Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * Generic hugetlb support. |
Nadia Yvette Chambers | 6d49e35 | 2012-12-06 10:39:54 +0100 | [diff] [blame] | 3 | * (C) Nadia Yvette Chambers, April 2004 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5 | #include <linux/list.h> |
| 6 | #include <linux/init.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 7 | #include <linux/mm.h> |
Alexey Dobriyan | e1759c2 | 2008-10-15 23:50:22 +0400 | [diff] [blame] | 8 | #include <linux/seq_file.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 9 | #include <linux/sysctl.h> |
| 10 | #include <linux/highmem.h> |
Andrea Arcangeli | cddb8a5 | 2008-07-28 15:46:29 -0700 | [diff] [blame] | 11 | #include <linux/mmu_notifier.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 12 | #include <linux/nodemask.h> |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 13 | #include <linux/pagemap.h> |
Christoph Lameter | 5da7ca8 | 2006-01-06 00:10:46 -0800 | [diff] [blame] | 14 | #include <linux/mempolicy.h> |
Gideon Israel Dsouza | 3b32123 | 2014-04-07 15:37:26 -0700 | [diff] [blame] | 15 | #include <linux/compiler.h> |
Christoph Lameter | aea47ff | 2006-01-08 01:00:57 -0800 | [diff] [blame] | 16 | #include <linux/cpuset.h> |
David Gibson | 3935baa | 2006-03-22 00:08:53 -0800 | [diff] [blame] | 17 | #include <linux/mutex.h> |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 18 | #include <linux/bootmem.h> |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 19 | #include <linux/sysfs.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 20 | #include <linux/slab.h> |
Ingo Molnar | 174cd4b | 2017-02-02 19:15:33 +0100 | [diff] [blame] | 21 | #include <linux/sched/signal.h> |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 22 | #include <linux/rmap.h> |
Naoya Horiguchi | fd6a03e | 2010-05-28 09:29:21 +0900 | [diff] [blame] | 23 | #include <linux/swap.h> |
| 24 | #include <linux/swapops.h> |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 25 | #include <linux/jhash.h> |
Linus Torvalds | d660668 | 2008-08-06 12:04:54 -0700 | [diff] [blame] | 26 | |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 27 | #include <asm/page.h> |
| 28 | #include <asm/pgtable.h> |
Aneesh Kumar K.V | 24669e5 | 2012-07-31 16:42:03 -0700 | [diff] [blame] | 29 | #include <asm/tlb.h> |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 30 | |
Aneesh Kumar K.V | 24669e5 | 2012-07-31 16:42:03 -0700 | [diff] [blame] | 31 | #include <linux/io.h> |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 32 | #include <linux/hugetlb.h> |
Aneesh Kumar K.V | 9dd540e | 2012-07-31 16:42:15 -0700 | [diff] [blame] | 33 | #include <linux/hugetlb_cgroup.h> |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 34 | #include <linux/node.h> |
Mike Kravetz | 1a1aad8 | 2017-02-22 15:43:01 -0800 | [diff] [blame] | 35 | #include <linux/userfaultfd_k.h> |
Nick Piggin | 7835e98 | 2006-03-22 00:08:40 -0800 | [diff] [blame] | 36 | #include "internal.h" |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 37 | |
Andrey Ryabinin | 753162c | 2015-02-10 14:11:36 -0800 | [diff] [blame] | 38 | int hugepages_treat_as_movable; |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 39 | |
Aneesh Kumar K.V | c3f38a3 | 2012-07-31 16:42:10 -0700 | [diff] [blame] | 40 | int hugetlb_max_hstate __read_mostly; |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 41 | unsigned int default_hstate_idx; |
| 42 | struct hstate hstates[HUGE_MAX_HSTATE]; |
Naoya Horiguchi | 641844f | 2015-06-24 16:56:59 -0700 | [diff] [blame] | 43 | /* |
| 44 | * Minimum page order among possible hugepage sizes, set to a proper value |
| 45 | * at boot time. |
| 46 | */ |
| 47 | static unsigned int minimum_order __read_mostly = UINT_MAX; |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 48 | |
Jon Tollefson | 53ba51d | 2008-07-23 21:27:52 -0700 | [diff] [blame] | 49 | __initdata LIST_HEAD(huge_boot_pages); |
| 50 | |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 51 | /* for command line parsing */ |
| 52 | static struct hstate * __initdata parsed_hstate; |
| 53 | static unsigned long __initdata default_hstate_max_huge_pages; |
Nick Piggin | e11bfbf | 2008-07-23 21:27:52 -0700 | [diff] [blame] | 54 | static unsigned long __initdata default_hstate_size; |
Vaishali Thakkar | 9fee021 | 2016-05-19 17:11:04 -0700 | [diff] [blame] | 55 | static bool __initdata parsed_valid_hugepagesz = true; |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 56 | |
David Gibson | 3935baa | 2006-03-22 00:08:53 -0800 | [diff] [blame] | 57 | /* |
Naoya Horiguchi | 31caf66 | 2013-09-11 14:21:59 -0700 | [diff] [blame] | 58 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
| 59 | * free_huge_pages, and surplus_huge_pages. |
David Gibson | 3935baa | 2006-03-22 00:08:53 -0800 | [diff] [blame] | 60 | */ |
Aneesh Kumar K.V | c3f38a3 | 2012-07-31 16:42:10 -0700 | [diff] [blame] | 61 | DEFINE_SPINLOCK(hugetlb_lock); |
Eric Paris | 0bd0f9f | 2005-11-21 21:32:28 -0800 | [diff] [blame] | 62 | |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 63 | /* |
| 64 | * Serializes faults on the same logical page. This is used to |
| 65 | * prevent spurious OOMs when the hugepage pool is fully utilized. |
| 66 | */ |
| 67 | static int num_fault_mutexes; |
Mike Kravetz | c672c7f | 2015-09-08 15:01:35 -0700 | [diff] [blame] | 68 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 69 | |
Mike Kravetz | 7ca02d0a | 2015-04-15 16:13:42 -0700 | [diff] [blame] | 70 | /* Forward declaration */ |
| 71 | static int hugetlb_acct_memory(struct hstate *h, long delta); |
| 72 | |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 73 | static 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 Kravetz | 7ca02d0a | 2015-04-15 16:13:42 -0700 | [diff] [blame] | 80 | * 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 Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 86 | kfree(spool); |
Mike Kravetz | 7ca02d0a | 2015-04-15 16:13:42 -0700 | [diff] [blame] | 87 | } |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 88 | } |
| 89 | |
Mike Kravetz | 7ca02d0a | 2015-04-15 16:13:42 -0700 | [diff] [blame] | 90 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
| 91 | long min_hpages) |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 92 | { |
| 93 | struct hugepage_subpool *spool; |
| 94 | |
Mike Kravetz | c6a9182 | 2015-04-15 16:13:36 -0700 | [diff] [blame] | 95 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 96 | if (!spool) |
| 97 | return NULL; |
| 98 | |
| 99 | spin_lock_init(&spool->lock); |
| 100 | spool->count = 1; |
Mike Kravetz | 7ca02d0a | 2015-04-15 16:13:42 -0700 | [diff] [blame] | 101 | 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 Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 110 | |
| 111 | return spool; |
| 112 | } |
| 113 | |
| 114 | void 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 Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 122 | /* |
| 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 | */ |
| 130 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 131 | long delta) |
| 132 | { |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 133 | long ret = delta; |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 134 | |
| 135 | if (!spool) |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 136 | return ret; |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 137 | |
| 138 | spin_lock(&spool->lock); |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 139 | |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 140 | 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 Kravetz | 09a95e2 | 2016-05-19 17:11:01 -0700 | [diff] [blame] | 149 | /* minimum size accounting */ |
| 150 | if (spool->min_hpages != -1 && spool->rsv_hpages) { |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 151 | 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 | |
| 164 | unlock_ret: |
| 165 | spin_unlock(&spool->lock); |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 166 | return ret; |
| 167 | } |
| 168 | |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 169 | /* |
| 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 | */ |
| 175 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 176 | long delta) |
| 177 | { |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 178 | long ret = delta; |
| 179 | |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 180 | if (!spool) |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 181 | return delta; |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 182 | |
| 183 | spin_lock(&spool->lock); |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 184 | |
| 185 | if (spool->max_hpages != -1) /* maximum size accounting */ |
| 186 | spool->used_hpages -= delta; |
| 187 | |
Mike Kravetz | 09a95e2 | 2016-05-19 17:11:01 -0700 | [diff] [blame] | 188 | /* minimum size accounting */ |
| 189 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 190 | 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 Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 204 | unlock_or_release_subpool(spool); |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 205 | |
| 206 | return ret; |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 207 | } |
| 208 | |
| 209 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) |
| 210 | { |
| 211 | return HUGETLBFS_SB(inode->i_sb)->spool; |
| 212 | } |
| 213 | |
| 214 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) |
| 215 | { |
Al Viro | 496ad9a | 2013-01-23 17:07:38 -0500 | [diff] [blame] | 216 | return subpool_inode(file_inode(vma->vm_file)); |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 217 | } |
| 218 | |
Andy Whitcroft | e7c4b0b | 2008-07-23 21:27:26 -0700 | [diff] [blame] | 219 | /* |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 220 | * Region tracking -- allows tracking of reservations and instantiated pages |
| 221 | * across the pages in a mapping. |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 222 | * |
Mike Kravetz | 1dd308a | 2015-06-24 16:57:52 -0700 | [diff] [blame] | 223 | * 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 Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 237 | */ |
| 238 | struct file_region { |
| 239 | struct list_head link; |
| 240 | long from; |
| 241 | long to; |
| 242 | }; |
| 243 | |
Mike Kravetz | 1dd308a | 2015-06-24 16:57:52 -0700 | [diff] [blame] | 244 | /* |
| 245 | * Add the huge page range represented by [f, t) to the reserve |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 246 | * 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 Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 254 | * |
| 255 | * Return the number of new huge pages added to the map. This |
| 256 | * number is greater than or equal to zero. |
Mike Kravetz | 1dd308a | 2015-06-24 16:57:52 -0700 | [diff] [blame] | 257 | */ |
Joonsoo Kim | 1406ec9 | 2014-04-03 14:47:26 -0700 | [diff] [blame] | 258 | static long region_add(struct resv_map *resv, long f, long t) |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 259 | { |
Joonsoo Kim | 1406ec9 | 2014-04-03 14:47:26 -0700 | [diff] [blame] | 260 | struct list_head *head = &resv->regions; |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 261 | struct file_region *rg, *nrg, *trg; |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 262 | long add = 0; |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 263 | |
Davidlohr Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 264 | spin_lock(&resv->lock); |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 265 | /* 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 Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 270 | /* |
| 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 Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 292 | /* 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 Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 310 | /* 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 Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 315 | list_del(&rg->link); |
| 316 | kfree(rg); |
| 317 | } |
| 318 | } |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 319 | |
| 320 | add += (nrg->from - f); /* Added to beginning of region */ |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 321 | nrg->from = f; |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 322 | add += t - nrg->to; /* Added to end of region */ |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 323 | nrg->to = t; |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 324 | |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 325 | out_locked: |
| 326 | resv->adds_in_progress--; |
Davidlohr Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 327 | spin_unlock(&resv->lock); |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 328 | VM_BUG_ON(add < 0); |
| 329 | return add; |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 330 | } |
| 331 | |
Mike Kravetz | 1dd308a | 2015-06-24 16:57:52 -0700 | [diff] [blame] | 332 | /* |
| 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 Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 345 | * 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 Kravetz | 1dd308a | 2015-06-24 16:57:52 -0700 | [diff] [blame] | 353 | */ |
Joonsoo Kim | 1406ec9 | 2014-04-03 14:47:26 -0700 | [diff] [blame] | 354 | static long region_chg(struct resv_map *resv, long f, long t) |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 355 | { |
Joonsoo Kim | 1406ec9 | 2014-04-03 14:47:26 -0700 | [diff] [blame] | 356 | struct list_head *head = &resv->regions; |
Davidlohr Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 357 | struct file_region *rg, *nrg = NULL; |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 358 | long chg = 0; |
| 359 | |
Davidlohr Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 360 | retry: |
| 361 | spin_lock(&resv->lock); |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 362 | retry_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 Kravetz | dbe409e | 2015-12-11 13:40:52 -0800 | [diff] [blame] | 378 | if (!trg) { |
| 379 | kfree(nrg); |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 380 | return -ENOMEM; |
Mike Kravetz | dbe409e | 2015-12-11 13:40:52 -0800 | [diff] [blame] | 381 | } |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 382 | |
| 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 Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 389 | /* 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 Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 398 | if (!nrg) { |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 399 | resv->adds_in_progress--; |
Davidlohr Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 400 | spin_unlock(&resv->lock); |
| 401 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); |
| 402 | if (!nrg) |
| 403 | return -ENOMEM; |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 404 | |
Davidlohr Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 405 | 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 Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 414 | } |
| 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 Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 426 | goto out; |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 427 | |
Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 428 | /* We overlap with this area, if it extends further than |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 429 | * 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 Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 437 | |
| 438 | out: |
| 439 | spin_unlock(&resv->lock); |
| 440 | /* We already know we raced and no longer need the new region */ |
| 441 | kfree(nrg); |
| 442 | return chg; |
| 443 | out_nrg: |
| 444 | spin_unlock(&resv->lock); |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 445 | return chg; |
| 446 | } |
| 447 | |
Mike Kravetz | 1dd308a | 2015-06-24 16:57:52 -0700 | [diff] [blame] | 448 | /* |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 449 | * 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 | */ |
| 459 | static 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 Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 468 | * 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 Kravetz | 1dd308a | 2015-06-24 16:57:52 -0700 | [diff] [blame] | 480 | */ |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 481 | static long region_del(struct resv_map *resv, long f, long t) |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 482 | { |
Joonsoo Kim | 1406ec9 | 2014-04-03 14:47:26 -0700 | [diff] [blame] | 483 | struct list_head *head = &resv->regions; |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 484 | struct file_region *rg, *trg; |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 485 | struct file_region *nrg = NULL; |
| 486 | long del = 0; |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 487 | |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 488 | retry: |
Davidlohr Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 489 | spin_lock(&resv->lock); |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 490 | list_for_each_entry_safe(rg, trg, head, link) { |
Mike Kravetz | dbe409e | 2015-12-11 13:40:52 -0800 | [diff] [blame] | 491 | /* |
| 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 Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 499 | continue; |
Mike Kravetz | dbe409e | 2015-12-11 13:40:52 -0800 | [diff] [blame] | 500 | |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 501 | if (rg->from >= t) |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 502 | break; |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 503 | |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 504 | 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 Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 555 | } |
| 556 | |
Davidlohr Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 557 | spin_unlock(&resv->lock); |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 558 | kfree(nrg); |
| 559 | return del; |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 560 | } |
| 561 | |
Mike Kravetz | 1dd308a | 2015-06-24 16:57:52 -0700 | [diff] [blame] | 562 | /* |
Mike Kravetz | b5cec28 | 2015-09-08 15:01:41 -0700 | [diff] [blame] | 563 | * 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 jiang | 72e2936 | 2016-10-07 17:02:01 -0700 | [diff] [blame] | 571 | void hugetlb_fix_reserve_counts(struct inode *inode) |
Mike Kravetz | b5cec28 | 2015-09-08 15:01:41 -0700 | [diff] [blame] | 572 | { |
| 573 | struct hugepage_subpool *spool = subpool_inode(inode); |
| 574 | long rsv_adjust; |
| 575 | |
| 576 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); |
zhong jiang | 72e2936 | 2016-10-07 17:02:01 -0700 | [diff] [blame] | 577 | if (rsv_adjust) { |
Mike Kravetz | b5cec28 | 2015-09-08 15:01:41 -0700 | [diff] [blame] | 578 | struct hstate *h = hstate_inode(inode); |
| 579 | |
| 580 | hugetlb_acct_memory(h, 1); |
| 581 | } |
| 582 | } |
| 583 | |
| 584 | /* |
Mike Kravetz | 1dd308a | 2015-06-24 16:57:52 -0700 | [diff] [blame] | 585 | * Count and return the number of huge pages in the reserve map |
| 586 | * that intersect with the range [f, t). |
| 587 | */ |
Joonsoo Kim | 1406ec9 | 2014-04-03 14:47:26 -0700 | [diff] [blame] | 588 | static long region_count(struct resv_map *resv, long f, long t) |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 589 | { |
Joonsoo Kim | 1406ec9 | 2014-04-03 14:47:26 -0700 | [diff] [blame] | 590 | struct list_head *head = &resv->regions; |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 591 | struct file_region *rg; |
| 592 | long chg = 0; |
| 593 | |
Davidlohr Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 594 | spin_lock(&resv->lock); |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 595 | /* Locate each segment we overlap with, and count that overlap. */ |
| 596 | list_for_each_entry(rg, head, link) { |
Wang Sheng-Hui | f2135a4 | 2012-05-29 15:06:17 -0700 | [diff] [blame] | 597 | long seg_from; |
| 598 | long seg_to; |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 599 | |
| 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 Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 610 | spin_unlock(&resv->lock); |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 611 | |
| 612 | return chg; |
| 613 | } |
| 614 | |
Andy Whitcroft | 9682290 | 2008-07-23 21:27:29 -0700 | [diff] [blame] | 615 | /* |
Andy Whitcroft | e7c4b0b | 2008-07-23 21:27:26 -0700 | [diff] [blame] | 616 | * Convert the address within this vma to the page offset within |
Andy Whitcroft | e7c4b0b | 2008-07-23 21:27:26 -0700 | [diff] [blame] | 617 | * the mapping, in pagecache page units; huge pages here. |
| 618 | */ |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 619 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
| 620 | struct vm_area_struct *vma, unsigned long address) |
Andy Whitcroft | e7c4b0b | 2008-07-23 21:27:26 -0700 | [diff] [blame] | 621 | { |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 622 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
| 623 | (vma->vm_pgoff >> huge_page_order(h)); |
Andy Whitcroft | e7c4b0b | 2008-07-23 21:27:26 -0700 | [diff] [blame] | 624 | } |
| 625 | |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 626 | pgoff_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 Williams | dee4107 | 2016-05-14 12:20:44 -0700 | [diff] [blame] | 631 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 632 | |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 633 | /* |
Mel Gorman | 08fba69 | 2009-01-06 14:38:53 -0800 | [diff] [blame] | 634 | * 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 | */ |
| 637 | unsigned 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 Li | 2415cf1 | 2013-07-03 15:02:43 -0700 | [diff] [blame] | 646 | return 1UL << huge_page_shift(hstate); |
Mel Gorman | 08fba69 | 2009-01-06 14:38:53 -0800 | [diff] [blame] | 647 | } |
Joerg Roedel | f340ca0 | 2009-06-19 15:16:22 +0200 | [diff] [blame] | 648 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
Mel Gorman | 08fba69 | 2009-01-06 14:38:53 -0800 | [diff] [blame] | 649 | |
| 650 | /* |
Mel Gorman | 3340289 | 2009-01-06 14:38:54 -0800 | [diff] [blame] | 651 | * 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 |
| 657 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
| 658 | { |
| 659 | return vma_kernel_pagesize(vma); |
| 660 | } |
| 661 | #endif |
| 662 | |
| 663 | /* |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 664 | * 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 Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 670 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 671 | |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 672 | /* |
| 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 Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 681 | * |
| 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 Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 690 | */ |
Andy Whitcroft | e7c4b0b | 2008-07-23 21:27:26 -0700 | [diff] [blame] | 691 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
| 692 | { |
| 693 | return (unsigned long)vma->vm_private_data; |
| 694 | } |
| 695 | |
| 696 | static 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 Kim | 9119a41 | 2014-04-03 14:47:25 -0700 | [diff] [blame] | 702 | struct resv_map *resv_map_alloc(void) |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 703 | { |
| 704 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 705 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
| 706 | |
| 707 | if (!resv_map || !rg) { |
| 708 | kfree(resv_map); |
| 709 | kfree(rg); |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 710 | return NULL; |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 711 | } |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 712 | |
| 713 | kref_init(&resv_map->refs); |
Davidlohr Bueso | 7b24d86 | 2014-04-03 14:47:27 -0700 | [diff] [blame] | 714 | spin_lock_init(&resv_map->lock); |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 715 | INIT_LIST_HEAD(&resv_map->regions); |
| 716 | |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 717 | 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 Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 723 | return resv_map; |
| 724 | } |
| 725 | |
Joonsoo Kim | 9119a41 | 2014-04-03 14:47:25 -0700 | [diff] [blame] | 726 | void resv_map_release(struct kref *ref) |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 727 | { |
| 728 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 729 | struct list_head *head = &resv_map->region_cache; |
| 730 | struct file_region *rg, *trg; |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 731 | |
| 732 | /* Clear out any active regions before we release the map. */ |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 733 | region_del(resv_map, 0, LONG_MAX); |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 734 | |
| 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 Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 743 | kfree(resv_map); |
| 744 | } |
| 745 | |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 746 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
| 747 | { |
| 748 | return inode->i_mapping->private_data; |
| 749 | } |
| 750 | |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 751 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 752 | { |
