blob: 7c02b9dadfb05b28e2aef363523f021f933bdb6d [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01003 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080016#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080017#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070018#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070019#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090021#include <linux/rmap.h>
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +090022#include <linux/swap.h>
23#include <linux/swapops.h>
Naoya Horiguchic8721bb2013-09-11 14:22:09 -070024#include <linux/page-isolation.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070025#include <linux/jhash.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070026
David Gibson63551ae2005-06-21 17:14:44 -070027#include <asm/page.h>
28#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070029#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070030
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070031#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070032#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070033#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a305232009-12-14 17:58:25 -080034#include <linux/node.h>
Nick Piggin7835e982006-03-22 00:08:40 -080035#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070036
37const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
Mel Gorman396faf02007-07-17 04:03:13 -070038unsigned long hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070039
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070040int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070041unsigned int default_hstate_idx;
42struct hstate hstates[HUGE_MAX_HSTATE];
43
Jon Tollefson53ba51d2008-07-23 21:27:52 -070044__initdata LIST_HEAD(huge_boot_pages);
45
Andi Kleene5ff2152008-07-23 21:27:42 -070046/* for command line parsing */
47static struct hstate * __initdata parsed_hstate;
48static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070049static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070050
David Gibson3935baa2006-03-22 00:08:53 -080051/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070052 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
53 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080054 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070055DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080056
Davidlohr Bueso8382d912014-04-03 14:47:31 -070057/*
58 * Serializes faults on the same logical page. This is used to
59 * prevent spurious OOMs when the hugepage pool is fully utilized.
60 */
61static int num_fault_mutexes;
62static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp;
63
David Gibson90481622012-03-21 16:34:12 -070064static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
65{
66 bool free = (spool->count == 0) && (spool->used_hpages == 0);
67
68 spin_unlock(&spool->lock);
69
70 /* If no pages are used, and no other handles to the subpool
71 * remain, free the subpool the subpool remain */
72 if (free)
73 kfree(spool);
74}
75
76struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
77{
78 struct hugepage_subpool *spool;
79
80 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
81 if (!spool)
82 return NULL;
83
84 spin_lock_init(&spool->lock);
85 spool->count = 1;
86 spool->max_hpages = nr_blocks;
87 spool->used_hpages = 0;
88
89 return spool;
90}
91
92void hugepage_put_subpool(struct hugepage_subpool *spool)
93{
94 spin_lock(&spool->lock);
95 BUG_ON(!spool->count);
96 spool->count--;
97 unlock_or_release_subpool(spool);
98}
99
100static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
101 long delta)
102{
103 int ret = 0;
104
105 if (!spool)
106 return 0;
107
108 spin_lock(&spool->lock);
109 if ((spool->used_hpages + delta) <= spool->max_hpages) {
110 spool->used_hpages += delta;
111 } else {
112 ret = -ENOMEM;
113 }
114 spin_unlock(&spool->lock);
115
116 return ret;
117}
118
119static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
120 long delta)
121{
122 if (!spool)
123 return;
124
125 spin_lock(&spool->lock);
126 spool->used_hpages -= delta;
127 /* If hugetlbfs_put_super couldn't free spool due to
128 * an outstanding quota reference, free it now. */
129 unlock_or_release_subpool(spool);
130}
131
132static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
133{
134 return HUGETLBFS_SB(inode->i_sb)->spool;
135}
136
137static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
138{
Al Viro496ad9a2013-01-23 17:07:38 -0500139 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700140}
141
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700142/*
Andy Whitcroft96822902008-07-23 21:27:29 -0700143 * Region tracking -- allows tracking of reservations and instantiated pages
144 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700145 *
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700146 * The region data structures are embedded into a resv_map and
147 * protected by a resv_map's lock
Andy Whitcroft96822902008-07-23 21:27:29 -0700148 */
149struct file_region {
150 struct list_head link;
151 long from;
152 long to;
153};
154
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700155static long region_add(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700156{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700157 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700158 struct file_region *rg, *nrg, *trg;
159
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700160 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700161 /* Locate the region we are either in or before. */
162 list_for_each_entry(rg, head, link)
163 if (f <= rg->to)
164 break;
165
166 /* Round our left edge to the current segment if it encloses us. */
167 if (f > rg->from)
168 f = rg->from;
169
170 /* Check for and consume any regions we now overlap with. */
171 nrg = rg;
172 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
173 if (&rg->link == head)
174 break;
175 if (rg->from > t)
176 break;
177
178 /* If this area reaches higher then extend our area to
179 * include it completely. If this is not the first area
180 * which we intend to reuse, free it. */
181 if (rg->to > t)
182 t = rg->to;
183 if (rg != nrg) {
184 list_del(&rg->link);
185 kfree(rg);
186 }
187 }
188 nrg->from = f;
189 nrg->to = t;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700190 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700191 return 0;
192}
193
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700194static long region_chg(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700195{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700196 struct list_head *head = &resv->regions;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700197 struct file_region *rg, *nrg = NULL;
Andy Whitcroft96822902008-07-23 21:27:29 -0700198 long chg = 0;
199
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700200retry:
201 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700202 /* Locate the region we are before or in. */
203 list_for_each_entry(rg, head, link)
204 if (f <= rg->to)
205 break;
206
207 /* If we are below the current region then a new region is required.
208 * Subtle, allocate a new region at the position but make it zero
209 * size such that we can guarantee to record the reservation. */
210 if (&rg->link == head || t < rg->from) {
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700211 if (!nrg) {
212 spin_unlock(&resv->lock);
213 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
214 if (!nrg)
215 return -ENOMEM;
Andy Whitcroft96822902008-07-23 21:27:29 -0700216
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700217 nrg->from = f;
218 nrg->to = f;
219 INIT_LIST_HEAD(&nrg->link);
220 goto retry;
221 }
222
223 list_add(&nrg->link, rg->link.prev);
224 chg = t - f;
225 goto out_nrg;
Andy Whitcroft96822902008-07-23 21:27:29 -0700226 }
227
228 /* Round our left edge to the current segment if it encloses us. */
229 if (f > rg->from)
230 f = rg->from;
231 chg = t - f;
232
233 /* Check for and consume any regions we now overlap with. */
234 list_for_each_entry(rg, rg->link.prev, link) {
235 if (&rg->link == head)
236 break;
237 if (rg->from > t)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700238 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700239
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300240 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700241 * us then we must extend ourselves. Account for its
242 * existing reservation. */
243 if (rg->to > t) {
244 chg += rg->to - t;
245 t = rg->to;
246 }
247 chg -= rg->to - rg->from;
248 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700249
250out:
251 spin_unlock(&resv->lock);
252 /* We already know we raced and no longer need the new region */
253 kfree(nrg);
254 return chg;
255out_nrg:
256 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700257 return chg;
258}
259
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700260static long region_truncate(struct resv_map *resv, long end)
Andy Whitcroft96822902008-07-23 21:27:29 -0700261{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700262 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700263 struct file_region *rg, *trg;
264 long chg = 0;
265
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700266 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700267 /* Locate the region we are either in or before. */
268 list_for_each_entry(rg, head, link)
269 if (end <= rg->to)
270 break;
271 if (&rg->link == head)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700272 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700273
274 /* If we are in the middle of a region then adjust it. */
275 if (end > rg->from) {
276 chg = rg->to - end;
277 rg->to = end;
278 rg = list_entry(rg->link.next, typeof(*rg), link);
279 }
280
281 /* Drop any remaining regions. */
282 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
283 if (&rg->link == head)
284 break;
285 chg += rg->to - rg->from;
286 list_del(&rg->link);
287 kfree(rg);
288 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700289
290out:
291 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700292 return chg;
293}
294
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700295static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700296{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700297 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700298 struct file_region *rg;
299 long chg = 0;
300
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700301 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700302 /* Locate each segment we overlap with, and count that overlap. */
303 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700304 long seg_from;
305 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700306
307 if (rg->to <= f)
308 continue;
309 if (rg->from >= t)
310 break;
311
312 seg_from = max(rg->from, f);
313 seg_to = min(rg->to, t);
314
315 chg += seg_to - seg_from;
316 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700317 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700318
319 return chg;
320}
321
Andy Whitcroft96822902008-07-23 21:27:29 -0700322/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700323 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700324 * the mapping, in pagecache page units; huge pages here.
325 */
Andi Kleena5516432008-07-23 21:27:41 -0700326static pgoff_t vma_hugecache_offset(struct hstate *h,
327 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700328{
Andi Kleena5516432008-07-23 21:27:41 -0700329 return ((address - vma->vm_start) >> huge_page_shift(h)) +
330 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700331}
332
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900333pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
334 unsigned long address)
335{
336 return vma_hugecache_offset(hstate_vma(vma), vma, address);
337}
338
Andy Whitcroft84afd992008-07-23 21:27:32 -0700339/*
Mel Gorman08fba692009-01-06 14:38:53 -0800340 * Return the size of the pages allocated when backing a VMA. In the majority
341 * cases this will be same size as used by the page table entries.
342 */
343unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
344{
345 struct hstate *hstate;
346
347 if (!is_vm_hugetlb_page(vma))
348 return PAGE_SIZE;
349
350 hstate = hstate_vma(vma);
351
Wanpeng Li2415cf12013-07-03 15:02:43 -0700352 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800353}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200354EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800355
356/*
Mel Gorman33402892009-01-06 14:38:54 -0800357 * Return the page size being used by the MMU to back a VMA. In the majority
358 * of cases, the page size used by the kernel matches the MMU size. On
359 * architectures where it differs, an architecture-specific version of this
360 * function is required.
361 */
362#ifndef vma_mmu_pagesize
363unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
364{
365 return vma_kernel_pagesize(vma);
366}
367#endif
368
369/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700370 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
371 * bits of the reservation map pointer, which are always clear due to
372 * alignment.
373 */
374#define HPAGE_RESV_OWNER (1UL << 0)
375#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700376#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700377
Mel Gormana1e78772008-07-23 21:27:23 -0700378/*
379 * These helpers are used to track how many pages are reserved for
380 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
381 * is guaranteed to have their future faults succeed.
382 *
383 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
384 * the reserve counters are updated with the hugetlb_lock held. It is safe
385 * to reset the VMA at fork() time as it is not in use yet and there is no
386 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700387 *
388 * The private mapping reservation is represented in a subtly different
389 * manner to a shared mapping. A shared mapping has a region map associated
390 * with the underlying file, this region map represents the backing file
391 * pages which have ever had a reservation assigned which this persists even
392 * after the page is instantiated. A private mapping has a region map
393 * associated with the original mmap which is attached to all VMAs which
394 * reference it, this region map represents those offsets which have consumed
395 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700396 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700397static unsigned long get_vma_private_data(struct vm_area_struct *vma)
398{
399 return (unsigned long)vma->vm_private_data;
400}
401
402static void set_vma_private_data(struct vm_area_struct *vma,
403 unsigned long value)
404{
405 vma->vm_private_data = (void *)value;
406}
407
Joonsoo Kim9119a412014-04-03 14:47:25 -0700408struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700409{
410 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
411 if (!resv_map)
412 return NULL;
413
414 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700415 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700416 INIT_LIST_HEAD(&resv_map->regions);
417
418 return resv_map;
419}
420
Joonsoo Kim9119a412014-04-03 14:47:25 -0700421void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700422{
423 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
424
425 /* Clear out any active regions before we release the map. */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700426 region_truncate(resv_map, 0);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700427 kfree(resv_map);
428}
429
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700430static inline struct resv_map *inode_resv_map(struct inode *inode)
431{
432 return inode->i_mapping->private_data;
433}
434
Andy Whitcroft84afd992008-07-23 21:27:32 -0700435static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700436{
437 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700438 if (vma->vm_flags & VM_MAYSHARE) {
439 struct address_space *mapping = vma->vm_file->f_mapping;
440 struct inode *inode = mapping->host;
441
442 return inode_resv_map(inode);
443
444 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700445 return (struct resv_map *)(get_vma_private_data(vma) &
446 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700447 }
Mel Gormana1e78772008-07-23 21:27:23 -0700448}
449
Andy Whitcroft84afd992008-07-23 21:27:32 -0700450static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700451{
452 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700453 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Mel Gormana1e78772008-07-23 21:27:23 -0700454
Andy Whitcroft84afd992008-07-23 21:27:32 -0700455 set_vma_private_data(vma, (get_vma_private_data(vma) &
456 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700457}
458
459static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
460{
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700461 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700462 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700463
464 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700465}
466
467static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
468{
469 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700470
471 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700472}
473
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700474/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700475void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
476{
477 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700478 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700479 vma->vm_private_data = (void *)0;
480}
481
482/* Returns true if the VMA has associated reserve pages */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700483static int vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700484{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700485 if (vma->vm_flags & VM_NORESERVE) {
486 /*
487 * This address is already reserved by other process(chg == 0),
488 * so, we should decrement reserved count. Without decrementing,
489 * reserve count remains after releasing inode, because this
490 * allocated page will go into page cache and is regarded as
491 * coming from reserved pool in releasing step. Currently, we
492 * don't have any other solution to deal with this situation
493 * properly, so add work-around here.
