]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blame - mm/hugetlb.c
net: dsa: mv88e6xxx: rename ATU MAC accessors
[mirror_ubuntu-artful-kernel.git] / mm / hugetlb.c
CommitLineData
1da177e4
LT
1/*
2 * Generic hugetlb support.
6d49e352 3 * (C) Nadia Yvette Chambers, April 2004
1da177e4 4 */
1da177e4
LT
5#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
e1759c21 9#include <linux/seq_file.h>
1da177e4
LT
10#include <linux/sysctl.h>
11#include <linux/highmem.h>
cddb8a5c 12#include <linux/mmu_notifier.h>
1da177e4 13#include <linux/nodemask.h>
63551ae0 14#include <linux/pagemap.h>
5da7ca86 15#include <linux/mempolicy.h>
3b32123d 16#include <linux/compiler.h>
aea47ff3 17#include <linux/cpuset.h>
3935baa9 18#include <linux/mutex.h>
aa888a74 19#include <linux/bootmem.h>
a3437870 20#include <linux/sysfs.h>
5a0e3ad6 21#include <linux/slab.h>
0fe6e20b 22#include <linux/rmap.h>
fd6a03ed
NH
23#include <linux/swap.h>
24#include <linux/swapops.h>
c8721bbb 25#include <linux/page-isolation.h>
8382d914 26#include <linux/jhash.h>
d6606683 27
63551ae0
DG
28#include <asm/page.h>
29#include <asm/pgtable.h>
24669e58 30#include <asm/tlb.h>
63551ae0 31
24669e58 32#include <linux/io.h>
63551ae0 33#include <linux/hugetlb.h>
9dd540e2 34#include <linux/hugetlb_cgroup.h>
9a305230 35#include <linux/node.h>
7835e98b 36#include "internal.h"
1da177e4 37
753162cd 38int hugepages_treat_as_movable;
a5516438 39
c3f38a38 40int hugetlb_max_hstate __read_mostly;
e5ff2159
AK
41unsigned int default_hstate_idx;
42struct hstate hstates[HUGE_MAX_HSTATE];
641844f5
NH
43/*
44 * Minimum page order among possible hugepage sizes, set to a proper value
45 * at boot time.
46 */
47static unsigned int minimum_order __read_mostly = UINT_MAX;
e5ff2159 48
53ba51d2
JT
49__initdata LIST_HEAD(huge_boot_pages);
50
e5ff2159
AK
51/* for command line parsing */
52static struct hstate * __initdata parsed_hstate;
53static unsigned long __initdata default_hstate_max_huge_pages;
e11bfbfc 54static unsigned long __initdata default_hstate_size;
e5ff2159 55
3935baa9 56/*
31caf665
NH
57 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
58 * free_huge_pages, and surplus_huge_pages.
3935baa9 59 */
c3f38a38 60DEFINE_SPINLOCK(hugetlb_lock);
0bd0f9fb 61
8382d914
DB
62/*
63 * Serializes faults on the same logical page. This is used to
64 * prevent spurious OOMs when the hugepage pool is fully utilized.
65 */
66static int num_fault_mutexes;
67static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp;
68
7ca02d0a
MK
69/* Forward declaration */
70static int hugetlb_acct_memory(struct hstate *h, long delta);
71
90481622
DG
72static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
73{
74 bool free = (spool->count == 0) && (spool->used_hpages == 0);
75
76 spin_unlock(&spool->lock);
77
78 /* If no pages are used, and no other handles to the subpool
7ca02d0a
MK
79 * remain, give up any reservations mased on minimum size and
80 * free the subpool */
81 if (free) {
82 if (spool->min_hpages != -1)
83 hugetlb_acct_memory(spool->hstate,
84 -spool->min_hpages);
90481622 85 kfree(spool);
7ca02d0a 86 }
90481622
DG
87}
88
7ca02d0a
MK
89struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
90 long min_hpages)
90481622
DG
91{
92 struct hugepage_subpool *spool;
93
c6a91820 94 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
90481622
DG
95 if (!spool)
96 return NULL;
97
98 spin_lock_init(&spool->lock);
99 spool->count = 1;
7ca02d0a
MK
100 spool->max_hpages = max_hpages;
101 spool->hstate = h;
102 spool->min_hpages = min_hpages;
103
104 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
105 kfree(spool);
106 return NULL;
107 }
108 spool->rsv_hpages = min_hpages;
90481622
DG
109
110 return spool;
111}
112
113void hugepage_put_subpool(struct hugepage_subpool *spool)
114{
115 spin_lock(&spool->lock);
116 BUG_ON(!spool->count);
117 spool->count--;
118 unlock_or_release_subpool(spool);
119}
120
1c5ecae3
MK
121/*
122 * Subpool accounting for allocating and reserving pages.
123 * Return -ENOMEM if there are not enough resources to satisfy the
124 * the request. Otherwise, return the number of pages by which the
125 * global pools must be adjusted (upward). The returned value may
126 * only be different than the passed value (delta) in the case where
127 * a subpool minimum size must be manitained.
128 */
129static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
90481622
DG
130 long delta)
131{
1c5ecae3 132 long ret = delta;
90481622
DG
133
134 if (!spool)
1c5ecae3 135 return ret;
90481622
DG
136
137 spin_lock(&spool->lock);
1c5ecae3
MK
138
139 if (spool->max_hpages != -1) { /* maximum size accounting */
140 if ((spool->used_hpages + delta) <= spool->max_hpages)
141 spool->used_hpages += delta;
142 else {
143 ret = -ENOMEM;
144 goto unlock_ret;
145 }
90481622 146 }
90481622 147
1c5ecae3
MK
148 if (spool->min_hpages != -1) { /* minimum size accounting */
149 if (delta > spool->rsv_hpages) {
150 /*
151 * Asking for more reserves than those already taken on
152 * behalf of subpool. Return difference.
153 */
154 ret = delta - spool->rsv_hpages;
155 spool->rsv_hpages = 0;
156 } else {
157 ret = 0; /* reserves already accounted for */
158 spool->rsv_hpages -= delta;
159 }
160 }
161
162unlock_ret:
163 spin_unlock(&spool->lock);
90481622
DG
164 return ret;
165}
166
1c5ecae3
MK
167/*
168 * Subpool accounting for freeing and unreserving pages.
169 * Return the number of global page reservations that must be dropped.
170 * The return value may only be different than the passed value (delta)
171 * in the case where a subpool minimum size must be maintained.
172 */
173static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
90481622
DG
174 long delta)
175{
1c5ecae3
MK
176 long ret = delta;
177
90481622 178 if (!spool)
1c5ecae3 179 return delta;
90481622
DG
180
181 spin_lock(&spool->lock);
1c5ecae3
MK
182
183 if (spool->max_hpages != -1) /* maximum size accounting */
184 spool->used_hpages -= delta;
185
186 if (spool->min_hpages != -1) { /* minimum size accounting */
187 if (spool->rsv_hpages + delta <= spool->min_hpages)
188 ret = 0;
189 else
190 ret = spool->rsv_hpages + delta - spool->min_hpages;
191
192 spool->rsv_hpages += delta;
193 if (spool->rsv_hpages > spool->min_hpages)
194 spool->rsv_hpages = spool->min_hpages;
195 }
196
197 /*
198 * If hugetlbfs_put_super couldn't free spool due to an outstanding
199 * quota reference, free it now.
200 */
90481622 201 unlock_or_release_subpool(spool);
1c5ecae3
MK
202
203 return ret;
90481622
DG
204}
205
206static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
207{
208 return HUGETLBFS_SB(inode->i_sb)->spool;
209}
210
211static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
212{
496ad9aa 213 return subpool_inode(file_inode(vma->vm_file));
90481622
DG
214}
215
96822904
AW
216/*
217 * Region tracking -- allows tracking of reservations and instantiated pages
218 * across the pages in a mapping.
84afd99b 219 *
1dd308a7
MK
220 * The region data structures are embedded into a resv_map and protected
221 * by a resv_map's lock. The set of regions within the resv_map represent
222 * reservations for huge pages, or huge pages that have already been
223 * instantiated within the map. The from and to elements are huge page
224 * indicies into the associated mapping. from indicates the starting index
225 * of the region. to represents the first index past the end of the region.
226 *
227 * For example, a file region structure with from == 0 and to == 4 represents
228 * four huge pages in a mapping. It is important to note that the to element
229 * represents the first element past the end of the region. This is used in
230 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
231 *
232 * Interval notation of the form [from, to) will be used to indicate that
233 * the endpoint from is inclusive and to is exclusive.
96822904
AW
234 */
235struct file_region {
236 struct list_head link;
237 long from;
238 long to;
239};
240
1dd308a7
MK
241/*
242 * Add the huge page range represented by [f, t) to the reserve
243 * map. Existing regions will be expanded to accommodate the
244 * specified range. We know only existing regions need to be
245 * expanded, because region_add is only called after region_chg
246 * with the same range. If a new file_region structure must
247 * be allocated, it is done in region_chg.
cf3ad20b
MK
248 *
249 * Return the number of new huge pages added to the map. This
250 * number is greater than or equal to zero.
1dd308a7 251 */
1406ec9b 252static long region_add(struct resv_map *resv, long f, long t)
96822904 253{
1406ec9b 254 struct list_head *head = &resv->regions;
96822904 255 struct file_region *rg, *nrg, *trg;
cf3ad20b 256 long add = 0;
96822904 257
7b24d861 258 spin_lock(&resv->lock);
96822904
AW
259 /* Locate the region we are either in or before. */
260 list_for_each_entry(rg, head, link)
261 if (f <= rg->to)
262 break;
263
264 /* Round our left edge to the current segment if it encloses us. */
265 if (f > rg->from)
266 f = rg->from;
267
268 /* Check for and consume any regions we now overlap with. */
269 nrg = rg;
270 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
271 if (&rg->link == head)
272 break;
273 if (rg->from > t)
274 break;
275
276 /* If this area reaches higher then extend our area to
277 * include it completely. If this is not the first area
278 * which we intend to reuse, free it. */
279 if (rg->to > t)
280 t = rg->to;
281 if (rg != nrg) {
cf3ad20b
MK
282 /* Decrement return value by the deleted range.
283 * Another range will span this area so that by
284 * end of routine add will be >= zero
285 */
286 add -= (rg->to - rg->from);
96822904
AW
287 list_del(&rg->link);
288 kfree(rg);
289 }
290 }
cf3ad20b
MK
291
292 add += (nrg->from - f); /* Added to beginning of region */
96822904 293 nrg->from = f;
cf3ad20b 294 add += t - nrg->to; /* Added to end of region */
96822904 295 nrg->to = t;
cf3ad20b 296
7b24d861 297 spin_unlock(&resv->lock);
cf3ad20b
MK
298 VM_BUG_ON(add < 0);
299 return add;
96822904
AW
300}
301
1dd308a7
MK
302/*
303 * Examine the existing reserve map and determine how many
304 * huge pages in the specified range [f, t) are NOT currently
305 * represented. This routine is called before a subsequent
306 * call to region_add that will actually modify the reserve
307 * map to add the specified range [f, t). region_chg does
308 * not change the number of huge pages represented by the
309 * map. However, if the existing regions in the map can not
310 * be expanded to represent the new range, a new file_region
311 * structure is added to the map as a placeholder. This is
312 * so that the subsequent region_add call will have all the
313 * regions it needs and will not fail.
314 *
315 * Returns the number of huge pages that need to be added
316 * to the existing reservation map for the range [f, t).
317 * This number is greater or equal to zero. -ENOMEM is
318 * returned if a new file_region structure is needed and can
319 * not be allocated.
320 */
1406ec9b 321static long region_chg(struct resv_map *resv, long f, long t)
96822904 322{
1406ec9b 323 struct list_head *head = &resv->regions;
7b24d861 324 struct file_region *rg, *nrg = NULL;
96822904
AW
325 long chg = 0;
326
7b24d861
DB
327retry:
328 spin_lock(&resv->lock);
96822904
AW
329 /* Locate the region we are before or in. */
330 list_for_each_entry(rg, head, link)
331 if (f <= rg->to)
332 break;
333
334 /* If we are below the current region then a new region is required.
