ac902018-01-31 16:20:48 -080043#include <linux/page_owner.h>
Nick Piggin7835e982006-03-22 00:08:40 -080044#include "internal.h"
Muchun Songf41f2ed2021-06-30 18:47:13 -070045#include "hugetlb_vmemmap.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070047int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070048unsigned int default_hstate_idx;
49struct hstate hstates[HUGE_MAX_HSTATE];
Roman Gushchincf11e852020-04-10 14:32:45 -070050
Barry Songdbda8fe2020-07-23 21:15:30 -070051#ifdef CONFIG_CMA
Roman Gushchincf11e852020-04-10 14:32:45 -070052static struct cma *hugetlb_cma[MAX_NUMNODES];
Barry Songdbda8fe2020-07-23 21:15:30 -070053#endif
54static unsigned long hugetlb_cma_size __initdata;
Roman Gushchincf11e852020-04-10 14:32:45 -070055
Naoya Horiguchi641844f2015-06-24 16:56:59 -070056/*
57 * Minimum page order among possible hugepage sizes, set to a proper value
58 * at boot time.
59 */
60static unsigned int minimum_order __read_mostly = UINT_MAX;
Andi Kleene5ff2152008-07-23 21:27:42 -070061
Jon Tollefson53ba51d2008-07-23 21:27:52 -070062__initdata LIST_HEAD(huge_boot_pages);
63
Andi Kleene5ff2152008-07-23 21:27:42 -070064/* for command line parsing */
65static struct hstate * __initdata parsed_hstate;
66static unsigned long __initdata default_hstate_max_huge_pages;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -070067static bool __initdata parsed_valid_hugepagesz = true;
Mike Kravetz282f4212020-06-03 16:00:46 -070068static bool __initdata parsed_default_hugepagesz;
Andi Kleene5ff2152008-07-23 21:27:42 -070069
David Gibson3935baa2006-03-22 00:08:53 -080070/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070071 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
72 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080073 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070074DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080075
Davidlohr Bueso8382d912014-04-03 14:47:31 -070076/*
77 * Serializes faults on the same logical page. This is used to
78 * prevent spurious OOMs when the hugepage pool is fully utilized.
79 */
80static int num_fault_mutexes;
Mike Kravetzc672c7f2015-09-08 15:01:35 -070081struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
Davidlohr Bueso8382d912014-04-03 14:47:31 -070082
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -070083/* Forward declaration */
84static int hugetlb_acct_memory(struct hstate *h, long delta);
85
Miaohe Lin1d884332021-02-24 12:07:05 -080086static inline bool subpool_is_free(struct hugepage_subpool *spool)
87{
88 if (spool->count)
89 return false;
90 if (spool->max_hpages != -1)
91 return spool->used_hpages == 0;
92 if (spool->min_hpages != -1)
93 return spool->rsv_hpages == spool->min_hpages;
94
95 return true;
96}
97
Mike Kravetzdb71ef72021-05-04 18:35:07 -070098static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
99 unsigned long irq_flags)
David Gibson90481622012-03-21 16:34:12 -0700100{
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700101 spin_unlock_irqrestore(&spool->lock, irq_flags);
David Gibson90481622012-03-21 16:34:12 -0700102
103 /* If no pages are used, and no other handles to the subpool
Ethon Paul7c8de352020-06-04 16:49:07 -0700104 * remain, give up any reservations based on minimum size and
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700105 * free the subpool */
Miaohe Lin1d884332021-02-24 12:07:05 -0800106 if (subpool_is_free(spool)) {
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700107 if (spool->min_hpages != -1)
108 hugetlb_acct_memory(spool->hstate,
109 -spool->min_hpages);
David Gibson90481622012-03-21 16:34:12 -0700110 kfree(spool);
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700111 }
David Gibson90481622012-03-21 16:34:12 -0700112}
113
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700114struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
115 long min_hpages)
David Gibson90481622012-03-21 16:34:12 -0700116{
117 struct hugepage_subpool *spool;
118
Mike Kravetzc6a91822015-04-15 16:13:36 -0700119 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
David Gibson90481622012-03-21 16:34:12 -0700120 if (!spool)
121 return NULL;
122
123 spin_lock_init(&spool->lock);
124 spool->count = 1;
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700125 spool->max_hpages = max_hpages;
126 spool->hstate = h;
127 spool->min_hpages = min_hpages;
128
129 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
130 kfree(spool);
131 return NULL;
132 }
133 spool->rsv_hpages = min_hpages;
David Gibson90481622012-03-21 16:34:12 -0700134
135 return spool;
136}
137
138void hugepage_put_subpool(struct hugepage_subpool *spool)
139{
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700140 unsigned long flags;
141
142 spin_lock_irqsave(&spool->lock, flags);
David Gibson90481622012-03-21 16:34:12 -0700143 BUG_ON(!spool->count);
144 spool->count--;
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700145 unlock_or_release_subpool(spool, flags);
David Gibson90481622012-03-21 16:34:12 -0700146}
147
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700148/*
149 * Subpool accounting for allocating and reserving pages.
150 * Return -ENOMEM if there are not enough resources to satisfy the
Randy Dunlap9e7ee402020-08-11 18:32:59 -0700151 * request. Otherwise, return the number of pages by which the
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700152 * global pools must be adjusted (upward). The returned value may
153 * only be different than the passed value (delta) in the case where
Ethon Paul7c8de352020-06-04 16:49:07 -0700154 * a subpool minimum size must be maintained.
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700155 */
156static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700157 long delta)
158{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700159 long ret = delta;
David Gibson90481622012-03-21 16:34:12 -0700160
161 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700162 return ret;
David Gibson90481622012-03-21 16:34:12 -0700163
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700164 spin_lock_irq(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700165
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700166 if (spool->max_hpages != -1) { /* maximum size accounting */
167 if ((spool->used_hpages + delta) <= spool->max_hpages)
168 spool->used_hpages += delta;
169 else {
170 ret = -ENOMEM;
171 goto unlock_ret;
172 }
173 }
174
Mike Kravetz09a95e22016-05-19 17:11:01 -0700175 /* minimum size accounting */
176 if (spool->min_hpages != -1 && spool->rsv_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700177 if (delta > spool->rsv_hpages) {
178 /*
179 * Asking for more reserves than those already taken on
180 * behalf of subpool. Return difference.
181 */
182 ret = delta - spool->rsv_hpages;
183 spool->rsv_hpages = 0;
184 } else {
185 ret = 0; /* reserves already accounted for */
186 spool->rsv_hpages -= delta;
187 }
188 }
189
190unlock_ret:
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700191 spin_unlock_irq(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700192 return ret;
193}
194
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700195/*
196 * Subpool accounting for freeing and unreserving pages.
197 * Return the number of global page reservations that must be dropped.
198 * The return value may only be different than the passed value (delta)
199 * in the case where a subpool minimum size must be maintained.
200 */
201static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700202 long delta)
203{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700204 long ret = delta;
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700205 unsigned long flags;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700206
David Gibson90481622012-03-21 16:34:12 -0700207 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700208 return delta;
David Gibson90481622012-03-21 16:34:12 -0700209
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700210 spin_lock_irqsave(&spool->lock, flags);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700211
212 if (spool->max_hpages != -1) /* maximum size accounting */
213 spool->used_hpages -= delta;
214
Mike Kravetz09a95e22016-05-19 17:11:01 -0700215 /* minimum size accounting */
216 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700217 if (spool->rsv_hpages + delta <= spool->min_hpages)
218 ret = 0;
219 else
220 ret = spool->rsv_hpages + delta - spool->min_hpages;
221
222 spool->rsv_hpages += delta;
223 if (spool->rsv_hpages > spool->min_hpages)
224 spool->rsv_hpages = spool->min_hpages;
225 }
226
227 /*
228 * If hugetlbfs_put_super couldn't free spool due to an outstanding
229 * quota reference, free it now.
230 */
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700231 unlock_or_release_subpool(spool, flags);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700232
233 return ret;
David Gibson90481622012-03-21 16:34:12 -0700234}
235
236static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
237{
238 return HUGETLBFS_SB(inode->i_sb)->spool;
239}
240
241static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
242{
Al Viro496ad9a2013-01-23 17:07:38 -0500243 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700244}
245
Mina Almasry0db9d742020-04-01 21:11:25 -0700246/* Helper that removes a struct file_region from the resv_map cache and returns
247 * it for use.
248 */
249static struct file_region *
250get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
251{
252 struct file_region *nrg = NULL;
253
254 VM_BUG_ON(resv->region_cache_count <= 0);
255
256 resv->region_cache_count--;
257 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
Mina Almasry0db9d742020-04-01 21:11:25 -0700258 list_del(&nrg->link);
259
260 nrg->from = from;
261 nrg->to = to;
262
263 return nrg;
264}
265
Mina Almasry075a61d2020-04-01 21:11:28 -0700266static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
267 struct file_region *rg)
268{
269#ifdef CONFIG_CGROUP_HUGETLB
270 nrg->reservation_counter = rg->reservation_counter;
271 nrg->css = rg->css;
272 if (rg->css)
273 css_get(rg->css);
274#endif
275}
276
277/* Helper that records hugetlb_cgroup uncharge info. */
278static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
279 struct hstate *h,
280 struct resv_map *resv,
281 struct file_region *nrg)
282{
283#ifdef CONFIG_CGROUP_HUGETLB
284 if (h_cg) {
285 nrg->reservation_counter =
286 &h_cg->rsvd_hugepage[hstate_index(h)];
287 nrg->css = &h_cg->css;
Miaohe Lind85aecf2021-03-24 21:37:17 -0700288 /*
289 * The caller will hold exactly one h_cg->css reference for the
290 * whole contiguous reservation region. But this area might be
291 * scattered when there are already some file_regions reside in
292 * it. As a result, many file_regions may share only one css
293 * reference. In order to ensure that one file_region must hold
294 * exactly one h_cg->css reference, we should do css_get for
295 * each file_region and leave the reference held by caller
296 * untouched.
297 */
298 css_get(&h_cg->css);
Mina Almasry075a61d2020-04-01 21:11:28 -0700299 if (!resv->pages_per_hpage)
300 resv->pages_per_hpage = pages_per_huge_page(h);
301 /* pages_per_hpage should be the same for all entries in
302 * a resv_map.
303 */
304 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
305 } else {
306 nrg->reservation_counter = NULL;
307 nrg->css = NULL;
308 }
309#endif
310}
311
Miaohe Lind85aecf2021-03-24 21:37:17 -0700312static void put_uncharge_info(struct file_region *rg)
313{
314#ifdef CONFIG_CGROUP_HUGETLB
315 if (rg->css)
316 css_put(rg->css);
317#endif
318}
319
Mina Almasrya9b3f862020-04-01 21:11:35 -0700320static bool has_same_uncharge_info(struct file_region *rg,
321 struct file_region *org)
322{
323#ifdef CONFIG_CGROUP_HUGETLB
324 return rg && org &&
325 rg->reservation_counter == org->reservation_counter &&
326 rg->css == org->css;
327
328#else
329 return true;
330#endif
331}
332
333static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
334{
335 struct file_region *nrg = NULL, *prg = NULL;
336
337 prg = list_prev_entry(rg, link);
338 if (&prg->link != &resv->regions && prg->to == rg->from &&
339 has_same_uncharge_info(prg, rg)) {
340 prg->to = rg->to;
341
342 list_del(&rg->link);
Miaohe Lind85aecf2021-03-24 21:37:17 -0700343 put_uncharge_info(rg);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700344 kfree(rg);
345
Wei Yang7db5e7b2020-10-13 16:56:20 -0700346 rg = prg;
Mina Almasrya9b3f862020-04-01 21:11:35 -0700347 }
348
349 nrg = list_next_entry(rg, link);
350 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
351 has_same_uncharge_info(nrg, rg)) {
352 nrg->from = rg->from;
353
354 list_del(&rg->link);
Miaohe Lind85aecf2021-03-24 21:37:17 -0700355 put_uncharge_info(rg);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700356 kfree(rg);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700357 }
358}
359
Peter Xu2103cf92021-03-12 21:07:18 -0800360static inline long
361hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from,
362 long to, struct hstate *h, struct hugetlb_cgroup *cg,
363 long *regions_needed)
364{
365 struct file_region *nrg;
366
367 if (!regions_needed) {
368 nrg = get_file_region_entry_from_cache(map, from, to);
369 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
370 list_add(&nrg->link, rg->link.prev);
371 coalesce_file_region(map, nrg);
372 } else
373 *regions_needed += 1;
374
375 return to - from;
376}
377
Wei Yang972a3da32020-10-13 16:56:30 -0700378/*
379 * Must be called with resv->lock held.
380 *
381 * Calling this with regions_needed != NULL will count the number of pages
382 * to be added but will not modify the linked list. And regions_needed will
383 * indicate the number of file_regions needed in the cache to carry out to add
384 * the regions for this range.
Mina Almasryd75c6af2019-11-30 17:56:59 -0800385 */
386static long add_reservation_in_range(struct resv_map *resv, long f, long t,
Mina Almasry075a61d2020-04-01 21:11:28 -0700387 struct hugetlb_cgroup *h_cg,
Wei Yang972a3da32020-10-13 16:56:30 -0700388 struct hstate *h, long *regions_needed)
Mina Almasryd75c6af2019-11-30 17:56:59 -0800389{
Mina Almasry0db9d742020-04-01 21:11:25 -0700390 long add = 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800391 struct list_head *head = &resv->regions;
Mina Almasry0db9d742020-04-01 21:11:25 -0700392 long last_accounted_offset = f;
Peter Xu2103cf92021-03-12 21:07:18 -0800393 struct file_region *rg = NULL, *trg = NULL;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800394
Mina Almasry0db9d742020-04-01 21:11:25 -0700395 if (regions_needed)
396 *regions_needed = 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800397
Mina Almasry0db9d742020-04-01 21:11:25 -0700398 /* In this loop, we essentially handle an entry for the range
399 * [last_accounted_offset, rg->from), at every iteration, with some
400 * bounds checking.
401 */
402 list_for_each_entry_safe(rg, trg, head, link) {
403 /* Skip irrelevant regions that start before our range. */
404 if (rg->from < f) {
405 /* If this region ends after the last accounted offset,
406 * then we need to update last_accounted_offset.
407 */
408 if (rg->to > last_accounted_offset)
409 last_accounted_offset = rg->to;
410 continue;
411 }
Mina Almasryd75c6af2019-11-30 17:56:59 -0800412
Mina Almasry0db9d742020-04-01 21:11:25 -0700413 /* When we find a region that starts beyond our range, we've
414 * finished.
415 */
Peter Xuca7e0452021-03-12 21:07:22 -0800416 if (rg->from >= t)
Mina Almasryd75c6af2019-11-30 17:56:59 -0800417 break;
418
Mina Almasry0db9d742020-04-01 21:11:25 -0700419 /* Add an entry for last_accounted_offset -> rg->from, and
420 * update last_accounted_offset.
Mina Almasryd75c6af2019-11-30 17:56:59 -0800421 */
Peter Xu2103cf92021-03-12 21:07:18 -0800422 if (rg->from > last_accounted_offset)
423 add += hugetlb_resv_map_add(resv, rg,
424 last_accounted_offset,
425 rg->from, h, h_cg,
426 regions_needed);
Mina Almasryd75c6af2019-11-30 17:56:59 -0800427
Mina Almasry0db9d742020-04-01 21:11:25 -0700428 last_accounted_offset = rg->to;
429 }
430
431 /* Handle the case where our range extends beyond
432 * last_accounted_offset.
433 */
Peter Xu2103cf92021-03-12 21:07:18 -0800434 if (last_accounted_offset < t)
435 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
436 t, h, h_cg, regions_needed);
Mina Almasry0db9d742020-04-01 21:11:25 -0700437
438 VM_BUG_ON(add < 0);
439 return add;
440}
441
442/* Must be called with resv->lock acquired. Will drop lock to allocate entries.
443 */
444static int allocate_file_region_entries(struct resv_map *resv,
445 int regions_needed)
446 __must_hold(&resv->lock)
447{
448 struct list_head allocated_regions;
449 int to_allocate = 0, i = 0;
450 struct file_region *trg = NULL, *rg = NULL;
451
452 VM_BUG_ON(regions_needed < 0);
453
454 INIT_LIST_HEAD(&allocated_regions);
455
456 /*
457 * Check for sufficient descriptors in the cache to accommodate
458 * the number of in progress add operations plus regions_needed.
459 *
460 * This is a while loop because when we drop the lock, some other call
461 * to region_add or region_del may have consumed some region_entries,
462 * so we keep looping here until we finally have enough entries for
463 * (adds_in_progress + regions_needed).
464 */
465 while (resv->region_cache_count <
466 (resv->adds_in_progress + regions_needed)) {
467 to_allocate = resv->adds_in_progress + regions_needed -
468 resv->region_cache_count;
469
470 /* At this point, we should have enough entries in the cache
Ingo Molnarf0953a12021-05-06 18:06:47 -0700471 * for all the existing adds_in_progress. We should only be
Mina Almasry0db9d742020-04-01 21:11:25 -0700472 * needing to allocate for regions_needed.
473 */
474 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
475
476 spin_unlock(&resv->lock);
477 for (i = 0; i < to_allocate; i++) {
478 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
479 if (!trg)
480 goto out_of_memory;
481 list_add(&trg->link, &allocated_regions);
482 }
483
484 spin_lock(&resv->lock);
485
Wei Yangd3ec7b62020-10-13 16:56:27 -0700486 list_splice(&allocated_regions, &resv->region_cache);
487 resv->region_cache_count += to_allocate;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800488 }
489
Mina Almasry0db9d742020-04-01 21:11:25 -0700490 return 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800491
Mina Almasry0db9d742020-04-01 21:11:25 -0700492out_of_memory:
493 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
494 list_del(&rg->link);
495 kfree(rg);
496 }
497 return -ENOMEM;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800498}
499
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700500/*
501 * Add the huge page range represented by [f, t) to the reserve
Mina Almasry0db9d742020-04-01 21:11:25 -0700502 * map. Regions will be taken from the cache to fill in this range.
503 * Sufficient regions should exist in the cache due to the previous
504 * call to region_chg with the same range, but in some cases the cache will not
505 * have sufficient entries due to races with other code doing region_add or
506 * region_del. The extra needed entries will be allocated.
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700507 *
Mina Almasry0db9d742020-04-01 21:11:25 -0700508 * regions_needed is the out value provided by a previous call to region_chg.
509 *
510 * Return the number of new huge pages added to the map. This number is greater
511 * than or equal to zero. If file_region entries needed to be allocated for
Ethon Paul7c8de352020-06-04 16:49:07 -0700512 * this operation and we were not able to allocate, it returns -ENOMEM.
Mina Almasry0db9d742020-04-01 21:11:25 -0700513 * region_add of regions of length 1 never allocate file_regions and cannot
514 * fail; region_chg will always allocate at least 1 entry and a region_add for
515 * 1 page will only require at most 1 entry.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700516 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700517static long region_add(struct resv_map *resv, long f, long t,
Mina Almasry075a61d2020-04-01 21:11:28 -0700518 long in_regions_needed, struct hstate *h,
519 struct hugetlb_cgroup *h_cg)
Andy Whitcroft96822902008-07-23 21:27:29 -0700520{
Mina Almasry0db9d742020-04-01 21:11:25 -0700521 long add = 0, actual_regions_needed = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700522
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700523 spin_lock(&resv->lock);
Mina Almasry0db9d742020-04-01 21:11:25 -0700524retry:
525
526 /* Count how many regions are actually needed to execute this add. */
Wei Yang972a3da32020-10-13 16:56:30 -0700527 add_reservation_in_range(resv, f, t, NULL, NULL,
528 &actual_regions_needed);
Andy Whitcroft96822902008-07-23 21:27:29 -0700529
Mike Kravetz5e911372015-09-08 15:01:28 -0700530 /*
Mina Almasry0db9d742020-04-01 21:11:25 -0700531 * Check for sufficient descriptors in the cache to accommodate
532 * this add operation. Note that actual_regions_needed may be greater
533 * than in_regions_needed, as the resv_map may have been modified since
534 * the region_chg call. In this case, we need to make sure that we
535 * allocate extra entries, such that we have enough for all the
536 * existing adds_in_progress, plus the excess needed for this
537 * operation.
Mike Kravetz5e911372015-09-08 15:01:28 -0700538 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700539 if (actual_regions_needed > in_regions_needed &&
540 resv->region_cache_count <
541 resv->adds_in_progress +
542 (actual_regions_needed - in_regions_needed)) {
543 /* region_add operation of range 1 should never need to
544 * allocate file_region entries.
545 */
546 VM_BUG_ON(t - f <= 1);
Mike Kravetz5e911372015-09-08 15:01:28 -0700547
Mina Almasry0db9d742020-04-01 21:11:25 -0700548 if (allocate_file_region_entries(
549 resv, actual_regions_needed - in_regions_needed)) {
550 return -ENOMEM;
551 }
Mike Kravetz5e911372015-09-08 15:01:28 -0700552
Mina Almasry0db9d742020-04-01 21:11:25 -0700553 goto retry;
Mike Kravetz5e911372015-09-08 15:01:28 -0700554 }
555
Wei Yang972a3da32020-10-13 16:56:30 -0700556 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700557
Mina Almasry0db9d742020-04-01 21:11:25 -0700558 resv->adds_in_progress -= in_regions_needed;
559
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700560 spin_unlock(&resv->lock);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700561 return add;
Andy Whitcroft96822902008-07-23 21:27:29 -0700562}
563
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700564/*
565 * Examine the existing reserve map and determine how many
566 * huge pages in the specified range [f, t) are NOT currently
567 * represented. This routine is called before a subsequent
568 * call to region_add that will actually modify the reserve
569 * map to add the specified range [f, t). region_chg does
570 * not change the number of huge pages represented by the
Mina Almasry0db9d742020-04-01 21:11:25 -0700571 * map. A number of new file_region structures is added to the cache as a
572 * placeholder, for the subsequent region_add call to use. At least 1
573 * file_region structure is added.
574 *
575 * out_regions_needed is the number of regions added to the
576 * resv->adds_in_progress. This value needs to be provided to a follow up call
577 * to region_add or region_abort for proper accounting.
Mike Kravetz5e911372015-09-08 15:01:28 -0700578 *
579 * Returns the number of huge pages that need to be added to the existing
580 * reservation map for the range [f, t). This number is greater or equal to
581 * zero. -ENOMEM is returned if a new file_region structure or cache entry
582 * is needed and can not be allocated.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700583 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700584static long region_chg(struct resv_map *resv, long f, long t,
585 long *out_regions_needed)
Andy Whitcroft96822902008-07-23 21:27:29 -0700586{
Andy Whitcroft96822902008-07-23 21:27:29 -0700587 long chg = 0;
588
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700589 spin_lock(&resv->lock);
Mike Kravetz5e911372015-09-08 15:01:28 -0700590
Wei Yang972a3da32020-10-13 16:56:30 -0700591 /* Count how many hugepages in this range are NOT represented. */
Mina Almasry075a61d2020-04-01 21:11:28 -0700592 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
Wei Yang972a3da32020-10-13 16:56:30 -0700593 out_regions_needed);
Mike Kravetz5e911372015-09-08 15:01:28 -0700594
Mina Almasry0db9d742020-04-01 21:11:25 -0700595 if (*out_regions_needed == 0)
596 *out_regions_needed = 1;
Mike Kravetz5e911372015-09-08 15:01:28 -0700597
Mina Almasry0db9d742020-04-01 21:11:25 -0700598 if (allocate_file_region_entries(resv, *out_regions_needed))
599 return -ENOMEM;
Mike Kravetz5e911372015-09-08 15:01:28 -0700600
Mina Almasry0db9d742020-04-01 21:11:25 -0700601 resv->adds_in_progress += *out_regions_needed;
Andy Whitcroft96822902008-07-23 21:27:29 -0700602
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700603 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700604 return chg;
605}
606
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700607/*
Mike Kravetz5e911372015-09-08 15:01:28 -0700608 * Abort the in progress add operation. The adds_in_progress field
609 * of the resv_map keeps track of the operations in progress between
610 * calls to region_chg and region_add. Operations are sometimes
611 * aborted after the call to region_chg. In such cases, region_abort
Mina Almasry0db9d742020-04-01 21:11:25 -0700612 * is called to decrement the adds_in_progress counter. regions_needed
613 * is the value returned by the region_chg call, it is used to decrement
614 * the adds_in_progress counter.
Mike Kravetz5e911372015-09-08 15:01:28 -0700615 *
616 * NOTE: The range arguments [f, t) are not needed or used in this
617 * routine. They are kept to make reading the calling code easier as
618 * arguments will match the associated region_chg call.
619 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700620static void region_abort(struct resv_map *resv, long f, long t,
621 long regions_needed)
Mike Kravetz5e911372015-09-08 15:01:28 -0700622{
623 spin_lock(&resv->lock);
624 VM_BUG_ON(!resv->region_cache_count);
Mina Almasry0db9d742020-04-01 21:11:25 -0700625 resv->adds_in_progress -= regions_needed;
Mike Kravetz5e911372015-09-08 15:01:28 -0700626 spin_unlock(&resv->lock);
627}
628
629/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700630 * Delete the specified range [f, t) from the reserve map. If the
631 * t parameter is LONG_MAX, this indicates that ALL regions after f
632 * should be deleted. Locate the regions which intersect [f, t)
633 * and either trim, delete or split the existing regions.
634 *
635 * Returns the number of huge pages deleted from the reserve map.
636 * In the normal case, the return value is zero or more. In the
637 * case where a region must be split, a new region descriptor must
638 * be allocated. If the allocation fails, -ENOMEM will be returned.
639 * NOTE: If the parameter t == LONG_MAX, then we will never split
640 * a region and possibly return -ENOMEM. Callers specifying
641 * t == LONG_MAX do not need to check for -ENOMEM error.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700642 */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700643static long region_del(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700644{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700645 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700646 struct file_region *rg, *trg;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700647 struct file_region *nrg = NULL;
648 long del = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700649
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700650retry:
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700651 spin_lock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700652 list_for_each_entry_safe(rg, trg, head, link) {
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800653 /*
654 * Skip regions before the range to be deleted. file_region
655 * ranges are normally of the form [from, to). However, there
656 * may be a "placeholder" entry in the map which is of the form
657 * (from, to) with from == to. Check for placeholder entries
658 * at the beginning of the range to be deleted.
659 */
660 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700661 continue;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800662
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700663 if (rg->from >= t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700664 break;
Andy Whitcroft96822902008-07-23 21:27:29 -0700665
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700666 if (f > rg->from && t < rg->to) { /* Must split region */
667 /*
668 * Check for an entry in the cache before dropping
669 * lock and attempting allocation.
670 */
671 if (!nrg &&
672 resv->region_cache_count > resv->adds_in_progress) {
673 nrg = list_first_entry(&resv->region_cache,
674 struct file_region,
675 link);
676 list_del(&nrg->link);
677 resv->region_cache_count--;
678 }
679
680 if (!nrg) {
681 spin_unlock(&resv->lock);
682 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
683 if (!nrg)
684 return -ENOMEM;
685 goto retry;
686 }
687
688 del += t - f;
Mike Kravetz79aa9252020-11-01 17:07:27 -0800689 hugetlb_cgroup_uncharge_file_region(
Miaohe Lind85aecf2021-03-24 21:37:17 -0700690 resv, rg, t - f, false);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700691
692 /* New entry for end of split region */
693 nrg->from = t;
694 nrg->to = rg->to;
Mina Almasry075a61d2020-04-01 21:11:28 -0700695
696 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
697
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700698 INIT_LIST_HEAD(&nrg->link);
699
700 /* Original entry is trimmed */
701 rg->to = f;
702
703 list_add(&nrg->link, &rg->link);
704 nrg = NULL;
705 break;
706 }
707
708 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
709 del += rg->to - rg->from;
Mina Almasry075a61d2020-04-01 21:11:28 -0700710 hugetlb_cgroup_uncharge_file_region(resv, rg,
Miaohe Lind85aecf2021-03-24 21:37:17 -0700711 rg->to - rg->from, true);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700712 list_del(&rg->link);
713 kfree(rg);
714 continue;
715 }
716
717 if (f <= rg->from) { /* Trim beginning of region */
Mina Almasry075a61d2020-04-01 21:11:28 -0700718 hugetlb_cgroup_uncharge_file_region(resv, rg,
Miaohe Lind85aecf2021-03-24 21:37:17 -0700719 t - rg->from, false);
Mina Almasry075a61d2020-04-01 21:11:28 -0700720
Mike Kravetz79aa9252020-11-01 17:07:27 -0800721 del += t - rg->from;
722 rg->from = t;
723 } else { /* Trim end of region */
Mina Almasry075a61d2020-04-01 21:11:28 -0700724 hugetlb_cgroup_uncharge_file_region(resv, rg,
Miaohe Lind85aecf2021-03-24 21:37:17 -0700725 rg->to - f, false);
Mike Kravetz79aa9252020-11-01 17:07:27 -0800726
727 del += rg->to - f;
728 rg->to = f;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700729 }
Andy Whitcroft96822902008-07-23 21:27:29 -0700730 }
731
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700732 spin_unlock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700733 kfree(nrg);
734 return del;
Andy Whitcroft96822902008-07-23 21:27:29 -0700735}
736
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700737/*
Mike Kravetzb5cec282015-09-08 15:01:41 -0700738 * A rare out of memory error was encountered which prevented removal of
739 * the reserve map region for a page. The huge page itself was free'ed
740 * and removed from the page cache. This routine will adjust the subpool
741 * usage count, and the global reserve count if needed. By incrementing
742 * these counts, the reserve map entry which could not be deleted will
743 * appear as a "reserved" entry instead of simply dangling with incorrect
744 * counts.
745 */
zhong jiang72e29362016-10-07 17:02:01 -0700746void hugetlb_fix_reserve_counts(struct inode *inode)
Mike Kravetzb5cec282015-09-08 15:01:41 -0700747{
748 struct hugepage_subpool *spool = subpool_inode(inode);
749 long rsv_adjust;
Miaohe Linda563882021-05-04 18:34:38 -0700750 bool reserved = false;
Mike Kravetzb5cec282015-09-08 15:01:41 -0700751
752 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
Miaohe Linda563882021-05-04 18:34:38 -0700753 if (rsv_adjust > 0) {
Mike Kravetzb5cec282015-09-08 15:01:41 -0700754 struct hstate *h = hstate_inode(inode);
755
Miaohe Linda563882021-05-04 18:34:38 -0700756 if (!hugetlb_acct_memory(h, 1))
757 reserved = true;
758 } else if (!rsv_adjust) {
759 reserved = true;
Mike Kravetzb5cec282015-09-08 15:01:41 -0700760 }
Miaohe Linda563882021-05-04 18:34:38 -0700761
762 if (!reserved)
763 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
Mike Kravetzb5cec282015-09-08 15:01:41 -0700764}
765
766/*
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700767 * Count and return the number of huge pages in the reserve map
768 * that intersect with the range [f, t).
769 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700770static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700771{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700772 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700773 struct file_region *rg;
774 long chg = 0;
775
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700776 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700777 /* Locate each segment we overlap with, and count that overlap. */
778 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700779 long seg_from;
780 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700781
782 if (rg->to <= f)
783 continue;
784 if (rg->from >= t)
785 break;
786
787 seg_from = max(rg->from, f);
788 seg_to = min(rg->to, t);
789
790 chg += seg_to - seg_from;
791 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700792 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700793
794 return chg;
795}
796
Andy Whitcroft96822902008-07-23 21:27:29 -0700797/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700798 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700799 * the mapping, in pagecache page units; huge pages here.
800 */
Andi Kleena5516432008-07-23 21:27:41 -0700801static pgoff_t vma_hugecache_offset(struct hstate *h,
802 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700803{
Andi Kleena5516432008-07-23 21:27:41 -0700804 return ((address - vma->vm_start) >> huge_page_shift(h)) +
805 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700806}
807
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900808pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
809 unsigned long address)
810{
811 return vma_hugecache_offset(hstate_vma(vma), vma, address);
812}
Dan Williamsdee41072016-05-14 12:20:44 -0700813EXPORT_SYMBOL_GPL(linear_hugepage_index);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900814
Andy Whitcroft84afd992008-07-23 21:27:32 -0700815/*
Mel Gorman08fba692009-01-06 14:38:53 -0800816 * Return the size of the pages allocated when backing a VMA. In the majority
817 * cases this will be same size as used by the page table entries.
