blob: 8254e8f6db6b6c534e910770542713a8d1ccdb8c [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01003 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07007#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04008#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/sysctl.h>
10#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070011#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070012#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070013#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080014#include <linux/mempolicy.h>
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -070015#include <linux/compiler.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080016#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080017#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070018#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070019#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Ingo Molnar174cd4b2017-02-02 19:15:33 +010021#include <linux/sched/signal.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090022#include <linux/rmap.h>
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +090023#include <linux/swap.h>
24#include <linux/swapops.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070025#include <linux/jhash.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070026
David Gibson63551ae2005-06-21 17:14:44 -070027#include <asm/page.h>
28#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070029#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070030
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070031#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070032#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070033#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a3052302009-12-14 17:58:25 -080034#include <linux/node.h>
Mike Kravetz1a1aad82017-02-22 15:43:01 -080035#include <linux/userfaultfd_k.h>
Nick Piggin7835e982006-03-22 00:08:40 -080036#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070037
Andrey Ryabinin753162c2015-02-10 14:11:36 -080038int hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070039
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070040int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070041unsigned int default_hstate_idx;
42struct hstate hstates[HUGE_MAX_HSTATE];
Naoya Horiguchi641844f2015-06-24 16:56:59 -070043/*
44 * Minimum page order among possible hugepage sizes, set to a proper value
45 * at boot time.
46 */
47static unsigned int minimum_order __read_mostly = UINT_MAX;
Andi Kleene5ff2152008-07-23 21:27:42 -070048
Jon Tollefson53ba51d2008-07-23 21:27:52 -070049__initdata LIST_HEAD(huge_boot_pages);
50
Andi Kleene5ff2152008-07-23 21:27:42 -070051/* for command line parsing */
52static struct hstate * __initdata parsed_hstate;
53static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070054static unsigned long __initdata default_hstate_size;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -070055static bool __initdata parsed_valid_hugepagesz = true;
Andi Kleene5ff2152008-07-23 21:27:42 -070056
David Gibson3935baa2006-03-22 00:08:53 -080057/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070058 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
59 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080060 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070061DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080062
Davidlohr Bueso8382d912014-04-03 14:47:31 -070063/*
64 * Serializes faults on the same logical page. This is used to
65 * prevent spurious OOMs when the hugepage pool is fully utilized.
66 */
67static int num_fault_mutexes;
Mike Kravetzc672c7f2015-09-08 15:01:35 -070068struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
Davidlohr Bueso8382d912014-04-03 14:47:31 -070069
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -070070/* Forward declaration */
71static int hugetlb_acct_memory(struct hstate *h, long delta);
72
David Gibson90481622012-03-21 16:34:12 -070073static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
74{
75 bool free = (spool->count == 0) && (spool->used_hpages == 0);
76
77 spin_unlock(&spool->lock);
78
79 /* If no pages are used, and no other handles to the subpool
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -070080 * remain, give up any reservations mased on minimum size and
81 * free the subpool */
82 if (free) {
83 if (spool->min_hpages != -1)
84 hugetlb_acct_memory(spool->hstate,
85 -spool->min_hpages);
David Gibson90481622012-03-21 16:34:12 -070086 kfree(spool);
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -070087 }
David Gibson90481622012-03-21 16:34:12 -070088}
89
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -070090struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
91 long min_hpages)
David Gibson90481622012-03-21 16:34:12 -070092{
93 struct hugepage_subpool *spool;
94
Mike Kravetzc6a91822015-04-15 16:13:36 -070095 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
David Gibson90481622012-03-21 16:34:12 -070096 if (!spool)
97 return NULL;
98
99 spin_lock_init(&spool->lock);
100 spool->count = 1;
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700101 spool->max_hpages = max_hpages;
102 spool->hstate = h;
103 spool->min_hpages = min_hpages;
104
105 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
106 kfree(spool);
107 return NULL;
108 }
109 spool->rsv_hpages = min_hpages;
David Gibson90481622012-03-21 16:34:12 -0700110
111 return spool;
112}
113
114void hugepage_put_subpool(struct hugepage_subpool *spool)
115{
116 spin_lock(&spool->lock);
117 BUG_ON(!spool->count);
118 spool->count--;
119 unlock_or_release_subpool(spool);
120}
121
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700122/*
123 * Subpool accounting for allocating and reserving pages.
124 * Return -ENOMEM if there are not enough resources to satisfy the
125 * the request. Otherwise, return the number of pages by which the
126 * global pools must be adjusted (upward). The returned value may
127 * only be different than the passed value (delta) in the case where
128 * a subpool minimum size must be manitained.
129 */
130static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700131 long delta)
132{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700133 long ret = delta;
David Gibson90481622012-03-21 16:34:12 -0700134
135 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700136 return ret;
David Gibson90481622012-03-21 16:34:12 -0700137
138 spin_lock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700139
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700140 if (spool->max_hpages != -1) { /* maximum size accounting */
141 if ((spool->used_hpages + delta) <= spool->max_hpages)
142 spool->used_hpages += delta;
143 else {
144 ret = -ENOMEM;
145 goto unlock_ret;
146 }
147 }
148
Mike Kravetz09a95e22016-05-19 17:11:01 -0700149 /* minimum size accounting */
150 if (spool->min_hpages != -1 && spool->rsv_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700151 if (delta > spool->rsv_hpages) {
152 /*
153 * Asking for more reserves than those already taken on
154 * behalf of subpool. Return difference.
155 */
156 ret = delta - spool->rsv_hpages;
157 spool->rsv_hpages = 0;
158 } else {
159 ret = 0; /* reserves already accounted for */
160 spool->rsv_hpages -= delta;
161 }
162 }
163
164unlock_ret:
165 spin_unlock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700166 return ret;
167}
168
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700169/*
170 * Subpool accounting for freeing and unreserving pages.
171 * Return the number of global page reservations that must be dropped.
172 * The return value may only be different than the passed value (delta)
173 * in the case where a subpool minimum size must be maintained.
174 */
175static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700176 long delta)
177{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700178 long ret = delta;
179
David Gibson90481622012-03-21 16:34:12 -0700180 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700181 return delta;
David Gibson90481622012-03-21 16:34:12 -0700182
183 spin_lock(&spool->lock);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700184
185 if (spool->max_hpages != -1) /* maximum size accounting */
186 spool->used_hpages -= delta;
187
Mike Kravetz09a95e22016-05-19 17:11:01 -0700188 /* minimum size accounting */
189 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700190 if (spool->rsv_hpages + delta <= spool->min_hpages)
191 ret = 0;
192 else
193 ret = spool->rsv_hpages + delta - spool->min_hpages;
194
195 spool->rsv_hpages += delta;
196 if (spool->rsv_hpages > spool->min_hpages)
197 spool->rsv_hpages = spool->min_hpages;
198 }
199
200 /*
201 * If hugetlbfs_put_super couldn't free spool due to an outstanding
202 * quota reference, free it now.
203 */
David Gibson90481622012-03-21 16:34:12 -0700204 unlock_or_release_subpool(spool);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700205
206 return ret;
David Gibson90481622012-03-21 16:34:12 -0700207}
208
209static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
210{
211 return HUGETLBFS_SB(inode->i_sb)->spool;
212}
213
214static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
215{
Al Viro496ad9a2013-01-23 17:07:38 -0500216 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700217}
218
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700219/*
Andy Whitcroft96822902008-07-23 21:27:29 -0700220 * Region tracking -- allows tracking of reservations and instantiated pages
221 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700222 *
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700223 * The region data structures are embedded into a resv_map and protected
224 * by a resv_map's lock. The set of regions within the resv_map represent
225 * reservations for huge pages, or huge pages that have already been
226 * instantiated within the map. The from and to elements are huge page
227 * indicies into the associated mapping. from indicates the starting index
228 * of the region. to represents the first index past the end of the region.
229 *
230 * For example, a file region structure with from == 0 and to == 4 represents
231 * four huge pages in a mapping. It is important to note that the to element
232 * represents the first element past the end of the region. This is used in
233 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
234 *
235 * Interval notation of the form [from, to) will be used to indicate that
236 * the endpoint from is inclusive and to is exclusive.
Andy Whitcroft96822902008-07-23 21:27:29 -0700237 */
238struct file_region {
239 struct list_head link;
240 long from;
241 long to;
242};
243
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700244/*
245 * Add the huge page range represented by [f, t) to the reserve
Mike Kravetz5e911372015-09-08 15:01:28 -0700246 * map. In the normal case, existing regions will be expanded
247 * to accommodate the specified range. Sufficient regions should
248 * exist for expansion due to the previous call to region_chg
249 * with the same range. However, it is possible that region_del
250 * could have been called after region_chg and modifed the map
251 * in such a way that no region exists to be expanded. In this
252 * case, pull a region descriptor from the cache associated with
253 * the map and use that for the new range.
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700254 *
255 * Return the number of new huge pages added to the map. This
256 * number is greater than or equal to zero.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700257 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700258static long region_add(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700259{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700260 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700261 struct file_region *rg, *nrg, *trg;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700262 long add = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700263
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700264 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700265 /* Locate the region we are either in or before. */
266 list_for_each_entry(rg, head, link)
267 if (f <= rg->to)
268 break;
269
Mike Kravetz5e911372015-09-08 15:01:28 -0700270 /*
271 * If no region exists which can be expanded to include the
272 * specified range, the list must have been modified by an
273 * interleving call to region_del(). Pull a region descriptor
274 * from the cache and use it for this range.
275 */
276 if (&rg->link == head || t < rg->from) {
277 VM_BUG_ON(resv->region_cache_count <= 0);
278
279 resv->region_cache_count--;
280 nrg = list_first_entry(&resv->region_cache, struct file_region,
281 link);
282 list_del(&nrg->link);
283
284 nrg->from = f;
285 nrg->to = t;
286 list_add(&nrg->link, rg->link.prev);
287
288 add += t - f;
289 goto out_locked;
290 }
291
Andy Whitcroft96822902008-07-23 21:27:29 -0700292 /* Round our left edge to the current segment if it encloses us. */
293 if (f > rg->from)
294 f = rg->from;
295
296 /* Check for and consume any regions we now overlap with. */
297 nrg = rg;
298 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
299 if (&rg->link == head)
300 break;
301 if (rg->from > t)
302 break;
303
304 /* If this area reaches higher then extend our area to
305 * include it completely. If this is not the first area
306 * which we intend to reuse, free it. */
307 if (rg->to > t)
308 t = rg->to;
309 if (rg != nrg) {
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700310 /* Decrement return value by the deleted range.
311 * Another range will span this area so that by
312 * end of routine add will be >= zero
313 */
314 add -= (rg->to - rg->from);
Andy Whitcroft96822902008-07-23 21:27:29 -0700315 list_del(&rg->link);
316 kfree(rg);
317 }
318 }
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700319
320 add += (nrg->from - f); /* Added to beginning of region */
Andy Whitcroft96822902008-07-23 21:27:29 -0700321 nrg->from = f;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700322 add += t - nrg->to; /* Added to end of region */
Andy Whitcroft96822902008-07-23 21:27:29 -0700323 nrg->to = t;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700324
Mike Kravetz5e911372015-09-08 15:01:28 -0700325out_locked:
326 resv->adds_in_progress--;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700327 spin_unlock(&resv->lock);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700328 VM_BUG_ON(add < 0);
329 return add;
Andy Whitcroft96822902008-07-23 21:27:29 -0700330}
331
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700332/*
333 * Examine the existing reserve map and determine how many
334 * huge pages in the specified range [f, t) are NOT currently
335 * represented. This routine is called before a subsequent
336 * call to region_add that will actually modify the reserve
337 * map to add the specified range [f, t). region_chg does
338 * not change the number of huge pages represented by the
339 * map. However, if the existing regions in the map can not
340 * be expanded to represent the new range, a new file_region
341 * structure is added to the map as a placeholder. This is
342 * so that the subsequent region_add call will have all the
343 * regions it needs and will not fail.
344 *
Mike Kravetz5e911372015-09-08 15:01:28 -0700345 * Upon entry, region_chg will also examine the cache of region descriptors
346 * associated with the map. If there are not enough descriptors cached, one
347 * will be allocated for the in progress add operation.
348 *
349 * Returns the number of huge pages that need to be added to the existing
350 * reservation map for the range [f, t). This number is greater or equal to
351 * zero. -ENOMEM is returned if a new file_region structure or cache entry
352 * is needed and can not be allocated.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700353 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700354static long region_chg(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700355{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700356 struct list_head *head = &resv->regions;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700357 struct file_region *rg, *nrg = NULL;
Andy Whitcroft96822902008-07-23 21:27:29 -0700358 long chg = 0;
359
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700360retry:
361 spin_lock(&resv->lock);
Mike Kravetz5e911372015-09-08 15:01:28 -0700362retry_locked:
363 resv->adds_in_progress++;
364
365 /*
366 * Check for sufficient descriptors in the cache to accommodate
367 * the number of in progress add operations.
368 */
369 if (resv->adds_in_progress > resv->region_cache_count) {
370 struct file_region *trg;
371
372 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
373 /* Must drop lock to allocate a new descriptor. */
374 resv->adds_in_progress--;
375 spin_unlock(&resv->lock);
376
377 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800378 if (!trg) {
379 kfree(nrg);
Mike Kravetz5e911372015-09-08 15:01:28 -0700380 return -ENOMEM;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800381 }
Mike Kravetz5e911372015-09-08 15:01:28 -0700382
383 spin_lock(&resv->lock);
384 list_add(&trg->link, &resv->region_cache);
385 resv->region_cache_count++;
386 goto retry_locked;
387 }
388
Andy Whitcroft96822902008-07-23 21:27:29 -0700389 /* Locate the region we are before or in. */
390 list_for_each_entry(rg, head, link)
391 if (f <= rg->to)
392 break;
393
394 /* If we are below the current region then a new region is required.
395 * Subtle, allocate a new region at the position but make it zero
396 * size such that we can guarantee to record the reservation. */
397 if (&rg->link == head || t < rg->from) {
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700398 if (!nrg) {
Mike Kravetz5e911372015-09-08 15:01:28 -0700399 resv->adds_in_progress--;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700400 spin_unlock(&resv->lock);
401 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
402 if (!nrg)
403 return -ENOMEM;
Andy Whitcroft96822902008-07-23 21:27:29 -0700404
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700405 nrg->from = f;
406 nrg->to = f;
407 INIT_LIST_HEAD(&nrg->link);
408 goto retry;
409 }
410
411 list_add(&nrg->link, rg->link.prev);
412 chg = t - f;
413 goto out_nrg;
Andy Whitcroft96822902008-07-23 21:27:29 -0700414 }
415
416 /* Round our left edge to the current segment if it encloses us. */
417 if (f > rg->from)
418 f = rg->from;
419 chg = t - f;
420
421 /* Check for and consume any regions we now overlap with. */
422 list_for_each_entry(rg, rg->link.prev, link) {
423 if (&rg->link == head)
424 break;
425 if (rg->from > t)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700426 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700427
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300428 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700429 * us then we must extend ourselves. Account for its
430 * existing reservation. */
431 if (rg->to > t) {
432 chg += rg->to - t;
433 t = rg->to;
434 }
435 chg -= rg->to - rg->from;
436 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700437
438out:
439 spin_unlock(&resv->lock);
440 /* We already know we raced and no longer need the new region */
441 kfree(nrg);
442 return chg;
443out_nrg:
444 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700445 return chg;
446}
447
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700448/*
Mike Kravetz5e911372015-09-08 15:01:28 -0700449 * Abort the in progress add operation. The adds_in_progress field
450 * of the resv_map keeps track of the operations in progress between
451 * calls to region_chg and region_add. Operations are sometimes
452 * aborted after the call to region_chg. In such cases, region_abort
453 * is called to decrement the adds_in_progress counter.
454 *
455 * NOTE: The range arguments [f, t) are not needed or used in this
456 * routine. They are kept to make reading the calling code easier as
457 * arguments will match the associated region_chg call.
458 */
459static void region_abort(struct resv_map *resv, long f, long t)
460{
461 spin_lock(&resv->lock);
462 VM_BUG_ON(!resv->region_cache_count);
463 resv->adds_in_progress--;
464 spin_unlock(&resv->lock);
465}
466
467/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700468 * Delete the specified range [f, t) from the reserve map. If the
469 * t parameter is LONG_MAX, this indicates that ALL regions after f
470 * should be deleted. Locate the regions which intersect [f, t)
471 * and either trim, delete or split the existing regions.
472 *
473 * Returns the number of huge pages deleted from the reserve map.
474 * In the normal case, the return value is zero or more. In the
475 * case where a region must be split, a new region descriptor must
476 * be allocated. If the allocation fails, -ENOMEM will be returned.
477 * NOTE: If the parameter t == LONG_MAX, then we will never split
478 * a region and possibly return -ENOMEM. Callers specifying
479 * t == LONG_MAX do not need to check for -ENOMEM error.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700480 */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700481static long region_del(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700482{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700483 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700484 struct file_region *rg, *trg;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700485 struct file_region *nrg = NULL;
486 long del = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700487
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700488retry:
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700489 spin_lock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700490 list_for_each_entry_safe(rg, trg, head, link) {
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800491 /*
492 * Skip regions before the range to be deleted. file_region
493 * ranges are normally of the form [from, to). However, there
494 * may be a "placeholder" entry in the map which is of the form
495 * (from, to) with from == to. Check for placeholder entries
496 * at the beginning of the range to be deleted.
497 */
498 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700499 continue;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800500
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700501 if (rg->from >= t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700502 break;
Andy Whitcroft96822902008-07-23 21:27:29 -0700503
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700504 if (f > rg->from && t < rg->to) { /* Must split region */
505 /*
506 * Check for an entry in the cache before dropping
507 * lock and attempting allocation.
508 */
509 if (!nrg &&
510 resv->region_cache_count > resv->adds_in_progress) {
511 nrg = list_first_entry(&resv->region_cache,
512 struct file_region,
513 link);
514 list_del(&nrg->link);
515 resv->region_cache_count--;
516 }
517
518 if (!nrg) {
519 spin_unlock(&resv->lock);
520 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
521 if (!nrg)
522 return -ENOMEM;
523 goto retry;
524 }
525
526 del += t - f;
527
528 /* New entry for end of split region */
529 nrg->from = t;
530 nrg->to = rg->to;
531 INIT_LIST_HEAD(&nrg->link);
532
533 /* Original entry is trimmed */
534 rg->to = f;
535
536 list_add(&nrg->link, &rg->link);
537 nrg = NULL;
538 break;
539 }
540
541 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
542 del += rg->to - rg->from;
543 list_del(&rg->link);
544 kfree(rg);
545 continue;
546 }
547
548 if (f <= rg->from) { /* Trim beginning of region */
549 del += t - rg->from;
550 rg->from = t;
551 } else { /* Trim end of region */
552 del += rg->to - f;
553 rg->to = f;
554 }
Andy Whitcroft96822902008-07-23 21:27:29 -0700555 }
556
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700557 spin_unlock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700558 kfree(nrg);
559 return del;
Andy Whitcroft96822902008-07-23 21:27:29 -0700560}
561
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700562/*
Mike Kravetzb5cec282015-09-08 15:01:41 -0700563 * A rare out of memory error was encountered which prevented removal of
564 * the reserve map region for a page. The huge page itself was free'ed
565 * and removed from the page cache. This routine will adjust the subpool
566 * usage count, and the global reserve count if needed. By incrementing
567 * these counts, the reserve map entry which could not be deleted will
568 * appear as a "reserved" entry instead of simply dangling with incorrect
569 * counts.