335 * Subtle, allocate a new region at the position but make it zero
336 * size such that we can guarantee to record the reservation. */
337 if (&rg->link == head || t < rg->from) {
7b24d861
DB
338 if (!nrg) {
339 spin_unlock(&resv->lock);
340 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
341 if (!nrg)
342 return -ENOMEM;
343
344 nrg->from = f;
345 nrg->to = f;
346 INIT_LIST_HEAD(&nrg->link);
347 goto retry;
348 }
96822904 349
7b24d861
DB
350 list_add(&nrg->link, rg->link.prev);
351 chg = t - f;
352 goto out_nrg;
96822904
AW
353 }
354
355 /* Round our left edge to the current segment if it encloses us. */
356 if (f > rg->from)
357 f = rg->from;
358 chg = t - f;
359
360 /* Check for and consume any regions we now overlap with. */
361 list_for_each_entry(rg, rg->link.prev, link) {
362 if (&rg->link == head)
363 break;
364 if (rg->from > t)
7b24d861 365 goto out;
96822904 366
25985edc 367 /* We overlap with this area, if it extends further than
96822904
AW
368 * us then we must extend ourselves. Account for its
369 * existing reservation. */
370 if (rg->to > t) {
371 chg += rg->to - t;
372 t = rg->to;
373 }
374 chg -= rg->to - rg->from;
375 }
7b24d861
DB
376
377out:
378 spin_unlock(&resv->lock);
379 /* We already know we raced and no longer need the new region */
380 kfree(nrg);
381 return chg;
382out_nrg:
383 spin_unlock(&resv->lock);
96822904
AW
384 return chg;
385}
386
1dd308a7
MK
387/*
388 * Truncate the reserve map at index 'end'. Modify/truncate any
389 * region which contains end. Delete any regions past end.
390 * Return the number of huge pages removed from the map.
391 */
1406ec9b 392static long region_truncate(struct resv_map *resv, long end)
96822904 393{
1406ec9b 394 struct list_head *head = &resv->regions;
96822904
AW
395 struct file_region *rg, *trg;
396 long chg = 0;
397
7b24d861 398 spin_lock(&resv->lock);
96822904
AW
399 /* Locate the region we are either in or before. */
400 list_for_each_entry(rg, head, link)
401 if (end <= rg->to)
402 break;
403 if (&rg->link == head)
7b24d861 404 goto out;
96822904
AW
405
406 /* If we are in the middle of a region then adjust it. */
407 if (end > rg->from) {
408 chg = rg->to - end;
409 rg->to = end;
410 rg = list_entry(rg->link.next, typeof(*rg), link);
411 }
412
413 /* Drop any remaining regions. */
414 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
415 if (&rg->link == head)
416 break;
417 chg += rg->to - rg->from;
418 list_del(&rg->link);
419 kfree(rg);
420 }
7b24d861
DB
421
422out:
423 spin_unlock(&resv->lock);
96822904
AW
424 return chg;
425}
426
1dd308a7
MK
427/*
428 * Count and return the number of huge pages in the reserve map
429 * that intersect with the range [f, t).
430 */
1406ec9b 431static long region_count(struct resv_map *resv, long f, long t)
84afd99b 432{
1406ec9b 433 struct list_head *head = &resv->regions;
84afd99b
AW
434 struct file_region *rg;
435 long chg = 0;
436
7b24d861 437 spin_lock(&resv->lock);
84afd99b
AW
438 /* Locate each segment we overlap with, and count that overlap. */
439 list_for_each_entry(rg, head, link) {
f2135a4a
WSH
440 long seg_from;
441 long seg_to;
84afd99b
AW
442
443 if (rg->to <= f)
444 continue;
445 if (rg->from >= t)
446 break;
447
448 seg_from = max(rg->from, f);
449 seg_to = min(rg->to, t);
450
451 chg += seg_to - seg_from;
452 }
7b24d861 453 spin_unlock(&resv->lock);
84afd99b
AW
454
455 return chg;
456}
457
e7c4b0bf
AW
458/*
459 * Convert the address within this vma to the page offset within
460 * the mapping, in pagecache page units; huge pages here.
461 */
a5516438
AK
462static pgoff_t vma_hugecache_offset(struct hstate *h,
463 struct vm_area_struct *vma, unsigned long address)
e7c4b0bf 464{
a5516438
AK
465 return ((address - vma->vm_start) >> huge_page_shift(h)) +
466 (vma->vm_pgoff >> huge_page_order(h));
e7c4b0bf
AW
467}
468
0fe6e20b
NH
469pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
470 unsigned long address)
471{
472 return vma_hugecache_offset(hstate_vma(vma), vma, address);
473}
474
08fba699
MG
475/*
476 * Return the size of the pages allocated when backing a VMA. In the majority
477 * cases this will be same size as used by the page table entries.
478 */
479unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
480{
481 struct hstate *hstate;
482
483 if (!is_vm_hugetlb_page(vma))
484 return PAGE_SIZE;
485
486 hstate = hstate_vma(vma);
487
2415cf12 488 return 1UL << huge_page_shift(hstate);
08fba699 489}
f340ca0f 490EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
08fba699 491
3340289d
MG
492/*
493 * Return the page size being used by the MMU to back a VMA. In the majority
494 * of cases, the page size used by the kernel matches the MMU size. On
495 * architectures where it differs, an architecture-specific version of this
496 * function is required.
497 */
498#ifndef vma_mmu_pagesize
499unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
500{
501 return vma_kernel_pagesize(vma);
502}
503#endif
504
84afd99b
AW
505/*
506 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
507 * bits of the reservation map pointer, which are always clear due to
508 * alignment.
509 */
510#define HPAGE_RESV_OWNER (1UL << 0)
511#define HPAGE_RESV_UNMAPPED (1UL << 1)
04f2cbe3 512#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
84afd99b 513
a1e78772
MG
514/*
515 * These helpers are used to track how many pages are reserved for
516 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
517 * is guaranteed to have their future faults succeed.
518 *
519 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
520 * the reserve counters are updated with the hugetlb_lock held. It is safe
521 * to reset the VMA at fork() time as it is not in use yet and there is no
522 * chance of the global counters getting corrupted as a result of the values.
84afd99b
AW
523 *
524 * The private mapping reservation is represented in a subtly different
525 * manner to a shared mapping. A shared mapping has a region map associated
526 * with the underlying file, this region map represents the backing file
527 * pages which have ever had a reservation assigned which this persists even
528 * after the page is instantiated. A private mapping has a region map
529 * associated with the original mmap which is attached to all VMAs which
530 * reference it, this region map represents those offsets which have consumed
531 * reservation ie. where pages have been instantiated.
a1e78772 532 */
e7c4b0bf
AW
533static unsigned long get_vma_private_data(struct vm_area_struct *vma)
534{
535 return (unsigned long)vma->vm_private_data;
536}
537
538static void set_vma_private_data(struct vm_area_struct *vma,
539 unsigned long value)
540{
541 vma->vm_private_data = (void *)value;
542}
543
9119a41e 544struct resv_map *resv_map_alloc(void)
84afd99b
AW
545{
546 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
547 if (!resv_map)
548 return NULL;
549
550 kref_init(&resv_map->refs);
7b24d861 551 spin_lock_init(&resv_map->lock);
84afd99b
AW
552 INIT_LIST_HEAD(&resv_map->regions);
553
554 return resv_map;
555}
556
9119a41e 557void resv_map_release(struct kref *ref)
84afd99b
AW
558{
559 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
560
561 /* Clear out any active regions before we release the map. */
1406ec9b 562 region_truncate(resv_map, 0);
84afd99b
AW
563 kfree(resv_map);
564}
565
4e35f483
JK
566static inline struct resv_map *inode_resv_map(struct inode *inode)
567{
568 return inode->i_mapping->private_data;
569}
570
84afd99b 571static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
a1e78772 572{
81d1b09c 573 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
4e35f483
JK
574 if (vma->vm_flags & VM_MAYSHARE) {
575 struct address_space *mapping = vma->vm_file->f_mapping;
576 struct inode *inode = mapping->host;
577
578 return inode_resv_map(inode);
579
580 } else {
84afd99b
AW
581 return (struct resv_map *)(get_vma_private_data(vma) &
582 ~HPAGE_RESV_MASK);
4e35f483 583 }
a1e78772
MG
584}
585
84afd99b 586static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
a1e78772 587{
81d1b09c
SL
588 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
589 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
a1e78772 590
84afd99b
AW
591 set_vma_private_data(vma, (get_vma_private_data(vma) &
592 HPAGE_RESV_MASK) | (unsigned long)map);
04f2cbe3
MG
593}
594
595static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
596{
81d1b09c
SL
597 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
598 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
e7c4b0bf
AW
599
600 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
04f2cbe3
MG
601}
602
603static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
604{
81d1b09c 605 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
e7c4b0bf
AW
606
607 return (get_vma_private_data(vma) & flag) != 0;
a1e78772
MG
608}
609
04f2cbe3 610/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
a1e78772
MG
611void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
612{
81d1b09c 613 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
f83a275d 614 if (!(vma->vm_flags & VM_MAYSHARE))
a1e78772
MG
615 vma->vm_private_data = (void *)0;
616}
617
618/* Returns true if the VMA has associated reserve pages */
af0ed73e 619static int vma_has_reserves(struct vm_area_struct *vma, long chg)
a1e78772 620{
af0ed73e
JK
621 if (vma->vm_flags & VM_NORESERVE) {
622 /*
623 * This address is already reserved by other process(chg == 0),
624 * so, we should decrement reserved count. Without decrementing,
625 * reserve count remains after releasing inode, because this
626 * allocated page will go into page cache and is regarded as
627 * coming from reserved pool in releasing step. Currently, we
628 * don't have any other solution to deal with this situation
629 * properly, so add work-around here.
630 */
631 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
632 return 1;
633 else
634 return 0;
635 }
a63884e9
JK
636
637 /* Shared mappings always use reserves */
f83a275d 638 if (vma->vm_flags & VM_MAYSHARE)
7f09ca51 639 return 1;
a63884e9
JK
640
641 /*
642 * Only the process that called mmap() has reserves for
643 * private mappings.
644 */
7f09ca51
MG
645 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
646 return 1;
a63884e9 647
7f09ca51 648 return 0;
a1e78772
MG
649}
650
a5516438 651static void enqueue_huge_page(struct hstate *h, struct page *page)
1da177e4
LT
652{
653 int nid = page_to_nid(page);
0edaecfa 654 list_move(&page->lru, &h->hugepage_freelists[nid]);
a5516438
AK
655 h->free_huge_pages++;
656 h->free_huge_pages_node[nid]++;
1da177e4
LT
657}
658
bf50bab2
NH
659static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
660{
661 struct page *page;
662
c8721bbb
NH
663 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
664 if (!is_migrate_isolate_page(page))
665 break;
666 /*
667 * if 'non-isolated free hugepage' not found on the list,
668 * the allocation fails.