494 */
495 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
496 return 1;
497 else
498 return 0;
499 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700500
501 /* Shared mappings always use reserves */
Mel Gormanf83a2752009-05-28 14:34:40 -0700502 if (vma->vm_flags & VM_MAYSHARE)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700503 return 1;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700504
505 /*
506 * Only the process that called mmap() has reserves for
507 * private mappings.
508 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700509 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
510 return 1;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700511
Mel Gorman7f09ca52008-07-23 21:27:58 -0700512 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700513}
514
Andi Kleena5516432008-07-23 21:27:41 -0700515static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516{
517 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700518 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700519 h->free_huge_pages++;
520 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521}
522
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900523static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
524{
525 struct page *page;
526
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700527 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
528 if (!is_migrate_isolate_page(page))
529 break;
530 /*
531 * if 'non-isolated free hugepage' not found on the list,
532 * the allocation fails.
533 */
534 if (&h->hugepage_freelists[nid] == &page->lru)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900535 return NULL;
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700536 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900537 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900538 h->free_huge_pages--;
539 h->free_huge_pages_node[nid]--;
540 return page;
541}
542
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700543/* Movability of hugepages depends on migration support. */
544static inline gfp_t htlb_alloc_mask(struct hstate *h)
545{
546 if (hugepages_treat_as_movable || hugepage_migration_support(h))
547 return GFP_HIGHUSER_MOVABLE;
548 else
549 return GFP_HIGHUSER;
550}
551
Andi Kleena5516432008-07-23 21:27:41 -0700552static struct page *dequeue_huge_page_vma(struct hstate *h,
553 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700554 unsigned long address, int avoid_reserve,
555 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700557 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700558 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700559 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700560 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700561 struct zone *zone;
562 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700563 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564
Mel Gormana1e78772008-07-23 21:27:23 -0700565 /*
566 * A child process with MAP_PRIVATE mappings created by their parent
567 * have no page reserves. This check ensures that reservations are
568 * not "stolen". The child may still get SIGKILLed
569 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700570 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -0700571 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700572 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700573
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700574 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700575 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700576 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700577
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700578retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -0700579 cpuset_mems_cookie = read_mems_allowed_begin();
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700580 zonelist = huge_zonelist(vma, address,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700581 htlb_alloc_mask(h), &mpol, &nodemask);
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700582
Mel Gorman19770b32008-04-28 02:12:18 -0700583 for_each_zone_zonelist_nodemask(zone, z, zonelist,
584 MAX_NR_ZONES - 1, nodemask) {
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700585 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900586 page = dequeue_huge_page_node(h, zone_to_nid(zone));
587 if (page) {
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700588 if (avoid_reserve)
589 break;
590 if (!vma_has_reserves(vma, chg))
591 break;
592
Joonsoo Kim07443a82013-09-11 14:21:58 -0700593 SetPagePrivate(page);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700594 h->resv_huge_pages--;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900595 break;
596 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700597 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700599
600 mpol_cond_put(mpol);
Mel Gormand26914d2014-04-03 14:47:24 -0700601 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -0700602 goto retry_cpuset;
603 return page;
604
Miao Xiec0ff7452010-05-24 14:32:08 -0700605err:
Mel Gormancc9a6c82012-03-21 16:34:11 -0700606 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607}
608
Andi Kleena5516432008-07-23 21:27:41 -0700609static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700610{
611 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700612
Andy Whitcroft18229df2008-11-06 12:53:27 -0800613 VM_BUG_ON(h->order >= MAX_ORDER);
614
Andi Kleena5516432008-07-23 21:27:41 -0700615 h->nr_huge_pages--;
616 h->nr_huge_pages_node[page_to_nid(page)]--;
617 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -0700618 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
619 1 << PG_referenced | 1 << PG_dirty |
620 1 << PG_active | 1 << PG_reserved |
621 1 << PG_private | 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -0700622 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800623 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Adam Litke6af2acb2007-10-16 01:26:16 -0700624 set_compound_page_dtor(page, NULL);
625 set_page_refcounted(page);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700626 arch_release_hugepage(page);
Andi Kleena5516432008-07-23 21:27:41 -0700627 __free_pages(page, huge_page_order(h));
Adam Litke6af2acb2007-10-16 01:26:16 -0700628}
629
Andi Kleene5ff2152008-07-23 21:27:42 -0700630struct hstate *size_to_hstate(unsigned long size)
631{
632 struct hstate *h;
633
634 for_each_hstate(h) {
635 if (huge_page_size(h) == size)
636 return h;
637 }
638 return NULL;
639}
640
David Gibson27a85ef2006-03-22 00:08:56 -0800641static void free_huge_page(struct page *page)
642{
Andi Kleena5516432008-07-23 21:27:41 -0700643 /*
644 * Can't pass hstate in here because it is called from the
645 * compound page destructor.
646 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700647 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700648 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -0700649 struct hugepage_subpool *spool =
650 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -0700651 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -0800652
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800653 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -0400654 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -0700655 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900656 BUG_ON(page_mapcount(page));
Joonsoo Kim07443a82013-09-11 14:21:58 -0700657 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -0700658 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -0800659
660 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -0700661 hugetlb_cgroup_uncharge_page(hstate_index(h),
662 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -0700663 if (restore_reserve)
664 h->resv_huge_pages++;
665
Andi Kleenaa888a72008-07-23 21:27:47 -0700666 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700667 /* remove the page from active list */
668 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700669 update_and_free_page(h, page);
670 h->surplus_huge_pages--;
671 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700672 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -0700673 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -0700674 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700675 }
David Gibson27a85ef2006-03-22 00:08:56 -0800676 spin_unlock(&hugetlb_lock);
David Gibson90481622012-03-21 16:34:12 -0700677 hugepage_subpool_put_pages(spool, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800678}
679
Andi Kleena5516432008-07-23 21:27:41 -0700680static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700681{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700682 INIT_LIST_HEAD(&page->lru);
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700683 set_compound_page_dtor(page, free_huge_page);
684 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700685 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -0700686 h->nr_huge_pages++;
687 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700688 spin_unlock(&hugetlb_lock);
689 put_page(page); /* free it into the hugepage allocator */
690}
691
David Rientjesf412c972014-04-03 14:47:59 -0700692static void __init prep_compound_gigantic_page(struct page *page,
693 unsigned long order)
Wu Fengguang20a03072009-06-16 15:32:22 -0700694{
695 int i;
696 int nr_pages = 1 << order;
697 struct page *p = page + 1;
698
699 /* we rely on prep_new_huge_page to set the destructor */
700 set_compound_order(page, order);
701 __SetPageHead(page);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -0700702 __ClearPageReserved(page);
Wu Fengguang20a03072009-06-16 15:32:22 -0700703 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
704 __SetPageTail(p);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -0700705 /*
706 * For gigantic hugepages allocated through bootmem at
707 * boot, it's safer to be consistent with the not-gigantic
708 * hugepages and clear the PG_reserved bit from all tail pages
709 * too. Otherwse drivers using get_user_pages() to access tail
710 * pages may get the reference counting wrong if they see
711 * PG_reserved set on a tail page (despite the head page not
712 * having PG_reserved set). Enforcing this consistency between
713 * head and tail pages allows drivers to optimize away a check
714 * on the head page when they need know if put_page() is needed
715 * after get_user_pages().
716 */
717 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -0800718 set_page_count(p, 0);
Wu Fengguang20a03072009-06-16 15:32:22 -0700719 p->first_page = page;
720 }
721}
722
Andrew Morton77959122012-10-08 16:34:11 -0700723/*
724 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
725 * transparent huge pages. See the PageTransHuge() documentation for more
726 * details.
727 */
Wu Fengguang20a03072009-06-16 15:32:22 -0700728int PageHuge(struct page *page)
729{
Wu Fengguang20a03072009-06-16 15:32:22 -0700730 if (!PageCompound(page))
731 return 0;
732
733 page = compound_head(page);
Andrew Morton758f66a2014-01-21 15:48:57 -0800734 return get_compound_page_dtor(page) == free_huge_page;
Wu Fengguang20a03072009-06-16 15:32:22 -0700735}
Naoya Horiguchi43131e12010-05-28 09:29:22 +0900736EXPORT_SYMBOL_GPL(PageHuge);
737
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800738/*
739 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
740 * normal or transparent huge pages.
741 */
742int PageHeadHuge(struct page *page_head)
743{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800744 if (!PageHead(page_head))
745 return 0;
746
Andrew Morton758f66a2014-01-21 15:48:57 -0800747 return get_compound_page_dtor(page_head) == free_huge_page;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800748}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800749
Zhang Yi13d60f42013-06-25 21:19:31 +0800750pgoff_t __basepage_index(struct page *page)
751{
752 struct page *page_head = compound_head(page);
753 pgoff_t index = page_index(page_head);
754 unsigned long compound_idx;
755
756 if (!PageHuge(page_head))
757 return page_index(page);
758
759 if (compound_order(page_head) >= MAX_ORDER)
760 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
761 else
762 compound_idx = page - page_head;
763
764 return (index << compound_order(page_head)) + compound_idx;
765}
766
Andi Kleena5516432008-07-23 21:27:41 -0700767static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700770
Andi Kleenaa888a72008-07-23 21:27:47 -0700771 if (h->order >= MAX_ORDER)
772 return NULL;
773
Mel Gorman6484eb32009-06-16 15:31:54 -0700774 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700775 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700776 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700777 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700779 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700780 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700781 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700782 }
Andi Kleena5516432008-07-23 21:27:41 -0700783 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700785
786 return page;
787}
788
Andi Kleen5ced66c2008-07-23 21:27:45 -0700789/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800790 * common helper functions for hstate_next_node_to_{alloc|free}.
791 * We may have allocated or freed a huge page based on a different
792 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
793 * be outside of *nodes_allowed. Ensure that we use an allowed
794 * node for alloc or free.
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800795 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800796static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800797{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800798 nid = next_node(nid, *nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800799 if (nid == MAX_NUMNODES)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800800 nid = first_node(*nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800801 VM_BUG_ON(nid >= MAX_NUMNODES);
802
803 return nid;
804}
805
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800806static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
Andi Kleen5ced66c2008-07-23 21:27:45 -0700807{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800808 if (!node_isset(nid, *nodes_allowed))
809 nid = next_node_allowed(nid, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800810 return nid;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700811}
812
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800813/*
814 * returns the previously saved node ["this node"] from which to
815 * allocate a persistent huge page for the pool and advance the
816 * next node from which to allocate, handling wrap at end of node
817 * mask.
818 */
819static int hstate_next_node_to_alloc(struct hstate *h,
820 nodemask_t *nodes_allowed)
821{
822 int nid;
823
824 VM_BUG_ON(!nodes_allowed);
825
826 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
827 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
828
829 return nid;
830}
831
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700832/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800833 * helper for free_pool_huge_page() - return the previously saved
834 * node ["this node"] from which to free a huge page. Advance the
835 * next node id whether or not we find a free huge page to free so
836 * that the next attempt to free addresses the next node.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700837 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800838static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700839{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800840 int nid;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800841
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800842 VM_BUG_ON(!nodes_allowed);
843
844 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
845 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
846
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800847 return nid;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700848}
849
Joonsoo Kimb2261022013-09-11 14:21:00 -0700850#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
851 for (nr_nodes = nodes_weight(*mask); \
852 nr_nodes > 0 && \
853 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
854 nr_nodes--)
855
856#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
857 for (nr_nodes = nodes_weight(*mask); \
858 nr_nodes > 0 && \
859 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
860 nr_nodes--)
861
862static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
863{
864 struct page *page;
865 int nr_nodes, node;
866 int ret = 0;
867
868 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
869 page = alloc_fresh_huge_page_node(h, node);
870 if (page) {
871 ret = 1;
872 break;
873 }
874 }
875
876 if (ret)
877 count_vm_event(HTLB_BUDDY_PGALLOC);
878 else
879 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
880
881 return ret;
882}
883
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700884/*
885 * Free huge page from pool from next node to free.
886 * Attempt to keep persistent huge pages more or less
887 * balanced over allowed nodes.
888 * Called with hugetlb_lock locked.
889 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800890static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
891 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700892{
Joonsoo Kimb2261022013-09-11 14:21:00 -0700893 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700894 int ret = 0;
895
Joonsoo Kimb2261022013-09-11 14:21:00 -0700896 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700897 /*
898 * If we're returning unused surplus pages, only examine
899 * nodes with surplus pages.