818 */
819unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
820{
Dan Williams05ea8862018-04-05 16:24:25 -0700821 if (vma->vm_ops && vma->vm_ops->pagesize)
822 return vma->vm_ops->pagesize(vma);
823 return PAGE_SIZE;
Mel Gorman08fba692009-01-06 14:38:53 -0800824}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200825EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800826
827/*
Mel Gorman33402892009-01-06 14:38:54 -0800828 * Return the page size being used by the MMU to back a VMA. In the majority
829 * of cases, the page size used by the kernel matches the MMU size. On
Dan Williams09135cc2018-04-05 16:24:21 -0700830 * architectures where it differs, an architecture-specific 'strong'
831 * version of this symbol is required.
Mel Gorman33402892009-01-06 14:38:54 -0800832 */
Dan Williams09135cc2018-04-05 16:24:21 -0700833__weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
Mel Gorman33402892009-01-06 14:38:54 -0800834{
835 return vma_kernel_pagesize(vma);
836}
Mel Gorman33402892009-01-06 14:38:54 -0800837
838/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700839 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
840 * bits of the reservation map pointer, which are always clear due to
841 * alignment.
842 */
843#define HPAGE_RESV_OWNER (1UL << 0)
844#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700845#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700846
Mel Gormana1e78772008-07-23 21:27:23 -0700847/*
848 * These helpers are used to track how many pages are reserved for
849 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
850 * is guaranteed to have their future faults succeed.
851 *
852 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
853 * the reserve counters are updated with the hugetlb_lock held. It is safe
854 * to reset the VMA at fork() time as it is not in use yet and there is no
855 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700856 *
857 * The private mapping reservation is represented in a subtly different
858 * manner to a shared mapping. A shared mapping has a region map associated
859 * with the underlying file, this region map represents the backing file
860 * pages which have ever had a reservation assigned which this persists even
861 * after the page is instantiated. A private mapping has a region map
862 * associated with the original mmap which is attached to all VMAs which
863 * reference it, this region map represents those offsets which have consumed
864 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700865 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700866static unsigned long get_vma_private_data(struct vm_area_struct *vma)
867{
868 return (unsigned long)vma->vm_private_data;
869}
870
871static void set_vma_private_data(struct vm_area_struct *vma,
872 unsigned long value)
873{
874 vma->vm_private_data = (void *)value;
875}
876
Mina Almasrye9fe92a2020-04-01 21:11:21 -0700877static void
878resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
879 struct hugetlb_cgroup *h_cg,
880 struct hstate *h)
881{
882#ifdef CONFIG_CGROUP_HUGETLB
883 if (!h_cg || !h) {
884 resv_map->reservation_counter = NULL;
885 resv_map->pages_per_hpage = 0;
886 resv_map->css = NULL;
887 } else {
888 resv_map->reservation_counter =
889 &h_cg->rsvd_hugepage[hstate_index(h)];
890 resv_map->pages_per_hpage = pages_per_huge_page(h);
891 resv_map->css = &h_cg->css;
892 }
893#endif
894}
895
Joonsoo Kim9119a412014-04-03 14:47:25 -0700896struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700897{
898 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700899 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
900
901 if (!resv_map || !rg) {
902 kfree(resv_map);
903 kfree(rg);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700904 return NULL;
Mike Kravetz5e911372015-09-08 15:01:28 -0700905 }
Andy Whitcroft84afd992008-07-23 21:27:32 -0700906
907 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700908 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700909 INIT_LIST_HEAD(&resv_map->regions);
910
Mike Kravetz5e911372015-09-08 15:01:28 -0700911 resv_map->adds_in_progress = 0;
Mina Almasrye9fe92a2020-04-01 21:11:21 -0700912 /*
913 * Initialize these to 0. On shared mappings, 0's here indicate these
914 * fields don't do cgroup accounting. On private mappings, these will be
915 * re-initialized to the proper values, to indicate that hugetlb cgroup
916 * reservations are to be un-charged from here.
917 */
918 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700919
920 INIT_LIST_HEAD(&resv_map->region_cache);
921 list_add(&rg->link, &resv_map->region_cache);
922 resv_map->region_cache_count = 1;
923
Andy Whitcroft84afd992008-07-23 21:27:32 -0700924 return resv_map;
925}
926
Joonsoo Kim9119a412014-04-03 14:47:25 -0700927void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700928{
929 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
Mike Kravetz5e911372015-09-08 15:01:28 -0700930 struct list_head *head = &resv_map->region_cache;
931 struct file_region *rg, *trg;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700932
933 /* Clear out any active regions before we release the map. */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700934 region_del(resv_map, 0, LONG_MAX);
Mike Kravetz5e911372015-09-08 15:01:28 -0700935
936 /* ... and any entries left in the cache */
937 list_for_each_entry_safe(rg, trg, head, link) {
938 list_del(&rg->link);
939 kfree(rg);
940 }
941
942 VM_BUG_ON(resv_map->adds_in_progress);
943
Andy Whitcroft84afd992008-07-23 21:27:32 -0700944 kfree(resv_map);
945}
946
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700947static inline struct resv_map *inode_resv_map(struct inode *inode)
948{
Mike Kravetzf27a5132019-05-13 17:22:55 -0700949 /*
950 * At inode evict time, i_mapping may not point to the original
951 * address space within the inode. This original address space
952 * contains the pointer to the resv_map. So, always use the
953 * address space embedded within the inode.
954 * The VERY common case is inode->mapping == &inode->i_data but,
955 * this may not be true for device special inodes.
956 */
957 return (struct resv_map *)(&inode->i_data)->private_data;
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700958}
959
Andy Whitcroft84afd992008-07-23 21:27:32 -0700960static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700961{
Sasha Levin81d1b092014-10-09 15:28:10 -0700962 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700963 if (vma->vm_flags & VM_MAYSHARE) {
964 struct address_space *mapping = vma->vm_file->f_mapping;
965 struct inode *inode = mapping->host;
966
967 return inode_resv_map(inode);
968
969 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700970 return (struct resv_map *)(get_vma_private_data(vma) &
971 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700972 }
Mel Gormana1e78772008-07-23 21:27:23 -0700973}
974
Andy Whitcroft84afd992008-07-23 21:27:32 -0700975static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700976{
Sasha Levin81d1b092014-10-09 15:28:10 -0700977 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
978 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700979
Andy Whitcroft84afd992008-07-23 21:27:32 -0700980 set_vma_private_data(vma, (get_vma_private_data(vma) &
981 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700982}
983
984static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
985{
Sasha Levin81d1b092014-10-09 15:28:10 -0700986 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
987 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700988
989 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700990}
991
992static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
993{
Sasha Levin81d1b092014-10-09 15:28:10 -0700994 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700995
996 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700997}
998
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700999/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -07001000void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
1001{
Sasha Levin81d1b092014-10-09 15:28:10 -07001002 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -07001003 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -07001004 vma->vm_private_data = (void *)0;
1005}
1006
1007/* Returns true if the VMA has associated reserve pages */
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001008static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -07001009{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001010 if (vma->vm_flags & VM_NORESERVE) {
1011 /*
1012 * This address is already reserved by other process(chg == 0),
1013 * so, we should decrement reserved count. Without decrementing,
1014 * reserve count remains after releasing inode, because this
1015 * allocated page will go into page cache and is regarded as
1016 * coming from reserved pool in releasing step. Currently, we
1017 * don't have any other solution to deal with this situation
1018 * properly, so add work-around here.
1019 */
1020 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001021 return true;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001022 else
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001023 return false;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001024 }
Joonsoo Kima63884e2013-09-11 14:21:07 -07001025
1026 /* Shared mappings always use reserves */
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -07001027 if (vma->vm_flags & VM_MAYSHARE) {
1028 /*
1029 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1030 * be a region map for all pages. The only situation where
1031 * there is no region map is if a hole was punched via
Ethon Paul7c8de352020-06-04 16:49:07 -07001032 * fallocate. In this case, there really are no reserves to
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -07001033 * use. This situation is indicated if chg != 0.
1034 */
1035 if (chg)
1036 return false;
1037 else
1038 return true;
1039 }
Joonsoo Kima63884e2013-09-11 14:21:07 -07001040
1041 /*
1042 * Only the process that called mmap() has reserves for
1043 * private mappings.
1044 */
Mike Kravetz67961f92016-06-08 15:33:42 -07001045 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1046 /*
1047 * Like the shared case above, a hole punch or truncate
1048 * could have been performed on the private mapping.
1049 * Examine the value of chg to determine if reserves
1050 * actually exist or were previously consumed.
1051 * Very Subtle - The value of chg comes from a previous
1052 * call to vma_needs_reserves(). The reserve map for
1053 * private mappings has different (opposite) semantics
1054 * than that of shared mappings. vma_needs_reserves()
1055 * has already taken this difference in semantics into
1056 * account. Therefore, the meaning of chg is the same
1057 * as in the shared case above. Code could easily be
1058 * combined, but keeping it separate draws attention to
1059 * subtle differences.
1060 */
1061 if (chg)
1062 return false;
1063 else
1064 return true;
1065 }
Joonsoo Kima63884e2013-09-11 14:21:07 -07001066
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001067 return false;
Mel Gormana1e78772008-07-23 21:27:23 -07001068}
1069
Andi Kleena5516432008-07-23 21:27:41 -07001070static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071{
1072 int nid = page_to_nid(page);
Mike Kravetz9487ca62021-05-04 18:35:10 -07001073
1074 lockdep_assert_held(&hugetlb_lock);
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07001075 VM_BUG_ON_PAGE(page_count(page), page);
1076
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001077 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -07001078 h->free_huge_pages++;
1079 h->free_huge_pages_node[nid]++;
Mike Kravetz6c037142021-02-24 12:09:04 -08001080 SetHPageFreed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081}
1082
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001083static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001084{
1085 struct page *page;
Pavel Tatashin1a08ae32021-05-04 18:38:53 -07001086 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001087
Mike Kravetz9487ca62021-05-04 18:35:10 -07001088 lockdep_assert_held(&hugetlb_lock);
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001089 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
Pavel Tatashin8e3560d2021-05-04 18:39:00 -07001090 if (pin && !is_pinnable_page(page))
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001091 continue;
1092
Wei Yang6664bfc2020-10-13 16:56:39 -07001093 if (PageHWPoison(page))
1094 continue;
1095
1096 list_move(&page->lru, &h->hugepage_activelist);
1097 set_page_refcounted(page);
Mike Kravetz6c037142021-02-24 12:09:04 -08001098 ClearHPageFreed(page);
Wei Yang6664bfc2020-10-13 16:56:39 -07001099 h->free_huge_pages--;
1100 h->free_huge_pages_node[nid]--;
1101 return page;
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001102 }
1103
Wei Yang6664bfc2020-10-13 16:56:39 -07001104 return NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001105}
1106
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001107static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1108 nodemask_t *nmask)
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001109{
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001110 unsigned int cpuset_mems_cookie;
1111 struct zonelist *zonelist;
1112 struct zone *zone;
1113 struct zoneref *z;
Anshuman Khandual98fa15f2019-03-05 15:42:58 -08001114 int node = NUMA_NO_NODE;
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001115
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001116 zonelist = node_zonelist(nid, gfp_mask);
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001117
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001118retry_cpuset:
1119 cpuset_mems_cookie = read_mems_allowed_begin();
1120 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1121 struct page *page;
1122
1123 if (!cpuset_zone_allowed(zone, gfp_mask))
1124 continue;
1125 /*
1126 * no need to ask again on the same node. Pool is node rather than
1127 * zone aware
1128 */
1129 if (zone_to_nid(zone) == node)
1130 continue;
1131 node = zone_to_nid(zone);
1132
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001133 page = dequeue_huge_page_node_exact(h, node);
1134 if (page)
1135 return page;
1136 }
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001137 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1138 goto retry_cpuset;
1139
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001140 return NULL;
1141}
1142
Andi Kleena5516432008-07-23 21:27:41 -07001143static struct page *dequeue_huge_page_vma(struct hstate *h,
1144 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001145 unsigned long address, int avoid_reserve,
1146 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147{
Ben Widawskycfcaa662021-09-02 15:00:13 -07001148 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001149 struct mempolicy *mpol;
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001150 gfp_t gfp_mask;
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001151 nodemask_t *nodemask;
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001152 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153
Mel Gormana1e78772008-07-23 21:27:23 -07001154 /*
1155 * A child process with MAP_PRIVATE mappings created by their parent
1156 * have no page reserves. This check ensures that reservations are
1157 * not "stolen". The child may still get SIGKILLed
1158 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001159 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -07001160 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -07001161 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -07001162
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001163 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -07001164 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -07001165 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001166
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001167 gfp_mask = htlb_alloc_mask(h);
1168 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
Ben Widawskycfcaa662021-09-02 15:00:13 -07001169
1170 if (mpol_is_preferred_many(mpol)) {
1171 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1172
1173 /* Fallback to all nodes if page==NULL */
1174 nodemask = NULL;
1175 }
1176
1177 if (!page)
1178 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1179
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001180 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
Mike Kravetzd6995da2021-02-24 12:08:51 -08001181 SetHPageRestoreReserve(page);
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001182 h->resv_huge_pages--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07001184
1185 mpol_cond_put(mpol);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001186 return page;
1187
Miao Xiec0ff7452010-05-24 14:32:08 -07001188err:
Mel Gormancc9a6c82012-03-21 16:34:11 -07001189 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190}
1191
Luiz Capitulino1cac6f22014-06-04 16:07:11 -07001192/*
1193 * common helper functions for hstate_next_node_to_{alloc|free}.
1194 * We may have allocated or freed a huge page based on a different
1195 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1196 * be outside of *nodes_allowed. Ensure that we use an allowed
1197 * node for alloc or free.
1198 */
1199static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1200{
Andrew Morton0edaf862016-05-19 17:10:58 -07001201 nid = next_node_in(nid, *nodes_allowed);
Luiz Capitulino1cac6f22014-06-04 16:07:11 -07001202 VM_BUG_ON(nid >= MAX_NUMNODES);
1203
1204 return nid;
1205}
1206
1207static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1208{
1209 if (!node_isset(nid, *nodes_allowed))
1210 nid = next_node_allowed(nid, nodes_allowed);
1211 return nid;
1212}
1213
1214/*
1215 * returns the previously saved node ["this node"] from which to
1216 * allocate a persistent huge page for the pool and advance the
1217 * next node from which to allocate, handling wrap at end of node
1218 * mask.
1219 */
1220static int hstate_next_node_to_alloc(struct hstate *h,
1221 nodemask_t *nodes_allowed)
1222{
1223 int nid;
1224
1225 VM_BUG_ON(!nodes_allowed);
1226
1227 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1228 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1229
1230 return nid;
1231}
1232
1233/*
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001234 * helper for remove_pool_huge_page() - return the previously saved
Luiz Capitulino1cac6f22014-06-04 16:07:11 -07001235 * node ["this node"] from which to free a huge page. Advance the
1236 * next node id whether or not we find a free huge page to free so
1237 * that the next attempt to free addresses the next node.
1238 */
1239static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1240{
1241 int nid;
1242
1243 VM_BUG_ON(!nodes_allowed);
1244
1245 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1246 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1247
1248 return nid;
1249}
1250
1251#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1252 for (nr_nodes = nodes_weight(*mask); \
1253 nr_nodes > 0 && \
1254 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1255 nr_nodes--)
1256
1257#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1258 for (nr_nodes = nodes_weight(*mask); \
1259 nr_nodes > 0 && \
1260 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1261 nr_nodes--)
1262
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001263#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001264static void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001265 unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001266{
1267 int i;
1268 int nr_pages = 1 << order;
1269 struct page *p = page + 1;
1270
Gerald Schaeferc8cc7082016-06-24 14:50:04 -07001271 atomic_set(compound_mapcount_ptr(page), 0);
Yanfei Xu5291c092021-02-24 12:07:22 -08001272 atomic_set(compound_pincount_ptr(page), 0);
John Hubbard47e29d32020-04-01 21:05:33 -07001273
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001274 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001275 clear_compound_head(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001276 set_page_refcounted(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001277 }
1278
1279 set_compound_order(page, 0);
Gerald Schaeferba9c1202020-12-11 13:36:53 -08001280 page[1].compound_nr = 0;
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001281 __ClearPageHead(page);
1282}
1283
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001284static void free_gigantic_page(struct page *page, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001285{
Roman Gushchincf11e852020-04-10 14:32:45 -07001286 /*
1287 * If the page isn't allocated using the cma allocator,
1288 * cma_release() returns false.
1289 */
Barry Songdbda8fe2020-07-23 21:15:30 -07001290#ifdef CONFIG_CMA
1291 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
Roman Gushchincf11e852020-04-10 14:32:45 -07001292 return;
Barry Songdbda8fe2020-07-23 21:15:30 -07001293#endif
Roman Gushchincf11e852020-04-10 14:32:45 -07001294
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001295 free_contig_range(page_to_pfn(page), 1 << order);
1296}
1297
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001298#ifdef CONFIG_CONTIG_ALLOC
Michal Hockod9cc948f2018-01-31 16:20:44 -08001299static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1300 int nid, nodemask_t *nodemask)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001301{
Miaohe Lin04adbc32021-05-04 18:33:22 -07001302 unsigned long nr_pages = pages_per_huge_page(h);
Li Xinhai953f0642020-09-04 16:36:10 -07001303 if (nid == NUMA_NO_NODE)
1304 nid = numa_mem_id();
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001305
Barry Songdbda8fe2020-07-23 21:15:30 -07001306#ifdef CONFIG_CMA
1307 {
Roman Gushchincf11e852020-04-10 14:32:45 -07001308 struct page *page;
1309 int node;
1310
Li Xinhai953f0642020-09-04 16:36:10 -07001311 if (hugetlb_cma[nid]) {
1312 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1313 huge_page_order(h), true);
Roman Gushchincf11e852020-04-10 14:32:45 -07001314 if (page)
1315 return page;
1316 }
Li Xinhai953f0642020-09-04 16:36:10 -07001317
1318 if (!(gfp_mask & __GFP_THISNODE)) {
1319 for_each_node_mask(node, *nodemask) {
1320 if (node == nid || !hugetlb_cma[node])
1321 continue;
1322
1323 page = cma_alloc(hugetlb_cma[node], nr_pages,
1324 huge_page_order(h), true);
1325 if (page)
1326 return page;
1327 }
1328 }
Roman Gushchincf11e852020-04-10 14:32:45 -07001329 }
Barry Songdbda8fe2020-07-23 21:15:30 -07001330#endif
Roman Gushchincf11e852020-04-10 14:32:45 -07001331
Anshuman Khandual5e27a2d2019-11-30 17:55:06 -08001332 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001333}
1334
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001335#else /* !CONFIG_CONTIG_ALLOC */
1336static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1337 int nid, nodemask_t *nodemask)
1338{
1339 return NULL;
1340}
1341#endif /* CONFIG_CONTIG_ALLOC */
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001342
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001343#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
Michal Hockod9cc948f2018-01-31 16:20:44 -08001344static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001345 int nid, nodemask_t *nodemask)
1346{
1347 return NULL;
1348}
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001349static inline void free_gigantic_page(struct page *page, unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001350static inline void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001351 unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001352#endif
1353
Mike Kravetz6eb4e882021-05-04 18:34:55 -07001354/*
1355 * Remove hugetlb page from lists, and update dtor so that page appears
1356 * as just a compound page. A reference is held on the page.
1357 *
1358 * Must be called with hugetlb lock held.
1359 */
1360static void remove_hugetlb_page(struct hstate *h, struct page *page,
1361 bool adjust_surplus)
1362{
1363 int nid = page_to_nid(page);
1364
1365 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1366 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1367
Mike Kravetz9487ca62021-05-04 18:35:10 -07001368 lockdep_assert_held(&hugetlb_lock);
Mike Kravetz6eb4e882021-05-04 18:34:55 -07001369 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1370 return;
1371
1372 list_del(&page->lru);
1373
1374 if (HPageFreed(page)) {
1375 h->free_huge_pages--;
1376 h->free_huge_pages_node[nid]--;
1377 }
1378 if (adjust_surplus) {
1379 h->surplus_huge_pages--;
1380 h->surplus_huge_pages_node[nid]--;
1381 }
1382
Mike Kravetze32d20c2021-09-02 14:58:50 -07001383 /*
1384 * Very subtle
1385 *
1386 * For non-gigantic pages set the destructor to the normal compound
1387 * page dtor. This is needed in case someone takes an additional
1388 * temporary ref to the page, and freeing is delayed until they drop
1389 * their reference.
1390 *
1391 * For gigantic pages set the destructor to the null dtor. This
1392 * destructor will never be called. Before freeing the gigantic
1393 * page destroy_compound_gigantic_page will turn the compound page
1394 * into a simple group of pages. After this the destructor does not
1395 * apply.
1396 *
1397 * This handles the case where more than one ref is held when and
1398 * after update_and_free_page is called.
1399 */
Mike Kravetz6eb4e882021-05-04 18:34:55 -07001400 set_page_refcounted(page);
Mike Kravetze32d20c2021-09-02 14:58:50 -07001401 if (hstate_is_gigantic(h))
1402 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1403 else
1404 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
Mike Kravetz6eb4e882021-05-04 18:34:55 -07001405
1406 h->nr_huge_pages--;
1407 h->nr_huge_pages_node[nid]--;
1408}
1409
Muchun Songad2fa372021-06-30 18:47:21 -07001410static void add_hugetlb_page(struct hstate *h, struct page *page,
1411 bool adjust_surplus)
1412{
1413 int zeroed;
1414 int nid = page_to_nid(page);
1415
1416 VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1417
1418 lockdep_assert_held(&hugetlb_lock);
1419
1420 INIT_LIST_HEAD(&page->lru);
1421 h->nr_huge_pages++;
1422 h->nr_huge_pages_node[nid]++;
1423
1424 if (adjust_surplus) {
1425 h->surplus_huge_pages++;
1426 h->surplus_huge_pages_node[nid]++;
1427 }
1428
1429 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1430 set_page_private(page, 0);
1431 SetHPageVmemmapOptimized(page);
1432
1433 /*
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07001434 * This page is about to be managed by the hugetlb allocator and
1435 * should have no users. Drop our reference, and check for others
1436 * just in case.
Muchun Songad2fa372021-06-30 18:47:21 -07001437 */
1438 zeroed = put_page_testzero(page);
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07001439 if (!zeroed)
1440 /*
1441 * It is VERY unlikely soneone else has taken a ref on
1442 * the page. In this case, we simply return as the
1443 * hugetlb destructor (free_huge_page) will be called
1444 * when this other ref is dropped.
1445 */
1446 return;
1447
Muchun Songad2fa372021-06-30 18:47:21 -07001448 arch_clear_hugepage_flags(page);
1449 enqueue_huge_page(h, page);
1450}
1451
Muchun Songb65d4ad2021-06-30 18:47:17 -07001452static void __update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -07001453{
1454 int i;
Mike Kravetzdbfee5a2021-02-24 12:07:50 -08001455 struct page *subpage = page;
Andi Kleena5516432008-07-23 21:27:41 -07001456
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001457 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001458 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -08001459
Muchun Songad2fa372021-06-30 18:47:21 -07001460 if (alloc_huge_page_vmemmap(h, page)) {
1461 spin_lock_irq(&hugetlb_lock);
1462 /*
1463 * If we cannot allocate vmemmap pages, just refuse to free the
1464 * page and put the page back on the hugetlb free list and treat
1465 * as a surplus page.
1466 */
1467 add_hugetlb_page(h, page, true);
1468 spin_unlock_irq(&hugetlb_lock);
1469 return;
1470 }
1471
Mike Kravetzdbfee5a2021-02-24 12:07:50 -08001472 for (i = 0; i < pages_per_huge_page(h);
1473 i++, subpage = mem_map_next(subpage, page, i)) {
1474 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
Chris Forbes32f84522011-07-25 17:12:14 -07001475 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -07001476 1 << PG_active | 1 << PG_private |
1477 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -07001478 }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001479 if (hstate_is_gigantic(h)) {
1480 destroy_compound_gigantic_page(page, huge_page_order(h));
1481 free_gigantic_page(page, huge_page_order(h));
1482 } else {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001483 __free_pages(page, huge_page_order(h));
1484 }
Adam Litke6af2acb2007-10-16 01:26:16 -07001485}
1486
Muchun Songb65d4ad2021-06-30 18:47:17 -07001487/*
1488 * As update_and_free_page() can be called under any context, so we cannot
1489 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1490 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1491 * the vmemmap pages.
1492 *
1493 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1494 * freed and frees them one-by-one. As the page->mapping pointer is going
1495 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1496 * structure of a lockless linked list of huge pages to be freed.
1497 */
1498static LLIST_HEAD(hpage_freelist);
1499
1500static void free_hpage_workfn(struct work_struct *work)
1501{
1502 struct llist_node *node;
1503
1504 node = llist_del_all(&hpage_freelist);
1505
1506 while (node) {
1507 struct page *page;
1508 struct hstate *h;
1509
1510 page = container_of((struct address_space **)node,
1511 struct page, mapping);
1512 node = node->next;
1513 page->mapping = NULL;
1514 /*
1515 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1516 * is going to trigger because a previous call to
1517 * remove_hugetlb_page() will set_compound_page_dtor(page,
1518 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1519 */
1520 h = size_to_hstate(page_size(page));
1521
1522 __update_and_free_page(h, page);
1523
1524 cond_resched();
1525 }
1526}
1527static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1528
1529static inline void flush_free_hpage_work(struct hstate *h)
1530{
1531 if (free_vmemmap_pages_per_hpage(h))
1532 flush_work(&free_hpage_work);
1533}
1534
1535static void update_and_free_page(struct hstate *h, struct page *page,
1536 bool atomic)
1537{
Muchun Songad2fa372021-06-30 18:47:21 -07001538 if (!HPageVmemmapOptimized(page) || !atomic) {
Muchun Songb65d4ad2021-06-30 18:47:17 -07001539 __update_and_free_page(h, page);
1540 return;
1541 }
1542
1543 /*
1544 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1545 *
1546 * Only call schedule_work() if hpage_freelist is previously
1547 * empty. Otherwise, schedule_work() had been called but the workfn
1548 * hasn't retrieved the list yet.
1549 */
1550 if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1551 schedule_work(&free_hpage_work);
1552}
1553
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001554static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1555{
1556 struct page *page, *t_page;
1557
1558 list_for_each_entry_safe(page, t_page, list, lru) {
Muchun Songb65d4ad2021-06-30 18:47:17 -07001559 update_and_free_page(h, page, false);
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001560 cond_resched();
1561 }
1562}
1563
Andi Kleene5ff2152008-07-23 21:27:42 -07001564struct hstate *size_to_hstate(unsigned long size)
1565{
1566 struct hstate *h;
1567
1568 for_each_hstate(h) {
1569 if (huge_page_size(h) == size)
1570 return h;
1571 }
1572 return NULL;
1573}
1574
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001575void free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -08001576{
Andi Kleena5516432008-07-23 21:27:41 -07001577 /*
1578 * Can't pass hstate in here because it is called from the
1579 * compound page destructor.
1580 */
Andi Kleene5ff2152008-07-23 21:27:42 -07001581 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001582 int nid = page_to_nid(page);
Mike Kravetzd6995da2021-02-24 12:08:51 -08001583 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001584 bool restore_reserve;
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001585 unsigned long flags;
David Gibson27a85ef2006-03-22 00:08:56 -08001586
Mike Kravetzb4330af2016-02-05 15:36:38 -08001587 VM_BUG_ON_PAGE(page_count(page), page);
1588 VM_BUG_ON_PAGE(page_mapcount(page), page);
Yongkai Wu8ace22b2018-12-14 14:17:10 -08001589
Mike Kravetzd6995da2021-02-24 12:08:51 -08001590 hugetlb_set_page_subpool(page, NULL);
Yongkai Wu8ace22b2018-12-14 14:17:10 -08001591 page->mapping = NULL;
Mike Kravetzd6995da2021-02-24 12:08:51 -08001592 restore_reserve = HPageRestoreReserve(page);
1593 ClearHPageRestoreReserve(page);
David Gibson27a85ef2006-03-22 00:08:56 -08001594
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001595 /*
Mike Kravetzd6995da2021-02-24 12:08:51 -08001596 * If HPageRestoreReserve was set on page, page allocation consumed a
Mike Kravetz0919e1b2019-05-13 17:19:38 -07001597 * reservation. If the page was associated with a subpool, there
1598 * would have been a page reserved in the subpool before allocation
1599 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
Miaohe Lin6c26d312021-02-24 12:07:19 -08001600 * reservation, do not call hugepage_subpool_put_pages() as this will
Mike Kravetz0919e1b2019-05-13 17:19:38 -07001601 * remove the reserved page from the subpool.
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001602 */
Mike Kravetz0919e1b2019-05-13 17:19:38 -07001603 if (!restore_reserve) {
1604 /*
1605 * A return code of zero implies that the subpool will be
1606 * under its minimum size if the reservation is not restored
1607 * after page is free. Therefore, force restore_reserve
1608 * operation.
1609 */
1610 if (hugepage_subpool_put_pages(spool, 1) == 0)
1611 restore_reserve = true;
1612 }
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001613
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001614 spin_lock_irqsave(&hugetlb_lock, flags);
Mike Kravetz8f251a32021-02-24 12:08:56 -08001615 ClearHPageMigratable(page);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001616 hugetlb_cgroup_uncharge_page(hstate_index(h),
1617 pages_per_huge_page(h), page);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07001618 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1619 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001620 if (restore_reserve)
1621 h->resv_huge_pages++;
1622
Mike Kravetz9157c3112021-02-24 12:09:00 -08001623 if (HPageTemporary(page)) {
Mike Kravetz6eb4e882021-05-04 18:34:55 -07001624 remove_hugetlb_page(h, page, false);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001625 spin_unlock_irqrestore(&hugetlb_lock, flags);
Muchun Songb65d4ad2021-06-30 18:47:17 -07001626 update_and_free_page(h, page, true);
Michal Hockoab5ac902018-01-31 16:20:48 -08001627 } else if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001628 /* remove the page from active list */
Mike Kravetz6eb4e882021-05-04 18:34:55 -07001629 remove_hugetlb_page(h, page, true);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001630 spin_unlock_irqrestore(&hugetlb_lock, flags);
Muchun Songb65d4ad2021-06-30 18:47:17 -07001631 update_and_free_page(h, page, true);
Adam Litke7893d1d2007-10-16 01:26:18 -07001632 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -07001633 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -07001634 enqueue_huge_page(h, page);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001635 spin_unlock_irqrestore(&hugetlb_lock, flags);
Adam Litke7893d1d2007-10-16 01:26:18 -07001636 }
David Gibson27a85ef2006-03-22 00:08:56 -08001637}
1638
Oscar Salvadord3d99fc2021-05-04 18:35:23 -07001639/*
1640 * Must be called with the hugetlb lock held
1641 */
1642static void __prep_account_new_huge_page(struct hstate *h, int nid)
1643{
1644 lockdep_assert_held(&hugetlb_lock);
1645 h->nr_huge_pages++;
1646 h->nr_huge_pages_node[nid]++;
1647}
1648
Muchun Songf41f2ed2021-06-30 18:47:13 -07001649static void __prep_new_huge_page(struct hstate *h, struct page *page)
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001650{
Muchun Songf41f2ed2021-06-30 18:47:13 -07001651 free_huge_page_vmemmap(h, page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001652 INIT_LIST_HEAD(&page->lru);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001653 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Mike Kravetzff546112021-02-24 12:09:11 -08001654 hugetlb_set_page_subpool(page, NULL);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001655 set_hugetlb_cgroup(page, NULL);
Mina Almasry1adc4d42020-04-01 21:11:15 -07001656 set_hugetlb_cgroup_rsvd(page, NULL);
Oscar Salvadord3d99fc2021-05-04 18:35:23 -07001657}
1658
1659static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1660{
Muchun Songf41f2ed2021-06-30 18:47:13 -07001661 __prep_new_huge_page(h, page);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001662 spin_lock_irq(&hugetlb_lock);
Oscar Salvadord3d99fc2021-05-04 18:35:23 -07001663 __prep_account_new_huge_page(h, nid);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001664 spin_unlock_irq(&hugetlb_lock);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001665}
1666
Mike Kravetz7118fc22021-06-30 18:48:34 -07001667static bool prep_compound_gigantic_page(struct page *page, unsigned int order)
Wu Fengguang20a03072009-06-16 15:32:22 -07001668{
Mike Kravetz7118fc22021-06-30 18:48:34 -07001669 int i, j;
Wu Fengguang20a03072009-06-16 15:32:22 -07001670 int nr_pages = 1 << order;
1671 struct page *p = page + 1;
1672
1673 /* we rely on prep_new_huge_page to set the destructor */
1674 set_compound_order(page, order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001675 __ClearPageReserved(page);
Kirill A. Shutemovde09d312016-01-15 16:51:42 -08001676 __SetPageHead(page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001677 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001678 /*
1679 * For gigantic hugepages allocated through bootmem at
1680 * boot, it's safer to be consistent with the not-gigantic
1681 * hugepages and clear the PG_reserved bit from all tail pages
Ethon Paul7c8de352020-06-04 16:49:07 -07001682 * too. Otherwise drivers using get_user_pages() to access tail
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001683 * pages may get the reference counting wrong if they see
1684 * PG_reserved set on a tail page (despite the head page not
1685 * having PG_reserved set). Enforcing this consistency between
1686 * head and tail pages allows drivers to optimize away a check
1687 * on the head page when they need know if put_page() is needed
1688 * after get_user_pages().