669 */
670 if (&h->hugepage_freelists[nid] == &page->lru)
bf50bab2 671 return NULL;
0edaecfa 672 list_move(&page->lru, &h->hugepage_activelist);
a9869b83 673 set_page_refcounted(page);
bf50bab2
NH
674 h->free_huge_pages--;
675 h->free_huge_pages_node[nid]--;
676 return page;
677}
678
86cdb465
NH
679/* Movability of hugepages depends on migration support. */
680static inline gfp_t htlb_alloc_mask(struct hstate *h)
681{
100873d7 682 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
86cdb465
NH
683 return GFP_HIGHUSER_MOVABLE;
684 else
685 return GFP_HIGHUSER;
686}
687
a5516438
AK
688static struct page *dequeue_huge_page_vma(struct hstate *h,
689 struct vm_area_struct *vma,
af0ed73e
JK
690 unsigned long address, int avoid_reserve,
691 long chg)
1da177e4 692{
b1c12cbc 693 struct page *page = NULL;
480eccf9 694 struct mempolicy *mpol;
19770b32 695 nodemask_t *nodemask;
c0ff7453 696 struct zonelist *zonelist;
dd1a239f
MG
697 struct zone *zone;
698 struct zoneref *z;
cc9a6c87 699 unsigned int cpuset_mems_cookie;
1da177e4 700
a1e78772
MG
701 /*
702 * A child process with MAP_PRIVATE mappings created by their parent
703 * have no page reserves. This check ensures that reservations are
704 * not "stolen". The child may still get SIGKILLed
705 */
af0ed73e 706 if (!vma_has_reserves(vma, chg) &&
a5516438 707 h->free_huge_pages - h->resv_huge_pages == 0)
c0ff7453 708 goto err;
a1e78772 709
04f2cbe3 710 /* If reserves cannot be used, ensure enough pages are in the pool */
a5516438 711 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
6eab04a8 712 goto err;
04f2cbe3 713
9966c4bb 714retry_cpuset:
d26914d1 715 cpuset_mems_cookie = read_mems_allowed_begin();
9966c4bb 716 zonelist = huge_zonelist(vma, address,
86cdb465 717 htlb_alloc_mask(h), &mpol, &nodemask);
9966c4bb 718
19770b32
MG
719 for_each_zone_zonelist_nodemask(zone, z, zonelist,
720 MAX_NR_ZONES - 1, nodemask) {
344736f2 721 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
bf50bab2
NH
722 page = dequeue_huge_page_node(h, zone_to_nid(zone));
723 if (page) {
af0ed73e
JK
724 if (avoid_reserve)
725 break;
726 if (!vma_has_reserves(vma, chg))
727 break;
728
07443a85 729 SetPagePrivate(page);
af0ed73e 730 h->resv_huge_pages--;
bf50bab2
NH
731 break;
732 }
3abf7afd 733 }
1da177e4 734 }
cc9a6c87 735
52cd3b07 736 mpol_cond_put(mpol);
d26914d1 737 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
cc9a6c87 738 goto retry_cpuset;
1da177e4 739 return page;
cc9a6c87
MG
740
741err:
cc9a6c87 742 return NULL;
1da177e4
LT
743}
744
1cac6f2c
LC
745/*
746 * common helper functions for hstate_next_node_to_{alloc|free}.
747 * We may have allocated or freed a huge page based on a different
748 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
749 * be outside of *nodes_allowed. Ensure that we use an allowed
750 * node for alloc or free.
751 */
752static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
753{
754 nid = next_node(nid, *nodes_allowed);
755 if (nid == MAX_NUMNODES)
756 nid = first_node(*nodes_allowed);
757 VM_BUG_ON(nid >= MAX_NUMNODES);
758
759 return nid;
760}
761
762static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
763{
764 if (!node_isset(nid, *nodes_allowed))
765 nid = next_node_allowed(nid, nodes_allowed);
766 return nid;
767}
768
769/*
770 * returns the previously saved node ["this node"] from which to
771 * allocate a persistent huge page for the pool and advance the
772 * next node from which to allocate, handling wrap at end of node
773 * mask.
774 */
775static int hstate_next_node_to_alloc(struct hstate *h,
776 nodemask_t *nodes_allowed)
777{
778 int nid;
779
780 VM_BUG_ON(!nodes_allowed);
781
782 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
783 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
784
785 return nid;
786}
787
788/*
789 * helper for free_pool_huge_page() - return the previously saved
790 * node ["this node"] from which to free a huge page. Advance the
791 * next node id whether or not we find a free huge page to free so
792 * that the next attempt to free addresses the next node.
793 */
794static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
795{
796 int nid;
797
798 VM_BUG_ON(!nodes_allowed);
799
800 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
801 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
802
803 return nid;
804}
805
806#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
807 for (nr_nodes = nodes_weight(*mask); \
808 nr_nodes > 0 && \
809 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
810 nr_nodes--)
811
812#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
813 for (nr_nodes = nodes_weight(*mask); \
814 nr_nodes > 0 && \
815 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
816 nr_nodes--)
817
944d9fec
LC
818#if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
819static void destroy_compound_gigantic_page(struct page *page,
820 unsigned long order)
821{
822 int i;
823 int nr_pages = 1 << order;
824 struct page *p = page + 1;
825
826 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
827 __ClearPageTail(p);
828 set_page_refcounted(p);
829 p->first_page = NULL;
830 }
831
832 set_compound_order(page, 0);
833 __ClearPageHead(page);
834}
835
836static void free_gigantic_page(struct page *page, unsigned order)
837{
838 free_contig_range(page_to_pfn(page), 1 << order);
839}
840
841static int __alloc_gigantic_page(unsigned long start_pfn,
842 unsigned long nr_pages)
843{
844 unsigned long end_pfn = start_pfn + nr_pages;
845 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
846}
847
848static bool pfn_range_valid_gigantic(unsigned long start_pfn,
849 unsigned long nr_pages)
850{
851 unsigned long i, end_pfn = start_pfn + nr_pages;
852 struct page *page;
853
854 for (i = start_pfn; i < end_pfn; i++) {
855 if (!pfn_valid(i))
856 return false;
857
858 page = pfn_to_page(i);
859
860 if (PageReserved(page))
861 return false;
862
863 if (page_count(page) > 0)
864 return false;
865
866 if (PageHuge(page))
867 return false;
868 }
869
870 return true;
871}
872
873static bool zone_spans_last_pfn(const struct zone *zone,
874 unsigned long start_pfn, unsigned long nr_pages)
875{
876 unsigned long last_pfn = start_pfn + nr_pages - 1;
877 return zone_spans_pfn(zone, last_pfn);
878}
879
880static struct page *alloc_gigantic_page(int nid, unsigned order)
881{
882 unsigned long nr_pages = 1 << order;
883 unsigned long ret, pfn, flags;
884 struct zone *z;
885
886 z = NODE_DATA(nid)->node_zones;
887 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
888 spin_lock_irqsave(&z->lock, flags);
889
890 pfn = ALIGN(z->zone_start_pfn, nr_pages);
891 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
892 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
893 /*
894 * We release the zone lock here because
895 * alloc_contig_range() will also lock the zone
896 * at some point. If there's an allocation
897 * spinning on this lock, it may win the race
898 * and cause alloc_contig_range() to fail...
899 */
900 spin_unlock_irqrestore(&z->lock, flags);
901 ret = __alloc_gigantic_page(pfn, nr_pages);
902 if (!ret)
903 return pfn_to_page(pfn);
904 spin_lock_irqsave(&z->lock, flags);
905 }
906 pfn += nr_pages;
907 }
908
909 spin_unlock_irqrestore(&z->lock, flags);
910 }
911
912 return NULL;
913}
914
915static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
916static void prep_compound_gigantic_page(struct page *page, unsigned long order);
917
918static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
919{
920 struct page *page;
921
922 page = alloc_gigantic_page(nid, huge_page_order(h));
923 if (page) {
924 prep_compound_gigantic_page(page, huge_page_order(h));
925 prep_new_huge_page(h, page, nid);
926 }
927
928 return page;
929}
930
931static int alloc_fresh_gigantic_page(struct hstate *h,
932 nodemask_t *nodes_allowed)
933{
934 struct page *page = NULL;
935 int nr_nodes, node;
936
937 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
938 page = alloc_fresh_gigantic_page_node(h, node);
939 if (page)
940 return 1;
941 }
942
943 return 0;
944}
945
946static inline bool gigantic_page_supported(void) { return true; }
947#else
948static inline bool gigantic_page_supported(void) { return false; }
949static inline void free_gigantic_page(struct page *page, unsigned order) { }
950static inline void destroy_compound_gigantic_page(struct page *page,
951 unsigned long order) { }
952static inline int alloc_fresh_gigantic_page(struct hstate *h,
953 nodemask_t *nodes_allowed) { return 0; }
954#endif
955
a5516438 956static void update_and_free_page(struct hstate *h, struct page *page)
6af2acb6
AL
957{
958 int i;
a5516438 959
944d9fec
LC
960 if (hstate_is_gigantic(h) && !gigantic_page_supported())
961 return;
18229df5 962
a5516438
AK
963 h->nr_huge_pages--;
964 h->nr_huge_pages_node[page_to_nid(page)]--;
965 for (i = 0; i < pages_per_huge_page(h); i++) {
32f84528
CF
966 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
967 1 << PG_referenced | 1 << PG_dirty |
a7407a27
LC
968 1 << PG_active | 1 << PG_private |
969 1 << PG_writeback);
6af2acb6 970 }
309381fe 971 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
6af2acb6
AL
972 set_compound_page_dtor(page, NULL);
973 set_page_refcounted(page);
944d9fec
LC
974 if (hstate_is_gigantic(h)) {
975 destroy_compound_gigantic_page(page, huge_page_order(h));
976 free_gigantic_page(page, huge_page_order(h));
977 } else {
944d9fec
LC
978 __free_pages(page, huge_page_order(h));
979 }
6af2acb6
AL
980}
981
e5ff2159
AK
982struct hstate *size_to_hstate(unsigned long size)
983{
984 struct hstate *h;
985
986 for_each_hstate(h) {
987 if (huge_page_size(h) == size)
988 return h;
989 }
990 return NULL;
991}
992
bcc54222
NH
993/*
994 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
995 * to hstate->hugepage_activelist.)
996 *
997 * This function can be called for tail pages, but never returns true for them.
998 */
999bool page_huge_active(struct page *page)
1000{
1001 VM_BUG_ON_PAGE(!PageHuge(page), page);
1002 return PageHead(page) && PagePrivate(&page[1]);
1003}
1004
1005/* never called for tail page */
1006static void set_page_huge_active(struct page *page)
1007{
1008 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1009 SetPagePrivate(&page[1]);
1010}
1011
1012static void clear_page_huge_active(struct page *page)
1013{
1014 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1015 ClearPagePrivate(&page[1]);
1016}
1017
8f1d26d0 1018void free_huge_page(struct page *page)
27a85ef1 1019{
a5516438
AK
1020 /*
1021 * Can't pass hstate in here because it is called from the
1022 * compound page destructor.
1023 */
e5ff2159 1024 struct hstate *h = page_hstate(page);
7893d1d5 1025 int nid = page_to_nid(page);
90481622
DG
1026 struct hugepage_subpool *spool =
1027 (struct hugepage_subpool *)page_private(page);
07443a85 1028 bool restore_reserve;
27a85ef1 1029
e5df70ab 1030 set_page_private(page, 0);
23be7468 1031 page->mapping = NULL;
7893d1d5 1032 BUG_ON(page_count(page));
0fe6e20b 1033 BUG_ON(page_mapcount(page));
07443a85 1034 restore_reserve = PagePrivate(page);
16c794b4 1035 ClearPagePrivate(page);
27a85ef1 1036
1c5ecae3
MK
1037 /*
1038 * A return code of zero implies that the subpool will be under its
1039 * minimum size if the reservation is not restored after page is free.
1040 * Therefore, force restore_reserve operation.