900 */
Joonsoo Kimb2261022013-09-11 14:21:00 -0700901 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
902 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700903 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -0700904 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700905 struct page, lru);
906 list_del(&page->lru);
907 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -0700908 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700909 if (acct_surplus) {
910 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -0700911 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700912 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700913 update_and_free_page(h, page);
914 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800915 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700916 }
Joonsoo Kimb2261022013-09-11 14:21:00 -0700917 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700918
919 return ret;
920}
921
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700922/*
923 * Dissolve a given free hugepage into free buddy pages. This function does
924 * nothing for in-use (including surplus) hugepages.
925 */
926static void dissolve_free_huge_page(struct page *page)
927{
928 spin_lock(&hugetlb_lock);
929 if (PageHuge(page) && !page_count(page)) {
930 struct hstate *h = page_hstate(page);
931 int nid = page_to_nid(page);
932 list_del(&page->lru);
933 h->free_huge_pages--;
934 h->free_huge_pages_node[nid]--;
935 update_and_free_page(h, page);
936 }
937 spin_unlock(&hugetlb_lock);
938}
939
940/*
941 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
942 * make specified memory blocks removable from the system.
943 * Note that start_pfn should aligned with (minimum) hugepage size.
944 */
945void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
946{
947 unsigned int order = 8 * sizeof(void *);
948 unsigned long pfn;
949 struct hstate *h;
950
951 /* Set scan step to minimum hugepage size */
952 for_each_hstate(h)
953 if (order > huge_page_order(h))
954 order = huge_page_order(h);
955 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
956 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
957 dissolve_free_huge_page(pfn_to_page(pfn));
958}
959
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900960static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -0700961{
962 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900963 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700964
Andi Kleenaa888a72008-07-23 21:27:47 -0700965 if (h->order >= MAX_ORDER)
966 return NULL;
967
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800968 /*
969 * Assume we will successfully allocate the surplus page to
970 * prevent racing processes from causing the surplus to exceed
971 * overcommit
972 *
973 * This however introduces a different race, where a process B
974 * tries to grow the static hugepage pool while alloc_pages() is
975 * called by process A. B will only examine the per-node
976 * counters in determining if surplus huge pages can be
977 * converted to normal huge pages in adjust_pool_surplus(). A
978 * won't be able to increment the per-node counter, until the
979 * lock is dropped by B, but B doesn't drop hugetlb_lock until
980 * no more huge pages can be converted from surplus to normal
981 * state (and doesn't try to convert again). Thus, we have a
982 * case where a surplus huge page exists, the pool is grown, and
983 * the surplus huge page still exists after, even though it
984 * should just have been converted to a normal huge page. This
985 * does not leak memory, though, as the hugepage will be freed
986 * once it is out of use. It also does not allow the counters to
987 * go out of whack in adjust_pool_surplus() as we don't modify
988 * the node values until we've gotten the hugepage and only the
989 * per-node value is checked there.
990 */
991 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700992 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800993 spin_unlock(&hugetlb_lock);
994 return NULL;
995 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700996 h->nr_huge_pages++;
997 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800998 }
999 spin_unlock(&hugetlb_lock);
1000
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001001 if (nid == NUMA_NO_NODE)
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001002 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001003 __GFP_REPEAT|__GFP_NOWARN,
1004 huge_page_order(h));
1005 else
1006 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001007 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001008 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001009
Gerald Schaefercaff3a22008-08-12 15:08:38 -07001010 if (page && arch_prepare_hugepage(page)) {
1011 __free_pages(page, huge_page_order(h));
Hillf Dantonea5768c2012-01-10 15:08:30 -08001012 page = NULL;
Gerald Schaefercaff3a22008-08-12 15:08:38 -07001013 }
1014
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001015 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001016 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001017 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001018 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001019 set_compound_page_dtor(page, free_huge_page);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001020 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001021 /*
1022 * We incremented the global counters already
1023 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001024 h->nr_huge_pages_node[r_nid]++;
1025 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001026 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001027 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001028 h->nr_huge_pages--;
1029 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001030 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001031 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001032 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001033
1034 return page;
1035}
1036
Adam Litkee4e574b2007-10-16 01:26:19 -07001037/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001038 * This allocation function is useful in the context where vma is irrelevant.
1039 * E.g. soft-offlining uses this function because it only cares physical
1040 * address of error page.
1041 */
1042struct page *alloc_huge_page_node(struct hstate *h, int nid)
1043{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001044 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001045
1046 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001047 if (h->free_huge_pages - h->resv_huge_pages > 0)
1048 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001049 spin_unlock(&hugetlb_lock);
1050
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001051 if (!page)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001052 page = alloc_buddy_huge_page(h, nid);
1053
1054 return page;
1055}
1056
1057/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001058 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001059 * of size 'delta'.
1060 */
Andi Kleena5516432008-07-23 21:27:41 -07001061static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001062{
1063 struct list_head surplus_list;
1064 struct page *page, *tmp;
1065 int ret, i;
1066 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001067 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001068
Andi Kleena5516432008-07-23 21:27:41 -07001069 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001070 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001071 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001072 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001073 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001074
1075 allocated = 0;
1076 INIT_LIST_HEAD(&surplus_list);
1077
1078 ret = -ENOMEM;
1079retry:
1080 spin_unlock(&hugetlb_lock);
1081 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001082 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001083 if (!page) {
1084 alloc_ok = false;
1085 break;
1086 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001087 list_add(&page->lru, &surplus_list);
1088 }
Hillf Danton28073b02012-03-21 16:34:00 -07001089 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001090
1091 /*
1092 * After retaking hugetlb_lock, we need to recalculate 'needed'
1093 * because either resv_huge_pages or free_huge_pages may have changed.
1094 */
1095 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001096 needed = (h->resv_huge_pages + delta) -
1097 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001098 if (needed > 0) {
1099 if (alloc_ok)
1100 goto retry;
1101 /*
1102 * We were not able to allocate enough pages to
1103 * satisfy the entire reservation so we free what
1104 * we've allocated so far.
1105 */
1106 goto free;
1107 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001108 /*
1109 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001110 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001111 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001112 * allocator. Commit the entire reservation here to prevent another
1113 * process from stealing the pages as they are added to the pool but
1114 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001115 */
1116 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001117 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001118 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001119
Adam Litke19fc3f02008-04-28 02:12:20 -07001120 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001121 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001122 if ((--needed) < 0)
1123 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001124 /*
1125 * This page is now managed by the hugetlb allocator and has
1126 * no users -- drop the buddy allocator's reference.
1127 */
1128 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001129 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001130 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001131 }
Hillf Danton28073b02012-03-21 16:34:00 -07001132free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001133 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001134
1135 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001136 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1137 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001138 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001139
1140 return ret;
1141}
1142
1143/*
1144 * When releasing a hugetlb pool reservation, any surplus pages that were
1145 * allocated to satisfy the reservation must be explicitly freed if they were
1146 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001147 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001148 */
Andi Kleena5516432008-07-23 21:27:41 -07001149static void return_unused_surplus_pages(struct hstate *h,
1150 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001151{
Adam Litkee4e574b2007-10-16 01:26:19 -07001152 unsigned long nr_pages;
1153
Adam Litkeac09b3a2008-03-04 14:29:38 -08001154 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001155 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001156
Andi Kleenaa888a72008-07-23 21:27:47 -07001157 /* Cannot return gigantic pages currently */
1158 if (h->order >= MAX_ORDER)
1159 return;
1160
Andi Kleena5516432008-07-23 21:27:41 -07001161 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001162
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001163 /*
1164 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001165 * evenly across all nodes with memory. Iterate across these nodes
1166 * until we can no longer free unreserved surplus pages. This occurs
1167 * when the nodes with surplus pages have no free pages.
1168 * free_pool_huge_page() will balance the the freed pages across the
1169 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001170 */
1171 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001172 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001173 break;
Adam Litkee4e574b2007-10-16 01:26:19 -07001174 }
1175}
1176
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001177/*
1178 * Determine if the huge page at addr within the vma has an associated
1179 * reservation. Where it does not we will need to logically increase
David Gibson90481622012-03-21 16:34:12 -07001180 * reservation and actually increase subpool usage before an allocation
1181 * can occur. Where any new reservation would be required the
1182 * reservation change is prepared, but not committed. Once the page
1183 * has been allocated from the subpool and instantiated the change should
1184 * be committed via vma_commit_reservation. No action is required on
1185 * failure.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001186 */
Roel Kluine2f17d92009-03-31 15:23:15 -07001187static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001188 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001189{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001190 struct resv_map *resv;
1191 pgoff_t idx;
1192 long chg;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001193
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001194 resv = vma_resv_map(vma);
1195 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001196 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001197
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001198 idx = vma_hugecache_offset(h, vma, addr);
1199 chg = region_chg(resv, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001200
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001201 if (vma->vm_flags & VM_MAYSHARE)
1202 return chg;
1203 else
1204 return chg < 0 ? chg : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001205}
Andi Kleena5516432008-07-23 21:27:41 -07001206static void vma_commit_reservation(struct hstate *h,
1207 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001208{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001209 struct resv_map *resv;
1210 pgoff_t idx;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001211
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001212 resv = vma_resv_map(vma);
1213 if (!resv)
1214 return;
Joonsoo Kim9119a412014-04-03 14:47:25 -07001215
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001216 idx = vma_hugecache_offset(h, vma, addr);
1217 region_add(resv, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001218}
1219
David Gibson27a85ef2006-03-22 00:08:56 -08001220static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001221 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222{
David Gibson90481622012-03-21 16:34:12 -07001223 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001224 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001225 struct page *page;
Roel Kluine2f17d92009-03-31 15:23:15 -07001226 long chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001227 int ret, idx;
1228 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001229
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001230 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001231 /*
David Gibson90481622012-03-21 16:34:12 -07001232 * Processes that did not create the mapping will have no
1233 * reserves and will not have accounted against subpool
1234 * limit. Check that the subpool limit can be made before
1235 * satisfying the allocation MAP_NORESERVE mappings may also
1236 * need pages and subpool limit allocated allocated if no reserve
1237 * mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001238 */
Andi Kleena5516432008-07-23 21:27:41 -07001239 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001240 if (chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001241 return ERR_PTR(-ENOMEM);
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001242 if (chg || avoid_reserve)
1243 if (hugepage_subpool_get_pages(spool, 1))
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001244 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001245
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001246 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1247 if (ret) {
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001248 if (chg || avoid_reserve)
1249 hugepage_subpool_put_pages(spool, 1);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001250 return ERR_PTR(-ENOSPC);
1251 }
Mel Gormana1e78772008-07-23 21:27:23 -07001252 spin_lock(&hugetlb_lock);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001253 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001254 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001255 spin_unlock(&hugetlb_lock);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001256 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Mel Gormana1e78772008-07-23 21:27:23 -07001257 if (!page) {
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001258 hugetlb_cgroup_uncharge_cgroup(idx,
1259 pages_per_huge_page(h),
1260 h_cg);
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001261 if (chg || avoid_reserve)
1262 hugepage_subpool_put_pages(spool, 1);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001263 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001264 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001265 spin_lock(&hugetlb_lock);
1266 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001267 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001268 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001269 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1270 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001271
David Gibson90481622012-03-21 16:34:12 -07001272 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001273
Andi Kleena5516432008-07-23 21:27:41 -07001274 vma_commit_reservation(h, vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -08001275 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001276}
1277
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001278/*
1279 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1280 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1281 * where no ERR_VALUE is expected to be returned.
1282 */
1283struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1284 unsigned long addr, int avoid_reserve)
1285{
1286 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1287 if (IS_ERR(page))
1288 page = NULL;
1289 return page;
1290}
1291
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001292int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001293{
1294 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001295 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07001296
Joonsoo Kimb2261022013-09-11 14:21:00 -07001297 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001298 void *addr;
1299
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001300 addr = memblock_virt_alloc_try_nid_nopanic(
1301 huge_page_size(h), huge_page_size(h),
1302 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07001303 if (addr) {
1304 /*
1305 * Use the beginning of the huge page to store the
1306 * huge_bootmem_page struct (until gather_bootmem
1307 * puts them into the mem_map).
1308 */
1309 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001310 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001311 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001312 }
1313 return 0;
1314
1315found:
1316 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1317 /* Put them into a private list first because mem_map is not up yet */
1318 list_add(&m->list, &huge_boot_pages);
1319 m->hstate = h;
1320 return 1;
1321}
1322
David Rientjesf412c972014-04-03 14:47:59 -07001323static void __init prep_compound_huge_page(struct page *page, int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08001324{
1325 if (unlikely(order > (MAX_ORDER - 1)))
1326 prep_compound_gigantic_page(page, order);
1327 else
1328 prep_compound_page(page, order);
1329}
1330
Andi Kleenaa888a72008-07-23 21:27:47 -07001331/* Put bootmem huge pages into the standard lists after mem_map is up */
1332static void __init gather_bootmem_prealloc(void)
1333{
1334 struct huge_bootmem_page *m;
1335
1336 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001337 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001338 struct page *page;
1339
1340#ifdef CONFIG_HIGHMEM
1341 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001342 memblock_free_late(__pa(m),
1343 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07001344#else
1345 page = virt_to_page(m);
1346#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001347 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001348 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001349 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07001350 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001351 /*
1352 * If we had gigantic hugepages allocated at boot time, we need
1353 * to restore the 'stolen' pages to totalram_pages in order to
1354 * fix confusing memory reports from free(1) and another
1355 * side-effects, like CommitLimit going negative.