Sasha Levin | 81d1b09 | 2014-10-09 15:28:10 -0700 | [diff] [blame] | 753 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 754 | 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 Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 761 | return (struct resv_map *)(get_vma_private_data(vma) & |
| 762 | ~HPAGE_RESV_MASK); |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 763 | } |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 764 | } |
| 765 | |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 766 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 767 | { |
Sasha Levin | 81d1b09 | 2014-10-09 15:28:10 -0700 | [diff] [blame] | 768 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
| 769 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 770 | |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 771 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
| 772 | HPAGE_RESV_MASK) | (unsigned long)map); |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 773 | } |
| 774 | |
| 775 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) |
| 776 | { |
Sasha Levin | 81d1b09 | 2014-10-09 15:28:10 -0700 | [diff] [blame] | 777 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
| 778 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); |
Andy Whitcroft | e7c4b0b | 2008-07-23 21:27:26 -0700 | [diff] [blame] | 779 | |
| 780 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 781 | } |
| 782 | |
| 783 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) |
| 784 | { |
Sasha Levin | 81d1b09 | 2014-10-09 15:28:10 -0700 | [diff] [blame] | 785 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
Andy Whitcroft | e7c4b0b | 2008-07-23 21:27:26 -0700 | [diff] [blame] | 786 | |
| 787 | return (get_vma_private_data(vma) & flag) != 0; |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 788 | } |
| 789 | |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 790 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 791 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
| 792 | { |
Sasha Levin | 81d1b09 | 2014-10-09 15:28:10 -0700 | [diff] [blame] | 793 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
Mel Gorman | f83a275 | 2009-05-28 14:34:40 -0700 | [diff] [blame] | 794 | if (!(vma->vm_flags & VM_MAYSHARE)) |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 795 | vma->vm_private_data = (void *)0; |
| 796 | } |
| 797 | |
| 798 | /* Returns true if the VMA has associated reserve pages */ |
Nicholas Krause | 559ec2f | 2015-09-04 15:48:27 -0700 | [diff] [blame] | 799 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 800 | { |
Joonsoo Kim | af0ed73 | 2013-09-11 14:21:18 -0700 | [diff] [blame] | 801 | 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 Krause | 559ec2f | 2015-09-04 15:48:27 -0700 | [diff] [blame] | 812 | return true; |
Joonsoo Kim | af0ed73 | 2013-09-11 14:21:18 -0700 | [diff] [blame] | 813 | else |
Nicholas Krause | 559ec2f | 2015-09-04 15:48:27 -0700 | [diff] [blame] | 814 | return false; |
Joonsoo Kim | af0ed73 | 2013-09-11 14:21:18 -0700 | [diff] [blame] | 815 | } |
Joonsoo Kim | a63884e | 2013-09-11 14:21:07 -0700 | [diff] [blame] | 816 | |
| 817 | /* Shared mappings always use reserves */ |
Mike Kravetz | 1fb1b0e | 2015-09-08 15:01:44 -0700 | [diff] [blame] | 818 | 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 Kim | a63884e | 2013-09-11 14:21:07 -0700 | [diff] [blame] | 831 | |
| 832 | /* |
| 833 | * Only the process that called mmap() has reserves for |
| 834 | * private mappings. |
| 835 | */ |
Mike Kravetz | 67961f9 | 2016-06-08 15:33:42 -0700 | [diff] [blame] | 836 | 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 Kim | a63884e | 2013-09-11 14:21:07 -0700 | [diff] [blame] | 857 | |
Nicholas Krause | 559ec2f | 2015-09-04 15:48:27 -0700 | [diff] [blame] | 858 | return false; |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 859 | } |
| 860 | |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 861 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 862 | { |
| 863 | int nid = page_to_nid(page); |
Aneesh Kumar K.V | 0edaecf | 2012-07-31 16:42:07 -0700 | [diff] [blame] | 864 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 865 | h->free_huge_pages++; |
| 866 | h->free_huge_pages_node[nid]++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 867 | } |
| 868 | |
Anshuman Khandual | 94310cb | 2017-07-06 15:38:38 -0700 | [diff] [blame] | 869 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
Naoya Horiguchi | bf50bab | 2010-09-08 10:19:33 +0900 | [diff] [blame] | 870 | { |
| 871 | struct page *page; |
| 872 | |
Naoya Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 873 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
Naoya Horiguchi | 243abd5 | 2017-07-10 15:47:32 -0700 | [diff] [blame] | 874 | if (!PageHWPoison(page)) |
Naoya Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 875 | 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 Horiguchi | bf50bab | 2010-09-08 10:19:33 +0900 | [diff] [blame] | 881 | return NULL; |
Aneesh Kumar K.V | 0edaecf | 2012-07-31 16:42:07 -0700 | [diff] [blame] | 882 | list_move(&page->lru, &h->hugepage_activelist); |
Naoya Horiguchi | a9869b8 | 2010-09-08 10:19:37 +0900 | [diff] [blame] | 883 | set_page_refcounted(page); |
Naoya Horiguchi | bf50bab | 2010-09-08 10:19:33 +0900 | [diff] [blame] | 884 | h->free_huge_pages--; |
| 885 | h->free_huge_pages_node[nid]--; |
| 886 | return page; |
| 887 | } |
| 888 | |
Anshuman Khandual | 94310cb | 2017-07-06 15:38:38 -0700 | [diff] [blame] | 889 | static 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 Horiguchi | 86cdb46 | 2013-09-11 14:22:13 -0700 | [diff] [blame] | 905 | /* Movability of hugepages depends on migration support. */ |
| 906 | static inline gfp_t htlb_alloc_mask(struct hstate *h) |
| 907 | { |
Naoya Horiguchi | 100873d | 2014-06-04 16:10:56 -0700 | [diff] [blame] | 908 | if (hugepages_treat_as_movable || hugepage_migration_supported(h)) |
Naoya Horiguchi | 86cdb46 | 2013-09-11 14:22:13 -0700 | [diff] [blame] | 909 | return GFP_HIGHUSER_MOVABLE; |
| 910 | else |
| 911 | return GFP_HIGHUSER; |
| 912 | } |
| 913 | |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 914 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
| 915 | struct vm_area_struct *vma, |
Joonsoo Kim | af0ed73 | 2013-09-11 14:21:18 -0700 | [diff] [blame] | 916 | unsigned long address, int avoid_reserve, |
| 917 | long chg) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 918 | { |
Konstantin Khlebnikov | b1c12cb | 2012-04-25 16:01:46 -0700 | [diff] [blame] | 919 | struct page *page = NULL; |
Lee Schermerhorn | 480eccf | 2007-09-18 22:46:47 -0700 | [diff] [blame] | 920 | struct mempolicy *mpol; |
Mel Gorman | 19770b3 | 2008-04-28 02:12:18 -0700 | [diff] [blame] | 921 | nodemask_t *nodemask; |
Vlastimil Babka | 04ec626 | 2017-07-06 15:40:03 -0700 | [diff] [blame] | 922 | gfp_t gfp_mask; |
| 923 | int nid; |
Miao Xie | c0ff745 | 2010-05-24 14:32:08 -0700 | [diff] [blame] | 924 | struct zonelist *zonelist; |
Mel Gorman | dd1a239 | 2008-04-28 02:12:17 -0700 | [diff] [blame] | 925 | struct zone *zone; |
| 926 | struct zoneref *z; |
Mel Gorman | cc9a6c8 | 2012-03-21 16:34:11 -0700 | [diff] [blame] | 927 | unsigned int cpuset_mems_cookie; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 928 | |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 929 | /* |
| 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 Kim | af0ed73 | 2013-09-11 14:21:18 -0700 | [diff] [blame] | 934 | if (!vma_has_reserves(vma, chg) && |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 935 | h->free_huge_pages - h->resv_huge_pages == 0) |
Miao Xie | c0ff745 | 2010-05-24 14:32:08 -0700 | [diff] [blame] | 936 | goto err; |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 937 | |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 938 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 939 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
Justin P. Mattock | 6eab04a | 2011-04-08 19:49:08 -0700 | [diff] [blame] | 940 | goto err; |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 941 | |
Joonsoo Kim | 9966c4bb | 2013-09-11 14:20:50 -0700 | [diff] [blame] | 942 | retry_cpuset: |
Mel Gorman | d26914d | 2014-04-03 14:47:24 -0700 | [diff] [blame] | 943 | cpuset_mems_cookie = read_mems_allowed_begin(); |
Vlastimil Babka | 04ec626 | 2017-07-06 15:40:03 -0700 | [diff] [blame] | 944 | gfp_mask = htlb_alloc_mask(h); |
| 945 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); |
| 946 | zonelist = node_zonelist(nid, gfp_mask); |
Joonsoo Kim | 9966c4bb | 2013-09-11 14:20:50 -0700 | [diff] [blame] | 947 | |
Mel Gorman | 19770b3 | 2008-04-28 02:12:18 -0700 | [diff] [blame] | 948 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
| 949 | MAX_NR_ZONES - 1, nodemask) { |
Vlastimil Babka | 04ec626 | 2017-07-06 15:40:03 -0700 | [diff] [blame] | 950 | if (cpuset_zone_allowed(zone, gfp_mask)) { |
Naoya Horiguchi | bf50bab | 2010-09-08 10:19:33 +0900 | [diff] [blame] | 951 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
| 952 | if (page) { |
Joonsoo Kim | af0ed73 | 2013-09-11 14:21:18 -0700 | [diff] [blame] | 953 | if (avoid_reserve) |
| 954 | break; |
| 955 | if (!vma_has_reserves(vma, chg)) |
| 956 | break; |
| 957 | |
Joonsoo Kim | 07443a8 | 2013-09-11 14:21:58 -0700 | [diff] [blame] | 958 | SetPagePrivate(page); |
Joonsoo Kim | af0ed73 | 2013-09-11 14:21:18 -0700 | [diff] [blame] | 959 | h->resv_huge_pages--; |
Naoya Horiguchi | bf50bab | 2010-09-08 10:19:33 +0900 | [diff] [blame] | 960 | break; |
| 961 | } |
Andrew Morton | 3abf7af | 2007-07-19 01:49:08 -0700 | [diff] [blame] | 962 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 963 | } |
Mel Gorman | cc9a6c8 | 2012-03-21 16:34:11 -0700 | [diff] [blame] | 964 | |
| 965 | mpol_cond_put(mpol); |
Mel Gorman | d26914d | 2014-04-03 14:47:24 -0700 | [diff] [blame] | 966 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
Mel Gorman | cc9a6c8 | 2012-03-21 16:34:11 -0700 | [diff] [blame] | 967 | goto retry_cpuset; |
| 968 | return page; |
| 969 | |
Miao Xie | c0ff745 | 2010-05-24 14:32:08 -0700 | [diff] [blame] | 970 | err: |
Mel Gorman | cc9a6c8 | 2012-03-21 16:34:11 -0700 | [diff] [blame] | 971 | return NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 972 | } |
| 973 | |
Luiz Capitulino | 1cac6f2 | 2014-06-04 16:07:11 -0700 | [diff] [blame] | 974 | /* |
| 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 | */ |
| 981 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) |
| 982 | { |
Andrew Morton | 0edaf86 | 2016-05-19 17:10:58 -0700 | [diff] [blame] | 983 | nid = next_node_in(nid, *nodes_allowed); |
Luiz Capitulino | 1cac6f2 | 2014-06-04 16:07:11 -0700 | [diff] [blame] | 984 | VM_BUG_ON(nid >= MAX_NUMNODES); |
| 985 | |
| 986 | return nid; |
| 987 | } |
| 988 | |
| 989 | static 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 | */ |
| 1002 | static 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 | */ |
| 1021 | static 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.V | e1073d1 | 2017-07-06 15:39:17 -0700 | [diff] [blame] | 1045 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1046 | static void destroy_compound_gigantic_page(struct page *page, |
Kirill A. Shutemov | d00181b | 2015-11-06 16:29:57 -0800 | [diff] [blame] | 1047 | unsigned int order) |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1048 | { |
| 1049 | int i; |
| 1050 | int nr_pages = 1 << order; |
| 1051 | struct page *p = page + 1; |
| 1052 | |
Gerald Schaefer | c8cc708 | 2016-06-24 14:50:04 -0700 | [diff] [blame] | 1053 | atomic_set(compound_mapcount_ptr(page), 0); |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1054 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
Kirill A. Shutemov | 1d798ca | 2015-11-06 16:29:54 -0800 | [diff] [blame] | 1055 | clear_compound_head(p); |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1056 | set_page_refcounted(p); |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1057 | } |
| 1058 | |
| 1059 | set_compound_order(page, 0); |
| 1060 | __ClearPageHead(page); |
| 1061 | } |
| 1062 | |
Kirill A. Shutemov | d00181b | 2015-11-06 16:29:57 -0800 | [diff] [blame] | 1063 | static void free_gigantic_page(struct page *page, unsigned int order) |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1064 | { |
| 1065 | free_contig_range(page_to_pfn(page), 1 << order); |
| 1066 | } |
| 1067 | |
| 1068 | static int __alloc_gigantic_page(unsigned long start_pfn, |
| 1069 | unsigned long nr_pages) |
| 1070 | { |
| 1071 | unsigned long end_pfn = start_pfn + nr_pages; |
Lucas Stach | ca96b62 | 2017-02-24 14:58:37 -0800 | [diff] [blame] | 1072 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, |
| 1073 | GFP_KERNEL); |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1074 | } |
| 1075 | |
Joonsoo Kim | f44b2dda8 | 2016-05-19 17:12:03 -0700 | [diff] [blame] | 1076 | static bool pfn_range_valid_gigantic(struct zone *z, |
| 1077 | unsigned long start_pfn, unsigned long nr_pages) |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1078 | { |
| 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 Kim | f44b2dda8 | 2016-05-19 17:12:03 -0700 | [diff] [blame] | 1088 | if (page_zone(page) != z) |
| 1089 | return false; |
| 1090 | |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1091 | 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 | |
| 1104 | static 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. Shutemov | d00181b | 2015-11-06 16:29:57 -0800 | [diff] [blame] | 1111 | static struct page *alloc_gigantic_page(int nid, unsigned int order) |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1112 | { |
| 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 Kim | f44b2dda8 | 2016-05-19 17:12:03 -0700 | [diff] [blame] | 1123 | if (pfn_range_valid_gigantic(z, pfn, nr_pages)) { |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1124 | /* |
| 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 | |
| 1146 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); |
Kirill A. Shutemov | d00181b | 2015-11-06 16:29:57 -0800 | [diff] [blame] | 1147 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1148 | |
| 1149 | static 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 | |
| 1162 | static 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.V | e1073d1 | 2017-07-06 15:39:17 -0700 | [diff] [blame] | 1177 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1178 | static inline bool gigantic_page_supported(void) { return false; } |
Kirill A. Shutemov | d00181b | 2015-11-06 16:29:57 -0800 | [diff] [blame] | 1179 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1180 | static inline void destroy_compound_gigantic_page(struct page *page, |
Kirill A. Shutemov | d00181b | 2015-11-06 16:29:57 -0800 | [diff] [blame] | 1181 | unsigned int order) { } |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1182 | static inline int alloc_fresh_gigantic_page(struct hstate *h, |
| 1183 | nodemask_t *nodes_allowed) { return 0; } |
| 1184 | #endif |
| 1185 | |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1186 | static void update_and_free_page(struct hstate *h, struct page *page) |
Adam Litke | 6af2acb | 2007-10-16 01:26:16 -0700 | [diff] [blame] | 1187 | { |
| 1188 | int i; |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1189 | |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1190 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
| 1191 | return; |
Andy Whitcroft | 18229df | 2008-11-06 12:53:27 -0800 | [diff] [blame] | 1192 | |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1193 | 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 Forbes | 32f8452 | 2011-07-25 17:12:14 -0700 | [diff] [blame] | 1196 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
| 1197 | 1 << PG_referenced | 1 << PG_dirty | |
Luiz Capitulino | a7407a2 | 2014-06-04 16:07:09 -0700 | [diff] [blame] | 1198 | 1 << PG_active | 1 << PG_private | |
| 1199 | 1 << PG_writeback); |
Adam Litke | 6af2acb | 2007-10-16 01:26:16 -0700 | [diff] [blame] | 1200 | } |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1201 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
Kirill A. Shutemov | f1e6155 | 2015-11-06 16:29:50 -0800 | [diff] [blame] | 1202 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
Adam Litke | 6af2acb | 2007-10-16 01:26:16 -0700 | [diff] [blame] | 1203 | set_page_refcounted(page); |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1204 | 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 Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1208 | __free_pages(page, huge_page_order(h)); |
| 1209 | } |
Adam Litke | 6af2acb | 2007-10-16 01:26:16 -0700 | [diff] [blame] | 1210 | } |
| 1211 | |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 1212 | struct 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 Horiguchi | bcc5422 | 2015-04-15 16:14:38 -0700 | [diff] [blame] | 1223 | /* |
| 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 | */ |
| 1229 | bool 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 */ |
| 1236 | static void set_page_huge_active(struct page *page) |
| 1237 | { |
| 1238 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); |
| 1239 | SetPagePrivate(&page[1]); |
| 1240 | } |
| 1241 | |
| 1242 | static void clear_page_huge_active(struct page *page) |
| 1243 | { |
| 1244 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); |
| 1245 | ClearPagePrivate(&page[1]); |
| 1246 | } |
| 1247 | |
Atsushi Kumagai | 8f1d26d | 2014-07-30 16:08:39 -0700 | [diff] [blame] | 1248 | void free_huge_page(struct page *page) |
David Gibson | 27a85ef | 2006-03-22 00:08:56 -0800 | [diff] [blame] | 1249 | { |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1250 | /* |
| 1251 | * Can't pass hstate in here because it is called from the |
| 1252 | * compound page destructor. |
| 1253 | */ |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 1254 | struct hstate *h = page_hstate(page); |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 1255 | int nid = page_to_nid(page); |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 1256 | struct hugepage_subpool *spool = |
| 1257 | (struct hugepage_subpool *)page_private(page); |
Joonsoo Kim | 07443a8 | 2013-09-11 14:21:58 -0700 | [diff] [blame] | 1258 | bool restore_reserve; |
David Gibson | 27a85ef | 2006-03-22 00:08:56 -0800 | [diff] [blame] | 1259 | |
Andy Whitcroft | e5df70a | 2008-02-23 15:23:32 -0800 | [diff] [blame] | 1260 | set_page_private(page, 0); |
Mel Gorman | 23be746 | 2010-04-23 13:17:56 -0400 | [diff] [blame] | 1261 | page->mapping = NULL; |
Mike Kravetz | b4330af | 2016-02-05 15:36:38 -0800 | [diff] [blame] | 1262 | VM_BUG_ON_PAGE(page_count(page), page); |
| 1263 | VM_BUG_ON_PAGE(page_mapcount(page), page); |
Joonsoo Kim | 07443a8 | 2013-09-11 14:21:58 -0700 | [diff] [blame] | 1264 | restore_reserve = PagePrivate(page); |
Joonsoo Kim | 16c794b | 2013-10-16 13:46:48 -0700 | [diff] [blame] | 1265 | ClearPagePrivate(page); |
David Gibson | 27a85ef | 2006-03-22 00:08:56 -0800 | [diff] [blame] | 1266 | |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 1267 | /* |
| 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 Gibson | 27a85ef | 2006-03-22 00:08:56 -0800 | [diff] [blame] | 1275 | spin_lock(&hugetlb_lock); |
Naoya Horiguchi | bcc5422 | 2015-04-15 16:14:38 -0700 | [diff] [blame] | 1276 | clear_page_huge_active(page); |
Aneesh Kumar K.V | 6d76dcf | 2012-07-31 16:42:18 -0700 | [diff] [blame] | 1277 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
| 1278 | pages_per_huge_page(h), page); |
Joonsoo Kim | 07443a8 | 2013-09-11 14:21:58 -0700 | [diff] [blame] | 1279 | if (restore_reserve) |
| 1280 | h->resv_huge_pages++; |
| 1281 | |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 1282 | if (h->surplus_huge_pages_node[nid]) { |
Aneesh Kumar K.V | 0edaecf | 2012-07-31 16:42:07 -0700 | [diff] [blame] | 1283 | /* remove the page from active list */ |
| 1284 | list_del(&page->lru); |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1285 | update_and_free_page(h, page); |
| 1286 | h->surplus_huge_pages--; |
| 1287 | h->surplus_huge_pages_node[nid]--; |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 1288 | } else { |
Will Deacon | 5d3a551 | 2012-10-08 16:29:32 -0700 | [diff] [blame] | 1289 | arch_clear_hugepage_flags(page); |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1290 | enqueue_huge_page(h, page); |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 1291 | } |
David Gibson | 27a85ef | 2006-03-22 00:08:56 -0800 | [diff] [blame] | 1292 | spin_unlock(&hugetlb_lock); |
| 1293 | } |
| 1294 | |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1295 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
Andi Kleen | b7ba30c | 2008-07-23 21:27:40 -0700 | [diff] [blame] | 1296 | { |
Aneesh Kumar K.V | 0edaecf | 2012-07-31 16:42:07 -0700 | [diff] [blame] | 1297 | INIT_LIST_HEAD(&page->lru); |
Kirill A. Shutemov | f1e6155 | 2015-11-06 16:29:50 -0800 | [diff] [blame] | 1298 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
Andi Kleen | b7ba30c | 2008-07-23 21:27:40 -0700 | [diff] [blame] | 1299 | spin_lock(&hugetlb_lock); |
Aneesh Kumar K.V | 9dd540e | 2012-07-31 16:42:15 -0700 | [diff] [blame] | 1300 | set_hugetlb_cgroup(page, NULL); |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1301 | h->nr_huge_pages++; |
| 1302 | h->nr_huge_pages_node[nid]++; |
Andi Kleen | b7ba30c | 2008-07-23 21:27:40 -0700 | [diff] [blame] | 1303 | spin_unlock(&hugetlb_lock); |
| 1304 | put_page(page); /* free it into the hugepage allocator */ |
| 1305 | } |
| 1306 | |
Kirill A. Shutemov | d00181b | 2015-11-06 16:29:57 -0800 | [diff] [blame] | 1307 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
Wu Fengguang | 20a0307 | 2009-06-16 15:32:22 -0700 | [diff] [blame] | 1308 | { |
| 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 Arcangeli | ef5a22b | 2013-10-16 13:46:56 -0700 | [diff] [blame] | 1315 | __ClearPageReserved(page); |
Kirill A. Shutemov | de09d31 | 2016-01-15 16:51:42 -0800 | [diff] [blame] | 1316 | __SetPageHead(page); |
Wu Fengguang | 20a0307 | 2009-06-16 15:32:22 -0700 | [diff] [blame] | 1317 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
Andrea Arcangeli | ef5a22b | 2013-10-16 13:46:56 -0700 | [diff] [blame] | 1318 | /* |
| 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 Song | 58a84aa | 2011-12-08 14:34:18 -0800 | [diff] [blame] | 1331 | set_page_count(p, 0); |
Kirill A. Shutemov | 1d798ca | 2015-11-06 16:29:54 -0800 | [diff] [blame] | 1332 | set_compound_head(p, page); |
Wu Fengguang | 20a0307 | 2009-06-16 15:32:22 -0700 | [diff] [blame] | 1333 | } |
Mike Kravetz | b4330af | 2016-02-05 15:36:38 -0800 | [diff] [blame] | 1334 | atomic_set(compound_mapcount_ptr(page), -1); |
Wu Fengguang | 20a0307 | 2009-06-16 15:32:22 -0700 | [diff] [blame] | 1335 | } |
| 1336 | |
Andrew Morton | 7795912 | 2012-10-08 16:34:11 -0700 | [diff] [blame] | 1337 | /* |
| 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 Fengguang | 20a0307 | 2009-06-16 15:32:22 -0700 | [diff] [blame] | 1342 | int PageHuge(struct page *page) |
| 1343 | { |
Wu Fengguang | 20a0307 | 2009-06-16 15:32:22 -0700 | [diff] [blame] | 1344 | if (!PageCompound(page)) |
| 1345 | return 0; |
| 1346 | |
| 1347 | page = compound_head(page); |
Kirill A. Shutemov | f1e6155 | 2015-11-06 16:29:50 -0800 | [diff] [blame] | 1348 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
Wu Fengguang | 20a0307 | 2009-06-16 15:32:22 -0700 | [diff] [blame] | 1349 | } |
Naoya Horiguchi | 43131e1 | 2010-05-28 09:29:22 +0900 | [diff] [blame] | 1350 | EXPORT_SYMBOL_GPL(PageHuge); |
| 1351 | |
Andrea Arcangeli | 27c73ae | 2013-11-21 14:32:02 -0800 | [diff] [blame] | 1352 | /* |
| 1353 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for |
| 1354 | * normal or transparent huge pages. |
| 1355 | */ |
| 1356 | int PageHeadHuge(struct page *page_head) |
| 1357 | { |
Andrea Arcangeli | 27c73ae | 2013-11-21 14:32:02 -0800 | [diff] [blame] | 1358 | if (!PageHead(page_head)) |
| 1359 | return 0; |
| 1360 | |
Andrew Morton | 758f66a | 2014-01-21 15:48:57 -0800 | [diff] [blame] | 1361 | return get_compound_page_dtor(page_head) == free_huge_page; |
Andrea Arcangeli | 27c73ae | 2013-11-21 14:32:02 -0800 | [diff] [blame] | 1362 | } |
Andrea Arcangeli | 27c73ae | 2013-11-21 14:32:02 -0800 | [diff] [blame] | 1363 | |
Zhang Yi | 13d60f4 | 2013-06-25 21:19:31 +0800 | [diff] [blame] | 1364 | pgoff_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 Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1381 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1382 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1383 | struct page *page; |
Joe Jin | f96efd5 | 2007-07-15 23:38:12 -0700 | [diff] [blame] | 1384 | |
Vlastimil Babka | 96db800 | 2015-09-08 15:03:50 -0700 | [diff] [blame] | 1385 | page = __alloc_pages_node(nid, |
Naoya Horiguchi | 86cdb46 | 2013-09-11 14:22:13 -0700 | [diff] [blame] | 1386 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
Nishanth Aravamudan | 551883a | 2008-04-29 00:58:26 -0700 | [diff] [blame] | 1387 | __GFP_REPEAT|__GFP_NOWARN, |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1388 | huge_page_order(h)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1389 | if (page) { |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1390 | prep_new_huge_page(h, page, nid); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1391 | } |
Nishanth Aravamudan | 63b4613 | 2007-10-16 01:26:24 -0700 | [diff] [blame] | 1392 | |