1041 */
1042 if (hugepage_subpool_put_pages(spool, 1) == 0)
1043 restore_reserve = true;
1044
27a85ef1 1045 spin_lock(&hugetlb_lock);
bcc54222 1046 clear_page_huge_active(page);
6d76dcf4
AK
1047 hugetlb_cgroup_uncharge_page(hstate_index(h),
1048 pages_per_huge_page(h), page);
07443a85
JK
1049 if (restore_reserve)
1050 h->resv_huge_pages++;
1051
944d9fec 1052 if (h->surplus_huge_pages_node[nid]) {
0edaecfa
AK
1053 /* remove the page from active list */
1054 list_del(&page->lru);
a5516438
AK
1055 update_and_free_page(h, page);
1056 h->surplus_huge_pages--;
1057 h->surplus_huge_pages_node[nid]--;
7893d1d5 1058 } else {
5d3a551c 1059 arch_clear_hugepage_flags(page);
a5516438 1060 enqueue_huge_page(h, page);
7893d1d5 1061 }
27a85ef1
DG
1062 spin_unlock(&hugetlb_lock);
1063}
1064
a5516438 1065static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
b7ba30c6 1066{
0edaecfa 1067 INIT_LIST_HEAD(&page->lru);
b7ba30c6
AK
1068 set_compound_page_dtor(page, free_huge_page);
1069 spin_lock(&hugetlb_lock);
9dd540e2 1070 set_hugetlb_cgroup(page, NULL);
a5516438
AK
1071 h->nr_huge_pages++;
1072 h->nr_huge_pages_node[nid]++;
b7ba30c6
AK
1073 spin_unlock(&hugetlb_lock);
1074 put_page(page); /* free it into the hugepage allocator */
1075}
1076
2906dd52 1077static void prep_compound_gigantic_page(struct page *page, unsigned long order)
20a0307c
WF
1078{
1079 int i;
1080 int nr_pages = 1 << order;
1081 struct page *p = page + 1;
1082
1083 /* we rely on prep_new_huge_page to set the destructor */
1084 set_compound_order(page, order);
1085 __SetPageHead(page);
ef5a22be 1086 __ClearPageReserved(page);
20a0307c 1087 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
ef5a22be
AA
1088 /*
1089 * For gigantic hugepages allocated through bootmem at
1090 * boot, it's safer to be consistent with the not-gigantic
1091 * hugepages and clear the PG_reserved bit from all tail pages
1092 * too. Otherwse drivers using get_user_pages() to access tail
1093 * pages may get the reference counting wrong if they see
1094 * PG_reserved set on a tail page (despite the head page not
1095 * having PG_reserved set). Enforcing this consistency between
1096 * head and tail pages allows drivers to optimize away a check
1097 * on the head page when they need know if put_page() is needed
1098 * after get_user_pages().
1099 */
1100 __ClearPageReserved(p);
58a84aa9 1101 set_page_count(p, 0);
20a0307c 1102 p->first_page = page;
44fc8057
DR
1103 /* Make sure p->first_page is always valid for PageTail() */
1104 smp_wmb();
1105 __SetPageTail(p);
20a0307c
WF
1106 }
1107}
1108
7795912c
AM
1109/*
1110 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1111 * transparent huge pages. See the PageTransHuge() documentation for more
1112 * details.
1113 */
20a0307c
WF
1114int PageHuge(struct page *page)
1115{
20a0307c
WF
1116 if (!PageCompound(page))
1117 return 0;
1118
1119 page = compound_head(page);
758f66a2 1120 return get_compound_page_dtor(page) == free_huge_page;
20a0307c 1121}
43131e14
NH
1122EXPORT_SYMBOL_GPL(PageHuge);
1123
27c73ae7
AA
1124/*
1125 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1126 * normal or transparent huge pages.
1127 */
1128int PageHeadHuge(struct page *page_head)
1129{
27c73ae7
AA
1130 if (!PageHead(page_head))
1131 return 0;
1132
758f66a2 1133 return get_compound_page_dtor(page_head) == free_huge_page;
27c73ae7 1134}
27c73ae7 1135
13d60f4b
ZY
1136pgoff_t __basepage_index(struct page *page)
1137{
1138 struct page *page_head = compound_head(page);
1139 pgoff_t index = page_index(page_head);
1140 unsigned long compound_idx;
1141
1142 if (!PageHuge(page_head))
1143 return page_index(page);
1144
1145 if (compound_order(page_head) >= MAX_ORDER)
1146 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1147 else
1148 compound_idx = page - page_head;
1149
1150 return (index << compound_order(page_head)) + compound_idx;
1151}
1152
a5516438 1153static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1da177e4 1154{
1da177e4 1155 struct page *page;
f96efd58 1156
6484eb3e 1157 page = alloc_pages_exact_node(nid,
86cdb465 1158 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
551883ae 1159 __GFP_REPEAT|__GFP_NOWARN,
a5516438 1160 huge_page_order(h));
1da177e4 1161 if (page) {
a5516438 1162 prep_new_huge_page(h, page, nid);
1da177e4 1163 }
63b4613c
NA
1164
1165 return page;
1166}
1167
b2261026
JK
1168static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1169{
1170 struct page *page;
1171 int nr_nodes, node;
1172 int ret = 0;
1173
1174 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1175 page = alloc_fresh_huge_page_node(h, node);
1176 if (page) {
1177 ret = 1;
1178 break;
1179 }
1180 }
1181
1182 if (ret)
1183 count_vm_event(HTLB_BUDDY_PGALLOC);
1184 else
1185 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1186
1187 return ret;
1188}
1189
e8c5c824
LS
1190/*
1191 * Free huge page from pool from next node to free.
1192 * Attempt to keep persistent huge pages more or less
1193 * balanced over allowed nodes.
1194 * Called with hugetlb_lock locked.
1195 */
6ae11b27
LS
1196static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1197 bool acct_surplus)
e8c5c824 1198{
b2261026 1199 int nr_nodes, node;
e8c5c824
LS
1200 int ret = 0;
1201
b2261026 1202 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
685f3457
LS
1203 /*
1204 * If we're returning unused surplus pages, only examine
1205 * nodes with surplus pages.
1206 */
b2261026
JK
1207 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1208 !list_empty(&h->hugepage_freelists[node])) {
e8c5c824 1209 struct page *page =
b2261026 1210 list_entry(h->hugepage_freelists[node].next,
e8c5c824
LS
1211 struct page, lru);
1212 list_del(&page->lru);
1213 h->free_huge_pages--;
b2261026 1214 h->free_huge_pages_node[node]--;
685f3457
LS
1215 if (acct_surplus) {
1216 h->surplus_huge_pages--;
b2261026 1217 h->surplus_huge_pages_node[node]--;
685f3457 1218 }
e8c5c824
LS
1219 update_and_free_page(h, page);
1220 ret = 1;
9a76db09 1221 break;
e8c5c824 1222 }
b2261026 1223 }
e8c5c824
LS
1224
1225 return ret;
1226}
1227
c8721bbb
NH
1228/*
1229 * Dissolve a given free hugepage into free buddy pages. This function does
1230 * nothing for in-use (including surplus) hugepages.
1231 */
1232static void dissolve_free_huge_page(struct page *page)
1233{
1234 spin_lock(&hugetlb_lock);
1235 if (PageHuge(page) && !page_count(page)) {
1236 struct hstate *h = page_hstate(page);
1237 int nid = page_to_nid(page);
1238 list_del(&page->lru);
1239 h->free_huge_pages--;
1240 h->free_huge_pages_node[nid]--;
1241 update_and_free_page(h, page);
1242 }
1243 spin_unlock(&hugetlb_lock);
1244}
1245
1246/*
1247 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1248 * make specified memory blocks removable from the system.
1249 * Note that start_pfn should aligned with (minimum) hugepage size.
1250 */
1251void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1252{
c8721bbb 1253 unsigned long pfn;
c8721bbb 1254
d0177639
LZ
1255 if (!hugepages_supported())
1256 return;
1257
641844f5
NH
1258 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1259 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
c8721bbb
NH
1260 dissolve_free_huge_page(pfn_to_page(pfn));
1261}
1262
bf50bab2 1263static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
7893d1d5
AL
1264{
1265 struct page *page;
bf50bab2 1266 unsigned int r_nid;
7893d1d5 1267
bae7f4ae 1268 if (hstate_is_gigantic(h))
aa888a74
AK
1269 return NULL;
1270
d1c3fb1f
NA
1271 /*
1272 * Assume we will successfully allocate the surplus page to
1273 * prevent racing processes from causing the surplus to exceed
1274 * overcommit
1275 *
1276 * This however introduces a different race, where a process B
1277 * tries to grow the static hugepage pool while alloc_pages() is
1278 * called by process A. B will only examine the per-node
1279 * counters in determining if surplus huge pages can be
1280 * converted to normal huge pages in adjust_pool_surplus(). A
1281 * won't be able to increment the per-node counter, until the
1282 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1283 * no more huge pages can be converted from surplus to normal
1284 * state (and doesn't try to convert again). Thus, we have a
1285 * case where a surplus huge page exists, the pool is grown, and
1286 * the surplus huge page still exists after, even though it
1287 * should just have been converted to a normal huge page. This
1288 * does not leak memory, though, as the hugepage will be freed
1289 * once it is out of use. It also does not allow the counters to
1290 * go out of whack in adjust_pool_surplus() as we don't modify
1291 * the node values until we've gotten the hugepage and only the
1292 * per-node value is checked there.
1293 */
1294 spin_lock(&hugetlb_lock);
a5516438 1295 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
d1c3fb1f
NA
1296 spin_unlock(&hugetlb_lock);
1297 return NULL;
1298 } else {
a5516438
AK
1299 h->nr_huge_pages++;
1300 h->surplus_huge_pages++;
d1c3fb1f
NA
1301 }
1302 spin_unlock(&hugetlb_lock);
1303
bf50bab2 1304 if (nid == NUMA_NO_NODE)
86cdb465 1305 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
bf50bab2
NH
1306 __GFP_REPEAT|__GFP_NOWARN,
1307 huge_page_order(h));
1308 else
1309 page = alloc_pages_exact_node(nid,
86cdb465 1310 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
bf50bab2 1311 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
d1c3fb1f
NA
1312
1313 spin_lock(&hugetlb_lock);
7893d1d5 1314 if (page) {
0edaecfa 1315 INIT_LIST_HEAD(&page->lru);
bf50bab2 1316 r_nid = page_to_nid(page);
7893d1d5 1317 set_compound_page_dtor(page, free_huge_page);
9dd540e2 1318 set_hugetlb_cgroup(page, NULL);
d1c3fb1f
NA
1319 /*
1320 * We incremented the global counters already
1321 */
bf50bab2
NH
1322 h->nr_huge_pages_node[r_nid]++;
1323 h->surplus_huge_pages_node[r_nid]++;
3b116300 1324 __count_vm_event(HTLB_BUDDY_PGALLOC);
d1c3fb1f 1325 } else {
a5516438
AK
1326 h->nr_huge_pages--;
1327 h->surplus_huge_pages--;
3b116300 1328 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
7893d1d5 1329 }
d1c3fb1f 1330 spin_unlock(&hugetlb_lock);
7893d1d5
AL
1331
1332 return page;
1333}
1334
bf50bab2
NH
1335/*
1336 * This allocation function is useful in the context where vma is irrelevant.
1337 * E.g. soft-offlining uses this function because it only cares physical
1338 * address of error page.
1339 */
1340struct page *alloc_huge_page_node(struct hstate *h, int nid)
1341{
4ef91848 1342 struct page *page = NULL;
bf50bab2
NH
1343
1344 spin_lock(&hugetlb_lock);
4ef91848
JK
1345 if (h->free_huge_pages - h->resv_huge_pages > 0)
1346 page = dequeue_huge_page_node(h, nid);
bf50bab2
NH
1347 spin_unlock(&hugetlb_lock);
1348
94ae8ba7 1349 if (!page)
bf50bab2
NH
1350 page = alloc_buddy_huge_page(h, nid);
1351
1352 return page;
1353}
1354
e4e574b7 1355/*
25985edc 1356 * Increase the hugetlb pool such that it can accommodate a reservation
e4e574b7
AL
1357 * of size 'delta'.