1356 */
1357 if (h->order > (MAX_ORDER - 1))
Jiang Liu3dcc0572013-07-03 15:03:21 -07001358 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001359 }
1360}
1361
Andi Kleen8faa8b02008-07-23 21:27:48 -07001362static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363{
1364 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365
Andi Kleene5ff2152008-07-23 21:27:42 -07001366 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001367 if (h->order >= MAX_ORDER) {
1368 if (!alloc_bootmem_huge_page(h))
1369 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001370 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001371 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001374 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001375}
1376
1377static void __init hugetlb_init_hstates(void)
1378{
1379 struct hstate *h;
1380
1381 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001382 /* oversize hugepages were init'ed in early boot */
1383 if (h->order < MAX_ORDER)
1384 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001385 }
1386}
1387
Andi Kleen4abd32d2008-07-23 21:27:49 -07001388static char * __init memfmt(char *buf, unsigned long n)
1389{
1390 if (n >= (1UL << 30))
1391 sprintf(buf, "%lu GB", n >> 30);
1392 else if (n >= (1UL << 20))
1393 sprintf(buf, "%lu MB", n >> 20);
1394 else
1395 sprintf(buf, "%lu KB", n >> 10);
1396 return buf;
1397}
1398
Andi Kleene5ff2152008-07-23 21:27:42 -07001399static void __init report_hugepages(void)
1400{
1401 struct hstate *h;
1402
1403 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001404 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08001405 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07001406 memfmt(buf, huge_page_size(h)),
1407 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001408 }
1409}
1410
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001412static void try_to_free_low(struct hstate *h, unsigned long count,
1413 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001415 int i;
1416
Andi Kleenaa888a72008-07-23 21:27:47 -07001417 if (h->order >= MAX_ORDER)
1418 return;
1419
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001420 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001422 struct list_head *freel = &h->hugepage_freelists[i];
1423 list_for_each_entry_safe(page, next, freel, lru) {
1424 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001425 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 if (PageHighMem(page))
1427 continue;
1428 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001429 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001430 h->free_huge_pages--;
1431 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 }
1433 }
1434}
1435#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001436static inline void try_to_free_low(struct hstate *h, unsigned long count,
1437 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438{
1439}
1440#endif
1441
Wu Fengguang20a03072009-06-16 15:32:22 -07001442/*
1443 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1444 * balanced by operating on them in a round-robin fashion.
1445 * Returns 1 if an adjustment was made.
1446 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001447static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1448 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001449{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001450 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07001451
1452 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001453
Joonsoo Kimb2261022013-09-11 14:21:00 -07001454 if (delta < 0) {
1455 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1456 if (h->surplus_huge_pages_node[node])
1457 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001458 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001459 } else {
1460 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1461 if (h->surplus_huge_pages_node[node] <
1462 h->nr_huge_pages_node[node])
1463 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001464 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001465 }
1466 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07001467
Joonsoo Kimb2261022013-09-11 14:21:00 -07001468found:
1469 h->surplus_huge_pages += delta;
1470 h->surplus_huge_pages_node[node] += delta;
1471 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07001472}
1473
Andi Kleena5516432008-07-23 21:27:41 -07001474#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001475static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1476 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477{
Adam Litke7893d1d2007-10-16 01:26:18 -07001478 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479
Andi Kleenaa888a72008-07-23 21:27:47 -07001480 if (h->order >= MAX_ORDER)
1481 return h->max_huge_pages;
1482
Adam Litke7893d1d2007-10-16 01:26:18 -07001483 /*
1484 * Increase the pool size
1485 * First take pages out of surplus state. Then make up the
1486 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001487 *
1488 * We might race with alloc_buddy_huge_page() here and be unable
1489 * to convert a surplus huge page to a normal huge page. That is
1490 * not critical, though, it just means the overall size of the
1491 * pool might be one hugepage larger than it needs to be, but
1492 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001493 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001495 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001496 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001497 break;
1498 }
1499
Andi Kleena5516432008-07-23 21:27:41 -07001500 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001501 /*
1502 * If this allocation races such that we no longer need the
1503 * page, free_huge_page will handle it by freeing the page
1504 * and reducing the surplus.
1505 */
1506 spin_unlock(&hugetlb_lock);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001507 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001508 spin_lock(&hugetlb_lock);
1509 if (!ret)
1510 goto out;
1511
Mel Gorman536240f22009-12-14 17:59:56 -08001512 /* Bail for signals. Probably ctrl-c from user */
1513 if (signal_pending(current))
1514 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001515 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001516
1517 /*
1518 * Decrease the pool size
1519 * First return free pages to the buddy allocator (being careful
1520 * to keep enough around to satisfy reservations). Then place
1521 * pages into surplus state as needed so the pool will shrink
1522 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001523 *
1524 * By placing pages into the surplus state independent of the
1525 * overcommit value, we are allowing the surplus pool size to
1526 * exceed overcommit. There are few sane options here. Since
1527 * alloc_buddy_huge_page() is checking the global counter,
1528 * though, we'll note that we're not allowed to exceed surplus
1529 * and won't grow the pool anywhere else. Not until one of the
1530 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001531 */
Andi Kleena5516432008-07-23 21:27:41 -07001532 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001533 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001534 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001535 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001536 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538 }
Andi Kleena5516432008-07-23 21:27:41 -07001539 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001540 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001541 break;
1542 }
1543out:
Andi Kleena5516432008-07-23 21:27:41 -07001544 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001546 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547}
1548
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001549#define HSTATE_ATTR_RO(_name) \
1550 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1551
1552#define HSTATE_ATTR(_name) \
1553 static struct kobj_attribute _name##_attr = \
1554 __ATTR(_name, 0644, _name##_show, _name##_store)
1555
1556static struct kobject *hugepages_kobj;
1557static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1558
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001559static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1560
1561static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001562{
1563 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001564
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001565 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001566 if (hstate_kobjs[i] == kobj) {
1567 if (nidp)
1568 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001569 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001570 }
1571
1572 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001573}
1574
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001575static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001576 struct kobj_attribute *attr, char *buf)
1577{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001578 struct hstate *h;
1579 unsigned long nr_huge_pages;
1580 int nid;
1581
1582 h = kobj_to_hstate(kobj, &nid);
1583 if (nid == NUMA_NO_NODE)
1584 nr_huge_pages = h->nr_huge_pages;
1585 else
1586 nr_huge_pages = h->nr_huge_pages_node[nid];
1587
1588 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001589}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001590
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001591static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1592 struct kobject *kobj, struct kobj_attribute *attr,
1593 const char *buf, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001594{
1595 int err;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001596 int nid;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001597 unsigned long count;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001598 struct hstate *h;
David Rientjesbad44b52009-12-14 17:58:38 -08001599 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001600
Jingoo Han3dbb95f2013-09-11 14:20:25 -07001601 err = kstrtoul(buf, 10, &count);
Eric B Munson73ae31e2011-01-13 15:47:28 -08001602 if (err)
Eric B Munsonadbe8722011-01-13 15:47:27 -08001603 goto out;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001604
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001605 h = kobj_to_hstate(kobj, &nid);
Eric B Munsonadbe8722011-01-13 15:47:27 -08001606 if (h->order >= MAX_ORDER) {
1607 err = -EINVAL;
1608 goto out;
1609 }
1610
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001611 if (nid == NUMA_NO_NODE) {
1612 /*
1613 * global hstate attribute
1614 */
1615 if (!(obey_mempolicy &&
1616 init_nodemask_of_mempolicy(nodes_allowed))) {
1617 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001618 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001619 }
1620 } else if (nodes_allowed) {
1621 /*
1622 * per node hstate attribute: adjust count to global,
1623 * but restrict alloc/free to the specified node.
1624 */
1625 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1626 init_nodemask_of_node(nodes_allowed, nid);
1627 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001628 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001629
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001630 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001631
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001632 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001633 NODEMASK_FREE(nodes_allowed);
1634
1635 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001636out:
1637 NODEMASK_FREE(nodes_allowed);
1638 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001639}
1640
1641static ssize_t nr_hugepages_show(struct kobject *kobj,
1642 struct kobj_attribute *attr, char *buf)
1643{
1644 return nr_hugepages_show_common(kobj, attr, buf);
1645}
1646
1647static ssize_t nr_hugepages_store(struct kobject *kobj,
1648 struct kobj_attribute *attr, const char *buf, size_t len)
1649{
1650 return nr_hugepages_store_common(false, kobj, attr, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001651}
1652HSTATE_ATTR(nr_hugepages);
1653
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001654#ifdef CONFIG_NUMA
1655
1656/*
1657 * hstate attribute for optionally mempolicy-based constraint on persistent
1658 * huge page alloc/free.
1659 */
1660static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1661 struct kobj_attribute *attr, char *buf)
1662{
1663 return nr_hugepages_show_common(kobj, attr, buf);
1664}
1665
1666static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1667 struct kobj_attribute *attr, const char *buf, size_t len)
1668{
1669 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1670}
1671HSTATE_ATTR(nr_hugepages_mempolicy);
1672#endif
1673
1674
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001675static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1676 struct kobj_attribute *attr, char *buf)
1677{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001678 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001679 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1680}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001681
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001682static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1683 struct kobj_attribute *attr, const char *buf, size_t count)
1684{
1685 int err;
1686 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001687 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001688
Eric B Munsonadbe8722011-01-13 15:47:27 -08001689 if (h->order >= MAX_ORDER)
1690 return -EINVAL;
1691
Jingoo Han3dbb95f2013-09-11 14:20:25 -07001692 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001693 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001694 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001695
1696 spin_lock(&hugetlb_lock);
1697 h->nr_overcommit_huge_pages = input;
1698 spin_unlock(&hugetlb_lock);
1699
1700 return count;
1701}
1702HSTATE_ATTR(nr_overcommit_hugepages);
1703
1704static ssize_t free_hugepages_show(struct kobject *kobj,
1705 struct kobj_attribute *attr, char *buf)
1706{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001707 struct hstate *h;
1708 unsigned long free_huge_pages;
1709 int nid;
1710
1711 h = kobj_to_hstate(kobj, &nid);
1712 if (nid == NUMA_NO_NODE)
1713 free_huge_pages = h->free_huge_pages;
1714 else
1715 free_huge_pages = h->free_huge_pages_node[nid];
1716
1717 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001718}
1719HSTATE_ATTR_RO(free_hugepages);
1720
1721static ssize_t resv_hugepages_show(struct kobject *kobj,
1722 struct kobj_attribute *attr, char *buf)
1723{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001724 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001725 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1726}
1727HSTATE_ATTR_RO(resv_hugepages);
1728
1729static ssize_t surplus_hugepages_show(struct kobject *kobj,
1730 struct kobj_attribute *attr, char *buf)
1731{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001732 struct hstate *h;
1733 unsigned long surplus_huge_pages;
1734 int nid;
1735
1736 h = kobj_to_hstate(kobj, &nid);
1737 if (nid == NUMA_NO_NODE)
1738 surplus_huge_pages = h->surplus_huge_pages;
1739 else
1740 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1741
1742 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001743}
1744HSTATE_ATTR_RO(surplus_hugepages);
1745
1746static struct attribute *hstate_attrs[] = {
1747 &nr_hugepages_attr.attr,
1748 &nr_overcommit_hugepages_attr.attr,
1749 &free_hugepages_attr.attr,
1750 &resv_hugepages_attr.attr,
1751 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001752#ifdef CONFIG_NUMA
1753 &nr_hugepages_mempolicy_attr.attr,
1754#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001755 NULL,
1756};
1757
1758static struct attribute_group hstate_attr_group = {
1759 .attrs = hstate_attrs,
1760};
1761
Jeff Mahoney094e9532010-02-02 13:44:14 -08001762static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1763 struct kobject **hstate_kobjs,
1764 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001765{
1766 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001767 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001768
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001769 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1770 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001771 return -ENOMEM;
1772
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001773 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001774 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001775 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001776
1777 return retval;
1778}
1779
1780static void __init hugetlb_sysfs_init(void)
1781{
1782 struct hstate *h;
1783 int err;
1784
1785 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1786 if (!hugepages_kobj)
1787 return;
1788
1789 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001790 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1791 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001792 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08001793 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001794 }
1795}
1796
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001797#ifdef CONFIG_NUMA
1798
1799/*
1800 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001801 * with node devices in node_devices[] using a parallel array. The array
1802 * index of a node device or _hstate == node id.