| 1393 | return page; |
| 1394 | } |
| 1395 | |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 1396 | static 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 Schermerhorn | e8c5c82 | 2009-09-21 17:01:22 -0700 | [diff] [blame] | 1418 | /* |
| 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 Schermerhorn | 6ae11b2 | 2009-12-14 17:58:16 -0800 | [diff] [blame] | 1424 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
| 1425 | bool acct_surplus) |
Lee Schermerhorn | e8c5c82 | 2009-09-21 17:01:22 -0700 | [diff] [blame] | 1426 | { |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 1427 | int nr_nodes, node; |
Lee Schermerhorn | e8c5c82 | 2009-09-21 17:01:22 -0700 | [diff] [blame] | 1428 | int ret = 0; |
| 1429 | |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 1430 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
Lee Schermerhorn | 685f345 | 2009-09-21 17:01:23 -0700 | [diff] [blame] | 1431 | /* |
| 1432 | * If we're returning unused surplus pages, only examine |
| 1433 | * nodes with surplus pages. |
| 1434 | */ |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 1435 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
| 1436 | !list_empty(&h->hugepage_freelists[node])) { |
Lee Schermerhorn | e8c5c82 | 2009-09-21 17:01:22 -0700 | [diff] [blame] | 1437 | struct page *page = |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 1438 | list_entry(h->hugepage_freelists[node].next, |
Lee Schermerhorn | e8c5c82 | 2009-09-21 17:01:22 -0700 | [diff] [blame] | 1439 | struct page, lru); |
| 1440 | list_del(&page->lru); |
| 1441 | h->free_huge_pages--; |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 1442 | h->free_huge_pages_node[node]--; |
Lee Schermerhorn | 685f345 | 2009-09-21 17:01:23 -0700 | [diff] [blame] | 1443 | if (acct_surplus) { |
| 1444 | h->surplus_huge_pages--; |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 1445 | h->surplus_huge_pages_node[node]--; |
Lee Schermerhorn | 685f345 | 2009-09-21 17:01:23 -0700 | [diff] [blame] | 1446 | } |
Lee Schermerhorn | e8c5c82 | 2009-09-21 17:01:22 -0700 | [diff] [blame] | 1447 | update_and_free_page(h, page); |
| 1448 | ret = 1; |
Lee Schermerhorn | 9a76db0 | 2009-12-14 17:58:15 -0800 | [diff] [blame] | 1449 | break; |
Lee Schermerhorn | e8c5c82 | 2009-09-21 17:01:22 -0700 | [diff] [blame] | 1450 | } |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 1451 | } |
Lee Schermerhorn | e8c5c82 | 2009-09-21 17:01:22 -0700 | [diff] [blame] | 1452 | |
| 1453 | return ret; |
| 1454 | } |
| 1455 | |
Naoya Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 1456 | /* |
| 1457 | * Dissolve a given free hugepage into free buddy pages. This function does |
Gerald Schaefer | 082d5b6 | 2016-10-07 17:01:10 -0700 | [diff] [blame] | 1458 | * 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 Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 1461 | */ |
Anshuman Khandual | c3114a8 | 2017-07-10 15:47:41 -0700 | [diff] [blame] | 1462 | int dissolve_free_huge_page(struct page *page) |
Naoya Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 1463 | { |
Gerald Schaefer | 082d5b6 | 2016-10-07 17:01:10 -0700 | [diff] [blame] | 1464 | int rc = 0; |
| 1465 | |
Naoya Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 1466 | spin_lock(&hugetlb_lock); |
| 1467 | if (PageHuge(page) && !page_count(page)) { |
Gerald Schaefer | 2247bb3 | 2016-10-07 17:01:07 -0700 | [diff] [blame] | 1468 | struct page *head = compound_head(page); |
| 1469 | struct hstate *h = page_hstate(head); |
| 1470 | int nid = page_to_nid(head); |
Gerald Schaefer | 082d5b6 | 2016-10-07 17:01:10 -0700 | [diff] [blame] | 1471 | if (h->free_huge_pages - h->resv_huge_pages == 0) { |
| 1472 | rc = -EBUSY; |
| 1473 | goto out; |
| 1474 | } |
Anshuman Khandual | c3114a8 | 2017-07-10 15:47:41 -0700 | [diff] [blame] | 1475 | /* |
| 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 Schaefer | 2247bb3 | 2016-10-07 17:01:07 -0700 | [diff] [blame] | 1483 | list_del(&head->lru); |
Naoya Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 1484 | h->free_huge_pages--; |
| 1485 | h->free_huge_pages_node[nid]--; |
zhong jiang | c1470b3 | 2016-08-11 15:32:55 -0700 | [diff] [blame] | 1486 | h->max_huge_pages--; |
Gerald Schaefer | 2247bb3 | 2016-10-07 17:01:07 -0700 | [diff] [blame] | 1487 | update_and_free_page(h, head); |
Naoya Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 1488 | } |
Gerald Schaefer | 082d5b6 | 2016-10-07 17:01:10 -0700 | [diff] [blame] | 1489 | out: |
Naoya Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 1490 | spin_unlock(&hugetlb_lock); |
Gerald Schaefer | 082d5b6 | 2016-10-07 17:01:10 -0700 | [diff] [blame] | 1491 | return rc; |
Naoya Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 1492 | } |
| 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 Schaefer | 2247bb3 | 2016-10-07 17:01:07 -0700 | [diff] [blame] | 1497 | * Note that this will dissolve a free gigantic hugepage completely, if any |
| 1498 | * part of it lies within the given range. |
Gerald Schaefer | 082d5b6 | 2016-10-07 17:01:10 -0700 | [diff] [blame] | 1499 | * 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 Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 1501 | */ |
Gerald Schaefer | 082d5b6 | 2016-10-07 17:01:10 -0700 | [diff] [blame] | 1502 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
Naoya Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 1503 | { |
Naoya Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 1504 | unsigned long pfn; |
Gerald Schaefer | eb03aa0 | 2016-10-07 17:01:13 -0700 | [diff] [blame] | 1505 | struct page *page; |
Gerald Schaefer | 082d5b6 | 2016-10-07 17:01:10 -0700 | [diff] [blame] | 1506 | int rc = 0; |
Naoya Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 1507 | |
Li Zhong | d017763 | 2014-08-06 16:07:56 -0700 | [diff] [blame] | 1508 | if (!hugepages_supported()) |
Gerald Schaefer | 082d5b6 | 2016-10-07 17:01:10 -0700 | [diff] [blame] | 1509 | return rc; |
Li Zhong | d017763 | 2014-08-06 16:07:56 -0700 | [diff] [blame] | 1510 | |
Gerald Schaefer | eb03aa0 | 2016-10-07 17:01:13 -0700 | [diff] [blame] | 1511 | 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 Schaefer | 082d5b6 | 2016-10-07 17:01:10 -0700 | [diff] [blame] | 1519 | |
| 1520 | return rc; |
Naoya Horiguchi | c8721bb | 2013-09-11 14:22:09 -0700 | [diff] [blame] | 1521 | } |
| 1522 | |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1523 | /* |
| 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 | */ |
| 1532 | static 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 Hansen | e0ec90e | 2015-11-05 18:50:20 -0800 | [diff] [blame] | 1541 | * 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 Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1547 | */ |
Dave Hansen | e0ec90e | 2015-11-05 18:50:20 -0800 | [diff] [blame] | 1548 | if (!IS_ENABLED(CONFIG_NUMA) || !vma) { |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1549 | /* |
| 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 Hansen | e0ec90e | 2015-11-05 18:50:20 -0800 | [diff] [blame] | 1565 | * allocate a huge page with it. We will only reach this |
| 1566 | * when CONFIG_NUMA=y. |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1567 | */ |
| 1568 | do { |
| 1569 | struct page *page; |
| 1570 | struct mempolicy *mpol; |
Vlastimil Babka | 04ec626 | 2017-07-06 15:40:03 -0700 | [diff] [blame] | 1571 | int nid; |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1572 | nodemask_t *nodemask; |
| 1573 | |
| 1574 | cpuset_mems_cookie = read_mems_allowed_begin(); |
Vlastimil Babka | 04ec626 | 2017-07-06 15:40:03 -0700 | [diff] [blame] | 1575 | nid = huge_node(vma, addr, gfp, &mpol, &nodemask); |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1576 | mpol_cond_put(mpol); |
Vlastimil Babka | 04ec626 | 2017-07-06 15:40:03 -0700 | [diff] [blame] | 1577 | page = __alloc_pages_nodemask(gfp, order, nid, nodemask); |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1578 | 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 | */ |
| 1597 | static struct page *__alloc_buddy_huge_page(struct hstate *h, |
| 1598 | struct vm_area_struct *vma, unsigned long addr, int nid) |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 1599 | { |
| 1600 | struct page *page; |
Naoya Horiguchi | bf50bab | 2010-09-08 10:19:33 +0900 | [diff] [blame] | 1601 | unsigned int r_nid; |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 1602 | |
Luiz Capitulino | bae7f4a | 2014-06-04 16:07:08 -0700 | [diff] [blame] | 1603 | if (hstate_is_gigantic(h)) |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 1604 | return NULL; |
| 1605 | |
Nishanth Aravamudan | d1c3fb1 | 2007-12-17 16:20:12 -0800 | [diff] [blame] | 1606 | /* |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1607 | * 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 Hansen | e0ec90e | 2015-11-05 18:50:20 -0800 | [diff] [blame] | 1612 | VM_WARN_ON_ONCE(addr == -1); |
| 1613 | VM_WARN_ON_ONCE(nid != NUMA_NO_NODE); |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1614 | } |
| 1615 | /* |
Nishanth Aravamudan | d1c3fb1 | 2007-12-17 16:20:12 -0800 | [diff] [blame] | 1616 | * 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 Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1639 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
Nishanth Aravamudan | d1c3fb1 | 2007-12-17 16:20:12 -0800 | [diff] [blame] | 1640 | spin_unlock(&hugetlb_lock); |
| 1641 | return NULL; |
| 1642 | } else { |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1643 | h->nr_huge_pages++; |
| 1644 | h->surplus_huge_pages++; |
Nishanth Aravamudan | d1c3fb1 | 2007-12-17 16:20:12 -0800 | [diff] [blame] | 1645 | } |
| 1646 | spin_unlock(&hugetlb_lock); |
| 1647 | |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1648 | page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid); |
Nishanth Aravamudan | d1c3fb1 | 2007-12-17 16:20:12 -0800 | [diff] [blame] | 1649 | |
| 1650 | spin_lock(&hugetlb_lock); |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 1651 | if (page) { |
Aneesh Kumar K.V | 0edaecf | 2012-07-31 16:42:07 -0700 | [diff] [blame] | 1652 | INIT_LIST_HEAD(&page->lru); |
Naoya Horiguchi | bf50bab | 2010-09-08 10:19:33 +0900 | [diff] [blame] | 1653 | r_nid = page_to_nid(page); |
Kirill A. Shutemov | f1e6155 | 2015-11-06 16:29:50 -0800 | [diff] [blame] | 1654 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
Aneesh Kumar K.V | 9dd540e | 2012-07-31 16:42:15 -0700 | [diff] [blame] | 1655 | set_hugetlb_cgroup(page, NULL); |
Nishanth Aravamudan | d1c3fb1 | 2007-12-17 16:20:12 -0800 | [diff] [blame] | 1656 | /* |
| 1657 | * We incremented the global counters already |
| 1658 | */ |
Naoya Horiguchi | bf50bab | 2010-09-08 10:19:33 +0900 | [diff] [blame] | 1659 | h->nr_huge_pages_node[r_nid]++; |
| 1660 | h->surplus_huge_pages_node[r_nid]++; |
Adam Litke | 3b11630 | 2008-04-28 02:13:06 -0700 | [diff] [blame] | 1661 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
Nishanth Aravamudan | d1c3fb1 | 2007-12-17 16:20:12 -0800 | [diff] [blame] | 1662 | } else { |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1663 | h->nr_huge_pages--; |
| 1664 | h->surplus_huge_pages--; |
Adam Litke | 3b11630 | 2008-04-28 02:13:06 -0700 | [diff] [blame] | 1665 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 1666 | } |
Nishanth Aravamudan | d1c3fb1 | 2007-12-17 16:20:12 -0800 | [diff] [blame] | 1667 | spin_unlock(&hugetlb_lock); |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 1668 | |
| 1669 | return page; |
| 1670 | } |
| 1671 | |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1672 | /* |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1673 | * 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 Hansen | e0ec90e | 2015-11-05 18:50:20 -0800 | [diff] [blame] | 1677 | static |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1678 | struct 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 Hansen | e0ec90e | 2015-11-05 18:50:20 -0800 | [diff] [blame] | 1688 | static |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1689 | struct 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 Horiguchi | bf50bab | 2010-09-08 10:19:33 +0900 | [diff] [blame] | 1696 | * 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 | */ |
| 1700 | struct page *alloc_huge_page_node(struct hstate *h, int nid) |
| 1701 | { |
Joonsoo Kim | 4ef9184 | 2013-09-11 14:21:51 -0700 | [diff] [blame] | 1702 | struct page *page = NULL; |
Naoya Horiguchi | bf50bab | 2010-09-08 10:19:33 +0900 | [diff] [blame] | 1703 | |
| 1704 | spin_lock(&hugetlb_lock); |
Joonsoo Kim | 4ef9184 | 2013-09-11 14:21:51 -0700 | [diff] [blame] | 1705 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
| 1706 | page = dequeue_huge_page_node(h, nid); |
Naoya Horiguchi | bf50bab | 2010-09-08 10:19:33 +0900 | [diff] [blame] | 1707 | spin_unlock(&hugetlb_lock); |
| 1708 | |
Aneesh Kumar K.V | 94ae8ba | 2012-07-31 16:42:35 -0700 | [diff] [blame] | 1709 | if (!page) |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1710 | page = __alloc_buddy_huge_page_no_mpol(h, nid); |
Naoya Horiguchi | bf50bab | 2010-09-08 10:19:33 +0900 | [diff] [blame] | 1711 | |
| 1712 | return page; |
| 1713 | } |
| 1714 | |
| 1715 | /* |
Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 1716 | * Increase the hugetlb pool such that it can accommodate a reservation |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1717 | * of size 'delta'. |
| 1718 | */ |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1719 | static int gather_surplus_pages(struct hstate *h, int delta) |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1720 | { |
| 1721 | struct list_head surplus_list; |
| 1722 | struct page *page, *tmp; |
| 1723 | int ret, i; |
| 1724 | int needed, allocated; |
Hillf Danton | 28073b0 | 2012-03-21 16:34:00 -0700 | [diff] [blame] | 1725 | bool alloc_ok = true; |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1726 | |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1727 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
Adam Litke | ac09b3a | 2008-03-04 14:29:38 -0800 | [diff] [blame] | 1728 | if (needed <= 0) { |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1729 | h->resv_huge_pages += delta; |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1730 | return 0; |
Adam Litke | ac09b3a | 2008-03-04 14:29:38 -0800 | [diff] [blame] | 1731 | } |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1732 | |
| 1733 | allocated = 0; |
| 1734 | INIT_LIST_HEAD(&surplus_list); |
| 1735 | |
| 1736 | ret = -ENOMEM; |
| 1737 | retry: |
| 1738 | spin_unlock(&hugetlb_lock); |
| 1739 | for (i = 0; i < needed; i++) { |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 1740 | page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE); |
Hillf Danton | 28073b0 | 2012-03-21 16:34:00 -0700 | [diff] [blame] | 1741 | if (!page) { |
| 1742 | alloc_ok = false; |
| 1743 | break; |
| 1744 | } |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1745 | list_add(&page->lru, &surplus_list); |
| 1746 | } |
Hillf Danton | 28073b0 | 2012-03-21 16:34:00 -0700 | [diff] [blame] | 1747 | allocated += i; |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1748 | |
| 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 Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1754 | needed = (h->resv_huge_pages + delta) - |
| 1755 | (h->free_huge_pages + allocated); |
Hillf Danton | 28073b0 | 2012-03-21 16:34:00 -0700 | [diff] [blame] | 1756 | 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 Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1766 | /* |
| 1767 | * The surplus_list now contains _at_least_ the number of extra pages |
Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 1768 | * needed to accommodate the reservation. Add the appropriate number |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1769 | * of pages to the hugetlb pool and free the extras back to the buddy |
Adam Litke | ac09b3a | 2008-03-04 14:29:38 -0800 | [diff] [blame] | 1770 | * 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 Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1773 | */ |
| 1774 | needed += allocated; |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1775 | h->resv_huge_pages += delta; |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1776 | ret = 0; |
Naoya Horiguchi | a9869b8 | 2010-09-08 10:19:37 +0900 | [diff] [blame] | 1777 | |
Adam Litke | 19fc3f0 | 2008-04-28 02:12:20 -0700 | [diff] [blame] | 1778 | /* Free the needed pages to the hugetlb pool */ |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1779 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
Adam Litke | 19fc3f0 | 2008-04-28 02:12:20 -0700 | [diff] [blame] | 1780 | if ((--needed) < 0) |
| 1781 | break; |
Naoya Horiguchi | a9869b8 | 2010-09-08 10:19:37 +0900 | [diff] [blame] | 1782 | /* |
| 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 Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 1787 | VM_BUG_ON_PAGE(page_count(page), page); |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1788 | enqueue_huge_page(h, page); |
Adam Litke | 19fc3f0 | 2008-04-28 02:12:20 -0700 | [diff] [blame] | 1789 | } |
Hillf Danton | 28073b0 | 2012-03-21 16:34:00 -0700 | [diff] [blame] | 1790 | free: |
Hillf Danton | b0365c8 | 2011-12-28 15:57:16 -0800 | [diff] [blame] | 1791 | spin_unlock(&hugetlb_lock); |
Adam Litke | 19fc3f0 | 2008-04-28 02:12:20 -0700 | [diff] [blame] | 1792 | |
| 1793 | /* Free unnecessary surplus pages to the buddy allocator */ |
Joonsoo Kim | c0d934b | 2013-09-11 14:21:02 -0700 | [diff] [blame] | 1794 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
| 1795 | put_page(page); |
Naoya Horiguchi | a9869b8 | 2010-09-08 10:19:37 +0900 | [diff] [blame] | 1796 | spin_lock(&hugetlb_lock); |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1797 | |
| 1798 | return ret; |
| 1799 | } |
| 1800 | |
| 1801 | /* |
Mike Kravetz | e5bbc8a | 2017-01-10 16:58:27 -0800 | [diff] [blame] | 1802 | * 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 Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1814 | */ |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1815 | static void return_unused_surplus_pages(struct hstate *h, |
| 1816 | unsigned long unused_resv_pages) |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1817 | { |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1818 | unsigned long nr_pages; |
| 1819 | |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 1820 | /* Cannot return gigantic pages currently */ |
Luiz Capitulino | bae7f4a | 2014-06-04 16:07:08 -0700 | [diff] [blame] | 1821 | if (hstate_is_gigantic(h)) |
Mike Kravetz | e5bbc8a | 2017-01-10 16:58:27 -0800 | [diff] [blame] | 1822 | goto out; |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 1823 | |
Mike Kravetz | e5bbc8a | 2017-01-10 16:58:27 -0800 | [diff] [blame] | 1824 | /* |
| 1825 | * Part (or even all) of the reservation could have been backed |
| 1826 | * by pre-allocated pages. Only free surplus pages. |
| 1827 | */ |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1828 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1829 | |
Lee Schermerhorn | 685f345 | 2009-09-21 17:01:23 -0700 | [diff] [blame] | 1830 | /* |
| 1831 | * We want to release as many surplus pages as possible, spread |
Lee Schermerhorn | 9b5e5d0 | 2009-12-14 17:58:32 -0800 | [diff] [blame] | 1832 | * 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 Kravetz | e5bbc8a | 2017-01-10 16:58:27 -0800 | [diff] [blame] | 1837 | * |
| 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 Schermerhorn | 685f345 | 2009-09-21 17:01:23 -0700 | [diff] [blame] | 1841 | */ |
| 1842 | while (nr_pages--) { |
Mike Kravetz | e5bbc8a | 2017-01-10 16:58:27 -0800 | [diff] [blame] | 1843 | h->resv_huge_pages--; |
| 1844 | unused_resv_pages--; |
Lai Jiangshan | 8cebfcd | 2012-12-12 13:51:36 -0800 | [diff] [blame] | 1845 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
Mike Kravetz | e5bbc8a | 2017-01-10 16:58:27 -0800 | [diff] [blame] | 1846 | goto out; |
Mizuma, Masayoshi | 7848a4b | 2014-04-18 15:07:18 -0700 | [diff] [blame] | 1847 | cond_resched_lock(&hugetlb_lock); |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1848 | } |
Mike Kravetz | e5bbc8a | 2017-01-10 16:58:27 -0800 | [diff] [blame] | 1849 | |
| 1850 | out: |
| 1851 | /* Fully uncommit the reservation */ |
| 1852 | h->resv_huge_pages -= unused_resv_pages; |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 1853 | } |
| 1854 | |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 1855 | |
Andy Whitcroft | c37f9fb | 2008-07-23 21:27:30 -0700 | [diff] [blame] | 1856 | /* |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 1857 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 1858 | * are used by the huge page allocation routines to manage reservations. |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 1859 | * |
| 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 Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 1865 | * 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 Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 1868 | * |
| 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 Kravetz | 96b96a9 | 2016-11-10 10:46:32 -0800 | [diff] [blame] | 1874 | * |
| 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 Whitcroft | c37f9fb | 2008-07-23 21:27:30 -0700 | [diff] [blame] | 1879 | */ |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 1880 | enum vma_resv_mode { |
| 1881 | VMA_NEEDS_RESV, |
| 1882 | VMA_COMMIT_RESV, |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 1883 | VMA_END_RESV, |
Mike Kravetz | 96b96a9 | 2016-11-10 10:46:32 -0800 | [diff] [blame] | 1884 | VMA_ADD_RESV, |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 1885 | }; |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 1886 | static long __vma_reservation_common(struct hstate *h, |
| 1887 | struct vm_area_struct *vma, unsigned long addr, |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 1888 | enum vma_resv_mode mode) |
Andy Whitcroft | c37f9fb | 2008-07-23 21:27:30 -0700 | [diff] [blame] | 1889 | { |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 1890 | struct resv_map *resv; |
| 1891 | pgoff_t idx; |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 1892 | long ret; |
Andy Whitcroft | c37f9fb | 2008-07-23 21:27:30 -0700 | [diff] [blame] | 1893 | |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 1894 | resv = vma_resv_map(vma); |
| 1895 | if (!resv) |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 1896 | return 1; |
Andy Whitcroft | c37f9fb | 2008-07-23 21:27:30 -0700 | [diff] [blame] | 1897 | |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 1898 | idx = vma_hugecache_offset(h, vma, addr); |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 1899 | switch (mode) { |
| 1900 | case VMA_NEEDS_RESV: |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 1901 | ret = region_chg(resv, idx, idx + 1); |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 1902 | break; |
| 1903 | case VMA_COMMIT_RESV: |
| 1904 | ret = region_add(resv, idx, idx + 1); |
| 1905 | break; |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 1906 | case VMA_END_RESV: |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 1907 | region_abort(resv, idx, idx + 1); |
| 1908 | ret = 0; |
| 1909 | break; |
Mike Kravetz | 96b96a9 | 2016-11-10 10:46:32 -0800 | [diff] [blame] | 1910 | 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 Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 1918 | default: |
| 1919 | BUG(); |
| 1920 | } |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 1921 | |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 1922 | if (vma->vm_flags & VM_MAYSHARE) |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 1923 | return ret; |
Mike Kravetz | 67961f9 | 2016-06-08 15:33:42 -0700 | [diff] [blame] | 1924 | 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 Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 1943 | else |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 1944 | return ret < 0 ? ret : 0; |
Andy Whitcroft | c37f9fb | 2008-07-23 21:27:30 -0700 | [diff] [blame] | 1945 | } |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 1946 | |
| 1947 | static long vma_needs_reservation(struct hstate *h, |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 1948 | struct vm_area_struct *vma, unsigned long addr) |
Andy Whitcroft | c37f9fb | 2008-07-23 21:27:30 -0700 | [diff] [blame] | 1949 | { |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 1950 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 1951 | } |
Andy Whitcroft | c37f9fb | 2008-07-23 21:27:30 -0700 | [diff] [blame] | 1952 | |
Mike Kravetz | cf3ad20 | 2015-06-24 16:57:55 -0700 | [diff] [blame] | 1953 | static long vma_commit_reservation(struct hstate *h, |
| 1954 | struct vm_area_struct *vma, unsigned long addr) |
| 1955 | { |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 1956 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
| 1957 | } |
| 1958 | |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 1959 | static void vma_end_reservation(struct hstate *h, |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 1960 | struct vm_area_struct *vma, unsigned long addr) |
| 1961 | { |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 1962 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
Andy Whitcroft | c37f9fb | 2008-07-23 21:27:30 -0700 | [diff] [blame] | 1963 | } |