1358 */
a5516438 1359static int gather_surplus_pages(struct hstate *h, int delta)
e4e574b7
AL
1360{
1361 struct list_head surplus_list;
1362 struct page *page, *tmp;
1363 int ret, i;
1364 int needed, allocated;
28073b02 1365 bool alloc_ok = true;
e4e574b7 1366
a5516438 1367 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
ac09b3a1 1368 if (needed <= 0) {
a5516438 1369 h->resv_huge_pages += delta;
e4e574b7 1370 return 0;
ac09b3a1 1371 }
e4e574b7
AL
1372
1373 allocated = 0;
1374 INIT_LIST_HEAD(&surplus_list);
1375
1376 ret = -ENOMEM;
1377retry:
1378 spin_unlock(&hugetlb_lock);
1379 for (i = 0; i < needed; i++) {
bf50bab2 1380 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
28073b02
HD
1381 if (!page) {
1382 alloc_ok = false;
1383 break;
1384 }
e4e574b7
AL
1385 list_add(&page->lru, &surplus_list);
1386 }
28073b02 1387 allocated += i;
e4e574b7
AL
1388
1389 /*
1390 * After retaking hugetlb_lock, we need to recalculate 'needed'
1391 * because either resv_huge_pages or free_huge_pages may have changed.
1392 */
1393 spin_lock(&hugetlb_lock);
a5516438
AK
1394 needed = (h->resv_huge_pages + delta) -
1395 (h->free_huge_pages + allocated);
28073b02
HD
1396 if (needed > 0) {
1397 if (alloc_ok)
1398 goto retry;
1399 /*
1400 * We were not able to allocate enough pages to
1401 * satisfy the entire reservation so we free what
1402 * we've allocated so far.
1403 */
1404 goto free;
1405 }
e4e574b7
AL
1406 /*
1407 * The surplus_list now contains _at_least_ the number of extra pages
25985edc 1408 * needed to accommodate the reservation. Add the appropriate number
e4e574b7 1409 * of pages to the hugetlb pool and free the extras back to the buddy
ac09b3a1
AL
1410 * allocator. Commit the entire reservation here to prevent another
1411 * process from stealing the pages as they are added to the pool but
1412 * before they are reserved.
e4e574b7
AL
1413 */
1414 needed += allocated;
a5516438 1415 h->resv_huge_pages += delta;
e4e574b7 1416 ret = 0;
a9869b83 1417
19fc3f0a 1418 /* Free the needed pages to the hugetlb pool */
e4e574b7 1419 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
19fc3f0a
AL
1420 if ((--needed) < 0)
1421 break;
a9869b83
NH
1422 /*
1423 * This page is now managed by the hugetlb allocator and has
1424 * no users -- drop the buddy allocator's reference.
1425 */
1426 put_page_testzero(page);
309381fe 1427 VM_BUG_ON_PAGE(page_count(page), page);
a5516438 1428 enqueue_huge_page(h, page);
19fc3f0a 1429 }
28073b02 1430free:
b0365c8d 1431 spin_unlock(&hugetlb_lock);
19fc3f0a
AL
1432
1433 /* Free unnecessary surplus pages to the buddy allocator */
c0d934ba
JK
1434 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1435 put_page(page);
a9869b83 1436 spin_lock(&hugetlb_lock);
e4e574b7
AL
1437
1438 return ret;
1439}
1440
1441/*
1442 * When releasing a hugetlb pool reservation, any surplus pages that were
1443 * allocated to satisfy the reservation must be explicitly freed if they were
1444 * never used.
685f3457 1445 * Called with hugetlb_lock held.
e4e574b7 1446 */
a5516438
AK
1447static void return_unused_surplus_pages(struct hstate *h,
1448 unsigned long unused_resv_pages)
e4e574b7 1449{
e4e574b7
AL
1450 unsigned long nr_pages;
1451
ac09b3a1 1452 /* Uncommit the reservation */
a5516438 1453 h->resv_huge_pages -= unused_resv_pages;
ac09b3a1 1454
aa888a74 1455 /* Cannot return gigantic pages currently */
bae7f4ae 1456 if (hstate_is_gigantic(h))
aa888a74
AK
1457 return;
1458
a5516438 1459 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
e4e574b7 1460
685f3457
LS
1461 /*
1462 * We want to release as many surplus pages as possible, spread
9b5e5d0f
LS
1463 * evenly across all nodes with memory. Iterate across these nodes
1464 * until we can no longer free unreserved surplus pages. This occurs
1465 * when the nodes with surplus pages have no free pages.
1466 * free_pool_huge_page() will balance the the freed pages across the
1467 * on-line nodes with memory and will handle the hstate accounting.
685f3457
LS
1468 */
1469 while (nr_pages--) {
8cebfcd0 1470 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
685f3457 1471 break;
7848a4bf 1472 cond_resched_lock(&hugetlb_lock);
e4e574b7
AL
1473 }
1474}
1475
c37f9fb1 1476/*
cf3ad20b
MK
1477 * vma_needs_reservation and vma_commit_reservation are used by the huge
1478 * page allocation routines to manage reservations.
1479 *
1480 * vma_needs_reservation is called to determine if the huge page at addr
1481 * within the vma has an associated reservation. If a reservation is
1482 * needed, the value 1 is returned. The caller is then responsible for
1483 * managing the global reservation and subpool usage counts. After
1484 * the huge page has been allocated, vma_commit_reservation is called
1485 * to add the page to the reservation map.
1486 *
1487 * In the normal case, vma_commit_reservation returns the same value
1488 * as the preceding vma_needs_reservation call. The only time this
1489 * is not the case is if a reserve map was changed between calls. It
1490 * is the responsibility of the caller to notice the difference and
1491 * take appropriate action.
c37f9fb1 1492 */
cf3ad20b
MK
1493static long __vma_reservation_common(struct hstate *h,
1494 struct vm_area_struct *vma, unsigned long addr,
1495 bool commit)
c37f9fb1 1496{
4e35f483
JK
1497 struct resv_map *resv;
1498 pgoff_t idx;
cf3ad20b 1499 long ret;
c37f9fb1 1500
4e35f483
JK
1501 resv = vma_resv_map(vma);
1502 if (!resv)
84afd99b 1503 return 1;
c37f9fb1 1504
4e35f483 1505 idx = vma_hugecache_offset(h, vma, addr);
cf3ad20b
MK
1506 if (commit)
1507 ret = region_add(resv, idx, idx + 1);
1508 else
1509 ret = region_chg(resv, idx, idx + 1);
84afd99b 1510
4e35f483 1511 if (vma->vm_flags & VM_MAYSHARE)
cf3ad20b 1512 return ret;
4e35f483 1513 else
cf3ad20b 1514 return ret < 0 ? ret : 0;
c37f9fb1 1515}
cf3ad20b
MK
1516
1517static long vma_needs_reservation(struct hstate *h,
a5516438 1518 struct vm_area_struct *vma, unsigned long addr)
c37f9fb1 1519{
cf3ad20b
MK
1520 return __vma_reservation_common(h, vma, addr, false);
1521}
84afd99b 1522
cf3ad20b
MK
1523static long vma_commit_reservation(struct hstate *h,
1524 struct vm_area_struct *vma, unsigned long addr)
1525{
1526 return __vma_reservation_common(h, vma, addr, true);
c37f9fb1
AW
1527}
1528
a1e78772 1529static struct page *alloc_huge_page(struct vm_area_struct *vma,
04f2cbe3 1530 unsigned long addr, int avoid_reserve)
1da177e4 1531{
90481622 1532 struct hugepage_subpool *spool = subpool_vma(vma);
a5516438 1533 struct hstate *h = hstate_vma(vma);
348ea204 1534 struct page *page;
33039678 1535 long chg, commit;
6d76dcf4
AK
1536 int ret, idx;
1537 struct hugetlb_cgroup *h_cg;
a1e78772 1538
6d76dcf4 1539 idx = hstate_index(h);
a1e78772 1540 /*
90481622
DG
1541 * Processes that did not create the mapping will have no
1542 * reserves and will not have accounted against subpool
1543 * limit. Check that the subpool limit can be made before
1544 * satisfying the allocation MAP_NORESERVE mappings may also
1545 * need pages and subpool limit allocated allocated if no reserve
1546 * mapping overlaps.
a1e78772 1547 */
a5516438 1548 chg = vma_needs_reservation(h, vma, addr);
c37f9fb1 1549 if (chg < 0)
76dcee75 1550 return ERR_PTR(-ENOMEM);
8bb3f12e 1551 if (chg || avoid_reserve)
1c5ecae3 1552 if (hugepage_subpool_get_pages(spool, 1) < 0)
76dcee75 1553 return ERR_PTR(-ENOSPC);
1da177e4 1554
6d76dcf4 1555 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
8f34af6f
JZ
1556 if (ret)
1557 goto out_subpool_put;
1558
1da177e4 1559 spin_lock(&hugetlb_lock);
af0ed73e 1560 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
81a6fcae 1561 if (!page) {
94ae8ba7 1562 spin_unlock(&hugetlb_lock);
bf50bab2 1563 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
8f34af6f
JZ
1564 if (!page)
1565 goto out_uncharge_cgroup;
1566
79dbb236
AK
1567 spin_lock(&hugetlb_lock);
1568 list_move(&page->lru, &h->hugepage_activelist);
81a6fcae 1569 /* Fall through */
68842c9b 1570 }
81a6fcae
JK
1571 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1572 spin_unlock(&hugetlb_lock);
348ea204 1573
90481622 1574 set_page_private(page, (unsigned long)spool);
90d8b7e6 1575
33039678
MK
1576 commit = vma_commit_reservation(h, vma, addr);
1577 if (unlikely(chg > commit)) {
1578 /*
1579 * The page was added to the reservation map between
1580 * vma_needs_reservation and vma_commit_reservation.
1581 * This indicates a race with hugetlb_reserve_pages.
1582 * Adjust for the subpool count incremented above AND
1583 * in hugetlb_reserve_pages for the same page. Also,
1584 * the reservation count added in hugetlb_reserve_pages
1585 * no longer applies.
1586 */
1587 long rsv_adjust;
1588
1589 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1590 hugetlb_acct_memory(h, -rsv_adjust);
1591 }
90d8b7e6 1592 return page;
8f34af6f
JZ
1593
1594out_uncharge_cgroup:
1595 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1596out_subpool_put:
1597 if (chg || avoid_reserve)
1598 hugepage_subpool_put_pages(spool, 1);
1599 return ERR_PTR(-ENOSPC);
b45b5bd6
DG
1600}
1601
74060e4d
NH
1602/*
1603 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1604 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1605 * where no ERR_VALUE is expected to be returned.
1606 */
1607struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1608 unsigned long addr, int avoid_reserve)
1609{
1610 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1611 if (IS_ERR(page))
1612 page = NULL;
1613 return page;
1614}
1615
91f47662 1616int __weak alloc_bootmem_huge_page(struct hstate *h)
aa888a74
AK
1617{
1618 struct huge_bootmem_page *m;
b2261026 1619 int nr_nodes, node;
aa888a74 1620
b2261026 1621 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
aa888a74
AK
1622 void *addr;
1623
8b89a116
GS
1624 addr = memblock_virt_alloc_try_nid_nopanic(
1625 huge_page_size(h), huge_page_size(h),
1626 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
aa888a74
AK
1627 if (addr) {
1628 /*
1629 * Use the beginning of the huge page to store the
1630 * huge_bootmem_page struct (until gather_bootmem
1631 * puts them into the mem_map).
1632 */
1633 m = addr;
91f47662 1634 goto found;
aa888a74 1635 }
aa888a74
AK
1636 }
1637 return 0;
1638
1639found:
df994ead 1640 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
aa888a74
AK
1641 /* Put them into a private list first because mem_map is not up yet */
1642 list_add(&m->list, &huge_boot_pages);
1643 m->hstate = h;
1644 return 1;
1645}
1646
f412c97a 1647static void __init prep_compound_huge_page(struct page *page, int order)
18229df5
AW
1648{
1649 if (unlikely(order > (MAX_ORDER - 1)))
1650 prep_compound_gigantic_page(page, order);
1651 else
1652 prep_compound_page(page, order);
1653}
1654
aa888a74
AK
1655/* Put bootmem huge pages into the standard lists after mem_map is up */
1656static void __init gather_bootmem_prealloc(void)
1657{
1658 struct huge_bootmem_page *m;
1659
1660 list_for_each_entry(m, &huge_boot_pages, list) {
aa888a74 1661 struct hstate *h = m->hstate;
ee8f248d
BB
1662 struct page *page;
1663
1664#ifdef CONFIG_HIGHMEM
1665 page = pfn_to_page(m->phys >> PAGE_SHIFT);
8b89a116
GS
1666 memblock_free_late(__pa(m),
1667 sizeof(struct huge_bootmem_page));
ee8f248d
BB
1668#else
1669 page = virt_to_page(m);
1670#endif
aa888a74 1671 WARN_ON(page_count(page) != 1);
18229df5 1672 prep_compound_huge_page(page, h->order);
ef5a22be 1673 WARN_ON(PageReserved(page));
aa888a74 1674 prep_new_huge_page(h, page, page_to_nid(page));
b0320c7b
RA
1675 /*
1676 * If we had gigantic hugepages allocated at boot time, we need
1677 * to restore the 'stolen' pages to totalram_pages in order to
1678 * fix confusing memory reports from free(1) and another
1679 * side-effects, like CommitLimit going negative.