1689 */
1690 __ClearPageReserved(p);
Mike Kravetz7118fc22021-06-30 18:48:34 -07001691 /*
1692 * Subtle and very unlikely
1693 *
1694 * Gigantic 'page allocators' such as memblock or cma will
1695 * return a set of pages with each page ref counted. We need
1696 * to turn this set of pages into a compound page with tail
1697 * page ref counts set to zero. Code such as speculative page
1698 * cache adding could take a ref on a 'to be' tail page.
1699 * We need to respect any increased ref count, and only set
1700 * the ref count to zero if count is currently 1. If count
Mike Kravetz416d85e2021-09-02 14:58:43 -07001701 * is not 1, we return an error. An error return indicates
1702 * the set of pages can not be converted to a gigantic page.
1703 * The caller who allocated the pages should then discard the
1704 * pages using the appropriate free interface.
Mike Kravetz7118fc22021-06-30 18:48:34 -07001705 */
1706 if (!page_ref_freeze(p, 1)) {
Mike Kravetz416d85e2021-09-02 14:58:43 -07001707 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1708 goto out_error;
Mike Kravetz7118fc22021-06-30 18:48:34 -07001709 }
Youquan Song58a84aa2011-12-08 14:34:18 -08001710 set_page_count(p, 0);
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001711 set_compound_head(p, page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001712 }
Mike Kravetzb4330af2016-02-05 15:36:38 -08001713 atomic_set(compound_mapcount_ptr(page), -1);
Yanfei Xu5291c092021-02-24 12:07:22 -08001714 atomic_set(compound_pincount_ptr(page), 0);
Mike Kravetz7118fc22021-06-30 18:48:34 -07001715 return true;
1716
1717out_error:
1718 /* undo tail page modifications made above */
1719 p = page + 1;
1720 for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) {
1721 clear_compound_head(p);
1722 set_page_refcounted(p);
1723 }
1724 /* need to clear PG_reserved on remaining tail pages */
1725 for (; j < nr_pages; j++, p = mem_map_next(p, page, j))
1726 __ClearPageReserved(p);
1727 set_compound_order(page, 0);
1728 page[1].compound_nr = 0;
1729 __ClearPageHead(page);
1730 return false;
Wu Fengguang20a03072009-06-16 15:32:22 -07001731}
1732
Andrew Morton77959122012-10-08 16:34:11 -07001733/*
1734 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1735 * transparent huge pages. See the PageTransHuge() documentation for more
1736 * details.
1737 */
Wu Fengguang20a03072009-06-16 15:32:22 -07001738int PageHuge(struct page *page)
1739{
Wu Fengguang20a03072009-06-16 15:32:22 -07001740 if (!PageCompound(page))
1741 return 0;
1742
1743 page = compound_head(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001744 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
Wu Fengguang20a03072009-06-16 15:32:22 -07001745}
Naoya Horiguchi43131e12010-05-28 09:29:22 +09001746EXPORT_SYMBOL_GPL(PageHuge);
1747
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001748/*
1749 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1750 * normal or transparent huge pages.
1751 */
1752int PageHeadHuge(struct page *page_head)
1753{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001754 if (!PageHead(page_head))
1755 return 0;
1756
Vlastimil Babkad4af73e2020-04-01 21:11:48 -07001757 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001758}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001759
Mike Kravetzc0d03812020-04-01 21:11:05 -07001760/*
Mike Kravetzc0d03812020-04-01 21:11:05 -07001761 * Find and lock address space (mapping) in write mode.
1762 *
Mike Kravetz336bf302020-11-13 22:52:16 -08001763 * Upon entry, the page is locked which means that page_mapping() is
1764 * stable. Due to locking order, we can only trylock_write. If we can
1765 * not get the lock, simply return NULL to caller.
Mike Kravetzc0d03812020-04-01 21:11:05 -07001766 */
1767struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1768{
Mike Kravetz336bf302020-11-13 22:52:16 -08001769 struct address_space *mapping = page_mapping(hpage);
Mike Kravetzc0d03812020-04-01 21:11:05 -07001770
Mike Kravetzc0d03812020-04-01 21:11:05 -07001771 if (!mapping)
1772 return mapping;
1773
Mike Kravetzc0d03812020-04-01 21:11:05 -07001774 if (i_mmap_trylock_write(mapping))
1775 return mapping;
1776
Mike Kravetz336bf302020-11-13 22:52:16 -08001777 return NULL;
Mike Kravetzc0d03812020-04-01 21:11:05 -07001778}
1779
Hugh Dickinsfe19bd32021-06-24 18:39:52 -07001780pgoff_t hugetlb_basepage_index(struct page *page)
Zhang Yi13d60f42013-06-25 21:19:31 +08001781{
1782 struct page *page_head = compound_head(page);
1783 pgoff_t index = page_index(page_head);
1784 unsigned long compound_idx;
1785
Zhang Yi13d60f42013-06-25 21:19:31 +08001786 if (compound_order(page_head) >= MAX_ORDER)
1787 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1788 else
1789 compound_idx = page - page_head;
1790
1791 return (index << compound_order(page_head)) + compound_idx;
1792}
1793
Michal Hocko0c397da2018-01-31 16:20:56 -08001794static struct page *alloc_buddy_huge_page(struct hstate *h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001795 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1796 nodemask_t *node_alloc_noretry)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797{
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001798 int order = huge_page_order(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 struct page *page;
Mike Kravetzf60858f2019-09-23 15:37:35 -07001800 bool alloc_try_hard = true;
Joe Jinf96efd52007-07-15 23:38:12 -07001801
Mike Kravetzf60858f2019-09-23 15:37:35 -07001802 /*
1803 * By default we always try hard to allocate the page with
1804 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1805 * a loop (to adjust global huge page counts) and previous allocation
1806 * failed, do not continue to try hard on the same node. Use the
1807 * node_alloc_noretry bitmap to manage this state information.
1808 */
1809 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1810 alloc_try_hard = false;
1811 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1812 if (alloc_try_hard)
1813 gfp_mask |= __GFP_RETRY_MAYFAIL;
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001814 if (nid == NUMA_NO_NODE)
1815 nid = numa_mem_id();
Matthew Wilcox (Oracle)84172f42021-04-29 23:01:15 -07001816 page = __alloc_pages(gfp_mask, order, nid, nmask);
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001817 if (page)
1818 __count_vm_event(HTLB_BUDDY_PGALLOC);
1819 else
1820 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001821
Mike Kravetzf60858f2019-09-23 15:37:35 -07001822 /*
1823 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1824 * indicates an overall state change. Clear bit so that we resume
1825 * normal 'try hard' allocations.
1826 */
1827 if (node_alloc_noretry && page && !alloc_try_hard)
1828 node_clear(nid, *node_alloc_noretry);
1829
1830 /*
1831 * If we tried hard to get a page but failed, set bit so that
1832 * subsequent attempts will not try as hard until there is an
1833 * overall state change.
1834 */
1835 if (node_alloc_noretry && !page && alloc_try_hard)
1836 node_set(nid, *node_alloc_noretry);
1837
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001838 return page;
1839}
1840
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001841/*
Michal Hocko0c397da2018-01-31 16:20:56 -08001842 * Common helper to allocate a fresh hugetlb page. All specific allocators
1843 * should use this function to get new hugetlb pages
1844 */
1845static struct page *alloc_fresh_huge_page(struct hstate *h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001846 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1847 nodemask_t *node_alloc_noretry)
Michal Hocko0c397da2018-01-31 16:20:56 -08001848{
1849 struct page *page;
Mike Kravetz7118fc22021-06-30 18:48:34 -07001850 bool retry = false;
Michal Hocko0c397da2018-01-31 16:20:56 -08001851
Mike Kravetz7118fc22021-06-30 18:48:34 -07001852retry:
Michal Hocko0c397da2018-01-31 16:20:56 -08001853 if (hstate_is_gigantic(h))
1854 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1855 else
1856 page = alloc_buddy_huge_page(h, gfp_mask,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001857 nid, nmask, node_alloc_noretry);
Michal Hocko0c397da2018-01-31 16:20:56 -08001858 if (!page)
1859 return NULL;
1860
Mike Kravetz7118fc22021-06-30 18:48:34 -07001861 if (hstate_is_gigantic(h)) {
1862 if (!prep_compound_gigantic_page(page, huge_page_order(h))) {
1863 /*
1864 * Rare failure to convert pages to compound page.
1865 * Free pages and try again - ONCE!
1866 */
1867 free_gigantic_page(page, huge_page_order(h));
1868 if (!retry) {
1869 retry = true;
1870 goto retry;
1871 }
Mike Kravetz7118fc22021-06-30 18:48:34 -07001872 return NULL;
1873 }
1874 }
Michal Hocko0c397da2018-01-31 16:20:56 -08001875 prep_new_huge_page(h, page, page_to_nid(page));
1876
1877 return page;
1878}
1879
1880/*
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001881 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1882 * manner.
1883 */
Mike Kravetzf60858f2019-09-23 15:37:35 -07001884static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1885 nodemask_t *node_alloc_noretry)
Joonsoo Kimb2261022013-09-11 14:21:00 -07001886{
1887 struct page *page;
1888 int nr_nodes, node;
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001889 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001890
1891 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
Mike Kravetzf60858f2019-09-23 15:37:35 -07001892 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1893 node_alloc_noretry);
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001894 if (page)
Joonsoo Kimb2261022013-09-11 14:21:00 -07001895 break;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001896 }
1897
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001898 if (!page)
1899 return 0;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001900
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001901 put_page(page); /* free it into the hugepage allocator */
1902
1903 return 1;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001904}
1905
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001906/*
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001907 * Remove huge page from pool from next node to free. Attempt to keep
1908 * persistent huge pages more or less balanced over allowed nodes.
1909 * This routine only 'removes' the hugetlb page. The caller must make
1910 * an additional call to free the page to low level allocators.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001911 * Called with hugetlb_lock locked.
1912 */
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001913static struct page *remove_pool_huge_page(struct hstate *h,
1914 nodemask_t *nodes_allowed,
1915 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001916{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001917 int nr_nodes, node;
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001918 struct page *page = NULL;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001919
Mike Kravetz9487ca62021-05-04 18:35:10 -07001920 lockdep_assert_held(&hugetlb_lock);
Joonsoo Kimb2261022013-09-11 14:21:00 -07001921 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001922 /*
1923 * If we're returning unused surplus pages, only examine
1924 * nodes with surplus pages.
1925 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001926 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1927 !list_empty(&h->hugepage_freelists[node])) {
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001928 page = list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001929 struct page, lru);
Mike Kravetz6eb4e882021-05-04 18:34:55 -07001930 remove_hugetlb_page(h, page, acct_surplus);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001931 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001932 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001933 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001934
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001935 return page;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001936}
1937
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001938/*
1939 * Dissolve a given free hugepage into free buddy pages. This function does
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001940 * nothing for in-use hugepages and non-hugepages.
1941 * This function returns values like below:
1942 *
Muchun Songad2fa372021-06-30 18:47:21 -07001943 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
1944 * when the system is under memory pressure and the feature of
1945 * freeing unused vmemmap pages associated with each hugetlb page
1946 * is enabled.
1947 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1948 * (allocated or reserved.)
1949 * 0: successfully dissolved free hugepages or the page is not a
1950 * hugepage (considered as already dissolved)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001951 */
Anshuman Khandualc3114a82017-07-10 15:47:41 -07001952int dissolve_free_huge_page(struct page *page)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001953{
Naoya Horiguchi6bc9b562018-08-23 17:00:38 -07001954 int rc = -EBUSY;
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001955
Muchun Song7ffddd42021-02-04 18:32:06 -08001956retry:
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001957 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1958 if (!PageHuge(page))
1959 return 0;
1960
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001961 spin_lock_irq(&hugetlb_lock);
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001962 if (!PageHuge(page)) {
1963 rc = 0;
1964 goto out;
1965 }
1966
1967 if (!page_count(page)) {
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001968 struct page *head = compound_head(page);
1969 struct hstate *h = page_hstate(head);
Naoya Horiguchi6bc9b562018-08-23 17:00:38 -07001970 if (h->free_huge_pages - h->resv_huge_pages == 0)
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001971 goto out;
Muchun Song7ffddd42021-02-04 18:32:06 -08001972
1973 /*
1974 * We should make sure that the page is already on the free list
1975 * when it is dissolved.
1976 */
Mike Kravetz6c037142021-02-24 12:09:04 -08001977 if (unlikely(!HPageFreed(head))) {
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001978 spin_unlock_irq(&hugetlb_lock);
Muchun Song7ffddd42021-02-04 18:32:06 -08001979 cond_resched();
1980
1981 /*
1982 * Theoretically, we should return -EBUSY when we
1983 * encounter this race. In fact, we have a chance
1984 * to successfully dissolve the page if we do a
1985 * retry. Because the race window is quite small.
1986 * If we seize this opportunity, it is an optimization
1987 * for increasing the success rate of dissolving page.
1988 */
1989 goto retry;
1990 }
1991
Naoya Horiguchi0c5da352021-06-04 20:01:27 -07001992 remove_hugetlb_page(h, head, false);
zhong jiangc1470b32016-08-11 15:32:55 -07001993 h->max_huge_pages--;
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001994 spin_unlock_irq(&hugetlb_lock);
Muchun Songad2fa372021-06-30 18:47:21 -07001995
1996 /*
1997 * Normally update_and_free_page will allocate required vmemmmap
1998 * before freeing the page. update_and_free_page will fail to
1999 * free the page if it can not allocate required vmemmap. We
2000 * need to adjust max_huge_pages if the page is not freed.
2001 * Attempt to allocate vmemmmap here so that we can take
2002 * appropriate action on failure.
2003 */
2004 rc = alloc_huge_page_vmemmap(h, head);
2005 if (!rc) {
2006 /*
2007 * Move PageHWPoison flag from head page to the raw
2008 * error page, which makes any subpages rather than
2009 * the error page reusable.
2010 */
2011 if (PageHWPoison(head) && page != head) {
2012 SetPageHWPoison(page);
2013 ClearPageHWPoison(head);
2014 }
2015 update_and_free_page(h, head, false);
2016 } else {
2017 spin_lock_irq(&hugetlb_lock);
2018 add_hugetlb_page(h, head, false);
2019 h->max_huge_pages++;
2020 spin_unlock_irq(&hugetlb_lock);
2021 }
2022
2023 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07002024 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07002025out:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002026 spin_unlock_irq(&hugetlb_lock);
Gerald Schaefer082d5b62016-10-07 17:01:10 -07002027 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07002028}
2029
2030/*
2031 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2032 * make specified memory blocks removable from the system.
Gerald Schaefer2247bb32016-10-07 17:01:07 -07002033 * Note that this will dissolve a free gigantic hugepage completely, if any
2034 * part of it lies within the given range.
Gerald Schaefer082d5b62016-10-07 17:01:10 -07002035 * Also note that if dissolve_free_huge_page() returns with an error, all
2036 * free hugepages that were dissolved before that error are lost.
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07002037 */
Gerald Schaefer082d5b62016-10-07 17:01:10 -07002038int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07002039{
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07002040 unsigned long pfn;
Gerald Schaefereb03aa02016-10-07 17:01:13 -07002041 struct page *page;
Gerald Schaefer082d5b62016-10-07 17:01:10 -07002042 int rc = 0;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07002043
Li Zhongd0177632014-08-06 16:07:56 -07002044 if (!hugepages_supported())
Gerald Schaefer082d5b62016-10-07 17:01:10 -07002045 return rc;
Li Zhongd0177632014-08-06 16:07:56 -07002046
Gerald Schaefereb03aa02016-10-07 17:01:13 -07002047 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
2048 page = pfn_to_page(pfn);
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07002049 rc = dissolve_free_huge_page(page);
2050 if (rc)
2051 break;
Gerald Schaefereb03aa02016-10-07 17:01:13 -07002052 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07002053
2054 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07002055}
2056
Michal Hockoab5ac902018-01-31 16:20:48 -08002057/*
2058 * Allocates a fresh surplus page from the page allocator.
2059 */
Michal Hocko0c397da2018-01-31 16:20:56 -08002060static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07002061 int nid, nodemask_t *nmask, bool zero_ref)
Adam Litke7893d1d2007-10-16 01:26:18 -07002062{
Michal Hocko9980d742018-01-31 16:20:52 -08002063 struct page *page = NULL;
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07002064 bool retry = false;
Adam Litke7893d1d2007-10-16 01:26:18 -07002065
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002066 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07002067 return NULL;
2068
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002069 spin_lock_irq(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08002070 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2071 goto out_unlock;
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002072 spin_unlock_irq(&hugetlb_lock);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002073
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07002074retry:
Mike Kravetzf60858f2019-09-23 15:37:35 -07002075 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
Michal Hocko9980d742018-01-31 16:20:52 -08002076 if (!page)
Michal Hocko0c397da2018-01-31 16:20:56 -08002077 return NULL;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002078
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002079 spin_lock_irq(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08002080 /*
2081 * We could have raced with the pool size change.
2082 * Double check that and simply deallocate the new page
2083 * if we would end up overcommiting the surpluses. Abuse
2084 * temporary page to workaround the nasty free_huge_page
2085 * codeflow
2086 */
2087 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Mike Kravetz9157c3112021-02-24 12:09:00 -08002088 SetHPageTemporary(page);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002089 spin_unlock_irq(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08002090 put_page(page);
Kai Shen2bf753e2019-05-13 17:15:37 -07002091 return NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -07002092 }
Michal Hocko9980d742018-01-31 16:20:52 -08002093
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07002094 if (zero_ref) {
2095 /*
2096 * Caller requires a page with zero ref count.
2097 * We will drop ref count here. If someone else is holding
2098 * a ref, the page will be freed when they drop it. Abuse
2099 * temporary page flag to accomplish this.
2100 */
2101 SetHPageTemporary(page);
2102 if (!put_page_testzero(page)) {
2103 /*
2104 * Unexpected inflated ref count on freshly allocated
2105 * huge. Retry once.
2106 */
2107 pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n");
2108 spin_unlock_irq(&hugetlb_lock);
2109 if (retry)
2110 return NULL;
2111
2112 retry = true;
2113 goto retry;
2114 }
2115 ClearHPageTemporary(page);
2116 }
2117
2118 h->surplus_huge_pages++;
2119 h->surplus_huge_pages_node[page_to_nid(page)]++;
2120
Michal Hocko9980d742018-01-31 16:20:52 -08002121out_unlock:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002122 spin_unlock_irq(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07002123
2124 return page;
2125}
2126
Joonsoo Kimbbe88752020-08-11 18:37:38 -07002127static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
Aneesh Kumar K.V9a4e9f32019-03-05 15:47:44 -08002128 int nid, nodemask_t *nmask)
Michal Hockoab5ac902018-01-31 16:20:48 -08002129{
2130 struct page *page;
2131
2132 if (hstate_is_gigantic(h))
2133 return NULL;
2134
Mike Kravetzf60858f2019-09-23 15:37:35 -07002135 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
Michal Hockoab5ac902018-01-31 16:20:48 -08002136 if (!page)
2137 return NULL;
2138
2139 /*
2140 * We do not account these pages as surplus because they are only
2141 * temporary and will be released properly on the last reference
2142 */
Mike Kravetz9157c3112021-02-24 12:09:00 -08002143 SetHPageTemporary(page);
Michal Hockoab5ac902018-01-31 16:20:48 -08002144
2145 return page;
2146}
2147
Adam Litkee4e574b2007-10-16 01:26:19 -07002148/*
Dave Hansen099730d2015-11-05 18:50:17 -08002149 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2150 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08002151static
Michal Hocko0c397da2018-01-31 16:20:56 -08002152struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
Dave Hansen099730d2015-11-05 18:50:17 -08002153 struct vm_area_struct *vma, unsigned long addr)
2154{
Ben Widawskycfcaa662021-09-02 15:00:13 -07002155 struct page *page = NULL;
Michal Hockoaaf14e42017-07-10 15:49:08 -07002156 struct mempolicy *mpol;
2157 gfp_t gfp_mask = htlb_alloc_mask(h);
2158 int nid;
2159 nodemask_t *nodemask;
2160
2161 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
Ben Widawskycfcaa662021-09-02 15:00:13 -07002162 if (mpol_is_preferred_many(mpol)) {
2163 gfp_t gfp = gfp_mask | __GFP_NOWARN;
Michal Hockoaaf14e42017-07-10 15:49:08 -07002164
Ben Widawskycfcaa662021-09-02 15:00:13 -07002165 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2166 page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false);
2167
2168 /* Fallback to all nodes if page==NULL */
2169 nodemask = NULL;
2170 }
2171
2172 if (!page)
2173 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false);
2174 mpol_cond_put(mpol);
Michal Hockoaaf14e42017-07-10 15:49:08 -07002175 return page;
Dave Hansen099730d2015-11-05 18:50:17 -08002176}
2177
Michal Hockoab5ac902018-01-31 16:20:48 -08002178/* page migration callback function */
Michal Hocko3e59fcb2017-07-10 15:49:11 -07002179struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
Joonsoo Kimd92bbc22020-08-11 18:37:17 -07002180 nodemask_t *nmask, gfp_t gfp_mask)
Michal Hocko4db9b2e2017-07-10 15:48:44 -07002181{
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002182 spin_lock_irq(&hugetlb_lock);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07002183 if (h->free_huge_pages - h->resv_huge_pages > 0) {
Michal Hocko3e59fcb2017-07-10 15:49:11 -07002184 struct page *page;
2185
2186 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2187 if (page) {
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002188 spin_unlock_irq(&hugetlb_lock);
Michal Hocko3e59fcb2017-07-10 15:49:11 -07002189 return page;
Michal Hocko4db9b2e2017-07-10 15:48:44 -07002190 }
2191 }
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002192 spin_unlock_irq(&hugetlb_lock);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07002193
Michal Hocko0c397da2018-01-31 16:20:56 -08002194 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07002195}
2196
Michal Hockoebd63722018-01-31 16:21:00 -08002197/* mempolicy aware migration callback */
Michal Hocko389c8172018-01-31 16:21:03 -08002198struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2199 unsigned long address)
Michal Hockoebd63722018-01-31 16:21:00 -08002200{
2201 struct mempolicy *mpol;
2202 nodemask_t *nodemask;
2203 struct page *page;
Michal Hockoebd63722018-01-31 16:21:00 -08002204 gfp_t gfp_mask;
2205 int node;
2206
Michal Hockoebd63722018-01-31 16:21:00 -08002207 gfp_mask = htlb_alloc_mask(h);
2208 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
Joonsoo Kimd92bbc22020-08-11 18:37:17 -07002209 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
Michal Hockoebd63722018-01-31 16:21:00 -08002210 mpol_cond_put(mpol);
2211
2212 return page;
2213}
2214
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09002215/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002216 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07002217 * of size 'delta'.
2218 */
Liu Xiang0a4f3d12020-12-14 19:12:05 -08002219static int gather_surplus_pages(struct hstate *h, long delta)
Jules Irenge1b2a1e72020-04-06 20:08:09 -07002220 __must_hold(&hugetlb_lock)
Adam Litkee4e574b2007-10-16 01:26:19 -07002221{
2222 struct list_head surplus_list;
2223 struct page *page, *tmp;
Liu Xiang0a4f3d12020-12-14 19:12:05 -08002224 int ret;
2225 long i;
2226 long needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07002227 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07002228
Mike Kravetz9487ca62021-05-04 18:35:10 -07002229 lockdep_assert_held(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002230 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08002231 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002232 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07002233 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08002234 }
Adam Litkee4e574b2007-10-16 01:26:19 -07002235
2236 allocated = 0;
2237 INIT_LIST_HEAD(&surplus_list);
2238
2239 ret = -ENOMEM;
2240retry:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002241 spin_unlock_irq(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07002242 for (i = 0; i < needed; i++) {
Michal Hocko0c397da2018-01-31 16:20:56 -08002243 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07002244 NUMA_NO_NODE, NULL, true);
Hillf Danton28073b02012-03-21 16:34:00 -07002245 if (!page) {
2246 alloc_ok = false;
2247 break;
2248 }
Adam Litkee4e574b2007-10-16 01:26:19 -07002249 list_add(&page->lru, &surplus_list);
David Rientjes69ed7792017-07-10 15:48:50 -07002250 cond_resched();
Adam Litkee4e574b2007-10-16 01:26:19 -07002251 }
Hillf Danton28073b02012-03-21 16:34:00 -07002252 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07002253
2254 /*
2255 * After retaking hugetlb_lock, we need to recalculate 'needed'
2256 * because either resv_huge_pages or free_huge_pages may have changed.
2257 */
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002258 spin_lock_irq(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002259 needed = (h->resv_huge_pages + delta) -
2260 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07002261 if (needed > 0) {
2262 if (alloc_ok)
2263 goto retry;
2264 /*
2265 * We were not able to allocate enough pages to
2266 * satisfy the entire reservation so we free what
2267 * we've allocated so far.
2268 */
2269 goto free;
2270 }
Adam Litkee4e574b2007-10-16 01:26:19 -07002271 /*
2272 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002273 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07002274 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08002275 * allocator. Commit the entire reservation here to prevent another
2276 * process from stealing the pages as they are added to the pool but
2277 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07002278 */
2279 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07002280 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07002281 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09002282
Adam Litke19fc3f02008-04-28 02:12:20 -07002283 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07002284 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07002285 if ((--needed) < 0)
2286 break;
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07002287 /* Add the page to the hugetlb allocator */
Andi Kleena5516432008-07-23 21:27:41 -07002288 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07002289 }
Hillf Danton28073b02012-03-21 16:34:00 -07002290free:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002291 spin_unlock_irq(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07002292
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07002293 /*
2294 * Free unnecessary surplus pages to the buddy allocator.
2295 * Pages have no ref count, call free_huge_page directly.
2296 */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07002297 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07002298 free_huge_page(page);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002299 spin_lock_irq(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07002300
2301 return ret;
2302}
2303
2304/*
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002305 * This routine has two main purposes:
2306 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2307 * in unused_resv_pages. This corresponds to the prior adjustments made
2308 * to the associated reservation map.
2309 * 2) Free any unused surplus pages that may have been allocated to satisfy
2310 * the reservation. As many as unused_resv_pages may be freed.
Adam Litkee4e574b2007-10-16 01:26:19 -07002311 */
Andi Kleena5516432008-07-23 21:27:41 -07002312static void return_unused_surplus_pages(struct hstate *h,
2313 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07002314{
Adam Litkee4e574b2007-10-16 01:26:19 -07002315 unsigned long nr_pages;
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002316 struct page *page;
2317 LIST_HEAD(page_list);
2318
Mike Kravetz9487ca62021-05-04 18:35:10 -07002319 lockdep_assert_held(&hugetlb_lock);
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002320 /* Uncommit the reservation */
2321 h->resv_huge_pages -= unused_resv_pages;
Adam Litkee4e574b2007-10-16 01:26:19 -07002322
Andi Kleenaa888a72008-07-23 21:27:47 -07002323 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002324 if (hstate_is_gigantic(h))
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002325 goto out;
Andi Kleenaa888a72008-07-23 21:27:47 -07002326
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002327 /*
2328 * Part (or even all) of the reservation could have been backed
2329 * by pre-allocated pages. Only free surplus pages.
2330 */
Andi Kleena5516432008-07-23 21:27:41 -07002331 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07002332
Lee Schermerhorn685f3452009-09-21 17:01:23 -07002333 /*
2334 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002335 * evenly across all nodes with memory. Iterate across these nodes
2336 * until we can no longer free unreserved surplus pages. This occurs
2337 * when the nodes with surplus pages have no free pages.
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002338 * remove_pool_huge_page() will balance the freed pages across the
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002339 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07002340 */
2341 while (nr_pages--) {
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002342 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2343 if (!page)
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002344 goto out;
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002345
2346 list_add(&page->lru, &page_list);
Adam Litkee4e574b2007-10-16 01:26:19 -07002347 }
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002348
2349out:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002350 spin_unlock_irq(&hugetlb_lock);
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002351 update_and_free_pages_bulk(h, &page_list);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002352 spin_lock_irq(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07002353}
2354
Mike Kravetz5e911372015-09-08 15:01:28 -07002355
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002356/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002357 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
Mike Kravetz5e911372015-09-08 15:01:28 -07002358 * are used by the huge page allocation routines to manage reservations.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002359 *
2360 * vma_needs_reservation is called to determine if the huge page at addr
2361 * within the vma has an associated reservation. If a reservation is
2362 * needed, the value 1 is returned. The caller is then responsible for
2363 * managing the global reservation and subpool usage counts. After
2364 * the huge page has been allocated, vma_commit_reservation is called
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002365 * to add the page to the reservation map. If the page allocation fails,
2366 * the reservation must be ended instead of committed. vma_end_reservation
2367 * is called in such cases.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002368 *
2369 * In the normal case, vma_commit_reservation returns the same value
2370 * as the preceding vma_needs_reservation call. The only time this
2371 * is not the case is if a reserve map was changed between calls. It
2372 * is the responsibility of the caller to notice the difference and
2373 * take appropriate action.
Mike Kravetz96b96a92016-11-10 10:46:32 -08002374 *
2375 * vma_add_reservation is used in error paths where a reservation must
2376 * be restored when a newly allocated huge page must be freed. It is
2377 * to be called after calling vma_needs_reservation to determine if a
2378 * reservation exists.
Mike Kravetz846be082021-06-15 18:23:29 -07002379 *
2380 * vma_del_reservation is used in error paths where an entry in the reserve
2381 * map was created during huge page allocation and must be removed. It is to
2382 * be called after calling vma_needs_reservation to determine if a reservation
2383 * exists.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002384 */
Mike Kravetz5e911372015-09-08 15:01:28 -07002385enum vma_resv_mode {
2386 VMA_NEEDS_RESV,
2387 VMA_COMMIT_RESV,
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002388 VMA_END_RESV,
Mike Kravetz96b96a92016-11-10 10:46:32 -08002389 VMA_ADD_RESV,
Mike Kravetz846be082021-06-15 18:23:29 -07002390 VMA_DEL_RESV,
Mike Kravetz5e911372015-09-08 15:01:28 -07002391};
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002392static long __vma_reservation_common(struct hstate *h,
2393 struct vm_area_struct *vma, unsigned long addr,
Mike Kravetz5e911372015-09-08 15:01:28 -07002394 enum vma_resv_mode mode)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002395{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002396 struct resv_map *resv;
2397 pgoff_t idx;
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002398 long ret;
Mina Almasry0db9d742020-04-01 21:11:25 -07002399 long dummy_out_regions_needed;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002400
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002401 resv = vma_resv_map(vma);
2402 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07002403 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002404
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002405 idx = vma_hugecache_offset(h, vma, addr);
Mike Kravetz5e911372015-09-08 15:01:28 -07002406 switch (mode) {
2407 case VMA_NEEDS_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002408 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2409 /* We assume that vma_reservation_* routines always operate on
2410 * 1 page, and that adding to resv map a 1 page entry can only
2411 * ever require 1 region.