1803 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001804 * the base kernel, on the hugetlb module.
1805 */
1806struct node_hstate {
1807 struct kobject *hugepages_kobj;
1808 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1809};
1810struct node_hstate node_hstates[MAX_NUMNODES];
1811
1812/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001813 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001814 */
1815static struct attribute *per_node_hstate_attrs[] = {
1816 &nr_hugepages_attr.attr,
1817 &free_hugepages_attr.attr,
1818 &surplus_hugepages_attr.attr,
1819 NULL,
1820};
1821
1822static struct attribute_group per_node_hstate_attr_group = {
1823 .attrs = per_node_hstate_attrs,
1824};
1825
1826/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001827 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001828 * Returns node id via non-NULL nidp.
1829 */
1830static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1831{
1832 int nid;
1833
1834 for (nid = 0; nid < nr_node_ids; nid++) {
1835 struct node_hstate *nhs = &node_hstates[nid];
1836 int i;
1837 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1838 if (nhs->hstate_kobjs[i] == kobj) {
1839 if (nidp)
1840 *nidp = nid;
1841 return &hstates[i];
1842 }
1843 }
1844
1845 BUG();
1846 return NULL;
1847}
1848
1849/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001850 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001851 * No-op if no hstate attributes attached.
1852 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02001853static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001854{
1855 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001856 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001857
1858 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001859 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001860
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001861 for_each_hstate(h) {
1862 int idx = hstate_index(h);
1863 if (nhs->hstate_kobjs[idx]) {
1864 kobject_put(nhs->hstate_kobjs[idx]);
1865 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001866 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001867 }
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001868
1869 kobject_put(nhs->hugepages_kobj);
1870 nhs->hugepages_kobj = NULL;
1871}
1872
1873/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001874 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001875 * that have them.
1876 */
1877static void hugetlb_unregister_all_nodes(void)
1878{
1879 int nid;
1880
1881 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001882 * disable node device registrations.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001883 */
1884 register_hugetlbfs_with_node(NULL, NULL);
1885
1886 /*
1887 * remove hstate attributes from any nodes that have them.
1888 */
1889 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08001890 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001891}
1892
1893/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001894 * Register hstate attributes for a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001895 * No-op if attributes already registered.
1896 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02001897static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001898{
1899 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001900 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001901 int err;
1902
1903 if (nhs->hugepages_kobj)
1904 return; /* already allocated */
1905
1906 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08001907 &node->dev.kobj);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001908 if (!nhs->hugepages_kobj)
1909 return;
1910
1911 for_each_hstate(h) {
1912 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1913 nhs->hstate_kobjs,
1914 &per_node_hstate_attr_group);
1915 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08001916 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1917 h->name, node->dev.id);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001918 hugetlb_unregister_node(node);
1919 break;
1920 }
1921 }
1922}
1923
1924/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001925 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08001926 * devices of nodes that have memory. All on-line nodes should have
1927 * registered their associated device by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001928 */
1929static void hugetlb_register_all_nodes(void)
1930{
1931 int nid;
1932
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001933 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08001934 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08001935 if (node->dev.id == nid)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001936 hugetlb_register_node(node);
1937 }
1938
1939 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001940 * Let the node device driver know we're here so it can
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001941 * [un]register hstate attributes on node hotplug.
1942 */
1943 register_hugetlbfs_with_node(hugetlb_register_node,
1944 hugetlb_unregister_node);
1945}
1946#else /* !CONFIG_NUMA */
1947
1948static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1949{
1950 BUG();
1951 if (nidp)
1952 *nidp = -1;
1953 return NULL;
1954}
1955
1956static void hugetlb_unregister_all_nodes(void) { }
1957
1958static void hugetlb_register_all_nodes(void) { }
1959
1960#endif
1961
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001962static void __exit hugetlb_exit(void)
1963{
1964 struct hstate *h;
1965
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001966 hugetlb_unregister_all_nodes();
1967
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001968 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001969 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001970 }
1971
1972 kobject_put(hugepages_kobj);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07001973 kfree(htlb_fault_mutex_table);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001974}
1975module_exit(hugetlb_exit);
1976
1977static int __init hugetlb_init(void)
1978{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07001979 int i;
1980
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001981 /* Some platform decide whether they support huge pages at boot
1982 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1983 * there is no such support
1984 */
1985 if (HPAGE_SHIFT == 0)
1986 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001987
Nick Piggine11bfbf2008-07-23 21:27:52 -07001988 if (!size_to_hstate(default_hstate_size)) {
1989 default_hstate_size = HPAGE_SIZE;
1990 if (!size_to_hstate(default_hstate_size))
1991 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001992 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001993 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07001994 if (default_hstate_max_huge_pages)
1995 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001996
1997 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07001998 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001999 report_hugepages();
2000
2001 hugetlb_sysfs_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002002 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002003 hugetlb_cgroup_file_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002004
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002005#ifdef CONFIG_SMP
2006 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2007#else
2008 num_fault_mutexes = 1;
2009#endif
2010 htlb_fault_mutex_table =
2011 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2012 BUG_ON(!htlb_fault_mutex_table);
2013
2014 for (i = 0; i < num_fault_mutexes; i++)
2015 mutex_init(&htlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002016 return 0;
2017}
2018module_init(hugetlb_init);
2019
2020/* Should be called on processing a hugepagesz=... option */
2021void __init hugetlb_add_hstate(unsigned order)
2022{
2023 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002024 unsigned long i;
2025
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002026 if (size_to_hstate(PAGE_SIZE << order)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002027 pr_warning("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002028 return;
2029 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002030 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002031 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002032 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002033 h->order = order;
2034 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002035 h->nr_huge_pages = 0;
2036 h->free_huge_pages = 0;
2037 for (i = 0; i < MAX_NUMNODES; ++i)
2038 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002039 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002040 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2041 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002042 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2043 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002044
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002045 parsed_hstate = h;
2046}
2047
Nick Piggine11bfbf2008-07-23 21:27:52 -07002048static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002049{
2050 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002051 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002052
2053 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002054 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002055 * so this hugepages= parameter goes to the "default hstate".
2056 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002057 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002058 mhp = &default_hstate_max_huge_pages;
2059 else
2060 mhp = &parsed_hstate->max_huge_pages;
2061
Andi Kleen8faa8b02008-07-23 21:27:48 -07002062 if (mhp == last_mhp) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002063 pr_warning("hugepages= specified twice without "
2064 "interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002065 return 1;
2066 }
2067
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002068 if (sscanf(s, "%lu", mhp) <= 0)
2069 *mhp = 0;
2070
Andi Kleen8faa8b02008-07-23 21:27:48 -07002071 /*
2072 * Global state is always initialized later in hugetlb_init.
2073 * But we need to allocate >= MAX_ORDER hstates here early to still
2074 * use the bootmem allocator.
2075 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002076 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002077 hugetlb_hstate_alloc_pages(parsed_hstate);
2078
2079 last_mhp = mhp;
2080
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002081 return 1;
2082}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002083__setup("hugepages=", hugetlb_nrpages_setup);
2084
2085static int __init hugetlb_default_setup(char *s)
2086{
2087 default_hstate_size = memparse(s, &s);
2088 return 1;
2089}
2090__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002091
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002092static unsigned int cpuset_mems_nr(unsigned int *array)
2093{
2094 int node;
2095 unsigned int nr = 0;
2096
2097 for_each_node_mask(node, cpuset_current_mems_allowed)
2098 nr += array[node];
2099
2100 return nr;
2101}
2102
2103#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002104static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2105 struct ctl_table *table, int write,
2106 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107{
Andi Kleene5ff2152008-07-23 21:27:42 -07002108 struct hstate *h = &default_hstate;
2109 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002110 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002111
Petr Holasekc033a932011-03-22 16:33:05 -07002112 tmp = h->max_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002113
Eric B Munsonadbe8722011-01-13 15:47:27 -08002114 if (write && h->order >= MAX_ORDER)
2115 return -EINVAL;
2116
Andi Kleene5ff2152008-07-23 21:27:42 -07002117 table->data = &tmp;
2118 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002119 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2120 if (ret)
2121 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002122
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002123 if (write) {
David Rientjesbad44b52009-12-14 17:58:38 -08002124 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2125 GFP_KERNEL | __GFP_NORETRY);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002126 if (!(obey_mempolicy &&
2127 init_nodemask_of_mempolicy(nodes_allowed))) {
2128 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002129 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002130 }
2131 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2132
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002133 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002134 NODEMASK_FREE(nodes_allowed);
2135 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002136out:
2137 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138}
Mel Gorman396faf02007-07-17 04:03:13 -07002139
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002140int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2141 void __user *buffer, size_t *length, loff_t *ppos)
2142{
2143
2144 return hugetlb_sysctl_handler_common(false, table, write,
2145 buffer, length, ppos);
2146}
2147
2148#ifdef CONFIG_NUMA
2149int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2150 void __user *buffer, size_t *length, loff_t *ppos)
2151{
2152 return hugetlb_sysctl_handler_common(true, table, write,
2153 buffer, length, ppos);
2154}
2155#endif /* CONFIG_NUMA */
2156
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002157int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002158 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002159 size_t *length, loff_t *ppos)
2160{
Andi Kleena5516432008-07-23 21:27:41 -07002161 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002162 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002163 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002164
Petr Holasekc033a932011-03-22 16:33:05 -07002165 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002166
Eric B Munsonadbe8722011-01-13 15:47:27 -08002167 if (write && h->order >= MAX_ORDER)
2168 return -EINVAL;
2169
Andi Kleene5ff2152008-07-23 21:27:42 -07002170 table->data = &tmp;
2171 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002172 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2173 if (ret)
2174 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002175
2176 if (write) {
2177 spin_lock(&hugetlb_lock);
2178 h->nr_overcommit_huge_pages = tmp;
2179 spin_unlock(&hugetlb_lock);
2180 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002181out:
2182 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002183}
2184
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185#endif /* CONFIG_SYSCTL */
2186
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002187void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188{
Andi Kleena5516432008-07-23 21:27:41 -07002189 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002190 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002191 "HugePages_Total: %5lu\n"
2192 "HugePages_Free: %5lu\n"
2193 "HugePages_Rsvd: %5lu\n"
2194 "HugePages_Surp: %5lu\n"
2195 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002196 h->nr_huge_pages,
2197 h->free_huge_pages,
2198 h->resv_huge_pages,
2199 h->surplus_huge_pages,
2200 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201}
2202
2203int hugetlb_report_node_meminfo(int nid, char *buf)
2204{
Andi Kleena5516432008-07-23 21:27:41 -07002205 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 return sprintf(buf,
2207 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002208 "Node %d HugePages_Free: %5u\n"
2209 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002210 nid, h->nr_huge_pages_node[nid],
2211 nid, h->free_huge_pages_node[nid],
2212 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213}
2214
David Rientjes949f7ec2013-04-29 15:07:48 -07002215void hugetlb_show_meminfo(void)
2216{
2217 struct hstate *h;
2218 int nid;
2219
2220 for_each_node_state(nid, N_MEMORY)
2221 for_each_hstate(h)
2222 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2223 nid,
2224 h->nr_huge_pages_node[nid],
2225 h->free_huge_pages_node[nid],
2226 h->surplus_huge_pages_node[nid],
2227 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2228}
2229
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2231unsigned long hugetlb_total_pages(void)
2232{
Wanpeng Lid0028582013-03-22 15:04:40 -07002233 struct hstate *h;
2234 unsigned long nr_total_pages = 0;
2235
2236 for_each_hstate(h)
2237 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2238 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240
Andi Kleena5516432008-07-23 21:27:41 -07002241static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002242{
2243 int ret = -ENOMEM;
2244
2245 spin_lock(&hugetlb_lock);
2246 /*
2247 * When cpuset is configured, it breaks the strict hugetlb page
2248 * reservation as the accounting is done on a global variable. Such
2249 * reservation is completely rubbish in the presence of cpuset because
2250 * the reservation is not checked against page availability for the
2251 * current cpuset. Application can still potentially OOM'ed by kernel
2252 * with lack of free htlb page in cpuset that the task is in.
2253 * Attempt to enforce strict accounting with cpuset is almost
2254 * impossible (or too ugly) because cpuset is too fluid that
2255 * task or memory node can be dynamically moved between cpusets.
2256 *
2257 * The change of semantics for shared hugetlb mapping with cpuset is
2258 * undesirable. However, in order to preserve some of the semantics,
2259 * we fall back to check against current free page availability as
2260 * a best attempt and hopefully to minimize the impact of changing
2261 * semantics that cpuset has.