1680 */
bae7f4ae 1681 if (hstate_is_gigantic(h))
3dcc0571 1682 adjust_managed_page_count(page, 1 << h->order);
aa888a74
AK
1683 }
1684}
1685
8faa8b07 1686static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1da177e4
LT
1687{
1688 unsigned long i;
a5516438 1689
e5ff2159 1690 for (i = 0; i < h->max_huge_pages; ++i) {
bae7f4ae 1691 if (hstate_is_gigantic(h)) {
aa888a74
AK
1692 if (!alloc_bootmem_huge_page(h))
1693 break;
9b5e5d0f 1694 } else if (!alloc_fresh_huge_page(h,
8cebfcd0 1695 &node_states[N_MEMORY]))
1da177e4 1696 break;
1da177e4 1697 }
8faa8b07 1698 h->max_huge_pages = i;
e5ff2159
AK
1699}
1700
1701static void __init hugetlb_init_hstates(void)
1702{
1703 struct hstate *h;
1704
1705 for_each_hstate(h) {
641844f5
NH
1706 if (minimum_order > huge_page_order(h))
1707 minimum_order = huge_page_order(h);
1708
8faa8b07 1709 /* oversize hugepages were init'ed in early boot */
bae7f4ae 1710 if (!hstate_is_gigantic(h))
8faa8b07 1711 hugetlb_hstate_alloc_pages(h);
e5ff2159 1712 }
641844f5 1713 VM_BUG_ON(minimum_order == UINT_MAX);
e5ff2159
AK
1714}
1715
4abd32db
AK
1716static char * __init memfmt(char *buf, unsigned long n)
1717{
1718 if (n >= (1UL << 30))
1719 sprintf(buf, "%lu GB", n >> 30);
1720 else if (n >= (1UL << 20))
1721 sprintf(buf, "%lu MB", n >> 20);
1722 else
1723 sprintf(buf, "%lu KB", n >> 10);
1724 return buf;
1725}
1726
e5ff2159
AK
1727static void __init report_hugepages(void)
1728{
1729 struct hstate *h;
1730
1731 for_each_hstate(h) {
4abd32db 1732 char buf[32];
ffb22af5 1733 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
4abd32db
AK
1734 memfmt(buf, huge_page_size(h)),
1735 h->free_huge_pages);
e5ff2159
AK
1736 }
1737}
1738
1da177e4 1739#ifdef CONFIG_HIGHMEM
6ae11b27
LS
1740static void try_to_free_low(struct hstate *h, unsigned long count,
1741 nodemask_t *nodes_allowed)
1da177e4 1742{
4415cc8d
CL
1743 int i;
1744
bae7f4ae 1745 if (hstate_is_gigantic(h))
aa888a74
AK
1746 return;
1747
6ae11b27 1748 for_each_node_mask(i, *nodes_allowed) {
1da177e4 1749 struct page *page, *next;
a5516438
AK
1750 struct list_head *freel = &h->hugepage_freelists[i];
1751 list_for_each_entry_safe(page, next, freel, lru) {
1752 if (count >= h->nr_huge_pages)
6b0c880d 1753 return;
1da177e4
LT
1754 if (PageHighMem(page))
1755 continue;
1756 list_del(&page->lru);
e5ff2159 1757 update_and_free_page(h, page);
a5516438
AK
1758 h->free_huge_pages--;
1759 h->free_huge_pages_node[page_to_nid(page)]--;
1da177e4
LT
1760 }
1761 }
1762}
1763#else
6ae11b27
LS
1764static inline void try_to_free_low(struct hstate *h, unsigned long count,
1765 nodemask_t *nodes_allowed)
1da177e4
LT
1766{
1767}
1768#endif
1769
20a0307c
WF
1770/*
1771 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1772 * balanced by operating on them in a round-robin fashion.
1773 * Returns 1 if an adjustment was made.
1774 */
6ae11b27
LS
1775static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1776 int delta)
20a0307c 1777{
b2261026 1778 int nr_nodes, node;
20a0307c
WF
1779
1780 VM_BUG_ON(delta != -1 && delta != 1);
20a0307c 1781
b2261026
JK
1782 if (delta < 0) {
1783 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1784 if (h->surplus_huge_pages_node[node])
1785 goto found;
e8c5c824 1786 }
b2261026
JK
1787 } else {
1788 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1789 if (h->surplus_huge_pages_node[node] <
1790 h->nr_huge_pages_node[node])
1791 goto found;
e8c5c824 1792 }
b2261026
JK
1793 }
1794 return 0;
20a0307c 1795
b2261026
JK
1796found:
1797 h->surplus_huge_pages += delta;
1798 h->surplus_huge_pages_node[node] += delta;
1799 return 1;
20a0307c
WF
1800}
1801
a5516438 1802#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
6ae11b27
LS
1803static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1804 nodemask_t *nodes_allowed)
1da177e4 1805{
7893d1d5 1806 unsigned long min_count, ret;
1da177e4 1807
944d9fec 1808 if (hstate_is_gigantic(h) && !gigantic_page_supported())
aa888a74
AK
1809 return h->max_huge_pages;
1810
7893d1d5
AL
1811 /*
1812 * Increase the pool size
1813 * First take pages out of surplus state. Then make up the
1814 * remaining difference by allocating fresh huge pages.
d1c3fb1f
NA
1815 *
1816 * We might race with alloc_buddy_huge_page() here and be unable
1817 * to convert a surplus huge page to a normal huge page. That is
1818 * not critical, though, it just means the overall size of the
1819 * pool might be one hugepage larger than it needs to be, but
1820 * within all the constraints specified by the sysctls.
7893d1d5 1821 */
1da177e4 1822 spin_lock(&hugetlb_lock);
a5516438 1823 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
6ae11b27 1824 if (!adjust_pool_surplus(h, nodes_allowed, -1))
7893d1d5
AL
1825 break;
1826 }
1827
a5516438 1828 while (count > persistent_huge_pages(h)) {
7893d1d5
AL
1829 /*
1830 * If this allocation races such that we no longer need the
1831 * page, free_huge_page will handle it by freeing the page
1832 * and reducing the surplus.
1833 */
1834 spin_unlock(&hugetlb_lock);
944d9fec
LC
1835 if (hstate_is_gigantic(h))
1836 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
1837 else
1838 ret = alloc_fresh_huge_page(h, nodes_allowed);
7893d1d5
AL
1839 spin_lock(&hugetlb_lock);
1840 if (!ret)
1841 goto out;
1842
536240f2
MG
1843 /* Bail for signals. Probably ctrl-c from user */
1844 if (signal_pending(current))
1845 goto out;
7893d1d5 1846 }
7893d1d5
AL
1847
1848 /*
1849 * Decrease the pool size
1850 * First return free pages to the buddy allocator (being careful
1851 * to keep enough around to satisfy reservations). Then place
1852 * pages into surplus state as needed so the pool will shrink
1853 * to the desired size as pages become free.
d1c3fb1f
NA
1854 *
1855 * By placing pages into the surplus state independent of the
1856 * overcommit value, we are allowing the surplus pool size to
1857 * exceed overcommit. There are few sane options here. Since
1858 * alloc_buddy_huge_page() is checking the global counter,
1859 * though, we'll note that we're not allowed to exceed surplus
1860 * and won't grow the pool anywhere else. Not until one of the
1861 * sysctls are changed, or the surplus pages go out of use.
7893d1d5 1862 */
a5516438 1863 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
6b0c880d 1864 min_count = max(count, min_count);
6ae11b27 1865 try_to_free_low(h, min_count, nodes_allowed);
a5516438 1866 while (min_count < persistent_huge_pages(h)) {
6ae11b27 1867 if (!free_pool_huge_page(h, nodes_allowed, 0))
1da177e4 1868 break;
55f67141 1869 cond_resched_lock(&hugetlb_lock);
1da177e4 1870 }
a5516438 1871 while (count < persistent_huge_pages(h)) {
6ae11b27 1872 if (!adjust_pool_surplus(h, nodes_allowed, 1))
7893d1d5
AL
1873 break;
1874 }
1875out:
a5516438 1876 ret = persistent_huge_pages(h);
1da177e4 1877 spin_unlock(&hugetlb_lock);
7893d1d5 1878 return ret;
1da177e4
LT
1879}
1880
a3437870
NA
1881#define HSTATE_ATTR_RO(_name) \
1882 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1883
1884#define HSTATE_ATTR(_name) \
1885 static struct kobj_attribute _name##_attr = \
1886 __ATTR(_name, 0644, _name##_show, _name##_store)
1887
1888static struct kobject *hugepages_kobj;
1889static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1890
9a305230
LS
1891static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1892
1893static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
a3437870
NA
1894{
1895 int i;
9a305230 1896
a3437870 1897 for (i = 0; i < HUGE_MAX_HSTATE; i++)
9a305230
LS
1898 if (hstate_kobjs[i] == kobj) {
1899 if (nidp)
1900 *nidp = NUMA_NO_NODE;
a3437870 1901 return &hstates[i];
9a305230
LS
1902 }
1903
1904 return kobj_to_node_hstate(kobj, nidp);
a3437870
NA
1905}
1906
06808b08 1907static ssize_t nr_hugepages_show_common(struct kobject *kobj,
a3437870
NA
1908 struct kobj_attribute *attr, char *buf)
1909{
9a305230
LS
1910 struct hstate *h;
1911 unsigned long nr_huge_pages;
1912 int nid;
1913
1914 h = kobj_to_hstate(kobj, &nid);
1915 if (nid == NUMA_NO_NODE)
1916 nr_huge_pages = h->nr_huge_pages;
1917 else
1918 nr_huge_pages = h->nr_huge_pages_node[nid];
1919
1920 return sprintf(buf, "%lu\n", nr_huge_pages);
a3437870 1921}
adbe8726 1922
238d3c13
DR
1923static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
1924 struct hstate *h, int nid,
1925 unsigned long count, size_t len)
a3437870
NA
1926{
1927 int err;
bad44b5b 1928 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
a3437870 1929
944d9fec 1930 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
adbe8726
EM
1931 err = -EINVAL;
1932 goto out;
1933 }
1934
9a305230
LS
1935 if (nid == NUMA_NO_NODE) {
1936 /*
1937 * global hstate attribute
1938 */
1939 if (!(obey_mempolicy &&
1940 init_nodemask_of_mempolicy(nodes_allowed))) {
1941 NODEMASK_FREE(nodes_allowed);
8cebfcd0 1942 nodes_allowed = &node_states[N_MEMORY];
9a305230
LS
1943 }
1944 } else if (nodes_allowed) {
1945 /*
1946 * per node hstate attribute: adjust count to global,
1947 * but restrict alloc/free to the specified node.