570 */
zhong jiang72e29362016-10-07 17:02:01 -0700571void hugetlb_fix_reserve_counts(struct inode *inode)
Mike Kravetzb5cec282015-09-08 15:01:41 -0700572{
573 struct hugepage_subpool *spool = subpool_inode(inode);
574 long rsv_adjust;
575
576 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
zhong jiang72e29362016-10-07 17:02:01 -0700577 if (rsv_adjust) {
Mike Kravetzb5cec282015-09-08 15:01:41 -0700578 struct hstate *h = hstate_inode(inode);
579
580 hugetlb_acct_memory(h, 1);
581 }
582}
583
584/*
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700585 * Count and return the number of huge pages in the reserve map
586 * that intersect with the range [f, t).
587 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700588static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700589{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700590 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700591 struct file_region *rg;
592 long chg = 0;
593
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700594 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700595 /* Locate each segment we overlap with, and count that overlap. */
596 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700597 long seg_from;
598 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700599
600 if (rg->to <= f)
601 continue;
602 if (rg->from >= t)
603 break;
604
605 seg_from = max(rg->from, f);
606 seg_to = min(rg->to, t);
607
608 chg += seg_to - seg_from;
609 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700610 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700611
612 return chg;
613}
614
Andy Whitcroft96822902008-07-23 21:27:29 -0700615/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700616 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700617 * the mapping, in pagecache page units; huge pages here.
618 */
Andi Kleena5516432008-07-23 21:27:41 -0700619static pgoff_t vma_hugecache_offset(struct hstate *h,
620 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700621{
Andi Kleena5516432008-07-23 21:27:41 -0700622 return ((address - vma->vm_start) >> huge_page_shift(h)) +
623 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700624}
625
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900626pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
627 unsigned long address)
628{
629 return vma_hugecache_offset(hstate_vma(vma), vma, address);
630}
Dan Williamsdee41072016-05-14 12:20:44 -0700631EXPORT_SYMBOL_GPL(linear_hugepage_index);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900632
Andy Whitcroft84afd992008-07-23 21:27:32 -0700633/*
Mel Gorman08fba692009-01-06 14:38:53 -0800634 * Return the size of the pages allocated when backing a VMA. In the majority
635 * cases this will be same size as used by the page table entries.
636 */
637unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
638{
639 struct hstate *hstate;
640
641 if (!is_vm_hugetlb_page(vma))
642 return PAGE_SIZE;
643
644 hstate = hstate_vma(vma);
645
Wanpeng Li2415cf12013-07-03 15:02:43 -0700646 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800647}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200648EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800649
650/*
Mel Gorman33402892009-01-06 14:38:54 -0800651 * Return the page size being used by the MMU to back a VMA. In the majority
652 * of cases, the page size used by the kernel matches the MMU size. On
653 * architectures where it differs, an architecture-specific version of this
654 * function is required.
655 */
656#ifndef vma_mmu_pagesize
657unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
658{
659 return vma_kernel_pagesize(vma);
660}
661#endif
662
663/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700664 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
665 * bits of the reservation map pointer, which are always clear due to
666 * alignment.
667 */
668#define HPAGE_RESV_OWNER (1UL << 0)
669#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700670#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700671
Mel Gormana1e78772008-07-23 21:27:23 -0700672/*
673 * These helpers are used to track how many pages are reserved for
674 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
675 * is guaranteed to have their future faults succeed.
676 *
677 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
678 * the reserve counters are updated with the hugetlb_lock held. It is safe
679 * to reset the VMA at fork() time as it is not in use yet and there is no
680 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700681 *
682 * The private mapping reservation is represented in a subtly different
683 * manner to a shared mapping. A shared mapping has a region map associated
684 * with the underlying file, this region map represents the backing file
685 * pages which have ever had a reservation assigned which this persists even
686 * after the page is instantiated. A private mapping has a region map
687 * associated with the original mmap which is attached to all VMAs which
688 * reference it, this region map represents those offsets which have consumed
689 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700690 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700691static unsigned long get_vma_private_data(struct vm_area_struct *vma)
692{
693 return (unsigned long)vma->vm_private_data;
694}
695
696static void set_vma_private_data(struct vm_area_struct *vma,
697 unsigned long value)
698{
699 vma->vm_private_data = (void *)value;
700}
701
Joonsoo Kim9119a412014-04-03 14:47:25 -0700702struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700703{
704 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700705 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
706
707 if (!resv_map || !rg) {
708 kfree(resv_map);
709 kfree(rg);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700710 return NULL;
Mike Kravetz5e911372015-09-08 15:01:28 -0700711 }
Andy Whitcroft84afd992008-07-23 21:27:32 -0700712
713 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700714 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700715 INIT_LIST_HEAD(&resv_map->regions);
716
Mike Kravetz5e911372015-09-08 15:01:28 -0700717 resv_map->adds_in_progress = 0;
718
719 INIT_LIST_HEAD(&resv_map->region_cache);
720 list_add(&rg->link, &resv_map->region_cache);
721 resv_map->region_cache_count = 1;
722
Andy Whitcroft84afd992008-07-23 21:27:32 -0700723 return resv_map;
724}
725
Joonsoo Kim9119a412014-04-03 14:47:25 -0700726void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700727{
728 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
Mike Kravetz5e911372015-09-08 15:01:28 -0700729 struct list_head *head = &resv_map->region_cache;
730 struct file_region *rg, *trg;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700731
732 /* Clear out any active regions before we release the map. */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700733 region_del(resv_map, 0, LONG_MAX);
Mike Kravetz5e911372015-09-08 15:01:28 -0700734
735 /* ... and any entries left in the cache */
736 list_for_each_entry_safe(rg, trg, head, link) {
737 list_del(&rg->link);
738 kfree(rg);
739 }
740
741 VM_BUG_ON(resv_map->adds_in_progress);
742
Andy Whitcroft84afd992008-07-23 21:27:32 -0700743 kfree(resv_map);
744}
745
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700746static inline struct resv_map *inode_resv_map(struct inode *inode)
747{
748 return inode->i_mapping->private_data;
749}
750
Andy Whitcroft84afd992008-07-23 21:27:32 -0700751static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700752{
Sasha Levin81d1b092014-10-09 15:28:10 -0700753 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700754 if (vma->vm_flags & VM_MAYSHARE) {
755 struct address_space *mapping = vma->vm_file->f_mapping;
756 struct inode *inode = mapping->host;
757
758 return inode_resv_map(inode);
759
760 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700761 return (struct resv_map *)(get_vma_private_data(vma) &
762 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700763 }
Mel Gormana1e78772008-07-23 21:27:23 -0700764}
765
Andy Whitcroft84afd992008-07-23 21:27:32 -0700766static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700767{
Sasha Levin81d1b092014-10-09 15:28:10 -0700768 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
769 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700770
Andy Whitcroft84afd992008-07-23 21:27:32 -0700771 set_vma_private_data(vma, (get_vma_private_data(vma) &
772 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700773}
774
775static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
776{
Sasha Levin81d1b092014-10-09 15:28:10 -0700777 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
778 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700779
780 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700781}
782
783static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
784{
Sasha Levin81d1b092014-10-09 15:28:10 -0700785 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700786
787 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700788}
789
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700790/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700791void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
792{
Sasha Levin81d1b092014-10-09 15:28:10 -0700793 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -0700794 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700795 vma->vm_private_data = (void *)0;
796}
797
798/* Returns true if the VMA has associated reserve pages */
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700799static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700800{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700801 if (vma->vm_flags & VM_NORESERVE) {
802 /*
803 * This address is already reserved by other process(chg == 0),
804 * so, we should decrement reserved count. Without decrementing,
805 * reserve count remains after releasing inode, because this
806 * allocated page will go into page cache and is regarded as
807 * coming from reserved pool in releasing step. Currently, we
808 * don't have any other solution to deal with this situation
809 * properly, so add work-around here.
810 */
811 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700812 return true;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700813 else
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700814 return false;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700815 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700816
817 /* Shared mappings always use reserves */
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -0700818 if (vma->vm_flags & VM_MAYSHARE) {
819 /*
820 * We know VM_NORESERVE is not set. Therefore, there SHOULD
821 * be a region map for all pages. The only situation where
822 * there is no region map is if a hole was punched via
823 * fallocate. In this case, there really are no reverves to
824 * use. This situation is indicated if chg != 0.
825 */
826 if (chg)
827 return false;
828 else
829 return true;
830 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700831
832 /*
833 * Only the process that called mmap() has reserves for
834 * private mappings.
835 */
Mike Kravetz67961f92016-06-08 15:33:42 -0700836 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
837 /*
838 * Like the shared case above, a hole punch or truncate
839 * could have been performed on the private mapping.
840 * Examine the value of chg to determine if reserves
841 * actually exist or were previously consumed.
842 * Very Subtle - The value of chg comes from a previous
843 * call to vma_needs_reserves(). The reserve map for
844 * private mappings has different (opposite) semantics
845 * than that of shared mappings. vma_needs_reserves()
846 * has already taken this difference in semantics into
847 * account. Therefore, the meaning of chg is the same
848 * as in the shared case above. Code could easily be
849 * combined, but keeping it separate draws attention to
850 * subtle differences.
851 */
852 if (chg)
853 return false;
854 else
855 return true;
856 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700857
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700858 return false;
Mel Gormana1e78772008-07-23 21:27:23 -0700859}
860
Andi Kleena5516432008-07-23 21:27:41 -0700861static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862{
863 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700864 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700865 h->free_huge_pages++;
866 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867}
868
Anshuman Khandual94310cb2017-07-06 15:38:38 -0700869static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900870{
871 struct page *page;
872
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700873 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
Naoya Horiguchi243abd52017-07-10 15:47:32 -0700874 if (!PageHWPoison(page))
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700875 break;
876 /*
877 * if 'non-isolated free hugepage' not found on the list,
878 * the allocation fails.
879 */
880 if (&h->hugepage_freelists[nid] == &page->lru)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900881 return NULL;
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700882 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900883 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900884 h->free_huge_pages--;
885 h->free_huge_pages_node[nid]--;
886 return page;
887}
888
Anshuman Khandual94310cb2017-07-06 15:38:38 -0700889static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
890{
891 struct page *page;
892 int node;
893
894 if (nid != NUMA_NO_NODE)
895 return dequeue_huge_page_node_exact(h, nid);
896
897 for_each_online_node(node) {
898 page = dequeue_huge_page_node_exact(h, node);
899 if (page)
900 return page;
901 }
902 return NULL;
903}
904
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700905/* Movability of hugepages depends on migration support. */
906static inline gfp_t htlb_alloc_mask(struct hstate *h)
907{
Naoya Horiguchi100873d2014-06-04 16:10:56 -0700908 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700909 return GFP_HIGHUSER_MOVABLE;
910 else
911 return GFP_HIGHUSER;
912}
913
Andi Kleena5516432008-07-23 21:27:41 -0700914static struct page *dequeue_huge_page_vma(struct hstate *h,
915 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700916 unsigned long address, int avoid_reserve,
917 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700919 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700920 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700921 nodemask_t *nodemask;
Vlastimil Babka04ec6262017-07-06 15:40:03 -0700922 gfp_t gfp_mask;
923 int nid;
Miao Xiec0ff7452010-05-24 14:32:08 -0700924 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700925 struct zone *zone;
926 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700927 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928
Mel Gormana1e78772008-07-23 21:27:23 -0700929 /*
930 * A child process with MAP_PRIVATE mappings created by their parent
931 * have no page reserves. This check ensures that reservations are
932 * not "stolen". The child may still get SIGKILLed
933 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700934 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -0700935 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700936 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700937
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700938 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700939 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700940 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700941
Joonsoo Kim9966c4bb2013-09-11 14:20:50 -0700942retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -0700943 cpuset_mems_cookie = read_mems_allowed_begin();
Vlastimil Babka04ec6262017-07-06 15:40:03 -0700944 gfp_mask = htlb_alloc_mask(h);
945 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
946 zonelist = node_zonelist(nid, gfp_mask);
Joonsoo Kim9966c4bb2013-09-11 14:20:50 -0700947
Mel Gorman19770b32008-04-28 02:12:18 -0700948 for_each_zone_zonelist_nodemask(zone, z, zonelist,
949 MAX_NR_ZONES - 1, nodemask) {
Vlastimil Babka04ec6262017-07-06 15:40:03 -0700950 if (cpuset_zone_allowed(zone, gfp_mask)) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900951 page = dequeue_huge_page_node(h, zone_to_nid(zone));
952 if (page) {
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700953 if (avoid_reserve)
954 break;
955 if (!vma_has_reserves(vma, chg))
956 break;
957
Joonsoo Kim07443a82013-09-11 14:21:58 -0700958 SetPagePrivate(page);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700959 h->resv_huge_pages--;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900960 break;
961 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700962 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700964
965 mpol_cond_put(mpol);
Mel Gormand26914d2014-04-03 14:47:24 -0700966 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -0700967 goto retry_cpuset;
968 return page;
969
Miao Xiec0ff7452010-05-24 14:32:08 -0700970err:
Mel Gormancc9a6c82012-03-21 16:34:11 -0700971 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972}
973
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700974/*
975 * common helper functions for hstate_next_node_to_{alloc|free}.
976 * We may have allocated or freed a huge page based on a different
977 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
978 * be outside of *nodes_allowed. Ensure that we use an allowed
979 * node for alloc or free.
980 */
981static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
982{
Andrew Morton0edaf862016-05-19 17:10:58 -0700983 nid = next_node_in(nid, *nodes_allowed);
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700984 VM_BUG_ON(nid >= MAX_NUMNODES);
985
986 return nid;
987}
988
989static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
990{
991 if (!node_isset(nid, *nodes_allowed))
992 nid = next_node_allowed(nid, nodes_allowed);
993 return nid;
994}
995
996/*
997 * returns the previously saved node ["this node"] from which to
998 * allocate a persistent huge page for the pool and advance the
999 * next node from which to allocate, handling wrap at end of node
1000 * mask.
1001 */
1002static int hstate_next_node_to_alloc(struct hstate *h,
1003 nodemask_t *nodes_allowed)
1004{
1005 int nid;
1006
1007 VM_BUG_ON(!nodes_allowed);
1008
1009 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1010 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1011
1012 return nid;
1013}
1014
1015/*
1016 * helper for free_pool_huge_page() - return the previously saved
1017 * node ["this node"] from which to free a huge page. Advance the
1018 * next node id whether or not we find a free huge page to free so
1019 * that the next attempt to free addresses the next node.
1020 */
1021static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1022{
1023 int nid;
1024
1025 VM_BUG_ON(!nodes_allowed);
1026
1027 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1028 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1029
1030 return nid;
1031}
1032
1033#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1034 for (nr_nodes = nodes_weight(*mask); \
1035 nr_nodes > 0 && \
1036 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1037 nr_nodes--)
1038
1039#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1040 for (nr_nodes = nodes_weight(*mask); \
1041 nr_nodes > 0 && \
1042 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1043 nr_nodes--)
1044
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001045#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001046static void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001047 unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001048{
1049 int i;
1050 int nr_pages = 1 << order;
1051 struct page *p = page + 1;
1052
Gerald Schaeferc8cc7082016-06-24 14:50:04 -07001053 atomic_set(compound_mapcount_ptr(page), 0);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001054 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001055 clear_compound_head(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001056 set_page_refcounted(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001057 }
1058
1059 set_compound_order(page, 0);
1060 __ClearPageHead(page);
1061}
1062
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001063static void free_gigantic_page(struct page *page, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001064{
1065 free_contig_range(page_to_pfn(page), 1 << order);
1066}
1067
1068static int __alloc_gigantic_page(unsigned long start_pfn,
1069 unsigned long nr_pages)
1070{
1071 unsigned long end_pfn = start_pfn + nr_pages;
Lucas Stachca96b622017-02-24 14:58:37 -08001072 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1073 GFP_KERNEL);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001074}
1075
Joonsoo Kimf44b2dda82016-05-19 17:12:03 -07001076static bool pfn_range_valid_gigantic(struct zone *z,
1077 unsigned long start_pfn, unsigned long nr_pages)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001078{
1079 unsigned long i, end_pfn = start_pfn + nr_pages;
1080 struct page *page;
1081
1082 for (i = start_pfn; i < end_pfn; i++) {
1083 if (!pfn_valid(i))
1084 return false;
1085
1086 page = pfn_to_page(i);
1087
Joonsoo Kimf44b2dda82016-05-19 17:12:03 -07001088 if (page_zone(page) != z)
1089 return false;
1090
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001091 if (PageReserved(page))
1092 return false;
1093
1094 if (page_count(page) > 0)
1095 return false;
1096
1097 if (PageHuge(page))
1098 return false;
1099 }
1100
1101 return true;
1102}
1103
1104static bool zone_spans_last_pfn(const struct zone *zone,
1105 unsigned long start_pfn, unsigned long nr_pages)
1106{
1107 unsigned long last_pfn = start_pfn + nr_pages - 1;
1108 return zone_spans_pfn(zone, last_pfn);
1109}
1110
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001111static struct page *alloc_gigantic_page(int nid, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001112{
1113 unsigned long nr_pages = 1 << order;
1114 unsigned long ret, pfn, flags;
1115 struct zone *z;
1116
1117 z = NODE_DATA(nid)->node_zones;
1118 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1119 spin_lock_irqsave(&z->lock, flags);
1120
1121 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1122 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
Joonsoo Kimf44b2dda82016-05-19 17:12:03 -07001123 if (pfn_range_valid_gigantic(z, pfn, nr_pages)) {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001124 /*
1125 * We release the zone lock here because
1126 * alloc_contig_range() will also lock the zone
1127 * at some point. If there's an allocation
1128 * spinning on this lock, it may win the race
1129 * and cause alloc_contig_range() to fail...