2412 */
2413 VM_BUG_ON(dummy_out_regions_needed != 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07002414 break;
2415 case VMA_COMMIT_RESV:
Mina Almasry075a61d2020-04-01 21:11:28 -07002416 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
Mina Almasry0db9d742020-04-01 21:11:25 -07002417 /* region_add calls of range 1 should never fail. */
2418 VM_BUG_ON(ret < 0);
Mike Kravetz5e911372015-09-08 15:01:28 -07002419 break;
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002420 case VMA_END_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002421 region_abort(resv, idx, idx + 1, 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07002422 ret = 0;
2423 break;
Mike Kravetz96b96a92016-11-10 10:46:32 -08002424 case VMA_ADD_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002425 if (vma->vm_flags & VM_MAYSHARE) {
Mina Almasry075a61d2020-04-01 21:11:28 -07002426 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
Mina Almasry0db9d742020-04-01 21:11:25 -07002427 /* region_add calls of range 1 should never fail. */
2428 VM_BUG_ON(ret < 0);
2429 } else {
2430 region_abort(resv, idx, idx + 1, 1);
Mike Kravetz96b96a92016-11-10 10:46:32 -08002431 ret = region_del(resv, idx, idx + 1);
2432 }
2433 break;
Mike Kravetz846be082021-06-15 18:23:29 -07002434 case VMA_DEL_RESV:
2435 if (vma->vm_flags & VM_MAYSHARE) {
2436 region_abort(resv, idx, idx + 1, 1);
2437 ret = region_del(resv, idx, idx + 1);
2438 } else {
2439 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2440 /* region_add calls of range 1 should never fail. */
2441 VM_BUG_ON(ret < 0);
2442 }
2443 break;
Mike Kravetz5e911372015-09-08 15:01:28 -07002444 default:
2445 BUG();
2446 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002447
Mike Kravetz846be082021-06-15 18:23:29 -07002448 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002449 return ret;
Miaohe Linbf3d12b2021-05-04 18:34:32 -07002450 /*
2451 * We know private mapping must have HPAGE_RESV_OWNER set.
2452 *
2453 * In most cases, reserves always exist for private mappings.
2454 * However, a file associated with mapping could have been
2455 * hole punched or truncated after reserves were consumed.
2456 * As subsequent fault on such a range will not use reserves.
2457 * Subtle - The reserve map for private mappings has the
2458 * opposite meaning than that of shared mappings. If NO
2459 * entry is in the reserve map, it means a reservation exists.
2460 * If an entry exists in the reserve map, it means the
2461 * reservation has already been consumed. As a result, the
2462 * return value of this routine is the opposite of the
2463 * value returned from reserve map manipulation routines above.
2464 */
2465 if (ret > 0)
2466 return 0;
2467 if (ret == 0)
2468 return 1;
2469 return ret;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002470}
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002471
2472static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07002473 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002474{
Mike Kravetz5e911372015-09-08 15:01:28 -07002475 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002476}
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002477
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002478static long vma_commit_reservation(struct hstate *h,
2479 struct vm_area_struct *vma, unsigned long addr)
2480{
Mike Kravetz5e911372015-09-08 15:01:28 -07002481 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2482}
2483
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002484static void vma_end_reservation(struct hstate *h,
Mike Kravetz5e911372015-09-08 15:01:28 -07002485 struct vm_area_struct *vma, unsigned long addr)
2486{
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002487 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002488}
2489
Mike Kravetz96b96a92016-11-10 10:46:32 -08002490static long vma_add_reservation(struct hstate *h,
2491 struct vm_area_struct *vma, unsigned long addr)
2492{
2493 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2494}
2495
Mike Kravetz846be082021-06-15 18:23:29 -07002496static long vma_del_reservation(struct hstate *h,
2497 struct vm_area_struct *vma, unsigned long addr)
Mike Kravetz96b96a92016-11-10 10:46:32 -08002498{
Mike Kravetz846be082021-06-15 18:23:29 -07002499 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2500}
Mike Kravetz96b96a92016-11-10 10:46:32 -08002501
Mike Kravetz846be082021-06-15 18:23:29 -07002502/*
2503 * This routine is called to restore reservation information on error paths.
2504 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2505 * the hugetlb mutex should remain held when calling this routine.
2506 *
2507 * It handles two specific cases:
2508 * 1) A reservation was in place and the page consumed the reservation.
2509 * HPageRestoreReserve is set in the page.
2510 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2511 * not set. However, alloc_huge_page always updates the reserve map.
2512 *
2513 * In case 1, free_huge_page later in the error path will increment the
2514 * global reserve count. But, free_huge_page does not have enough context
2515 * to adjust the reservation map. This case deals primarily with private
2516 * mappings. Adjust the reserve map here to be consistent with global
2517 * reserve count adjustments to be made by free_huge_page. Make sure the
2518 * reserve map indicates there is a reservation present.
2519 *
2520 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2521 */
2522void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2523 unsigned long address, struct page *page)
2524{
2525 long rc = vma_needs_reservation(h, vma, address);
2526
2527 if (HPageRestoreReserve(page)) {
2528 if (unlikely(rc < 0))
Mike Kravetz96b96a92016-11-10 10:46:32 -08002529 /*
2530 * Rare out of memory condition in reserve map
Mike Kravetzd6995da2021-02-24 12:08:51 -08002531 * manipulation. Clear HPageRestoreReserve so that
Mike Kravetz96b96a92016-11-10 10:46:32 -08002532 * global reserve count will not be incremented
2533 * by free_huge_page. This will make it appear
2534 * as though the reservation for this page was
2535 * consumed. This may prevent the task from
2536 * faulting in the page at a later time. This
2537 * is better than inconsistent global huge page
2538 * accounting of reserve counts.
2539 */
Mike Kravetzd6995da2021-02-24 12:08:51 -08002540 ClearHPageRestoreReserve(page);
Mike Kravetz846be082021-06-15 18:23:29 -07002541 else if (rc)
2542 (void)vma_add_reservation(h, vma, address);
2543 else
Mike Kravetz96b96a92016-11-10 10:46:32 -08002544 vma_end_reservation(h, vma, address);
Mike Kravetz846be082021-06-15 18:23:29 -07002545 } else {
2546 if (!rc) {
2547 /*
2548 * This indicates there is an entry in the reserve map
Mike Kravetzc7b18502021-08-19 19:04:33 -07002549 * not added by alloc_huge_page. We know it was added
Mike Kravetz846be082021-06-15 18:23:29 -07002550 * before the alloc_huge_page call, otherwise
2551 * HPageRestoreReserve would be set on the page.
2552 * Remove the entry so that a subsequent allocation
2553 * does not consume a reservation.
2554 */
2555 rc = vma_del_reservation(h, vma, address);
2556 if (rc < 0)
2557 /*
2558 * VERY rare out of memory condition. Since
2559 * we can not delete the entry, set
2560 * HPageRestoreReserve so that the reserve
2561 * count will be incremented when the page
2562 * is freed. This reserve will be consumed
2563 * on a subsequent allocation.
2564 */
2565 SetHPageRestoreReserve(page);
2566 } else if (rc < 0) {
2567 /*
2568 * Rare out of memory condition from
2569 * vma_needs_reservation call. Memory allocation is
2570 * only attempted if a new entry is needed. Therefore,
2571 * this implies there is not an entry in the
2572 * reserve map.
2573 *
2574 * For shared mappings, no entry in the map indicates
2575 * no reservation. We are done.
2576 */
2577 if (!(vma->vm_flags & VM_MAYSHARE))
2578 /*
2579 * For private mappings, no entry indicates
2580 * a reservation is present. Since we can
2581 * not add an entry, set SetHPageRestoreReserve
2582 * on the page so reserve count will be
2583 * incremented when freed. This reserve will
2584 * be consumed on a subsequent allocation.
2585 */
2586 SetHPageRestoreReserve(page);
2587 } else
2588 /*
2589 * No reservation present, do nothing
2590 */
2591 vma_end_reservation(h, vma, address);
Mike Kravetz96b96a92016-11-10 10:46:32 -08002592 }
2593}
2594
Oscar Salvador369fa222021-05-04 18:35:26 -07002595/*
2596 * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2597 * @h: struct hstate old page belongs to
2598 * @old_page: Old page to dissolve
Oscar Salvadorae37c7f2021-05-04 18:35:29 -07002599 * @list: List to isolate the page in case we need to
Oscar Salvador369fa222021-05-04 18:35:26 -07002600 * Returns 0 on success, otherwise negated error.
2601 */
Oscar Salvadorae37c7f2021-05-04 18:35:29 -07002602static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2603 struct list_head *list)
Oscar Salvador369fa222021-05-04 18:35:26 -07002604{
2605 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2606 int nid = page_to_nid(old_page);
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07002607 bool alloc_retry = false;
Oscar Salvador369fa222021-05-04 18:35:26 -07002608 struct page *new_page;
2609 int ret = 0;
2610
2611 /*
2612 * Before dissolving the page, we need to allocate a new one for the
Muchun Songf41f2ed2021-06-30 18:47:13 -07002613 * pool to remain stable. Here, we allocate the page and 'prep' it
2614 * by doing everything but actually updating counters and adding to
2615 * the pool. This simplifies and let us do most of the processing
2616 * under the lock.
Oscar Salvador369fa222021-05-04 18:35:26 -07002617 */
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07002618alloc_retry:
Oscar Salvador369fa222021-05-04 18:35:26 -07002619 new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2620 if (!new_page)
2621 return -ENOMEM;
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07002622 /*
2623 * If all goes well, this page will be directly added to the free
2624 * list in the pool. For this the ref count needs to be zero.
2625 * Attempt to drop now, and retry once if needed. It is VERY
2626 * unlikely there is another ref on the page.
2627 *
2628 * If someone else has a reference to the page, it will be freed
2629 * when they drop their ref. Abuse temporary page flag to accomplish
2630 * this. Retry once if there is an inflated ref count.
2631 */
2632 SetHPageTemporary(new_page);
2633 if (!put_page_testzero(new_page)) {
2634 if (alloc_retry)
2635 return -EBUSY;
2636
2637 alloc_retry = true;
2638 goto alloc_retry;
2639 }
2640 ClearHPageTemporary(new_page);
2641
Muchun Songf41f2ed2021-06-30 18:47:13 -07002642 __prep_new_huge_page(h, new_page);
Oscar Salvador369fa222021-05-04 18:35:26 -07002643
2644retry:
2645 spin_lock_irq(&hugetlb_lock);
2646 if (!PageHuge(old_page)) {
2647 /*
2648 * Freed from under us. Drop new_page too.
2649 */
2650 goto free_new;
2651 } else if (page_count(old_page)) {
2652 /*
Oscar Salvadorae37c7f2021-05-04 18:35:29 -07002653 * Someone has grabbed the page, try to isolate it here.
2654 * Fail with -EBUSY if not possible.
Oscar Salvador369fa222021-05-04 18:35:26 -07002655 */
Oscar Salvadorae37c7f2021-05-04 18:35:29 -07002656 spin_unlock_irq(&hugetlb_lock);
2657 if (!isolate_huge_page(old_page, list))
2658 ret = -EBUSY;
2659 spin_lock_irq(&hugetlb_lock);
Oscar Salvador369fa222021-05-04 18:35:26 -07002660 goto free_new;
2661 } else if (!HPageFreed(old_page)) {
2662 /*
2663 * Page's refcount is 0 but it has not been enqueued in the
2664 * freelist yet. Race window is small, so we can succeed here if
2665 * we retry.
2666 */
2667 spin_unlock_irq(&hugetlb_lock);
2668 cond_resched();
2669 goto retry;
2670 } else {
2671 /*
2672 * Ok, old_page is still a genuine free hugepage. Remove it from
2673 * the freelist and decrease the counters. These will be
2674 * incremented again when calling __prep_account_new_huge_page()
2675 * and enqueue_huge_page() for new_page. The counters will remain
2676 * stable since this happens under the lock.
2677 */
2678 remove_hugetlb_page(h, old_page, false);
2679
2680 /*
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07002681 * Ref count on new page is already zero as it was dropped
2682 * earlier. It can be directly added to the pool free list.
Oscar Salvador369fa222021-05-04 18:35:26 -07002683 */
Oscar Salvador369fa222021-05-04 18:35:26 -07002684 __prep_account_new_huge_page(h, nid);
Oscar Salvador369fa222021-05-04 18:35:26 -07002685 enqueue_huge_page(h, new_page);
2686
2687 /*
2688 * Pages have been replaced, we can safely free the old one.
2689 */
2690 spin_unlock_irq(&hugetlb_lock);
Muchun Songb65d4ad2021-06-30 18:47:17 -07002691 update_and_free_page(h, old_page, false);
Oscar Salvador369fa222021-05-04 18:35:26 -07002692 }
2693
2694 return ret;
2695
2696free_new:
2697 spin_unlock_irq(&hugetlb_lock);
Mike Kravetzb65a4ed2021-09-02 14:58:47 -07002698 /* Page has a zero ref count, but needs a ref to be freed */
2699 set_page_refcounted(new_page);
Muchun Songb65d4ad2021-06-30 18:47:17 -07002700 update_and_free_page(h, new_page, false);
Oscar Salvador369fa222021-05-04 18:35:26 -07002701
2702 return ret;
2703}
2704
Oscar Salvadorae37c7f2021-05-04 18:35:29 -07002705int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
Oscar Salvador369fa222021-05-04 18:35:26 -07002706{
2707 struct hstate *h;
2708 struct page *head;
Oscar Salvadorae37c7f2021-05-04 18:35:29 -07002709 int ret = -EBUSY;
Oscar Salvador369fa222021-05-04 18:35:26 -07002710
2711 /*
2712 * The page might have been dissolved from under our feet, so make sure
2713 * to carefully check the state under the lock.
2714 * Return success when racing as if we dissolved the page ourselves.
2715 */
2716 spin_lock_irq(&hugetlb_lock);
2717 if (PageHuge(page)) {
2718 head = compound_head(page);
2719 h = page_hstate(head);
2720 } else {
2721 spin_unlock_irq(&hugetlb_lock);
2722 return 0;
2723 }
2724 spin_unlock_irq(&hugetlb_lock);
2725
2726 /*
2727 * Fence off gigantic pages as there is a cyclic dependency between
2728 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2729 * of bailing out right away without further retrying.
2730 */
2731 if (hstate_is_gigantic(h))
2732 return -ENOMEM;
2733
Oscar Salvadorae37c7f2021-05-04 18:35:29 -07002734 if (page_count(head) && isolate_huge_page(head, list))
2735 ret = 0;
2736 else if (!page_count(head))
2737 ret = alloc_and_dissolve_huge_page(h, head, list);
2738
2739 return ret;
Oscar Salvador369fa222021-05-04 18:35:26 -07002740}
2741
Mike Kravetz70c35472015-09-08 15:01:54 -07002742struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002743 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744{
David Gibson90481622012-03-21 16:34:12 -07002745 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07002746 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08002747 struct page *page;
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002748 long map_chg, map_commit;
2749 long gbl_chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002750 int ret, idx;
2751 struct hugetlb_cgroup *h_cg;
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002752 bool deferred_reserve;
Adam Litke2fc39ce2007-11-14 16:59:39 -08002753
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002754 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07002755 /*
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002756 * Examine the region/reserve map to determine if the process
2757 * has a reservation for the page to be allocated. A return
2758 * code of zero indicates a reservation exists (no change).
Mel Gormana1e78772008-07-23 21:27:23 -07002759 */
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002760 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2761 if (map_chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002762 return ERR_PTR(-ENOMEM);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002763
2764 /*
2765 * Processes that did not create the mapping will have no
2766 * reserves as indicated by the region/reserve map. Check
2767 * that the allocation will not exceed the subpool limit.
2768 * Allocations for MAP_NORESERVE mappings also need to be
2769 * checked against any subpool limit.
2770 */
2771 if (map_chg || avoid_reserve) {
2772 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2773 if (gbl_chg < 0) {
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002774 vma_end_reservation(h, vma, addr);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002775 return ERR_PTR(-ENOSPC);
Mike Kravetz5e911372015-09-08 15:01:28 -07002776 }
Mel Gormana1e78772008-07-23 21:27:23 -07002777
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002778 /*
2779 * Even though there was no reservation in the region/reserve
2780 * map, there could be reservations associated with the
2781 * subpool that can be used. This would be indicated if the
2782 * return value of hugepage_subpool_get_pages() is zero.
2783 * However, if avoid_reserve is specified we still avoid even
2784 * the subpool reservations.
2785 */
2786 if (avoid_reserve)
2787 gbl_chg = 1;
2788 }
2789
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002790 /* If this allocation is not consuming a reservation, charge it now.
2791 */
Miaohe Lin6501fe52021-05-04 18:33:16 -07002792 deferred_reserve = map_chg || avoid_reserve;
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002793 if (deferred_reserve) {
2794 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2795 idx, pages_per_huge_page(h), &h_cg);
2796 if (ret)
2797 goto out_subpool_put;
2798 }
2799
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002800 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002801 if (ret)
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002802 goto out_uncharge_cgroup_reservation;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002803
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002804 spin_lock_irq(&hugetlb_lock);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002805 /*
2806 * glb_chg is passed to indicate whether or not a page must be taken
2807 * from the global free pool (global change). gbl_chg == 0 indicates
2808 * a reservation exists for the allocation.
2809 */
2810 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002811 if (!page) {
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002812 spin_unlock_irq(&hugetlb_lock);
Michal Hocko0c397da2018-01-31 16:20:56 -08002813 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002814 if (!page)
2815 goto out_uncharge_cgroup;
Naoya Horiguchia88c7692015-12-11 13:40:24 -08002816 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
Mike Kravetzd6995da2021-02-24 12:08:51 -08002817 SetHPageRestoreReserve(page);
Naoya Horiguchia88c7692015-12-11 13:40:24 -08002818 h->resv_huge_pages--;
2819 }
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002820 spin_lock_irq(&hugetlb_lock);
Wei Yang15a8d682020-10-13 16:56:33 -07002821 list_add(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002822 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07002823 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002824 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002825 /* If allocation is not consuming a reservation, also store the
2826 * hugetlb_cgroup pointer on the page.
2827 */
2828 if (deferred_reserve) {
2829 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2830 h_cg, page);
2831 }
2832
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002833 spin_unlock_irq(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07002834
Mike Kravetzd6995da2021-02-24 12:08:51 -08002835 hugetlb_set_page_subpool(page, spool);
Mel Gormana1e78772008-07-23 21:27:23 -07002836
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002837 map_commit = vma_commit_reservation(h, vma, addr);
2838 if (unlikely(map_chg > map_commit)) {
Mike Kravetz33039672015-06-24 16:57:58 -07002839 /*
2840 * The page was added to the reservation map between
2841 * vma_needs_reservation and vma_commit_reservation.
2842 * This indicates a race with hugetlb_reserve_pages.
2843 * Adjust for the subpool count incremented above AND
2844 * in hugetlb_reserve_pages for the same page. Also,
2845 * the reservation count added in hugetlb_reserve_pages
2846 * no longer applies.
2847 */
2848 long rsv_adjust;
2849
2850 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2851 hugetlb_acct_memory(h, -rsv_adjust);
Mike Kravetz79aa9252020-11-01 17:07:27 -08002852 if (deferred_reserve)
2853 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2854 pages_per_huge_page(h), page);
Mike Kravetz33039672015-06-24 16:57:58 -07002855 }
Adam Litke90d8b7e2007-11-14 16:59:42 -08002856 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002857
2858out_uncharge_cgroup:
2859 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002860out_uncharge_cgroup_reservation:
2861 if (deferred_reserve)
2862 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2863 h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002864out_subpool_put:
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002865 if (map_chg || avoid_reserve)
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002866 hugepage_subpool_put_pages(spool, 1);
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002867 vma_end_reservation(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002868 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08002869}
2870
Aneesh Kumar K.Ve24a1302017-07-28 10:31:25 +05302871int alloc_bootmem_huge_page(struct hstate *h)
2872 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2873int __alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07002874{
2875 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07002876 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07002877
Joonsoo Kimb2261022013-09-11 14:21:00 -07002878 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002879 void *addr;
2880
Mike Rapoporteb31d552018-10-30 15:08:04 -07002881 addr = memblock_alloc_try_nid_raw(
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002882 huge_page_size(h), huge_page_size(h),
Mike Rapoport97ad1082018-10-30 15:09:44 -07002883 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07002884 if (addr) {
2885 /*
2886 * Use the beginning of the huge page to store the
2887 * huge_bootmem_page struct (until gather_bootmem
2888 * puts them into the mem_map).
2889 */
2890 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08002891 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07002892 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002893 }
2894 return 0;
2895
2896found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08002897 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07002898 /* Put them into a private list first because mem_map is not up yet */
Cannon Matthews330d6e42018-08-17 15:49:17 -07002899 INIT_LIST_HEAD(&m->list);
Andi Kleenaa888a72008-07-23 21:27:47 -07002900 list_add(&m->list, &huge_boot_pages);
2901 m->hstate = h;
2902 return 1;
2903}
2904
Mike Kravetz48b8d742021-06-30 18:48:31 -07002905/*
2906 * Put bootmem huge pages into the standard lists after mem_map is up.
2907 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
2908 */
Andi Kleenaa888a72008-07-23 21:27:47 -07002909static void __init gather_bootmem_prealloc(void)
2910{
2911 struct huge_bootmem_page *m;
2912
2913 list_for_each_entry(m, &huge_boot_pages, list) {
Mike Kravetz40d18eb2018-08-17 15:49:07 -07002914 struct page *page = virt_to_page(m);
Andi Kleenaa888a72008-07-23 21:27:47 -07002915 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07002916
Mike Kravetz48b8d742021-06-30 18:48:31 -07002917 VM_BUG_ON(!hstate_is_gigantic(h));
Andi Kleenaa888a72008-07-23 21:27:47 -07002918 WARN_ON(page_count(page) != 1);
Mike Kravetz7118fc22021-06-30 18:48:34 -07002919 if (prep_compound_gigantic_page(page, huge_page_order(h))) {
2920 WARN_ON(PageReserved(page));
2921 prep_new_huge_page(h, page, page_to_nid(page));
2922 put_page(page); /* add to the hugepage allocator */
2923 } else {
Mike Kravetz416d85e2021-09-02 14:58:43 -07002924 /* VERY unlikely inflated ref count on a tail page */
Mike Kravetz7118fc22021-06-30 18:48:34 -07002925 free_gigantic_page(page, huge_page_order(h));
Mike Kravetz7118fc22021-06-30 18:48:34 -07002926 }
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08002927
Rafael Aquinib0320c72011-06-15 15:08:39 -07002928 /*
Mike Kravetz48b8d742021-06-30 18:48:31 -07002929 * We need to restore the 'stolen' pages to totalram_pages
2930 * in order to fix confusing memory reports from free(1) and
2931 * other side-effects, like CommitLimit going negative.
Rafael Aquinib0320c72011-06-15 15:08:39 -07002932 */
Mike Kravetz48b8d742021-06-30 18:48:31 -07002933 adjust_managed_page_count(page, pages_per_huge_page(h));
Cannon Matthews520495f2018-07-03 17:02:43 -07002934 cond_resched();
Andi Kleenaa888a72008-07-23 21:27:47 -07002935 }
2936}
2937
Andi Kleen8faa8b02008-07-23 21:27:48 -07002938static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939{
2940 unsigned long i;
Mike Kravetzf60858f2019-09-23 15:37:35 -07002941 nodemask_t *node_alloc_noretry;
2942
2943 if (!hstate_is_gigantic(h)) {
2944 /*
2945 * Bit mask controlling how hard we retry per-node allocations.
2946 * Ignore errors as lower level routines can deal with
2947 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2948 * time, we are likely in bigger trouble.
2949 */
2950 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2951 GFP_KERNEL);
2952 } else {
2953 /* allocations done at boot time */
2954 node_alloc_noretry = NULL;
2955 }
2956
2957 /* bit mask controlling how hard we retry per-node allocations */
2958 if (node_alloc_noretry)
2959 nodes_clear(*node_alloc_noretry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960
Andi Kleene5ff2152008-07-23 21:27:42 -07002961 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002962 if (hstate_is_gigantic(h)) {
Barry Songdbda8fe2020-07-23 21:15:30 -07002963 if (hugetlb_cma_size) {
Roman Gushchincf11e852020-04-10 14:32:45 -07002964 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
Chen Wandun7ecc9562021-02-24 12:07:58 -08002965 goto free;
Roman Gushchincf11e852020-04-10 14:32:45 -07002966 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002967 if (!alloc_bootmem_huge_page(h))
2968 break;
Michal Hocko0c397da2018-01-31 16:20:56 -08002969 } else if (!alloc_pool_huge_page(h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07002970 &node_states[N_MEMORY],
2971 node_alloc_noretry))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002972 break;
David Rientjes69ed7792017-07-10 15:48:50 -07002973 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974 }
Liam R. Howlettd715cf82017-07-10 15:48:15 -07002975 if (i < h->max_huge_pages) {
2976 char buf[32];
2977
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002978 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
Liam R. Howlettd715cf82017-07-10 15:48:15 -07002979 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2980 h->max_huge_pages, buf, i);
2981 h->max_huge_pages = i;
2982 }
Chen Wandun7ecc9562021-02-24 12:07:58 -08002983free:
Mike Kravetzf60858f2019-09-23 15:37:35 -07002984 kfree(node_alloc_noretry);
Andi Kleene5ff2152008-07-23 21:27:42 -07002985}
2986
2987static void __init hugetlb_init_hstates(void)
2988{
2989 struct hstate *h;
2990
2991 for_each_hstate(h) {
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002992 if (minimum_order > huge_page_order(h))
2993 minimum_order = huge_page_order(h);
2994
Andi Kleen8faa8b02008-07-23 21:27:48 -07002995 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002996 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07002997 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07002998 }
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002999 VM_BUG_ON(minimum_order == UINT_MAX);
Andi Kleene5ff2152008-07-23 21:27:42 -07003000}
3001
3002static void __init report_hugepages(void)
3003{
3004 struct hstate *h;
3005
3006 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07003007 char buf[32];
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07003008
3009 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
Andrew Mortonffb22af2013-02-22 16:32:08 -08003010 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07003011 buf, h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07003012 }
3013}
3014
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08003016static void try_to_free_low(struct hstate *h, unsigned long count,
3017 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018{
Christoph Lameter4415cc82006-09-25 23:31:55 -07003019 int i;
Mike Kravetz11218282021-05-04 18:34:59 -07003020 LIST_HEAD(page_list);
Christoph Lameter4415cc82006-09-25 23:31:55 -07003021
Mike Kravetz9487ca62021-05-04 18:35:10 -07003022 lockdep_assert_held(&hugetlb_lock);
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07003023 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07003024 return;
3025
Mike Kravetz11218282021-05-04 18:34:59 -07003026 /*
3027 * Collect pages to be freed on a list, and free after dropping lock
3028 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08003029 for_each_node_mask(i, *nodes_allowed) {
Mike Kravetz10c6ec42021-05-04 18:35:03 -07003030 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07003031 struct list_head *freel = &h->hugepage_freelists[i];
3032 list_for_each_entry_safe(page, next, freel, lru) {
3033 if (count >= h->nr_huge_pages)
Mike Kravetz11218282021-05-04 18:34:59 -07003034 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 if (PageHighMem(page))
3036 continue;
Mike Kravetz6eb4e882021-05-04 18:34:55 -07003037 remove_hugetlb_page(h, page, false);
Mike Kravetz11218282021-05-04 18:34:59 -07003038 list_add(&page->lru, &page_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039 }
3040 }
Mike Kravetz11218282021-05-04 18:34:59 -07003041
3042out:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003043 spin_unlock_irq(&hugetlb_lock);
Mike Kravetz10c6ec42021-05-04 18:35:03 -07003044 update_and_free_pages_bulk(h, &page_list);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003045 spin_lock_irq(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003046}
3047#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08003048static inline void try_to_free_low(struct hstate *h, unsigned long count,
3049 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050{
3051}
3052#endif
3053
Wu Fengguang20a03072009-06-16 15:32:22 -07003054/*
3055 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3056 * balanced by operating on them in a round-robin fashion.
3057 * Returns 1 if an adjustment was made.
3058 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08003059static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3060 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07003061{
Joonsoo Kimb2261022013-09-11 14:21:00 -07003062 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07003063
Mike Kravetz9487ca62021-05-04 18:35:10 -07003064 lockdep_assert_held(&hugetlb_lock);
Wu Fengguang20a03072009-06-16 15:32:22 -07003065 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07003066
Joonsoo Kimb2261022013-09-11 14:21:00 -07003067 if (delta < 0) {
3068 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3069 if (h->surplus_huge_pages_node[node])
3070 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07003071 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07003072 } else {
3073 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3074 if (h->surplus_huge_pages_node[node] <
3075 h->nr_huge_pages_node[node])
3076 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07003077 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07003078 }
3079 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07003080
Joonsoo Kimb2261022013-09-11 14:21:00 -07003081found:
3082 h->surplus_huge_pages += delta;
3083 h->surplus_huge_pages_node[node] += delta;
3084 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07003085}
3086
Andi Kleena5516432008-07-23 21:27:41 -07003087#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Mike Kravetzfd875dc2019-05-13 17:19:20 -07003088static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07003089 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090{
Adam Litke7893d1d2007-10-16 01:26:18 -07003091 unsigned long min_count, ret;
Mike Kravetz10c6ec42021-05-04 18:35:03 -07003092 struct page *page;
3093 LIST_HEAD(page_list);
Mike Kravetzf60858f2019-09-23 15:37:35 -07003094 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3095
3096 /*
3097 * Bit mask controlling how hard we retry per-node allocations.
3098 * If we can not allocate the bit mask, do not attempt to allocate
3099 * the requested huge pages.
3100 */
3101 if (node_alloc_noretry)
3102 nodes_clear(*node_alloc_noretry);
3103 else
3104 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003105
Mike Kravetz29383962021-05-04 18:34:52 -07003106 /*
3107 * resize_lock mutex prevents concurrent adjustments to number of
3108 * pages in hstate via the proc/sysfs interfaces.
3109 */
3110 mutex_lock(&h->resize_lock);
Muchun Songb65d4ad2021-06-30 18:47:17 -07003111 flush_free_hpage_work(h);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003112 spin_lock_irq(&hugetlb_lock);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07003113
3114 /*
Mike Kravetzfd875dc2019-05-13 17:19:20 -07003115 * Check for a node specific request.
3116 * Changing node specific huge page count may require a corresponding
3117 * change to the global count. In any case, the passed node mask
3118 * (nodes_allowed) will restrict alloc/free to the specified node.
3119 */
3120 if (nid != NUMA_NO_NODE) {
3121 unsigned long old_count = count;
3122
3123 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3124 /*
3125 * User may have specified a large count value which caused the
3126 * above calculation to overflow. In this case, they wanted
3127 * to allocate as many huge pages as possible. Set count to
3128 * largest possible value to align with their intention.
3129 */
3130 if (count < old_count)
3131 count = ULONG_MAX;
3132 }
3133
3134 /*
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07003135 * Gigantic pages runtime allocation depend on the capability for large
3136 * page range allocation.
3137 * If the system does not provide this feature, return an error when
3138 * the user tries to allocate gigantic pages but let the user free the
3139 * boottime allocated gigantic pages.
3140 */
3141 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3142 if (count > persistent_huge_pages(h)) {
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003143 spin_unlock_irq(&hugetlb_lock);
Mike Kravetz29383962021-05-04 18:34:52 -07003144 mutex_unlock(&h->resize_lock);
Mike Kravetzf60858f2019-09-23 15:37:35 -07003145 NODEMASK_FREE(node_alloc_noretry);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07003146 return -EINVAL;
3147 }
3148 /* Fall through to decrease pool */
3149 }
Andi Kleenaa888a72008-07-23 21:27:47 -07003150
Adam Litke7893d1d2007-10-16 01:26:18 -07003151 /*
3152 * Increase the pool size
3153 * First take pages out of surplus state. Then make up the
3154 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08003155 *
Michal Hocko0c397da2018-01-31 16:20:56 -08003156 * We might race with alloc_surplus_huge_page() here and be unable
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08003157 * to convert a surplus huge page to a normal huge page. That is
3158 * not critical, though, it just means the overall size of the
3159 * pool might be one hugepage larger than it needs to be, but
3160 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07003161 */
Andi Kleena5516432008-07-23 21:27:41 -07003162 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08003163 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07003164 break;
3165 }
3166
Andi Kleena5516432008-07-23 21:27:41 -07003167 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07003168 /*
3169 * If this allocation races such that we no longer need the
3170 * page, free_huge_page will handle it by freeing the page
3171 * and reducing the surplus.