1948 */
1949 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1950 init_nodemask_of_node(nodes_allowed, nid);
1951 } else
8cebfcd0 1952 nodes_allowed = &node_states[N_MEMORY];
9a305230 1953
06808b08 1954 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
a3437870 1955
8cebfcd0 1956 if (nodes_allowed != &node_states[N_MEMORY])
06808b08
LS
1957 NODEMASK_FREE(nodes_allowed);
1958
1959 return len;
adbe8726
EM
1960out:
1961 NODEMASK_FREE(nodes_allowed);
1962 return err;
06808b08
LS
1963}
1964
238d3c13
DR
1965static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1966 struct kobject *kobj, const char *buf,
1967 size_t len)
1968{
1969 struct hstate *h;
1970 unsigned long count;
1971 int nid;
1972 int err;
1973
1974 err = kstrtoul(buf, 10, &count);
1975 if (err)
1976 return err;
1977
1978 h = kobj_to_hstate(kobj, &nid);
1979 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
1980}
1981
06808b08
LS
1982static ssize_t nr_hugepages_show(struct kobject *kobj,
1983 struct kobj_attribute *attr, char *buf)
1984{
1985 return nr_hugepages_show_common(kobj, attr, buf);
1986}
1987
1988static ssize_t nr_hugepages_store(struct kobject *kobj,
1989 struct kobj_attribute *attr, const char *buf, size_t len)
1990{
238d3c13 1991 return nr_hugepages_store_common(false, kobj, buf, len);
a3437870
NA
1992}
1993HSTATE_ATTR(nr_hugepages);
1994
06808b08
LS
1995#ifdef CONFIG_NUMA
1996
1997/*
1998 * hstate attribute for optionally mempolicy-based constraint on persistent
1999 * huge page alloc/free.
2000 */
2001static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2002 struct kobj_attribute *attr, char *buf)
2003{
2004 return nr_hugepages_show_common(kobj, attr, buf);
2005}
2006
2007static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2008 struct kobj_attribute *attr, const char *buf, size_t len)
2009{
238d3c13 2010 return nr_hugepages_store_common(true, kobj, buf, len);
06808b08
LS
2011}
2012HSTATE_ATTR(nr_hugepages_mempolicy);
2013#endif
2014
2015
a3437870
NA
2016static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2017 struct kobj_attribute *attr, char *buf)
2018{
9a305230 2019 struct hstate *h = kobj_to_hstate(kobj, NULL);
a3437870
NA
2020 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2021}
adbe8726 2022
a3437870
NA
2023static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2024 struct kobj_attribute *attr, const char *buf, size_t count)
2025{
2026 int err;
2027 unsigned long input;
9a305230 2028 struct hstate *h = kobj_to_hstate(kobj, NULL);
a3437870 2029
bae7f4ae 2030 if (hstate_is_gigantic(h))
adbe8726
EM
2031 return -EINVAL;
2032
3dbb95f7 2033 err = kstrtoul(buf, 10, &input);
a3437870 2034 if (err)
73ae31e5 2035 return err;
a3437870
NA
2036
2037 spin_lock(&hugetlb_lock);
2038 h->nr_overcommit_huge_pages = input;
2039 spin_unlock(&hugetlb_lock);
2040
2041 return count;
2042}
2043HSTATE_ATTR(nr_overcommit_hugepages);
2044
2045static ssize_t free_hugepages_show(struct kobject *kobj,
2046 struct kobj_attribute *attr, char *buf)
2047{
9a305230
LS
2048 struct hstate *h;
2049 unsigned long free_huge_pages;
2050 int nid;
2051
2052 h = kobj_to_hstate(kobj, &nid);
2053 if (nid == NUMA_NO_NODE)
2054 free_huge_pages = h->free_huge_pages;
2055 else
2056 free_huge_pages = h->free_huge_pages_node[nid];
2057
2058 return sprintf(buf, "%lu\n", free_huge_pages);
a3437870
NA
2059}
2060HSTATE_ATTR_RO(free_hugepages);
2061
2062static ssize_t resv_hugepages_show(struct kobject *kobj,
2063 struct kobj_attribute *attr, char *buf)
2064{
9a305230 2065 struct hstate *h = kobj_to_hstate(kobj, NULL);
a3437870
NA
2066 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2067}
2068HSTATE_ATTR_RO(resv_hugepages);
2069
2070static ssize_t surplus_hugepages_show(struct kobject *kobj,
2071 struct kobj_attribute *attr, char *buf)
2072{
9a305230
LS
2073 struct hstate *h;
2074 unsigned long surplus_huge_pages;
2075 int nid;
2076
2077 h = kobj_to_hstate(kobj, &nid);
2078 if (nid == NUMA_NO_NODE)
2079 surplus_huge_pages = h->surplus_huge_pages;
2080 else
2081 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2082
2083 return sprintf(buf, "%lu\n", surplus_huge_pages);
a3437870
NA
2084}
2085HSTATE_ATTR_RO(surplus_hugepages);
2086
2087static struct attribute *hstate_attrs[] = {
2088 &nr_hugepages_attr.attr,
2089 &nr_overcommit_hugepages_attr.attr,
2090 &free_hugepages_attr.attr,
2091 &resv_hugepages_attr.attr,
2092 &surplus_hugepages_attr.attr,
06808b08
LS
2093#ifdef CONFIG_NUMA
2094 &nr_hugepages_mempolicy_attr.attr,
2095#endif
a3437870
NA
2096 NULL,
2097};
2098
2099static struct attribute_group hstate_attr_group = {
2100 .attrs = hstate_attrs,
2101};
2102
094e9539
JM
2103static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2104 struct kobject **hstate_kobjs,
2105 struct attribute_group *hstate_attr_group)
a3437870
NA
2106{
2107 int retval;
972dc4de 2108 int hi = hstate_index(h);
a3437870 2109
9a305230
LS
2110 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2111 if (!hstate_kobjs[hi])
a3437870
NA
2112 return -ENOMEM;
2113
9a305230 2114 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
a3437870 2115 if (retval)
9a305230 2116 kobject_put(hstate_kobjs[hi]);
a3437870
NA
2117
2118 return retval;
2119}
2120
2121static void __init hugetlb_sysfs_init(void)
2122{
2123 struct hstate *h;
2124 int err;
2125
2126 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2127 if (!hugepages_kobj)
2128 return;
2129
2130 for_each_hstate(h) {
9a305230
LS
2131 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2132 hstate_kobjs, &hstate_attr_group);
a3437870 2133 if (err)
ffb22af5 2134 pr_err("Hugetlb: Unable to add hstate %s", h->name);
a3437870
NA
2135 }
2136}
2137
9a305230
LS
2138#ifdef CONFIG_NUMA
2139
2140/*
2141 * node_hstate/s - associate per node hstate attributes, via their kobjects,
10fbcf4c
KS
2142 * with node devices in node_devices[] using a parallel array. The array
2143 * index of a node device or _hstate == node id.
2144 * This is here to avoid any static dependency of the node device driver, in
9a305230
LS
2145 * the base kernel, on the hugetlb module.
2146 */
2147struct node_hstate {
2148 struct kobject *hugepages_kobj;
2149 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2150};
2151struct node_hstate node_hstates[MAX_NUMNODES];
2152
2153/*
10fbcf4c 2154 * A subset of global hstate attributes for node devices
9a305230
LS
2155 */
2156static struct attribute *per_node_hstate_attrs[] = {
2157 &nr_hugepages_attr.attr,
2158 &free_hugepages_attr.attr,
2159 &surplus_hugepages_attr.attr,
2160 NULL,
2161};
2162
2163static struct attribute_group per_node_hstate_attr_group = {
2164 .attrs = per_node_hstate_attrs,
2165};
2166
2167/*
10fbcf4c 2168 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
9a305230
LS
2169 * Returns node id via non-NULL nidp.
2170 */
2171static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2172{
2173 int nid;
2174
2175 for (nid = 0; nid < nr_node_ids; nid++) {
2176 struct node_hstate *nhs = &node_hstates[nid];
2177 int i;
2178 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2179 if (nhs->hstate_kobjs[i] == kobj) {
2180 if (nidp)
2181 *nidp = nid;
2182 return &hstates[i];
2183 }
2184 }
2185
2186 BUG();
2187 return NULL;
2188}
2189
2190/*
10fbcf4c 2191 * Unregister hstate attributes from a single node device.
9a305230
LS
2192 * No-op if no hstate attributes attached.
2193 */
3cd8b44f 2194static void hugetlb_unregister_node(struct node *node)
9a305230
LS
2195{
2196 struct hstate *h;
10fbcf4c 2197 struct node_hstate *nhs = &node_hstates[node->dev.id];
9a305230
LS
2198
2199 if (!nhs->hugepages_kobj)
9b5e5d0f 2200 return; /* no hstate attributes */
9a305230 2201
972dc4de
AK
2202 for_each_hstate(h) {
2203 int idx = hstate_index(h);
2204 if (nhs->hstate_kobjs[idx]) {
2205 kobject_put(nhs->hstate_kobjs[idx]);
2206 nhs->hstate_kobjs[idx] = NULL;
9a305230 2207 }
972dc4de 2208 }
9a305230
LS
2209
2210 kobject_put(nhs->hugepages_kobj);
2211 nhs->hugepages_kobj = NULL;
2212}
2213
2214/*
10fbcf4c 2215 * hugetlb module exit: unregister hstate attributes from node devices
9a305230
LS
2216 * that have them.
2217 */
2218static void hugetlb_unregister_all_nodes(void)
2219{
2220 int nid;
2221
2222 /*
10fbcf4c 2223 * disable node device registrations.
9a305230
LS
2224 */
2225 register_hugetlbfs_with_node(NULL, NULL);
2226
2227 /*
2228 * remove hstate attributes from any nodes that have them.
2229 */
2230 for (nid = 0; nid < nr_node_ids; nid++)
8732794b 2231 hugetlb_unregister_node(node_devices[nid]);
9a305230
LS
2232}
2233
2234/*
10fbcf4c 2235 * Register hstate attributes for a single node device.
9a305230
LS
2236 * No-op if attributes already registered.
2237 */
3cd8b44f 2238static void hugetlb_register_node(struct node *node)
9a305230
LS
2239{
2240 struct hstate *h;
10fbcf4c 2241 struct node_hstate *nhs = &node_hstates[node->dev.id];
9a305230
LS
2242 int err;
2243
2244 if (nhs->hugepages_kobj)
2245 return; /* already allocated */
2246
2247 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
10fbcf4c 2248 &node->dev.kobj);
9a305230
LS
2249 if (!nhs->hugepages_kobj)
2250 return;
2251
2252 for_each_hstate(h) {
2253 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2254 nhs->hstate_kobjs,
2255 &per_node_hstate_attr_group);
2256 if (err) {
ffb22af5
AM
2257 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2258 h->name, node->dev.id);
9a305230
LS
2259 hugetlb_unregister_node(node);
2260 break;
2261 }
2262 }
2263}
2264
2265/*
9b5e5d0f 2266 * hugetlb init time: register hstate attributes for all registered node
10fbcf4c
KS
2267 * devices of nodes that have memory. All on-line nodes should have
2268 * registered their associated device by this time.
9a305230 2269 */
7d9ca000 2270static void __init hugetlb_register_all_nodes(void)
9a305230
LS
2271{
2272 int nid;
2273
8cebfcd0 2274 for_each_node_state(nid, N_MEMORY) {
8732794b 2275 struct node *node = node_devices[nid];
10fbcf4c 2276 if (node->dev.id == nid)
9a305230
LS
2277 hugetlb_register_node(node);
2278 }
2279
2280 /*
10fbcf4c 2281 * Let the node device driver know we're here so it can
9a305230
LS
2282 * [un]register hstate attributes on node hotplug.