| 1964 | |
Mike Kravetz | 96b96a9 | 2016-11-10 10:46:32 -0800 | [diff] [blame] | 1965 | static 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 | */ |
| 1982 | static 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 Kravetz | 70c3547 | 2015-09-08 15:01:54 -0700 | [diff] [blame] | 2015 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 2016 | unsigned long addr, int avoid_reserve) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2017 | { |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 2018 | struct hugepage_subpool *spool = subpool_vma(vma); |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 2019 | struct hstate *h = hstate_vma(vma); |
Adam Litke | 348ea20 | 2007-11-14 16:59:37 -0800 | [diff] [blame] | 2020 | struct page *page; |
Mike Kravetz | d85f69b | 2015-09-08 15:01:47 -0700 | [diff] [blame] | 2021 | long map_chg, map_commit; |
| 2022 | long gbl_chg; |
Aneesh Kumar K.V | 6d76dcf | 2012-07-31 16:42:18 -0700 | [diff] [blame] | 2023 | int ret, idx; |
| 2024 | struct hugetlb_cgroup *h_cg; |
Adam Litke | 2fc39ce | 2007-11-14 16:59:39 -0800 | [diff] [blame] | 2025 | |
Aneesh Kumar K.V | 6d76dcf | 2012-07-31 16:42:18 -0700 | [diff] [blame] | 2026 | idx = hstate_index(h); |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 2027 | /* |
Mike Kravetz | d85f69b | 2015-09-08 15:01:47 -0700 | [diff] [blame] | 2028 | * 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 Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 2031 | */ |
Mike Kravetz | d85f69b | 2015-09-08 15:01:47 -0700 | [diff] [blame] | 2032 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
| 2033 | if (map_chg < 0) |
Aneesh Kumar K.V | 76dcee7 | 2012-07-31 16:41:57 -0700 | [diff] [blame] | 2034 | return ERR_PTR(-ENOMEM); |
Mike Kravetz | d85f69b | 2015-09-08 15:01:47 -0700 | [diff] [blame] | 2035 | |
| 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 Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 2046 | vma_end_reservation(h, vma, addr); |
Aneesh Kumar K.V | 76dcee7 | 2012-07-31 16:41:57 -0700 | [diff] [blame] | 2047 | return ERR_PTR(-ENOSPC); |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 2048 | } |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 2049 | |
Mike Kravetz | d85f69b | 2015-09-08 15:01:47 -0700 | [diff] [blame] | 2050 | /* |
| 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.V | 6d76dcf | 2012-07-31 16:42:18 -0700 | [diff] [blame] | 2062 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
Jianyu Zhan | 8f34af6 | 2014-06-04 16:10:36 -0700 | [diff] [blame] | 2063 | if (ret) |
| 2064 | goto out_subpool_put; |
| 2065 | |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 2066 | spin_lock(&hugetlb_lock); |
Mike Kravetz | d85f69b | 2015-09-08 15:01:47 -0700 | [diff] [blame] | 2067 | /* |
| 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 Kim | 81a6fca | 2013-09-11 14:20:58 -0700 | [diff] [blame] | 2073 | if (!page) { |
Aneesh Kumar K.V | 94ae8ba | 2012-07-31 16:42:35 -0700 | [diff] [blame] | 2074 | spin_unlock(&hugetlb_lock); |
Dave Hansen | 099730d | 2015-11-05 18:50:17 -0800 | [diff] [blame] | 2075 | page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); |
Jianyu Zhan | 8f34af6 | 2014-06-04 16:10:36 -0700 | [diff] [blame] | 2076 | if (!page) |
| 2077 | goto out_uncharge_cgroup; |
Naoya Horiguchi | a88c769 | 2015-12-11 13:40:24 -0800 | [diff] [blame] | 2078 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
| 2079 | SetPagePrivate(page); |
| 2080 | h->resv_huge_pages--; |
| 2081 | } |
Aneesh Kumar K.V | 79dbb23 | 2012-07-31 16:42:32 -0700 | [diff] [blame] | 2082 | spin_lock(&hugetlb_lock); |
| 2083 | list_move(&page->lru, &h->hugepage_activelist); |
Joonsoo Kim | 81a6fca | 2013-09-11 14:20:58 -0700 | [diff] [blame] | 2084 | /* Fall through */ |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 2085 | } |
Joonsoo Kim | 81a6fca | 2013-09-11 14:20:58 -0700 | [diff] [blame] | 2086 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
| 2087 | spin_unlock(&hugetlb_lock); |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 2088 | |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 2089 | set_page_private(page, (unsigned long)spool); |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 2090 | |
Mike Kravetz | d85f69b | 2015-09-08 15:01:47 -0700 | [diff] [blame] | 2091 | map_commit = vma_commit_reservation(h, vma, addr); |
| 2092 | if (unlikely(map_chg > map_commit)) { |
Mike Kravetz | 3303967 | 2015-06-24 16:57:58 -0700 | [diff] [blame] | 2093 | /* |
| 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 Litke | 90d8b7e | 2007-11-14 16:59:42 -0800 | [diff] [blame] | 2107 | return page; |
Jianyu Zhan | 8f34af6 | 2014-06-04 16:10:36 -0700 | [diff] [blame] | 2108 | |
| 2109 | out_uncharge_cgroup: |
| 2110 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); |
| 2111 | out_subpool_put: |
Mike Kravetz | d85f69b | 2015-09-08 15:01:47 -0700 | [diff] [blame] | 2112 | if (map_chg || avoid_reserve) |
Jianyu Zhan | 8f34af6 | 2014-06-04 16:10:36 -0700 | [diff] [blame] | 2113 | hugepage_subpool_put_pages(spool, 1); |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 2114 | vma_end_reservation(h, vma, addr); |
Jianyu Zhan | 8f34af6 | 2014-06-04 16:10:36 -0700 | [diff] [blame] | 2115 | return ERR_PTR(-ENOSPC); |
David Gibson | b45b5bd | 2006-03-22 00:08:55 -0800 | [diff] [blame] | 2116 | } |
| 2117 | |
Naoya Horiguchi | 74060e4 | 2013-09-11 14:22:06 -0700 | [diff] [blame] | 2118 | /* |
| 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 | */ |
| 2123 | struct 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 Gorcunov | 91f4766 | 2009-01-06 14:40:33 -0800 | [diff] [blame] | 2132 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2133 | { |
| 2134 | struct huge_bootmem_page *m; |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 2135 | int nr_nodes, node; |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2136 | |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 2137 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2138 | void *addr; |
| 2139 | |
Grygorii Strashko | 8b89a11 | 2014-01-21 15:50:36 -0800 | [diff] [blame] | 2140 | addr = memblock_virt_alloc_try_nid_nopanic( |
| 2141 | huge_page_size(h), huge_page_size(h), |
| 2142 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2143 | 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 Gorcunov | 91f4766 | 2009-01-06 14:40:33 -0800 | [diff] [blame] | 2150 | goto found; |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2151 | } |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2152 | } |
| 2153 | return 0; |
| 2154 | |
| 2155 | found: |
Luiz Capitulino | df994ea | 2014-12-12 16:55:21 -0800 | [diff] [blame] | 2156 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2157 | /* 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. Shutemov | d00181b | 2015-11-06 16:29:57 -0800 | [diff] [blame] | 2163 | static void __init prep_compound_huge_page(struct page *page, |
| 2164 | unsigned int order) |
Andy Whitcroft | 18229df | 2008-11-06 12:53:27 -0800 | [diff] [blame] | 2165 | { |
| 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 Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2172 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
| 2173 | static 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 Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2178 | struct hstate *h = m->hstate; |
Becky Bruce | ee8f248 | 2011-07-25 17:11:50 -0700 | [diff] [blame] | 2179 | struct page *page; |
| 2180 | |
| 2181 | #ifdef CONFIG_HIGHMEM |
| 2182 | page = pfn_to_page(m->phys >> PAGE_SHIFT); |
Grygorii Strashko | 8b89a11 | 2014-01-21 15:50:36 -0800 | [diff] [blame] | 2183 | memblock_free_late(__pa(m), |
| 2184 | sizeof(struct huge_bootmem_page)); |
Becky Bruce | ee8f248 | 2011-07-25 17:11:50 -0700 | [diff] [blame] | 2185 | #else |
| 2186 | page = virt_to_page(m); |
| 2187 | #endif |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2188 | WARN_ON(page_count(page) != 1); |
Andy Whitcroft | 18229df | 2008-11-06 12:53:27 -0800 | [diff] [blame] | 2189 | prep_compound_huge_page(page, h->order); |
Andrea Arcangeli | ef5a22b | 2013-10-16 13:46:56 -0700 | [diff] [blame] | 2190 | WARN_ON(PageReserved(page)); |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2191 | prep_new_huge_page(h, page, page_to_nid(page)); |
Rafael Aquini | b0320c7 | 2011-06-15 15:08:39 -0700 | [diff] [blame] | 2192 | /* |
| 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 Capitulino | bae7f4a | 2014-06-04 16:07:08 -0700 | [diff] [blame] | 2198 | if (hstate_is_gigantic(h)) |
Jiang Liu | 3dcc057 | 2013-07-03 15:03:21 -0700 | [diff] [blame] | 2199 | adjust_managed_page_count(page, 1 << h->order); |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2200 | } |
| 2201 | } |
| 2202 | |
Andi Kleen | 8faa8b0 | 2008-07-23 21:27:48 -0700 | [diff] [blame] | 2203 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2204 | { |
| 2205 | unsigned long i; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2206 | |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2207 | for (i = 0; i < h->max_huge_pages; ++i) { |
Luiz Capitulino | bae7f4a | 2014-06-04 16:07:08 -0700 | [diff] [blame] | 2208 | if (hstate_is_gigantic(h)) { |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2209 | if (!alloc_bootmem_huge_page(h)) |
| 2210 | break; |
Lee Schermerhorn | 9b5e5d0 | 2009-12-14 17:58:32 -0800 | [diff] [blame] | 2211 | } else if (!alloc_fresh_huge_page(h, |
Lai Jiangshan | 8cebfcd | 2012-12-12 13:51:36 -0800 | [diff] [blame] | 2212 | &node_states[N_MEMORY])) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2213 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2214 | } |
Andi Kleen | 8faa8b0 | 2008-07-23 21:27:48 -0700 | [diff] [blame] | 2215 | h->max_huge_pages = i; |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2216 | } |
| 2217 | |
| 2218 | static void __init hugetlb_init_hstates(void) |
| 2219 | { |
| 2220 | struct hstate *h; |
| 2221 | |
| 2222 | for_each_hstate(h) { |
Naoya Horiguchi | 641844f | 2015-06-24 16:56:59 -0700 | [diff] [blame] | 2223 | if (minimum_order > huge_page_order(h)) |
| 2224 | minimum_order = huge_page_order(h); |
| 2225 | |
Andi Kleen | 8faa8b0 | 2008-07-23 21:27:48 -0700 | [diff] [blame] | 2226 | /* oversize hugepages were init'ed in early boot */ |
Luiz Capitulino | bae7f4a | 2014-06-04 16:07:08 -0700 | [diff] [blame] | 2227 | if (!hstate_is_gigantic(h)) |
Andi Kleen | 8faa8b0 | 2008-07-23 21:27:48 -0700 | [diff] [blame] | 2228 | hugetlb_hstate_alloc_pages(h); |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2229 | } |
Naoya Horiguchi | 641844f | 2015-06-24 16:56:59 -0700 | [diff] [blame] | 2230 | VM_BUG_ON(minimum_order == UINT_MAX); |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2231 | } |
| 2232 | |
Andi Kleen | 4abd32d | 2008-07-23 21:27:49 -0700 | [diff] [blame] | 2233 | static 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 Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2244 | static void __init report_hugepages(void) |
| 2245 | { |
| 2246 | struct hstate *h; |
| 2247 | |
| 2248 | for_each_hstate(h) { |
Andi Kleen | 4abd32d | 2008-07-23 21:27:49 -0700 | [diff] [blame] | 2249 | char buf[32]; |
Andrew Morton | ffb22af | 2013-02-22 16:32:08 -0800 | [diff] [blame] | 2250 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
Andi Kleen | 4abd32d | 2008-07-23 21:27:49 -0700 | [diff] [blame] | 2251 | memfmt(buf, huge_page_size(h)), |
| 2252 | h->free_huge_pages); |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2253 | } |
| 2254 | } |
| 2255 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2256 | #ifdef CONFIG_HIGHMEM |
Lee Schermerhorn | 6ae11b2 | 2009-12-14 17:58:16 -0800 | [diff] [blame] | 2257 | static void try_to_free_low(struct hstate *h, unsigned long count, |
| 2258 | nodemask_t *nodes_allowed) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2259 | { |
Christoph Lameter | 4415cc8 | 2006-09-25 23:31:55 -0700 | [diff] [blame] | 2260 | int i; |
| 2261 | |
Luiz Capitulino | bae7f4a | 2014-06-04 16:07:08 -0700 | [diff] [blame] | 2262 | if (hstate_is_gigantic(h)) |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2263 | return; |
| 2264 | |
Lee Schermerhorn | 6ae11b2 | 2009-12-14 17:58:16 -0800 | [diff] [blame] | 2265 | for_each_node_mask(i, *nodes_allowed) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2266 | struct page *page, *next; |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 2267 | 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 Litke | 6b0c880 | 2007-10-16 01:26:23 -0700 | [diff] [blame] | 2270 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2271 | if (PageHighMem(page)) |
| 2272 | continue; |
| 2273 | list_del(&page->lru); |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2274 | update_and_free_page(h, page); |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 2275 | h->free_huge_pages--; |
| 2276 | h->free_huge_pages_node[page_to_nid(page)]--; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2277 | } |
| 2278 | } |
| 2279 | } |
| 2280 | #else |
Lee Schermerhorn | 6ae11b2 | 2009-12-14 17:58:16 -0800 | [diff] [blame] | 2281 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
| 2282 | nodemask_t *nodes_allowed) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2283 | { |
| 2284 | } |
| 2285 | #endif |
| 2286 | |
Wu Fengguang | 20a0307 | 2009-06-16 15:32:22 -0700 | [diff] [blame] | 2287 | /* |
| 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 Schermerhorn | 6ae11b2 | 2009-12-14 17:58:16 -0800 | [diff] [blame] | 2292 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
| 2293 | int delta) |
Wu Fengguang | 20a0307 | 2009-06-16 15:32:22 -0700 | [diff] [blame] | 2294 | { |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 2295 | int nr_nodes, node; |
Wu Fengguang | 20a0307 | 2009-06-16 15:32:22 -0700 | [diff] [blame] | 2296 | |
| 2297 | VM_BUG_ON(delta != -1 && delta != 1); |
Wu Fengguang | 20a0307 | 2009-06-16 15:32:22 -0700 | [diff] [blame] | 2298 | |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 2299 | 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 Schermerhorn | e8c5c82 | 2009-09-21 17:01:22 -0700 | [diff] [blame] | 2303 | } |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 2304 | } 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 Schermerhorn | e8c5c82 | 2009-09-21 17:01:22 -0700 | [diff] [blame] | 2309 | } |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 2310 | } |
| 2311 | return 0; |
Wu Fengguang | 20a0307 | 2009-06-16 15:32:22 -0700 | [diff] [blame] | 2312 | |
Joonsoo Kim | b226102 | 2013-09-11 14:21:00 -0700 | [diff] [blame] | 2313 | found: |
| 2314 | h->surplus_huge_pages += delta; |
| 2315 | h->surplus_huge_pages_node[node] += delta; |
| 2316 | return 1; |
Wu Fengguang | 20a0307 | 2009-06-16 15:32:22 -0700 | [diff] [blame] | 2317 | } |
| 2318 | |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 2319 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
Lee Schermerhorn | 6ae11b2 | 2009-12-14 17:58:16 -0800 | [diff] [blame] | 2320 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
| 2321 | nodemask_t *nodes_allowed) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2322 | { |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 2323 | unsigned long min_count, ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2324 | |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 2325 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2326 | return h->max_huge_pages; |
| 2327 | |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 2328 | /* |
| 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 Aravamudan | d1c3fb1 | 2007-12-17 16:20:12 -0800 | [diff] [blame] | 2332 | * |
Naoya Horiguchi | d15c7c0 | 2015-11-10 14:45:11 -0800 | [diff] [blame] | 2333 | * We might race with __alloc_buddy_huge_page() here and be unable |
Nishanth Aravamudan | d1c3fb1 | 2007-12-17 16:20:12 -0800 | [diff] [blame] | 2334 | * 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 Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 2338 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2339 | spin_lock(&hugetlb_lock); |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 2340 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
Lee Schermerhorn | 6ae11b2 | 2009-12-14 17:58:16 -0800 | [diff] [blame] | 2341 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 2342 | break; |
| 2343 | } |
| 2344 | |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 2345 | while (count > persistent_huge_pages(h)) { |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 2346 | /* |
| 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 He | 649920c | 2016-08-02 14:02:31 -0700 | [diff] [blame] | 2352 | |
| 2353 | /* yield cpu to avoid soft lockup */ |
| 2354 | cond_resched(); |
| 2355 | |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 2356 | 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 Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 2360 | spin_lock(&hugetlb_lock); |
| 2361 | if (!ret) |
| 2362 | goto out; |
| 2363 | |
Mel Gorman | 536240f2 | 2009-12-14 17:59:56 -0800 | [diff] [blame] | 2364 | /* Bail for signals. Probably ctrl-c from user */ |
| 2365 | if (signal_pending(current)) |
| 2366 | goto out; |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 2367 | } |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 2368 | |
| 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 Aravamudan | d1c3fb1 | 2007-12-17 16:20:12 -0800 | [diff] [blame] | 2375 | * |
| 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 Horiguchi | d15c7c0 | 2015-11-10 14:45:11 -0800 | [diff] [blame] | 2379 | * __alloc_buddy_huge_page() is checking the global counter, |
Nishanth Aravamudan | d1c3fb1 | 2007-12-17 16:20:12 -0800 | [diff] [blame] | 2380 | * 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 Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 2383 | */ |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 2384 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
Adam Litke | 6b0c880 | 2007-10-16 01:26:23 -0700 | [diff] [blame] | 2385 | min_count = max(count, min_count); |
Lee Schermerhorn | 6ae11b2 | 2009-12-14 17:58:16 -0800 | [diff] [blame] | 2386 | try_to_free_low(h, min_count, nodes_allowed); |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 2387 | while (min_count < persistent_huge_pages(h)) { |
Lee Schermerhorn | 6ae11b2 | 2009-12-14 17:58:16 -0800 | [diff] [blame] | 2388 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2389 | break; |
Mizuma, Masayoshi | 55f6714 | 2014-04-07 15:37:54 -0700 | [diff] [blame] | 2390 | cond_resched_lock(&hugetlb_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2391 | } |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 2392 | while (count < persistent_huge_pages(h)) { |
Lee Schermerhorn | 6ae11b2 | 2009-12-14 17:58:16 -0800 | [diff] [blame] | 2393 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 2394 | break; |
| 2395 | } |
| 2396 | out: |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 2397 | ret = persistent_huge_pages(h); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2398 | spin_unlock(&hugetlb_lock); |
Adam Litke | 7893d1d | 2007-10-16 01:26:18 -0700 | [diff] [blame] | 2399 | return ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2400 | } |
| 2401 | |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2402 | #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 | |
| 2409 | static struct kobject *hugepages_kobj; |
| 2410 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; |
| 2411 | |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2412 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
| 2413 | |
| 2414 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2415 | { |
| 2416 | int i; |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2417 | |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2418 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2419 | if (hstate_kobjs[i] == kobj) { |
| 2420 | if (nidp) |
| 2421 | *nidp = NUMA_NO_NODE; |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2422 | return &hstates[i]; |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2423 | } |
| 2424 | |
| 2425 | return kobj_to_node_hstate(kobj, nidp); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2426 | } |
| 2427 | |
Lee Schermerhorn | 06808b0 | 2009-12-14 17:58:21 -0800 | [diff] [blame] | 2428 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2429 | struct kobj_attribute *attr, char *buf) |
| 2430 | { |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2431 | 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 Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2442 | } |
Eric B Munson | adbe872 | 2011-01-13 15:47:27 -0800 | [diff] [blame] | 2443 | |
David Rientjes | 238d3c1 | 2014-08-06 16:06:51 -0700 | [diff] [blame] | 2444 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
| 2445 | struct hstate *h, int nid, |
| 2446 | unsigned long count, size_t len) |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2447 | { |
| 2448 | int err; |
David Rientjes | bad44b5 | 2009-12-14 17:58:38 -0800 | [diff] [blame] | 2449 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2450 | |
Luiz Capitulino | 944d9fe | 2014-06-04 16:07:13 -0700 | [diff] [blame] | 2451 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
Eric B Munson | adbe872 | 2011-01-13 15:47:27 -0800 | [diff] [blame] | 2452 | err = -EINVAL; |
| 2453 | goto out; |
| 2454 | } |
| 2455 | |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2456 | 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 Jiangshan | 8cebfcd | 2012-12-12 13:51:36 -0800 | [diff] [blame] | 2463 | nodes_allowed = &node_states[N_MEMORY]; |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2464 | } |
| 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 Jiangshan | 8cebfcd | 2012-12-12 13:51:36 -0800 | [diff] [blame] | 2473 | nodes_allowed = &node_states[N_MEMORY]; |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2474 | |
Lee Schermerhorn | 06808b0 | 2009-12-14 17:58:21 -0800 | [diff] [blame] | 2475 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2476 | |
Lai Jiangshan | 8cebfcd | 2012-12-12 13:51:36 -0800 | [diff] [blame] | 2477 | if (nodes_allowed != &node_states[N_MEMORY]) |
Lee Schermerhorn | 06808b0 | 2009-12-14 17:58:21 -0800 | [diff] [blame] | 2478 | NODEMASK_FREE(nodes_allowed); |
| 2479 | |
| 2480 | return len; |
Eric B Munson | adbe872 | 2011-01-13 15:47:27 -0800 | [diff] [blame] | 2481 | out: |
| 2482 | NODEMASK_FREE(nodes_allowed); |
| 2483 | return err; |
Lee Schermerhorn | 06808b0 | 2009-12-14 17:58:21 -0800 | [diff] [blame] | 2484 | } |
| 2485 | |
David Rientjes | 238d3c1 | 2014-08-06 16:06:51 -0700 | [diff] [blame] | 2486 | static 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 Schermerhorn | 06808b0 | 2009-12-14 17:58:21 -0800 | [diff] [blame] | 2503 | static 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 | |
| 2509 | static ssize_t nr_hugepages_store(struct kobject *kobj, |
| 2510 | struct kobj_attribute *attr, const char *buf, size_t len) |
| 2511 | { |
David Rientjes | 238d3c1 | 2014-08-06 16:06:51 -0700 | [diff] [blame] | 2512 | return nr_hugepages_store_common(false, kobj, buf, len); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2513 | } |
| 2514 | HSTATE_ATTR(nr_hugepages); |
| 2515 | |
Lee Schermerhorn | 06808b0 | 2009-12-14 17:58:21 -0800 | [diff] [blame] | 2516 | #ifdef CONFIG_NUMA |
| 2517 | |
| 2518 | /* |
| 2519 | * hstate attribute for optionally mempolicy-based constraint on persistent |
| 2520 | * huge page alloc/free. |
| 2521 | */ |
| 2522 | static 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 | |
| 2528 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, |
| 2529 | struct kobj_attribute *attr, const char *buf, size_t len) |
| 2530 | { |
David Rientjes | 238d3c1 | 2014-08-06 16:06:51 -0700 | [diff] [blame] | 2531 | return nr_hugepages_store_common(true, kobj, buf, len); |
Lee Schermerhorn | 06808b0 | 2009-12-14 17:58:21 -0800 | [diff] [blame] | 2532 | } |
| 2533 | HSTATE_ATTR(nr_hugepages_mempolicy); |
| 2534 | #endif |
| 2535 | |
| 2536 | |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2537 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
| 2538 | struct kobj_attribute *attr, char *buf) |
| 2539 | { |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2540 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2541 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
| 2542 | } |
Eric B Munson | adbe872 | 2011-01-13 15:47:27 -0800 | [diff] [blame] | 2543 | |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2544 | static 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 Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2549 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2550 | |
Luiz Capitulino | bae7f4a | 2014-06-04 16:07:08 -0700 | [diff] [blame] | 2551 | if (hstate_is_gigantic(h)) |
Eric B Munson | adbe872 | 2011-01-13 15:47:27 -0800 | [diff] [blame] | 2552 | return -EINVAL; |
| 2553 | |
Jingoo Han | 3dbb95f | 2013-09-11 14:20:25 -0700 | [diff] [blame] | 2554 | err = kstrtoul(buf, 10, &input); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2555 | if (err) |
Eric B Munson | 73ae31e | 2011-01-13 15:47:28 -0800 | [diff] [blame] | 2556 | return err; |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2557 | |
| 2558 | spin_lock(&hugetlb_lock); |
| 2559 | h->nr_overcommit_huge_pages = input; |
| 2560 | spin_unlock(&hugetlb_lock); |
| 2561 | |
| 2562 | return count; |
| 2563 | } |