2262 */
2263 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002264 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002265 goto out;
2266
Andi Kleena5516432008-07-23 21:27:41 -07002267 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2268 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002269 goto out;
2270 }
2271 }
2272
2273 ret = 0;
2274 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002275 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002276
2277out:
2278 spin_unlock(&hugetlb_lock);
2279 return ret;
2280}
2281
Andy Whitcroft84afd992008-07-23 21:27:32 -07002282static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2283{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002284 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002285
2286 /*
2287 * This new VMA should share its siblings reservation map if present.
2288 * The VMA will only ever have a valid reservation map pointer where
2289 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002290 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002291 * after this open call completes. It is therefore safe to take a
2292 * new reference here without additional locking.
2293 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002294 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002295 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002296}
2297
Mel Gormana1e78772008-07-23 21:27:23 -07002298static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2299{
Andi Kleena5516432008-07-23 21:27:41 -07002300 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002301 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002302 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002303 unsigned long reserve, start, end;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002304
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002305 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2306 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002307
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002308 start = vma_hugecache_offset(h, vma, vma->vm_start);
2309 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002310
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002311 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002312
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002313 kref_put(&resv->refs, resv_map_release);
2314
2315 if (reserve) {
2316 hugetlb_acct_memory(h, -reserve);
2317 hugepage_subpool_put_pages(spool, reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002318 }
Mel Gormana1e78772008-07-23 21:27:23 -07002319}
2320
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321/*
2322 * We cannot handle pagefaults against hugetlb pages at all. They cause
2323 * handle_mm_fault() to try to instantiate regular-sized pages in the
2324 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2325 * this far.
2326 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002327static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328{
2329 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002330 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331}
2332
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002333const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002334 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002335 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002336 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337};
2338
David Gibson1e8f8892006-01-06 00:10:44 -08002339static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2340 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002341{
2342 pte_t entry;
2343
David Gibson1e8f8892006-01-06 00:10:44 -08002344 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002345 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2346 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07002347 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002348 entry = huge_pte_wrprotect(mk_huge_pte(page,
2349 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002350 }
2351 entry = pte_mkyoung(entry);
2352 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04002353 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07002354
2355 return entry;
2356}
2357
David Gibson1e8f8892006-01-06 00:10:44 -08002358static void set_huge_ptep_writable(struct vm_area_struct *vma,
2359 unsigned long address, pte_t *ptep)
2360{
2361 pte_t entry;
2362
Gerald Schaefer106c9922013-04-29 15:07:23 -07002363 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002364 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002365 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002366}
2367
2368
David Gibson63551ae2005-06-21 17:14:44 -07002369int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2370 struct vm_area_struct *vma)
2371{
2372 pte_t *src_pte, *dst_pte, entry;
2373 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002374 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002375 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002376 struct hstate *h = hstate_vma(vma);
2377 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002378 unsigned long mmun_start; /* For mmu_notifiers */
2379 unsigned long mmun_end; /* For mmu_notifiers */
2380 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002381
2382 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002383
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002384 mmun_start = vma->vm_start;
2385 mmun_end = vma->vm_end;
2386 if (cow)
2387 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2388
Andi Kleena5516432008-07-23 21:27:41 -07002389 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002390 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002391 src_pte = huge_pte_offset(src, addr);
2392 if (!src_pte)
2393 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002394 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002395 if (!dst_pte) {
2396 ret = -ENOMEM;
2397 break;
2398 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08002399
2400 /* If the pagetables are shared don't copy or take references */
2401 if (dst_pte == src_pte)
2402 continue;
2403
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002404 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2405 src_ptl = huge_pte_lockptr(h, src, src_pte);
2406 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002407 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002408 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002409 huge_ptep_set_wrprotect(src, addr, src_pte);
2410 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002411 ptepage = pte_page(entry);
2412 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002413 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002414 set_huge_pte_at(dst, addr, dst_pte, entry);
2415 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002416 spin_unlock(src_ptl);
2417 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002418 }
David Gibson63551ae2005-06-21 17:14:44 -07002419
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002420 if (cow)
2421 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2422
2423 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07002424}
2425
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002426static int is_hugetlb_entry_migration(pte_t pte)
2427{
2428 swp_entry_t swp;
2429
2430 if (huge_pte_none(pte) || pte_present(pte))
2431 return 0;
2432 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002433 if (non_swap_entry(swp) && is_migration_entry(swp))
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002434 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002435 else
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002436 return 0;
2437}
2438
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002439static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2440{
2441 swp_entry_t swp;
2442
2443 if (huge_pte_none(pte) || pte_present(pte))
2444 return 0;
2445 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002446 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002447 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002448 else
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002449 return 0;
2450}
2451
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002452void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2453 unsigned long start, unsigned long end,
2454 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002455{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002456 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07002457 struct mm_struct *mm = vma->vm_mm;
2458 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002459 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002460 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002461 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07002462 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002463 struct hstate *h = hstate_vma(vma);
2464 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002465 const unsigned long mmun_start = start; /* For mmu_notifiers */
2466 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07002467
David Gibson63551ae2005-06-21 17:14:44 -07002468 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002469 BUG_ON(start & ~huge_page_mask(h));
2470 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002471
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002472 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002473 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002474again:
Andi Kleena5516432008-07-23 21:27:41 -07002475 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002476 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002477 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002478 continue;
2479
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002480 ptl = huge_pte_lock(h, mm, ptep);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002481 if (huge_pmd_unshare(mm, &address, ptep))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002482 goto unlock;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002483
Hillf Danton66293262012-03-23 15:01:48 -07002484 pte = huge_ptep_get(ptep);
2485 if (huge_pte_none(pte))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002486 goto unlock;
Hillf Danton66293262012-03-23 15:01:48 -07002487
2488 /*
2489 * HWPoisoned hugepage is already unmapped and dropped reference
2490 */
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002491 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002492 huge_pte_clear(mm, address, ptep);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002493 goto unlock;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002494 }
Hillf Danton66293262012-03-23 15:01:48 -07002495
2496 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002497 /*
2498 * If a reference page is supplied, it is because a specific
2499 * page is being unmapped, not a range. Ensure the page we
2500 * are about to unmap is the actual page of interest.
2501 */
2502 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002503 if (page != ref_page)
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002504 goto unlock;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002505
2506 /*
2507 * Mark the VMA as having unmapped its page so that
2508 * future faults in this VMA will fail rather than
2509 * looking like data was lost
2510 */
2511 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2512 }
2513
David Gibsonc7546f82005-08-05 11:59:35 -07002514 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002515 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07002516 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08002517 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07002518
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002519 page_remove_rmap(page);
2520 force_flush = !__tlb_remove_page(tlb, page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002521 if (force_flush) {
2522 spin_unlock(ptl);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002523 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002524 }
Hillf Danton9e811302012-03-21 16:34:03 -07002525 /* Bail out after unmapping reference page if supplied */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002526 if (ref_page) {
2527 spin_unlock(ptl);
Hillf Danton9e811302012-03-21 16:34:03 -07002528 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002529 }
2530unlock:
2531 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002532 }
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002533 /*
2534 * mmu_gather ran out of room to batch pages, we break out of
2535 * the PTE lock to avoid doing the potential expensive TLB invalidate
2536 * and page-free while holding it.
2537 */
2538 if (force_flush) {
2539 force_flush = 0;
2540 tlb_flush_mmu(tlb);
2541 if (address < end && !ref_page)
2542 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002543 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002544 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002545 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546}
David Gibson63551ae2005-06-21 17:14:44 -07002547
Mel Gormand8333522012-07-31 16:46:20 -07002548void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2549 struct vm_area_struct *vma, unsigned long start,
2550 unsigned long end, struct page *ref_page)
2551{
2552 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2553
2554 /*
2555 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2556 * test will fail on a vma being torn down, and not grab a page table
2557 * on its way out. We're lucky that the flag has such an appropriate
2558 * name, and can in fact be safely cleared here. We could clear it
2559 * before the __unmap_hugepage_range above, but all that's necessary
2560 * is to clear it before releasing the i_mmap_mutex. This works
2561 * because in the context this is called, the VMA is about to be
2562 * destroyed and the i_mmap_mutex is held.
2563 */
2564 vma->vm_flags &= ~VM_MAYSHARE;
2565}
2566
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002567void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002568 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002569{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002570 struct mm_struct *mm;
2571 struct mmu_gather tlb;
2572
2573 mm = vma->vm_mm;
2574
Linus Torvalds2b047252013-08-15 11:42:25 -07002575 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002576 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2577 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002578}
2579
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002580/*
2581 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2582 * mappping it owns the reserve page for. The intention is to unmap the page
2583 * from other VMAs and let the children be SIGKILLed if they are faulting the
2584 * same region.
2585 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07002586static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2587 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002588{
Adam Litke75266742008-11-12 13:24:56 -08002589 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002590 struct vm_area_struct *iter_vma;
2591 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002592 pgoff_t pgoff;
2593
2594 /*
2595 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2596 * from page cache lookup which is in HPAGE_SIZE units.
2597 */
Adam Litke75266742008-11-12 13:24:56 -08002598 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07002599 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2600 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05002601 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002602
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002603 /*
2604 * Take the mapping lock for the duration of the table walk. As
2605 * this mapping should be shared between all the VMAs,
2606 * __unmap_hugepage_range() is called as the lock is already held
2607 */
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002608 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002609 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002610 /* Do not unmap the current VMA */
2611 if (iter_vma == vma)
2612 continue;
2613
2614 /*
2615 * Unmap the page from other VMAs without their own reserves.
2616 * They get marked to be SIGKILLed if they fault in these
2617 * areas. This is because a future no-page fault on this VMA
2618 * could insert a zeroed page instead of the data existing
2619 * from the time of fork. This would look like data corruption
2620 */
2621 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002622 unmap_hugepage_range(iter_vma, address,
2623 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002624 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002625 mutex_unlock(&mapping->i_mmap_mutex);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002626
2627 return 1;
2628}
2629
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002630/*
2631 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08002632 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2633 * cannot race with other handlers or page migration.
2634 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002635 */
David Gibson1e8f8892006-01-06 00:10:44 -08002636static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002637 unsigned long address, pte_t *ptep, pte_t pte,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002638 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08002639{
Andi Kleena5516432008-07-23 21:27:41 -07002640 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002641 struct page *old_page, *new_page;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002642 int outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002643 unsigned long mmun_start; /* For mmu_notifiers */
2644 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08002645
2646 old_page = pte_page(pte);
2647
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002648retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002649 /* If no-one else is actually using this page, avoid the copy
2650 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07002651 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2652 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002653 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002654 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002655 }
2656
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002657 /*
2658 * If the process that created a MAP_PRIVATE mapping is about to
2659 * perform a COW due to a shared page count, attempt to satisfy
2660 * the allocation without using the existing reserves. The pagecache
2661 * page is used to determine if the reserve at this address was
2662 * consumed or not. If reserves were used, a partial faulted mapping
2663 * at the time of fork() could consume its reserves on COW instead
2664 * of the full address range.
2665 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07002666 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002667 old_page != pagecache_page)
2668 outside_reserve = 1;
2669
David Gibson1e8f8892006-01-06 00:10:44 -08002670 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002671
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002672 /* Drop page table lock as buddy allocator may be called */
2673 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002674 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002675
Adam Litke2fc39ce2007-11-14 16:59:39 -08002676 if (IS_ERR(new_page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002677 long err = PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002678 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002679
2680 /*
2681 * If a process owning a MAP_PRIVATE mapping fails to COW,
2682 * it is due to references held by a child and an insufficient
2683 * huge page pool. To guarantee the original mappers
2684 * reliability, unmap the page from child processes. The child
2685 * may get SIGKILLed if it later faults.
2686 */
2687 if (outside_reserve) {
2688 BUG_ON(huge_pte_none(pte));
2689 if (unmap_ref_private(mm, vma, old_page, address)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002690 BUG_ON(huge_pte_none(pte));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002691 spin_lock(ptl);
Hillf Dantona734bcc2012-01-10 15:07:20 -08002692 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2693 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2694 goto retry_avoidcopy;
2695 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002696 * race occurs while re-acquiring page table
2697 * lock, and our job is done.
Hillf Dantona734bcc2012-01-10 15:07:20 -08002698 */
2699 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002700 }
2701 WARN_ON_ONCE(1);
2702 }
2703
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002704 /* Caller expects lock to be held */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002705 spin_lock(ptl);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002706 if (err == -ENOMEM)
2707 return VM_FAULT_OOM;
2708 else
2709 return VM_FAULT_SIGBUS;
David Gibson1e8f8892006-01-06 00:10:44 -08002710 }
2711
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002712 /*
2713 * When the original hugepage is shared one, it does not have
2714 * anon_vma prepared.