1130 */
1131 spin_unlock_irqrestore(&z->lock, flags);
1132 ret = __alloc_gigantic_page(pfn, nr_pages);
1133 if (!ret)
1134 return pfn_to_page(pfn);
1135 spin_lock_irqsave(&z->lock, flags);
1136 }
1137 pfn += nr_pages;
1138 }
1139
1140 spin_unlock_irqrestore(&z->lock, flags);
1141 }
1142
1143 return NULL;
1144}
1145
1146static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001147static void prep_compound_gigantic_page(struct page *page, unsigned int order);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001148
1149static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1150{
1151 struct page *page;
1152
1153 page = alloc_gigantic_page(nid, huge_page_order(h));
1154 if (page) {
1155 prep_compound_gigantic_page(page, huge_page_order(h));
1156 prep_new_huge_page(h, page, nid);
1157 }
1158
1159 return page;
1160}
1161
1162static int alloc_fresh_gigantic_page(struct hstate *h,
1163 nodemask_t *nodes_allowed)
1164{
1165 struct page *page = NULL;
1166 int nr_nodes, node;
1167
1168 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1169 page = alloc_fresh_gigantic_page_node(h, node);
1170 if (page)
1171 return 1;
1172 }
1173
1174 return 0;
1175}
1176
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001177#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001178static inline bool gigantic_page_supported(void) { return false; }
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001179static inline void free_gigantic_page(struct page *page, unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001180static inline void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001181 unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001182static inline int alloc_fresh_gigantic_page(struct hstate *h,
1183 nodemask_t *nodes_allowed) { return 0; }
1184#endif
1185
Andi Kleena5516432008-07-23 21:27:41 -07001186static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -07001187{
1188 int i;
Andi Kleena5516432008-07-23 21:27:41 -07001189
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001190 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1191 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -08001192
Andi Kleena5516432008-07-23 21:27:41 -07001193 h->nr_huge_pages--;
1194 h->nr_huge_pages_node[page_to_nid(page)]--;
1195 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -07001196 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1197 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -07001198 1 << PG_active | 1 << PG_private |
1199 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -07001200 }
Sasha Levin309381fea2014-01-23 15:52:54 -08001201 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001202 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
Adam Litke6af2acb2007-10-16 01:26:16 -07001203 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001204 if (hstate_is_gigantic(h)) {
1205 destroy_compound_gigantic_page(page, huge_page_order(h));
1206 free_gigantic_page(page, huge_page_order(h));
1207 } else {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001208 __free_pages(page, huge_page_order(h));
1209 }
Adam Litke6af2acb2007-10-16 01:26:16 -07001210}
1211
Andi Kleene5ff2152008-07-23 21:27:42 -07001212struct hstate *size_to_hstate(unsigned long size)
1213{
1214 struct hstate *h;
1215
1216 for_each_hstate(h) {
1217 if (huge_page_size(h) == size)
1218 return h;
1219 }
1220 return NULL;
1221}
1222
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001223/*
1224 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1225 * to hstate->hugepage_activelist.)
1226 *
1227 * This function can be called for tail pages, but never returns true for them.
1228 */
1229bool page_huge_active(struct page *page)
1230{
1231 VM_BUG_ON_PAGE(!PageHuge(page), page);
1232 return PageHead(page) && PagePrivate(&page[1]);
1233}
1234
1235/* never called for tail page */
1236static void set_page_huge_active(struct page *page)
1237{
1238 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1239 SetPagePrivate(&page[1]);
1240}
1241
1242static void clear_page_huge_active(struct page *page)
1243{
1244 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1245 ClearPagePrivate(&page[1]);
1246}
1247
Atsushi Kumagai8f1d26d2014-07-30 16:08:39 -07001248void free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -08001249{
Andi Kleena5516432008-07-23 21:27:41 -07001250 /*
1251 * Can't pass hstate in here because it is called from the
1252 * compound page destructor.
1253 */
Andi Kleene5ff2152008-07-23 21:27:42 -07001254 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001255 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -07001256 struct hugepage_subpool *spool =
1257 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001258 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -08001259
Andy Whitcrofte5df70a2008-02-23 15:23:32 -08001260 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -04001261 page->mapping = NULL;
Mike Kravetzb4330af2016-02-05 15:36:38 -08001262 VM_BUG_ON_PAGE(page_count(page), page);
1263 VM_BUG_ON_PAGE(page_mapcount(page), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001264 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -07001265 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -08001266
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001267 /*
1268 * A return code of zero implies that the subpool will be under its
1269 * minimum size if the reservation is not restored after page is free.
1270 * Therefore, force restore_reserve operation.
1271 */
1272 if (hugepage_subpool_put_pages(spool, 1) == 0)
1273 restore_reserve = true;
1274
David Gibson27a85ef2006-03-22 00:08:56 -08001275 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001276 clear_page_huge_active(page);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001277 hugetlb_cgroup_uncharge_page(hstate_index(h),
1278 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001279 if (restore_reserve)
1280 h->resv_huge_pages++;
1281
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001282 if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001283 /* remove the page from active list */
1284 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -07001285 update_and_free_page(h, page);
1286 h->surplus_huge_pages--;
1287 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -07001288 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -07001289 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -07001290 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001291 }
David Gibson27a85ef2006-03-22 00:08:56 -08001292 spin_unlock(&hugetlb_lock);
1293}
1294
Andi Kleena5516432008-07-23 21:27:41 -07001295static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001296{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001297 INIT_LIST_HEAD(&page->lru);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001298 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001299 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001300 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -07001301 h->nr_huge_pages++;
1302 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001303 spin_unlock(&hugetlb_lock);
1304 put_page(page); /* free it into the hugepage allocator */
1305}
1306
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001307static void prep_compound_gigantic_page(struct page *page, unsigned int order)
Wu Fengguang20a03072009-06-16 15:32:22 -07001308{
1309 int i;
1310 int nr_pages = 1 << order;
1311 struct page *p = page + 1;
1312
1313 /* we rely on prep_new_huge_page to set the destructor */
1314 set_compound_order(page, order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001315 __ClearPageReserved(page);
Kirill A. Shutemovde09d312016-01-15 16:51:42 -08001316 __SetPageHead(page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001317 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001318 /*
1319 * For gigantic hugepages allocated through bootmem at
1320 * boot, it's safer to be consistent with the not-gigantic
1321 * hugepages and clear the PG_reserved bit from all tail pages
1322 * too. Otherwse drivers using get_user_pages() to access tail
1323 * pages may get the reference counting wrong if they see
1324 * PG_reserved set on a tail page (despite the head page not
1325 * having PG_reserved set). Enforcing this consistency between
1326 * head and tail pages allows drivers to optimize away a check
1327 * on the head page when they need know if put_page() is needed
1328 * after get_user_pages().
1329 */
1330 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -08001331 set_page_count(p, 0);
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001332 set_compound_head(p, page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001333 }
Mike Kravetzb4330af2016-02-05 15:36:38 -08001334 atomic_set(compound_mapcount_ptr(page), -1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001335}
1336
Andrew Morton77959122012-10-08 16:34:11 -07001337/*
1338 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1339 * transparent huge pages. See the PageTransHuge() documentation for more
1340 * details.
1341 */
Wu Fengguang20a03072009-06-16 15:32:22 -07001342int PageHuge(struct page *page)
1343{
Wu Fengguang20a03072009-06-16 15:32:22 -07001344 if (!PageCompound(page))
1345 return 0;
1346
1347 page = compound_head(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001348 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
Wu Fengguang20a03072009-06-16 15:32:22 -07001349}
Naoya Horiguchi43131e12010-05-28 09:29:22 +09001350EXPORT_SYMBOL_GPL(PageHuge);
1351
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001352/*
1353 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1354 * normal or transparent huge pages.
1355 */
1356int PageHeadHuge(struct page *page_head)
1357{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001358 if (!PageHead(page_head))
1359 return 0;
1360
Andrew Morton758f66a2014-01-21 15:48:57 -08001361 return get_compound_page_dtor(page_head) == free_huge_page;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001362}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001363
Zhang Yi13d60f42013-06-25 21:19:31 +08001364pgoff_t __basepage_index(struct page *page)
1365{
1366 struct page *page_head = compound_head(page);
1367 pgoff_t index = page_index(page_head);
1368 unsigned long compound_idx;
1369
1370 if (!PageHuge(page_head))
1371 return page_index(page);
1372
1373 if (compound_order(page_head) >= MAX_ORDER)
1374 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1375 else
1376 compound_idx = page - page_head;
1377
1378 return (index << compound_order(page_head)) + compound_idx;
1379}
1380
Andi Kleena5516432008-07-23 21:27:41 -07001381static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -07001384
Vlastimil Babka96db8002015-09-08 15:03:50 -07001385 page = __alloc_pages_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001386 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Nishanth Aravamudan551883a2008-04-29 00:58:26 -07001387 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -07001388 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 if (page) {
Andi Kleena5516432008-07-23 21:27:41 -07001390 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001392
1393 return page;
1394}
1395
Joonsoo Kimb2261022013-09-11 14:21:00 -07001396static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1397{
1398 struct page *page;
1399 int nr_nodes, node;
1400 int ret = 0;
1401
1402 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1403 page = alloc_fresh_huge_page_node(h, node);
1404 if (page) {
1405 ret = 1;
1406 break;
1407 }
1408 }
1409
1410 if (ret)
1411 count_vm_event(HTLB_BUDDY_PGALLOC);
1412 else
1413 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1414
1415 return ret;
1416}
1417
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001418/*
1419 * Free huge page from pool from next node to free.
1420 * Attempt to keep persistent huge pages more or less
1421 * balanced over allowed nodes.
1422 * Called with hugetlb_lock locked.
1423 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001424static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1425 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001426{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001427 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001428 int ret = 0;
1429
Joonsoo Kimb2261022013-09-11 14:21:00 -07001430 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001431 /*
1432 * If we're returning unused surplus pages, only examine
1433 * nodes with surplus pages.
1434 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001435 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1436 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001437 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001438 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001439 struct page, lru);
1440 list_del(&page->lru);
1441 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001442 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001443 if (acct_surplus) {
1444 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001445 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001446 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001447 update_and_free_page(h, page);
1448 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001449 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001450 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001451 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001452
1453 return ret;
1454}
1455
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001456/*
1457 * Dissolve a given free hugepage into free buddy pages. This function does
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001458 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1459 * number of free hugepages would be reduced below the number of reserved
1460 * hugepages.
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001461 */
Anshuman Khandualc3114a82017-07-10 15:47:41 -07001462int dissolve_free_huge_page(struct page *page)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001463{
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001464 int rc = 0;
1465
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001466 spin_lock(&hugetlb_lock);
1467 if (PageHuge(page) && !page_count(page)) {
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001468 struct page *head = compound_head(page);
1469 struct hstate *h = page_hstate(head);
1470 int nid = page_to_nid(head);
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001471 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1472 rc = -EBUSY;
1473 goto out;
1474 }
Anshuman Khandualc3114a82017-07-10 15:47:41 -07001475 /*
1476 * Move PageHWPoison flag from head page to the raw error page,
1477 * which makes any subpages rather than the error page reusable.
1478 */
1479 if (PageHWPoison(head) && page != head) {
1480 SetPageHWPoison(page);
1481 ClearPageHWPoison(head);
1482 }
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001483 list_del(&head->lru);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001484 h->free_huge_pages--;
1485 h->free_huge_pages_node[nid]--;
zhong jiangc1470b32016-08-11 15:32:55 -07001486 h->max_huge_pages--;
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001487 update_and_free_page(h, head);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001488 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001489out:
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001490 spin_unlock(&hugetlb_lock);
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001491 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001492}
1493
1494/*
1495 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1496 * make specified memory blocks removable from the system.
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001497 * Note that this will dissolve a free gigantic hugepage completely, if any
1498 * part of it lies within the given range.
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001499 * Also note that if dissolve_free_huge_page() returns with an error, all
1500 * free hugepages that were dissolved before that error are lost.
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001501 */
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001502int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001503{
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001504 unsigned long pfn;
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001505 struct page *page;
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001506 int rc = 0;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001507
Li Zhongd0177632014-08-06 16:07:56 -07001508 if (!hugepages_supported())
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001509 return rc;
Li Zhongd0177632014-08-06 16:07:56 -07001510
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001511 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1512 page = pfn_to_page(pfn);
1513 if (PageHuge(page) && !page_count(page)) {
1514 rc = dissolve_free_huge_page(page);
1515 if (rc)
1516 break;
1517 }
1518 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001519
1520 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001521}
1522
Dave Hansen099730d2015-11-05 18:50:17 -08001523/*
1524 * There are 3 ways this can get called:
1525 * 1. With vma+addr: we use the VMA's memory policy
1526 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1527 * page from any node, and let the buddy allocator itself figure
1528 * it out.
1529 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1530 * strictly from 'nid'
1531 */
1532static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1533 struct vm_area_struct *vma, unsigned long addr, int nid)
1534{
1535 int order = huge_page_order(h);
1536 gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1537 unsigned int cpuset_mems_cookie;
1538
1539 /*
1540 * We need a VMA to get a memory policy. If we do not
Dave Hansene0ec90e2015-11-05 18:50:20 -08001541 * have one, we use the 'nid' argument.
1542 *
1543 * The mempolicy stuff below has some non-inlined bits
1544 * and calls ->vm_ops. That makes it hard to optimize at
1545 * compile-time, even when NUMA is off and it does
1546 * nothing. This helps the compiler optimize it out.
Dave Hansen099730d2015-11-05 18:50:17 -08001547 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001548 if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
Dave Hansen099730d2015-11-05 18:50:17 -08001549 /*
1550 * If a specific node is requested, make sure to
1551 * get memory from there, but only when a node
1552 * is explicitly specified.
1553 */
1554 if (nid != NUMA_NO_NODE)
1555 gfp |= __GFP_THISNODE;
1556 /*
1557 * Make sure to call something that can handle
1558 * nid=NUMA_NO_NODE
1559 */
1560 return alloc_pages_node(nid, gfp, order);
1561 }
1562
1563 /*
1564 * OK, so we have a VMA. Fetch the mempolicy and try to
Dave Hansene0ec90e2015-11-05 18:50:20 -08001565 * allocate a huge page with it. We will only reach this
1566 * when CONFIG_NUMA=y.
Dave Hansen099730d2015-11-05 18:50:17 -08001567 */
1568 do {
1569 struct page *page;
1570 struct mempolicy *mpol;
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001571 int nid;
Dave Hansen099730d2015-11-05 18:50:17 -08001572 nodemask_t *nodemask;
1573
1574 cpuset_mems_cookie = read_mems_allowed_begin();
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001575 nid = huge_node(vma, addr, gfp, &mpol, &nodemask);
Dave Hansen099730d2015-11-05 18:50:17 -08001576 mpol_cond_put(mpol);
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001577 page = __alloc_pages_nodemask(gfp, order, nid, nodemask);
Dave Hansen099730d2015-11-05 18:50:17 -08001578 if (page)
1579 return page;
1580 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1581
1582 return NULL;
1583}
1584
1585/*
1586 * There are two ways to allocate a huge page:
1587 * 1. When you have a VMA and an address (like a fault)
1588 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1589 *
1590 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1591 * this case which signifies that the allocation should be done with
1592 * respect for the VMA's memory policy.
1593 *
1594 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1595 * implies that memory policies will not be taken in to account.
1596 */
1597static struct page *__alloc_buddy_huge_page(struct hstate *h,
1598 struct vm_area_struct *vma, unsigned long addr, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -07001599{
1600 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001601 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -07001602
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001603 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001604 return NULL;
1605
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001606 /*
Dave Hansen099730d2015-11-05 18:50:17 -08001607 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1608 * This makes sure the caller is picking _one_ of the modes with which
1609 * we can call this function, not both.
1610 */
1611 if (vma || (addr != -1)) {
Dave Hansene0ec90e2015-11-05 18:50:20 -08001612 VM_WARN_ON_ONCE(addr == -1);
1613 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
Dave Hansen099730d2015-11-05 18:50:17 -08001614 }
1615 /*
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001616 * Assume we will successfully allocate the surplus page to
1617 * prevent racing processes from causing the surplus to exceed
1618 * overcommit
1619 *
1620 * This however introduces a different race, where a process B
1621 * tries to grow the static hugepage pool while alloc_pages() is
1622 * called by process A. B will only examine the per-node
1623 * counters in determining if surplus huge pages can be
1624 * converted to normal huge pages in adjust_pool_surplus(). A
1625 * won't be able to increment the per-node counter, until the
1626 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1627 * no more huge pages can be converted from surplus to normal
1628 * state (and doesn't try to convert again). Thus, we have a
1629 * case where a surplus huge page exists, the pool is grown, and
1630 * the surplus huge page still exists after, even though it
1631 * should just have been converted to a normal huge page. This
1632 * does not leak memory, though, as the hugepage will be freed
1633 * once it is out of use. It also does not allow the counters to
1634 * go out of whack in adjust_pool_surplus() as we don't modify
1635 * the node values until we've gotten the hugepage and only the
1636 * per-node value is checked there.
1637 */
1638 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001639 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001640 spin_unlock(&hugetlb_lock);
1641 return NULL;
1642 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001643 h->nr_huge_pages++;
1644 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001645 }
1646 spin_unlock(&hugetlb_lock);
1647
Dave Hansen099730d2015-11-05 18:50:17 -08001648 page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001649
1650 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001651 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001652 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001653 r_nid = page_to_nid(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001654 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001655 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001656 /*
1657 * We incremented the global counters already
1658 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001659 h->nr_huge_pages_node[r_nid]++;
1660 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001661 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001662 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001663 h->nr_huge_pages--;
1664 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001665 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001666 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001667 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001668
1669 return page;
1670}
1671
Adam Litkee4e574b2007-10-16 01:26:19 -07001672/*
Dave Hansen099730d2015-11-05 18:50:17 -08001673 * Allocate a huge page from 'nid'. Note, 'nid' may be
1674 * NUMA_NO_NODE, which means that it may be allocated
1675 * anywhere.
1676 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001677static
Dave Hansen099730d2015-11-05 18:50:17 -08001678struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1679{
1680 unsigned long addr = -1;
1681
1682 return __alloc_buddy_huge_page(h, NULL, addr, nid);
1683}
1684
1685/*
1686 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1687 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001688static
Dave Hansen099730d2015-11-05 18:50:17 -08001689struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1690 struct vm_area_struct *vma, unsigned long addr)
1691{
1692 return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1693}
1694
1695/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001696 * This allocation function is useful in the context where vma is irrelevant.
1697 * E.g. soft-offlining uses this function because it only cares physical
1698 * address of error page.