| 2564 | HSTATE_ATTR(nr_overcommit_hugepages); |
| 2565 | |
| 2566 | static ssize_t free_hugepages_show(struct kobject *kobj, |
| 2567 | struct kobj_attribute *attr, char *buf) |
| 2568 | { |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2569 | 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 Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2580 | } |
| 2581 | HSTATE_ATTR_RO(free_hugepages); |
| 2582 | |
| 2583 | static ssize_t resv_hugepages_show(struct kobject *kobj, |
| 2584 | struct kobj_attribute *attr, char *buf) |
| 2585 | { |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2586 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2587 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
| 2588 | } |
| 2589 | HSTATE_ATTR_RO(resv_hugepages); |
| 2590 | |
| 2591 | static ssize_t surplus_hugepages_show(struct kobject *kobj, |
| 2592 | struct kobj_attribute *attr, char *buf) |
| 2593 | { |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2594 | 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 Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2605 | } |
| 2606 | HSTATE_ATTR_RO(surplus_hugepages); |
| 2607 | |
| 2608 | static 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 Schermerhorn | 06808b0 | 2009-12-14 17:58:21 -0800 | [diff] [blame] | 2614 | #ifdef CONFIG_NUMA |
| 2615 | &nr_hugepages_mempolicy_attr.attr, |
| 2616 | #endif |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2617 | NULL, |
| 2618 | }; |
| 2619 | |
| 2620 | static struct attribute_group hstate_attr_group = { |
| 2621 | .attrs = hstate_attrs, |
| 2622 | }; |
| 2623 | |
Jeff Mahoney | 094e953 | 2010-02-02 13:44:14 -0800 | [diff] [blame] | 2624 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
| 2625 | struct kobject **hstate_kobjs, |
| 2626 | struct attribute_group *hstate_attr_group) |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2627 | { |
| 2628 | int retval; |
Aneesh Kumar K.V | 972dc4d | 2012-07-31 16:42:00 -0700 | [diff] [blame] | 2629 | int hi = hstate_index(h); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2630 | |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2631 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
| 2632 | if (!hstate_kobjs[hi]) |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2633 | return -ENOMEM; |
| 2634 | |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2635 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2636 | if (retval) |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2637 | kobject_put(hstate_kobjs[hi]); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2638 | |
| 2639 | return retval; |
| 2640 | } |
| 2641 | |
| 2642 | static 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 Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2652 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
| 2653 | hstate_kobjs, &hstate_attr_group); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2654 | if (err) |
Andrew Morton | ffb22af | 2013-02-22 16:32:08 -0800 | [diff] [blame] | 2655 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2656 | } |
| 2657 | } |
| 2658 | |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2659 | #ifdef CONFIG_NUMA |
| 2660 | |
| 2661 | /* |
| 2662 | * node_hstate/s - associate per node hstate attributes, via their kobjects, |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 2663 | * 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 Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2666 | * the base kernel, on the hugetlb module. |
| 2667 | */ |
| 2668 | struct node_hstate { |
| 2669 | struct kobject *hugepages_kobj; |
| 2670 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; |
| 2671 | }; |
Alexander Kuleshov | b4e289a | 2015-11-05 18:50:14 -0800 | [diff] [blame] | 2672 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2673 | |
| 2674 | /* |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 2675 | * A subset of global hstate attributes for node devices |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2676 | */ |
| 2677 | static 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 | |
| 2684 | static struct attribute_group per_node_hstate_attr_group = { |
| 2685 | .attrs = per_node_hstate_attrs, |
| 2686 | }; |
| 2687 | |
| 2688 | /* |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 2689 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2690 | * Returns node id via non-NULL nidp. |
| 2691 | */ |
| 2692 | static 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 Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 2712 | * Unregister hstate attributes from a single node device. |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2713 | * No-op if no hstate attributes attached. |
| 2714 | */ |
Claudiu Ghioc | 3cd8b44 | 2013-03-04 12:46:15 +0200 | [diff] [blame] | 2715 | static void hugetlb_unregister_node(struct node *node) |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2716 | { |
| 2717 | struct hstate *h; |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 2718 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2719 | |
| 2720 | if (!nhs->hugepages_kobj) |
Lee Schermerhorn | 9b5e5d0 | 2009-12-14 17:58:32 -0800 | [diff] [blame] | 2721 | return; /* no hstate attributes */ |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2722 | |
Aneesh Kumar K.V | 972dc4d | 2012-07-31 16:42:00 -0700 | [diff] [blame] | 2723 | 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 Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2728 | } |
Aneesh Kumar K.V | 972dc4d | 2012-07-31 16:42:00 -0700 | [diff] [blame] | 2729 | } |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2730 | |
| 2731 | kobject_put(nhs->hugepages_kobj); |
| 2732 | nhs->hugepages_kobj = NULL; |
| 2733 | } |
| 2734 | |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2735 | |
| 2736 | /* |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 2737 | * Register hstate attributes for a single node device. |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2738 | * No-op if attributes already registered. |
| 2739 | */ |
Claudiu Ghioc | 3cd8b44 | 2013-03-04 12:46:15 +0200 | [diff] [blame] | 2740 | static void hugetlb_register_node(struct node *node) |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2741 | { |
| 2742 | struct hstate *h; |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 2743 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2744 | int err; |
| 2745 | |
| 2746 | if (nhs->hugepages_kobj) |
| 2747 | return; /* already allocated */ |
| 2748 | |
| 2749 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 2750 | &node->dev.kobj); |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2751 | 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 Morton | ffb22af | 2013-02-22 16:32:08 -0800 | [diff] [blame] | 2759 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
| 2760 | h->name, node->dev.id); |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2761 | hugetlb_unregister_node(node); |
| 2762 | break; |
| 2763 | } |
| 2764 | } |
| 2765 | } |
| 2766 | |
| 2767 | /* |
Lee Schermerhorn | 9b5e5d0 | 2009-12-14 17:58:32 -0800 | [diff] [blame] | 2768 | * hugetlb init time: register hstate attributes for all registered node |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 2769 | * devices of nodes that have memory. All on-line nodes should have |
| 2770 | * registered their associated device by this time. |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2771 | */ |
Luiz Capitulino | 7d9ca00 | 2014-12-12 16:55:24 -0800 | [diff] [blame] | 2772 | static void __init hugetlb_register_all_nodes(void) |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2773 | { |
| 2774 | int nid; |
| 2775 | |
Lai Jiangshan | 8cebfcd | 2012-12-12 13:51:36 -0800 | [diff] [blame] | 2776 | for_each_node_state(nid, N_MEMORY) { |
Wen Congyang | 8732794 | 2012-12-11 16:00:56 -0800 | [diff] [blame] | 2777 | struct node *node = node_devices[nid]; |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 2778 | if (node->dev.id == nid) |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2779 | hugetlb_register_node(node); |
| 2780 | } |
| 2781 | |
| 2782 | /* |
Kay Sievers | 10fbcf4 | 2011-12-21 14:48:43 -0800 | [diff] [blame] | 2783 | * Let the node device driver know we're here so it can |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2784 | * [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 | |
| 2791 | static 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 Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2799 | static void hugetlb_register_all_nodes(void) { } |
| 2800 | |
| 2801 | #endif |
| 2802 | |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2803 | static int __init hugetlb_init(void) |
| 2804 | { |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 2805 | int i; |
| 2806 | |
Nishanth Aravamudan | 457c1b2 | 2014-05-06 12:50:00 -0700 | [diff] [blame] | 2807 | if (!hugepages_supported()) |
Benjamin Herrenschmidt | 0ef89d2 | 2008-07-31 00:07:30 -0700 | [diff] [blame] | 2808 | return 0; |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2809 | |
Nick Piggin | e11bfbf | 2008-07-23 21:27:52 -0700 | [diff] [blame] | 2810 | 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 Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2814 | } |
Aneesh Kumar K.V | 972dc4d | 2012-07-31 16:42:00 -0700 | [diff] [blame] | 2815 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
Vaishali Thakkar | f8b7481 | 2016-02-17 13:11:26 -0800 | [diff] [blame] | 2816 | 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 Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2820 | |
| 2821 | hugetlb_init_hstates(); |
Andi Kleen | aa888a7 | 2008-07-23 21:27:47 -0700 | [diff] [blame] | 2822 | gather_bootmem_prealloc(); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2823 | report_hugepages(); |
| 2824 | |
| 2825 | hugetlb_sysfs_init(); |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2826 | hugetlb_register_all_nodes(); |
Jianguo Wu | 7179e7b | 2012-12-18 14:23:19 -0800 | [diff] [blame] | 2827 | hugetlb_cgroup_file_init(); |
Lee Schermerhorn | 9a305230 | 2009-12-14 17:58:25 -0800 | [diff] [blame] | 2828 | |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 2829 | #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 Kravetz | c672c7f | 2015-09-08 15:01:35 -0700 | [diff] [blame] | 2834 | hugetlb_fault_mutex_table = |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 2835 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); |
Mike Kravetz | c672c7f | 2015-09-08 15:01:35 -0700 | [diff] [blame] | 2836 | BUG_ON(!hugetlb_fault_mutex_table); |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 2837 | |
| 2838 | for (i = 0; i < num_fault_mutexes; i++) |
Mike Kravetz | c672c7f | 2015-09-08 15:01:35 -0700 | [diff] [blame] | 2839 | mutex_init(&hugetlb_fault_mutex_table[i]); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2840 | return 0; |
| 2841 | } |
Paul Gortmaker | 3e89e1c | 2016-01-14 15:21:52 -0800 | [diff] [blame] | 2842 | subsys_initcall(hugetlb_init); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2843 | |
| 2844 | /* Should be called on processing a hugepagesz=... option */ |
Vaishali Thakkar | 9fee021 | 2016-05-19 17:11:04 -0700 | [diff] [blame] | 2845 | void __init hugetlb_bad_size(void) |
| 2846 | { |
| 2847 | parsed_valid_hugepagesz = false; |
| 2848 | } |
| 2849 | |
Kirill A. Shutemov | d00181b | 2015-11-06 16:29:57 -0800 | [diff] [blame] | 2850 | void __init hugetlb_add_hstate(unsigned int order) |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2851 | { |
| 2852 | struct hstate *h; |
Andi Kleen | 8faa8b0 | 2008-07-23 21:27:48 -0700 | [diff] [blame] | 2853 | unsigned long i; |
| 2854 | |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2855 | if (size_to_hstate(PAGE_SIZE << order)) { |
Joe Perches | 598d809 | 2016-03-17 14:19:44 -0700 | [diff] [blame] | 2856 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2857 | return; |
| 2858 | } |
Aneesh Kumar K.V | 47d3834 | 2012-07-31 16:41:54 -0700 | [diff] [blame] | 2859 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2860 | BUG_ON(order == 0); |
Aneesh Kumar K.V | 47d3834 | 2012-07-31 16:41:54 -0700 | [diff] [blame] | 2861 | h = &hstates[hugetlb_max_hstate++]; |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2862 | h->order = order; |
| 2863 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); |
Andi Kleen | 8faa8b0 | 2008-07-23 21:27:48 -0700 | [diff] [blame] | 2864 | 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.V | 0edaecf | 2012-07-31 16:42:07 -0700 | [diff] [blame] | 2868 | INIT_LIST_HEAD(&h->hugepage_activelist); |
Andrew Morton | 54f18d3 | 2016-05-19 17:11:40 -0700 | [diff] [blame] | 2869 | h->next_nid_to_alloc = first_memory_node; |
| 2870 | h->next_nid_to_free = first_memory_node; |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2871 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
| 2872 | huge_page_size(h)/1024); |
Andi Kleen | 8faa8b0 | 2008-07-23 21:27:48 -0700 | [diff] [blame] | 2873 | |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2874 | parsed_hstate = h; |
| 2875 | } |
| 2876 | |
Nick Piggin | e11bfbf | 2008-07-23 21:27:52 -0700 | [diff] [blame] | 2877 | static int __init hugetlb_nrpages_setup(char *s) |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2878 | { |
| 2879 | unsigned long *mhp; |
Andi Kleen | 8faa8b0 | 2008-07-23 21:27:48 -0700 | [diff] [blame] | 2880 | static unsigned long *last_mhp; |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2881 | |
Vaishali Thakkar | 9fee021 | 2016-05-19 17:11:04 -0700 | [diff] [blame] | 2882 | 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 Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2888 | /* |
Aneesh Kumar K.V | 47d3834 | 2012-07-31 16:41:54 -0700 | [diff] [blame] | 2889 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2890 | * so this hugepages= parameter goes to the "default hstate". |
| 2891 | */ |
Vaishali Thakkar | 9fee021 | 2016-05-19 17:11:04 -0700 | [diff] [blame] | 2892 | else if (!hugetlb_max_hstate) |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2893 | mhp = &default_hstate_max_huge_pages; |
| 2894 | else |
| 2895 | mhp = &parsed_hstate->max_huge_pages; |
| 2896 | |
Andi Kleen | 8faa8b0 | 2008-07-23 21:27:48 -0700 | [diff] [blame] | 2897 | if (mhp == last_mhp) { |
Joe Perches | 598d809 | 2016-03-17 14:19:44 -0700 | [diff] [blame] | 2898 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
Andi Kleen | 8faa8b0 | 2008-07-23 21:27:48 -0700 | [diff] [blame] | 2899 | return 1; |
| 2900 | } |
| 2901 | |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2902 | if (sscanf(s, "%lu", mhp) <= 0) |
| 2903 | *mhp = 0; |
| 2904 | |
Andi Kleen | 8faa8b0 | 2008-07-23 21:27:48 -0700 | [diff] [blame] | 2905 | /* |
| 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.V | 47d3834 | 2012-07-31 16:41:54 -0700 | [diff] [blame] | 2910 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
Andi Kleen | 8faa8b0 | 2008-07-23 21:27:48 -0700 | [diff] [blame] | 2911 | hugetlb_hstate_alloc_pages(parsed_hstate); |
| 2912 | |
| 2913 | last_mhp = mhp; |
| 2914 | |
Nishanth Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2915 | return 1; |
| 2916 | } |
Nick Piggin | e11bfbf | 2008-07-23 21:27:52 -0700 | [diff] [blame] | 2917 | __setup("hugepages=", hugetlb_nrpages_setup); |
| 2918 | |
| 2919 | static 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 Aravamudan | a343787 | 2008-07-23 21:27:44 -0700 | [diff] [blame] | 2925 | |
Nishanth Aravamudan | 8a21346 | 2008-07-25 19:44:37 -0700 | [diff] [blame] | 2926 | static 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 Schermerhorn | 06808b0 | 2009-12-14 17:58:21 -0800 | [diff] [blame] | 2938 | static 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 Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2941 | { |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2942 | struct hstate *h = &default_hstate; |
David Rientjes | 238d3c1 | 2014-08-06 16:06:51 -0700 | [diff] [blame] | 2943 | unsigned long tmp = h->max_huge_pages; |
Michal Hocko | 08d4a24 | 2011-01-13 15:47:26 -0800 | [diff] [blame] | 2944 | int ret; |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2945 | |
Nishanth Aravamudan | 457c1b2 | 2014-05-06 12:50:00 -0700 | [diff] [blame] | 2946 | if (!hugepages_supported()) |
Jan Stancek | 8661362 | 2016-03-09 14:08:35 -0800 | [diff] [blame] | 2947 | return -EOPNOTSUPP; |
Nishanth Aravamudan | 457c1b2 | 2014-05-06 12:50:00 -0700 | [diff] [blame] | 2948 | |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2949 | table->data = &tmp; |
| 2950 | table->maxlen = sizeof(unsigned long); |
Michal Hocko | 08d4a24 | 2011-01-13 15:47:26 -0800 | [diff] [blame] | 2951 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
| 2952 | if (ret) |
| 2953 | goto out; |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2954 | |
David Rientjes | 238d3c1 | 2014-08-06 16:06:51 -0700 | [diff] [blame] | 2955 | if (write) |
| 2956 | ret = __nr_hugepages_store_common(obey_mempolicy, h, |
| 2957 | NUMA_NO_NODE, tmp, *length); |
Michal Hocko | 08d4a24 | 2011-01-13 15:47:26 -0800 | [diff] [blame] | 2958 | out: |
| 2959 | return ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2960 | } |
Mel Gorman | 396faf0 | 2007-07-17 04:03:13 -0700 | [diff] [blame] | 2961 | |
Lee Schermerhorn | 06808b0 | 2009-12-14 17:58:21 -0800 | [diff] [blame] | 2962 | int 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 |
| 2971 | int 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 Aravamudan | a3d0c6a | 2008-02-08 04:18:18 -0800 | [diff] [blame] | 2979 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
Alexey Dobriyan | 8d65af7 | 2009-09-23 15:57:19 -0700 | [diff] [blame] | 2980 | void __user *buffer, |
Nishanth Aravamudan | a3d0c6a | 2008-02-08 04:18:18 -0800 | [diff] [blame] | 2981 | size_t *length, loff_t *ppos) |
| 2982 | { |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 2983 | struct hstate *h = &default_hstate; |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2984 | unsigned long tmp; |
Michal Hocko | 08d4a24 | 2011-01-13 15:47:26 -0800 | [diff] [blame] | 2985 | int ret; |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2986 | |
Nishanth Aravamudan | 457c1b2 | 2014-05-06 12:50:00 -0700 | [diff] [blame] | 2987 | if (!hugepages_supported()) |
Jan Stancek | 8661362 | 2016-03-09 14:08:35 -0800 | [diff] [blame] | 2988 | return -EOPNOTSUPP; |
Nishanth Aravamudan | 457c1b2 | 2014-05-06 12:50:00 -0700 | [diff] [blame] | 2989 | |
Petr Holasek | c033a93 | 2011-03-22 16:33:05 -0700 | [diff] [blame] | 2990 | tmp = h->nr_overcommit_huge_pages; |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2991 | |
Luiz Capitulino | bae7f4a | 2014-06-04 16:07:08 -0700 | [diff] [blame] | 2992 | if (write && hstate_is_gigantic(h)) |
Eric B Munson | adbe872 | 2011-01-13 15:47:27 -0800 | [diff] [blame] | 2993 | return -EINVAL; |
| 2994 | |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 2995 | table->data = &tmp; |
| 2996 | table->maxlen = sizeof(unsigned long); |
Michal Hocko | 08d4a24 | 2011-01-13 15:47:26 -0800 | [diff] [blame] | 2997 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
| 2998 | if (ret) |
| 2999 | goto out; |
Andi Kleen | e5ff215 | 2008-07-23 21:27:42 -0700 | [diff] [blame] | 3000 | |
| 3001 | if (write) { |
| 3002 | spin_lock(&hugetlb_lock); |
| 3003 | h->nr_overcommit_huge_pages = tmp; |
| 3004 | spin_unlock(&hugetlb_lock); |
| 3005 | } |
Michal Hocko | 08d4a24 | 2011-01-13 15:47:26 -0800 | [diff] [blame] | 3006 | out: |
| 3007 | return ret; |
Nishanth Aravamudan | a3d0c6a | 2008-02-08 04:18:18 -0800 | [diff] [blame] | 3008 | } |
| 3009 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3010 | #endif /* CONFIG_SYSCTL */ |
| 3011 | |
Alexey Dobriyan | e1759c2 | 2008-10-15 23:50:22 +0400 | [diff] [blame] | 3012 | void hugetlb_report_meminfo(struct seq_file *m) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3013 | { |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3014 | struct hstate *h = &default_hstate; |
Nishanth Aravamudan | 457c1b2 | 2014-05-06 12:50:00 -0700 | [diff] [blame] | 3015 | if (!hugepages_supported()) |
| 3016 | return; |
Alexey Dobriyan | e1759c2 | 2008-10-15 23:50:22 +0400 | [diff] [blame] | 3017 | seq_printf(m, |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 3018 | "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 Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3023 | 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 Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3028 | } |
| 3029 | |
| 3030 | int hugetlb_report_node_meminfo(int nid, char *buf) |
| 3031 | { |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3032 | struct hstate *h = &default_hstate; |
Nishanth Aravamudan | 457c1b2 | 2014-05-06 12:50:00 -0700 | [diff] [blame] | 3033 | if (!hugepages_supported()) |
| 3034 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3035 | return sprintf(buf, |
| 3036 | "Node %d HugePages_Total: %5u\n" |
Nishanth Aravamudan | a1de091 | 2008-03-26 14:37:53 -0700 | [diff] [blame] | 3037 | "Node %d HugePages_Free: %5u\n" |
| 3038 | "Node %d HugePages_Surp: %5u\n", |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3039 | nid, h->nr_huge_pages_node[nid], |
| 3040 | nid, h->free_huge_pages_node[nid], |
| 3041 | nid, h->surplus_huge_pages_node[nid]); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3042 | } |
| 3043 | |
David Rientjes | 949f7ec | 2013-04-29 15:07:48 -0700 | [diff] [blame] | 3044 | void hugetlb_show_meminfo(void) |
| 3045 | { |
| 3046 | struct hstate *h; |
| 3047 | int nid; |
| 3048 | |
Nishanth Aravamudan | 457c1b2 | 2014-05-06 12:50:00 -0700 | [diff] [blame] | 3049 | if (!hugepages_supported()) |
| 3050 | return; |
| 3051 | |
David Rientjes | 949f7ec | 2013-04-29 15:07:48 -0700 | [diff] [blame] | 3052 | 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 Horiguchi | 5d317b2 | 2015-11-05 18:47:14 -0800 | [diff] [blame] | 3062 | void 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 Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3068 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
| 3069 | unsigned long hugetlb_total_pages(void) |
| 3070 | { |
Wanpeng Li | d002858 | 2013-03-22 15:04:40 -0700 | [diff] [blame] | 3071 | 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 Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3077 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3078 | |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3079 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
Mel Gorman | fc1b8a7 | 2008-07-23 21:27:22 -0700 | [diff] [blame] | 3080 | { |
| 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 Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3102 | if (gather_surplus_pages(h, delta) < 0) |
Mel Gorman | fc1b8a7 | 2008-07-23 21:27:22 -0700 | [diff] [blame] | 3103 | goto out; |
| 3104 | |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3105 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
| 3106 | return_unused_surplus_pages(h, delta); |
Mel Gorman | fc1b8a7 | 2008-07-23 21:27:22 -0700 | [diff] [blame] | 3107 | goto out; |
| 3108 | } |
| 3109 | } |
| 3110 | |
| 3111 | ret = 0; |
| 3112 | if (delta < 0) |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3113 | return_unused_surplus_pages(h, (unsigned long) -delta); |
Mel Gorman | fc1b8a7 | 2008-07-23 21:27:22 -0700 | [diff] [blame] | 3114 | |
| 3115 | out: |
| 3116 | spin_unlock(&hugetlb_lock); |
| 3117 | return ret; |
| 3118 | } |
| 3119 | |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 3120 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
| 3121 | { |
Joonsoo Kim | f522c3a | 2013-09-11 14:21:53 -0700 | [diff] [blame] | 3122 | struct resv_map *resv = vma_resv_map(vma); |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 3123 | |
| 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 Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 3128 | * has a reference to the reservation map it cannot disappear until |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 3129 | * after this open call completes. It is therefore safe to take a |
| 3130 | * new reference here without additional locking. |
| 3131 | */ |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 3132 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
Joonsoo Kim | f522c3a | 2013-09-11 14:21:53 -0700 | [diff] [blame] | 3133 | kref_get(&resv->refs); |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 3134 | } |
| 3135 | |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 3136 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
| 3137 | { |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3138 | struct hstate *h = hstate_vma(vma); |
Joonsoo Kim | f522c3a | 2013-09-11 14:21:53 -0700 | [diff] [blame] | 3139 | struct resv_map *resv = vma_resv_map(vma); |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 3140 | struct hugepage_subpool *spool = subpool_vma(vma); |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 3141 | unsigned long reserve, start, end; |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 3142 | long gbl_reserve; |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 3143 | |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 3144 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
| 3145 | return; |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 3146 | |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 3147 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