3172 */
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003173 spin_unlock_irq(&hugetlb_lock);
Jia He649920c2016-08-02 14:02:31 -07003174
3175 /* yield cpu to avoid soft lockup */
3176 cond_resched();
3177
Mike Kravetzf60858f2019-09-23 15:37:35 -07003178 ret = alloc_pool_huge_page(h, nodes_allowed,
3179 node_alloc_noretry);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003180 spin_lock_irq(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07003181 if (!ret)
3182 goto out;
3183
Mel Gorman536240f22009-12-14 17:59:56 -08003184 /* Bail for signals. Probably ctrl-c from user */
3185 if (signal_pending(current))
3186 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07003187 }
Adam Litke7893d1d2007-10-16 01:26:18 -07003188
3189 /*
3190 * Decrease the pool size
3191 * First return free pages to the buddy allocator (being careful
3192 * to keep enough around to satisfy reservations). Then place
3193 * pages into surplus state as needed so the pool will shrink
3194 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08003195 *
3196 * By placing pages into the surplus state independent of the
3197 * overcommit value, we are allowing the surplus pool size to
3198 * exceed overcommit. There are few sane options here. Since
Michal Hocko0c397da2018-01-31 16:20:56 -08003199 * alloc_surplus_huge_page() is checking the global counter,
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08003200 * though, we'll note that we're not allowed to exceed surplus
3201 * and won't grow the pool anywhere else. Not until one of the
3202 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07003203 */
Andi Kleena5516432008-07-23 21:27:41 -07003204 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07003205 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08003206 try_to_free_low(h, min_count, nodes_allowed);
Mike Kravetz10c6ec42021-05-04 18:35:03 -07003207
3208 /*
3209 * Collect pages to be removed on list without dropping lock
3210 */
Andi Kleena5516432008-07-23 21:27:41 -07003211 while (min_count < persistent_huge_pages(h)) {
Mike Kravetz10c6ec42021-05-04 18:35:03 -07003212 page = remove_pool_huge_page(h, nodes_allowed, 0);
3213 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003214 break;
Mike Kravetz10c6ec42021-05-04 18:35:03 -07003215
3216 list_add(&page->lru, &page_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003217 }
Mike Kravetz10c6ec42021-05-04 18:35:03 -07003218 /* free the pages after dropping lock */
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003219 spin_unlock_irq(&hugetlb_lock);
Mike Kravetz10c6ec42021-05-04 18:35:03 -07003220 update_and_free_pages_bulk(h, &page_list);
Muchun Songb65d4ad2021-06-30 18:47:17 -07003221 flush_free_hpage_work(h);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003222 spin_lock_irq(&hugetlb_lock);
Mike Kravetz10c6ec42021-05-04 18:35:03 -07003223
Andi Kleena5516432008-07-23 21:27:41 -07003224 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08003225 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07003226 break;
3227 }
3228out:
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07003229 h->max_huge_pages = persistent_huge_pages(h);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003230 spin_unlock_irq(&hugetlb_lock);
Mike Kravetz29383962021-05-04 18:34:52 -07003231 mutex_unlock(&h->resize_lock);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07003232
Mike Kravetzf60858f2019-09-23 15:37:35 -07003233 NODEMASK_FREE(node_alloc_noretry);
3234
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07003235 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236}
3237
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003238#define HSTATE_ATTR_RO(_name) \
3239 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3240
3241#define HSTATE_ATTR(_name) \
3242 static struct kobj_attribute _name##_attr = \
3243 __ATTR(_name, 0644, _name##_show, _name##_store)
3244
3245static struct kobject *hugepages_kobj;
3246static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3247
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003248static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3249
3250static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003251{
3252 int i;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003253
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003254 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003255 if (hstate_kobjs[i] == kobj) {
3256 if (nidp)
3257 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003258 return &hstates[i];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003259 }
3260
3261 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003262}
3263
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003264static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003265 struct kobj_attribute *attr, char *buf)
3266{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003267 struct hstate *h;
3268 unsigned long nr_huge_pages;
3269 int nid;
3270
3271 h = kobj_to_hstate(kobj, &nid);
3272 if (nid == NUMA_NO_NODE)
3273 nr_huge_pages = h->nr_huge_pages;
3274 else
3275 nr_huge_pages = h->nr_huge_pages_node[nid];
3276
Joe Perchesae7a9272020-12-14 19:14:42 -08003277 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003278}
Eric B Munsonadbe8722011-01-13 15:47:27 -08003279
David Rientjes238d3c12014-08-06 16:06:51 -07003280static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3281 struct hstate *h, int nid,
3282 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003283{
3284 int err;
Oscar Salvador2d0adf72019-05-13 17:19:23 -07003285 nodemask_t nodes_allowed, *n_mask;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003286
Oscar Salvador2d0adf72019-05-13 17:19:23 -07003287 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3288 return -EINVAL;
Eric B Munsonadbe8722011-01-13 15:47:27 -08003289
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003290 if (nid == NUMA_NO_NODE) {
3291 /*
3292 * global hstate attribute
3293 */
3294 if (!(obey_mempolicy &&
Oscar Salvador2d0adf72019-05-13 17:19:23 -07003295 init_nodemask_of_mempolicy(&nodes_allowed)))
3296 n_mask = &node_states[N_MEMORY];
3297 else
3298 n_mask = &nodes_allowed;
3299 } else {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003300 /*
Mike Kravetzfd875dc2019-05-13 17:19:20 -07003301 * Node specific request. count adjustment happens in
3302 * set_max_huge_pages() after acquiring hugetlb_lock.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003303 */
Oscar Salvador2d0adf72019-05-13 17:19:23 -07003304 init_nodemask_of_node(&nodes_allowed, nid);
3305 n_mask = &nodes_allowed;
Mike Kravetzfd875dc2019-05-13 17:19:20 -07003306 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003307
Oscar Salvador2d0adf72019-05-13 17:19:23 -07003308 err = set_max_huge_pages(h, count, nid, n_mask);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003309
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07003310 return err ? err : len;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003311}
3312
David Rientjes238d3c12014-08-06 16:06:51 -07003313static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3314 struct kobject *kobj, const char *buf,
3315 size_t len)
3316{
3317 struct hstate *h;
3318 unsigned long count;
3319 int nid;
3320 int err;
3321
3322 err = kstrtoul(buf, 10, &count);
3323 if (err)
3324 return err;
3325
3326 h = kobj_to_hstate(kobj, &nid);
3327 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3328}
3329
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003330static ssize_t nr_hugepages_show(struct kobject *kobj,
3331 struct kobj_attribute *attr, char *buf)
3332{
3333 return nr_hugepages_show_common(kobj, attr, buf);
3334}
3335
3336static ssize_t nr_hugepages_store(struct kobject *kobj,
3337 struct kobj_attribute *attr, const char *buf, size_t len)
3338{
David Rientjes238d3c12014-08-06 16:06:51 -07003339 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003340}
3341HSTATE_ATTR(nr_hugepages);
3342
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003343#ifdef CONFIG_NUMA
3344
3345/*
3346 * hstate attribute for optionally mempolicy-based constraint on persistent
3347 * huge page alloc/free.
3348 */
3349static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
Joe Perchesae7a9272020-12-14 19:14:42 -08003350 struct kobj_attribute *attr,
3351 char *buf)
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003352{
3353 return nr_hugepages_show_common(kobj, attr, buf);
3354}
3355
3356static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3357 struct kobj_attribute *attr, const char *buf, size_t len)
3358{
David Rientjes238d3c12014-08-06 16:06:51 -07003359 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003360}
3361HSTATE_ATTR(nr_hugepages_mempolicy);
3362#endif
3363
3364
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003365static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3366 struct kobj_attribute *attr, char *buf)
3367{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003368 struct hstate *h = kobj_to_hstate(kobj, NULL);
Joe Perchesae7a9272020-12-14 19:14:42 -08003369 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003370}
Eric B Munsonadbe8722011-01-13 15:47:27 -08003371
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003372static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3373 struct kobj_attribute *attr, const char *buf, size_t count)
3374{
3375 int err;
3376 unsigned long input;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003377 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003378
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07003379 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08003380 return -EINVAL;
3381
Jingoo Han3dbb95f2013-09-11 14:20:25 -07003382 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003383 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08003384 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003385
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003386 spin_lock_irq(&hugetlb_lock);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003387 h->nr_overcommit_huge_pages = input;
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003388 spin_unlock_irq(&hugetlb_lock);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003389
3390 return count;
3391}
3392HSTATE_ATTR(nr_overcommit_hugepages);
3393
3394static ssize_t free_hugepages_show(struct kobject *kobj,
3395 struct kobj_attribute *attr, char *buf)
3396{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003397 struct hstate *h;
3398 unsigned long free_huge_pages;
3399 int nid;
3400
3401 h = kobj_to_hstate(kobj, &nid);
3402 if (nid == NUMA_NO_NODE)
3403 free_huge_pages = h->free_huge_pages;
3404 else
3405 free_huge_pages = h->free_huge_pages_node[nid];
3406
Joe Perchesae7a9272020-12-14 19:14:42 -08003407 return sysfs_emit(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003408}
3409HSTATE_ATTR_RO(free_hugepages);
3410
3411static ssize_t resv_hugepages_show(struct kobject *kobj,
3412 struct kobj_attribute *attr, char *buf)
3413{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003414 struct hstate *h = kobj_to_hstate(kobj, NULL);
Joe Perchesae7a9272020-12-14 19:14:42 -08003415 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003416}
3417HSTATE_ATTR_RO(resv_hugepages);
3418
3419static ssize_t surplus_hugepages_show(struct kobject *kobj,
3420 struct kobj_attribute *attr, char *buf)
3421{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003422 struct hstate *h;
3423 unsigned long surplus_huge_pages;
3424 int nid;
3425
3426 h = kobj_to_hstate(kobj, &nid);
3427 if (nid == NUMA_NO_NODE)
3428 surplus_huge_pages = h->surplus_huge_pages;
3429 else
3430 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3431
Joe Perchesae7a9272020-12-14 19:14:42 -08003432 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003433}
3434HSTATE_ATTR_RO(surplus_hugepages);
3435
3436static struct attribute *hstate_attrs[] = {
3437 &nr_hugepages_attr.attr,
3438 &nr_overcommit_hugepages_attr.attr,
3439 &free_hugepages_attr.attr,
3440 &resv_hugepages_attr.attr,
3441 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003442#ifdef CONFIG_NUMA
3443 &nr_hugepages_mempolicy_attr.attr,
3444#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003445 NULL,
3446};
3447
Arvind Yadav67e5ed92017-09-06 16:22:06 -07003448static const struct attribute_group hstate_attr_group = {
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003449 .attrs = hstate_attrs,
3450};
3451
Jeff Mahoney094e9532010-02-02 13:44:14 -08003452static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3453 struct kobject **hstate_kobjs,
Arvind Yadav67e5ed92017-09-06 16:22:06 -07003454 const struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003455{
3456 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003457 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003458
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003459 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3460 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003461 return -ENOMEM;
3462
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003463 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Miaohe Lincc2205a2021-02-24 12:06:50 -08003464 if (retval) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003465 kobject_put(hstate_kobjs[hi]);
Miaohe Lincc2205a2021-02-24 12:06:50 -08003466 hstate_kobjs[hi] = NULL;
3467 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003468
3469 return retval;
3470}
3471
3472static void __init hugetlb_sysfs_init(void)
3473{
3474 struct hstate *h;
3475 int err;
3476
3477 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3478 if (!hugepages_kobj)
3479 return;
3480
3481 for_each_hstate(h) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003482 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3483 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003484 if (err)
Mike Kravetz282f4212020-06-03 16:00:46 -07003485 pr_err("HugeTLB: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003486 }
3487}
3488
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003489#ifdef CONFIG_NUMA
3490
3491/*
3492 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08003493 * with node devices in node_devices[] using a parallel array. The array
3494 * index of a node device or _hstate == node id.
3495 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003496 * the base kernel, on the hugetlb module.
3497 */
3498struct node_hstate {
3499 struct kobject *hugepages_kobj;
3500 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3501};
Alexander Kuleshovb4e289a2015-11-05 18:50:14 -08003502static struct node_hstate node_hstates[MAX_NUMNODES];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003503
3504/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003505 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003506 */
3507static struct attribute *per_node_hstate_attrs[] = {
3508 &nr_hugepages_attr.attr,
3509 &free_hugepages_attr.attr,
3510 &surplus_hugepages_attr.attr,
3511 NULL,
3512};
3513
Arvind Yadav67e5ed92017-09-06 16:22:06 -07003514static const struct attribute_group per_node_hstate_attr_group = {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003515 .attrs = per_node_hstate_attrs,
3516};
3517
3518/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003519 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003520 * Returns node id via non-NULL nidp.
3521 */
3522static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3523{
3524 int nid;
3525
3526 for (nid = 0; nid < nr_node_ids; nid++) {
3527 struct node_hstate *nhs = &node_hstates[nid];
3528 int i;
3529 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3530 if (nhs->hstate_kobjs[i] == kobj) {
3531 if (nidp)
3532 *nidp = nid;
3533 return &hstates[i];
3534 }
3535 }
3536
3537 BUG();
3538 return NULL;
3539}
3540
3541/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003542 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003543 * No-op if no hstate attributes attached.
3544 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02003545static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003546{
3547 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08003548 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003549
3550 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08003551 return; /* no hstate attributes */
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003552
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003553 for_each_hstate(h) {
3554 int idx = hstate_index(h);
3555 if (nhs->hstate_kobjs[idx]) {
3556 kobject_put(nhs->hstate_kobjs[idx]);
3557 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003558 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003559 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003560
3561 kobject_put(nhs->hugepages_kobj);
3562 nhs->hugepages_kobj = NULL;
3563}
3564
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003565
3566/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003567 * Register hstate attributes for a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003568 * No-op if attributes already registered.
3569 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02003570static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003571{
3572 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08003573 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003574 int err;
3575
3576 if (nhs->hugepages_kobj)
3577 return; /* already allocated */
3578
3579 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08003580 &node->dev.kobj);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003581 if (!nhs->hugepages_kobj)
3582 return;
3583
3584 for_each_hstate(h) {
3585 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3586 nhs->hstate_kobjs,
3587 &per_node_hstate_attr_group);
3588 if (err) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003589 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08003590 h->name, node->dev.id);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003591 hugetlb_unregister_node(node);
3592 break;
3593 }
3594 }
3595}
3596
3597/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08003598 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08003599 * devices of nodes that have memory. All on-line nodes should have
3600 * registered their associated device by this time.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003601 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08003602static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003603{
3604 int nid;
3605
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08003606 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08003607 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08003608 if (node->dev.id == nid)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003609 hugetlb_register_node(node);
3610 }
3611
3612 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003613 * Let the node device driver know we're here so it can
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003614 * [un]register hstate attributes on node hotplug.
3615 */
3616 register_hugetlbfs_with_node(hugetlb_register_node,
3617 hugetlb_unregister_node);
3618}
3619#else /* !CONFIG_NUMA */
3620
3621static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3622{
3623 BUG();
3624 if (nidp)
3625 *nidp = -1;
3626 return NULL;
3627}
3628
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003629static void hugetlb_register_all_nodes(void) { }
3630
3631#endif
3632
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003633static int __init hugetlb_init(void)
3634{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003635 int i;
3636
Mike Kravetzd6995da2021-02-24 12:08:51 -08003637 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3638 __NR_HPAGEFLAGS);
3639
Mike Kravetzc2833a52020-06-03 16:00:50 -07003640 if (!hugepages_supported()) {
3641 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3642 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07003643 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003644 }
Vaishali Thakkarf8b74812016-02-17 13:11:26 -08003645
Mike Kravetz282f4212020-06-03 16:00:46 -07003646 /*
3647 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3648 * architectures depend on setup being done here.
3649 */
3650 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3651 if (!parsed_default_hugepagesz) {
3652 /*
3653 * If we did not parse a default huge page size, set
3654 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3655 * number of huge pages for this default size was implicitly
3656 * specified, set that here as well.
3657 * Note that the implicit setting will overwrite an explicit
3658 * setting. A warning will be printed in this case.
3659 */
3660 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3661 if (default_hstate_max_huge_pages) {
3662 if (default_hstate.max_huge_pages) {
3663 char buf[32];
Andi Kleenaa888a72008-07-23 21:27:47 -07003664
Mike Kravetz282f4212020-06-03 16:00:46 -07003665 string_get_size(huge_page_size(&default_hstate),
3666 1, STRING_UNITS_2, buf, 32);
3667 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3668 default_hstate.max_huge_pages, buf);
3669 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3670 default_hstate_max_huge_pages);
3671 }
3672 default_hstate.max_huge_pages =
3673 default_hstate_max_huge_pages;
3674 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003675 }
Andi Kleenaa888a72008-07-23 21:27:47 -07003676
Roman Gushchincf11e852020-04-10 14:32:45 -07003677 hugetlb_cma_check();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003678 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07003679 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003680 report_hugepages();
3681
3682 hugetlb_sysfs_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003683 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08003684 hugetlb_cgroup_file_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003685
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003686#ifdef CONFIG_SMP
3687 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3688#else
3689 num_fault_mutexes = 1;
3690#endif
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003691 hugetlb_fault_mutex_table =
Kees Cook6da2ec52018-06-12 13:55:00 -07003692 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3693 GFP_KERNEL);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003694 BUG_ON(!hugetlb_fault_mutex_table);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003695
3696 for (i = 0; i < num_fault_mutexes; i++)
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003697 mutex_init(&hugetlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003698 return 0;
3699}
Paul Gortmaker3e89e1c2016-01-14 15:21:52 -08003700subsys_initcall(hugetlb_init);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003701
Mike Kravetzae94da82020-06-03 16:00:34 -07003702/* Overwritten by architectures with more huge page sizes */
3703bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003704{
Mike Kravetzae94da82020-06-03 16:00:34 -07003705 return size == HPAGE_SIZE;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003706}
3707
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08003708void __init hugetlb_add_hstate(unsigned int order)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003709{
3710 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003711 unsigned long i;
3712
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003713 if (size_to_hstate(PAGE_SIZE << order)) {
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003714 return;
3715 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07003716 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003717 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07003718 h = &hstates[hugetlb_max_hstate++];
Mike Kravetz29383962021-05-04 18:34:52 -07003719 mutex_init(&h->resize_lock);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003720 h->order = order;
Miaohe Linaca78302021-02-24 12:07:46 -08003721 h->mask = ~(huge_page_size(h) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07003722 for (i = 0; i < MAX_NUMNODES; ++i)
3723 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07003724 INIT_LIST_HEAD(&h->hugepage_activelist);
Andrew Morton54f18d32016-05-19 17:11:40 -07003725 h->next_nid_to_alloc = first_memory_node;
3726 h->next_nid_to_free = first_memory_node;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003727 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3728 huge_page_size(h)/1024);
Muchun Song77490582021-06-30 18:47:33 -07003729 hugetlb_vmemmap_init(h);
Andi Kleen8faa8b02008-07-23 21:27:48 -07003730
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003731 parsed_hstate = h;
3732}
3733
Mike Kravetz282f4212020-06-03 16:00:46 -07003734/*
3735 * hugepages command line processing
3736 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3737 * specification. If not, ignore the hugepages value. hugepages can also
3738 * be the first huge page command line option in which case it implicitly
3739 * specifies the number of huge pages for the default size.
3740 */
3741static int __init hugepages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003742{
3743 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003744 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003745
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003746 if (!parsed_valid_hugepagesz) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003747 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003748 parsed_valid_hugepagesz = true;
Mike Kravetz282f4212020-06-03 16:00:46 -07003749 return 0;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003750 }
Mike Kravetz282f4212020-06-03 16:00:46 -07003751
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003752 /*
Mike Kravetz282f4212020-06-03 16:00:46 -07003753 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3754 * yet, so this hugepages= parameter goes to the "default hstate".
3755 * Otherwise, it goes with the previously parsed hugepagesz or
3756 * default_hugepagesz.
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003757 */
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003758 else if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003759 mhp = &default_hstate_max_huge_pages;
3760 else
3761 mhp = &parsed_hstate->max_huge_pages;
3762
Andi Kleen8faa8b02008-07-23 21:27:48 -07003763 if (mhp == last_mhp) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003764 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3765 return 0;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003766 }
3767
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003768 if (sscanf(s, "%lu", mhp) <= 0)
3769 *mhp = 0;
3770
Andi Kleen8faa8b02008-07-23 21:27:48 -07003771 /*
3772 * Global state is always initialized later in hugetlb_init.
Miaohe Lin04adbc32021-05-04 18:33:22 -07003773 * But we need to allocate gigantic hstates here early to still
Andi Kleen8faa8b02008-07-23 21:27:48 -07003774 * use the bootmem allocator.
3775 */
Miaohe Lin04adbc32021-05-04 18:33:22 -07003776 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
Andi Kleen8faa8b02008-07-23 21:27:48 -07003777 hugetlb_hstate_alloc_pages(parsed_hstate);
3778
3779 last_mhp = mhp;
3780
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003781 return 1;
3782}
Mike Kravetz282f4212020-06-03 16:00:46 -07003783__setup("hugepages=", hugepages_setup);
Nick Piggine11bfbf2008-07-23 21:27:52 -07003784
Mike Kravetz282f4212020-06-03 16:00:46 -07003785/*
3786 * hugepagesz command line processing
3787 * A specific huge page size can only be specified once with hugepagesz.
3788 * hugepagesz is followed by hugepages on the command line. The global
3789 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3790 * hugepagesz argument was valid.
3791 */
Mike Kravetz359f2542020-06-03 16:00:38 -07003792static int __init hugepagesz_setup(char *s)
Nick Piggine11bfbf2008-07-23 21:27:52 -07003793{
Mike Kravetz359f2542020-06-03 16:00:38 -07003794 unsigned long size;
Mike Kravetz282f4212020-06-03 16:00:46 -07003795 struct hstate *h;
3796
3797 parsed_valid_hugepagesz = false;
Mike Kravetz359f2542020-06-03 16:00:38 -07003798 size = (unsigned long)memparse(s, NULL);
3799
3800 if (!arch_hugetlb_valid_size(size)) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003801 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
Mike Kravetz359f2542020-06-03 16:00:38 -07003802 return 0;
3803 }
3804
Mike Kravetz282f4212020-06-03 16:00:46 -07003805 h = size_to_hstate(size);
3806 if (h) {
3807 /*
3808 * hstate for this size already exists. This is normally
3809 * an error, but is allowed if the existing hstate is the
3810 * default hstate. More specifically, it is only allowed if
3811 * the number of huge pages for the default hstate was not
3812 * previously specified.
3813 */
3814 if (!parsed_default_hugepagesz || h != &default_hstate ||
3815 default_hstate.max_huge_pages) {
3816 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3817 return 0;
3818 }
3819
3820 /*
3821 * No need to call hugetlb_add_hstate() as hstate already
3822 * exists. But, do set parsed_hstate so that a following
3823 * hugepages= parameter will be applied to this hstate.
3824 */
3825 parsed_hstate = h;
3826 parsed_valid_hugepagesz = true;
3827 return 1;
Mike Kravetz38237832020-06-03 16:00:42 -07003828 }
3829
Mike Kravetz359f2542020-06-03 16:00:38 -07003830 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
Mike Kravetz282f4212020-06-03 16:00:46 -07003831 parsed_valid_hugepagesz = true;
Nick Piggine11bfbf2008-07-23 21:27:52 -07003832 return 1;
3833}
Mike Kravetz359f2542020-06-03 16:00:38 -07003834__setup("hugepagesz=", hugepagesz_setup);
3835
Mike Kravetz282f4212020-06-03 16:00:46 -07003836/*
3837 * default_hugepagesz command line input
3838 * Only one instance of default_hugepagesz allowed on command line.
3839 */
Mike Kravetzae94da82020-06-03 16:00:34 -07003840static int __init default_hugepagesz_setup(char *s)
Nick Piggine11bfbf2008-07-23 21:27:52 -07003841{
Mike Kravetzae94da82020-06-03 16:00:34 -07003842 unsigned long size;
3843
Mike Kravetz282f4212020-06-03 16:00:46 -07003844 parsed_valid_hugepagesz = false;
Mike Kravetz282f4212020-06-03 16:00:46 -07003845 if (parsed_default_hugepagesz) {
3846 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3847 return 0;
3848 }
3849
3850 size = (unsigned long)memparse(s, NULL);
3851
3852 if (!arch_hugetlb_valid_size(size)) {
3853 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3854 return 0;
3855 }
3856
3857 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3858 parsed_valid_hugepagesz = true;
3859 parsed_default_hugepagesz = true;
3860 default_hstate_idx = hstate_index(size_to_hstate(size));
3861
3862 /*
3863 * The number of default huge pages (for this size) could have been
3864 * specified as the first hugetlb parameter: hugepages=X. If so,
3865 * then default_hstate_max_huge_pages is set. If the default huge
3866 * page size is gigantic (>= MAX_ORDER), then the pages must be
3867 * allocated here from bootmem allocator.
3868 */
3869 if (default_hstate_max_huge_pages) {
3870 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3871 if (hstate_is_gigantic(&default_hstate))
3872 hugetlb_hstate_alloc_pages(&default_hstate);
3873 default_hstate_max_huge_pages = 0;
3874 }
3875
Nick Piggine11bfbf2008-07-23 21:27:52 -07003876 return 1;
3877}
Mike Kravetzae94da82020-06-03 16:00:34 -07003878__setup("default_hugepagesz=", default_hugepagesz_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003879
Muchun Song8ca39e62020-08-11 18:30:32 -07003880static unsigned int allowed_mems_nr(struct hstate *h)
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003881{
3882 int node;
3883 unsigned int nr = 0;
Muchun Song8ca39e62020-08-11 18:30:32 -07003884 nodemask_t *mpol_allowed;
3885 unsigned int *array = h->free_huge_pages_node;
3886 gfp_t gfp_mask = htlb_alloc_mask(h);
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003887
Muchun Song8ca39e62020-08-11 18:30:32 -07003888 mpol_allowed = policy_nodemask_current(gfp_mask);
3889
3890 for_each_node_mask(node, cpuset_current_mems_allowed) {
Jiapeng Zhongc93b0a92021-02-24 12:07:09 -08003891 if (!mpol_allowed || node_isset(node, *mpol_allowed))
Muchun Song8ca39e62020-08-11 18:30:32 -07003892 nr += array[node];
3893 }
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003894
3895 return nr;
3896}
3897
3898#ifdef CONFIG_SYSCTL
Muchun Song17743792020-09-04 16:36:13 -07003899static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3900 void *buffer, size_t *length,
3901 loff_t *ppos, unsigned long *out)
3902{
3903 struct ctl_table dup_table;
3904
3905 /*
3906 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3907 * can duplicate the @table and alter the duplicate of it.
3908 */
3909 dup_table = *table;
3910 dup_table.data = out;
3911
3912 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3913}
3914
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003915static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3916 struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003917 void *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003918{
Andi Kleene5ff2152008-07-23 21:27:42 -07003919 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07003920 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08003921 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07003922
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003923 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08003924 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003925
Muchun Song17743792020-09-04 16:36:13 -07003926 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3927 &tmp);
Michal Hocko08d4a242011-01-13 15:47:26 -08003928 if (ret)
3929 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07003930
David Rientjes238d3c12014-08-06 16:06:51 -07003931 if (write)
3932 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3933 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08003934out:
3935 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936}
Mel Gorman396faf02007-07-17 04:03:13 -07003937
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003938int hugetlb_sysctl_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003939 void *buffer, size_t *length, loff_t *ppos)
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003940{
3941
3942 return hugetlb_sysctl_handler_common(false, table, write,
3943 buffer, length, ppos);
3944}
3945
3946#ifdef CONFIG_NUMA
3947int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003948 void *buffer, size_t *length, loff_t *ppos)
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003949{
3950 return hugetlb_sysctl_handler_common(true, table, write,
3951 buffer, length, ppos);
3952}
3953#endif /* CONFIG_NUMA */
3954
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003955int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003956 void *buffer, size_t *length, loff_t *ppos)
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003957{
Andi Kleena5516432008-07-23 21:27:41 -07003958 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07003959 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08003960 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07003961
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003962 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08003963 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003964
Petr Holasekc033a932011-03-22 16:33:05 -07003965 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07003966
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07003967 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08003968 return -EINVAL;
3969
Muchun Song17743792020-09-04 16:36:13 -07003970 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3971 &tmp);
Michal Hocko08d4a242011-01-13 15:47:26 -08003972 if (ret)
3973 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07003974
3975 if (write) {
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003976 spin_lock_irq(&hugetlb_lock);
Andi Kleene5ff2152008-07-23 21:27:42 -07003977 h->nr_overcommit_huge_pages = tmp;
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003978 spin_unlock_irq(&hugetlb_lock);
Andi Kleene5ff2152008-07-23 21:27:42 -07003979 }
Michal Hocko08d4a242011-01-13 15:47:26 -08003980out:
3981 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003982}
3983
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984#endif /* CONFIG_SYSCTL */
3985
Alexey Dobriyane1759c22008-10-15 23:50:22 +04003986void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003987{
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08003988 struct hstate *h;
3989 unsigned long total = 0;
3990
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003991 if (!hugepages_supported())
3992 return;
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08003993
3994 for_each_hstate(h) {
3995 unsigned long count = h->nr_huge_pages;
3996
Miaohe Linaca78302021-02-24 12:07:46 -08003997 total += huge_page_size(h) * count;
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08003998
3999 if (h == &default_hstate)
4000 seq_printf(m,
4001 "HugePages_Total: %5lu\n"
4002 "HugePages_Free: %5lu\n"
4003 "HugePages_Rsvd: %5lu\n"
4004 "HugePages_Surp: %5lu\n"
4005 "Hugepagesize: %8lu kB\n",
4006 count,
4007 h->free_huge_pages,
4008 h->resv_huge_pages,
4009 h->surplus_huge_pages,
Miaohe Linaca78302021-02-24 12:07:46 -08004010 huge_page_size(h) / SZ_1K);
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08004011 }
4012
Miaohe Linaca78302021-02-24 12:07:46 -08004013 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004014}
4015
Joe Perches79815932020-09-16 13:40:43 -07004016int hugetlb_report_node_meminfo(char *buf, int len, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017{
Andi Kleena5516432008-07-23 21:27:41 -07004018 struct hstate *h = &default_hstate;
Joe Perches79815932020-09-16 13:40:43 -07004019
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07004020 if (!hugepages_supported())
4021 return 0;
Joe Perches79815932020-09-16 13:40:43 -07004022
4023 return sysfs_emit_at(buf, len,
4024 "Node %d HugePages_Total: %5u\n"
4025 "Node %d HugePages_Free: %5u\n"
4026 "Node %d HugePages_Surp: %5u\n",
4027 nid, h->nr_huge_pages_node[nid],
4028 nid, h->free_huge_pages_node[nid],
4029 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030}
4031
David Rientjes949f7ec2013-04-29 15:07:48 -07004032void hugetlb_show_meminfo(void)
4033{
4034 struct hstate *h;
4035 int nid;
4036
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07004037 if (!hugepages_supported())
4038 return;
4039
David Rientjes949f7ec2013-04-29 15:07:48 -07004040 for_each_node_state(nid, N_MEMORY)
4041 for_each_hstate(h)
4042 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4043 nid,
4044 h->nr_huge_pages_node[nid],
4045 h->free_huge_pages_node[nid],
4046 h->surplus_huge_pages_node[nid],
Miaohe Linaca78302021-02-24 12:07:46 -08004047 huge_page_size(h) / SZ_1K);
David Rientjes949f7ec2013-04-29 15:07:48 -07004048}
4049
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08004050void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4051{
4052 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4053 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4054}
4055
Linus Torvalds1da177e2005-04-16 15:20:36 -07004056/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4057unsigned long hugetlb_total_pages(void)
4058{
Wanpeng Lid0028582013-03-22 15:04:40 -07004059 struct hstate *h;
4060 unsigned long nr_total_pages = 0;
4061
4062 for_each_hstate(h)
4063 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4064 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004066
Andi Kleena5516432008-07-23 21:27:41 -07004067static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07004068{
4069 int ret = -ENOMEM;
4070
Miaohe Lin0aa7f352021-02-24 12:06:57 -08004071 if (!delta)
4072 return 0;
4073
Mike Kravetzdb71ef72021-05-04 18:35:07 -07004074 spin_lock_irq(&hugetlb_lock);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07004075 /*
4076 * When cpuset is configured, it breaks the strict hugetlb page
4077 * reservation as the accounting is done on a global variable. Such
4078 * reservation is completely rubbish in the presence of cpuset because
4079 * the reservation is not checked against page availability for the
4080 * current cpuset. Application can still potentially OOM'ed by kernel
4081 * with lack of free htlb page in cpuset that the task is in.
4082 * Attempt to enforce strict accounting with cpuset is almost
4083 * impossible (or too ugly) because cpuset is too fluid that
4084 * task or memory node can be dynamically moved between cpusets.