1699 */
1700struct page *alloc_huge_page_node(struct hstate *h, int nid)
1701{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001702 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001703
1704 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001705 if (h->free_huge_pages - h->resv_huge_pages > 0)
1706 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001707 spin_unlock(&hugetlb_lock);
1708
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001709 if (!page)
Dave Hansen099730d2015-11-05 18:50:17 -08001710 page = __alloc_buddy_huge_page_no_mpol(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001711
1712 return page;
1713}
1714
1715/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001716 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001717 * of size 'delta'.
1718 */
Andi Kleena5516432008-07-23 21:27:41 -07001719static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001720{
1721 struct list_head surplus_list;
1722 struct page *page, *tmp;
1723 int ret, i;
1724 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001725 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001726
Andi Kleena5516432008-07-23 21:27:41 -07001727 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001728 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001729 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001730 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001731 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001732
1733 allocated = 0;
1734 INIT_LIST_HEAD(&surplus_list);
1735
1736 ret = -ENOMEM;
1737retry:
1738 spin_unlock(&hugetlb_lock);
1739 for (i = 0; i < needed; i++) {
Dave Hansen099730d2015-11-05 18:50:17 -08001740 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001741 if (!page) {
1742 alloc_ok = false;
1743 break;
1744 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001745 list_add(&page->lru, &surplus_list);
1746 }
Hillf Danton28073b02012-03-21 16:34:00 -07001747 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001748
1749 /*
1750 * After retaking hugetlb_lock, we need to recalculate 'needed'
1751 * because either resv_huge_pages or free_huge_pages may have changed.
1752 */
1753 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001754 needed = (h->resv_huge_pages + delta) -
1755 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001756 if (needed > 0) {
1757 if (alloc_ok)
1758 goto retry;
1759 /*
1760 * We were not able to allocate enough pages to
1761 * satisfy the entire reservation so we free what
1762 * we've allocated so far.
1763 */
1764 goto free;
1765 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001766 /*
1767 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001768 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001769 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001770 * allocator. Commit the entire reservation here to prevent another
1771 * process from stealing the pages as they are added to the pool but
1772 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001773 */
1774 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001775 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001776 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001777
Adam Litke19fc3f02008-04-28 02:12:20 -07001778 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001779 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001780 if ((--needed) < 0)
1781 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001782 /*
1783 * This page is now managed by the hugetlb allocator and has
1784 * no users -- drop the buddy allocator's reference.
1785 */
1786 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001787 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001788 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001789 }
Hillf Danton28073b02012-03-21 16:34:00 -07001790free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001791 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001792
1793 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001794 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1795 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001796 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001797
1798 return ret;
1799}
1800
1801/*
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001802 * This routine has two main purposes:
1803 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1804 * in unused_resv_pages. This corresponds to the prior adjustments made
1805 * to the associated reservation map.
1806 * 2) Free any unused surplus pages that may have been allocated to satisfy
1807 * the reservation. As many as unused_resv_pages may be freed.
1808 *
1809 * Called with hugetlb_lock held. However, the lock could be dropped (and
1810 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1811 * we must make sure nobody else can claim pages we are in the process of
1812 * freeing. Do this by ensuring resv_huge_page always is greater than the
1813 * number of huge pages we plan to free when dropping the lock.
Adam Litkee4e574b2007-10-16 01:26:19 -07001814 */
Andi Kleena5516432008-07-23 21:27:41 -07001815static void return_unused_surplus_pages(struct hstate *h,
1816 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001817{
Adam Litkee4e574b2007-10-16 01:26:19 -07001818 unsigned long nr_pages;
1819
Andi Kleenaa888a72008-07-23 21:27:47 -07001820 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001821 if (hstate_is_gigantic(h))
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001822 goto out;
Andi Kleenaa888a72008-07-23 21:27:47 -07001823
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001824 /*
1825 * Part (or even all) of the reservation could have been backed
1826 * by pre-allocated pages. Only free surplus pages.
1827 */
Andi Kleena5516432008-07-23 21:27:41 -07001828 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001829
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001830 /*
1831 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001832 * evenly across all nodes with memory. Iterate across these nodes
1833 * until we can no longer free unreserved surplus pages. This occurs
1834 * when the nodes with surplus pages have no free pages.
1835 * free_pool_huge_page() will balance the the freed pages across the
1836 * on-line nodes with memory and will handle the hstate accounting.
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001837 *
1838 * Note that we decrement resv_huge_pages as we free the pages. If
1839 * we drop the lock, resv_huge_pages will still be sufficiently large
1840 * to cover subsequent pages we may free.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001841 */
1842 while (nr_pages--) {
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001843 h->resv_huge_pages--;
1844 unused_resv_pages--;
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001845 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001846 goto out;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07001847 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001848 }
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001849
1850out:
1851 /* Fully uncommit the reservation */
1852 h->resv_huge_pages -= unused_resv_pages;
Adam Litkee4e574b2007-10-16 01:26:19 -07001853}
1854
Mike Kravetz5e911372015-09-08 15:01:28 -07001855
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001856/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001857 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
Mike Kravetz5e911372015-09-08 15:01:28 -07001858 * are used by the huge page allocation routines to manage reservations.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001859 *
1860 * vma_needs_reservation is called to determine if the huge page at addr
1861 * within the vma has an associated reservation. If a reservation is
1862 * needed, the value 1 is returned. The caller is then responsible for
1863 * managing the global reservation and subpool usage counts. After
1864 * the huge page has been allocated, vma_commit_reservation is called
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001865 * to add the page to the reservation map. If the page allocation fails,
1866 * the reservation must be ended instead of committed. vma_end_reservation
1867 * is called in such cases.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001868 *
1869 * In the normal case, vma_commit_reservation returns the same value
1870 * as the preceding vma_needs_reservation call. The only time this
1871 * is not the case is if a reserve map was changed between calls. It
1872 * is the responsibility of the caller to notice the difference and
1873 * take appropriate action.
Mike Kravetz96b96a92016-11-10 10:46:32 -08001874 *
1875 * vma_add_reservation is used in error paths where a reservation must
1876 * be restored when a newly allocated huge page must be freed. It is
1877 * to be called after calling vma_needs_reservation to determine if a
1878 * reservation exists.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001879 */
Mike Kravetz5e911372015-09-08 15:01:28 -07001880enum vma_resv_mode {
1881 VMA_NEEDS_RESV,
1882 VMA_COMMIT_RESV,
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001883 VMA_END_RESV,
Mike Kravetz96b96a92016-11-10 10:46:32 -08001884 VMA_ADD_RESV,
Mike Kravetz5e911372015-09-08 15:01:28 -07001885};
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001886static long __vma_reservation_common(struct hstate *h,
1887 struct vm_area_struct *vma, unsigned long addr,
Mike Kravetz5e911372015-09-08 15:01:28 -07001888 enum vma_resv_mode mode)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001889{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001890 struct resv_map *resv;
1891 pgoff_t idx;
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001892 long ret;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001893
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001894 resv = vma_resv_map(vma);
1895 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001896 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001897
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001898 idx = vma_hugecache_offset(h, vma, addr);
Mike Kravetz5e911372015-09-08 15:01:28 -07001899 switch (mode) {
1900 case VMA_NEEDS_RESV:
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001901 ret = region_chg(resv, idx, idx + 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07001902 break;
1903 case VMA_COMMIT_RESV:
1904 ret = region_add(resv, idx, idx + 1);
1905 break;
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001906 case VMA_END_RESV:
Mike Kravetz5e911372015-09-08 15:01:28 -07001907 region_abort(resv, idx, idx + 1);
1908 ret = 0;
1909 break;
Mike Kravetz96b96a92016-11-10 10:46:32 -08001910 case VMA_ADD_RESV:
1911 if (vma->vm_flags & VM_MAYSHARE)
1912 ret = region_add(resv, idx, idx + 1);
1913 else {
1914 region_abort(resv, idx, idx + 1);
1915 ret = region_del(resv, idx, idx + 1);
1916 }
1917 break;
Mike Kravetz5e911372015-09-08 15:01:28 -07001918 default:
1919 BUG();
1920 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07001921
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001922 if (vma->vm_flags & VM_MAYSHARE)
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001923 return ret;
Mike Kravetz67961f92016-06-08 15:33:42 -07001924 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1925 /*
1926 * In most cases, reserves always exist for private mappings.
1927 * However, a file associated with mapping could have been
1928 * hole punched or truncated after reserves were consumed.
1929 * As subsequent fault on such a range will not use reserves.
1930 * Subtle - The reserve map for private mappings has the
1931 * opposite meaning than that of shared mappings. If NO
1932 * entry is in the reserve map, it means a reservation exists.
1933 * If an entry exists in the reserve map, it means the
1934 * reservation has already been consumed. As a result, the
1935 * return value of this routine is the opposite of the
1936 * value returned from reserve map manipulation routines above.
1937 */
1938 if (ret)
1939 return 0;
1940 else
1941 return 1;
1942 }
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001943 else
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001944 return ret < 0 ? ret : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001945}
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001946
1947static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001948 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001949{
Mike Kravetz5e911372015-09-08 15:01:28 -07001950 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001951}
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001952
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001953static long vma_commit_reservation(struct hstate *h,
1954 struct vm_area_struct *vma, unsigned long addr)
1955{
Mike Kravetz5e911372015-09-08 15:01:28 -07001956 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1957}
1958
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001959static void vma_end_reservation(struct hstate *h,
Mike Kravetz5e911372015-09-08 15:01:28 -07001960 struct vm_area_struct *vma, unsigned long addr)
1961{
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001962 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001963}
1964
Mike Kravetz96b96a92016-11-10 10:46:32 -08001965static long vma_add_reservation(struct hstate *h,
1966 struct vm_area_struct *vma, unsigned long addr)
1967{
1968 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1969}
1970
1971/*
1972 * This routine is called to restore a reservation on error paths. In the
1973 * specific error paths, a huge page was allocated (via alloc_huge_page)
1974 * and is about to be freed. If a reservation for the page existed,
1975 * alloc_huge_page would have consumed the reservation and set PagePrivate
1976 * in the newly allocated page. When the page is freed via free_huge_page,
1977 * the global reservation count will be incremented if PagePrivate is set.
1978 * However, free_huge_page can not adjust the reserve map. Adjust the
1979 * reserve map here to be consistent with global reserve count adjustments
1980 * to be made by free_huge_page.
1981 */
1982static void restore_reserve_on_error(struct hstate *h,
1983 struct vm_area_struct *vma, unsigned long address,
1984 struct page *page)
1985{
1986 if (unlikely(PagePrivate(page))) {
1987 long rc = vma_needs_reservation(h, vma, address);
1988
1989 if (unlikely(rc < 0)) {
1990 /*
1991 * Rare out of memory condition in reserve map
1992 * manipulation. Clear PagePrivate so that
1993 * global reserve count will not be incremented
1994 * by free_huge_page. This will make it appear
1995 * as though the reservation for this page was
1996 * consumed. This may prevent the task from
1997 * faulting in the page at a later time. This
1998 * is better than inconsistent global huge page
1999 * accounting of reserve counts.
2000 */
2001 ClearPagePrivate(page);
2002 } else if (rc) {
2003 rc = vma_add_reservation(h, vma, address);
2004 if (unlikely(rc < 0))
2005 /*
2006 * See above comment about rare out of
2007 * memory condition.
2008 */
2009 ClearPagePrivate(page);
2010 } else
2011 vma_end_reservation(h, vma, address);
2012 }
2013}
2014
Mike Kravetz70c35472015-09-08 15:01:54 -07002015struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002016 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002017{
David Gibson90481622012-03-21 16:34:12 -07002018 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07002019 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08002020 struct page *page;
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002021 long map_chg, map_commit;
2022 long gbl_chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002023 int ret, idx;
2024 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08002025
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002026 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07002027 /*
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002028 * Examine the region/reserve map to determine if the process
2029 * has a reservation for the page to be allocated. A return
2030 * code of zero indicates a reservation exists (no change).
Mel Gormana1e78772008-07-23 21:27:23 -07002031 */
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002032 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2033 if (map_chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002034 return ERR_PTR(-ENOMEM);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002035
2036 /*
2037 * Processes that did not create the mapping will have no
2038 * reserves as indicated by the region/reserve map. Check
2039 * that the allocation will not exceed the subpool limit.
2040 * Allocations for MAP_NORESERVE mappings also need to be
2041 * checked against any subpool limit.
2042 */
2043 if (map_chg || avoid_reserve) {
2044 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2045 if (gbl_chg < 0) {
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002046 vma_end_reservation(h, vma, addr);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002047 return ERR_PTR(-ENOSPC);
Mike Kravetz5e911372015-09-08 15:01:28 -07002048 }
Mel Gormana1e78772008-07-23 21:27:23 -07002049
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002050 /*
2051 * Even though there was no reservation in the region/reserve
2052 * map, there could be reservations associated with the
2053 * subpool that can be used. This would be indicated if the
2054 * return value of hugepage_subpool_get_pages() is zero.
2055 * However, if avoid_reserve is specified we still avoid even
2056 * the subpool reservations.
2057 */
2058 if (avoid_reserve)
2059 gbl_chg = 1;
2060 }
2061
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002062 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002063 if (ret)
2064 goto out_subpool_put;
2065
Mel Gormana1e78772008-07-23 21:27:23 -07002066 spin_lock(&hugetlb_lock);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002067 /*
2068 * glb_chg is passed to indicate whether or not a page must be taken
2069 * from the global free pool (global change). gbl_chg == 0 indicates
2070 * a reservation exists for the allocation.
2071 */
2072 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002073 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07002074 spin_unlock(&hugetlb_lock);
Dave Hansen099730d2015-11-05 18:50:17 -08002075 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002076 if (!page)
2077 goto out_uncharge_cgroup;
Naoya Horiguchia88c7692015-12-11 13:40:24 -08002078 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2079 SetPagePrivate(page);
2080 h->resv_huge_pages--;
2081 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07002082 spin_lock(&hugetlb_lock);
2083 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002084 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07002085 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002086 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2087 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07002088
David Gibson90481622012-03-21 16:34:12 -07002089 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07002090
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002091 map_commit = vma_commit_reservation(h, vma, addr);
2092 if (unlikely(map_chg > map_commit)) {
Mike Kravetz33039672015-06-24 16:57:58 -07002093 /*
2094 * The page was added to the reservation map between
2095 * vma_needs_reservation and vma_commit_reservation.
2096 * This indicates a race with hugetlb_reserve_pages.
2097 * Adjust for the subpool count incremented above AND
2098 * in hugetlb_reserve_pages for the same page. Also,
2099 * the reservation count added in hugetlb_reserve_pages
2100 * no longer applies.
2101 */
2102 long rsv_adjust;
2103
2104 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2105 hugetlb_acct_memory(h, -rsv_adjust);
2106 }
Adam Litke90d8b7e2007-11-14 16:59:42 -08002107 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002108
2109out_uncharge_cgroup:
2110 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2111out_subpool_put:
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002112 if (map_chg || avoid_reserve)
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002113 hugepage_subpool_put_pages(spool, 1);
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002114 vma_end_reservation(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002115 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08002116}
2117
Naoya Horiguchi74060e42013-09-11 14:22:06 -07002118/*
2119 * alloc_huge_page()'s wrapper which simply returns the page if allocation
2120 * succeeds, otherwise NULL. This function is called from new_vma_page(),
2121 * where no ERR_VALUE is expected to be returned.
2122 */
2123struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
2124 unsigned long addr, int avoid_reserve)
2125{
2126 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
2127 if (IS_ERR(page))
2128 page = NULL;
2129 return page;
2130}
2131
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08002132int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07002133{
2134 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07002135 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07002136
Joonsoo Kimb2261022013-09-11 14:21:00 -07002137 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002138 void *addr;
2139
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002140 addr = memblock_virt_alloc_try_nid_nopanic(
2141 huge_page_size(h), huge_page_size(h),
2142 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07002143 if (addr) {
2144 /*
2145 * Use the beginning of the huge page to store the
2146 * huge_bootmem_page struct (until gather_bootmem
2147 * puts them into the mem_map).
2148 */
2149 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08002150 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07002151 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002152 }
2153 return 0;
2154
2155found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08002156 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07002157 /* Put them into a private list first because mem_map is not up yet */
2158 list_add(&m->list, &huge_boot_pages);
2159 m->hstate = h;
2160 return 1;
2161}
2162
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08002163static void __init prep_compound_huge_page(struct page *page,
2164 unsigned int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08002165{
2166 if (unlikely(order > (MAX_ORDER - 1)))
2167 prep_compound_gigantic_page(page, order);
2168 else
2169 prep_compound_page(page, order);
2170}
2171
Andi Kleenaa888a72008-07-23 21:27:47 -07002172/* Put bootmem huge pages into the standard lists after mem_map is up */
2173static void __init gather_bootmem_prealloc(void)
2174{
2175 struct huge_bootmem_page *m;
2176
2177 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002178 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07002179 struct page *page;
2180
2181#ifdef CONFIG_HIGHMEM
2182 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002183 memblock_free_late(__pa(m),
2184 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07002185#else
2186 page = virt_to_page(m);
2187#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07002188 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08002189 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07002190 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07002191 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07002192 /*
2193 * If we had gigantic hugepages allocated at boot time, we need
2194 * to restore the 'stolen' pages to totalram_pages in order to
2195 * fix confusing memory reports from free(1) and another
2196 * side-effects, like CommitLimit going negative.
2197 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002198 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07002199 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07002200 }
2201}
2202
Andi Kleen8faa8b02008-07-23 21:27:48 -07002203static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204{
2205 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206
Andi Kleene5ff2152008-07-23 21:27:42 -07002207 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002208 if (hstate_is_gigantic(h)) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002209 if (!alloc_bootmem_huge_page(h))
2210 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002211 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002212 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07002215 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07002216}
2217
2218static void __init hugetlb_init_hstates(void)
2219{
2220 struct hstate *h;
2221
2222 for_each_hstate(h) {
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002223 if (minimum_order > huge_page_order(h))
2224 minimum_order = huge_page_order(h);
2225
Andi Kleen8faa8b02008-07-23 21:27:48 -07002226 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002227 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07002228 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07002229 }
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002230 VM_BUG_ON(minimum_order == UINT_MAX);
Andi Kleene5ff2152008-07-23 21:27:42 -07002231}
2232
Andi Kleen4abd32d2008-07-23 21:27:49 -07002233static char * __init memfmt(char *buf, unsigned long n)
2234{
2235 if (n >= (1UL << 30))
2236 sprintf(buf, "%lu GB", n >> 30);
2237 else if (n >= (1UL << 20))
2238 sprintf(buf, "%lu MB", n >> 20);
2239 else
2240 sprintf(buf, "%lu KB", n >> 10);
2241 return buf;
2242}
2243
Andi Kleene5ff2152008-07-23 21:27:42 -07002244static void __init report_hugepages(void)
2245{
2246 struct hstate *h;
2247
2248 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07002249 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08002250 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07002251 memfmt(buf, huge_page_size(h)),
2252 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07002253 }
2254}
2255
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002257static void try_to_free_low(struct hstate *h, unsigned long count,
2258 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259{
Christoph Lameter4415cc82006-09-25 23:31:55 -07002260 int i;
2261
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002262 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07002263 return;
2264
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002265 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07002267 struct list_head *freel = &h->hugepage_freelists[i];
2268 list_for_each_entry_safe(page, next, freel, lru) {
2269 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07002270 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 if (PageHighMem(page))
2272 continue;
2273 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07002274 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07002275 h->free_huge_pages--;
2276 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277 }
2278 }
2279}
2280#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002281static inline void try_to_free_low(struct hstate *h, unsigned long count,
2282 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283{
2284}
2285#endif
2286
Wu Fengguang20a03072009-06-16 15:32:22 -07002287/*
2288 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2289 * balanced by operating on them in a round-robin fashion.