| 3148 | end = vma_hugecache_offset(h, vma, vma->vm_end); |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 3149 | |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 3150 | reserve = (end - start) - region_count(resv, start, end); |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 3151 | |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 3152 | kref_put(&resv->refs, resv_map_release); |
| 3153 | |
| 3154 | if (reserve) { |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 3155 | /* |
| 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 Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 3161 | } |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 3162 | } |
| 3163 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3164 | /* |
| 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 Jiang | 11bac80 | 2017-02-24 14:56:41 -0800 | [diff] [blame] | 3170 | static int hugetlb_vm_op_fault(struct vm_fault *vmf) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3171 | { |
| 3172 | BUG(); |
Nick Piggin | d0217ac | 2007-07-19 01:47:03 -0700 | [diff] [blame] | 3173 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3174 | } |
| 3175 | |
Alexey Dobriyan | f0f37e2f | 2009-09-27 22:29:37 +0400 | [diff] [blame] | 3176 | const struct vm_operations_struct hugetlb_vm_ops = { |
Nick Piggin | d0217ac | 2007-07-19 01:47:03 -0700 | [diff] [blame] | 3177 | .fault = hugetlb_vm_op_fault, |
Andy Whitcroft | 84afd99 | 2008-07-23 21:27:32 -0700 | [diff] [blame] | 3178 | .open = hugetlb_vm_op_open, |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 3179 | .close = hugetlb_vm_op_close, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3180 | }; |
| 3181 | |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3182 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
| 3183 | int writable) |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3184 | { |
| 3185 | pte_t entry; |
| 3186 | |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3187 | if (writable) { |
Gerald Schaefer | 106c992 | 2013-04-29 15:07:23 -0700 | [diff] [blame] | 3188 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
| 3189 | vma->vm_page_prot))); |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3190 | } else { |
Gerald Schaefer | 106c992 | 2013-04-29 15:07:23 -0700 | [diff] [blame] | 3191 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
| 3192 | vma->vm_page_prot)); |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3193 | } |
| 3194 | entry = pte_mkyoung(entry); |
| 3195 | entry = pte_mkhuge(entry); |
Chris Metcalf | d9ed9fa | 2012-04-01 14:01:34 -0400 | [diff] [blame] | 3196 | entry = arch_make_huge_pte(entry, vma, page, writable); |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3197 | |
| 3198 | return entry; |
| 3199 | } |
| 3200 | |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3201 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
| 3202 | unsigned long address, pte_t *ptep) |
| 3203 | { |
| 3204 | pte_t entry; |
| 3205 | |
Gerald Schaefer | 106c992 | 2013-04-29 15:07:23 -0700 | [diff] [blame] | 3206 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
Chris Forbes | 32f8452 | 2011-07-25 17:12:14 -0700 | [diff] [blame] | 3207 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
Russell King | 4b3073e | 2009-12-18 16:40:18 +0000 | [diff] [blame] | 3208 | update_mmu_cache(vma, address, ptep); |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3209 | } |
| 3210 | |
Aneesh Kumar K.V | d5ed744 | 2017-07-06 15:38:47 -0700 | [diff] [blame] | 3211 | bool is_hugetlb_entry_migration(pte_t pte) |
Naoya Horiguchi | 4a705fe | 2014-06-23 13:22:03 -0700 | [diff] [blame] | 3212 | { |
| 3213 | swp_entry_t swp; |
| 3214 | |
| 3215 | if (huge_pte_none(pte) || pte_present(pte)) |
Aneesh Kumar K.V | d5ed744 | 2017-07-06 15:38:47 -0700 | [diff] [blame] | 3216 | return false; |
Naoya Horiguchi | 4a705fe | 2014-06-23 13:22:03 -0700 | [diff] [blame] | 3217 | swp = pte_to_swp_entry(pte); |
| 3218 | if (non_swap_entry(swp) && is_migration_entry(swp)) |
Aneesh Kumar K.V | d5ed744 | 2017-07-06 15:38:47 -0700 | [diff] [blame] | 3219 | return true; |
Naoya Horiguchi | 4a705fe | 2014-06-23 13:22:03 -0700 | [diff] [blame] | 3220 | else |
Aneesh Kumar K.V | d5ed744 | 2017-07-06 15:38:47 -0700 | [diff] [blame] | 3221 | return false; |
Naoya Horiguchi | 4a705fe | 2014-06-23 13:22:03 -0700 | [diff] [blame] | 3222 | } |
| 3223 | |
| 3224 | static 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 Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3236 | |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3237 | int 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 Dickins | 1c59827 | 2005-10-19 21:23:43 -0700 | [diff] [blame] | 3242 | unsigned long addr; |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3243 | int cow; |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3244 | struct hstate *h = hstate_vma(vma); |
| 3245 | unsigned long sz = huge_page_size(h); |
Andreas Sandberg | e8569dd | 2014-01-21 15:49:09 -0800 | [diff] [blame] | 3246 | unsigned long mmun_start; /* For mmu_notifiers */ |
| 3247 | unsigned long mmun_end; /* For mmu_notifiers */ |
| 3248 | int ret = 0; |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3249 | |
| 3250 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3251 | |
Andreas Sandberg | e8569dd | 2014-01-21 15:49:09 -0800 | [diff] [blame] | 3252 | 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 Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3257 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3258 | spinlock_t *src_ptl, *dst_ptl; |
Punit Agrawal | 7868a20 | 2017-07-06 15:39:42 -0700 | [diff] [blame] | 3259 | src_pte = huge_pte_offset(src, addr, sz); |
Hugh Dickins | c74df32 | 2005-10-29 18:16:23 -0700 | [diff] [blame] | 3260 | if (!src_pte) |
| 3261 | continue; |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3262 | dst_pte = huge_pte_alloc(dst, addr, sz); |
Andreas Sandberg | e8569dd | 2014-01-21 15:49:09 -0800 | [diff] [blame] | 3263 | if (!dst_pte) { |
| 3264 | ret = -ENOMEM; |
| 3265 | break; |
| 3266 | } |
Larry Woodman | c5c9942 | 2008-01-24 05:49:25 -0800 | [diff] [blame] | 3267 | |
| 3268 | /* If the pagetables are shared don't copy or take references */ |
| 3269 | if (dst_pte == src_pte) |
| 3270 | continue; |
| 3271 | |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3272 | 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 Horiguchi | 4a705fe | 2014-06-23 13:22:03 -0700 | [diff] [blame] | 3275 | 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 Agrawal | e5251fd | 2017-07-06 15:39:50 -0700 | [diff] [blame] | 3289 | set_huge_swap_pte_at(src, addr, src_pte, |
| 3290 | entry, sz); |
Naoya Horiguchi | 4a705fe | 2014-06-23 13:22:03 -0700 | [diff] [blame] | 3291 | } |
Punit Agrawal | e5251fd | 2017-07-06 15:39:50 -0700 | [diff] [blame] | 3292 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
Naoya Horiguchi | 4a705fe | 2014-06-23 13:22:03 -0700 | [diff] [blame] | 3293 | } else { |
Joerg Roedel | 34ee645 | 2014-11-13 13:46:09 +1100 | [diff] [blame] | 3294 | if (cow) { |
Gerald Schaefer | 7f2e952 | 2008-04-28 02:13:29 -0700 | [diff] [blame] | 3295 | huge_ptep_set_wrprotect(src, addr, src_pte); |
Joerg Roedel | 34ee645 | 2014-11-13 13:46:09 +1100 | [diff] [blame] | 3296 | mmu_notifier_invalidate_range(src, mmun_start, |
| 3297 | mmun_end); |
| 3298 | } |
Naoya Horiguchi | 0253d63 | 2014-07-23 14:00:19 -0700 | [diff] [blame] | 3299 | entry = huge_ptep_get(src_pte); |
Hugh Dickins | 1c59827 | 2005-10-19 21:23:43 -0700 | [diff] [blame] | 3300 | ptepage = pte_page(entry); |
| 3301 | get_page(ptepage); |
Kirill A. Shutemov | 53f9263 | 2016-01-15 16:53:42 -0800 | [diff] [blame] | 3302 | page_dup_rmap(ptepage, true); |
Hugh Dickins | 1c59827 | 2005-10-19 21:23:43 -0700 | [diff] [blame] | 3303 | set_huge_pte_at(dst, addr, dst_pte, entry); |
Naoya Horiguchi | 5d317b2 | 2015-11-05 18:47:14 -0800 | [diff] [blame] | 3304 | hugetlb_count_add(pages_per_huge_page(h), dst); |
Hugh Dickins | 1c59827 | 2005-10-19 21:23:43 -0700 | [diff] [blame] | 3305 | } |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3306 | spin_unlock(src_ptl); |
| 3307 | spin_unlock(dst_ptl); |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3308 | } |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3309 | |
Andreas Sandberg | e8569dd | 2014-01-21 15:49:09 -0800 | [diff] [blame] | 3310 | if (cow) |
| 3311 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); |
| 3312 | |
| 3313 | return ret; |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3314 | } |
| 3315 | |
Aneesh Kumar K.V | 24669e5 | 2012-07-31 16:42:03 -0700 | [diff] [blame] | 3316 | void __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 Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3319 | { |
| 3320 | struct mm_struct *mm = vma->vm_mm; |
| 3321 | unsigned long address; |
David Gibson | c7546f8 | 2005-08-05 11:59:35 -0700 | [diff] [blame] | 3322 | pte_t *ptep; |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3323 | pte_t pte; |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3324 | spinlock_t *ptl; |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3325 | struct page *page; |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3326 | struct hstate *h = hstate_vma(vma); |
| 3327 | unsigned long sz = huge_page_size(h); |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 3328 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
| 3329 | const unsigned long mmun_end = end; /* For mmu_notifiers */ |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3330 | |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3331 | WARN_ON(!is_vm_hugetlb_page(vma)); |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3332 | BUG_ON(start & ~huge_page_mask(h)); |
| 3333 | BUG_ON(end & ~huge_page_mask(h)); |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3334 | |
Aneesh Kumar K.V | 07e3266 | 2016-12-12 16:42:40 -0800 | [diff] [blame] | 3335 | /* |
| 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.V | 24669e5 | 2012-07-31 16:42:03 -0700 | [diff] [blame] | 3340 | tlb_start_vma(tlb, vma); |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 3341 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
Hillf Danton | 569f48b8 | 2014-12-10 15:44:41 -0800 | [diff] [blame] | 3342 | address = start; |
Hillf Danton | 569f48b8 | 2014-12-10 15:44:41 -0800 | [diff] [blame] | 3343 | for (; address < end; address += sz) { |
Punit Agrawal | 7868a20 | 2017-07-06 15:39:42 -0700 | [diff] [blame] | 3344 | ptep = huge_pte_offset(mm, address, sz); |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 3345 | if (!ptep) |
David Gibson | c7546f8 | 2005-08-05 11:59:35 -0700 | [diff] [blame] | 3346 | continue; |
| 3347 | |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3348 | ptl = huge_pte_lock(h, mm, ptep); |
Aneesh Kumar K.V | 31d49da | 2016-07-26 15:24:06 -0700 | [diff] [blame] | 3349 | if (huge_pmd_unshare(mm, &address, ptep)) { |
| 3350 | spin_unlock(ptl); |
| 3351 | continue; |
| 3352 | } |
Chen, Kenneth W | 39dde65 | 2006-12-06 20:32:03 -0800 | [diff] [blame] | 3353 | |
Hillf Danton | 6629326 | 2012-03-23 15:01:48 -0700 | [diff] [blame] | 3354 | pte = huge_ptep_get(ptep); |
Aneesh Kumar K.V | 31d49da | 2016-07-26 15:24:06 -0700 | [diff] [blame] | 3355 | if (huge_pte_none(pte)) { |
| 3356 | spin_unlock(ptl); |
| 3357 | continue; |
| 3358 | } |
Hillf Danton | 6629326 | 2012-03-23 15:01:48 -0700 | [diff] [blame] | 3359 | |
| 3360 | /* |
Naoya Horiguchi | 9fbc1f6 | 2015-02-11 15:25:32 -0800 | [diff] [blame] | 3361 | * Migrating hugepage or HWPoisoned hugepage is already |
| 3362 | * unmapped and its refcount is dropped, so just clear pte here. |
Hillf Danton | 6629326 | 2012-03-23 15:01:48 -0700 | [diff] [blame] | 3363 | */ |
Naoya Horiguchi | 9fbc1f6 | 2015-02-11 15:25:32 -0800 | [diff] [blame] | 3364 | if (unlikely(!pte_present(pte))) { |
Punit Agrawal | 9386fac | 2017-07-06 15:39:46 -0700 | [diff] [blame] | 3365 | huge_pte_clear(mm, address, ptep, sz); |
Aneesh Kumar K.V | 31d49da | 2016-07-26 15:24:06 -0700 | [diff] [blame] | 3366 | spin_unlock(ptl); |
| 3367 | continue; |
Naoya Horiguchi | 8c4894c | 2012-12-12 13:52:28 -0800 | [diff] [blame] | 3368 | } |
Hillf Danton | 6629326 | 2012-03-23 15:01:48 -0700 | [diff] [blame] | 3369 | |
| 3370 | page = pte_page(pte); |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3371 | /* |
| 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.V | 31d49da | 2016-07-26 15:24:06 -0700 | [diff] [blame] | 3377 | if (page != ref_page) { |
| 3378 | spin_unlock(ptl); |
| 3379 | continue; |
| 3380 | } |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3381 | /* |
| 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 Gibson | c7546f8 | 2005-08-05 11:59:35 -0700 | [diff] [blame] | 3389 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
Aneesh Kumar K.V | b528e4b | 2016-12-12 16:42:37 -0800 | [diff] [blame] | 3390 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
Gerald Schaefer | 106c992 | 2013-04-29 15:07:23 -0700 | [diff] [blame] | 3391 | if (huge_pte_dirty(pte)) |
Ken Chen | 6649a38 | 2007-02-08 14:20:27 -0800 | [diff] [blame] | 3392 | set_page_dirty(page); |
Hillf Danton | 9e81130 | 2012-03-21 16:34:03 -0700 | [diff] [blame] | 3393 | |
Naoya Horiguchi | 5d317b2 | 2015-11-05 18:47:14 -0800 | [diff] [blame] | 3394 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
Kirill A. Shutemov | d281ee6 | 2016-01-15 16:52:16 -0800 | [diff] [blame] | 3395 | page_remove_rmap(page, true); |
Aneesh Kumar K.V | 31d49da | 2016-07-26 15:24:06 -0700 | [diff] [blame] | 3396 | |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3397 | spin_unlock(ptl); |
Aneesh Kumar K.V | e77b085 | 2016-07-26 15:24:12 -0700 | [diff] [blame] | 3398 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
Aneesh Kumar K.V | 31d49da | 2016-07-26 15:24:06 -0700 | [diff] [blame] | 3399 | /* |
| 3400 | * Bail out after unmapping reference page if supplied |
| 3401 | */ |
| 3402 | if (ref_page) |
| 3403 | break; |
Chen, Kenneth W | fe1668a | 2006-10-04 02:15:24 -0700 | [diff] [blame] | 3404 | } |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 3405 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
Aneesh Kumar K.V | 24669e5 | 2012-07-31 16:42:03 -0700 | [diff] [blame] | 3406 | tlb_end_vma(tlb, vma); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3407 | } |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 3408 | |
Mel Gorman | d833352 | 2012-07-31 16:46:20 -0700 | [diff] [blame] | 3409 | void __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 Bueso | c8c06ef | 2014-12-12 16:54:24 -0800 | [diff] [blame] | 3421 | * is to clear it before releasing the i_mmap_rwsem. This works |
Mel Gorman | d833352 | 2012-07-31 16:46:20 -0700 | [diff] [blame] | 3422 | * because in the context this is called, the VMA is about to be |
Davidlohr Bueso | c8c06ef | 2014-12-12 16:54:24 -0800 | [diff] [blame] | 3423 | * destroyed and the i_mmap_rwsem is held. |
Mel Gorman | d833352 | 2012-07-31 16:46:20 -0700 | [diff] [blame] | 3424 | */ |
| 3425 | vma->vm_flags &= ~VM_MAYSHARE; |
| 3426 | } |
| 3427 | |
Chen, Kenneth W | 502717f | 2006-10-11 01:20:46 -0700 | [diff] [blame] | 3428 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3429 | unsigned long end, struct page *ref_page) |
Chen, Kenneth W | 502717f | 2006-10-11 01:20:46 -0700 | [diff] [blame] | 3430 | { |
Aneesh Kumar K.V | 24669e5 | 2012-07-31 16:42:03 -0700 | [diff] [blame] | 3431 | struct mm_struct *mm; |
| 3432 | struct mmu_gather tlb; |
| 3433 | |
| 3434 | mm = vma->vm_mm; |
| 3435 | |
Linus Torvalds | 2b04725 | 2013-08-15 11:42:25 -0700 | [diff] [blame] | 3436 | tlb_gather_mmu(&tlb, mm, start, end); |
Aneesh Kumar K.V | 24669e5 | 2012-07-31 16:42:03 -0700 | [diff] [blame] | 3437 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
| 3438 | tlb_finish_mmu(&tlb, start, end); |
Chen, Kenneth W | 502717f | 2006-10-11 01:20:46 -0700 | [diff] [blame] | 3439 | } |
| 3440 | |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3441 | /* |
| 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 Bueso | 2f4612a | 2014-08-06 16:06:45 -0700 | [diff] [blame] | 3447 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
| 3448 | struct page *page, unsigned long address) |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3449 | { |
Adam Litke | 7526674 | 2008-11-12 13:24:56 -0800 | [diff] [blame] | 3450 | struct hstate *h = hstate_vma(vma); |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3451 | struct vm_area_struct *iter_vma; |
| 3452 | struct address_space *mapping; |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3453 | 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 Litke | 7526674 | 2008-11-12 13:24:56 -0800 | [diff] [blame] | 3459 | address = address & huge_page_mask(h); |
Michal Hocko | 36e4f20 | 2012-10-08 16:33:31 -0700 | [diff] [blame] | 3460 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
| 3461 | vma->vm_pgoff; |
Al Viro | 93c76a3 | 2015-12-04 23:45:44 -0500 | [diff] [blame] | 3462 | mapping = vma->vm_file->f_mapping; |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3463 | |
Mel Gorman | 4eb2b1d | 2009-12-14 17:59:53 -0800 | [diff] [blame] | 3464 | /* |
| 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 Bueso | 83cde9e | 2014-12-12 16:54:21 -0800 | [diff] [blame] | 3469 | i_mmap_lock_write(mapping); |
Michel Lespinasse | 6b2dbba | 2012-10-08 16:31:25 -0700 | [diff] [blame] | 3470 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3471 | /* Do not unmap the current VMA */ |
| 3472 | if (iter_vma == vma) |
| 3473 | continue; |
| 3474 | |
| 3475 | /* |
Mel Gorman | 2f84a89 | 2015-10-01 15:36:57 -0700 | [diff] [blame] | 3476 | * 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 Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3484 | * 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.V | 24669e5 | 2012-07-31 16:42:03 -0700 | [diff] [blame] | 3491 | unmap_hugepage_range(iter_vma, address, |
| 3492 | address + huge_page_size(h), page); |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3493 | } |
Davidlohr Bueso | 83cde9e | 2014-12-12 16:54:21 -0800 | [diff] [blame] | 3494 | i_mmap_unlock_write(mapping); |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3495 | } |
| 3496 | |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 3497 | /* |
| 3498 | * Hugetlb_cow() should be called with page lock of the original hugepage held. |
Michal Hocko | ef009b2 | 2012-01-10 15:07:21 -0800 | [diff] [blame] | 3499 | * 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 Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 3502 | */ |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3503 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
Aneesh Kumar K.V | 3999f52 | 2016-12-12 16:41:56 -0800 | [diff] [blame] | 3504 | unsigned long address, pte_t *ptep, |
| 3505 | struct page *pagecache_page, spinlock_t *ptl) |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3506 | { |
Aneesh Kumar K.V | 3999f52 | 2016-12-12 16:41:56 -0800 | [diff] [blame] | 3507 | pte_t pte; |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3508 | struct hstate *h = hstate_vma(vma); |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3509 | struct page *old_page, *new_page; |
Davidlohr Bueso | ad4404a | 2014-08-06 16:06:47 -0700 | [diff] [blame] | 3510 | int ret = 0, outside_reserve = 0; |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 3511 | unsigned long mmun_start; /* For mmu_notifiers */ |
| 3512 | unsigned long mmun_end; /* For mmu_notifiers */ |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3513 | |
Aneesh Kumar K.V | 3999f52 | 2016-12-12 16:41:56 -0800 | [diff] [blame] | 3514 | pte = huge_ptep_get(ptep); |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3515 | old_page = pte_page(pte); |
| 3516 | |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3517 | retry_avoidcopy: |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3518 | /* If no-one else is actually using this page, avoid the copy |
| 3519 | * and just make the page writable */ |
Joonsoo Kim | 37a2140 | 2013-09-11 14:21:04 -0700 | [diff] [blame] | 3520 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
Hugh Dickins | 5a49973 | 2016-07-14 12:07:38 -0700 | [diff] [blame] | 3521 | page_move_anon_rmap(old_page, vma); |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3522 | set_huge_ptep_writable(vma, address, ptep); |
Nick Piggin | 83c5407 | 2007-07-19 01:47:05 -0700 | [diff] [blame] | 3523 | return 0; |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3524 | } |
| 3525 | |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3526 | /* |
| 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 Kim | 5944d01 | 2013-09-11 14:21:55 -0700 | [diff] [blame] | 3535 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3536 | old_page != pagecache_page) |
| 3537 | outside_reserve = 1; |
| 3538 | |
Kirill A. Shutemov | 09cbfea | 2016-04-01 15:29:47 +0300 | [diff] [blame] | 3539 | get_page(old_page); |
Larry Woodman | b76c8cf | 2009-12-14 17:59:37 -0800 | [diff] [blame] | 3540 | |
Davidlohr Bueso | ad4404a | 2014-08-06 16:06:47 -0700 | [diff] [blame] | 3541 | /* |
| 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. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3545 | spin_unlock(ptl); |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3546 | new_page = alloc_huge_page(vma, address, outside_reserve); |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3547 | |
Adam Litke | 2fc39ce | 2007-11-14 16:59:39 -0800 | [diff] [blame] | 3548 | if (IS_ERR(new_page)) { |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3549 | /* |
| 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. Shutemov | 09cbfea | 2016-04-01 15:29:47 +0300 | [diff] [blame] | 3557 | put_page(old_page); |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3558 | BUG_ON(huge_pte_none(pte)); |
Davidlohr Bueso | 2f4612a | 2014-08-06 16:06:45 -0700 | [diff] [blame] | 3559 | unmap_ref_private(mm, vma, old_page, address); |
| 3560 | BUG_ON(huge_pte_none(pte)); |
| 3561 | spin_lock(ptl); |
Punit Agrawal | 7868a20 | 2017-07-06 15:39:42 -0700 | [diff] [blame] | 3562 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
| 3563 | huge_page_size(h)); |
Davidlohr Bueso | 2f4612a | 2014-08-06 16:06:45 -0700 | [diff] [blame] | 3564 | 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 Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3572 | } |
| 3573 | |
Davidlohr Bueso | ad4404a | 2014-08-06 16:06:47 -0700 | [diff] [blame] | 3574 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
| 3575 | VM_FAULT_OOM : VM_FAULT_SIGBUS; |
| 3576 | goto out_release_old; |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3577 | } |
| 3578 | |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 3579 | /* |
| 3580 | * When the original hugepage is shared one, it does not have |
| 3581 | * anon_vma prepared. |
| 3582 | */ |
Dean Nelson | 44e2aa9 | 2010-10-26 14:22:08 -0700 | [diff] [blame] | 3583 | if (unlikely(anon_vma_prepare(vma))) { |
Davidlohr Bueso | ad4404a | 2014-08-06 16:06:47 -0700 | [diff] [blame] | 3584 | ret = VM_FAULT_OOM; |
| 3585 | goto out_release_all; |
Dean Nelson | 44e2aa9 | 2010-10-26 14:22:08 -0700 | [diff] [blame] | 3586 | } |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 3587 | |
Andrea Arcangeli | 47ad847 | 2011-01-13 15:46:47 -0800 | [diff] [blame] | 3588 | copy_user_huge_page(new_page, old_page, address, vma, |
| 3589 | pages_per_huge_page(h)); |
Nick Piggin | 0ed361d | 2008-02-04 22:29:34 -0800 | [diff] [blame] | 3590 | __SetPageUptodate(new_page); |
Naoya Horiguchi | bcc5422 | 2015-04-15 16:14:38 -0700 | [diff] [blame] | 3591 | set_page_huge_active(new_page); |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3592 | |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 3593 | 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 Bueso | ad4404a | 2014-08-06 16:06:47 -0700 | [diff] [blame] | 3596 | |
Larry Woodman | b76c8cf | 2009-12-14 17:59:37 -0800 | [diff] [blame] | 3597 | /* |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3598 | * Retake the page table lock to check for racing updates |
Larry Woodman | b76c8cf | 2009-12-14 17:59:37 -0800 | [diff] [blame] | 3599 | * before the page tables are altered |
| 3600 | */ |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3601 | spin_lock(ptl); |
Punit Agrawal | 7868a20 | 2017-07-06 15:39:42 -0700 | [diff] [blame] | 3602 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
| 3603 | huge_page_size(h)); |
Naoya Horiguchi | a9af0c5 | 2014-04-07 15:36:54 -0700 | [diff] [blame] | 3604 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
Joonsoo Kim | 07443a8 | 2013-09-11 14:21:58 -0700 | [diff] [blame] | 3605 | ClearPagePrivate(new_page); |
| 3606 | |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3607 | /* Break COW */ |
Gerald Schaefer | 8fe627e | 2008-04-28 02:13:28 -0700 | [diff] [blame] | 3608 | huge_ptep_clear_flush(vma, address, ptep); |
Joerg Roedel | 34ee645 | 2014-11-13 13:46:09 +1100 | [diff] [blame] | 3609 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3610 | set_huge_pte_at(mm, address, ptep, |
| 3611 | make_huge_pte(vma, new_page, 1)); |
Kirill A. Shutemov | d281ee6 | 2016-01-15 16:52:16 -0800 | [diff] [blame] | 3612 | page_remove_rmap(old_page, true); |
Naoya Horiguchi | cd67f0d | 2010-09-10 13:23:04 +0900 | [diff] [blame] | 3613 | hugepage_add_new_anon_rmap(new_page, vma, address); |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3614 | /* Make the old page be freed below */ |