4085 *
4086 * The change of semantics for shared hugetlb mapping with cpuset is
4087 * undesirable. However, in order to preserve some of the semantics,
4088 * we fall back to check against current free page availability as
4089 * a best attempt and hopefully to minimize the impact of changing
4090 * semantics that cpuset has.
Muchun Song8ca39e62020-08-11 18:30:32 -07004091 *
4092 * Apart from cpuset, we also have memory policy mechanism that
4093 * also determines from which node the kernel will allocate memory
4094 * in a NUMA system. So similar to cpuset, we also should consider
4095 * the memory policy of the current task. Similar to the description
4096 * above.
Mel Gormanfc1b8a72008-07-23 21:27:22 -07004097 */
4098 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07004099 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07004100 goto out;
4101
Muchun Song8ca39e62020-08-11 18:30:32 -07004102 if (delta > allowed_mems_nr(h)) {
Andi Kleena5516432008-07-23 21:27:41 -07004103 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07004104 goto out;
4105 }
4106 }
4107
4108 ret = 0;
4109 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07004110 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07004111
4112out:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07004113 spin_unlock_irq(&hugetlb_lock);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07004114 return ret;
4115}
4116
Andy Whitcroft84afd992008-07-23 21:27:32 -07004117static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4118{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07004119 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07004120
4121 /*
4122 * This new VMA should share its siblings reservation map if present.
4123 * The VMA will only ever have a valid reservation map pointer where
4124 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03004125 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07004126 * after this open call completes. It is therefore safe to take a
4127 * new reference here without additional locking.
4128 */
Mike Kravetz09a26e82021-09-02 14:58:53 -07004129 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4130 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07004131 kref_get(&resv->refs);
Mike Kravetz09a26e82021-09-02 14:58:53 -07004132 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07004133}
4134
Mel Gormana1e78772008-07-23 21:27:23 -07004135static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4136{
Andi Kleena5516432008-07-23 21:27:41 -07004137 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07004138 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07004139 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004140 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004141 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07004142
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004143 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4144 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07004145
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004146 start = vma_hugecache_offset(h, vma, vma->vm_start);
4147 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07004148
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004149 reserve = (end - start) - region_count(resv, start, end);
Mina Almasrye9fe92a2020-04-01 21:11:21 -07004150 hugetlb_cgroup_uncharge_counter(resv, start, end);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004151 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004152 /*
4153 * Decrement reserve counts. The global reserve count may be
4154 * adjusted if the subpool has a minimum size.
4155 */
4156 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4157 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07004158 }
Mina Almasrye9fe92a2020-04-01 21:11:21 -07004159
4160 kref_put(&resv->refs, resv_map_release);
Mel Gormana1e78772008-07-23 21:27:23 -07004161}
4162
Dan Williams31383c62017-11-29 16:10:28 -08004163static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4164{
4165 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4166 return -EINVAL;
4167 return 0;
4168}
4169
Dan Williams05ea8862018-04-05 16:24:25 -07004170static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4171{
Miaohe Linaca78302021-02-24 12:07:46 -08004172 return huge_page_size(hstate_vma(vma));
Dan Williams05ea8862018-04-05 16:24:25 -07004173}
4174
Linus Torvalds1da177e2005-04-16 15:20:36 -07004175/*
4176 * We cannot handle pagefaults against hugetlb pages at all. They cause
4177 * handle_mm_fault() to try to instantiate regular-sized pages in the
Miaohe Lin6c26d312021-02-24 12:07:19 -08004178 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
Linus Torvalds1da177e2005-04-16 15:20:36 -07004179 * this far.
4180 */
Souptick Joarderb3ec9f32018-06-07 17:08:04 -07004181static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004182{
4183 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07004184 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004185}
4186
Jane Chueec36362018-08-02 15:36:05 -07004187/*
4188 * When a new function is introduced to vm_operations_struct and added
4189 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4190 * This is because under System V memory model, mappings created via
4191 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4192 * their original vm_ops are overwritten with shm_vm_ops.
4193 */
Alexey Dobriyanf0f37e2f2009-09-27 22:29:37 +04004194const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07004195 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07004196 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07004197 .close = hugetlb_vm_op_close,
Dmitry Safonovdd3b6142020-12-14 19:08:17 -08004198 .may_split = hugetlb_vm_op_split,
Dan Williams05ea8862018-04-05 16:24:25 -07004199 .pagesize = hugetlb_vm_op_pagesize,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004200};
4201
David Gibson1e8f8892006-01-06 00:10:44 -08004202static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4203 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07004204{
4205 pte_t entry;
Christophe Leroy79c1c592021-06-30 18:48:00 -07004206 unsigned int shift = huge_page_shift(hstate_vma(vma));
David Gibson63551ae2005-06-21 17:14:44 -07004207
David Gibson1e8f8892006-01-06 00:10:44 -08004208 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07004209 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4210 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07004211 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07004212 entry = huge_pte_wrprotect(mk_huge_pte(page,
4213 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07004214 }
4215 entry = pte_mkyoung(entry);
4216 entry = pte_mkhuge(entry);
Christophe Leroy79c1c592021-06-30 18:48:00 -07004217 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
David Gibson63551ae2005-06-21 17:14:44 -07004218
4219 return entry;
4220}
4221
David Gibson1e8f8892006-01-06 00:10:44 -08004222static void set_huge_ptep_writable(struct vm_area_struct *vma,
4223 unsigned long address, pte_t *ptep)
4224{
4225 pte_t entry;
4226
Gerald Schaefer106c9922013-04-29 15:07:23 -07004227 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07004228 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00004229 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08004230}
4231
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07004232bool is_hugetlb_entry_migration(pte_t pte)
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004233{
4234 swp_entry_t swp;
4235
4236 if (huge_pte_none(pte) || pte_present(pte))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07004237 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004238 swp = pte_to_swp_entry(pte);
Baoquan Hed79d1762020-10-13 16:56:14 -07004239 if (is_migration_entry(swp))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07004240 return true;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004241 else
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07004242 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004243}
4244
Baoquan He3e5c3602020-10-13 16:56:10 -07004245static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004246{
4247 swp_entry_t swp;
4248
4249 if (huge_pte_none(pte) || pte_present(pte))
Baoquan He3e5c3602020-10-13 16:56:10 -07004250 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004251 swp = pte_to_swp_entry(pte);
Baoquan Hed79d1762020-10-13 16:56:14 -07004252 if (is_hwpoison_entry(swp))
Baoquan He3e5c3602020-10-13 16:56:10 -07004253 return true;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004254 else
Baoquan He3e5c3602020-10-13 16:56:10 -07004255 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004256}
David Gibson1e8f8892006-01-06 00:10:44 -08004257
Peter Xu4eae4ef2021-03-12 21:07:33 -08004258static void
4259hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4260 struct page *new_page)
4261{
4262 __SetPageUptodate(new_page);
4263 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4264 hugepage_add_new_anon_rmap(new_page, vma, addr);
4265 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4266 ClearHPageRestoreReserve(new_page);
4267 SetHPageMigratable(new_page);
4268}
4269
David Gibson63551ae2005-06-21 17:14:44 -07004270int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4271 struct vm_area_struct *vma)
4272{
Mike Kravetz5e415402018-11-16 15:08:04 -08004273 pte_t *src_pte, *dst_pte, entry, dst_entry;
David Gibson63551ae2005-06-21 17:14:44 -07004274 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07004275 unsigned long addr;
Peter Xuca6eb142021-03-12 21:07:30 -08004276 bool cow = is_cow_mapping(vma->vm_flags);
Andi Kleena5516432008-07-23 21:27:41 -07004277 struct hstate *h = hstate_vma(vma);
4278 unsigned long sz = huge_page_size(h);
Peter Xu4eae4ef2021-03-12 21:07:33 -08004279 unsigned long npages = pages_per_huge_page(h);
Mike Kravetzc0d03812020-04-01 21:11:05 -07004280 struct address_space *mapping = vma->vm_file->f_mapping;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004281 struct mmu_notifier_range range;
Andreas Sandberge8569dd2014-01-21 15:49:09 -08004282 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08004283
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004284 if (cow) {
Jérôme Glisse7269f992019-05-13 17:20:53 -07004285 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07004286 vma->vm_start,
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004287 vma->vm_end);
4288 mmu_notifier_invalidate_range_start(&range);
Mike Kravetzc0d03812020-04-01 21:11:05 -07004289 } else {
4290 /*
4291 * For shared mappings i_mmap_rwsem must be held to call
4292 * huge_pte_alloc, otherwise the returned ptep could go
4293 * away if part of a shared pmd and another thread calls
4294 * huge_pmd_unshare.
4295 */
4296 i_mmap_lock_read(mapping);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004297 }
Andreas Sandberge8569dd2014-01-21 15:49:09 -08004298
Andi Kleena5516432008-07-23 21:27:41 -07004299 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004300 spinlock_t *src_ptl, *dst_ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07004301 src_pte = huge_pte_offset(src, addr, sz);
Hugh Dickinsc74df322005-10-29 18:16:23 -07004302 if (!src_pte)
4303 continue;
Peter Xuaec44e02021-05-04 18:33:00 -07004304 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08004305 if (!dst_pte) {
4306 ret = -ENOMEM;
4307 break;
4308 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08004309
Mike Kravetz5e415402018-11-16 15:08:04 -08004310 /*
4311 * If the pagetables are shared don't copy or take references.
4312 * dst_pte == src_pte is the common case of src/dest sharing.
4313 *
4314 * However, src could have 'unshared' and dst shares with
4315 * another vma. If dst_pte !none, this implies sharing.
4316 * Check here before taking page table lock, and once again
4317 * after taking the lock below.
4318 */
4319 dst_entry = huge_ptep_get(dst_pte);
4320 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
Larry Woodmanc5c99422008-01-24 05:49:25 -08004321 continue;
4322
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004323 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4324 src_ptl = huge_pte_lockptr(h, src, src_pte);
4325 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004326 entry = huge_ptep_get(src_pte);
Mike Kravetz5e415402018-11-16 15:08:04 -08004327 dst_entry = huge_ptep_get(dst_pte);
Peter Xu4eae4ef2021-03-12 21:07:33 -08004328again:
Mike Kravetz5e415402018-11-16 15:08:04 -08004329 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
4330 /*
4331 * Skip if src entry none. Also, skip in the
4332 * unlikely case dst entry !none as this implies
4333 * sharing with another vma.
4334 */
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004335 ;
4336 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
4337 is_hugetlb_entry_hwpoisoned(entry))) {
4338 swp_entry_t swp_entry = pte_to_swp_entry(entry);
4339
Alistair Popple4dd845b2021-06-30 18:54:09 -07004340 if (is_writable_migration_entry(swp_entry) && cow) {
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004341 /*
4342 * COW mappings require pages in both
4343 * parent and child to be set to read.
4344 */
Alistair Popple4dd845b2021-06-30 18:54:09 -07004345 swp_entry = make_readable_migration_entry(
4346 swp_offset(swp_entry));
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004347 entry = swp_entry_to_pte(swp_entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07004348 set_huge_swap_pte_at(src, addr, src_pte,
4349 entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004350 }
Punit Agrawale5251fd2017-07-06 15:39:50 -07004351 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07004352 } else {
Peter Xu4eae4ef2021-03-12 21:07:33 -08004353 entry = huge_ptep_get(src_pte);
4354 ptepage = pte_page(entry);
4355 get_page(ptepage);
4356
4357 /*
4358 * This is a rare case where we see pinned hugetlb
4359 * pages while they're prone to COW. We need to do the
4360 * COW earlier during fork.
4361 *
4362 * When pre-allocating the page or copying data, we
4363 * need to be without the pgtable locks since we could
4364 * sleep during the process.
4365 */
4366 if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
4367 pte_t src_pte_old = entry;
4368 struct page *new;
4369
4370 spin_unlock(src_ptl);
4371 spin_unlock(dst_ptl);
4372 /* Do not use reserve as it's private owned */
4373 new = alloc_huge_page(vma, addr, 1);
4374 if (IS_ERR(new)) {
4375 put_page(ptepage);
4376 ret = PTR_ERR(new);
4377 break;
4378 }
4379 copy_user_huge_page(new, ptepage, addr, vma,
4380 npages);
4381 put_page(ptepage);
4382
4383 /* Install the new huge page if src pte stable */
4384 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4385 src_ptl = huge_pte_lockptr(h, src, src_pte);
4386 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4387 entry = huge_ptep_get(src_pte);
4388 if (!pte_same(src_pte_old, entry)) {
Mike Kravetz846be082021-06-15 18:23:29 -07004389 restore_reserve_on_error(h, vma, addr,
4390 new);
Peter Xu4eae4ef2021-03-12 21:07:33 -08004391 put_page(new);
4392 /* dst_entry won't change as in child */
4393 goto again;
4394 }
4395 hugetlb_install_page(vma, dst_pte, addr, new);
4396 spin_unlock(src_ptl);
4397 spin_unlock(dst_ptl);
4398 continue;
4399 }
4400
Joerg Roedel34ee6452014-11-13 13:46:09 +11004401 if (cow) {
Jérôme Glisse0f108512017-11-15 17:34:07 -08004402 /*
4403 * No need to notify as we are downgrading page
4404 * table protection not changing it to point
4405 * to a new page.
4406 *
Mike Rapoportad56b732018-03-21 21:22:47 +02004407 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08004408 */
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004409 huge_ptep_set_wrprotect(src, addr, src_pte);
Peter Xu84894e12021-05-14 17:27:07 -07004410 entry = huge_pte_wrprotect(entry);
Joerg Roedel34ee6452014-11-13 13:46:09 +11004411 }
Peter Xu4eae4ef2021-03-12 21:07:33 -08004412
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08004413 page_dup_rmap(ptepage, true);
Hugh Dickins1c598272005-10-19 21:23:43 -07004414 set_huge_pte_at(dst, addr, dst_pte, entry);
Peter Xu4eae4ef2021-03-12 21:07:33 -08004415 hugetlb_count_add(npages, dst);
Hugh Dickins1c598272005-10-19 21:23:43 -07004416 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004417 spin_unlock(src_ptl);
4418 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07004419 }
David Gibson63551ae2005-06-21 17:14:44 -07004420
Andreas Sandberge8569dd2014-01-21 15:49:09 -08004421 if (cow)
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004422 mmu_notifier_invalidate_range_end(&range);
Mike Kravetzc0d03812020-04-01 21:11:05 -07004423 else
4424 i_mmap_unlock_read(mapping);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08004425
4426 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07004427}
4428
Peter Xu73c54762021-11-05 13:41:17 -07004429static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
4430 unsigned long start, unsigned long end,
4431 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07004432{
4433 struct mm_struct *mm = vma->vm_mm;
4434 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07004435 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07004436 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004437 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07004438 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07004439 struct hstate *h = hstate_vma(vma);
4440 unsigned long sz = huge_page_size(h);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004441 struct mmu_notifier_range range;
Andi Kleena5516432008-07-23 21:27:41 -07004442
David Gibson63551ae2005-06-21 17:14:44 -07004443 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07004444 BUG_ON(start & ~huge_page_mask(h));
4445 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07004446
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08004447 /*
4448 * This is a hugetlb vma, all the pte entries should point
4449 * to huge page.
4450 */
Peter Zijlstraed6a7932018-08-31 14:46:08 +02004451 tlb_change_page_size(tlb, sz);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004452 tlb_start_vma(tlb, vma);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07004453
4454 /*
4455 * If sharing possible, alert mmu notifiers of worst case.
4456 */
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07004457 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
4458 end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004459 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4460 mmu_notifier_invalidate_range_start(&range);
Hillf Danton569f48b82014-12-10 15:44:41 -08004461 address = start;
Hillf Danton569f48b82014-12-10 15:44:41 -08004462 for (; address < end; address += sz) {
Punit Agrawal7868a202017-07-06 15:39:42 -07004463 ptep = huge_pte_offset(mm, address, sz);
Adam Litke4c887262005-10-29 18:16:46 -07004464 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07004465 continue;
4466
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004467 ptl = huge_pte_lock(h, mm, ptep);
Mike Kravetz34ae2042020-08-11 18:31:38 -07004468 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07004469 spin_unlock(ptl);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07004470 /*
4471 * We just unmapped a page of PMDs by clearing a PUD.
4472 * The caller's TLB flush range should cover this area.
4473 */
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07004474 continue;
4475 }
Chen, Kenneth W39dde652006-12-06 20:32:03 -08004476
Hillf Danton66293262012-03-23 15:01:48 -07004477 pte = huge_ptep_get(ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07004478 if (huge_pte_none(pte)) {
4479 spin_unlock(ptl);
4480 continue;
4481 }
Hillf Danton66293262012-03-23 15:01:48 -07004482
4483 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08004484 * Migrating hugepage or HWPoisoned hugepage is already
4485 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07004486 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08004487 if (unlikely(!pte_present(pte))) {
Punit Agrawal9386fac2017-07-06 15:39:46 -07004488 huge_pte_clear(mm, address, ptep, sz);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07004489 spin_unlock(ptl);
4490 continue;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08004491 }
Hillf Danton66293262012-03-23 15:01:48 -07004492
4493 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004494 /*
4495 * If a reference page is supplied, it is because a specific
4496 * page is being unmapped, not a range. Ensure the page we
4497 * are about to unmap is the actual page of interest.
4498 */
4499 if (ref_page) {
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07004500 if (page != ref_page) {
4501 spin_unlock(ptl);
4502 continue;
4503 }
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004504 /*
4505 * Mark the VMA as having unmapped its page so that
4506 * future faults in this VMA will fail rather than
4507 * looking like data was lost
4508 */
4509 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4510 }
4511
David Gibsonc7546f82005-08-05 11:59:35 -07004512 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.Vb528e4b2016-12-12 16:42:37 -08004513 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07004514 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08004515 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07004516
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08004517 hugetlb_count_sub(pages_per_huge_page(h), mm);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08004518 page_remove_rmap(page, true);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07004519
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004520 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07004521 tlb_remove_page_size(tlb, page, huge_page_size(h));
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07004522 /*
4523 * Bail out after unmapping reference page if supplied
4524 */
4525 if (ref_page)
4526 break;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07004527 }
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004528 mmu_notifier_invalidate_range_end(&range);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004529 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004530}
David Gibson63551ae2005-06-21 17:14:44 -07004531
Mel Gormand8333522012-07-31 16:46:20 -07004532void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4533 struct vm_area_struct *vma, unsigned long start,
4534 unsigned long end, struct page *ref_page)
4535{
4536 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4537
4538 /*
4539 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4540 * test will fail on a vma being torn down, and not grab a page table
4541 * on its way out. We're lucky that the flag has such an appropriate
4542 * name, and can in fact be safely cleared here. We could clear it
4543 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004544 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07004545 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004546 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07004547 */
4548 vma->vm_flags &= ~VM_MAYSHARE;
4549}
4550
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004551void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004552 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004553{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004554 struct mmu_gather tlb;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07004555
Will Deacona72afd82021-01-27 23:53:45 +00004556 tlb_gather_mmu(&tlb, vma->vm_mm);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004557 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
Will Deaconae8eba82021-01-27 23:53:43 +00004558 tlb_finish_mmu(&tlb);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004559}
4560
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004561/*
4562 * This is called when the original mapper is failing to COW a MAP_PRIVATE
Zhiyuan Dai578b7722021-02-24 12:07:26 -08004563 * mapping it owns the reserve page for. The intention is to unmap the page
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004564 * from other VMAs and let the children be SIGKILLed if they are faulting the
4565 * same region.
4566 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004567static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4568 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004569{
Adam Litke75266742008-11-12 13:24:56 -08004570 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004571 struct vm_area_struct *iter_vma;
4572 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004573 pgoff_t pgoff;
4574
4575 /*
4576 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4577 * from page cache lookup which is in HPAGE_SIZE units.
4578 */
Adam Litke75266742008-11-12 13:24:56 -08004579 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07004580 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4581 vma->vm_pgoff;
Al Viro93c76a32015-12-04 23:45:44 -05004582 mapping = vma->vm_file->f_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004583
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08004584 /*
4585 * Take the mapping lock for the duration of the table walk. As
4586 * this mapping should be shared between all the VMAs,
4587 * __unmap_hugepage_range() is called as the lock is already held
4588 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004589 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07004590 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004591 /* Do not unmap the current VMA */
4592 if (iter_vma == vma)
4593 continue;
4594
4595 /*
Mel Gorman2f84a892015-10-01 15:36:57 -07004596 * Shared VMAs have their own reserves and do not affect
4597 * MAP_PRIVATE accounting but it is possible that a shared
4598 * VMA is using the same page so check and skip such VMAs.
4599 */
4600 if (iter_vma->vm_flags & VM_MAYSHARE)
4601 continue;
4602
4603 /*
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004604 * Unmap the page from other VMAs without their own reserves.
4605 * They get marked to be SIGKILLed if they fault in these
4606 * areas. This is because a future no-page fault on this VMA
4607 * could insert a zeroed page instead of the data existing
4608 * from the time of fork. This would look like data corruption
4609 */
4610 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004611 unmap_hugepage_range(iter_vma, address,
4612 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004613 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004614 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004615}
4616
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004617/*
4618 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08004619 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4620 * cannot race with other handlers or page migration.
4621 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004622 */
Souptick Joarder2b740302018-08-23 17:01:36 -07004623static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Huang Ying974e6d62018-08-17 15:45:57 -07004624 unsigned long address, pte_t *ptep,
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004625 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08004626{
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004627 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07004628 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08004629 struct page *old_page, *new_page;
Souptick Joarder2b740302018-08-23 17:01:36 -07004630 int outside_reserve = 0;
4631 vm_fault_t ret = 0;
Huang Ying974e6d62018-08-17 15:45:57 -07004632 unsigned long haddr = address & huge_page_mask(h);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004633 struct mmu_notifier_range range;
David Gibson1e8f8892006-01-06 00:10:44 -08004634
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004635 pte = huge_ptep_get(ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08004636 old_page = pte_page(pte);
4637
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004638retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08004639 /* If no-one else is actually using this page, avoid the copy
4640 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07004641 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
Hugh Dickins5a499732016-07-14 12:07:38 -07004642 page_move_anon_rmap(old_page, vma);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004643 set_huge_ptep_writable(vma, haddr, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07004644 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08004645 }
4646
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004647 /*
4648 * If the process that created a MAP_PRIVATE mapping is about to
4649 * perform a COW due to a shared page count, attempt to satisfy
4650 * the allocation without using the existing reserves. The pagecache
4651 * page is used to determine if the reserve at this address was
4652 * consumed or not. If reserves were used, a partial faulted mapping
4653 * at the time of fork() could consume its reserves on COW instead
4654 * of the full address range.
4655 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07004656 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004657 old_page != pagecache_page)
4658 outside_reserve = 1;
4659
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004660 get_page(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004661
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004662 /*
4663 * Drop page table lock as buddy allocator may be called. It will
4664 * be acquired again before returning to the caller, as expected.
4665 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004666 spin_unlock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004667 new_page = alloc_huge_page(vma, haddr, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08004668
Adam Litke2fc39ce2007-11-14 16:59:39 -08004669 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004670 /*
4671 * If a process owning a MAP_PRIVATE mapping fails to COW,
4672 * it is due to references held by a child and an insufficient
4673 * huge page pool. To guarantee the original mappers
4674 * reliability, unmap the page from child processes. The child
4675 * may get SIGKILLed if it later faults.
4676 */
4677 if (outside_reserve) {
Mike Kravetze7dd91c2020-12-29 15:14:25 -08004678 struct address_space *mapping = vma->vm_file->f_mapping;
4679 pgoff_t idx;
4680 u32 hash;
4681
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004682 put_page(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004683 BUG_ON(huge_pte_none(pte));
Mike Kravetze7dd91c2020-12-29 15:14:25 -08004684 /*
4685 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4686 * unmapping. unmapping needs to hold i_mmap_rwsem
4687 * in write mode. Dropping i_mmap_rwsem in read mode
4688 * here is OK as COW mappings do not interact with
4689 * PMD sharing.
4690 *
4691 * Reacquire both after unmap operation.
4692 */
4693 idx = vma_hugecache_offset(h, vma, haddr);
4694 hash = hugetlb_fault_mutex_hash(mapping, idx);
4695 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4696 i_mmap_unlock_read(mapping);
4697
Huang Ying5b7a1d42018-08-17 15:45:53 -07004698 unmap_ref_private(mm, vma, old_page, haddr);
Mike Kravetze7dd91c2020-12-29 15:14:25 -08004699
4700 i_mmap_lock_read(mapping);
4701 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004702 spin_lock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004703 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004704 if (likely(ptep &&
4705 pte_same(huge_ptep_get(ptep), pte)))
4706 goto retry_avoidcopy;
4707 /*
4708 * race occurs while re-acquiring page table
4709 * lock, and our job is done.
4710 */
4711 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004712 }
4713
Souptick Joarder2b740302018-08-23 17:01:36 -07004714 ret = vmf_error(PTR_ERR(new_page));
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004715 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08004716 }
4717
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004718 /*
4719 * When the original hugepage is shared one, it does not have
4720 * anon_vma prepared.
4721 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07004722 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004723 ret = VM_FAULT_OOM;
4724 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07004725 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004726
Huang Ying974e6d62018-08-17 15:45:57 -07004727 copy_user_huge_page(new_page, old_page, address, vma,
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004728 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08004729 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08004730
Jérôme Glisse7269f992019-05-13 17:20:53 -07004731 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07004732 haddr + huge_page_size(h));
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004733 mmu_notifier_invalidate_range_start(&range);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004734
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004735 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004736 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004737 * before the page tables are altered
4738 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004739 spin_lock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004740 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07004741 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Mike Kravetzd6995da2021-02-24 12:08:51 -08004742 ClearHPageRestoreReserve(new_page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07004743
David Gibson1e8f8892006-01-06 00:10:44 -08004744 /* Break COW */
Huang Ying5b7a1d42018-08-17 15:45:53 -07004745 huge_ptep_clear_flush(vma, haddr, ptep);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004746 mmu_notifier_invalidate_range(mm, range.start, range.end);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004747 set_huge_pte_at(mm, haddr, ptep,
David Gibson1e8f8892006-01-06 00:10:44 -08004748 make_huge_pte(vma, new_page, 1));
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08004749 page_remove_rmap(old_page, true);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004750 hugepage_add_new_anon_rmap(new_page, vma, haddr);
Mike Kravetz8f251a32021-02-24 12:08:56 -08004751 SetHPageMigratable(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08004752 /* Make the old page be freed below */
4753 new_page = old_page;
4754 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004755 spin_unlock(ptl);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004756 mmu_notifier_invalidate_range_end(&range);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004757out_release_all:
Mike Kravetzc7b18502021-08-19 19:04:33 -07004758 /* No restore in case of successful pagetable update (Break COW) */
4759 if (new_page != old_page)
4760 restore_reserve_on_error(h, vma, haddr, new_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004761 put_page(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004762out_release_old:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004763 put_page(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07004764
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004765 spin_lock(ptl); /* Caller expects lock to be held */
4766 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08004767}
4768
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004769/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07004770static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4771 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004772{
4773 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07004774 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004775
4776 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07004777 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004778
4779 return find_lock_page(mapping, idx);
4780}
4781
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004782/*
4783 * Return whether there is a pagecache page to back given address within VMA.
4784 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4785 */
4786static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004787 struct vm_area_struct *vma, unsigned long address)
4788{
4789 struct address_space *mapping;
4790 pgoff_t idx;
4791 struct page *page;
4792
4793 mapping = vma->vm_file->f_mapping;
4794 idx = vma_hugecache_offset(h, vma, address);
4795
4796 page = find_get_page(mapping, idx);
4797 if (page)
4798 put_page(page);
4799 return page != NULL;
4800}
4801
Mike Kravetzab76ad52015-09-08 15:01:50 -07004802int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4803 pgoff_t idx)
4804{
4805 struct inode *inode = mapping->host;
4806 struct hstate *h = hstate_inode(inode);
4807 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4808
4809 if (err)
4810 return err;
Mike Kravetzd6995da2021-02-24 12:08:51 -08004811 ClearHPageRestoreReserve(page);
Mike Kravetzab76ad52015-09-08 15:01:50 -07004812
Mike Kravetz22146c32018-10-26 15:10:58 -07004813 /*
4814 * set page dirty so that it will not be removed from cache/file
4815 * by non-hugetlbfs specific code paths.
4816 */
4817 set_page_dirty(page);
4818
Mike Kravetzab76ad52015-09-08 15:01:50 -07004819 spin_lock(&inode->i_lock);
4820 inode->i_blocks += blocks_per_huge_page(h);
4821 spin_unlock(&inode->i_lock);
4822 return 0;
4823}
4824
Axel Rasmussen7677f7f2021-05-04 18:35:36 -07004825static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
4826 struct address_space *mapping,
4827 pgoff_t idx,
4828 unsigned int flags,
4829 unsigned long haddr,
4830 unsigned long reason)
4831{
4832 vm_fault_t ret;
4833 u32 hash;
4834 struct vm_fault vmf = {
4835 .vma = vma,
4836 .address = haddr,
4837 .flags = flags,
4838
4839 /*
4840 * Hard to debug if it ends up being
4841 * used by a callee that assumes
4842 * something about the other
4843 * uninitialized fields... same as in
4844 * memory.c
4845 */
4846 };
4847
4848 /*
4849 * hugetlb_fault_mutex and i_mmap_rwsem must be
4850 * dropped before handling userfault. Reacquire
4851 * after handling fault to make calling code simpler.
4852 */
4853 hash = hugetlb_fault_mutex_hash(mapping, idx);
4854 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4855 i_mmap_unlock_read(mapping);
4856 ret = handle_userfault(&vmf, reason);
4857 i_mmap_lock_read(mapping);
4858 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4859
4860 return ret;
4861}
4862
Souptick Joarder2b740302018-08-23 17:01:36 -07004863static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4864 struct vm_area_struct *vma,
4865 struct address_space *mapping, pgoff_t idx,
4866 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004867{
Andi Kleena5516432008-07-23 21:27:41 -07004868 struct hstate *h = hstate_vma(vma);
Souptick Joarder2b740302018-08-23 17:01:36 -07004869 vm_fault_t ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08004870 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07004871 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07004872 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08004873 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004874 spinlock_t *ptl;
Huang Ying285b8dc2018-06-07 17:08:08 -07004875 unsigned long haddr = address & huge_page_mask(h);
Mike Kravetzc7b18502021-08-19 19:04:33 -07004876 bool new_page, new_pagecache_page = false;
Adam Litke4c887262005-10-29 18:16:46 -07004877
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004878 /*
4879 * Currently, we are forced to kill the process in the event the
4880 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03004881 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004882 */
4883 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Geoffrey Thomas910154d2016-03-09 14:08:04 -08004884 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08004885 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004886 return ret;
4887 }
4888
Adam Litke4c887262005-10-29 18:16:46 -07004889 /*
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004890 * We can not race with truncation due to holding i_mmap_rwsem.
4891 * i_size is modified when holding i_mmap_rwsem, so check here
4892 * once for faults beyond end of file.
Adam Litke4c887262005-10-29 18:16:46 -07004893 */
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004894 size = i_size_read(mapping->host) >> huge_page_shift(h);
4895 if (idx >= size)
4896 goto out;
4897
Christoph Lameter6bda6662006-01-06 00:10:49 -08004898retry:
Mike Kravetzc7b18502021-08-19 19:04:33 -07004899 new_page = false;
Christoph Lameter6bda6662006-01-06 00:10:49 -08004900 page = find_lock_page(mapping, idx);
4901 if (!page) {
Axel Rasmussen7677f7f2021-05-04 18:35:36 -07004902 /* Check for page in userfault range */
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004903 if (userfaultfd_missing(vma)) {
Axel Rasmussen7677f7f2021-05-04 18:35:36 -07004904 ret = hugetlb_handle_userfault(vma, mapping, idx,
4905 flags, haddr,
4906 VM_UFFD_MISSING);
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004907 goto out;
4908 }
4909
Huang Ying285b8dc2018-06-07 17:08:08 -07004910 page = alloc_huge_page(vma, haddr, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08004911 if (IS_ERR(page)) {
Mike Kravetz4643d672019-08-13 15:38:00 -07004912 /*
4913 * Returning error will result in faulting task being
4914 * sent SIGBUS. The hugetlb fault mutex prevents two
4915 * tasks from racing to fault in the same page which
4916 * could result in false unable to allocate errors.
4917 * Page migration does not take the fault mutex, but
4918 * does a clear then write of pte's under page table
4919 * lock. Page fault code could race with migration,
4920 * notice the clear pte and try to allocate a page
4921 * here. Before returning error, get ptl and make
4922 * sure there really is no pte entry.