2290 * Returns 1 if an adjustment was made.
2291 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002292static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2293 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07002294{
Joonsoo Kimb2261022013-09-11 14:21:00 -07002295 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07002296
2297 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07002298
Joonsoo Kimb2261022013-09-11 14:21:00 -07002299 if (delta < 0) {
2300 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2301 if (h->surplus_huge_pages_node[node])
2302 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002303 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002304 } else {
2305 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2306 if (h->surplus_huge_pages_node[node] <
2307 h->nr_huge_pages_node[node])
2308 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002309 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002310 }
2311 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07002312
Joonsoo Kimb2261022013-09-11 14:21:00 -07002313found:
2314 h->surplus_huge_pages += delta;
2315 h->surplus_huge_pages_node[node] += delta;
2316 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07002317}
2318
Andi Kleena5516432008-07-23 21:27:41 -07002319#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002320static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2321 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322{
Adam Litke7893d1d2007-10-16 01:26:18 -07002323 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002325 if (hstate_is_gigantic(h) && !gigantic_page_supported())
Andi Kleenaa888a72008-07-23 21:27:47 -07002326 return h->max_huge_pages;
2327
Adam Litke7893d1d2007-10-16 01:26:18 -07002328 /*
2329 * Increase the pool size
2330 * First take pages out of surplus state. Then make up the
2331 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002332 *
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002333 * We might race with __alloc_buddy_huge_page() here and be unable
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002334 * to convert a surplus huge page to a normal huge page. That is
2335 * not critical, though, it just means the overall size of the
2336 * pool might be one hugepage larger than it needs to be, but
2337 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07002338 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002340 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002341 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002342 break;
2343 }
2344
Andi Kleena5516432008-07-23 21:27:41 -07002345 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07002346 /*
2347 * If this allocation races such that we no longer need the
2348 * page, free_huge_page will handle it by freeing the page
2349 * and reducing the surplus.
2350 */
2351 spin_unlock(&hugetlb_lock);
Jia He649920c2016-08-02 14:02:31 -07002352
2353 /* yield cpu to avoid soft lockup */
2354 cond_resched();
2355
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002356 if (hstate_is_gigantic(h))
2357 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2358 else
2359 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07002360 spin_lock(&hugetlb_lock);
2361 if (!ret)
2362 goto out;
2363
Mel Gorman536240f22009-12-14 17:59:56 -08002364 /* Bail for signals. Probably ctrl-c from user */
2365 if (signal_pending(current))
2366 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07002367 }
Adam Litke7893d1d2007-10-16 01:26:18 -07002368
2369 /*
2370 * Decrease the pool size
2371 * First return free pages to the buddy allocator (being careful
2372 * to keep enough around to satisfy reservations). Then place
2373 * pages into surplus state as needed so the pool will shrink
2374 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002375 *
2376 * By placing pages into the surplus state independent of the
2377 * overcommit value, we are allowing the surplus pool size to
2378 * exceed overcommit. There are few sane options here. Since
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002379 * __alloc_buddy_huge_page() is checking the global counter,
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002380 * though, we'll note that we're not allowed to exceed surplus
2381 * and won't grow the pool anywhere else. Not until one of the
2382 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07002383 */
Andi Kleena5516432008-07-23 21:27:41 -07002384 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07002385 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002386 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07002387 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002388 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07002390 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 }
Andi Kleena5516432008-07-23 21:27:41 -07002392 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002393 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002394 break;
2395 }
2396out:
Andi Kleena5516432008-07-23 21:27:41 -07002397 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07002399 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400}
2401
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002402#define HSTATE_ATTR_RO(_name) \
2403 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2404
2405#define HSTATE_ATTR(_name) \
2406 static struct kobj_attribute _name##_attr = \
2407 __ATTR(_name, 0644, _name##_show, _name##_store)
2408
2409static struct kobject *hugepages_kobj;
2410static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2411
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002412static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2413
2414static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002415{
2416 int i;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002417
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002418 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002419 if (hstate_kobjs[i] == kobj) {
2420 if (nidp)
2421 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002422 return &hstates[i];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002423 }
2424
2425 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002426}
2427
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002428static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002429 struct kobj_attribute *attr, char *buf)
2430{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002431 struct hstate *h;
2432 unsigned long nr_huge_pages;
2433 int nid;
2434
2435 h = kobj_to_hstate(kobj, &nid);
2436 if (nid == NUMA_NO_NODE)
2437 nr_huge_pages = h->nr_huge_pages;
2438 else
2439 nr_huge_pages = h->nr_huge_pages_node[nid];
2440
2441 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002442}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002443
David Rientjes238d3c12014-08-06 16:06:51 -07002444static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2445 struct hstate *h, int nid,
2446 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002447{
2448 int err;
David Rientjesbad44b52009-12-14 17:58:38 -08002449 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002450
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002451 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
Eric B Munsonadbe8722011-01-13 15:47:27 -08002452 err = -EINVAL;
2453 goto out;
2454 }
2455
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002456 if (nid == NUMA_NO_NODE) {
2457 /*
2458 * global hstate attribute
2459 */
2460 if (!(obey_mempolicy &&
2461 init_nodemask_of_mempolicy(nodes_allowed))) {
2462 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002463 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002464 }
2465 } else if (nodes_allowed) {
2466 /*
2467 * per node hstate attribute: adjust count to global,
2468 * but restrict alloc/free to the specified node.
2469 */
2470 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2471 init_nodemask_of_node(nodes_allowed, nid);
2472 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002473 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002474
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002475 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002476
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002477 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002478 NODEMASK_FREE(nodes_allowed);
2479
2480 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08002481out:
2482 NODEMASK_FREE(nodes_allowed);
2483 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002484}
2485
David Rientjes238d3c12014-08-06 16:06:51 -07002486static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2487 struct kobject *kobj, const char *buf,
2488 size_t len)
2489{
2490 struct hstate *h;
2491 unsigned long count;
2492 int nid;
2493 int err;
2494
2495 err = kstrtoul(buf, 10, &count);
2496 if (err)
2497 return err;
2498
2499 h = kobj_to_hstate(kobj, &nid);
2500 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2501}
2502
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002503static ssize_t nr_hugepages_show(struct kobject *kobj,
2504 struct kobj_attribute *attr, char *buf)
2505{
2506 return nr_hugepages_show_common(kobj, attr, buf);
2507}
2508
2509static ssize_t nr_hugepages_store(struct kobject *kobj,
2510 struct kobj_attribute *attr, const char *buf, size_t len)
2511{
David Rientjes238d3c12014-08-06 16:06:51 -07002512 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002513}
2514HSTATE_ATTR(nr_hugepages);
2515
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002516#ifdef CONFIG_NUMA
2517
2518/*
2519 * hstate attribute for optionally mempolicy-based constraint on persistent
2520 * huge page alloc/free.
2521 */
2522static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2523 struct kobj_attribute *attr, char *buf)
2524{
2525 return nr_hugepages_show_common(kobj, attr, buf);
2526}
2527
2528static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2529 struct kobj_attribute *attr, const char *buf, size_t len)
2530{
David Rientjes238d3c12014-08-06 16:06:51 -07002531 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002532}
2533HSTATE_ATTR(nr_hugepages_mempolicy);
2534#endif
2535
2536
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002537static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2538 struct kobj_attribute *attr, char *buf)
2539{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002540 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002541 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2542}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002543
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002544static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2545 struct kobj_attribute *attr, const char *buf, size_t count)
2546{
2547 int err;
2548 unsigned long input;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002549 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002550
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002551 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002552 return -EINVAL;
2553
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002554 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002555 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08002556 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002557
2558 spin_lock(&hugetlb_lock);
2559 h->nr_overcommit_huge_pages = input;
2560 spin_unlock(&hugetlb_lock);
2561
2562 return count;
2563}
2564HSTATE_ATTR(nr_overcommit_hugepages);
2565
2566static ssize_t free_hugepages_show(struct kobject *kobj,
2567 struct kobj_attribute *attr, char *buf)
2568{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002569 struct hstate *h;
2570 unsigned long free_huge_pages;
2571 int nid;
2572
2573 h = kobj_to_hstate(kobj, &nid);
2574 if (nid == NUMA_NO_NODE)
2575 free_huge_pages = h->free_huge_pages;
2576 else
2577 free_huge_pages = h->free_huge_pages_node[nid];
2578
2579 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002580}
2581HSTATE_ATTR_RO(free_hugepages);
2582
2583static ssize_t resv_hugepages_show(struct kobject *kobj,
2584 struct kobj_attribute *attr, char *buf)
2585{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002586 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002587 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2588}
2589HSTATE_ATTR_RO(resv_hugepages);
2590
2591static ssize_t surplus_hugepages_show(struct kobject *kobj,
2592 struct kobj_attribute *attr, char *buf)
2593{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002594 struct hstate *h;
2595 unsigned long surplus_huge_pages;
2596 int nid;
2597
2598 h = kobj_to_hstate(kobj, &nid);
2599 if (nid == NUMA_NO_NODE)
2600 surplus_huge_pages = h->surplus_huge_pages;
2601 else
2602 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2603
2604 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002605}
2606HSTATE_ATTR_RO(surplus_hugepages);
2607
2608static struct attribute *hstate_attrs[] = {
2609 &nr_hugepages_attr.attr,
2610 &nr_overcommit_hugepages_attr.attr,
2611 &free_hugepages_attr.attr,
2612 &resv_hugepages_attr.attr,
2613 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002614#ifdef CONFIG_NUMA
2615 &nr_hugepages_mempolicy_attr.attr,
2616#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002617 NULL,
2618};
2619
2620static struct attribute_group hstate_attr_group = {
2621 .attrs = hstate_attrs,
2622};
2623
Jeff Mahoney094e9532010-02-02 13:44:14 -08002624static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2625 struct kobject **hstate_kobjs,
2626 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002627{
2628 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002629 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002630
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002631 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2632 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002633 return -ENOMEM;
2634
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002635 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002636 if (retval)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002637 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002638
2639 return retval;
2640}
2641
2642static void __init hugetlb_sysfs_init(void)
2643{
2644 struct hstate *h;
2645 int err;
2646
2647 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2648 if (!hugepages_kobj)
2649 return;
2650
2651 for_each_hstate(h) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002652 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2653 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002654 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08002655 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002656 }
2657}
2658
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002659#ifdef CONFIG_NUMA
2660
2661/*
2662 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08002663 * with node devices in node_devices[] using a parallel array. The array
2664 * index of a node device or _hstate == node id.
2665 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002666 * the base kernel, on the hugetlb module.
2667 */
2668struct node_hstate {
2669 struct kobject *hugepages_kobj;
2670 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2671};
Alexander Kuleshovb4e289a2015-11-05 18:50:14 -08002672static struct node_hstate node_hstates[MAX_NUMNODES];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002673
2674/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002675 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002676 */
2677static struct attribute *per_node_hstate_attrs[] = {
2678 &nr_hugepages_attr.attr,
2679 &free_hugepages_attr.attr,
2680 &surplus_hugepages_attr.attr,
2681 NULL,
2682};
2683
2684static struct attribute_group per_node_hstate_attr_group = {
2685 .attrs = per_node_hstate_attrs,
2686};
2687
2688/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002689 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002690 * Returns node id via non-NULL nidp.
2691 */
2692static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2693{
2694 int nid;
2695
2696 for (nid = 0; nid < nr_node_ids; nid++) {
2697 struct node_hstate *nhs = &node_hstates[nid];
2698 int i;
2699 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2700 if (nhs->hstate_kobjs[i] == kobj) {
2701 if (nidp)
2702 *nidp = nid;
2703 return &hstates[i];
2704 }
2705 }
2706
2707 BUG();
2708 return NULL;
2709}
2710
2711/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002712 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002713 * No-op if no hstate attributes attached.
2714 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002715static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002716{
2717 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002718 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002719
2720 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002721 return; /* no hstate attributes */
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002722
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002723 for_each_hstate(h) {
2724 int idx = hstate_index(h);
2725 if (nhs->hstate_kobjs[idx]) {
2726 kobject_put(nhs->hstate_kobjs[idx]);
2727 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002728 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002729 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002730
2731 kobject_put(nhs->hugepages_kobj);
2732 nhs->hugepages_kobj = NULL;
2733}
2734
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002735
2736/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002737 * Register hstate attributes for a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002738 * No-op if attributes already registered.
2739 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002740static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002741{
2742 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002743 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002744 int err;
2745
2746 if (nhs->hugepages_kobj)
2747 return; /* already allocated */
2748
2749 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08002750 &node->dev.kobj);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002751 if (!nhs->hugepages_kobj)
2752 return;
2753
2754 for_each_hstate(h) {
2755 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2756 nhs->hstate_kobjs,
2757 &per_node_hstate_attr_group);
2758 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002759 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2760 h->name, node->dev.id);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002761 hugetlb_unregister_node(node);
2762 break;
2763 }
2764 }
2765}
2766
2767/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002768 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08002769 * devices of nodes that have memory. All on-line nodes should have
2770 * registered their associated device by this time.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002771 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08002772static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002773{
2774 int nid;
2775
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002776 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08002777 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08002778 if (node->dev.id == nid)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002779 hugetlb_register_node(node);
2780 }
2781
2782 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002783 * Let the node device driver know we're here so it can
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002784 * [un]register hstate attributes on node hotplug.
2785 */
2786 register_hugetlbfs_with_node(hugetlb_register_node,
2787 hugetlb_unregister_node);
2788}
2789#else /* !CONFIG_NUMA */
2790
2791static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2792{
2793 BUG();
2794 if (nidp)
2795 *nidp = -1;
2796 return NULL;
2797}
2798
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002799static void hugetlb_register_all_nodes(void) { }
2800
2801#endif
2802
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002803static int __init hugetlb_init(void)
2804{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002805 int i;
2806
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002807 if (!hugepages_supported())
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07002808 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002809
Nick Piggine11bfbf2008-07-23 21:27:52 -07002810 if (!size_to_hstate(default_hstate_size)) {
2811 default_hstate_size = HPAGE_SIZE;
2812 if (!size_to_hstate(default_hstate_size))
2813 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002814 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002815 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Vaishali Thakkarf8b74812016-02-17 13:11:26 -08002816 if (default_hstate_max_huge_pages) {
2817 if (!default_hstate.max_huge_pages)
2818 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2819 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002820
2821 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07002822 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002823 report_hugepages();
2824
2825 hugetlb_sysfs_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002826 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002827 hugetlb_cgroup_file_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002828
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002829#ifdef CONFIG_SMP
2830 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2831#else
2832 num_fault_mutexes = 1;
2833#endif
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002834 hugetlb_fault_mutex_table =
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002835 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002836 BUG_ON(!hugetlb_fault_mutex_table);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002837
2838 for (i = 0; i < num_fault_mutexes; i++)
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002839 mutex_init(&hugetlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002840 return 0;
2841}
Paul Gortmaker3e89e1c2016-01-14 15:21:52 -08002842subsys_initcall(hugetlb_init);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002843
2844/* Should be called on processing a hugepagesz=... option */
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07002845void __init hugetlb_bad_size(void)
2846{
2847 parsed_valid_hugepagesz = false;
2848}
2849
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08002850void __init hugetlb_add_hstate(unsigned int order)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002851{
2852 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002853 unsigned long i;
2854
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002855 if (size_to_hstate(PAGE_SIZE << order)) {
Joe Perches598d8092016-03-17 14:19:44 -07002856 pr_warn("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002857 return;
2858 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002859 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002860 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002861 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002862 h->order = order;
2863 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002864 h->nr_huge_pages = 0;
2865 h->free_huge_pages = 0;
2866 for (i = 0; i < MAX_NUMNODES; ++i)
2867 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002868 INIT_LIST_HEAD(&h->hugepage_activelist);
Andrew Morton54f18d32016-05-19 17:11:40 -07002869 h->next_nid_to_alloc = first_memory_node;
2870 h->next_nid_to_free = first_memory_node;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002871 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2872 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002873
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002874 parsed_hstate = h;
2875}
2876
Nick Piggine11bfbf2008-07-23 21:27:52 -07002877static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002878{
2879 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002880 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002881
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07002882 if (!parsed_valid_hugepagesz) {
2883 pr_warn("hugepages = %s preceded by "
2884 "an unsupported hugepagesz, ignoring\n", s);
2885 parsed_valid_hugepagesz = true;
2886 return 1;
2887 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002888 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002889 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002890 * so this hugepages= parameter goes to the "default hstate".
2891 */
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07002892 else if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002893 mhp = &default_hstate_max_huge_pages;
2894 else
2895 mhp = &parsed_hstate->max_huge_pages;
2896
Andi Kleen8faa8b02008-07-23 21:27:48 -07002897 if (mhp == last_mhp) {
Joe Perches598d8092016-03-17 14:19:44 -07002898 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002899 return 1;
2900 }
2901
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002902 if (sscanf(s, "%lu", mhp) <= 0)
2903 *mhp = 0;
2904
Andi Kleen8faa8b02008-07-23 21:27:48 -07002905 /*
2906 * Global state is always initialized later in hugetlb_init.
2907 * But we need to allocate >= MAX_ORDER hstates here early to still
2908 * use the bootmem allocator.