2715 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002716 if (unlikely(anon_vma_prepare(vma))) {
Hillf Dantonea4039a2011-11-15 14:36:12 -08002717 page_cache_release(new_page);
2718 page_cache_release(old_page);
Dean Nelson44e2aa92010-10-26 14:22:08 -07002719 /* Caller expects lock to be held */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002720 spin_lock(ptl);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002721 return VM_FAULT_OOM;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002722 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002723
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002724 copy_user_huge_page(new_page, old_page, address, vma,
2725 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002726 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002727
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002728 mmun_start = address & huge_page_mask(h);
2729 mmun_end = mmun_start + huge_page_size(h);
2730 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002731 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002732 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002733 * before the page tables are altered
2734 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002735 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07002736 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002737 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07002738 ClearPagePrivate(new_page);
2739
David Gibson1e8f8892006-01-06 00:10:44 -08002740 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002741 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002742 set_huge_pte_at(mm, address, ptep,
2743 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002744 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002745 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002746 /* Make the old page be freed below */
2747 new_page = old_page;
2748 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002749 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002750 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08002751 page_cache_release(new_page);
2752 page_cache_release(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07002753
2754 /* Caller expects lock to be held */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002755 spin_lock(ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002756 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002757}
2758
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002759/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002760static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2761 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002762{
2763 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002764 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002765
2766 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002767 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002768
2769 return find_lock_page(mapping, idx);
2770}
2771
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002772/*
2773 * Return whether there is a pagecache page to back given address within VMA.
2774 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2775 */
2776static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002777 struct vm_area_struct *vma, unsigned long address)
2778{
2779 struct address_space *mapping;
2780 pgoff_t idx;
2781 struct page *page;
2782
2783 mapping = vma->vm_file->f_mapping;
2784 idx = vma_hugecache_offset(h, vma, address);
2785
2786 page = find_get_page(mapping, idx);
2787 if (page)
2788 put_page(page);
2789 return page != NULL;
2790}
2791
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002792static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002793 struct address_space *mapping, pgoff_t idx,
2794 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002795{
Andi Kleena5516432008-07-23 21:27:41 -07002796 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002797 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08002798 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002799 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002800 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08002801 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002802 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07002803
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002804 /*
2805 * Currently, we are forced to kill the process in the event the
2806 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002807 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002808 */
2809 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002810 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2811 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002812 return ret;
2813 }
2814
Adam Litke4c887262005-10-29 18:16:46 -07002815 /*
2816 * Use page lock to guard against racing truncation
2817 * before we get page_table_lock.
2818 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002819retry:
2820 page = find_lock_page(mapping, idx);
2821 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002822 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002823 if (idx >= size)
2824 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002825 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002826 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002827 ret = PTR_ERR(page);
2828 if (ret == -ENOMEM)
2829 ret = VM_FAULT_OOM;
2830 else
2831 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002832 goto out;
2833 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002834 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002835 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002836
Mel Gormanf83a2752009-05-28 14:34:40 -07002837 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002838 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002839 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002840
2841 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2842 if (err) {
2843 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002844 if (err == -EEXIST)
2845 goto retry;
2846 goto out;
2847 }
Joonsoo Kim07443a82013-09-11 14:21:58 -07002848 ClearPagePrivate(page);
Ken Chen45c682a2007-11-14 16:59:44 -08002849
2850 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002851 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002852 spin_unlock(&inode->i_lock);
Mel Gorman23be7462010-04-23 13:17:56 -04002853 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002854 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002855 if (unlikely(anon_vma_prepare(vma))) {
2856 ret = VM_FAULT_OOM;
2857 goto backout_unlocked;
2858 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08002859 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04002860 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002861 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002862 /*
2863 * If memory error occurs between mmap() and fault, some process
2864 * don't have hwpoisoned swap entry for errored virtual address.
2865 * So we need to block hugepage fault by PG_hwpoison bit check.
2866 */
2867 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07002868 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002869 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002870 goto backout_unlocked;
2871 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08002872 }
David Gibson1e8f8892006-01-06 00:10:44 -08002873
Andy Whitcroft57303d82008-08-12 15:08:47 -07002874 /*
2875 * If we are going to COW a private mapping later, we examine the
2876 * pending reservations for this page now. This will ensure that
2877 * any allocations necessary to record that reservation occur outside
2878 * the spinlock.
2879 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002880 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002881 if (vma_needs_reservation(h, vma, address) < 0) {
2882 ret = VM_FAULT_OOM;
2883 goto backout_unlocked;
2884 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002885
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002886 ptl = huge_pte_lockptr(h, mm, ptep);
2887 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07002888 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002889 if (idx >= size)
2890 goto backout;
2891
Nick Piggin83c54072007-07-19 01:47:05 -07002892 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002893 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002894 goto backout;
2895
Joonsoo Kim07443a82013-09-11 14:21:58 -07002896 if (anon_rmap) {
2897 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08002898 hugepage_add_new_anon_rmap(page, vma, address);
Joonsoo Kim07443a82013-09-11 14:21:58 -07002899 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08002900 else
2901 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08002902 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2903 && (vma->vm_flags & VM_SHARED)));
2904 set_huge_pte_at(mm, address, ptep, new_pte);
2905
Hugh Dickins788c7df2009-06-23 13:49:05 +01002906 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002907 /* Optimization, do the COW without a second fault */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002908 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08002909 }
2910
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002911 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07002912 unlock_page(page);
2913out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002914 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002915
2916backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002917 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002918backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002919 unlock_page(page);
2920 put_page(page);
2921 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002922}
2923
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002924#ifdef CONFIG_SMP
2925static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
2926 struct vm_area_struct *vma,
2927 struct address_space *mapping,
2928 pgoff_t idx, unsigned long address)
2929{
2930 unsigned long key[2];
2931 u32 hash;
2932
2933 if (vma->vm_flags & VM_SHARED) {
2934 key[0] = (unsigned long) mapping;
2935 key[1] = idx;
2936 } else {
2937 key[0] = (unsigned long) mm;
2938 key[1] = address >> huge_page_shift(h);
2939 }
2940
2941 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
2942
2943 return hash & (num_fault_mutexes - 1);
2944}
2945#else
2946/*
2947 * For uniprocesor systems we always use a single mutex, so just
2948 * return 0 and avoid the hashing overhead.
2949 */
2950static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
2951 struct vm_area_struct *vma,
2952 struct address_space *mapping,
2953 pgoff_t idx, unsigned long address)
2954{
2955 return 0;
2956}
2957#endif
2958
Adam Litke86e52162006-01-06 00:10:43 -08002959int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002960 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08002961{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002962 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002963 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08002964 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002965 u32 hash;
2966 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002967 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002968 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07002969 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002970 struct address_space *mapping;
Adam Litke86e52162006-01-06 00:10:43 -08002971
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08002972 address &= huge_page_mask(h);
2973
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002974 ptep = huge_pte_offset(mm, address);
2975 if (ptep) {
2976 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002977 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002978 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002979 return 0;
2980 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07002981 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002982 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002983 }
2984
Andi Kleena5516432008-07-23 21:27:41 -07002985 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002986 if (!ptep)
2987 return VM_FAULT_OOM;
2988
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002989 mapping = vma->vm_file->f_mapping;
2990 idx = vma_hugecache_offset(h, vma, address);
2991
David Gibson3935baa2006-03-22 00:08:53 -08002992 /*
2993 * Serialize hugepage allocation and instantiation, so that we don't
2994 * get spurious allocation failures if two CPUs race to instantiate
2995 * the same page in the page cache.
2996 */
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002997 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
2998 mutex_lock(&htlb_fault_mutex_table[hash]);
2999
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003000 entry = huge_ptep_get(ptep);
3001 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003002 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003003 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003004 }
Adam Litke86e52162006-01-06 00:10:43 -08003005
Nick Piggin83c54072007-07-19 01:47:05 -07003006 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003007
Andy Whitcroft57303d82008-08-12 15:08:47 -07003008 /*
3009 * If we are going to COW the mapping later, we examine the pending
3010 * reservations for this page now. This will ensure that any
3011 * allocations necessary to record that reservation occur outside the
3012 * spinlock. For private mappings, we also lookup the pagecache
3013 * page now as it is used to determine if a reservation has been
3014 * consumed.
3015 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003016 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003017 if (vma_needs_reservation(h, vma, address) < 0) {
3018 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003019 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003020 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003021
Mel Gormanf83a2752009-05-28 14:34:40 -07003022 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003023 pagecache_page = hugetlbfs_pagecache_page(h,
3024 vma, address);
3025 }
3026
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09003027 /*
3028 * hugetlb_cow() requires page locks of pte_page(entry) and
3029 * pagecache_page, so here we need take the former one
3030 * when page != pagecache_page or !pagecache_page.
3031 * Note that locking order is always pagecache_page -> page,
3032 * so no worry about deadlock.
3033 */
3034 page = pte_page(entry);
Chris Metcalf66aebce2012-04-12 12:49:15 -07003035 get_page(page);
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09003036 if (page != pagecache_page)
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003037 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003038
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003039 ptl = huge_pte_lockptr(h, mm, ptep);
3040 spin_lock(ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003041 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003042 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003043 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003044
3045
Hugh Dickins788c7df2009-06-23 13:49:05 +01003046 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003047 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07003048 ret = hugetlb_cow(mm, vma, address, ptep, entry,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003049 pagecache_page, ptl);
3050 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003051 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003052 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003053 }
3054 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003055 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3056 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003057 update_mmu_cache(vma, address, ptep);
David Gibsonb4d1d992008-10-15 22:01:11 -07003058
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003059out_ptl:
3060 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003061
3062 if (pagecache_page) {
3063 unlock_page(pagecache_page);
3064 put_page(pagecache_page);
3065 }
Dean Nelson1f64d692010-12-02 14:31:12 -08003066 if (page != pagecache_page)
3067 unlock_page(page);
Chris Metcalf66aebce2012-04-12 12:49:15 -07003068 put_page(page);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003069
David Gibsonb4d1d992008-10-15 22:01:11 -07003070out_mutex:
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003071 mutex_unlock(&htlb_fault_mutex_table[hash]);
David Gibson1e8f8892006-01-06 00:10:44 -08003072 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003073}
3074
Michel Lespinasse28a35712013-02-22 16:35:55 -08003075long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3076 struct page **pages, struct vm_area_struct **vmas,
3077 unsigned long *position, unsigned long *nr_pages,
3078 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003079{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003080 unsigned long pfn_offset;
3081 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003082 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003083 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003084
David Gibson63551ae2005-06-21 17:14:44 -07003085 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003086 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003087 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003088 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003089 struct page *page;
3090
3091 /*
3092 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003093 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07003094 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003095 *
3096 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07003097 */
Andi Kleena5516432008-07-23 21:27:41 -07003098 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003099 if (pte)
3100 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003101 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07003102
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003103 /*
3104 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003105 * an error where there's an empty slot with no huge pagecache
3106 * to back it. This way, we avoid allocating a hugepage, and
3107 * the sparse dumpfile avoids allocating disk blocks, but its
3108 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003109 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003110 if (absent && (flags & FOLL_DUMP) &&
3111 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003112 if (pte)
3113 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003114 remainder = 0;
3115 break;
3116 }
3117
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07003118 /*
3119 * We need call hugetlb_fault for both hugepages under migration
3120 * (in which case hugetlb_fault waits for the migration,) and
3121 * hwpoisoned hugepages (in which case we need to prevent the
3122 * caller from accessing to them.) In order to do this, we use
3123 * here is_swap_pte instead of is_hugetlb_entry_migration and
3124 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3125 * both cases, and because we can't follow correct pages
3126 * directly from any kind of swap entries.
3127 */
3128 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07003129 ((flags & FOLL_WRITE) &&
3130 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003131 int ret;
3132
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003133 if (pte)
3134 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003135 ret = hugetlb_fault(mm, vma, vaddr,
3136 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07003137 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003138 continue;
3139
3140 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003141 break;
3142 }
David Gibson63551ae2005-06-21 17:14:44 -07003143
Andi Kleena5516432008-07-23 21:27:41 -07003144 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003145 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003146same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003147 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003148 pages[i] = mem_map_offset(page, pfn_offset);
Andrea Arcangelia0368d42014-01-21 15:48:49 -08003149 get_page_foll(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003150 }
David Gibson63551ae2005-06-21 17:14:44 -07003151
3152 if (vmas)
3153 vmas[i] = vma;
3154
3155 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003156 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003157 --remainder;
3158 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003159 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003160 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003161 /*
3162 * We use pfn_offset to avoid touching the pageframes
3163 * of this compound page.