4923 */
4924 ptl = huge_pte_lock(h, mm, ptep);
Miaohe Lind83e6c8a2021-05-04 18:33:31 -07004925 ret = 0;
4926 if (huge_pte_none(huge_ptep_get(ptep)))
4927 ret = vmf_error(PTR_ERR(page));
Mike Kravetz4643d672019-08-13 15:38:00 -07004928 spin_unlock(ptl);
Christoph Lameter6bda6662006-01-06 00:10:49 -08004929 goto out;
4930 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004931 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08004932 __SetPageUptodate(page);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004933 new_page = true;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004934
Mel Gormanf83a2752009-05-28 14:34:40 -07004935 if (vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzab76ad52015-09-08 15:01:50 -07004936 int err = huge_add_to_page_cache(page, mapping, idx);
Christoph Lameter6bda6662006-01-06 00:10:49 -08004937 if (err) {
4938 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08004939 if (err == -EEXIST)
4940 goto retry;
4941 goto out;
4942 }
Mike Kravetzc7b18502021-08-19 19:04:33 -07004943 new_pagecache_page = true;
Mel Gorman23be7462010-04-23 13:17:56 -04004944 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08004945 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004946 if (unlikely(anon_vma_prepare(vma))) {
4947 ret = VM_FAULT_OOM;
4948 goto backout_unlocked;
4949 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08004950 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04004951 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004952 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09004953 /*
4954 * If memory error occurs between mmap() and fault, some process
4955 * don't have hwpoisoned swap entry for errored virtual address.
4956 * So we need to block hugepage fault by PG_hwpoison bit check.
4957 */
4958 if (unlikely(PageHWPoison(page))) {
Miaohe Lin0eb98f12021-01-12 15:49:24 -08004959 ret = VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07004960 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09004961 goto backout_unlocked;
4962 }
Axel Rasmussen7677f7f2021-05-04 18:35:36 -07004963
4964 /* Check for page in userfault range. */
4965 if (userfaultfd_minor(vma)) {
4966 unlock_page(page);
4967 put_page(page);
4968 ret = hugetlb_handle_userfault(vma, mapping, idx,
4969 flags, haddr,
4970 VM_UFFD_MINOR);
4971 goto out;
4972 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08004973 }
David Gibson1e8f8892006-01-06 00:10:44 -08004974
Andy Whitcroft57303d82008-08-12 15:08:47 -07004975 /*
4976 * If we are going to COW a private mapping later, we examine the
4977 * pending reservations for this page now. This will ensure that
4978 * any allocations necessary to record that reservation occur outside
4979 * the spinlock.
4980 */
Mike Kravetz5e911372015-09-08 15:01:28 -07004981 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
Huang Ying285b8dc2018-06-07 17:08:08 -07004982 if (vma_needs_reservation(h, vma, haddr) < 0) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07004983 ret = VM_FAULT_OOM;
4984 goto backout_unlocked;
4985 }
Mike Kravetz5e911372015-09-08 15:01:28 -07004986 /* Just decrements count, does not deallocate */
Huang Ying285b8dc2018-06-07 17:08:08 -07004987 vma_end_reservation(h, vma, haddr);
Mike Kravetz5e911372015-09-08 15:01:28 -07004988 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07004989
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08004990 ptl = huge_pte_lock(h, mm, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07004991 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004992 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07004993 goto backout;
4994
Joonsoo Kim07443a82013-09-11 14:21:58 -07004995 if (anon_rmap) {
Mike Kravetzd6995da2021-02-24 12:08:51 -08004996 ClearHPageRestoreReserve(page);
Huang Ying285b8dc2018-06-07 17:08:08 -07004997 hugepage_add_new_anon_rmap(page, vma, haddr);
Choi Gi-yongac714902014-04-07 15:37:36 -07004998 } else
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08004999 page_dup_rmap(page, true);
David Gibson1e8f8892006-01-06 00:10:44 -08005000 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
5001 && (vma->vm_flags & VM_SHARED)));
Huang Ying285b8dc2018-06-07 17:08:08 -07005002 set_huge_pte_at(mm, haddr, ptep, new_pte);
David Gibson1e8f8892006-01-06 00:10:44 -08005003
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08005004 hugetlb_count_add(pages_per_huge_page(h), mm);
Hugh Dickins788c7df2009-06-23 13:49:05 +01005005 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08005006 /* Optimization, do the COW without a second fault */
Huang Ying974e6d62018-08-17 15:45:57 -07005007 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08005008 }
5009
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005010 spin_unlock(ptl);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08005011
5012 /*
Mike Kravetz8f251a32021-02-24 12:08:56 -08005013 * Only set HPageMigratable in newly allocated pages. Existing pages
5014 * found in the pagecache may not have HPageMigratableset if they have
5015 * been isolated for migration.
Mike Kravetzcb6acd02019-02-28 16:22:02 -08005016 */
5017 if (new_page)
Mike Kravetz8f251a32021-02-24 12:08:56 -08005018 SetHPageMigratable(page);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08005019
Adam Litke4c887262005-10-29 18:16:46 -07005020 unlock_page(page);
5021out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01005022 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07005023
5024backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005025 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07005026backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07005027 unlock_page(page);
Mike Kravetzc7b18502021-08-19 19:04:33 -07005028 /* restore reserve for newly allocated pages not in page cache */
5029 if (new_page && !new_pagecache_page)
5030 restore_reserve_on_error(h, vma, haddr, page);
Adam Litke4c887262005-10-29 18:16:46 -07005031 put_page(page);
5032 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01005033}
5034
Davidlohr Bueso8382d912014-04-03 14:47:31 -07005035#ifdef CONFIG_SMP
Wei Yang188b04a2019-11-30 17:57:02 -08005036u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
Davidlohr Bueso8382d912014-04-03 14:47:31 -07005037{
5038 unsigned long key[2];
5039 u32 hash;
5040
Mike Kravetz1b426ba2019-05-13 17:19:41 -07005041 key[0] = (unsigned long) mapping;
5042 key[1] = idx;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07005043
Mike Kravetz55254632019-11-30 17:56:30 -08005044 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07005045
5046 return hash & (num_fault_mutexes - 1);
5047}
5048#else
5049/*
Miaohe Lin6c26d312021-02-24 12:07:19 -08005050 * For uniprocessor systems we always use a single mutex, so just
Davidlohr Bueso8382d912014-04-03 14:47:31 -07005051 * return 0 and avoid the hashing overhead.
5052 */
Wei Yang188b04a2019-11-30 17:57:02 -08005053u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
Davidlohr Bueso8382d912014-04-03 14:47:31 -07005054{
5055 return 0;
5056}
5057#endif
5058
Souptick Joarder2b740302018-08-23 17:01:36 -07005059vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01005060 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08005061{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07005062 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005063 spinlock_t *ptl;
Souptick Joarder2b740302018-08-23 17:01:36 -07005064 vm_fault_t ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07005065 u32 hash;
5066 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09005067 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07005068 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07005069 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07005070 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08005071 int need_wait_lock = 0;
Huang Ying285b8dc2018-06-07 17:08:08 -07005072 unsigned long haddr = address & huge_page_mask(h);
Adam Litke86e52162006-01-06 00:10:43 -08005073
Huang Ying285b8dc2018-06-07 17:08:08 -07005074 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09005075 if (ptep) {
Mike Kravetzc0d03812020-04-01 21:11:05 -07005076 /*
5077 * Since we hold no locks, ptep could be stale. That is
5078 * OK as we are only making decisions based on content and
5079 * not actually modifying content here.
5080 */
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09005081 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09005082 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005083 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09005084 return 0;
5085 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07005086 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07005087 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09005088 }
5089
Mike Kravetzc0d03812020-04-01 21:11:05 -07005090 /*
5091 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
Mike Kravetz87bf91d2020-04-01 21:11:08 -07005092 * until finished with ptep. This serves two purposes:
5093 * 1) It prevents huge_pmd_unshare from being called elsewhere
5094 * and making the ptep no longer valid.
5095 * 2) It synchronizes us with i_size modifications during truncation.
Mike Kravetzc0d03812020-04-01 21:11:05 -07005096 *
5097 * ptep could have already be assigned via huge_pte_offset. That
5098 * is OK, as huge_pte_alloc will return the same value unless
5099 * something has changed.
5100 */
Davidlohr Bueso8382d912014-04-03 14:47:31 -07005101 mapping = vma->vm_file->f_mapping;
Mike Kravetzc0d03812020-04-01 21:11:05 -07005102 i_mmap_lock_read(mapping);
Peter Xuaec44e02021-05-04 18:33:00 -07005103 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
Mike Kravetzc0d03812020-04-01 21:11:05 -07005104 if (!ptep) {
5105 i_mmap_unlock_read(mapping);
5106 return VM_FAULT_OOM;
5107 }
Davidlohr Bueso8382d912014-04-03 14:47:31 -07005108
David Gibson3935baa2006-03-22 00:08:53 -08005109 /*
5110 * Serialize hugepage allocation and instantiation, so that we don't
5111 * get spurious allocation failures if two CPUs race to instantiate
5112 * the same page in the page cache.
5113 */
Mike Kravetzc0d03812020-04-01 21:11:05 -07005114 idx = vma_hugecache_offset(h, vma, haddr);
Wei Yang188b04a2019-11-30 17:57:02 -08005115 hash = hugetlb_fault_mutex_hash(mapping, idx);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07005116 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07005117
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07005118 entry = huge_ptep_get(ptep);
5119 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07005120 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07005121 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08005122 }
Adam Litke86e52162006-01-06 00:10:43 -08005123
Nick Piggin83c54072007-07-19 01:47:05 -07005124 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08005125
Andy Whitcroft57303d82008-08-12 15:08:47 -07005126 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08005127 * entry could be a migration/hwpoison entry at this point, so this
5128 * check prevents the kernel from going below assuming that we have
Ethon Paul7c8de352020-06-04 16:49:07 -07005129 * an active hugepage in pagecache. This goto expects the 2nd page
5130 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
5131 * properly handle it.
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08005132 */
5133 if (!pte_present(entry))
5134 goto out_mutex;
5135
5136 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07005137 * If we are going to COW the mapping later, we examine the pending
5138 * reservations for this page now. This will ensure that any
5139 * allocations necessary to record that reservation occur outside the
5140 * spinlock. For private mappings, we also lookup the pagecache
5141 * page now as it is used to determine if a reservation has been
5142 * consumed.
5143 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07005144 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Huang Ying285b8dc2018-06-07 17:08:08 -07005145 if (vma_needs_reservation(h, vma, haddr) < 0) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07005146 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07005147 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07005148 }
Mike Kravetz5e911372015-09-08 15:01:28 -07005149 /* Just decrements count, does not deallocate */
Huang Ying285b8dc2018-06-07 17:08:08 -07005150 vma_end_reservation(h, vma, haddr);
Andy Whitcroft57303d82008-08-12 15:08:47 -07005151
Mel Gormanf83a2752009-05-28 14:34:40 -07005152 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07005153 pagecache_page = hugetlbfs_pagecache_page(h,
Huang Ying285b8dc2018-06-07 17:08:08 -07005154 vma, haddr);
Andy Whitcroft57303d82008-08-12 15:08:47 -07005155 }
5156
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08005157 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09005158
David Gibson1e8f8892006-01-06 00:10:44 -08005159 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07005160 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005161 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07005162
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08005163 /*
5164 * hugetlb_cow() requires page locks of pte_page(entry) and
5165 * pagecache_page, so here we need take the former one
5166 * when page != pagecache_page or !pagecache_page.
5167 */
5168 page = pte_page(entry);
5169 if (page != pagecache_page)
5170 if (!trylock_page(page)) {
5171 need_wait_lock = 1;
5172 goto out_ptl;
5173 }
5174
5175 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07005176
Hugh Dickins788c7df2009-06-23 13:49:05 +01005177 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07005178 if (!huge_pte_write(entry)) {
Huang Ying974e6d62018-08-17 15:45:57 -07005179 ret = hugetlb_cow(mm, vma, address, ptep,
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08005180 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08005181 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07005182 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07005183 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07005184 }
5185 entry = pte_mkyoung(entry);
Huang Ying285b8dc2018-06-07 17:08:08 -07005186 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
Hugh Dickins788c7df2009-06-23 13:49:05 +01005187 flags & FAULT_FLAG_WRITE))
Huang Ying285b8dc2018-06-07 17:08:08 -07005188 update_mmu_cache(vma, haddr, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08005189out_put_page:
5190 if (page != pagecache_page)
5191 unlock_page(page);
5192 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005193out_ptl:
5194 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07005195
5196 if (pagecache_page) {
5197 unlock_page(pagecache_page);
5198 put_page(pagecache_page);
5199 }
David Gibsonb4d1d992008-10-15 22:01:11 -07005200out_mutex:
Mike Kravetzc672c7f2015-09-08 15:01:35 -07005201 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Mike Kravetzc0d03812020-04-01 21:11:05 -07005202 i_mmap_unlock_read(mapping);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08005203 /*
5204 * Generally it's safe to hold refcount during waiting page lock. But
5205 * here we just wait to defer the next page fault to avoid busy loop and
5206 * the page is not used after unlocked before returning from the current
5207 * page fault. So we are safe from accessing freed page, even if we wait
5208 * here without taking refcount.
5209 */
5210 if (need_wait_lock)
5211 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08005212 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08005213}
5214
Axel Rasmussen714c1892021-05-04 18:35:45 -07005215#ifdef CONFIG_USERFAULTFD
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005216/*
5217 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
5218 * modifications for huge pages.
5219 */
5220int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
5221 pte_t *dst_pte,
5222 struct vm_area_struct *dst_vma,
5223 unsigned long dst_addr,
5224 unsigned long src_addr,
Axel Rasmussenf6191472021-05-04 18:35:49 -07005225 enum mcopy_atomic_mode mode,
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005226 struct page **pagep)
5227{
Axel Rasmussenf6191472021-05-04 18:35:49 -07005228 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
Mina Almasry8cc5fcb2021-06-30 18:48:19 -07005229 struct hstate *h = hstate_vma(dst_vma);
5230 struct address_space *mapping = dst_vma->vm_file->f_mapping;
5231 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07005232 unsigned long size;
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08005233 int vm_shared = dst_vma->vm_flags & VM_SHARED;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005234 pte_t _dst_pte;
5235 spinlock_t *ptl;
Mina Almasry8cc5fcb2021-06-30 18:48:19 -07005236 int ret = -ENOMEM;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005237 struct page *page;
Axel Rasmussenf6191472021-05-04 18:35:49 -07005238 int writable;
Mike Kravetzc7b18502021-08-19 19:04:33 -07005239 bool new_pagecache_page = false;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005240
Axel Rasmussenf6191472021-05-04 18:35:49 -07005241 if (is_continue) {
5242 ret = -EFAULT;
5243 page = find_lock_page(mapping, idx);
5244 if (!page)
5245 goto out;
5246 } else if (!*pagep) {
Mina Almasryd84cf062021-06-04 20:01:36 -07005247 /* If a page already exists, then it's UFFDIO_COPY for
5248 * a non-missing case. Return -EEXIST.
5249 */
5250 if (vm_shared &&
5251 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5252 ret = -EEXIST;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005253 goto out;
Mina Almasryd84cf062021-06-04 20:01:36 -07005254 }
5255
5256 page = alloc_huge_page(dst_vma, dst_addr, 0);
5257 if (IS_ERR(page)) {
5258 ret = -ENOMEM;
5259 goto out;
5260 }
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005261
5262 ret = copy_huge_page_from_user(page,
5263 (const void __user *) src_addr,
Mike Kravetz810a56b2017-02-22 15:42:58 -08005264 pages_per_huge_page(h), false);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005265
Michel Lespinassec1e8d7c2020-06-08 21:33:54 -07005266 /* fallback to copy_from_user outside mmap_lock */
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005267 if (unlikely(ret)) {
Andrea Arcangeli9e368252018-11-30 14:09:25 -08005268 ret = -ENOENT;
Mina Almasry8cc5fcb2021-06-30 18:48:19 -07005269 /* Free the allocated page which may have
5270 * consumed a reservation.
5271 */
5272 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5273 put_page(page);
5274
5275 /* Allocate a temporary page to hold the copied
5276 * contents.
5277 */
5278 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
5279 if (!page) {
5280 ret = -ENOMEM;
5281 goto out;
5282 }
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005283 *pagep = page;
Mina Almasry8cc5fcb2021-06-30 18:48:19 -07005284 /* Set the outparam pagep and return to the caller to
5285 * copy the contents outside the lock. Don't free the
5286 * page.
5287 */
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005288 goto out;
5289 }
5290 } else {
Mina Almasry8cc5fcb2021-06-30 18:48:19 -07005291 if (vm_shared &&
5292 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5293 put_page(*pagep);
5294 ret = -EEXIST;
5295 *pagep = NULL;
5296 goto out;
5297 }
5298
5299 page = alloc_huge_page(dst_vma, dst_addr, 0);
5300 if (IS_ERR(page)) {
5301 ret = -ENOMEM;
5302 *pagep = NULL;
5303 goto out;
5304 }
5305 copy_huge_page(page, *pagep);
5306 put_page(*pagep);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005307 *pagep = NULL;
5308 }
5309
5310 /*
5311 * The memory barrier inside __SetPageUptodate makes sure that
5312 * preceding stores to the page contents become visible before
5313 * the set_pte_at() write.
5314 */
5315 __SetPageUptodate(page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005316
Axel Rasmussenf6191472021-05-04 18:35:49 -07005317 /* Add shared, newly allocated pages to the page cache. */
5318 if (vm_shared && !is_continue) {
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07005319 size = i_size_read(mapping->host) >> huge_page_shift(h);
5320 ret = -EFAULT;
5321 if (idx >= size)
5322 goto out_release_nounlock;
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08005323
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07005324 /*
5325 * Serialization between remove_inode_hugepages() and
5326 * huge_add_to_page_cache() below happens through the
5327 * hugetlb_fault_mutex_table that here must be hold by
5328 * the caller.
5329 */
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08005330 ret = huge_add_to_page_cache(page, mapping, idx);
5331 if (ret)
5332 goto out_release_nounlock;
Mike Kravetzc7b18502021-08-19 19:04:33 -07005333 new_pagecache_page = true;
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08005334 }
5335
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005336 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
5337 spin_lock(ptl);
5338
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07005339 /*
5340 * Recheck the i_size after holding PT lock to make sure not
5341 * to leave any page mapped (as page_mapped()) beyond the end
5342 * of the i_size (remove_inode_hugepages() is strict about
5343 * enforcing that). If we bail out here, we'll also leave a
5344 * page in the radix tree in the vm_shared case beyond the end
5345 * of the i_size, but remove_inode_hugepages() will take care
5346 * of it as soon as we drop the hugetlb_fault_mutex_table.
5347 */
5348 size = i_size_read(mapping->host) >> huge_page_shift(h);
5349 ret = -EFAULT;
5350 if (idx >= size)
5351 goto out_release_unlock;
5352
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005353 ret = -EEXIST;
5354 if (!huge_pte_none(huge_ptep_get(dst_pte)))
5355 goto out_release_unlock;
5356
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08005357 if (vm_shared) {
5358 page_dup_rmap(page, true);
5359 } else {
Mike Kravetzd6995da2021-02-24 12:08:51 -08005360 ClearHPageRestoreReserve(page);
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08005361 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
5362 }
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005363
Axel Rasmussenf6191472021-05-04 18:35:49 -07005364 /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */
5365 if (is_continue && !vm_shared)
5366 writable = 0;
5367 else
5368 writable = dst_vma->vm_flags & VM_WRITE;
5369
5370 _dst_pte = make_huge_pte(dst_vma, page, writable);
5371 if (writable)
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005372 _dst_pte = huge_pte_mkdirty(_dst_pte);
5373 _dst_pte = pte_mkyoung(_dst_pte);
5374
5375 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
5376
5377 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
5378 dst_vma->vm_flags & VM_WRITE);
5379 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
5380
5381 /* No need to invalidate - it was non-present before */
5382 update_mmu_cache(dst_vma, dst_addr, dst_pte);
5383
5384 spin_unlock(ptl);
Axel Rasmussenf6191472021-05-04 18:35:49 -07005385 if (!is_continue)
5386 SetHPageMigratable(page);
5387 if (vm_shared || is_continue)
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08005388 unlock_page(page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005389 ret = 0;
5390out:
5391 return ret;
5392out_release_unlock:
5393 spin_unlock(ptl);
Axel Rasmussenf6191472021-05-04 18:35:49 -07005394 if (vm_shared || is_continue)
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08005395 unlock_page(page);
Andrea Arcangeli5af10df2017-08-10 15:23:38 -07005396out_release_nounlock:
Mike Kravetzc7b18502021-08-19 19:04:33 -07005397 if (!new_pagecache_page)
5398 restore_reserve_on_error(h, dst_vma, dst_addr, page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005399 put_page(page);
5400 goto out;
5401}
Axel Rasmussen714c1892021-05-04 18:35:45 -07005402#endif /* CONFIG_USERFAULTFD */
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08005403
Joao Martins82e5d372021-02-24 12:07:16 -08005404static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
5405 int refs, struct page **pages,
5406 struct vm_area_struct **vmas)
5407{
5408 int nr;
5409
5410 for (nr = 0; nr < refs; nr++) {
5411 if (likely(pages))
5412 pages[nr] = mem_map_offset(page, nr);
5413 if (vmas)
5414 vmas[nr] = vma;
5415 }
5416}
5417
Michel Lespinasse28a35712013-02-22 16:35:55 -08005418long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
5419 struct page **pages, struct vm_area_struct **vmas,
5420 unsigned long *position, unsigned long *nr_pages,
Peter Xu4f6da932020-04-01 21:07:58 -07005421 long i, unsigned int flags, int *locked)
David Gibson63551ae2005-06-21 17:14:44 -07005422{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08005423 unsigned long pfn_offset;
5424 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08005425 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07005426 struct hstate *h = hstate_vma(vma);
Joao Martins0fa5bc42021-02-24 12:07:12 -08005427 int err = -EFAULT, refs;
David Gibson63551ae2005-06-21 17:14:44 -07005428
David Gibson63551ae2005-06-21 17:14:44 -07005429 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07005430 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005431 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07005432 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07005433 struct page *page;
5434
5435 /*
David Rientjes02057962015-04-14 15:48:24 -07005436 * If we have a pending SIGKILL, don't keep faulting pages and
5437 * potentially allocating memory.
5438 */
Davidlohr Buesofa45f112019-01-03 15:28:55 -08005439 if (fatal_signal_pending(current)) {
David Rientjes02057962015-04-14 15:48:24 -07005440 remainder = 0;
5441 break;
5442 }
5443
5444 /*
Adam Litke4c887262005-10-29 18:16:46 -07005445 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07005446 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07005447 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005448 *
5449 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07005450 */
Punit Agrawal7868a202017-07-06 15:39:42 -07005451 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
5452 huge_page_size(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005453 if (pte)
5454 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07005455 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07005456
Hugh Dickins2a15efc2009-09-21 17:03:27 -07005457 /*
5458 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07005459 * an error where there's an empty slot with no huge pagecache
5460 * to back it. This way, we avoid allocating a hugepage, and
5461 * the sparse dumpfile avoids allocating disk blocks, but its
5462 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07005463 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07005464 if (absent && (flags & FOLL_DUMP) &&
5465 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005466 if (pte)
5467 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07005468 remainder = 0;
5469 break;
5470 }
5471
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07005472 /*
5473 * We need call hugetlb_fault for both hugepages under migration
5474 * (in which case hugetlb_fault waits for the migration,) and
5475 * hwpoisoned hugepages (in which case we need to prevent the
5476 * caller from accessing to them.) In order to do this, we use
5477 * here is_swap_pte instead of is_hugetlb_entry_migration and
5478 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
5479 * both cases, and because we can't follow correct pages
5480 * directly from any kind of swap entries.
5481 */
5482 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07005483 ((flags & FOLL_WRITE) &&
5484 !huge_pte_write(huge_ptep_get(pte)))) {
Souptick Joarder2b740302018-08-23 17:01:36 -07005485 vm_fault_t ret;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08005486 unsigned int fault_flags = 0;
Adam Litke4c887262005-10-29 18:16:46 -07005487
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005488 if (pte)
5489 spin_unlock(ptl);
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08005490 if (flags & FOLL_WRITE)
5491 fault_flags |= FAULT_FLAG_WRITE;
Peter Xu4f6da932020-04-01 21:07:58 -07005492 if (locked)
Peter Xu71335f32020-04-01 21:08:53 -07005493 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5494 FAULT_FLAG_KILLABLE;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08005495 if (flags & FOLL_NOWAIT)
5496 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5497 FAULT_FLAG_RETRY_NOWAIT;
5498 if (flags & FOLL_TRIED) {
Peter Xu4426e942020-04-01 21:08:49 -07005499 /*
5500 * Note: FAULT_FLAG_ALLOW_RETRY and
5501 * FAULT_FLAG_TRIED can co-exist
5502 */
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08005503 fault_flags |= FAULT_FLAG_TRIED;
5504 }
5505 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
5506 if (ret & VM_FAULT_ERROR) {
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07005507 err = vm_fault_to_errno(ret, flags);
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08005508 remainder = 0;
5509 break;
5510 }
5511 if (ret & VM_FAULT_RETRY) {
Peter Xu4f6da932020-04-01 21:07:58 -07005512 if (locked &&
Andrea Arcangeli1ac25012019-02-01 14:20:16 -08005513 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
Peter Xu4f6da932020-04-01 21:07:58 -07005514 *locked = 0;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08005515 *nr_pages = 0;
5516 /*
5517 * VM_FAULT_RETRY must not return an
5518 * error, it will return zero
5519 * instead.
5520 *
5521 * No need to update "position" as the
5522 * caller will not check it after
5523 * *nr_pages is set to 0.
5524 */
5525 return i;
5526 }
5527 continue;
Adam Litke4c887262005-10-29 18:16:46 -07005528 }
David Gibson63551ae2005-06-21 17:14:44 -07005529
Andi Kleena5516432008-07-23 21:27:41 -07005530 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07005531 page = pte_page(huge_ptep_get(pte));
Linus Torvalds8fde12c2019-04-11 10:49:19 -07005532
5533 /*
Zhigang Luacbfb082019-11-30 17:57:06 -08005534 * If subpage information not requested, update counters
5535 * and skip the same_page loop below.
5536 */
5537 if (!pages && !vmas && !pfn_offset &&
5538 (vaddr + huge_page_size(h) < vma->vm_end) &&
5539 (remainder >= pages_per_huge_page(h))) {
5540 vaddr += huge_page_size(h);
5541 remainder -= pages_per_huge_page(h);
5542 i += pages_per_huge_page(h);
5543 spin_unlock(ptl);
5544 continue;
5545 }
5546
Joao Martinsd08af0a2021-07-14 21:27:11 -07005547 /* vaddr may not be aligned to PAGE_SIZE */
5548 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
5549 (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
Joao Martins0fa5bc42021-02-24 12:07:12 -08005550
Joao Martins82e5d372021-02-24 12:07:16 -08005551 if (pages || vmas)
5552 record_subpages_vmas(mem_map_offset(page, pfn_offset),
5553 vma, refs,
5554 likely(pages) ? pages + i : NULL,
5555 vmas ? vmas + i : NULL);
David Gibson63551ae2005-06-21 17:14:44 -07005556
Joao Martins82e5d372021-02-24 12:07:16 -08005557 if (pages) {
Joao Martins0fa5bc42021-02-24 12:07:12 -08005558 /*
5559 * try_grab_compound_head() should always succeed here,
5560 * because: a) we hold the ptl lock, and b) we've just
5561 * checked that the huge page is present in the page
5562 * tables. If the huge page is present, then the tail
5563 * pages must also be present. The ptl prevents the
5564 * head page and tail pages from being rearranged in
5565 * any way. So this page must be available at this
5566 * point, unless the page refcount overflowed:
5567 */
Joao Martins82e5d372021-02-24 12:07:16 -08005568 if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
Joao Martins0fa5bc42021-02-24 12:07:12 -08005569 refs,
5570 flags))) {
5571 spin_unlock(ptl);
5572 remainder = 0;
5573 err = -ENOMEM;
5574 break;
5575 }
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08005576 }
Joao Martins82e5d372021-02-24 12:07:16 -08005577
5578 vaddr += (refs << PAGE_SHIFT);
5579 remainder -= refs;
5580 i += refs;
5581
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005582 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07005583 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08005584 *nr_pages = remainder;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08005585 /*
5586 * setting position is actually required only if remainder is
5587 * not zero but it's faster not to add a "if (remainder)"
5588 * branch.
5589 */
David Gibson63551ae2005-06-21 17:14:44 -07005590 *position = vaddr;
5591
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07005592 return i ? i : err;
David Gibson63551ae2005-06-21 17:14:44 -07005593}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005594
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005595unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005596 unsigned long address, unsigned long end, pgprot_t newprot)
5597{
5598 struct mm_struct *mm = vma->vm_mm;
5599 unsigned long start = address;
5600 pte_t *ptep;
5601 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07005602 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005603 unsigned long pages = 0;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005604 bool shared_pmd = false;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005605 struct mmu_notifier_range range;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005606
5607 /*
5608 * In the case of shared PMDs, the area to flush could be beyond
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005609 * start/end. Set range.start/range.end to cover the maximum possible
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005610 * range if PMD sharing is possible.
5611 */
Jérôme Glisse7269f992019-05-13 17:20:53 -07005612 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5613 0, vma, mm, start, end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005614 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005615
5616 BUG_ON(address >= end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005617 flush_cache_range(vma, range.start, range.end);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005618
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005619 mmu_notifier_invalidate_range_start(&range);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08005620 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07005621 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005622 spinlock_t *ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07005623 ptep = huge_pte_offset(mm, address, huge_page_size(h));
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005624 if (!ptep)
5625 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005626 ptl = huge_pte_lock(h, mm, ptep);
Mike Kravetz34ae2042020-08-11 18:31:38 -07005627 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005628 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005629 spin_unlock(ptl);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005630 shared_pmd = true;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08005631 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005632 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08005633 pte = huge_ptep_get(ptep);
5634 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5635 spin_unlock(ptl);
5636 continue;
5637 }
5638 if (unlikely(is_hugetlb_entry_migration(pte))) {
5639 swp_entry_t entry = pte_to_swp_entry(pte);
5640
Alistair Popple4dd845b2021-06-30 18:54:09 -07005641 if (is_writable_migration_entry(entry)) {
Naoya Horiguchia8bda282015-02-11 15:25:28 -08005642 pte_t newpte;
5643
Alistair Popple4dd845b2021-06-30 18:54:09 -07005644 entry = make_readable_migration_entry(
5645 swp_offset(entry));
Naoya Horiguchia8bda282015-02-11 15:25:28 -08005646 newpte = swp_entry_to_pte(entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07005647 set_huge_swap_pte_at(mm, address, ptep,
5648 newpte, huge_page_size(h));
Naoya Horiguchia8bda282015-02-11 15:25:28 -08005649 pages++;
5650 }
5651 spin_unlock(ptl);
5652 continue;
5653 }
5654 if (!huge_pte_none(pte)) {
Aneesh Kumar K.V023bdd02019-03-05 15:46:37 -08005655 pte_t old_pte;
Christophe Leroy79c1c592021-06-30 18:48:00 -07005656 unsigned int shift = huge_page_shift(hstate_vma(vma));
Aneesh Kumar K.V023bdd02019-03-05 15:46:37 -08005657
5658 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5659 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
Christophe Leroy79c1c592021-06-30 18:48:00 -07005660 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
Aneesh Kumar K.V023bdd02019-03-05 15:46:37 -08005661 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005662 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005663 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005664 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005665 }
Mel Gormand8333522012-07-31 16:46:20 -07005666 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08005667 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07005668 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08005669 * once we release i_mmap_rwsem, another task can do the final put_page
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005670 * and that page table be reused and filled with junk. If we actually
5671 * did unshare a page of pmds, flush the range corresponding to the pud.
Mel Gormand8333522012-07-31 16:46:20 -07005672 */
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005673 if (shared_pmd)
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005674 flush_hugetlb_tlb_range(vma, range.start, range.end);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005675 else
5676 flush_hugetlb_tlb_range(vma, start, end);
Jérôme Glisse0f108512017-11-15 17:34:07 -08005677 /*
5678 * No need to call mmu_notifier_invalidate_range() we are downgrading
5679 * page table protection not changing it to point to a new page.