| 3615 | new_page = old_page; |
| 3616 | } |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3617 | spin_unlock(ptl); |
Sagi Grimberg | 2ec74c3 | 2012-10-08 16:33:33 -0700 | [diff] [blame] | 3618 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
Davidlohr Bueso | ad4404a | 2014-08-06 16:06:47 -0700 | [diff] [blame] | 3619 | out_release_all: |
Mike Kravetz | 96b96a9 | 2016-11-10 10:46:32 -0800 | [diff] [blame] | 3620 | restore_reserve_on_error(h, vma, address, new_page); |
Kirill A. Shutemov | 09cbfea | 2016-04-01 15:29:47 +0300 | [diff] [blame] | 3621 | put_page(new_page); |
Davidlohr Bueso | ad4404a | 2014-08-06 16:06:47 -0700 | [diff] [blame] | 3622 | out_release_old: |
Kirill A. Shutemov | 09cbfea | 2016-04-01 15:29:47 +0300 | [diff] [blame] | 3623 | put_page(old_page); |
Joonsoo Kim | 8312034 | 2013-09-11 14:21:57 -0700 | [diff] [blame] | 3624 | |
Davidlohr Bueso | ad4404a | 2014-08-06 16:06:47 -0700 | [diff] [blame] | 3625 | spin_lock(ptl); /* Caller expects lock to be held */ |
| 3626 | return ret; |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3627 | } |
| 3628 | |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3629 | /* Return the pagecache page at a given address within a VMA */ |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3630 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
| 3631 | struct vm_area_struct *vma, unsigned long address) |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3632 | { |
| 3633 | struct address_space *mapping; |
Andy Whitcroft | e7c4b0b | 2008-07-23 21:27:26 -0700 | [diff] [blame] | 3634 | pgoff_t idx; |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3635 | |
| 3636 | mapping = vma->vm_file->f_mapping; |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3637 | idx = vma_hugecache_offset(h, vma, address); |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3638 | |
| 3639 | return find_lock_page(mapping, idx); |
| 3640 | } |
| 3641 | |
Hugh Dickins | 3ae77f4 | 2009-09-21 17:03:33 -0700 | [diff] [blame] | 3642 | /* |
| 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 | */ |
| 3646 | static bool hugetlbfs_pagecache_present(struct hstate *h, |
Hugh Dickins | 2a15efc | 2009-09-21 17:03:27 -0700 | [diff] [blame] | 3647 | 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 Kravetz | ab76ad5 | 2015-09-08 15:01:50 -0700 | [diff] [blame] | 3662 | int 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. Day | a1ed3dd | 2007-07-17 04:03:33 -0700 | [diff] [blame] | 3679 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 3680 | struct address_space *mapping, pgoff_t idx, |
| 3681 | unsigned long address, pte_t *ptep, unsigned int flags) |
Hugh Dickins | ac9b9c6 | 2005-10-20 16:24:28 +0100 | [diff] [blame] | 3682 | { |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3683 | struct hstate *h = hstate_vma(vma); |
Hugh Dickins | ac9b9c6 | 2005-10-20 16:24:28 +0100 | [diff] [blame] | 3684 | int ret = VM_FAULT_SIGBUS; |
Hillf Danton | 409eb8c | 2012-01-20 14:34:13 -0800 | [diff] [blame] | 3685 | int anon_rmap = 0; |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 3686 | unsigned long size; |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 3687 | struct page *page; |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3688 | pte_t new_pte; |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3689 | spinlock_t *ptl; |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 3690 | |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3691 | /* |
| 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 Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 3694 | * COW. Warn that such a situation has occurred as it may not be obvious |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3695 | */ |
| 3696 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { |
Geoffrey Thomas | 910154d | 2016-03-09 14:08:04 -0800 | [diff] [blame] | 3697 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
Andrew Morton | ffb22af | 2013-02-22 16:32:08 -0800 | [diff] [blame] | 3698 | current->pid); |
Mel Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3699 | return ret; |
| 3700 | } |
| 3701 | |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 3702 | /* |
| 3703 | * Use page lock to guard against racing truncation |
| 3704 | * before we get page_table_lock. |
| 3705 | */ |
Christoph Lameter | 6bda666 | 2006-01-06 00:10:49 -0800 | [diff] [blame] | 3706 | retry: |
| 3707 | page = find_lock_page(mapping, idx); |
| 3708 | if (!page) { |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3709 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
Hugh Dickins | ebed4bf | 2006-10-28 10:38:43 -0700 | [diff] [blame] | 3710 | if (idx >= size) |
| 3711 | goto out; |
Mike Kravetz | 1a1aad8 | 2017-02-22 15:43:01 -0800 | [diff] [blame] | 3712 | |
| 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 Gorman | 04f2cbe | 2008-07-23 21:27:25 -0700 | [diff] [blame] | 3744 | page = alloc_huge_page(vma, address, 0); |
Adam Litke | 2fc39ce | 2007-11-14 16:59:39 -0800 | [diff] [blame] | 3745 | if (IS_ERR(page)) { |
Aneesh Kumar K.V | 76dcee7 | 2012-07-31 16:41:57 -0700 | [diff] [blame] | 3746 | ret = PTR_ERR(page); |
| 3747 | if (ret == -ENOMEM) |
| 3748 | ret = VM_FAULT_OOM; |
| 3749 | else |
| 3750 | ret = VM_FAULT_SIGBUS; |
Christoph Lameter | 6bda666 | 2006-01-06 00:10:49 -0800 | [diff] [blame] | 3751 | goto out; |
| 3752 | } |
Andrea Arcangeli | 47ad847 | 2011-01-13 15:46:47 -0800 | [diff] [blame] | 3753 | clear_huge_page(page, address, pages_per_huge_page(h)); |
Nick Piggin | 0ed361d | 2008-02-04 22:29:34 -0800 | [diff] [blame] | 3754 | __SetPageUptodate(page); |
Naoya Horiguchi | bcc5422 | 2015-04-15 16:14:38 -0700 | [diff] [blame] | 3755 | set_page_huge_active(page); |
Hugh Dickins | ac9b9c6 | 2005-10-20 16:24:28 +0100 | [diff] [blame] | 3756 | |
Mel Gorman | f83a275 | 2009-05-28 14:34:40 -0700 | [diff] [blame] | 3757 | if (vma->vm_flags & VM_MAYSHARE) { |
Mike Kravetz | ab76ad5 | 2015-09-08 15:01:50 -0700 | [diff] [blame] | 3758 | int err = huge_add_to_page_cache(page, mapping, idx); |
Christoph Lameter | 6bda666 | 2006-01-06 00:10:49 -0800 | [diff] [blame] | 3759 | if (err) { |
| 3760 | put_page(page); |
Christoph Lameter | 6bda666 | 2006-01-06 00:10:49 -0800 | [diff] [blame] | 3761 | if (err == -EEXIST) |
| 3762 | goto retry; |
| 3763 | goto out; |
| 3764 | } |
Mel Gorman | 23be746 | 2010-04-23 13:17:56 -0400 | [diff] [blame] | 3765 | } else { |
Christoph Lameter | 6bda666 | 2006-01-06 00:10:49 -0800 | [diff] [blame] | 3766 | lock_page(page); |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 3767 | if (unlikely(anon_vma_prepare(vma))) { |
| 3768 | ret = VM_FAULT_OOM; |
| 3769 | goto backout_unlocked; |
| 3770 | } |
Hillf Danton | 409eb8c | 2012-01-20 14:34:13 -0800 | [diff] [blame] | 3771 | anon_rmap = 1; |
Mel Gorman | 23be746 | 2010-04-23 13:17:56 -0400 | [diff] [blame] | 3772 | } |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 3773 | } else { |
Naoya Horiguchi | 998b438 | 2010-09-08 10:19:32 +0900 | [diff] [blame] | 3774 | /* |
| 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 Forbes | 32f8452 | 2011-07-25 17:12:14 -0700 | [diff] [blame] | 3780 | ret = VM_FAULT_HWPOISON | |
Aneesh Kumar K.V | 972dc4d | 2012-07-31 16:42:00 -0700 | [diff] [blame] | 3781 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
Naoya Horiguchi | 998b438 | 2010-09-08 10:19:32 +0900 | [diff] [blame] | 3782 | goto backout_unlocked; |
| 3783 | } |
Christoph Lameter | 6bda666 | 2006-01-06 00:10:49 -0800 | [diff] [blame] | 3784 | } |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3785 | |
Andy Whitcroft | 57303d8 | 2008-08-12 15:08:47 -0700 | [diff] [blame] | 3786 | /* |
| 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 Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 3792 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
Andy Whitcroft | 2b26736 | 2008-08-12 15:08:49 -0700 | [diff] [blame] | 3793 | if (vma_needs_reservation(h, vma, address) < 0) { |
| 3794 | ret = VM_FAULT_OOM; |
| 3795 | goto backout_unlocked; |
| 3796 | } |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 3797 | /* Just decrements count, does not deallocate */ |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 3798 | vma_end_reservation(h, vma, address); |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 3799 | } |
Andy Whitcroft | 57303d8 | 2008-08-12 15:08:47 -0700 | [diff] [blame] | 3800 | |
Aneesh Kumar K.V | 8bea805 | 2016-12-12 16:41:59 -0800 | [diff] [blame] | 3801 | ptl = huge_pte_lock(h, mm, ptep); |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3802 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 3803 | if (idx >= size) |
| 3804 | goto backout; |
| 3805 | |
Nick Piggin | 83c5407 | 2007-07-19 01:47:05 -0700 | [diff] [blame] | 3806 | ret = 0; |
Gerald Schaefer | 7f2e952 | 2008-04-28 02:13:29 -0700 | [diff] [blame] | 3807 | if (!huge_pte_none(huge_ptep_get(ptep))) |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 3808 | goto backout; |
| 3809 | |
Joonsoo Kim | 07443a8 | 2013-09-11 14:21:58 -0700 | [diff] [blame] | 3810 | if (anon_rmap) { |
| 3811 | ClearPagePrivate(page); |
Hillf Danton | 409eb8c | 2012-01-20 14:34:13 -0800 | [diff] [blame] | 3812 | hugepage_add_new_anon_rmap(page, vma, address); |
Choi Gi-yong | ac71490 | 2014-04-07 15:37:36 -0700 | [diff] [blame] | 3813 | } else |
Kirill A. Shutemov | 53f9263 | 2016-01-15 16:53:42 -0800 | [diff] [blame] | 3814 | page_dup_rmap(page, true); |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3815 | 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 Horiguchi | 5d317b2 | 2015-11-05 18:47:14 -0800 | [diff] [blame] | 3819 | hugetlb_count_add(pages_per_huge_page(h), mm); |
Hugh Dickins | 788c7df | 2009-06-23 13:49:05 +0100 | [diff] [blame] | 3820 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3821 | /* Optimization, do the COW without a second fault */ |
Aneesh Kumar K.V | 3999f52 | 2016-12-12 16:41:56 -0800 | [diff] [blame] | 3822 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3823 | } |
| 3824 | |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3825 | spin_unlock(ptl); |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 3826 | unlock_page(page); |
| 3827 | out: |
Hugh Dickins | ac9b9c6 | 2005-10-20 16:24:28 +0100 | [diff] [blame] | 3828 | return ret; |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 3829 | |
| 3830 | backout: |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3831 | spin_unlock(ptl); |
Andy Whitcroft | 2b26736 | 2008-08-12 15:08:49 -0700 | [diff] [blame] | 3832 | backout_unlocked: |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 3833 | unlock_page(page); |
Mike Kravetz | 96b96a9 | 2016-11-10 10:46:32 -0800 | [diff] [blame] | 3834 | restore_reserve_on_error(h, vma, address, page); |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 3835 | put_page(page); |
| 3836 | goto out; |
Hugh Dickins | ac9b9c6 | 2005-10-20 16:24:28 +0100 | [diff] [blame] | 3837 | } |
| 3838 | |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 3839 | #ifdef CONFIG_SMP |
Mike Kravetz | c672c7f | 2015-09-08 15:01:35 -0700 | [diff] [blame] | 3840 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 3841 | 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 Kravetz | c672c7f | 2015-09-08 15:01:35 -0700 | [diff] [blame] | 3865 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 3866 | 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 Litke | 86e5216 | 2006-01-06 00:10:43 -0800 | [diff] [blame] | 3874 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
Hugh Dickins | 788c7df | 2009-06-23 13:49:05 +0100 | [diff] [blame] | 3875 | unsigned long address, unsigned int flags) |
Adam Litke | 86e5216 | 2006-01-06 00:10:43 -0800 | [diff] [blame] | 3876 | { |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 3877 | pte_t *ptep, entry; |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3878 | spinlock_t *ptl; |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3879 | int ret; |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 3880 | u32 hash; |
| 3881 | pgoff_t idx; |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 3882 | struct page *page = NULL; |
Andy Whitcroft | 57303d8 | 2008-08-12 15:08:47 -0700 | [diff] [blame] | 3883 | struct page *pagecache_page = NULL; |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 3884 | struct hstate *h = hstate_vma(vma); |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 3885 | struct address_space *mapping; |
Naoya Horiguchi | 0f792cf | 2015-02-11 15:25:25 -0800 | [diff] [blame] | 3886 | int need_wait_lock = 0; |
Adam Litke | 86e5216 | 2006-01-06 00:10:43 -0800 | [diff] [blame] | 3887 | |
KAMEZAWA Hiroyuki | 1e16a53 | 2012-01-10 15:07:22 -0800 | [diff] [blame] | 3888 | address &= huge_page_mask(h); |
| 3889 | |
Punit Agrawal | 7868a20 | 2017-07-06 15:39:42 -0700 | [diff] [blame] | 3890 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
Naoya Horiguchi | fd6a03e | 2010-05-28 09:29:21 +0900 | [diff] [blame] | 3891 | if (ptep) { |
| 3892 | entry = huge_ptep_get(ptep); |
Naoya Horiguchi | 290408d | 2010-09-08 10:19:35 +0900 | [diff] [blame] | 3893 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3894 | migration_entry_wait_huge(vma, mm, ptep); |
Naoya Horiguchi | 290408d | 2010-09-08 10:19:35 +0900 | [diff] [blame] | 3895 | return 0; |
| 3896 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) |
Chris Forbes | 32f8452 | 2011-07-25 17:12:14 -0700 | [diff] [blame] | 3897 | return VM_FAULT_HWPOISON_LARGE | |
Aneesh Kumar K.V | 972dc4d | 2012-07-31 16:42:00 -0700 | [diff] [blame] | 3898 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
Naoya Horiguchi | 0d777df | 2015-12-11 13:40:49 -0800 | [diff] [blame] | 3899 | } else { |
| 3900 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
| 3901 | if (!ptep) |
| 3902 | return VM_FAULT_OOM; |
Naoya Horiguchi | fd6a03e | 2010-05-28 09:29:21 +0900 | [diff] [blame] | 3903 | } |
| 3904 | |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 3905 | mapping = vma->vm_file->f_mapping; |
| 3906 | idx = vma_hugecache_offset(h, vma, address); |
| 3907 | |
David Gibson | 3935baa | 2006-03-22 00:08:53 -0800 | [diff] [blame] | 3908 | /* |
| 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 Kravetz | c672c7f | 2015-09-08 15:01:35 -0700 | [diff] [blame] | 3913 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); |
| 3914 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 3915 | |
Gerald Schaefer | 7f2e952 | 2008-04-28 02:13:29 -0700 | [diff] [blame] | 3916 | entry = huge_ptep_get(ptep); |
| 3917 | if (huge_pte_none(entry)) { |
Davidlohr Bueso | 8382d91 | 2014-04-03 14:47:31 -0700 | [diff] [blame] | 3918 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
David Gibson | b4d1d99 | 2008-10-15 22:01:11 -0700 | [diff] [blame] | 3919 | goto out_mutex; |
David Gibson | 3935baa | 2006-03-22 00:08:53 -0800 | [diff] [blame] | 3920 | } |
Adam Litke | 86e5216 | 2006-01-06 00:10:43 -0800 | [diff] [blame] | 3921 | |
Nick Piggin | 83c5407 | 2007-07-19 01:47:05 -0700 | [diff] [blame] | 3922 | ret = 0; |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3923 | |
Andy Whitcroft | 57303d8 | 2008-08-12 15:08:47 -0700 | [diff] [blame] | 3924 | /* |
Naoya Horiguchi | 0f792cf | 2015-02-11 15:25:25 -0800 | [diff] [blame] | 3925 | * 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 Whitcroft | 57303d8 | 2008-08-12 15:08:47 -0700 | [diff] [blame] | 3935 | * 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 Schaefer | 106c992 | 2013-04-29 15:07:23 -0700 | [diff] [blame] | 3942 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
Andy Whitcroft | 2b26736 | 2008-08-12 15:08:49 -0700 | [diff] [blame] | 3943 | if (vma_needs_reservation(h, vma, address) < 0) { |
| 3944 | ret = VM_FAULT_OOM; |
David Gibson | b4d1d99 | 2008-10-15 22:01:11 -0700 | [diff] [blame] | 3945 | goto out_mutex; |
Andy Whitcroft | 2b26736 | 2008-08-12 15:08:49 -0700 | [diff] [blame] | 3946 | } |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 3947 | /* Just decrements count, does not deallocate */ |
Mike Kravetz | feba16e | 2015-09-08 15:01:31 -0700 | [diff] [blame] | 3948 | vma_end_reservation(h, vma, address); |
Andy Whitcroft | 57303d8 | 2008-08-12 15:08:47 -0700 | [diff] [blame] | 3949 | |
Mel Gorman | f83a275 | 2009-05-28 14:34:40 -0700 | [diff] [blame] | 3950 | if (!(vma->vm_flags & VM_MAYSHARE)) |
Andy Whitcroft | 57303d8 | 2008-08-12 15:08:47 -0700 | [diff] [blame] | 3951 | pagecache_page = hugetlbfs_pagecache_page(h, |
| 3952 | vma, address); |
| 3953 | } |
| 3954 | |
Naoya Horiguchi | 0f792cf | 2015-02-11 15:25:25 -0800 | [diff] [blame] | 3955 | ptl = huge_pte_lock(h, mm, ptep); |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 3956 | |
David Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 3957 | /* Check for a racing update before calling hugetlb_cow */ |
David Gibson | b4d1d99 | 2008-10-15 22:01:11 -0700 | [diff] [blame] | 3958 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3959 | goto out_ptl; |
David Gibson | b4d1d99 | 2008-10-15 22:01:11 -0700 | [diff] [blame] | 3960 | |
Naoya Horiguchi | 0f792cf | 2015-02-11 15:25:25 -0800 | [diff] [blame] | 3961 | /* |
| 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 Gibson | b4d1d99 | 2008-10-15 22:01:11 -0700 | [diff] [blame] | 3974 | |
Hugh Dickins | 788c7df | 2009-06-23 13:49:05 +0100 | [diff] [blame] | 3975 | if (flags & FAULT_FLAG_WRITE) { |
Gerald Schaefer | 106c992 | 2013-04-29 15:07:23 -0700 | [diff] [blame] | 3976 | if (!huge_pte_write(entry)) { |
Aneesh Kumar K.V | 3999f52 | 2016-12-12 16:41:56 -0800 | [diff] [blame] | 3977 | ret = hugetlb_cow(mm, vma, address, ptep, |
| 3978 | pagecache_page, ptl); |
Naoya Horiguchi | 0f792cf | 2015-02-11 15:25:25 -0800 | [diff] [blame] | 3979 | goto out_put_page; |
David Gibson | b4d1d99 | 2008-10-15 22:01:11 -0700 | [diff] [blame] | 3980 | } |
Gerald Schaefer | 106c992 | 2013-04-29 15:07:23 -0700 | [diff] [blame] | 3981 | entry = huge_pte_mkdirty(entry); |
David Gibson | b4d1d99 | 2008-10-15 22:01:11 -0700 | [diff] [blame] | 3982 | } |
| 3983 | entry = pte_mkyoung(entry); |
Hugh Dickins | 788c7df | 2009-06-23 13:49:05 +0100 | [diff] [blame] | 3984 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
| 3985 | flags & FAULT_FLAG_WRITE)) |
Russell King | 4b3073e | 2009-12-18 16:40:18 +0000 | [diff] [blame] | 3986 | update_mmu_cache(vma, address, ptep); |
Naoya Horiguchi | 0f792cf | 2015-02-11 15:25:25 -0800 | [diff] [blame] | 3987 | out_put_page: |
| 3988 | if (page != pagecache_page) |
| 3989 | unlock_page(page); |
| 3990 | put_page(page); |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 3991 | out_ptl: |
| 3992 | spin_unlock(ptl); |
Andy Whitcroft | 57303d8 | 2008-08-12 15:08:47 -0700 | [diff] [blame] | 3993 | |
| 3994 | if (pagecache_page) { |
| 3995 | unlock_page(pagecache_page); |
| 3996 | put_page(pagecache_page); |
| 3997 | } |
David Gibson | b4d1d99 | 2008-10-15 22:01:11 -0700 | [diff] [blame] | 3998 | out_mutex: |
Mike Kravetz | c672c7f | 2015-09-08 15:01:35 -0700 | [diff] [blame] | 3999 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
Naoya Horiguchi | 0f792cf | 2015-02-11 15:25:25 -0800 | [diff] [blame] | 4000 | /* |
| 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 Gibson | 1e8f889 | 2006-01-06 00:10:44 -0800 | [diff] [blame] | 4009 | return ret; |
Adam Litke | 86e5216 | 2006-01-06 00:10:43 -0800 | [diff] [blame] | 4010 | } |
| 4011 | |
Mike Kravetz | 8fb5deb | 2017-02-22 15:42:52 -0800 | [diff] [blame] | 4012 | /* |
| 4013 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with |
| 4014 | * modifications for huge pages. |
| 4015 | */ |
| 4016 | int 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 Kravetz | 1c9e8de | 2017-02-22 15:43:43 -0800 | [diff] [blame] | 4023 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
Mike Kravetz | 8fb5deb | 2017-02-22 15:42:52 -0800 | [diff] [blame] | 4024 | 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 Kravetz | 810a56b | 2017-02-22 15:42:58 -0800 | [diff] [blame] | 4038 | pages_per_huge_page(h), false); |
Mike Kravetz | 8fb5deb | 2017-02-22 15:42:52 -0800 | [diff] [blame] | 4039 | |
| 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 Kravetz | 1c9e8de | 2017-02-22 15:43:43 -0800 | [diff] [blame] | 4060 | /* |
| 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 Kravetz | 8fb5deb | 2017-02-22 15:42:52 -0800 | [diff] [blame] | 4072 | 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 Kravetz | 1c9e8de | 2017-02-22 15:43:43 -0800 | [diff] [blame] | 4079 | 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 Kravetz | 8fb5deb | 2017-02-22 15:42:52 -0800 | [diff] [blame] | 4085 | |
| 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 Kravetz | 1c9e8de | 2017-02-22 15:43:43 -0800 | [diff] [blame] | 4101 | if (vm_shared) |
| 4102 | unlock_page(page); |
Mike Kravetz | 8fb5deb | 2017-02-22 15:42:52 -0800 | [diff] [blame] | 4103 | ret = 0; |
| 4104 | out: |
| 4105 | return ret; |
| 4106 | out_release_unlock: |
| 4107 | spin_unlock(ptl); |
Mike Kravetz | 1c9e8de | 2017-02-22 15:43:43 -0800 | [diff] [blame] | 4108 | out_release_nounlock: |
| 4109 | if (vm_shared) |
| 4110 | unlock_page(page); |
Mike Kravetz | 8fb5deb | 2017-02-22 15:42:52 -0800 | [diff] [blame] | 4111 | put_page(page); |
| 4112 | goto out; |
| 4113 | } |
| 4114 | |
Michel Lespinasse | 28a3571 | 2013-02-22 16:35:55 -0800 | [diff] [blame] | 4115 | long 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 Arcangeli | 87ffc11 | 2017-02-22 15:43:13 -0800 | [diff] [blame] | 4118 | long i, unsigned int flags, int *nonblocking) |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 4119 | { |
Chen, Kenneth W | d5d4b0a | 2006-03-22 00:09:03 -0800 | [diff] [blame] | 4120 | unsigned long pfn_offset; |
| 4121 | unsigned long vaddr = *position; |
Michel Lespinasse | 28a3571 | 2013-02-22 16:35:55 -0800 | [diff] [blame] | 4122 | unsigned long remainder = *nr_pages; |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 4123 | struct hstate *h = hstate_vma(vma); |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 4124 | |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 4125 | while (vaddr < vma->vm_end && remainder) { |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 4126 | pte_t *pte; |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 4127 | spinlock_t *ptl = NULL; |
Hugh Dickins | 2a15efc | 2009-09-21 17:03:27 -0700 | [diff] [blame] | 4128 | int absent; |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 4129 | struct page *page; |
| 4130 | |
| 4131 | /* |
David Rientjes | 0205796 | 2015-04-14 15:48:24 -0700 | [diff] [blame] | 4132 | * 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 Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 4141 | * Some archs (sparc64, sh*) have multiple pte_ts to |
Hugh Dickins | 2a15efc | 2009-09-21 17:03:27 -0700 | [diff] [blame] | 4142 | * each hugepage. We have to make sure we get the |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 4143 | * first, for the page indexing below to work. |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 4144 | * |
| 4145 | * Note that page table lock is not held when pte is null. |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 4146 | */ |
Punit Agrawal | 7868a20 | 2017-07-06 15:39:42 -0700 | [diff] [blame] | 4147 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
| 4148 | huge_page_size(h)); |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 4149 | if (pte) |
| 4150 | ptl = huge_pte_lock(h, mm, pte); |
Hugh Dickins | 2a15efc | 2009-09-21 17:03:27 -0700 | [diff] [blame] | 4151 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 4152 | |
Hugh Dickins | 2a15efc | 2009-09-21 17:03:27 -0700 | [diff] [blame] | 4153 | /* |
| 4154 | * When coredumping, it suits get_dump_page if we just return |
Hugh Dickins | 3ae77f4 | 2009-09-21 17:03:33 -0700 | [diff] [blame] | 4155 | * 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 Dickins | 2a15efc | 2009-09-21 17:03:27 -0700 | [diff] [blame] | 4159 | */ |
Hugh Dickins | 3ae77f4 | 2009-09-21 17:03:33 -0700 | [diff] [blame] | 4160 | if (absent && (flags & FOLL_DUMP) && |
| 4161 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 4162 | if (pte) |
| 4163 | spin_unlock(ptl); |
Hugh Dickins | 2a15efc | 2009-09-21 17:03:27 -0700 | [diff] [blame] | 4164 | remainder = 0; |
| 4165 | break; |
| 4166 | } |
| 4167 | |
Naoya Horiguchi | 9cc3a5b | 2013-04-17 15:58:30 -0700 | [diff] [blame] | 4168 | /* |
| 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 Schaefer | 106c992 | 2013-04-29 15:07:23 -0700 | [diff] [blame] | 4179 | ((flags & FOLL_WRITE) && |
| 4180 | !huge_pte_write(huge_ptep_get(pte)))) { |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 4181 | int ret; |
Andrea Arcangeli | 87ffc11 | 2017-02-22 15:43:13 -0800 | [diff] [blame] | 4182 | unsigned int fault_flags = 0; |
Adam Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 4183 | |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 4184 | if (pte) |
| 4185 | spin_unlock(ptl); |
Andrea Arcangeli | 87ffc11 | 2017-02-22 15:43:13 -0800 | [diff] [blame] | 4186 | 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 Morse | 9a291a7 | 2017-06-02 14:46:46 -0700 | [diff] [blame] | 4200 | int err = vm_fault_to_errno(ret, flags); |
| 4201 | |
| 4202 | if (err) |
| 4203 | return err; |
| 4204 | |