3164 */
3165 goto same_page;
3166 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003167 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003168 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08003169 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003170 *position = vaddr;
3171
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003172 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003173}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003174
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003175unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003176 unsigned long address, unsigned long end, pgprot_t newprot)
3177{
3178 struct mm_struct *mm = vma->vm_mm;
3179 unsigned long start = address;
3180 pte_t *ptep;
3181 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003182 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003183 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003184
3185 BUG_ON(address >= end);
3186 flush_cache_range(vma, address, end);
3187
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07003188 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07003189 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003190 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003191 ptep = huge_pte_offset(mm, address);
3192 if (!ptep)
3193 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003194 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003195 if (huge_pmd_unshare(mm, &address, ptep)) {
3196 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003197 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003198 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003199 }
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003200 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003201 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003202 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003203 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003204 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003205 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003206 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003207 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003208 }
Mel Gormand8333522012-07-31 16:46:20 -07003209 /*
3210 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3211 * may have cleared our pud entry and done put_page on the page table:
3212 * once we release i_mmap_mutex, another task can do the final put_page
3213 * and that page table be reused and filled with junk.
3214 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003215 flush_tlb_range(vma, start, end);
Mel Gormand8333522012-07-31 16:46:20 -07003216 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003217
3218 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003219}
3220
Mel Gormana1e78772008-07-23 21:27:23 -07003221int hugetlb_reserve_pages(struct inode *inode,
3222 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003223 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003224 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003225{
Mel Gorman17c9d122009-02-11 16:34:16 +00003226 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003227 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003228 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003229 struct resv_map *resv_map;
Adam Litkee4e574b2007-10-16 01:26:19 -07003230
Mel Gormana1e78772008-07-23 21:27:23 -07003231 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003232 * Only apply hugepage reservation if asked. At fault time, an
3233 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003234 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003235 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003236 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003237 return 0;
3238
3239 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003240 * Shared mappings base their reservation on the number of pages that
3241 * are already allocated on behalf of the file. Private mappings need
3242 * to reserve the full area even if read-only as mprotect() may be
3243 * called to make the mapping read-write. Assume !vma is a shm mapping
3244 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07003245 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003246 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003247
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003248 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003249
3250 } else {
3251 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003252 if (!resv_map)
3253 return -ENOMEM;
3254
Mel Gorman17c9d122009-02-11 16:34:16 +00003255 chg = to - from;
3256
Mel Gorman5a6fe122009-02-10 14:02:27 +00003257 set_vma_resv_map(vma, resv_map);
3258 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3259 }
3260
Dave Hansenc50ac052012-05-29 15:06:46 -07003261 if (chg < 0) {
3262 ret = chg;
3263 goto out_err;
3264 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003265
David Gibson90481622012-03-21 16:34:12 -07003266 /* There must be enough pages in the subpool for the mapping */
Dave Hansenc50ac052012-05-29 15:06:46 -07003267 if (hugepage_subpool_get_pages(spool, chg)) {
3268 ret = -ENOSPC;
3269 goto out_err;
3270 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003271
3272 /*
3273 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07003274 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00003275 */
3276 ret = hugetlb_acct_memory(h, chg);
3277 if (ret < 0) {
David Gibson90481622012-03-21 16:34:12 -07003278 hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07003279 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00003280 }
3281
3282 /*
3283 * Account for the reservations made. Shared mappings record regions
3284 * that have reservations as they are shared by multiple VMAs.
3285 * When the last VMA disappears, the region map says how much
3286 * the reservation was and the page cache tells how much of
3287 * the reservation was consumed. Private mappings are per-VMA and
3288 * only the consumed reservations are tracked. When the VMA
3289 * disappears, the original reservation is the VMA size and the
3290 * consumed reservations are stored in the map. Hence, nothing
3291 * else has to be done for private mappings here
3292 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003293 if (!vma || vma->vm_flags & VM_MAYSHARE)
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003294 region_add(resv_map, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003295 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07003296out_err:
Joonsoo Kimf031dd22014-04-03 14:47:28 -07003297 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3298 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07003299 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003300}
3301
3302void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3303{
Andi Kleena5516432008-07-23 21:27:41 -07003304 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003305 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003306 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07003307 struct hugepage_subpool *spool = subpool_inode(inode);
Ken Chen45c682a2007-11-14 16:59:44 -08003308
Joonsoo Kim9119a412014-04-03 14:47:25 -07003309 if (resv_map)
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003310 chg = region_truncate(resv_map, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08003311 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07003312 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08003313 spin_unlock(&inode->i_lock);
3314
David Gibson90481622012-03-21 16:34:12 -07003315 hugepage_subpool_put_pages(spool, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07003316 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003317}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003318
Steve Capper3212b532013-04-23 12:35:02 +01003319#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3320static unsigned long page_table_shareable(struct vm_area_struct *svma,
3321 struct vm_area_struct *vma,
3322 unsigned long addr, pgoff_t idx)
3323{
3324 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3325 svma->vm_start;
3326 unsigned long sbase = saddr & PUD_MASK;
3327 unsigned long s_end = sbase + PUD_SIZE;
3328
3329 /* Allow segments to share if only one is marked locked */
3330 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3331 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3332
3333 /*
3334 * match the virtual addresses, permission and the alignment of the
3335 * page table page.
3336 */
3337 if (pmd_index(addr) != pmd_index(saddr) ||
3338 vm_flags != svm_flags ||
3339 sbase < svma->vm_start || svma->vm_end < s_end)
3340 return 0;
3341
3342 return saddr;
3343}
3344
3345static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3346{
3347 unsigned long base = addr & PUD_MASK;
3348 unsigned long end = base + PUD_SIZE;
3349
3350 /*
3351 * check on proper vm_flags and page table alignment
3352 */
3353 if (vma->vm_flags & VM_MAYSHARE &&
3354 vma->vm_start <= base && end <= vma->vm_end)
3355 return 1;
3356 return 0;
3357}
3358
3359/*
3360 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3361 * and returns the corresponding pte. While this is not necessary for the
3362 * !shared pmd case because we can allocate the pmd later as well, it makes the
3363 * code much cleaner. pmd allocation is essential for the shared case because
3364 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3365 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3366 * bad pmd for sharing.
3367 */
3368pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3369{
3370 struct vm_area_struct *vma = find_vma(mm, addr);
3371 struct address_space *mapping = vma->vm_file->f_mapping;
3372 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3373 vma->vm_pgoff;
3374 struct vm_area_struct *svma;
3375 unsigned long saddr;
3376 pte_t *spte = NULL;
3377 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003378 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01003379
3380 if (!vma_shareable(vma, addr))
3381 return (pte_t *)pmd_alloc(mm, pud, addr);
3382
3383 mutex_lock(&mapping->i_mmap_mutex);
3384 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3385 if (svma == vma)
3386 continue;
3387
3388 saddr = page_table_shareable(svma, vma, addr, idx);
3389 if (saddr) {
3390 spte = huge_pte_offset(svma->vm_mm, saddr);
3391 if (spte) {
3392 get_page(virt_to_page(spte));
3393 break;
3394 }
3395 }
3396 }
3397
3398 if (!spte)
3399 goto out;
3400
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003401 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3402 spin_lock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01003403 if (pud_none(*pud))
3404 pud_populate(mm, pud,
3405 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3406 else
3407 put_page(virt_to_page(spte));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003408 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01003409out:
3410 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3411 mutex_unlock(&mapping->i_mmap_mutex);
3412 return pte;
3413}
3414
3415/*
3416 * unmap huge page backed by shared pte.
3417 *
3418 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3419 * indicated by page_count > 1, unmap is achieved by clearing pud and
3420 * decrementing the ref count. If count == 1, the pte page is not shared.
3421 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003422 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01003423 *
3424 * returns: 1 successfully unmapped a shared pte page
3425 * 0 the underlying pte page is not shared, or it is the last user
3426 */
3427int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3428{
3429 pgd_t *pgd = pgd_offset(mm, *addr);
3430 pud_t *pud = pud_offset(pgd, *addr);
3431
3432 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3433 if (page_count(virt_to_page(ptep)) == 1)
3434 return 0;
3435
3436 pud_clear(pud);
3437 put_page(virt_to_page(ptep));
3438 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3439 return 1;
3440}
Steve Capper9e5fc742013-04-30 08:02:03 +01003441#define want_pmd_share() (1)
3442#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3443pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3444{
3445 return NULL;
3446}
3447#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01003448#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3449
Steve Capper9e5fc742013-04-30 08:02:03 +01003450#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3451pte_t *huge_pte_alloc(struct mm_struct *mm,
3452 unsigned long addr, unsigned long sz)
3453{
3454 pgd_t *pgd;
3455 pud_t *pud;
3456 pte_t *pte = NULL;
3457
3458 pgd = pgd_offset(mm, addr);
3459 pud = pud_alloc(mm, pgd, addr);
3460 if (pud) {
3461 if (sz == PUD_SIZE) {
3462 pte = (pte_t *)pud;
3463 } else {
3464 BUG_ON(sz != PMD_SIZE);
3465 if (want_pmd_share() && pud_none(*pud))
3466 pte = huge_pmd_share(mm, addr, pud);
3467 else
3468 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3469 }
3470 }
3471 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3472
3473 return pte;
3474}
3475
3476pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3477{
3478 pgd_t *pgd;
3479 pud_t *pud;
3480 pmd_t *pmd = NULL;
3481
3482 pgd = pgd_offset(mm, addr);
3483 if (pgd_present(*pgd)) {
3484 pud = pud_offset(pgd, addr);
3485 if (pud_present(*pud)) {
3486 if (pud_huge(*pud))
3487 return (pte_t *)pud;
3488 pmd = pmd_offset(pud, addr);
3489 }
3490 }
3491 return (pte_t *) pmd;
3492}
3493
3494struct page *
3495follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3496 pmd_t *pmd, int write)
3497{
3498 struct page *page;
3499
3500 page = pte_page(*(pte_t *)pmd);
3501 if (page)
3502 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3503 return page;
3504}
3505
3506struct page *
3507follow_huge_pud(struct mm_struct *mm, unsigned long address,
3508 pud_t *pud, int write)
3509{
3510 struct page *page;
3511
3512 page = pte_page(*(pte_t *)pud);
3513 if (page)
3514 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3515 return page;
3516}
3517
3518#else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3519
3520/* Can be overriden by architectures */
3521__attribute__((weak)) struct page *
3522follow_huge_pud(struct mm_struct *mm, unsigned long address,
3523 pud_t *pud, int write)
3524{
3525 BUG();
3526 return NULL;
3527}
3528
3529#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3530
Andi Kleend5bd9102010-09-27 09:00:12 +02003531#ifdef CONFIG_MEMORY_FAILURE
3532
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003533/* Should be called in hugetlb_lock */
3534static int is_hugepage_on_freelist(struct page *hpage)
3535{
3536 struct page *page;
3537 struct page *tmp;
3538 struct hstate *h = page_hstate(hpage);
3539 int nid = page_to_nid(hpage);
3540
3541 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3542 if (page == hpage)
3543 return 1;
3544 return 0;
3545}
3546
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003547/*
3548 * This function is called from memory failure code.
3549 * Assume the caller holds page lock of the head page.
3550 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003551int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003552{
3553 struct hstate *h = page_hstate(hpage);
3554 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003555 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003556
3557 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003558 if (is_hugepage_on_freelist(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08003559 /*
3560 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3561 * but dangling hpage->lru can trigger list-debug warnings
3562 * (this happens when we call unpoison_memory() on it),
3563 * so let it point to itself with list_del_init().
3564 */
3565 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09003566 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003567 h->free_huge_pages--;
3568 h->free_huge_pages_node[nid]--;
3569 ret = 0;
3570 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003571 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003572 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003573}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003574#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003575
3576bool isolate_huge_page(struct page *page, struct list_head *list)
3577{
Sasha Levin309381fea2014-01-23 15:52:54 -08003578 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003579 if (!get_page_unless_zero(page))
3580 return false;
3581 spin_lock(&hugetlb_lock);
3582 list_move_tail(&page->lru, list);
3583 spin_unlock(&hugetlb_lock);
3584 return true;
3585}
3586
3587void putback_active_hugepage(struct page *page)
3588{
Sasha Levin309381fea2014-01-23 15:52:54 -08003589 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003590 spin_lock(&hugetlb_lock);
3591 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3592 spin_unlock(&hugetlb_lock);
3593 put_page(page);
3594}
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07003595
3596bool is_hugepage_active(struct page *page)
3597{
Sasha Levin309381fea2014-01-23 15:52:54 -08003598 VM_BUG_ON_PAGE(!PageHuge(page), page);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07003599 /*
3600 * This function can be called for a tail page because the caller,
3601 * scan_movable_pages, scans through a given pfn-range which typically
3602 * covers one memory block. In systems using gigantic hugepage (1GB
3603 * for x86_64,) a hugepage is larger than a memory block, and we don't
3604 * support migrating such large hugepages for now, so return false
3605 * when called for tail pages.
3606 */
3607 if (PageTail(page))
3608 return false;
3609 /*
3610 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3611 * so we should return false for them.
3612 */
3613 if (unlikely(PageHWPoison(page)))
3614 return false;
3615 return page_count(page) > 0;
3616}