2909 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002910 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002911 hugetlb_hstate_alloc_pages(parsed_hstate);
2912
2913 last_mhp = mhp;
2914
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002915 return 1;
2916}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002917__setup("hugepages=", hugetlb_nrpages_setup);
2918
2919static int __init hugetlb_default_setup(char *s)
2920{
2921 default_hstate_size = memparse(s, &s);
2922 return 1;
2923}
2924__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002925
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002926static unsigned int cpuset_mems_nr(unsigned int *array)
2927{
2928 int node;
2929 unsigned int nr = 0;
2930
2931 for_each_node_mask(node, cpuset_current_mems_allowed)
2932 nr += array[node];
2933
2934 return nr;
2935}
2936
2937#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002938static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2939 struct ctl_table *table, int write,
2940 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941{
Andi Kleene5ff2152008-07-23 21:27:42 -07002942 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07002943 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08002944 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002945
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002946 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08002947 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002948
Andi Kleene5ff2152008-07-23 21:27:42 -07002949 table->data = &tmp;
2950 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002951 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2952 if (ret)
2953 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002954
David Rientjes238d3c12014-08-06 16:06:51 -07002955 if (write)
2956 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2957 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08002958out:
2959 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960}
Mel Gorman396faf02007-07-17 04:03:13 -07002961
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002962int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2963 void __user *buffer, size_t *length, loff_t *ppos)
2964{
2965
2966 return hugetlb_sysctl_handler_common(false, table, write,
2967 buffer, length, ppos);
2968}
2969
2970#ifdef CONFIG_NUMA
2971int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2972 void __user *buffer, size_t *length, loff_t *ppos)
2973{
2974 return hugetlb_sysctl_handler_common(true, table, write,
2975 buffer, length, ppos);
2976}
2977#endif /* CONFIG_NUMA */
2978
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002979int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002980 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002981 size_t *length, loff_t *ppos)
2982{
Andi Kleena5516432008-07-23 21:27:41 -07002983 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002984 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002985 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002986
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002987 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08002988 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002989
Petr Holasekc033a932011-03-22 16:33:05 -07002990 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002991
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002992 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002993 return -EINVAL;
2994
Andi Kleene5ff2152008-07-23 21:27:42 -07002995 table->data = &tmp;
2996 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002997 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2998 if (ret)
2999 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07003000
3001 if (write) {
3002 spin_lock(&hugetlb_lock);
3003 h->nr_overcommit_huge_pages = tmp;
3004 spin_unlock(&hugetlb_lock);
3005 }
Michal Hocko08d4a242011-01-13 15:47:26 -08003006out:
3007 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003008}
3009
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010#endif /* CONFIG_SYSCTL */
3011
Alexey Dobriyane1759c22008-10-15 23:50:22 +04003012void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013{
Andi Kleena5516432008-07-23 21:27:41 -07003014 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003015 if (!hugepages_supported())
3016 return;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04003017 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003018 "HugePages_Total: %5lu\n"
3019 "HugePages_Free: %5lu\n"
3020 "HugePages_Rsvd: %5lu\n"
3021 "HugePages_Surp: %5lu\n"
3022 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07003023 h->nr_huge_pages,
3024 h->free_huge_pages,
3025 h->resv_huge_pages,
3026 h->surplus_huge_pages,
3027 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028}
3029
3030int hugetlb_report_node_meminfo(int nid, char *buf)
3031{
Andi Kleena5516432008-07-23 21:27:41 -07003032 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003033 if (!hugepages_supported())
3034 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 return sprintf(buf,
3036 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07003037 "Node %d HugePages_Free: %5u\n"
3038 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07003039 nid, h->nr_huge_pages_node[nid],
3040 nid, h->free_huge_pages_node[nid],
3041 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042}
3043
David Rientjes949f7ec2013-04-29 15:07:48 -07003044void hugetlb_show_meminfo(void)
3045{
3046 struct hstate *h;
3047 int nid;
3048
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003049 if (!hugepages_supported())
3050 return;
3051
David Rientjes949f7ec2013-04-29 15:07:48 -07003052 for_each_node_state(nid, N_MEMORY)
3053 for_each_hstate(h)
3054 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3055 nid,
3056 h->nr_huge_pages_node[nid],
3057 h->free_huge_pages_node[nid],
3058 h->surplus_huge_pages_node[nid],
3059 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3060}
3061
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003062void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3063{
3064 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3065 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3066}
3067
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3069unsigned long hugetlb_total_pages(void)
3070{
Wanpeng Lid0028582013-03-22 15:04:40 -07003071 struct hstate *h;
3072 unsigned long nr_total_pages = 0;
3073
3074 for_each_hstate(h)
3075 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3076 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078
Andi Kleena5516432008-07-23 21:27:41 -07003079static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003080{
3081 int ret = -ENOMEM;
3082
3083 spin_lock(&hugetlb_lock);
3084 /*
3085 * When cpuset is configured, it breaks the strict hugetlb page
3086 * reservation as the accounting is done on a global variable. Such
3087 * reservation is completely rubbish in the presence of cpuset because
3088 * the reservation is not checked against page availability for the
3089 * current cpuset. Application can still potentially OOM'ed by kernel
3090 * with lack of free htlb page in cpuset that the task is in.
3091 * Attempt to enforce strict accounting with cpuset is almost
3092 * impossible (or too ugly) because cpuset is too fluid that
3093 * task or memory node can be dynamically moved between cpusets.
3094 *
3095 * The change of semantics for shared hugetlb mapping with cpuset is
3096 * undesirable. However, in order to preserve some of the semantics,
3097 * we fall back to check against current free page availability as
3098 * a best attempt and hopefully to minimize the impact of changing
3099 * semantics that cpuset has.
3100 */
3101 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07003102 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003103 goto out;
3104
Andi Kleena5516432008-07-23 21:27:41 -07003105 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3106 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003107 goto out;
3108 }
3109 }
3110
3111 ret = 0;
3112 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07003113 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003114
3115out:
3116 spin_unlock(&hugetlb_lock);
3117 return ret;
3118}
3119
Andy Whitcroft84afd992008-07-23 21:27:32 -07003120static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3121{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003122 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003123
3124 /*
3125 * This new VMA should share its siblings reservation map if present.
3126 * The VMA will only ever have a valid reservation map pointer where
3127 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003128 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07003129 * after this open call completes. It is therefore safe to take a
3130 * new reference here without additional locking.
3131 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003132 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003133 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003134}
3135
Mel Gormana1e78772008-07-23 21:27:23 -07003136static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3137{
Andi Kleena5516432008-07-23 21:27:41 -07003138 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003139 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07003140 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003141 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003142 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003143
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003144 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3145 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003146
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003147 start = vma_hugecache_offset(h, vma, vma->vm_start);
3148 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003149
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003150 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003151
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003152 kref_put(&resv->refs, resv_map_release);
3153
3154 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003155 /*
3156 * Decrement reserve counts. The global reserve count may be
3157 * adjusted if the subpool has a minimum size.
3158 */
3159 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3160 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003161 }
Mel Gormana1e78772008-07-23 21:27:23 -07003162}
3163
Linus Torvalds1da177e2005-04-16 15:20:36 -07003164/*
3165 * We cannot handle pagefaults against hugetlb pages at all. They cause
3166 * handle_mm_fault() to try to instantiate regular-sized pages in the
3167 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3168 * this far.
3169 */
Dave Jiang11bac802017-02-24 14:56:41 -08003170static int hugetlb_vm_op_fault(struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003171{
3172 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07003173 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174}
3175
Alexey Dobriyanf0f37e2f2009-09-27 22:29:37 +04003176const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07003177 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07003178 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07003179 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180};
3181
David Gibson1e8f8892006-01-06 00:10:44 -08003182static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3183 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07003184{
3185 pte_t entry;
3186
David Gibson1e8f8892006-01-06 00:10:44 -08003187 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003188 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3189 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07003190 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003191 entry = huge_pte_wrprotect(mk_huge_pte(page,
3192 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07003193 }
3194 entry = pte_mkyoung(entry);
3195 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04003196 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07003197
3198 return entry;
3199}
3200
David Gibson1e8f8892006-01-06 00:10:44 -08003201static void set_huge_ptep_writable(struct vm_area_struct *vma,
3202 unsigned long address, pte_t *ptep)
3203{
3204 pte_t entry;
3205
Gerald Schaefer106c9922013-04-29 15:07:23 -07003206 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07003207 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00003208 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003209}
3210
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003211bool is_hugetlb_entry_migration(pte_t pte)
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003212{
3213 swp_entry_t swp;
3214
3215 if (huge_pte_none(pte) || pte_present(pte))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003216 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003217 swp = pte_to_swp_entry(pte);
3218 if (non_swap_entry(swp) && is_migration_entry(swp))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003219 return true;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003220 else
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003221 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003222}
3223
3224static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3225{
3226 swp_entry_t swp;
3227
3228 if (huge_pte_none(pte) || pte_present(pte))
3229 return 0;
3230 swp = pte_to_swp_entry(pte);
3231 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3232 return 1;
3233 else
3234 return 0;
3235}
David Gibson1e8f8892006-01-06 00:10:44 -08003236
David Gibson63551ae2005-06-21 17:14:44 -07003237int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3238 struct vm_area_struct *vma)
3239{
3240 pte_t *src_pte, *dst_pte, entry;
3241 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07003242 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08003243 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07003244 struct hstate *h = hstate_vma(vma);
3245 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003246 unsigned long mmun_start; /* For mmu_notifiers */
3247 unsigned long mmun_end; /* For mmu_notifiers */
3248 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003249
3250 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07003251
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003252 mmun_start = vma->vm_start;
3253 mmun_end = vma->vm_end;
3254 if (cow)
3255 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3256
Andi Kleena5516432008-07-23 21:27:41 -07003257 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003258 spinlock_t *src_ptl, *dst_ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07003259 src_pte = huge_pte_offset(src, addr, sz);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003260 if (!src_pte)
3261 continue;
Andi Kleena5516432008-07-23 21:27:41 -07003262 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003263 if (!dst_pte) {
3264 ret = -ENOMEM;
3265 break;
3266 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08003267
3268 /* If the pagetables are shared don't copy or take references */
3269 if (dst_pte == src_pte)
3270 continue;
3271
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003272 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3273 src_ptl = huge_pte_lockptr(h, src, src_pte);
3274 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003275 entry = huge_ptep_get(src_pte);
3276 if (huge_pte_none(entry)) { /* skip none entry */
3277 ;
3278 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3279 is_hugetlb_entry_hwpoisoned(entry))) {
3280 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3281
3282 if (is_write_migration_entry(swp_entry) && cow) {
3283 /*
3284 * COW mappings require pages in both
3285 * parent and child to be set to read.
3286 */
3287 make_migration_entry_read(&swp_entry);
3288 entry = swp_entry_to_pte(swp_entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07003289 set_huge_swap_pte_at(src, addr, src_pte,
3290 entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003291 }
Punit Agrawale5251fd2017-07-06 15:39:50 -07003292 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003293 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11003294 if (cow) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003295 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003296 mmu_notifier_invalidate_range(src, mmun_start,
3297 mmun_end);
3298 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07003299 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07003300 ptepage = pte_page(entry);
3301 get_page(ptepage);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003302 page_dup_rmap(ptepage, true);
Hugh Dickins1c598272005-10-19 21:23:43 -07003303 set_huge_pte_at(dst, addr, dst_pte, entry);
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003304 hugetlb_count_add(pages_per_huge_page(h), dst);
Hugh Dickins1c598272005-10-19 21:23:43 -07003305 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003306 spin_unlock(src_ptl);
3307 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003308 }
David Gibson63551ae2005-06-21 17:14:44 -07003309
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003310 if (cow)
3311 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3312
3313 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07003314}
3315
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003316void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3317 unsigned long start, unsigned long end,
3318 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07003319{
3320 struct mm_struct *mm = vma->vm_mm;
3321 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07003322 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07003323 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003324 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07003325 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07003326 struct hstate *h = hstate_vma(vma);
3327 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003328 const unsigned long mmun_start = start; /* For mmu_notifiers */
3329 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07003330
David Gibson63551ae2005-06-21 17:14:44 -07003331 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07003332 BUG_ON(start & ~huge_page_mask(h));
3333 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07003334
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08003335 /*
3336 * This is a hugetlb vma, all the pte entries should point
3337 * to huge page.
3338 */
3339 tlb_remove_check_page_size_change(tlb, sz);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003340 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003341 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Hillf Danton569f48b82014-12-10 15:44:41 -08003342 address = start;
Hillf Danton569f48b82014-12-10 15:44:41 -08003343 for (; address < end; address += sz) {
Punit Agrawal7868a202017-07-06 15:39:42 -07003344 ptep = huge_pte_offset(mm, address, sz);
Adam Litke4c887262005-10-29 18:16:46 -07003345 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07003346 continue;
3347
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003348 ptl = huge_pte_lock(h, mm, ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003349 if (huge_pmd_unshare(mm, &address, ptep)) {
3350 spin_unlock(ptl);
3351 continue;
3352 }
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003353
Hillf Danton66293262012-03-23 15:01:48 -07003354 pte = huge_ptep_get(ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003355 if (huge_pte_none(pte)) {
3356 spin_unlock(ptl);
3357 continue;
3358 }
Hillf Danton66293262012-03-23 15:01:48 -07003359
3360 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003361 * Migrating hugepage or HWPoisoned hugepage is already
3362 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07003363 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003364 if (unlikely(!pte_present(pte))) {
Punit Agrawal9386fac2017-07-06 15:39:46 -07003365 huge_pte_clear(mm, address, ptep, sz);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003366 spin_unlock(ptl);
3367 continue;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08003368 }
Hillf Danton66293262012-03-23 15:01:48 -07003369
3370 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003371 /*
3372 * If a reference page is supplied, it is because a specific
3373 * page is being unmapped, not a range. Ensure the page we
3374 * are about to unmap is the actual page of interest.
3375 */
3376 if (ref_page) {
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003377 if (page != ref_page) {
3378 spin_unlock(ptl);
3379 continue;
3380 }
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003381 /*
3382 * Mark the VMA as having unmapped its page so that
3383 * future faults in this VMA will fail rather than
3384 * looking like data was lost
3385 */
3386 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3387 }
3388
David Gibsonc7546f82005-08-05 11:59:35 -07003389 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.Vb528e4b2016-12-12 16:42:37 -08003390 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003391 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08003392 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07003393
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003394 hugetlb_count_sub(pages_per_huge_page(h), mm);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08003395 page_remove_rmap(page, true);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003396
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003397 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07003398 tlb_remove_page_size(tlb, page, huge_page_size(h));
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003399 /*
3400 * Bail out after unmapping reference page if supplied
3401 */
3402 if (ref_page)
3403 break;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07003404 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003405 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003406 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407}
David Gibson63551ae2005-06-21 17:14:44 -07003408
Mel Gormand8333522012-07-31 16:46:20 -07003409void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3410 struct vm_area_struct *vma, unsigned long start,
3411 unsigned long end, struct page *ref_page)
3412{
3413 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3414
3415 /*
3416 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3417 * test will fail on a vma being torn down, and not grab a page table
3418 * on its way out. We're lucky that the flag has such an appropriate
3419 * name, and can in fact be safely cleared here. We could clear it
3420 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003421 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07003422 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003423 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07003424 */
3425 vma->vm_flags &= ~VM_MAYSHARE;
3426}
3427
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003428void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003429 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003430{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003431 struct mm_struct *mm;
3432 struct mmu_gather tlb;
3433
3434 mm = vma->vm_mm;
3435
Linus Torvalds2b047252013-08-15 11:42:25 -07003436 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003437 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3438 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003439}
3440
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003441/*
3442 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3443 * mappping it owns the reserve page for. The intention is to unmap the page
3444 * from other VMAs and let the children be SIGKILLed if they are faulting the
3445 * same region.
3446 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003447static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3448 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003449{
Adam Litke75266742008-11-12 13:24:56 -08003450 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003451 struct vm_area_struct *iter_vma;
3452 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003453 pgoff_t pgoff;
3454
3455 /*
3456 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3457 * from page cache lookup which is in HPAGE_SIZE units.
3458 */
Adam Litke75266742008-11-12 13:24:56 -08003459 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07003460 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3461 vma->vm_pgoff;
Al Viro93c76a32015-12-04 23:45:44 -05003462 mapping = vma->vm_file->f_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003463
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08003464 /*
3465 * Take the mapping lock for the duration of the table walk. As
3466 * this mapping should be shared between all the VMAs,
3467 * __unmap_hugepage_range() is called as the lock is already held
3468 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003469 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07003470 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003471 /* Do not unmap the current VMA */
3472 if (iter_vma == vma)
3473 continue;
3474
3475 /*
Mel Gorman2f84a892015-10-01 15:36:57 -07003476 * Shared VMAs have their own reserves and do not affect
3477 * MAP_PRIVATE accounting but it is possible that a shared
3478 * VMA is using the same page so check and skip such VMAs.
3479 */
3480 if (iter_vma->vm_flags & VM_MAYSHARE)
3481 continue;
3482
3483 /*
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003484 * Unmap the page from other VMAs without their own reserves.
3485 * They get marked to be SIGKILLed if they fault in these
3486 * areas. This is because a future no-page fault on this VMA
3487 * could insert a zeroed page instead of the data existing
3488 * from the time of fork. This would look like data corruption
3489 */
3490 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003491 unmap_hugepage_range(iter_vma, address,
3492 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003493 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003494 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003495}
3496
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003497/*
3498 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08003499 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3500 * cannot race with other handlers or page migration.
3501 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003502 */
David Gibson1e8f8892006-01-06 00:10:44 -08003503static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08003504 unsigned long address, pte_t *ptep,
3505 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08003506{
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08003507 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003508 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08003509 struct page *old_page, *new_page;
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003510 int ret = 0, outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003511 unsigned long mmun_start; /* For mmu_notifiers */
3512 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08003513
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08003514 pte = huge_ptep_get(ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003515 old_page = pte_page(pte);
3516
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003517retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08003518 /* If no-one else is actually using this page, avoid the copy
3519 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07003520 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
Hugh Dickins5a499732016-07-14 12:07:38 -07003521 page_move_anon_rmap(old_page, vma);
David Gibson1e8f8892006-01-06 00:10:44 -08003522 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07003523 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003524 }
3525
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003526 /*
3527 * If the process that created a MAP_PRIVATE mapping is about to
3528 * perform a COW due to a shared page count, attempt to satisfy
3529 * the allocation without using the existing reserves. The pagecache
3530 * page is used to determine if the reserve at this address was
3531 * consumed or not. If reserves were used, a partial faulted mapping
3532 * at the time of fork() could consume its reserves on COW instead
3533 * of the full address range.
3534 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07003535 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003536 old_page != pagecache_page)
3537 outside_reserve = 1;
3538
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003539 get_page(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003540
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003541 /*
3542 * Drop page table lock as buddy allocator may be called. It will
3543 * be acquired again before returning to the caller, as expected.
3544 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003545 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003546 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08003547
Adam Litke2fc39ce2007-11-14 16:59:39 -08003548 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003549 /*
3550 * If a process owning a MAP_PRIVATE mapping fails to COW,
3551 * it is due to references held by a child and an insufficient
3552 * huge page pool. To guarantee the original mappers
3553 * reliability, unmap the page from child processes. The child
3554 * may get SIGKILLed if it later faults.