Andrea Arcangeli | 87ffc11 | 2017-02-22 15:43:13 -0800 | [diff] [blame] | 4205 | 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 Litke | 4c88726 | 2005-10-29 18:16:46 -0700 | [diff] [blame] | 4224 | } |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 4225 | |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 4226 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
Gerald Schaefer | 7f2e952 | 2008-04-28 02:13:29 -0700 | [diff] [blame] | 4227 | page = pte_page(huge_ptep_get(pte)); |
Chen, Kenneth W | d5d4b0a | 2006-03-22 00:09:03 -0800 | [diff] [blame] | 4228 | same_page: |
Chen, Kenneth W | d669218 | 2006-03-31 02:29:57 -0800 | [diff] [blame] | 4229 | if (pages) { |
Hugh Dickins | 2a15efc | 2009-09-21 17:03:27 -0700 | [diff] [blame] | 4230 | pages[i] = mem_map_offset(page, pfn_offset); |
Kirill A. Shutemov | ddc58f2 | 2016-01-15 16:52:56 -0800 | [diff] [blame] | 4231 | get_page(pages[i]); |
Chen, Kenneth W | d669218 | 2006-03-31 02:29:57 -0800 | [diff] [blame] | 4232 | } |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 4233 | |
| 4234 | if (vmas) |
| 4235 | vmas[i] = vma; |
| 4236 | |
| 4237 | vaddr += PAGE_SIZE; |
Chen, Kenneth W | d5d4b0a | 2006-03-22 00:09:03 -0800 | [diff] [blame] | 4238 | ++pfn_offset; |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 4239 | --remainder; |
| 4240 | ++i; |
Chen, Kenneth W | d5d4b0a | 2006-03-22 00:09:03 -0800 | [diff] [blame] | 4241 | if (vaddr < vma->vm_end && remainder && |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 4242 | pfn_offset < pages_per_huge_page(h)) { |
Chen, Kenneth W | d5d4b0a | 2006-03-22 00:09:03 -0800 | [diff] [blame] | 4243 | /* |
| 4244 | * We use pfn_offset to avoid touching the pageframes |
| 4245 | * of this compound page. |
| 4246 | */ |
| 4247 | goto same_page; |
| 4248 | } |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 4249 | spin_unlock(ptl); |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 4250 | } |
Michel Lespinasse | 28a3571 | 2013-02-22 16:35:55 -0800 | [diff] [blame] | 4251 | *nr_pages = remainder; |
Andrea Arcangeli | 87ffc11 | 2017-02-22 15:43:13 -0800 | [diff] [blame] | 4252 | /* |
| 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 Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 4257 | *position = vaddr; |
| 4258 | |
Hugh Dickins | 2a15efc | 2009-09-21 17:03:27 -0700 | [diff] [blame] | 4259 | return i ? i : -EFAULT; |
David Gibson | 63551ae | 2005-06-21 17:14:44 -0700 | [diff] [blame] | 4260 | } |
Zhang, Yanmin | 8f86059 | 2006-03-22 00:08:50 -0800 | [diff] [blame] | 4261 | |
Aneesh Kumar K.V | 5491ae7 | 2016-07-13 15:06:43 +0530 | [diff] [blame] | 4262 | #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 Zijlstra | 7da4d64 | 2012-11-19 03:14:23 +0100 | [diff] [blame] | 4270 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
Zhang, Yanmin | 8f86059 | 2006-03-22 00:08:50 -0800 | [diff] [blame] | 4271 | 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 Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 4277 | struct hstate *h = hstate_vma(vma); |
Peter Zijlstra | 7da4d64 | 2012-11-19 03:14:23 +0100 | [diff] [blame] | 4278 | unsigned long pages = 0; |
Zhang, Yanmin | 8f86059 | 2006-03-22 00:08:50 -0800 | [diff] [blame] | 4279 | |
| 4280 | BUG_ON(address >= end); |
| 4281 | flush_cache_range(vma, address, end); |
| 4282 | |
Rik van Riel | a533809 | 2014-04-07 15:36:57 -0700 | [diff] [blame] | 4283 | mmu_notifier_invalidate_range_start(mm, start, end); |
Davidlohr Bueso | 83cde9e | 2014-12-12 16:54:21 -0800 | [diff] [blame] | 4284 | i_mmap_lock_write(vma->vm_file->f_mapping); |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 4285 | for (; address < end; address += huge_page_size(h)) { |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 4286 | spinlock_t *ptl; |
Punit Agrawal | 7868a20 | 2017-07-06 15:39:42 -0700 | [diff] [blame] | 4287 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
Zhang, Yanmin | 8f86059 | 2006-03-22 00:08:50 -0800 | [diff] [blame] | 4288 | if (!ptep) |
| 4289 | continue; |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 4290 | ptl = huge_pte_lock(h, mm, ptep); |
Peter Zijlstra | 7da4d64 | 2012-11-19 03:14:23 +0100 | [diff] [blame] | 4291 | if (huge_pmd_unshare(mm, &address, ptep)) { |
| 4292 | pages++; |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 4293 | spin_unlock(ptl); |
Chen, Kenneth W | 39dde65 | 2006-12-06 20:32:03 -0800 | [diff] [blame] | 4294 | continue; |
Peter Zijlstra | 7da4d64 | 2012-11-19 03:14:23 +0100 | [diff] [blame] | 4295 | } |
Naoya Horiguchi | a8bda28 | 2015-02-11 15:25:28 -0800 | [diff] [blame] | 4296 | 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 Agrawal | e5251fd | 2017-07-06 15:39:50 -0700 | [diff] [blame] | 4309 | set_huge_swap_pte_at(mm, address, ptep, |
| 4310 | newpte, huge_page_size(h)); |
Naoya Horiguchi | a8bda28 | 2015-02-11 15:25:28 -0800 | [diff] [blame] | 4311 | pages++; |
| 4312 | } |
| 4313 | spin_unlock(ptl); |
| 4314 | continue; |
| 4315 | } |
| 4316 | if (!huge_pte_none(pte)) { |
Zhang, Yanmin | 8f86059 | 2006-03-22 00:08:50 -0800 | [diff] [blame] | 4317 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
Gerald Schaefer | 106c992 | 2013-04-29 15:07:23 -0700 | [diff] [blame] | 4318 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
Tony Lu | be7517d | 2013-02-04 14:28:46 -0800 | [diff] [blame] | 4319 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
Zhang, Yanmin | 8f86059 | 2006-03-22 00:08:50 -0800 | [diff] [blame] | 4320 | set_huge_pte_at(mm, address, ptep, pte); |
Peter Zijlstra | 7da4d64 | 2012-11-19 03:14:23 +0100 | [diff] [blame] | 4321 | pages++; |
Zhang, Yanmin | 8f86059 | 2006-03-22 00:08:50 -0800 | [diff] [blame] | 4322 | } |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 4323 | spin_unlock(ptl); |
Zhang, Yanmin | 8f86059 | 2006-03-22 00:08:50 -0800 | [diff] [blame] | 4324 | } |
Mel Gorman | d833352 | 2012-07-31 16:46:20 -0700 | [diff] [blame] | 4325 | /* |
Davidlohr Bueso | c8c06ef | 2014-12-12 16:54:24 -0800 | [diff] [blame] | 4326 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
Mel Gorman | d833352 | 2012-07-31 16:46:20 -0700 | [diff] [blame] | 4327 | * may have cleared our pud entry and done put_page on the page table: |
Davidlohr Bueso | c8c06ef | 2014-12-12 16:54:24 -0800 | [diff] [blame] | 4328 | * once we release i_mmap_rwsem, another task can do the final put_page |
Mel Gorman | d833352 | 2012-07-31 16:46:20 -0700 | [diff] [blame] | 4329 | * and that page table be reused and filled with junk. |
| 4330 | */ |
Aneesh Kumar K.V | 5491ae7 | 2016-07-13 15:06:43 +0530 | [diff] [blame] | 4331 | flush_hugetlb_tlb_range(vma, start, end); |
Joerg Roedel | 34ee645 | 2014-11-13 13:46:09 +1100 | [diff] [blame] | 4332 | mmu_notifier_invalidate_range(mm, start, end); |
Davidlohr Bueso | 83cde9e | 2014-12-12 16:54:21 -0800 | [diff] [blame] | 4333 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
Rik van Riel | a533809 | 2014-04-07 15:36:57 -0700 | [diff] [blame] | 4334 | mmu_notifier_invalidate_range_end(mm, start, end); |
Peter Zijlstra | 7da4d64 | 2012-11-19 03:14:23 +0100 | [diff] [blame] | 4335 | |
| 4336 | return pages << h->order; |
Zhang, Yanmin | 8f86059 | 2006-03-22 00:08:50 -0800 | [diff] [blame] | 4337 | } |
| 4338 | |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 4339 | int hugetlb_reserve_pages(struct inode *inode, |
| 4340 | long from, long to, |
Mel Gorman | 5a6fe12 | 2009-02-10 14:02:27 +0000 | [diff] [blame] | 4341 | struct vm_area_struct *vma, |
KOSAKI Motohiro | ca16d14 | 2011-05-26 19:16:19 +0900 | [diff] [blame] | 4342 | vm_flags_t vm_flags) |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 4343 | { |
Mel Gorman | 17c9d12 | 2009-02-11 16:34:16 +0000 | [diff] [blame] | 4344 | long ret, chg; |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 4345 | struct hstate *h = hstate_inode(inode); |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 4346 | struct hugepage_subpool *spool = subpool_inode(inode); |
Joonsoo Kim | 9119a41 | 2014-04-03 14:47:25 -0700 | [diff] [blame] | 4347 | struct resv_map *resv_map; |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 4348 | long gbl_reserve; |
Adam Litke | e4e574b | 2007-10-16 01:26:19 -0700 | [diff] [blame] | 4349 | |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 4350 | /* |
Mel Gorman | 17c9d12 | 2009-02-11 16:34:16 +0000 | [diff] [blame] | 4351 | * Only apply hugepage reservation if asked. At fault time, an |
| 4352 | * attempt will be made for VM_NORESERVE to allocate a page |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 4353 | * without using reserves |
Mel Gorman | 17c9d12 | 2009-02-11 16:34:16 +0000 | [diff] [blame] | 4354 | */ |
KOSAKI Motohiro | ca16d14 | 2011-05-26 19:16:19 +0900 | [diff] [blame] | 4355 | if (vm_flags & VM_NORESERVE) |
Mel Gorman | 17c9d12 | 2009-02-11 16:34:16 +0000 | [diff] [blame] | 4356 | return 0; |
| 4357 | |
| 4358 | /* |
Mel Gorman | a1e7877 | 2008-07-23 21:27:23 -0700 | [diff] [blame] | 4359 | * 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 Kim | 9119a41 | 2014-04-03 14:47:25 -0700 | [diff] [blame] | 4364 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 4365 | resv_map = inode_resv_map(inode); |
Joonsoo Kim | 9119a41 | 2014-04-03 14:47:25 -0700 | [diff] [blame] | 4366 | |
Joonsoo Kim | 1406ec9 | 2014-04-03 14:47:26 -0700 | [diff] [blame] | 4367 | chg = region_chg(resv_map, from, to); |
Joonsoo Kim | 9119a41 | 2014-04-03 14:47:25 -0700 | [diff] [blame] | 4368 | |
| 4369 | } else { |
| 4370 | resv_map = resv_map_alloc(); |
Mel Gorman | 5a6fe12 | 2009-02-10 14:02:27 +0000 | [diff] [blame] | 4371 | if (!resv_map) |
| 4372 | return -ENOMEM; |
| 4373 | |
Mel Gorman | 17c9d12 | 2009-02-11 16:34:16 +0000 | [diff] [blame] | 4374 | chg = to - from; |
| 4375 | |
Mel Gorman | 5a6fe12 | 2009-02-10 14:02:27 +0000 | [diff] [blame] | 4376 | set_vma_resv_map(vma, resv_map); |
| 4377 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); |
| 4378 | } |
| 4379 | |
Dave Hansen | c50ac05 | 2012-05-29 15:06:46 -0700 | [diff] [blame] | 4380 | if (chg < 0) { |
| 4381 | ret = chg; |
| 4382 | goto out_err; |
| 4383 | } |
Mel Gorman | 17c9d12 | 2009-02-11 16:34:16 +0000 | [diff] [blame] | 4384 | |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 4385 | /* |
| 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 Hansen | c50ac05 | 2012-05-29 15:06:46 -0700 | [diff] [blame] | 4392 | ret = -ENOSPC; |
| 4393 | goto out_err; |
| 4394 | } |
Mel Gorman | 17c9d12 | 2009-02-11 16:34:16 +0000 | [diff] [blame] | 4395 | |
| 4396 | /* |
| 4397 | * Check enough hugepages are available for the reservation. |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 4398 | * Hand the pages back to the subpool if there are not |
Mel Gorman | 17c9d12 | 2009-02-11 16:34:16 +0000 | [diff] [blame] | 4399 | */ |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 4400 | ret = hugetlb_acct_memory(h, gbl_reserve); |
Mel Gorman | 17c9d12 | 2009-02-11 16:34:16 +0000 | [diff] [blame] | 4401 | if (ret < 0) { |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 4402 | /* put back original number of pages, chg */ |
| 4403 | (void)hugepage_subpool_put_pages(spool, chg); |
Dave Hansen | c50ac05 | 2012-05-29 15:06:46 -0700 | [diff] [blame] | 4404 | goto out_err; |
Mel Gorman | 17c9d12 | 2009-02-11 16:34:16 +0000 | [diff] [blame] | 4405 | } |
| 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 Kravetz | 3303967 | 2015-06-24 16:57:58 -0700 | [diff] [blame] | 4418 | 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 W | a43a8c3 | 2006-06-23 02:03:15 -0700 | [diff] [blame] | 4436 | return 0; |
Dave Hansen | c50ac05 | 2012-05-29 15:06:46 -0700 | [diff] [blame] | 4437 | out_err: |
Mike Kravetz | 5e91137 | 2015-09-08 15:01:28 -0700 | [diff] [blame] | 4438 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
Mike Kravetz | ff8c0c5 | 2017-03-31 15:12:07 -0700 | [diff] [blame] | 4439 | /* Don't call region_abort if region_chg failed */ |
| 4440 | if (chg >= 0) |
| 4441 | region_abort(resv_map, from, to); |
Joonsoo Kim | f031dd2 | 2014-04-03 14:47:28 -0700 | [diff] [blame] | 4442 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
| 4443 | kref_put(&resv_map->refs, resv_map_release); |
Dave Hansen | c50ac05 | 2012-05-29 15:06:46 -0700 | [diff] [blame] | 4444 | return ret; |
Chen, Kenneth W | a43a8c3 | 2006-06-23 02:03:15 -0700 | [diff] [blame] | 4445 | } |
| 4446 | |
Mike Kravetz | b5cec28 | 2015-09-08 15:01:41 -0700 | [diff] [blame] | 4447 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
| 4448 | long freed) |
Chen, Kenneth W | a43a8c3 | 2006-06-23 02:03:15 -0700 | [diff] [blame] | 4449 | { |
Andi Kleen | a551643 | 2008-07-23 21:27:41 -0700 | [diff] [blame] | 4450 | struct hstate *h = hstate_inode(inode); |
Joonsoo Kim | 4e35f48 | 2014-04-03 14:47:30 -0700 | [diff] [blame] | 4451 | struct resv_map *resv_map = inode_resv_map(inode); |
Joonsoo Kim | 9119a41 | 2014-04-03 14:47:25 -0700 | [diff] [blame] | 4452 | long chg = 0; |
David Gibson | 9048162 | 2012-03-21 16:34:12 -0700 | [diff] [blame] | 4453 | struct hugepage_subpool *spool = subpool_inode(inode); |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 4454 | long gbl_reserve; |
Ken Chen | 45c682a | 2007-11-14 16:59:44 -0800 | [diff] [blame] | 4455 | |
Mike Kravetz | b5cec28 | 2015-09-08 15:01:41 -0700 | [diff] [blame] | 4456 | 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 Chen | 45c682a | 2007-11-14 16:59:44 -0800 | [diff] [blame] | 4467 | spin_lock(&inode->i_lock); |
Eric Sandeen | e4c6f8b | 2009-07-29 15:02:16 -0700 | [diff] [blame] | 4468 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
Ken Chen | 45c682a | 2007-11-14 16:59:44 -0800 | [diff] [blame] | 4469 | spin_unlock(&inode->i_lock); |
| 4470 | |
Mike Kravetz | 1c5ecae | 2015-04-15 16:13:39 -0700 | [diff] [blame] | 4471 | /* |
| 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 Kravetz | b5cec28 | 2015-09-08 15:01:41 -0700 | [diff] [blame] | 4477 | |
| 4478 | return 0; |
Chen, Kenneth W | a43a8c3 | 2006-06-23 02:03:15 -0700 | [diff] [blame] | 4479 | } |
Naoya Horiguchi | 93f70f9 | 2010-05-28 09:29:20 +0900 | [diff] [blame] | 4480 | |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4481 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
| 4482 | static 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 Munson | de60f5f | 2015-11-05 18:51:36 -0800 | [diff] [blame] | 4492 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
| 4493 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4494 | |
| 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 Krause | 31aafb4 | 2015-09-04 15:47:58 -0700 | [diff] [blame] | 4507 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4508 | { |
| 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 Krause | 31aafb4 | 2015-09-04 15:47:58 -0700 | [diff] [blame] | 4517 | return true; |
| 4518 | return false; |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4519 | } |
| 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 Bueso | c8c06ef | 2014-12-12 16:54:24 -0800 | [diff] [blame] | 4526 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4527 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
| 4528 | * bad pmd for sharing. |
| 4529 | */ |
| 4530 | pte_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. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 4540 | spinlock_t *ptl; |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4541 | |
| 4542 | if (!vma_shareable(vma, addr)) |
| 4543 | return (pte_t *)pmd_alloc(mm, pud, addr); |
| 4544 | |
Davidlohr Bueso | 83cde9e | 2014-12-12 16:54:21 -0800 | [diff] [blame] | 4545 | i_mmap_lock_write(mapping); |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4546 | 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 Agrawal | 7868a20 | 2017-07-06 15:39:42 -0700 | [diff] [blame] | 4552 | spte = huge_pte_offset(svma->vm_mm, saddr, |
| 4553 | vma_mmu_pagesize(svma)); |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4554 | 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.V | 8bea805 | 2016-12-12 16:41:59 -0800 | [diff] [blame] | 4564 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
Kirill A. Shutemov | dc6c9a3 | 2015-02-11 15:26:50 -0800 | [diff] [blame] | 4565 | if (pud_none(*pud)) { |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4566 | pud_populate(mm, pud, |
| 4567 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); |
Kirill A. Shutemov | c17b1f4 | 2016-06-24 14:49:51 -0700 | [diff] [blame] | 4568 | mm_inc_nr_pmds(mm); |
Kirill A. Shutemov | dc6c9a3 | 2015-02-11 15:26:50 -0800 | [diff] [blame] | 4569 | } else { |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4570 | put_page(virt_to_page(spte)); |
Kirill A. Shutemov | dc6c9a3 | 2015-02-11 15:26:50 -0800 | [diff] [blame] | 4571 | } |
Kirill A. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 4572 | spin_unlock(ptl); |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4573 | out: |
| 4574 | pte = (pte_t *)pmd_alloc(mm, pud, addr); |
Davidlohr Bueso | 83cde9e | 2014-12-12 16:54:21 -0800 | [diff] [blame] | 4575 | i_mmap_unlock_write(mapping); |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4576 | 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. Shutemov | cb900f4 | 2013-11-14 14:31:02 -0800 | [diff] [blame] | 4586 | * called with page table lock held. |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4587 | * |
| 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 | */ |
| 4591 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) |
| 4592 | { |
| 4593 | pgd_t *pgd = pgd_offset(mm, *addr); |
Kirill A. Shutemov | c2febaf | 2017-03-09 17:24:07 +0300 | [diff] [blame] | 4594 | p4d_t *p4d = p4d_offset(pgd, *addr); |
| 4595 | pud_t *pud = pud_offset(p4d, *addr); |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4596 | |
| 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. Shutemov | dc6c9a3 | 2015-02-11 15:26:50 -0800 | [diff] [blame] | 4603 | mm_dec_nr_pmds(mm); |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4604 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
| 4605 | return 1; |
| 4606 | } |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4607 | #define want_pmd_share() (1) |
| 4608 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
| 4609 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) |
| 4610 | { |
| 4611 | return NULL; |
| 4612 | } |
Zhang Zhen | e81f2d2 | 2015-06-24 16:56:13 -0700 | [diff] [blame] | 4613 | |
| 4614 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) |
| 4615 | { |
| 4616 | return 0; |
| 4617 | } |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4618 | #define want_pmd_share() (0) |
Steve Capper | 3212b53 | 2013-04-23 12:35:02 +0100 | [diff] [blame] | 4619 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
| 4620 | |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4621 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
| 4622 | pte_t *huge_pte_alloc(struct mm_struct *mm, |
| 4623 | unsigned long addr, unsigned long sz) |
| 4624 | { |
| 4625 | pgd_t *pgd; |
Kirill A. Shutemov | c2febaf | 2017-03-09 17:24:07 +0300 | [diff] [blame] | 4626 | p4d_t *p4d; |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4627 | pud_t *pud; |
| 4628 | pte_t *pte = NULL; |
| 4629 | |
| 4630 | pgd = pgd_offset(mm, addr); |
Kirill A. Shutemov | c2febaf | 2017-03-09 17:24:07 +0300 | [diff] [blame] | 4631 | p4d = p4d_offset(pgd, addr); |
| 4632 | pud = pud_alloc(mm, p4d, addr); |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4633 | 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 Hocko | 4e66631 | 2016-08-02 14:02:34 -0700 | [diff] [blame] | 4644 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4645 | |
| 4646 | return pte; |
| 4647 | } |
| 4648 | |
Punit Agrawal | 7868a20 | 2017-07-06 15:39:42 -0700 | [diff] [blame] | 4649 | pte_t *huge_pte_offset(struct mm_struct *mm, |
| 4650 | unsigned long addr, unsigned long sz) |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4651 | { |
| 4652 | pgd_t *pgd; |
Kirill A. Shutemov | c2febaf | 2017-03-09 17:24:07 +0300 | [diff] [blame] | 4653 | p4d_t *p4d; |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4654 | pud_t *pud; |
Kirill A. Shutemov | c2febaf | 2017-03-09 17:24:07 +0300 | [diff] [blame] | 4655 | pmd_t *pmd; |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4656 | |
| 4657 | pgd = pgd_offset(mm, addr); |
Kirill A. Shutemov | c2febaf | 2017-03-09 17:24:07 +0300 | [diff] [blame] | 4658 | 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 Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4669 | return (pte_t *) pmd; |
| 4670 | } |
| 4671 | |
Naoya Horiguchi | 61f77ed | 2015-02-11 15:25:15 -0800 | [diff] [blame] | 4672 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
| 4673 | |
| 4674 | /* |
| 4675 | * These functions are overwritable if your architecture needs its own |
| 4676 | * behavior. |
| 4677 | */ |
| 4678 | struct page * __weak |
| 4679 | follow_huge_addr(struct mm_struct *mm, unsigned long address, |
| 4680 | int write) |
| 4681 | { |
| 4682 | return ERR_PTR(-EINVAL); |
| 4683 | } |
| 4684 | |
| 4685 | struct page * __weak |
Aneesh Kumar K.V | 4dc7145 | 2017-07-06 15:38:56 -0700 | [diff] [blame] | 4686 | follow_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 | |
| 4693 | struct page * __weak |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4694 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
Naoya Horiguchi | e66f17f | 2015-02-11 15:25:22 -0800 | [diff] [blame] | 4695 | pmd_t *pmd, int flags) |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4696 | { |
Naoya Horiguchi | e66f17f | 2015-02-11 15:25:22 -0800 | [diff] [blame] | 4697 | struct page *page = NULL; |
| 4698 | spinlock_t *ptl; |
Naoya Horiguchi | c9d398f | 2017-03-31 15:11:55 -0700 | [diff] [blame] | 4699 | pte_t pte; |
Naoya Horiguchi | e66f17f | 2015-02-11 15:25:22 -0800 | [diff] [blame] | 4700 | retry: |
| 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 Horiguchi | c9d398f | 2017-03-31 15:11:55 -0700 | [diff] [blame] | 4709 | pte = huge_ptep_get((pte_t *)pmd); |
| 4710 | if (pte_present(pte)) { |
Gerald Schaefer | 9753412 | 2015-04-14 15:42:30 -0700 | [diff] [blame] | 4711 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
Naoya Horiguchi | e66f17f | 2015-02-11 15:25:22 -0800 | [diff] [blame] | 4712 | if (flags & FOLL_GET) |
| 4713 | get_page(page); |
| 4714 | } else { |
Naoya Horiguchi | c9d398f | 2017-03-31 15:11:55 -0700 | [diff] [blame] | 4715 | if (is_hugetlb_entry_migration(pte)) { |
Naoya Horiguchi | e66f17f | 2015-02-11 15:25:22 -0800 | [diff] [blame] | 4716 | 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 | } |
| 4725 | out: |
| 4726 | spin_unlock(ptl); |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4727 | return page; |
| 4728 | } |
| 4729 | |
Naoya Horiguchi | 61f77ed | 2015-02-11 15:25:15 -0800 | [diff] [blame] | 4730 | struct page * __weak |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4731 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
Naoya Horiguchi | e66f17f | 2015-02-11 15:25:22 -0800 | [diff] [blame] | 4732 | pud_t *pud, int flags) |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4733 | { |
Naoya Horiguchi | e66f17f | 2015-02-11 15:25:22 -0800 | [diff] [blame] | 4734 | if (flags & FOLL_GET) |
| 4735 | return NULL; |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4736 | |
Naoya Horiguchi | e66f17f | 2015-02-11 15:25:22 -0800 | [diff] [blame] | 4737 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
Steve Capper | 9e5fc74 | 2013-04-30 08:02:03 +0100 | [diff] [blame] | 4738 | } |
| 4739 | |
Anshuman Khandual | faaa5b6 | 2017-07-06 15:38:50 -0700 | [diff] [blame] | 4740 | struct page * __weak |
| 4741 | follow_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 Horiguchi | 31caf66 | 2013-09-11 14:21:59 -0700 | [diff] [blame] | 4749 | bool isolate_huge_page(struct page *page, struct list_head *list) |
| 4750 | { |
Naoya Horiguchi | bcc5422 | 2015-04-15 16:14:38 -0700 | [diff] [blame] | 4751 | bool ret = true; |
| 4752 | |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 4753 | VM_BUG_ON_PAGE(!PageHead(page), page); |
Naoya Horiguchi | 31caf66 | 2013-09-11 14:21:59 -0700 | [diff] [blame] | 4754 | spin_lock(&hugetlb_lock); |
Naoya Horiguchi | bcc5422 | 2015-04-15 16:14:38 -0700 | [diff] [blame] | 4755 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
| 4756 | ret = false; |
| 4757 | goto unlock; |
| 4758 | } |
| 4759 | clear_page_huge_active(page); |
Naoya Horiguchi | 31caf66 | 2013-09-11 14:21:59 -0700 | [diff] [blame] | 4760 | list_move_tail(&page->lru, list); |
Naoya Horiguchi | bcc5422 | 2015-04-15 16:14:38 -0700 | [diff] [blame] | 4761 | unlock: |
Naoya Horiguchi | 31caf66 | 2013-09-11 14:21:59 -0700 | [diff] [blame] | 4762 | spin_unlock(&hugetlb_lock); |
Naoya Horiguchi | bcc5422 | 2015-04-15 16:14:38 -0700 | [diff] [blame] | 4763 | return ret; |
Naoya Horiguchi | 31caf66 | 2013-09-11 14:21:59 -0700 | [diff] [blame] | 4764 | } |
| 4765 | |
| 4766 | void putback_active_hugepage(struct page *page) |
| 4767 | { |
Sasha Levin | 309381fea | 2014-01-23 15:52:54 -0800 | [diff] [blame] | 4768 | VM_BUG_ON_PAGE(!PageHead(page), page); |
Naoya Horiguchi | 31caf66 | 2013-09-11 14:21:59 -0700 | [diff] [blame] | 4769 | spin_lock(&hugetlb_lock); |
Naoya Horiguchi | bcc5422 | 2015-04-15 16:14:38 -0700 | [diff] [blame] | 4770 | set_page_huge_active(page); |
Naoya Horiguchi | 31caf66 | 2013-09-11 14:21:59 -0700 | [diff] [blame] | 4771 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
| 4772 | spin_unlock(&hugetlb_lock); |
| 4773 | put_page(page); |
| 4774 | } |