5680 *
Mike Rapoportad56b732018-03-21 21:22:47 +02005681 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08005682 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08005683 i_mmap_unlock_write(vma->vm_file->f_mapping);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005684 mmu_notifier_invalidate_range_end(&range);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005685
5686 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005687}
5688
Mike Kravetz33b8f842021-02-24 12:09:54 -08005689/* Return true if reservation was successful, false otherwise. */
5690bool hugetlb_reserve_pages(struct inode *inode,
Mel Gormana1e78772008-07-23 21:27:23 -07005691 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00005692 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09005693 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07005694{
Mike Kravetz33b8f842021-02-24 12:09:54 -08005695 long chg, add = -1;
Andi Kleena5516432008-07-23 21:27:41 -07005696 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07005697 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005698 struct resv_map *resv_map;
Mina Almasry075a61d2020-04-01 21:11:28 -07005699 struct hugetlb_cgroup *h_cg = NULL;
Mina Almasry0db9d742020-04-01 21:11:25 -07005700 long gbl_reserve, regions_needed = 0;
Adam Litkee4e574b2007-10-16 01:26:19 -07005701
Mike Kravetz63489f82018-03-22 16:17:13 -07005702 /* This should never happen */
5703 if (from > to) {
5704 VM_WARN(1, "%s called with a negative range\n", __func__);
Mike Kravetz33b8f842021-02-24 12:09:54 -08005705 return false;
Mike Kravetz63489f82018-03-22 16:17:13 -07005706 }
5707
Mel Gormana1e78772008-07-23 21:27:23 -07005708 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00005709 * Only apply hugepage reservation if asked. At fault time, an
5710 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07005711 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00005712 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09005713 if (vm_flags & VM_NORESERVE)
Mike Kravetz33b8f842021-02-24 12:09:54 -08005714 return true;
Mel Gorman17c9d122009-02-11 16:34:16 +00005715
5716 /*
Mel Gormana1e78772008-07-23 21:27:23 -07005717 * Shared mappings base their reservation on the number of pages that
5718 * are already allocated on behalf of the file. Private mappings need
5719 * to reserve the full area even if read-only as mprotect() may be
5720 * called to make the mapping read-write. Assume !vma is a shm mapping
5721 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07005722 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzf27a5132019-05-13 17:22:55 -07005723 /*
5724 * resv_map can not be NULL as hugetlb_reserve_pages is only
5725 * called for inodes for which resv_maps were created (see
5726 * hugetlbfs_get_inode).
5727 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07005728 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005729
Mina Almasry0db9d742020-04-01 21:11:25 -07005730 chg = region_chg(resv_map, from, to, &regions_needed);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005731
5732 } else {
Mina Almasrye9fe92a2020-04-01 21:11:21 -07005733 /* Private mapping. */
Joonsoo Kim9119a412014-04-03 14:47:25 -07005734 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00005735 if (!resv_map)
Mike Kravetz33b8f842021-02-24 12:09:54 -08005736 return false;
Mel Gorman5a6fe122009-02-10 14:02:27 +00005737
Mel Gorman17c9d122009-02-11 16:34:16 +00005738 chg = to - from;
5739
Mel Gorman5a6fe122009-02-10 14:02:27 +00005740 set_vma_resv_map(vma, resv_map);
5741 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5742 }
5743
Mike Kravetz33b8f842021-02-24 12:09:54 -08005744 if (chg < 0)
Dave Hansenc50ac052012-05-29 15:06:46 -07005745 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00005746
Mike Kravetz33b8f842021-02-24 12:09:54 -08005747 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5748 chg * pages_per_huge_page(h), &h_cg) < 0)
Mina Almasry075a61d2020-04-01 21:11:28 -07005749 goto out_err;
Mina Almasry075a61d2020-04-01 21:11:28 -07005750
5751 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5752 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5753 * of the resv_map.
5754 */
5755 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5756 }
5757
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005758 /*
5759 * There must be enough pages in the subpool for the mapping. If
5760 * the subpool has a minimum size, there may be some global
5761 * reservations already in place (gbl_reserve).
5762 */
5763 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
Mike Kravetz33b8f842021-02-24 12:09:54 -08005764 if (gbl_reserve < 0)
Mina Almasry075a61d2020-04-01 21:11:28 -07005765 goto out_uncharge_cgroup;
Mel Gorman17c9d122009-02-11 16:34:16 +00005766
5767 /*
5768 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07005769 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00005770 */
Mike Kravetz33b8f842021-02-24 12:09:54 -08005771 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
Mina Almasry075a61d2020-04-01 21:11:28 -07005772 goto out_put_pages;
Mel Gorman17c9d122009-02-11 16:34:16 +00005773
5774 /*
5775 * Account for the reservations made. Shared mappings record regions
5776 * that have reservations as they are shared by multiple VMAs.
5777 * When the last VMA disappears, the region map says how much
5778 * the reservation was and the page cache tells how much of
5779 * the reservation was consumed. Private mappings are per-VMA and
5780 * only the consumed reservations are tracked. When the VMA
5781 * disappears, the original reservation is the VMA size and the
5782 * consumed reservations are stored in the map. Hence, nothing
5783 * else has to be done for private mappings here
5784 */
Mike Kravetz33039672015-06-24 16:57:58 -07005785 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Mina Almasry075a61d2020-04-01 21:11:28 -07005786 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
Mike Kravetz33039672015-06-24 16:57:58 -07005787
Mina Almasry0db9d742020-04-01 21:11:25 -07005788 if (unlikely(add < 0)) {
5789 hugetlb_acct_memory(h, -gbl_reserve);
Mina Almasry075a61d2020-04-01 21:11:28 -07005790 goto out_put_pages;
Mina Almasry0db9d742020-04-01 21:11:25 -07005791 } else if (unlikely(chg > add)) {
Mike Kravetz33039672015-06-24 16:57:58 -07005792 /*
5793 * pages in this range were added to the reserve
5794 * map between region_chg and region_add. This
5795 * indicates a race with alloc_huge_page. Adjust
5796 * the subpool and reserve counts modified above
5797 * based on the difference.
5798 */
5799 long rsv_adjust;
5800
Miaohe Lind85aecf2021-03-24 21:37:17 -07005801 /*
5802 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5803 * reference to h_cg->css. See comment below for detail.
5804 */
Mina Almasry075a61d2020-04-01 21:11:28 -07005805 hugetlb_cgroup_uncharge_cgroup_rsvd(
5806 hstate_index(h),
5807 (chg - add) * pages_per_huge_page(h), h_cg);
5808
Mike Kravetz33039672015-06-24 16:57:58 -07005809 rsv_adjust = hugepage_subpool_put_pages(spool,
5810 chg - add);
5811 hugetlb_acct_memory(h, -rsv_adjust);
Miaohe Lind85aecf2021-03-24 21:37:17 -07005812 } else if (h_cg) {
5813 /*
5814 * The file_regions will hold their own reference to
5815 * h_cg->css. So we should release the reference held
5816 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5817 * done.
5818 */
5819 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
Mike Kravetz33039672015-06-24 16:57:58 -07005820 }
5821 }
Mike Kravetz33b8f842021-02-24 12:09:54 -08005822 return true;
5823
Mina Almasry075a61d2020-04-01 21:11:28 -07005824out_put_pages:
5825 /* put back original number of pages, chg */
5826 (void)hugepage_subpool_put_pages(spool, chg);
5827out_uncharge_cgroup:
5828 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5829 chg * pages_per_huge_page(h), h_cg);
Dave Hansenc50ac052012-05-29 15:06:46 -07005830out_err:
Mike Kravetz5e911372015-09-08 15:01:28 -07005831 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mina Almasry0db9d742020-04-01 21:11:25 -07005832 /* Only call region_abort if the region_chg succeeded but the
5833 * region_add failed or didn't run.
5834 */
5835 if (chg >= 0 && add < 0)
5836 region_abort(resv_map, from, to, regions_needed);
Joonsoo Kimf031dd22014-04-03 14:47:28 -07005837 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5838 kref_put(&resv_map->refs, resv_map_release);
Mike Kravetz33b8f842021-02-24 12:09:54 -08005839 return false;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005840}
5841
Mike Kravetzb5cec282015-09-08 15:01:41 -07005842long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5843 long freed)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005844{
Andi Kleena5516432008-07-23 21:27:41 -07005845 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07005846 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005847 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07005848 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005849 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08005850
Mike Kravetzf27a5132019-05-13 17:22:55 -07005851 /*
5852 * Since this routine can be called in the evict inode path for all
5853 * hugetlbfs inodes, resv_map could be NULL.
5854 */
Mike Kravetzb5cec282015-09-08 15:01:41 -07005855 if (resv_map) {
5856 chg = region_del(resv_map, start, end);
5857 /*
5858 * region_del() can fail in the rare case where a region
5859 * must be split and another region descriptor can not be
5860 * allocated. If end == LONG_MAX, it will not fail.
5861 */
5862 if (chg < 0)
5863 return chg;
5864 }
5865
Ken Chen45c682a2007-11-14 16:59:44 -08005866 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07005867 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08005868 spin_unlock(&inode->i_lock);
5869
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005870 /*
5871 * If the subpool has a minimum size, the number of global
5872 * reservations to be released may be adjusted.
Miaohe Lindddf31a2021-05-04 18:34:35 -07005873 *
5874 * Note that !resv_map implies freed == 0. So (chg - freed)
5875 * won't go negative.
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005876 */
5877 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5878 hugetlb_acct_memory(h, -gbl_reserve);
Mike Kravetzb5cec282015-09-08 15:01:41 -07005879
5880 return 0;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005881}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09005882
Steve Capper3212b532013-04-23 12:35:02 +01005883#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5884static unsigned long page_table_shareable(struct vm_area_struct *svma,
5885 struct vm_area_struct *vma,
5886 unsigned long addr, pgoff_t idx)
5887{
5888 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5889 svma->vm_start;
5890 unsigned long sbase = saddr & PUD_MASK;
5891 unsigned long s_end = sbase + PUD_SIZE;
5892
5893 /* Allow segments to share if only one is marked locked */
Eric B Munsonde60f5f2015-11-05 18:51:36 -08005894 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5895 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
Steve Capper3212b532013-04-23 12:35:02 +01005896
5897 /*
5898 * match the virtual addresses, permission and the alignment of the
5899 * page table page.
5900 */
5901 if (pmd_index(addr) != pmd_index(saddr) ||
5902 vm_flags != svm_flags ||
Miaohe Lin07e51ed2021-02-24 12:07:39 -08005903 !range_in_vma(svma, sbase, s_end))
Steve Capper3212b532013-04-23 12:35:02 +01005904 return 0;
5905
5906 return saddr;
5907}
5908
Nicholas Krause31aafb42015-09-04 15:47:58 -07005909static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
Steve Capper3212b532013-04-23 12:35:02 +01005910{
5911 unsigned long base = addr & PUD_MASK;
5912 unsigned long end = base + PUD_SIZE;
5913
5914 /*
5915 * check on proper vm_flags and page table alignment
5916 */
Mike Kravetz017b1662018-10-05 15:51:29 -07005917 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
Nicholas Krause31aafb42015-09-04 15:47:58 -07005918 return true;
5919 return false;
Steve Capper3212b532013-04-23 12:35:02 +01005920}
5921
Peter Xuc1991e02021-05-04 18:33:04 -07005922bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5923{
5924#ifdef CONFIG_USERFAULTFD
5925 if (uffd_disable_huge_pmd_share(vma))
5926 return false;
5927#endif
5928 return vma_shareable(vma, addr);
5929}
5930
Steve Capper3212b532013-04-23 12:35:02 +01005931/*
Mike Kravetz017b1662018-10-05 15:51:29 -07005932 * Determine if start,end range within vma could be mapped by shared pmd.
5933 * If yes, adjust start and end to cover range associated with possible
5934 * shared pmd mappings.
5935 */
5936void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5937 unsigned long *start, unsigned long *end)
5938{
Li Xinhaia1ba9da2021-02-24 12:06:54 -08005939 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5940 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
Mike Kravetz017b1662018-10-05 15:51:29 -07005941
Li Xinhaia1ba9da2021-02-24 12:06:54 -08005942 /*
Ingo Molnarf0953a12021-05-06 18:06:47 -07005943 * vma needs to span at least one aligned PUD size, and the range
5944 * must be at least partially within in.
Li Xinhaia1ba9da2021-02-24 12:06:54 -08005945 */
5946 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5947 (*end <= v_start) || (*start >= v_end))
Mike Kravetz017b1662018-10-05 15:51:29 -07005948 return;
5949
Peter Xu75802ca62020-08-06 23:26:11 -07005950 /* Extend the range to be PUD aligned for a worst case scenario */
Li Xinhaia1ba9da2021-02-24 12:06:54 -08005951 if (*start > v_start)
5952 *start = ALIGN_DOWN(*start, PUD_SIZE);
Mike Kravetz017b1662018-10-05 15:51:29 -07005953
Li Xinhaia1ba9da2021-02-24 12:06:54 -08005954 if (*end < v_end)
5955 *end = ALIGN(*end, PUD_SIZE);
Mike Kravetz017b1662018-10-05 15:51:29 -07005956}
5957
5958/*
Steve Capper3212b532013-04-23 12:35:02 +01005959 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5960 * and returns the corresponding pte. While this is not necessary for the
5961 * !shared pmd case because we can allocate the pmd later as well, it makes the
Mike Kravetzc0d03812020-04-01 21:11:05 -07005962 * code much cleaner.
5963 *
Mike Kravetz0bf7b642020-10-13 16:56:42 -07005964 * This routine must be called with i_mmap_rwsem held in at least read mode if
5965 * sharing is possible. For hugetlbfs, this prevents removal of any page
5966 * table entries associated with the address space. This is important as we
5967 * are setting up sharing based on existing page table entries (mappings).
5968 *
5969 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5970 * huge_pte_alloc know that sharing is not possible and do not take
5971 * i_mmap_rwsem as a performance optimization. This is handled by the
5972 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5973 * only required for subsequent processing.
Steve Capper3212b532013-04-23 12:35:02 +01005974 */
Peter Xuaec44e02021-05-04 18:33:00 -07005975pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5976 unsigned long addr, pud_t *pud)
Steve Capper3212b532013-04-23 12:35:02 +01005977{
Steve Capper3212b532013-04-23 12:35:02 +01005978 struct address_space *mapping = vma->vm_file->f_mapping;
5979 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5980 vma->vm_pgoff;
5981 struct vm_area_struct *svma;
5982 unsigned long saddr;
5983 pte_t *spte = NULL;
5984 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005985 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01005986
Mike Kravetz0bf7b642020-10-13 16:56:42 -07005987 i_mmap_assert_locked(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01005988 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5989 if (svma == vma)
5990 continue;
5991
5992 saddr = page_table_shareable(svma, vma, addr, idx);
5993 if (saddr) {
Punit Agrawal7868a202017-07-06 15:39:42 -07005994 spte = huge_pte_offset(svma->vm_mm, saddr,
5995 vma_mmu_pagesize(svma));
Steve Capper3212b532013-04-23 12:35:02 +01005996 if (spte) {
5997 get_page(virt_to_page(spte));
5998 break;
5999 }
6000 }
6001 }
6002
6003 if (!spte)
6004 goto out;
6005
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08006006 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08006007 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01006008 pud_populate(mm, pud,
6009 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovc17b1f42016-06-24 14:49:51 -07006010 mm_inc_nr_pmds(mm);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08006011 } else {
Steve Capper3212b532013-04-23 12:35:02 +01006012 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08006013 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08006014 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01006015out:
6016 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Steve Capper3212b532013-04-23 12:35:02 +01006017 return pte;
6018}
6019
6020/*
6021 * unmap huge page backed by shared pte.
6022 *
6023 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
6024 * indicated by page_count > 1, unmap is achieved by clearing pud and
6025 * decrementing the ref count. If count == 1, the pte page is not shared.
6026 *
Mike Kravetzc0d03812020-04-01 21:11:05 -07006027 * Called with page table lock held and i_mmap_rwsem held in write mode.
Steve Capper3212b532013-04-23 12:35:02 +01006028 *
6029 * returns: 1 successfully unmapped a shared pte page
6030 * 0 the underlying pte page is not shared, or it is the last user
6031 */
Mike Kravetz34ae2042020-08-11 18:31:38 -07006032int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6033 unsigned long *addr, pte_t *ptep)
Steve Capper3212b532013-04-23 12:35:02 +01006034{
6035 pgd_t *pgd = pgd_offset(mm, *addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03006036 p4d_t *p4d = p4d_offset(pgd, *addr);
6037 pud_t *pud = pud_offset(p4d, *addr);
Steve Capper3212b532013-04-23 12:35:02 +01006038
Mike Kravetz34ae2042020-08-11 18:31:38 -07006039 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
Steve Capper3212b532013-04-23 12:35:02 +01006040 BUG_ON(page_count(virt_to_page(ptep)) == 0);
6041 if (page_count(virt_to_page(ptep)) == 1)
6042 return 0;
6043
6044 pud_clear(pud);
6045 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08006046 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01006047 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
6048 return 1;
6049}
Peter Xuc1991e02021-05-04 18:33:04 -07006050
Steve Capper9e5fc742013-04-30 08:02:03 +01006051#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
Peter Xuaec44e02021-05-04 18:33:00 -07006052pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6053 unsigned long addr, pud_t *pud)
Steve Capper9e5fc742013-04-30 08:02:03 +01006054{
6055 return NULL;
6056}
Zhang Zhene81f2d22015-06-24 16:56:13 -07006057
Mike Kravetz34ae2042020-08-11 18:31:38 -07006058int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6059 unsigned long *addr, pte_t *ptep)
Zhang Zhene81f2d22015-06-24 16:56:13 -07006060{
6061 return 0;
6062}
Mike Kravetz017b1662018-10-05 15:51:29 -07006063
6064void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6065 unsigned long *start, unsigned long *end)
6066{
6067}
Peter Xuc1991e02021-05-04 18:33:04 -07006068
6069bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6070{
6071 return false;
6072}
Steve Capper3212b532013-04-23 12:35:02 +01006073#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6074
Steve Capper9e5fc742013-04-30 08:02:03 +01006075#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
Peter Xuaec44e02021-05-04 18:33:00 -07006076pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
Steve Capper9e5fc742013-04-30 08:02:03 +01006077 unsigned long addr, unsigned long sz)
6078{
6079 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03006080 p4d_t *p4d;
Steve Capper9e5fc742013-04-30 08:02:03 +01006081 pud_t *pud;
6082 pte_t *pte = NULL;
6083
6084 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovf4f0a3d2017-11-29 16:11:30 -08006085 p4d = p4d_alloc(mm, pgd, addr);
6086 if (!p4d)
6087 return NULL;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03006088 pud = pud_alloc(mm, p4d, addr);
Steve Capper9e5fc742013-04-30 08:02:03 +01006089 if (pud) {
6090 if (sz == PUD_SIZE) {
6091 pte = (pte_t *)pud;
6092 } else {
6093 BUG_ON(sz != PMD_SIZE);
Peter Xuc1991e02021-05-04 18:33:04 -07006094 if (want_pmd_share(vma, addr) && pud_none(*pud))
Peter Xuaec44e02021-05-04 18:33:00 -07006095 pte = huge_pmd_share(mm, vma, addr, pud);
Steve Capper9e5fc742013-04-30 08:02:03 +01006096 else
6097 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6098 }
6099 }
Michal Hocko4e666312016-08-02 14:02:34 -07006100 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
Steve Capper9e5fc742013-04-30 08:02:03 +01006101
6102 return pte;
6103}
6104
Punit Agrawal9b19df22017-09-06 16:21:01 -07006105/*
6106 * huge_pte_offset() - Walk the page table to resolve the hugepage
6107 * entry at address @addr
6108 *
Li Xinhai8ac0b812020-06-03 16:00:53 -07006109 * Return: Pointer to page table entry (PUD or PMD) for
6110 * address @addr, or NULL if a !p*d_present() entry is encountered and the
Punit Agrawal9b19df22017-09-06 16:21:01 -07006111 * size @sz doesn't match the hugepage size at this level of the page
6112 * table.
6113 */
Punit Agrawal7868a202017-07-06 15:39:42 -07006114pte_t *huge_pte_offset(struct mm_struct *mm,
6115 unsigned long addr, unsigned long sz)
Steve Capper9e5fc742013-04-30 08:02:03 +01006116{
6117 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03006118 p4d_t *p4d;
Li Xinhai8ac0b812020-06-03 16:00:53 -07006119 pud_t *pud;
6120 pmd_t *pmd;
Steve Capper9e5fc742013-04-30 08:02:03 +01006121
6122 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03006123 if (!pgd_present(*pgd))
6124 return NULL;
6125 p4d = p4d_offset(pgd, addr);
6126 if (!p4d_present(*p4d))
6127 return NULL;
Punit Agrawal9b19df22017-09-06 16:21:01 -07006128
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03006129 pud = pud_offset(p4d, addr);
Li Xinhai8ac0b812020-06-03 16:00:53 -07006130 if (sz == PUD_SIZE)
6131 /* must be pud huge, non-present or none */
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03006132 return (pte_t *)pud;
Li Xinhai8ac0b812020-06-03 16:00:53 -07006133 if (!pud_present(*pud))
6134 return NULL;
6135 /* must have a valid entry and size to go further */
Punit Agrawal9b19df22017-09-06 16:21:01 -07006136
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03006137 pmd = pmd_offset(pud, addr);
Li Xinhai8ac0b812020-06-03 16:00:53 -07006138 /* must be pmd huge, non-present or none */
6139 return (pte_t *)pmd;
Steve Capper9e5fc742013-04-30 08:02:03 +01006140}
6141
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08006142#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
6143
6144/*
6145 * These functions are overwritable if your architecture needs its own
6146 * behavior.
6147 */
6148struct page * __weak
6149follow_huge_addr(struct mm_struct *mm, unsigned long address,
6150 int write)
6151{
6152 return ERR_PTR(-EINVAL);
6153}
6154
6155struct page * __weak
Aneesh Kumar K.V4dc71452017-07-06 15:38:56 -07006156follow_huge_pd(struct vm_area_struct *vma,
6157 unsigned long address, hugepd_t hpd, int flags, int pdshift)
6158{
6159 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6160 return NULL;
6161}
6162
6163struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01006164follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08006165 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01006166{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08006167 struct page *page = NULL;
6168 spinlock_t *ptl;
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07006169 pte_t pte;
John Hubbard3faa52c2020-04-01 21:05:29 -07006170
6171 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
6172 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
6173 (FOLL_PIN | FOLL_GET)))
6174 return NULL;
6175
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08006176retry:
6177 ptl = pmd_lockptr(mm, pmd);
6178 spin_lock(ptl);
6179 /*
6180 * make sure that the address range covered by this pmd is not
6181 * unmapped from other threads.
6182 */
6183 if (!pmd_huge(*pmd))
6184 goto out;
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07006185 pte = huge_ptep_get((pte_t *)pmd);
6186 if (pte_present(pte)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07006187 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
John Hubbard3faa52c2020-04-01 21:05:29 -07006188 /*
6189 * try_grab_page() should always succeed here, because: a) we
6190 * hold the pmd (ptl) lock, and b) we've just checked that the
6191 * huge pmd (head) page is present in the page tables. The ptl
6192 * prevents the head page and tail pages from being rearranged
6193 * in any way. So this page must be available at this point,
6194 * unless the page refcount overflowed:
6195 */
6196 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
6197 page = NULL;
6198 goto out;
6199 }
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08006200 } else {
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07006201 if (is_hugetlb_entry_migration(pte)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08006202 spin_unlock(ptl);
6203 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
6204 goto retry;
6205 }
6206 /*
6207 * hwpoisoned entry is treated as no_page_table in
6208 * follow_page_mask().
6209 */
6210 }
6211out:
6212 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01006213 return page;
6214}
6215
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08006216struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01006217follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08006218 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01006219{
John Hubbard3faa52c2020-04-01 21:05:29 -07006220 if (flags & (FOLL_GET | FOLL_PIN))
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08006221 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01006222
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08006223 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01006224}
6225
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -07006226struct page * __weak
6227follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
6228{
John Hubbard3faa52c2020-04-01 21:05:29 -07006229 if (flags & (FOLL_GET | FOLL_PIN))
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -07006230 return NULL;
6231
6232 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
6233}
6234
Naoya Horiguchi31caf662013-09-11 14:21:59 -07006235bool isolate_huge_page(struct page *page, struct list_head *list)
6236{
Naoya Horiguchibcc54222015-04-15 16:14:38 -07006237 bool ret = true;
6238
Mike Kravetzdb71ef72021-05-04 18:35:07 -07006239 spin_lock_irq(&hugetlb_lock);
Mike Kravetz8f251a32021-02-24 12:08:56 -08006240 if (!PageHeadHuge(page) ||
6241 !HPageMigratable(page) ||
Muchun Song0eb2df22021-02-04 18:32:10 -08006242 !get_page_unless_zero(page)) {
Naoya Horiguchibcc54222015-04-15 16:14:38 -07006243 ret = false;
6244 goto unlock;
6245 }
Mike Kravetz8f251a32021-02-24 12:08:56 -08006246 ClearHPageMigratable(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07006247 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07006248unlock:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07006249 spin_unlock_irq(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07006250 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07006251}
6252
Naoya Horiguchi25182f02021-06-15 18:23:13 -07006253int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
6254{
6255 int ret = 0;
6256
6257 *hugetlb = false;
6258 spin_lock_irq(&hugetlb_lock);
6259 if (PageHeadHuge(page)) {
6260 *hugetlb = true;
6261 if (HPageFreed(page) || HPageMigratable(page))
6262 ret = get_page_unless_zero(page);
Naoya Horiguchi0ed950d2021-06-28 19:43:17 -07006263 else
6264 ret = -EBUSY;
Naoya Horiguchi25182f02021-06-15 18:23:13 -07006265 }
6266 spin_unlock_irq(&hugetlb_lock);
6267 return ret;
6268}
6269
Naoya Horiguchi31caf662013-09-11 14:21:59 -07006270void putback_active_hugepage(struct page *page)
6271{
Mike Kravetzdb71ef72021-05-04 18:35:07 -07006272 spin_lock_irq(&hugetlb_lock);
Mike Kravetz8f251a32021-02-24 12:08:56 -08006273 SetHPageMigratable(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07006274 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07006275 spin_unlock_irq(&hugetlb_lock);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07006276 put_page(page);
6277}
Michal Hockoab5ac902018-01-31 16:20:48 -08006278
6279void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
6280{
6281 struct hstate *h = page_hstate(oldpage);
6282
6283 hugetlb_cgroup_migrate(oldpage, newpage);
6284 set_page_owner_migrate_reason(newpage, reason);
6285
6286 /*
6287 * transfer temporary state of the new huge page. This is
6288 * reverse to other transitions because the newpage is going to
6289 * be final while the old one will be freed so it takes over
6290 * the temporary status.
6291 *
6292 * Also note that we have to transfer the per-node surplus state
6293 * here as well otherwise the global surplus count will not match
6294 * the per-node's.
6295 */
Mike Kravetz9157c3112021-02-24 12:09:00 -08006296 if (HPageTemporary(newpage)) {
Michal Hockoab5ac902018-01-31 16:20:48 -08006297 int old_nid = page_to_nid(oldpage);
6298 int new_nid = page_to_nid(newpage);
6299
Mike Kravetz9157c3112021-02-24 12:09:00 -08006300 SetHPageTemporary(oldpage);
6301 ClearHPageTemporary(newpage);
Michal Hockoab5ac902018-01-31 16:20:48 -08006302
Miaohe Lin5af1ab12021-05-04 18:33:25 -07006303 /*
6304 * There is no need to transfer the per-node surplus state
6305 * when we do not cross the node.
6306 */
6307 if (new_nid == old_nid)
6308 return;
Mike Kravetzdb71ef72021-05-04 18:35:07 -07006309 spin_lock_irq(&hugetlb_lock);
Michal Hockoab5ac902018-01-31 16:20:48 -08006310 if (h->surplus_huge_pages_node[old_nid]) {
6311 h->surplus_huge_pages_node[old_nid]--;
6312 h->surplus_huge_pages_node[new_nid]++;
6313 }
Mike Kravetzdb71ef72021-05-04 18:35:07 -07006314 spin_unlock_irq(&hugetlb_lock);
Michal Hockoab5ac902018-01-31 16:20:48 -08006315 }
6316}
Roman Gushchincf11e852020-04-10 14:32:45 -07006317
Peter Xu6dfeaff2021-05-04 18:33:13 -07006318/*
6319 * This function will unconditionally remove all the shared pmd pgtable entries
6320 * within the specific vma for a hugetlbfs memory range.
6321 */
6322void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
6323{
6324 struct hstate *h = hstate_vma(vma);
6325 unsigned long sz = huge_page_size(h);
6326 struct mm_struct *mm = vma->vm_mm;
6327 struct mmu_notifier_range range;
6328 unsigned long address, start, end;
6329 spinlock_t *ptl;
6330 pte_t *ptep;
6331
6332 if (!(vma->vm_flags & VM_MAYSHARE))
6333 return;
6334
6335 start = ALIGN(vma->vm_start, PUD_SIZE);
6336 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6337
6338 if (start >= end)
6339 return;
6340
6341 /*
6342 * No need to call adjust_range_if_pmd_sharing_possible(), because
6343 * we have already done the PUD_SIZE alignment.
6344 */
6345 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
6346 start, end);
6347 mmu_notifier_invalidate_range_start(&range);
6348 i_mmap_lock_write(vma->vm_file->f_mapping);
6349 for (address = start; address < end; address += PUD_SIZE) {
6350 unsigned long tmp = address;
6351
6352 ptep = huge_pte_offset(mm, address, sz);
6353 if (!ptep)
6354 continue;
6355 ptl = huge_pte_lock(h, mm, ptep);
6356 /* We don't want 'address' to be changed */
6357 huge_pmd_unshare(mm, vma, &tmp, ptep);
6358 spin_unlock(ptl);
6359 }
6360 flush_hugetlb_tlb_range(vma, start, end);
6361 i_mmap_unlock_write(vma->vm_file->f_mapping);
6362 /*
6363 * No need to call mmu_notifier_invalidate_range(), see
6364 * Documentation/vm/mmu_notifier.rst.
6365 */
6366 mmu_notifier_invalidate_range_end(&range);
6367}
6368
Roman Gushchincf11e852020-04-10 14:32:45 -07006369#ifdef CONFIG_CMA
Roman Gushchincf11e852020-04-10 14:32:45 -07006370static bool cma_reserve_called __initdata;
6371
6372static int __init cmdline_parse_hugetlb_cma(char *p)
6373{
6374 hugetlb_cma_size = memparse(p, &p);
6375 return 0;
6376}
6377
6378early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
6379
6380void __init hugetlb_cma_reserve(int order)
6381{
6382 unsigned long size, reserved, per_node;
6383 int nid;
6384
6385 cma_reserve_called = true;
6386
6387 if (!hugetlb_cma_size)
6388 return;
6389
6390 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
6391 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
6392 (PAGE_SIZE << order) / SZ_1M);
6393 return;
6394 }
6395
6396 /*
6397 * If 3 GB area is requested on a machine with 4 numa nodes,
6398 * let's allocate 1 GB on first three nodes and ignore the last one.
6399 */
6400 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
6401 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
6402 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
6403
6404 reserved = 0;
6405 for_each_node_state(nid, N_ONLINE) {
6406 int res;
Barry Song2281f792020-08-24 11:03:09 +12006407 char name[CMA_MAX_NAME];
Roman Gushchincf11e852020-04-10 14:32:45 -07006408
6409 size = min(per_node, hugetlb_cma_size - reserved);
6410 size = round_up(size, PAGE_SIZE << order);
6411
Barry Song2281f792020-08-24 11:03:09 +12006412 snprintf(name, sizeof(name), "hugetlb%d", nid);
Roman Gushchincf11e852020-04-10 14:32:45 -07006413 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
Barry Song29d0f412020-08-11 18:32:00 -07006414 0, false, name,
Roman Gushchincf11e852020-04-10 14:32:45 -07006415 &hugetlb_cma[nid], nid);
6416 if (res) {
6417 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
6418 res, nid);
6419 continue;
6420 }
6421
6422 reserved += size;
6423 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
6424 size / SZ_1M, nid);
6425
6426 if (reserved >= hugetlb_cma_size)
6427 break;
6428 }
6429}
6430
6431void __init hugetlb_cma_check(void)
6432{
6433 if (!hugetlb_cma_size || cma_reserve_called)
6434 return;
6435
6436 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
6437}
6438
6439#endif /* CONFIG_CMA */