3555 */
3556 if (outside_reserve) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003557 put_page(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003558 BUG_ON(huge_pte_none(pte));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003559 unmap_ref_private(mm, vma, old_page, address);
3560 BUG_ON(huge_pte_none(pte));
3561 spin_lock(ptl);
Punit Agrawal7868a202017-07-06 15:39:42 -07003562 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3563 huge_page_size(h));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003564 if (likely(ptep &&
3565 pte_same(huge_ptep_get(ptep), pte)))
3566 goto retry_avoidcopy;
3567 /*
3568 * race occurs while re-acquiring page table
3569 * lock, and our job is done.
3570 */
3571 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003572 }
3573
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003574 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3575 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3576 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08003577 }
3578
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003579 /*
3580 * When the original hugepage is shared one, it does not have
3581 * anon_vma prepared.
3582 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07003583 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003584 ret = VM_FAULT_OOM;
3585 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07003586 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003587
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003588 copy_user_huge_page(new_page, old_page, address, vma,
3589 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003590 __SetPageUptodate(new_page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003591 set_page_huge_active(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08003592
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003593 mmun_start = address & huge_page_mask(h);
3594 mmun_end = mmun_start + huge_page_size(h);
3595 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003596
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003597 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003598 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003599 * before the page tables are altered
3600 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003601 spin_lock(ptl);
Punit Agrawal7868a202017-07-06 15:39:42 -07003602 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3603 huge_page_size(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07003604 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07003605 ClearPagePrivate(new_page);
3606
David Gibson1e8f8892006-01-06 00:10:44 -08003607 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07003608 huge_ptep_clear_flush(vma, address, ptep);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003609 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08003610 set_huge_pte_at(mm, address, ptep,
3611 make_huge_pte(vma, new_page, 1));
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08003612 page_remove_rmap(old_page, true);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09003613 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08003614 /* Make the old page be freed below */
3615 new_page = old_page;
3616 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003617 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003618 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003619out_release_all:
Mike Kravetz96b96a92016-11-10 10:46:32 -08003620 restore_reserve_on_error(h, vma, address, new_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003621 put_page(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003622out_release_old:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003623 put_page(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07003624
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003625 spin_lock(ptl); /* Caller expects lock to be held */
3626 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08003627}
3628
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003629/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07003630static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3631 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003632{
3633 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07003634 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003635
3636 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07003637 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003638
3639 return find_lock_page(mapping, idx);
3640}
3641
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003642/*
3643 * Return whether there is a pagecache page to back given address within VMA.
3644 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3645 */
3646static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003647 struct vm_area_struct *vma, unsigned long address)
3648{
3649 struct address_space *mapping;
3650 pgoff_t idx;
3651 struct page *page;
3652
3653 mapping = vma->vm_file->f_mapping;
3654 idx = vma_hugecache_offset(h, vma, address);
3655
3656 page = find_get_page(mapping, idx);
3657 if (page)
3658 put_page(page);
3659 return page != NULL;
3660}
3661
Mike Kravetzab76ad52015-09-08 15:01:50 -07003662int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3663 pgoff_t idx)
3664{
3665 struct inode *inode = mapping->host;
3666 struct hstate *h = hstate_inode(inode);
3667 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3668
3669 if (err)
3670 return err;
3671 ClearPagePrivate(page);
3672
3673 spin_lock(&inode->i_lock);
3674 inode->i_blocks += blocks_per_huge_page(h);
3675 spin_unlock(&inode->i_lock);
3676 return 0;
3677}
3678
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07003679static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003680 struct address_space *mapping, pgoff_t idx,
3681 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003682{
Andi Kleena5516432008-07-23 21:27:41 -07003683 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003684 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08003685 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003686 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07003687 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08003688 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003689 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07003690
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003691 /*
3692 * Currently, we are forced to kill the process in the event the
3693 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003694 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003695 */
3696 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Geoffrey Thomas910154d2016-03-09 14:08:04 -08003697 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08003698 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003699 return ret;
3700 }
3701
Adam Litke4c887262005-10-29 18:16:46 -07003702 /*
3703 * Use page lock to guard against racing truncation
3704 * before we get page_table_lock.
3705 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08003706retry:
3707 page = find_lock_page(mapping, idx);
3708 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07003709 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07003710 if (idx >= size)
3711 goto out;
Mike Kravetz1a1aad82017-02-22 15:43:01 -08003712
3713 /*
3714 * Check for page in userfault range
3715 */
3716 if (userfaultfd_missing(vma)) {
3717 u32 hash;
3718 struct vm_fault vmf = {
3719 .vma = vma,
3720 .address = address,
3721 .flags = flags,
3722 /*
3723 * Hard to debug if it ends up being
3724 * used by a callee that assumes
3725 * something about the other
3726 * uninitialized fields... same as in
3727 * memory.c
3728 */
3729 };
3730
3731 /*
3732 * hugetlb_fault_mutex must be dropped before
3733 * handling userfault. Reacquire after handling
3734 * fault to make calling code simpler.
3735 */
3736 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3737 idx, address);
3738 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3739 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3740 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3741 goto out;
3742 }
3743
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003744 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08003745 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07003746 ret = PTR_ERR(page);
3747 if (ret == -ENOMEM)
3748 ret = VM_FAULT_OOM;
3749 else
3750 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003751 goto out;
3752 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003753 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003754 __SetPageUptodate(page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003755 set_page_huge_active(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003756
Mel Gormanf83a2752009-05-28 14:34:40 -07003757 if (vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzab76ad52015-09-08 15:01:50 -07003758 int err = huge_add_to_page_cache(page, mapping, idx);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003759 if (err) {
3760 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003761 if (err == -EEXIST)
3762 goto retry;
3763 goto out;
3764 }
Mel Gorman23be7462010-04-23 13:17:56 -04003765 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003766 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003767 if (unlikely(anon_vma_prepare(vma))) {
3768 ret = VM_FAULT_OOM;
3769 goto backout_unlocked;
3770 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08003771 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04003772 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003773 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003774 /*
3775 * If memory error occurs between mmap() and fault, some process
3776 * don't have hwpoisoned swap entry for errored virtual address.
3777 * So we need to block hugepage fault by PG_hwpoison bit check.
3778 */
3779 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07003780 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003781 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003782 goto backout_unlocked;
3783 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08003784 }
David Gibson1e8f8892006-01-06 00:10:44 -08003785
Andy Whitcroft57303d82008-08-12 15:08:47 -07003786 /*
3787 * If we are going to COW a private mapping later, we examine the
3788 * pending reservations for this page now. This will ensure that
3789 * any allocations necessary to record that reservation occur outside
3790 * the spinlock.
3791 */
Mike Kravetz5e911372015-09-08 15:01:28 -07003792 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003793 if (vma_needs_reservation(h, vma, address) < 0) {
3794 ret = VM_FAULT_OOM;
3795 goto backout_unlocked;
3796 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003797 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003798 vma_end_reservation(h, vma, address);
Mike Kravetz5e911372015-09-08 15:01:28 -07003799 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003800
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08003801 ptl = huge_pte_lock(h, mm, ptep);
Andi Kleena5516432008-07-23 21:27:41 -07003802 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07003803 if (idx >= size)
3804 goto backout;
3805
Nick Piggin83c54072007-07-19 01:47:05 -07003806 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003807 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07003808 goto backout;
3809
Joonsoo Kim07443a82013-09-11 14:21:58 -07003810 if (anon_rmap) {
3811 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08003812 hugepage_add_new_anon_rmap(page, vma, address);
Choi Gi-yongac714902014-04-07 15:37:36 -07003813 } else
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003814 page_dup_rmap(page, true);
David Gibson1e8f8892006-01-06 00:10:44 -08003815 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3816 && (vma->vm_flags & VM_SHARED)));
3817 set_huge_pte_at(mm, address, ptep, new_pte);
3818
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003819 hugetlb_count_add(pages_per_huge_page(h), mm);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003820 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08003821 /* Optimization, do the COW without a second fault */
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08003822 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003823 }
3824
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003825 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07003826 unlock_page(page);
3827out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003828 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07003829
3830backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003831 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07003832backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07003833 unlock_page(page);
Mike Kravetz96b96a92016-11-10 10:46:32 -08003834 restore_reserve_on_error(h, vma, address, page);
Adam Litke4c887262005-10-29 18:16:46 -07003835 put_page(page);
3836 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003837}
3838
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003839#ifdef CONFIG_SMP
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003840u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003841 struct vm_area_struct *vma,
3842 struct address_space *mapping,
3843 pgoff_t idx, unsigned long address)
3844{
3845 unsigned long key[2];
3846 u32 hash;
3847
3848 if (vma->vm_flags & VM_SHARED) {
3849 key[0] = (unsigned long) mapping;
3850 key[1] = idx;
3851 } else {
3852 key[0] = (unsigned long) mm;
3853 key[1] = address >> huge_page_shift(h);
3854 }
3855
3856 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3857
3858 return hash & (num_fault_mutexes - 1);
3859}
3860#else
3861/*
3862 * For uniprocesor systems we always use a single mutex, so just
3863 * return 0 and avoid the hashing overhead.
3864 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003865u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003866 struct vm_area_struct *vma,
3867 struct address_space *mapping,
3868 pgoff_t idx, unsigned long address)
3869{
3870 return 0;
3871}
3872#endif
3873
Adam Litke86e52162006-01-06 00:10:43 -08003874int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01003875 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08003876{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003877 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003878 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08003879 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003880 u32 hash;
3881 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003882 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07003883 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07003884 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003885 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003886 int need_wait_lock = 0;
Adam Litke86e52162006-01-06 00:10:43 -08003887
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08003888 address &= huge_page_mask(h);
3889
Punit Agrawal7868a202017-07-06 15:39:42 -07003890 ptep = huge_pte_offset(mm, address, huge_page_size(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09003891 if (ptep) {
3892 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003893 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003894 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003895 return 0;
3896 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07003897 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003898 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi0d777df2015-12-11 13:40:49 -08003899 } else {
3900 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3901 if (!ptep)
3902 return VM_FAULT_OOM;
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09003903 }
3904
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003905 mapping = vma->vm_file->f_mapping;
3906 idx = vma_hugecache_offset(h, vma, address);
3907
David Gibson3935baa2006-03-22 00:08:53 -08003908 /*
3909 * Serialize hugepage allocation and instantiation, so that we don't
3910 * get spurious allocation failures if two CPUs race to instantiate
3911 * the same page in the page cache.
3912 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003913 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3914 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003915
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003916 entry = huge_ptep_get(ptep);
3917 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003918 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003919 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003920 }
Adam Litke86e52162006-01-06 00:10:43 -08003921
Nick Piggin83c54072007-07-19 01:47:05 -07003922 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003923
Andy Whitcroft57303d82008-08-12 15:08:47 -07003924 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003925 * entry could be a migration/hwpoison entry at this point, so this
3926 * check prevents the kernel from going below assuming that we have
3927 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3928 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3929 * handle it.
3930 */
3931 if (!pte_present(entry))
3932 goto out_mutex;
3933
3934 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07003935 * If we are going to COW the mapping later, we examine the pending
3936 * reservations for this page now. This will ensure that any
3937 * allocations necessary to record that reservation occur outside the
3938 * spinlock. For private mappings, we also lookup the pagecache
3939 * page now as it is used to determine if a reservation has been
3940 * consumed.
3941 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003942 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003943 if (vma_needs_reservation(h, vma, address) < 0) {
3944 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003945 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003946 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003947 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003948 vma_end_reservation(h, vma, address);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003949
Mel Gormanf83a2752009-05-28 14:34:40 -07003950 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003951 pagecache_page = hugetlbfs_pagecache_page(h,
3952 vma, address);
3953 }
3954
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003955 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003956
David Gibson1e8f8892006-01-06 00:10:44 -08003957 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003958 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003959 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003960
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003961 /*
3962 * hugetlb_cow() requires page locks of pte_page(entry) and
3963 * pagecache_page, so here we need take the former one
3964 * when page != pagecache_page or !pagecache_page.
3965 */
3966 page = pte_page(entry);
3967 if (page != pagecache_page)
3968 if (!trylock_page(page)) {
3969 need_wait_lock = 1;
3970 goto out_ptl;
3971 }
3972
3973 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07003974
Hugh Dickins788c7df2009-06-23 13:49:05 +01003975 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003976 if (!huge_pte_write(entry)) {
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08003977 ret = hugetlb_cow(mm, vma, address, ptep,
3978 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003979 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07003980 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003981 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003982 }
3983 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003984 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3985 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003986 update_mmu_cache(vma, address, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003987out_put_page:
3988 if (page != pagecache_page)
3989 unlock_page(page);
3990 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003991out_ptl:
3992 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003993
3994 if (pagecache_page) {
3995 unlock_page(pagecache_page);
3996 put_page(pagecache_page);
3997 }
David Gibsonb4d1d992008-10-15 22:01:11 -07003998out_mutex:
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003999 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004000 /*
4001 * Generally it's safe to hold refcount during waiting page lock. But
4002 * here we just wait to defer the next page fault to avoid busy loop and
4003 * the page is not used after unlocked before returning from the current
4004 * page fault. So we are safe from accessing freed page, even if we wait
4005 * here without taking refcount.
4006 */
4007 if (need_wait_lock)
4008 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08004009 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08004010}
4011
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004012/*
4013 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4014 * modifications for huge pages.
4015 */
4016int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4017 pte_t *dst_pte,
4018 struct vm_area_struct *dst_vma,
4019 unsigned long dst_addr,
4020 unsigned long src_addr,
4021 struct page **pagep)
4022{
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004023 int vm_shared = dst_vma->vm_flags & VM_SHARED;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004024 struct hstate *h = hstate_vma(dst_vma);
4025 pte_t _dst_pte;
4026 spinlock_t *ptl;
4027 int ret;
4028 struct page *page;
4029
4030 if (!*pagep) {
4031 ret = -ENOMEM;
4032 page = alloc_huge_page(dst_vma, dst_addr, 0);
4033 if (IS_ERR(page))
4034 goto out;
4035
4036 ret = copy_huge_page_from_user(page,
4037 (const void __user *) src_addr,
Mike Kravetz810a56b2017-02-22 15:42:58 -08004038 pages_per_huge_page(h), false);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004039
4040 /* fallback to copy_from_user outside mmap_sem */
4041 if (unlikely(ret)) {
4042 ret = -EFAULT;
4043 *pagep = page;
4044 /* don't free the page */
4045 goto out;
4046 }
4047 } else {
4048 page = *pagep;
4049 *pagep = NULL;
4050 }
4051
4052 /*
4053 * The memory barrier inside __SetPageUptodate makes sure that
4054 * preceding stores to the page contents become visible before
4055 * the set_pte_at() write.
4056 */
4057 __SetPageUptodate(page);
4058 set_page_huge_active(page);
4059
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004060 /*
4061 * If shared, add to page cache
4062 */
4063 if (vm_shared) {
4064 struct address_space *mapping = dst_vma->vm_file->f_mapping;
4065 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4066
4067 ret = huge_add_to_page_cache(page, mapping, idx);
4068 if (ret)
4069 goto out_release_nounlock;
4070 }
4071
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004072 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4073 spin_lock(ptl);
4074
4075 ret = -EEXIST;
4076 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4077 goto out_release_unlock;
4078
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004079 if (vm_shared) {
4080 page_dup_rmap(page, true);
4081 } else {
4082 ClearPagePrivate(page);
4083 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4084 }
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004085
4086 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4087 if (dst_vma->vm_flags & VM_WRITE)
4088 _dst_pte = huge_pte_mkdirty(_dst_pte);
4089 _dst_pte = pte_mkyoung(_dst_pte);
4090
4091 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4092
4093 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4094 dst_vma->vm_flags & VM_WRITE);
4095 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4096
4097 /* No need to invalidate - it was non-present before */
4098 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4099
4100 spin_unlock(ptl);
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004101 if (vm_shared)
4102 unlock_page(page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004103 ret = 0;
4104out:
4105 return ret;
4106out_release_unlock:
4107 spin_unlock(ptl);
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004108out_release_nounlock:
4109 if (vm_shared)
4110 unlock_page(page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004111 put_page(page);
4112 goto out;
4113}
4114
Michel Lespinasse28a35712013-02-22 16:35:55 -08004115long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4116 struct page **pages, struct vm_area_struct **vmas,
4117 unsigned long *position, unsigned long *nr_pages,
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004118 long i, unsigned int flags, int *nonblocking)
David Gibson63551ae2005-06-21 17:14:44 -07004119{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004120 unsigned long pfn_offset;
4121 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08004122 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07004123 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07004124
David Gibson63551ae2005-06-21 17:14:44 -07004125 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07004126 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004127 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004128 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07004129 struct page *page;
4130
4131 /*
David Rientjes02057962015-04-14 15:48:24 -07004132 * If we have a pending SIGKILL, don't keep faulting pages and
4133 * potentially allocating memory.
4134 */
4135 if (unlikely(fatal_signal_pending(current))) {
4136 remainder = 0;
4137 break;
4138 }
4139
4140 /*
Adam Litke4c887262005-10-29 18:16:46 -07004141 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004142 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07004143 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004144 *
4145 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07004146 */
Punit Agrawal7868a202017-07-06 15:39:42 -07004147 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4148 huge_page_size(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004149 if (pte)
4150 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004151 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07004152
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004153 /*
4154 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004155 * an error where there's an empty slot with no huge pagecache
4156 * to back it. This way, we avoid allocating a hugepage, and
4157 * the sparse dumpfile avoids allocating disk blocks, but its
4158 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004159 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004160 if (absent && (flags & FOLL_DUMP) &&
4161 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004162 if (pte)
4163 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004164 remainder = 0;
4165 break;
4166 }
4167
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07004168 /*
4169 * We need call hugetlb_fault for both hugepages under migration
4170 * (in which case hugetlb_fault waits for the migration,) and
4171 * hwpoisoned hugepages (in which case we need to prevent the
4172 * caller from accessing to them.) In order to do this, we use
4173 * here is_swap_pte instead of is_hugetlb_entry_migration and
4174 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4175 * both cases, and because we can't follow correct pages
4176 * directly from any kind of swap entries.
4177 */
4178 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07004179 ((flags & FOLL_WRITE) &&
4180 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07004181 int ret;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004182 unsigned int fault_flags = 0;
Adam Litke4c887262005-10-29 18:16:46 -07004183
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004184 if (pte)
4185 spin_unlock(ptl);
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004186 if (flags & FOLL_WRITE)
4187 fault_flags |= FAULT_FLAG_WRITE;
4188 if (nonblocking)
4189 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4190 if (flags & FOLL_NOWAIT)
4191 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4192 FAULT_FLAG_RETRY_NOWAIT;
4193 if (flags & FOLL_TRIED) {
4194 VM_WARN_ON_ONCE(fault_flags &
4195 FAULT_FLAG_ALLOW_RETRY);
4196 fault_flags |= FAULT_FLAG_TRIED;
4197 }
4198 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4199 if (ret & VM_FAULT_ERROR) {
James Morse9a291a72017-06-02 14:46:46 -07004200 int err = vm_fault_to_errno(ret, flags);
4201
4202 if (err)
4203 return err;
4204
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004205 remainder = 0;
4206 break;
4207 }
4208 if (ret & VM_FAULT_RETRY) {
4209 if (nonblocking)
4210 *nonblocking = 0;
4211 *nr_pages = 0;
4212 /*
4213 * VM_FAULT_RETRY must not return an
4214 * error, it will return zero
4215 * instead.
4216 *
4217 * No need to update "position" as the
4218 * caller will not check it after
4219 * *nr_pages is set to 0.
4220 */
4221 return i;
4222 }
4223 continue;
Adam Litke4c887262005-10-29 18:16:46 -07004224 }
David Gibson63551ae2005-06-21 17:14:44 -07004225
Andi Kleena5516432008-07-23 21:27:41 -07004226 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004227 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004228same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08004229 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004230 pages[i] = mem_map_offset(page, pfn_offset);
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08004231 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08004232 }
David Gibson63551ae2005-06-21 17:14:44 -07004233
4234 if (vmas)
4235 vmas[i] = vma;
4236
4237 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004238 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07004239 --remainder;
4240 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004241 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07004242 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004243 /*
4244 * We use pfn_offset to avoid touching the pageframes
4245 * of this compound page.
4246 */
4247 goto same_page;
4248 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004249 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07004250 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08004251 *nr_pages = remainder;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004252 /*
4253 * setting position is actually required only if remainder is
4254 * not zero but it's faster not to add a "if (remainder)"
4255 * branch.
4256 */
David Gibson63551ae2005-06-21 17:14:44 -07004257 *position = vaddr;
4258
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004259 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07004260}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004261
Aneesh Kumar K.V5491ae72016-07-13 15:06:43 +05304262#ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4263/*
4264 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4265 * implement this.
4266 */
4267#define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4268#endif
4269
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004270unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004271 unsigned long address, unsigned long end, pgprot_t newprot)
4272{
4273 struct mm_struct *mm = vma->vm_mm;
4274 unsigned long start = address;
4275 pte_t *ptep;
4276 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07004277 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004278 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004279
4280 BUG_ON(address >= end);
4281 flush_cache_range(vma, address, end);
4282
Rik van Riela5338092014-04-07 15:36:57 -07004283 mmu_notifier_invalidate_range_start(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004284 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07004285 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004286 spinlock_t *ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07004287 ptep = huge_pte_offset(mm, address, huge_page_size(h));
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004288 if (!ptep)
4289 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004290 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004291 if (huge_pmd_unshare(mm, &address, ptep)) {
4292 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004293 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08004294 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004295 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08004296 pte = huge_ptep_get(ptep);
4297 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4298 spin_unlock(ptl);
4299 continue;
4300 }
4301 if (unlikely(is_hugetlb_entry_migration(pte))) {
4302 swp_entry_t entry = pte_to_swp_entry(pte);
4303
4304 if (is_write_migration_entry(entry)) {
4305 pte_t newpte;
4306
4307 make_migration_entry_read(&entry);
4308 newpte = swp_entry_to_pte(entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07004309 set_huge_swap_pte_at(mm, address, ptep,
4310 newpte, huge_page_size(h));
Naoya Horiguchia8bda282015-02-11 15:25:28 -08004311 pages++;
4312 }
4313 spin_unlock(ptl);
4314 continue;
4315 }
4316 if (!huge_pte_none(pte)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004317 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07004318 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08004319 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004320 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004321 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004322 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004323 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004324 }
Mel Gormand8333522012-07-31 16:46:20 -07004325 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004326 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07004327 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004328 * once we release i_mmap_rwsem, another task can do the final put_page
Mel Gormand8333522012-07-31 16:46:20 -07004329 * and that page table be reused and filled with junk.
4330 */
Aneesh Kumar K.V5491ae72016-07-13 15:06:43 +05304331 flush_hugetlb_tlb_range(vma, start, end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11004332 mmu_notifier_invalidate_range(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004333 i_mmap_unlock_write(vma->vm_file->f_mapping);
Rik van Riela5338092014-04-07 15:36:57 -07004334 mmu_notifier_invalidate_range_end(mm, start, end);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004335
4336 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004337}
4338
Mel Gormana1e78772008-07-23 21:27:23 -07004339int hugetlb_reserve_pages(struct inode *inode,
4340 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00004341 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09004342 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07004343{
Mel Gorman17c9d122009-02-11 16:34:16 +00004344 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07004345 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07004346 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004347 struct resv_map *resv_map;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004348 long gbl_reserve;
Adam Litkee4e574b2007-10-16 01:26:19 -07004349
Mel Gormana1e78772008-07-23 21:27:23 -07004350 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00004351 * Only apply hugepage reservation if asked. At fault time, an
4352 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07004353 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00004354 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09004355 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00004356 return 0;
4357
4358 /*
Mel Gormana1e78772008-07-23 21:27:23 -07004359 * Shared mappings base their reservation on the number of pages that
4360 * are already allocated on behalf of the file. Private mappings need
4361 * to reserve the full area even if read-only as mprotect() may be
4362 * called to make the mapping read-write. Assume !vma is a shm mapping
4363 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07004364 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004365 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004366
Joonsoo Kim1406ec92014-04-03 14:47:26 -07004367 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004368
4369 } else {
4370 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00004371 if (!resv_map)
4372 return -ENOMEM;
4373
Mel Gorman17c9d122009-02-11 16:34:16 +00004374 chg = to - from;
4375
Mel Gorman5a6fe122009-02-10 14:02:27 +00004376 set_vma_resv_map(vma, resv_map);
4377 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4378 }
4379
Dave Hansenc50ac052012-05-29 15:06:46 -07004380 if (chg < 0) {
4381 ret = chg;
4382 goto out_err;
4383 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004384
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004385 /*
4386 * There must be enough pages in the subpool for the mapping. If
4387 * the subpool has a minimum size, there may be some global
4388 * reservations already in place (gbl_reserve).
4389 */
4390 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4391 if (gbl_reserve < 0) {
Dave Hansenc50ac052012-05-29 15:06:46 -07004392 ret = -ENOSPC;
4393 goto out_err;
4394 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004395
4396 /*
4397 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07004398 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00004399 */
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004400 ret = hugetlb_acct_memory(h, gbl_reserve);
Mel Gorman17c9d122009-02-11 16:34:16 +00004401 if (ret < 0) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004402 /* put back original number of pages, chg */
4403 (void)hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07004404 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00004405 }
4406
4407 /*
4408 * Account for the reservations made. Shared mappings record regions
4409 * that have reservations as they are shared by multiple VMAs.
4410 * When the last VMA disappears, the region map says how much
4411 * the reservation was and the page cache tells how much of
4412 * the reservation was consumed. Private mappings are per-VMA and
4413 * only the consumed reservations are tracked. When the VMA
4414 * disappears, the original reservation is the VMA size and the
4415 * consumed reservations are stored in the map. Hence, nothing
4416 * else has to be done for private mappings here
4417 */
Mike Kravetz33039672015-06-24 16:57:58 -07004418 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4419 long add = region_add(resv_map, from, to);
4420
4421 if (unlikely(chg > add)) {
4422 /*
4423 * pages in this range were added to the reserve
4424 * map between region_chg and region_add. This
4425 * indicates a race with alloc_huge_page. Adjust
4426 * the subpool and reserve counts modified above
4427 * based on the difference.
4428 */
4429 long rsv_adjust;
4430
4431 rsv_adjust = hugepage_subpool_put_pages(spool,
4432 chg - add);
4433 hugetlb_acct_memory(h, -rsv_adjust);
4434 }
4435 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004436 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07004437out_err:
Mike Kravetz5e911372015-09-08 15:01:28 -07004438 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mike Kravetzff8c0c52017-03-31 15:12:07 -07004439 /* Don't call region_abort if region_chg failed */
4440 if (chg >= 0)
4441 region_abort(resv_map, from, to);
Joonsoo Kimf031dd22014-04-03 14:47:28 -07004442 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4443 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07004444 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004445}
4446
Mike Kravetzb5cec282015-09-08 15:01:41 -07004447long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4448 long freed)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004449{
Andi Kleena5516432008-07-23 21:27:41 -07004450 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004451 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004452 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07004453 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004454 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08004455
Mike Kravetzb5cec282015-09-08 15:01:41 -07004456 if (resv_map) {
4457 chg = region_del(resv_map, start, end);
4458 /*
4459 * region_del() can fail in the rare case where a region
4460 * must be split and another region descriptor can not be
4461 * allocated. If end == LONG_MAX, it will not fail.
4462 */
4463 if (chg < 0)
4464 return chg;
4465 }
4466
Ken Chen45c682a2007-11-14 16:59:44 -08004467 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07004468 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08004469 spin_unlock(&inode->i_lock);
4470
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004471 /*
4472 * If the subpool has a minimum size, the number of global
4473 * reservations to be released may be adjusted.
4474 */
4475 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4476 hugetlb_acct_memory(h, -gbl_reserve);
Mike Kravetzb5cec282015-09-08 15:01:41 -07004477
4478 return 0;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004479}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004480
Steve Capper3212b532013-04-23 12:35:02 +01004481#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4482static unsigned long page_table_shareable(struct vm_area_struct *svma,
4483 struct vm_area_struct *vma,
4484 unsigned long addr, pgoff_t idx)
4485{
4486 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4487 svma->vm_start;
4488 unsigned long sbase = saddr & PUD_MASK;
4489 unsigned long s_end = sbase + PUD_SIZE;
4490
4491 /* Allow segments to share if only one is marked locked */
Eric B Munsonde60f5f2015-11-05 18:51:36 -08004492 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4493 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
Steve Capper3212b532013-04-23 12:35:02 +01004494
4495 /*
4496 * match the virtual addresses, permission and the alignment of the
4497 * page table page.
4498 */
4499 if (pmd_index(addr) != pmd_index(saddr) ||
4500 vm_flags != svm_flags ||
4501 sbase < svma->vm_start || svma->vm_end < s_end)
4502 return 0;
4503
4504 return saddr;
4505}
4506
Nicholas Krause31aafb42015-09-04 15:47:58 -07004507static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
Steve Capper3212b532013-04-23 12:35:02 +01004508{
4509 unsigned long base = addr & PUD_MASK;
4510 unsigned long end = base + PUD_SIZE;
4511
4512 /*
4513 * check on proper vm_flags and page table alignment
4514 */
4515 if (vma->vm_flags & VM_MAYSHARE &&
4516 vma->vm_start <= base && end <= vma->vm_end)
Nicholas Krause31aafb42015-09-04 15:47:58 -07004517 return true;
4518 return false;
Steve Capper3212b532013-04-23 12:35:02 +01004519}
4520
4521/*
4522 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4523 * and returns the corresponding pte. While this is not necessary for the
4524 * !shared pmd case because we can allocate the pmd later as well, it makes the
4525 * code much cleaner. pmd allocation is essential for the shared case because
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004526 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
Steve Capper3212b532013-04-23 12:35:02 +01004527 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4528 * bad pmd for sharing.
4529 */
4530pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4531{
4532 struct vm_area_struct *vma = find_vma(mm, addr);
4533 struct address_space *mapping = vma->vm_file->f_mapping;
4534 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4535 vma->vm_pgoff;
4536 struct vm_area_struct *svma;
4537 unsigned long saddr;
4538 pte_t *spte = NULL;
4539 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004540 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01004541
4542 if (!vma_shareable(vma, addr))
4543 return (pte_t *)pmd_alloc(mm, pud, addr);
4544
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004545 i_mmap_lock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004546 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4547 if (svma == vma)
4548 continue;
4549
4550 saddr = page_table_shareable(svma, vma, addr, idx);
4551 if (saddr) {
Punit Agrawal7868a202017-07-06 15:39:42 -07004552 spte = huge_pte_offset(svma->vm_mm, saddr,
4553 vma_mmu_pagesize(svma));
Steve Capper3212b532013-04-23 12:35:02 +01004554 if (spte) {
4555 get_page(virt_to_page(spte));
4556 break;
4557 }
4558 }
4559 }
4560
4561 if (!spte)
4562 goto out;
4563
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08004564 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004565 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01004566 pud_populate(mm, pud,
4567 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovc17b1f42016-06-24 14:49:51 -07004568 mm_inc_nr_pmds(mm);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004569 } else {
Steve Capper3212b532013-04-23 12:35:02 +01004570 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004571 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004572 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01004573out:
4574 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004575 i_mmap_unlock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004576 return pte;
4577}
4578
4579/*
4580 * unmap huge page backed by shared pte.
4581 *
4582 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4583 * indicated by page_count > 1, unmap is achieved by clearing pud and
4584 * decrementing the ref count. If count == 1, the pte page is not shared.
4585 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004586 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01004587 *
4588 * returns: 1 successfully unmapped a shared pte page
4589 * 0 the underlying pte page is not shared, or it is the last user
4590 */
4591int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4592{
4593 pgd_t *pgd = pgd_offset(mm, *addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004594 p4d_t *p4d = p4d_offset(pgd, *addr);
4595 pud_t *pud = pud_offset(p4d, *addr);
Steve Capper3212b532013-04-23 12:35:02 +01004596
4597 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4598 if (page_count(virt_to_page(ptep)) == 1)
4599 return 0;
4600
4601 pud_clear(pud);
4602 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004603 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01004604 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4605 return 1;
4606}
Steve Capper9e5fc742013-04-30 08:02:03 +01004607#define want_pmd_share() (1)
4608#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4609pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4610{
4611 return NULL;
4612}
Zhang Zhene81f2d22015-06-24 16:56:13 -07004613
4614int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4615{
4616 return 0;
4617}
Steve Capper9e5fc742013-04-30 08:02:03 +01004618#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01004619#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4620
Steve Capper9e5fc742013-04-30 08:02:03 +01004621#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4622pte_t *huge_pte_alloc(struct mm_struct *mm,
4623 unsigned long addr, unsigned long sz)
4624{
4625 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004626 p4d_t *p4d;
Steve Capper9e5fc742013-04-30 08:02:03 +01004627 pud_t *pud;
4628 pte_t *pte = NULL;
4629
4630 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004631 p4d = p4d_offset(pgd, addr);
4632 pud = pud_alloc(mm, p4d, addr);
Steve Capper9e5fc742013-04-30 08:02:03 +01004633 if (pud) {
4634 if (sz == PUD_SIZE) {
4635 pte = (pte_t *)pud;
4636 } else {
4637 BUG_ON(sz != PMD_SIZE);
4638 if (want_pmd_share() && pud_none(*pud))
4639 pte = huge_pmd_share(mm, addr, pud);
4640 else
4641 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4642 }
4643 }
Michal Hocko4e666312016-08-02 14:02:34 -07004644 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
Steve Capper9e5fc742013-04-30 08:02:03 +01004645
4646 return pte;
4647}
4648
Punit Agrawal7868a202017-07-06 15:39:42 -07004649pte_t *huge_pte_offset(struct mm_struct *mm,
4650 unsigned long addr, unsigned long sz)
Steve Capper9e5fc742013-04-30 08:02:03 +01004651{
4652 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004653 p4d_t *p4d;
Steve Capper9e5fc742013-04-30 08:02:03 +01004654 pud_t *pud;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004655 pmd_t *pmd;
Steve Capper9e5fc742013-04-30 08:02:03 +01004656
4657 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004658 if (!pgd_present(*pgd))
4659 return NULL;
4660 p4d = p4d_offset(pgd, addr);
4661 if (!p4d_present(*p4d))
4662 return NULL;
4663 pud = pud_offset(p4d, addr);
4664 if (!pud_present(*pud))
4665 return NULL;
4666 if (pud_huge(*pud))
4667 return (pte_t *)pud;
4668 pmd = pmd_offset(pud, addr);
Steve Capper9e5fc742013-04-30 08:02:03 +01004669 return (pte_t *) pmd;
4670}
4671
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004672#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4673
4674/*
4675 * These functions are overwritable if your architecture needs its own
4676 * behavior.
4677 */
4678struct page * __weak
4679follow_huge_addr(struct mm_struct *mm, unsigned long address,
4680 int write)
4681{
4682 return ERR_PTR(-EINVAL);
4683}
4684
4685struct page * __weak
Aneesh Kumar K.V4dc71452017-07-06 15:38:56 -07004686follow_huge_pd(struct vm_area_struct *vma,
4687 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4688{
4689 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4690 return NULL;
4691}
4692
4693struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004694follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004695 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004696{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004697 struct page *page = NULL;
4698 spinlock_t *ptl;
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07004699 pte_t pte;
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004700retry:
4701 ptl = pmd_lockptr(mm, pmd);
4702 spin_lock(ptl);
4703 /*
4704 * make sure that the address range covered by this pmd is not
4705 * unmapped from other threads.
4706 */
4707 if (!pmd_huge(*pmd))
4708 goto out;
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07004709 pte = huge_ptep_get((pte_t *)pmd);
4710 if (pte_present(pte)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07004711 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004712 if (flags & FOLL_GET)
4713 get_page(page);
4714 } else {
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07004715 if (is_hugetlb_entry_migration(pte)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004716 spin_unlock(ptl);
4717 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4718 goto retry;
4719 }
4720 /*
4721 * hwpoisoned entry is treated as no_page_table in
4722 * follow_page_mask().
4723 */
4724 }
4725out:
4726 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01004727 return page;
4728}
4729
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004730struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004731follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004732 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004733{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004734 if (flags & FOLL_GET)
4735 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01004736
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004737 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01004738}
4739
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -07004740struct page * __weak
4741follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4742{
4743 if (flags & FOLL_GET)
4744 return NULL;
4745
4746 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4747}
4748
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004749bool isolate_huge_page(struct page *page, struct list_head *list)
4750{
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004751 bool ret = true;
4752
Sasha Levin309381fea2014-01-23 15:52:54 -08004753 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004754 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004755 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4756 ret = false;
4757 goto unlock;
4758 }
4759 clear_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004760 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004761unlock:
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004762 spin_unlock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004763 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004764}
4765
4766void putback_active_hugepage(struct page *page)
4767{
Sasha Levin309381fea2014-01-23 15:52:54 -08004768 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004769 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004770 set_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004771 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4772 spin_unlock(&hugetlb_lock);
4773 put_page(page);
4774}