2283 */
2284 register_hugetlbfs_with_node(hugetlb_register_node,
2285 hugetlb_unregister_node);
2286}
2287#else /* !CONFIG_NUMA */
2288
2289static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2290{
2291 BUG();
2292 if (nidp)
2293 *nidp = -1;
2294 return NULL;
2295}
2296
2297static void hugetlb_unregister_all_nodes(void) { }
2298
2299static void hugetlb_register_all_nodes(void) { }
2300
2301#endif
2302
a3437870
NA
2303static void __exit hugetlb_exit(void)
2304{
2305 struct hstate *h;
2306
9a305230
LS
2307 hugetlb_unregister_all_nodes();
2308
a3437870 2309 for_each_hstate(h) {
972dc4de 2310 kobject_put(hstate_kobjs[hstate_index(h)]);
a3437870
NA
2311 }
2312
2313 kobject_put(hugepages_kobj);
8382d914 2314 kfree(htlb_fault_mutex_table);
a3437870
NA
2315}
2316module_exit(hugetlb_exit);
2317
2318static int __init hugetlb_init(void)
2319{
8382d914
DB
2320 int i;
2321
457c1b27 2322 if (!hugepages_supported())
0ef89d25 2323 return 0;
a3437870 2324
e11bfbfc
NP
2325 if (!size_to_hstate(default_hstate_size)) {
2326 default_hstate_size = HPAGE_SIZE;
2327 if (!size_to_hstate(default_hstate_size))
2328 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
a3437870 2329 }
972dc4de 2330 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
e11bfbfc
NP
2331 if (default_hstate_max_huge_pages)
2332 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
a3437870
NA
2333
2334 hugetlb_init_hstates();
aa888a74 2335 gather_bootmem_prealloc();
a3437870
NA
2336 report_hugepages();
2337
2338 hugetlb_sysfs_init();
9a305230 2339 hugetlb_register_all_nodes();
7179e7bf 2340 hugetlb_cgroup_file_init();
9a305230 2341
8382d914
DB
2342#ifdef CONFIG_SMP
2343 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2344#else
2345 num_fault_mutexes = 1;
2346#endif
2347 htlb_fault_mutex_table =
2348 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2349 BUG_ON(!htlb_fault_mutex_table);
2350
2351 for (i = 0; i < num_fault_mutexes; i++)
2352 mutex_init(&htlb_fault_mutex_table[i]);
a3437870
NA
2353 return 0;
2354}
2355module_init(hugetlb_init);
2356
2357/* Should be called on processing a hugepagesz=... option */
2358void __init hugetlb_add_hstate(unsigned order)
2359{
2360 struct hstate *h;
8faa8b07
AK
2361 unsigned long i;
2362
a3437870 2363 if (size_to_hstate(PAGE_SIZE << order)) {
ffb22af5 2364 pr_warning("hugepagesz= specified twice, ignoring\n");
a3437870
NA
2365 return;
2366 }
47d38344 2367 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
a3437870 2368 BUG_ON(order == 0);
47d38344 2369 h = &hstates[hugetlb_max_hstate++];
a3437870
NA
2370 h->order = order;
2371 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
8faa8b07
AK
2372 h->nr_huge_pages = 0;
2373 h->free_huge_pages = 0;
2374 for (i = 0; i < MAX_NUMNODES; ++i)
2375 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
0edaecfa 2376 INIT_LIST_HEAD(&h->hugepage_activelist);
8cebfcd0
LJ
2377 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2378 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
a3437870
NA
2379 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2380 huge_page_size(h)/1024);
8faa8b07 2381
a3437870
NA
2382 parsed_hstate = h;
2383}
2384
e11bfbfc 2385static int __init hugetlb_nrpages_setup(char *s)
a3437870
NA
2386{
2387 unsigned long *mhp;
8faa8b07 2388 static unsigned long *last_mhp;
a3437870
NA
2389
2390 /*
47d38344 2391 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
a3437870
NA
2392 * so this hugepages= parameter goes to the "default hstate".
2393 */
47d38344 2394 if (!hugetlb_max_hstate)
a3437870
NA
2395 mhp = &default_hstate_max_huge_pages;
2396 else
2397 mhp = &parsed_hstate->max_huge_pages;
2398
8faa8b07 2399 if (mhp == last_mhp) {
ffb22af5
AM
2400 pr_warning("hugepages= specified twice without "
2401 "interleaving hugepagesz=, ignoring\n");
8faa8b07
AK
2402 return 1;
2403 }
2404
a3437870
NA
2405 if (sscanf(s, "%lu", mhp) <= 0)
2406 *mhp = 0;
2407
8faa8b07
AK
2408 /*
2409 * Global state is always initialized later in hugetlb_init.
2410 * But we need to allocate >= MAX_ORDER hstates here early to still
2411 * use the bootmem allocator.
2412 */
47d38344 2413 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
8faa8b07
AK
2414 hugetlb_hstate_alloc_pages(parsed_hstate);
2415
2416 last_mhp = mhp;
2417
a3437870
NA
2418 return 1;
2419}
e11bfbfc
NP
2420__setup("hugepages=", hugetlb_nrpages_setup);
2421
2422static int __init hugetlb_default_setup(char *s)
2423{
2424 default_hstate_size = memparse(s, &s);
2425 return 1;
2426}
2427__setup("default_hugepagesz=", hugetlb_default_setup);
a3437870 2428
8a213460
NA
2429static unsigned int cpuset_mems_nr(unsigned int *array)
2430{
2431 int node;
2432 unsigned int nr = 0;
2433
2434 for_each_node_mask(node, cpuset_current_mems_allowed)
2435 nr += array[node];
2436
2437 return nr;
2438}
2439
2440#ifdef CONFIG_SYSCTL
06808b08
LS
2441static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2442 struct ctl_table *table, int write,
2443 void __user *buffer, size_t *length, loff_t *ppos)
1da177e4 2444{
e5ff2159 2445 struct hstate *h = &default_hstate;
238d3c13 2446 unsigned long tmp = h->max_huge_pages;
08d4a246 2447 int ret;
e5ff2159 2448
457c1b27
NA
2449 if (!hugepages_supported())
2450 return -ENOTSUPP;
2451
e5ff2159
AK
2452 table->data = &tmp;
2453 table->maxlen = sizeof(unsigned long);
08d4a246
MH
2454 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2455 if (ret)
2456 goto out;
e5ff2159 2457
238d3c13
DR
2458 if (write)
2459 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2460 NUMA_NO_NODE, tmp, *length);
08d4a246
MH
2461out:
2462 return ret;
1da177e4 2463}
396faf03 2464
06808b08
LS
2465int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2466 void __user *buffer, size_t *length, loff_t *ppos)
2467{
2468
2469 return hugetlb_sysctl_handler_common(false, table, write,
2470 buffer, length, ppos);
2471}
2472
2473#ifdef CONFIG_NUMA
2474int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2475 void __user *buffer, size_t *length, loff_t *ppos)
2476{
2477 return hugetlb_sysctl_handler_common(true, table, write,
2478 buffer, length, ppos);
2479}
2480#endif /* CONFIG_NUMA */
2481
a3d0c6aa 2482int hugetlb_overcommit_handler(struct ctl_table *table, int write,
8d65af78 2483 void __user *buffer,
a3d0c6aa
NA
2484 size_t *length, loff_t *ppos)
2485{
a5516438 2486 struct hstate *h = &default_hstate;
e5ff2159 2487 unsigned long tmp;
08d4a246 2488 int ret;
e5ff2159 2489
457c1b27
NA
2490 if (!hugepages_supported())
2491 return -ENOTSUPP;
2492
c033a93c 2493 tmp = h->nr_overcommit_huge_pages;
e5ff2159 2494
bae7f4ae 2495 if (write && hstate_is_gigantic(h))
adbe8726
EM
2496 return -EINVAL;
2497
e5ff2159
AK
2498 table->data = &tmp;
2499 table->maxlen = sizeof(unsigned long);
08d4a246
MH
2500 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2501 if (ret)
2502 goto out;
e5ff2159
AK
2503
2504 if (write) {
2505 spin_lock(&hugetlb_lock);
2506 h->nr_overcommit_huge_pages = tmp;
2507 spin_unlock(&hugetlb_lock);
2508 }
08d4a246
MH
2509out:
2510 return ret;
a3d0c6aa
NA
2511}
2512
1da177e4
LT
2513#endif /* CONFIG_SYSCTL */
2514
e1759c21 2515void hugetlb_report_meminfo(struct seq_file *m)
1da177e4 2516{
a5516438 2517 struct hstate *h = &default_hstate;
457c1b27
NA
2518 if (!hugepages_supported())
2519 return;
e1759c21 2520 seq_printf(m,
4f98a2fe
RR
2521 "HugePages_Total: %5lu\n"
2522 "HugePages_Free: %5lu\n"
2523 "HugePages_Rsvd: %5lu\n"
2524 "HugePages_Surp: %5lu\n"
2525 "Hugepagesize: %8lu kB\n",
a5516438
AK
2526 h->nr_huge_pages,
2527 h->free_huge_pages,
2528 h->resv_huge_pages,
2529 h->surplus_huge_pages,
2530 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1da177e4
LT
2531}
2532
2533int hugetlb_report_node_meminfo(int nid, char *buf)
2534{
a5516438 2535 struct hstate *h = &default_hstate;
457c1b27
NA
2536 if (!hugepages_supported())
2537 return 0;
1da177e4
LT
2538 return sprintf(buf,
2539 "Node %d HugePages_Total: %5u\n"
a1de0919
NA
2540 "Node %d HugePages_Free: %5u\n"
2541 "Node %d HugePages_Surp: %5u\n",
a5516438
AK
2542 nid, h->nr_huge_pages_node[nid],
2543 nid, h->free_huge_pages_node[nid],
2544 nid, h->surplus_huge_pages_node[nid]);
1da177e4
LT
2545}
2546
949f7ec5
DR
2547void hugetlb_show_meminfo(void)
2548{
2549 struct hstate *h;
2550 int nid;
2551
457c1b27
NA
2552 if (!hugepages_supported())
2553 return;
2554
949f7ec5
DR
2555 for_each_node_state(nid, N_MEMORY)
2556 for_each_hstate(h)
2557 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2558 nid,
2559 h->nr_huge_pages_node[nid],
2560 h->free_huge_pages_node[nid],
2561 h->surplus_huge_pages_node[nid],
2562 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2563}
2564
1da177e4
LT
2565/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2566unsigned long hugetlb_total_pages(void)
2567{
d0028588
WL
2568 struct hstate *h;
2569 unsigned long nr_total_pages = 0;
2570
2571 for_each_hstate(h)
2572 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2573 return nr_total_pages;
1da177e4 2574}
1da177e4 2575
a5516438 2576static int hugetlb_acct_memory(struct hstate *h, long delta)
fc1b8a73
MG
2577{
2578 int ret = -ENOMEM;
2579
2580 spin_lock(&hugetlb_lock);
2581 /*
2582 * When cpuset is configured, it breaks the strict hugetlb page
2583 * reservation as the accounting is done on a global variable. Such
2584 * reservation is completely rubbish in the presence of cpuset because
2585 * the reservation is not checked against page availability for the
2586 * current cpuset. Application can still potentially OOM'ed by kernel
2587 * with lack of free htlb page in cpuset that the task is in.
2588 * Attempt to enforce strict accounting with cpuset is almost
2589 * impossible (or too ugly) because cpuset is too fluid that
2590 * task or memory node can be dynamically moved between cpusets.
2591 *
2592 * The change of semantics for shared hugetlb mapping with cpuset is
2593 * undesirable. However, in order to preserve some of the semantics,
2594 * we fall back to check against current free page availability as
2595 * a best attempt and hopefully to minimize the impact of changing
2596 * semantics that cpuset has.
2597 */
2598 if (delta > 0) {
a5516438 2599 if (gather_surplus_pages(h, delta) < 0)
fc1b8a73
MG
2600 goto out;
2601
a5516438
AK
2602 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2603 return_unused_surplus_pages(h, delta);
fc1b8a73
MG
2604 goto out;
2605 }
2606 }
2607
2608 ret = 0;
2609 if (delta < 0)
a5516438 2610 return_unused_surplus_pages(h, (unsigned long) -delta);
fc1b8a73
MG
2611
2612out:
2613 spin_unlock(&hugetlb_lock);
2614 return ret;
2615}
2616
84afd99b
AW
2617static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2618{
f522c3ac 2619 struct resv_map *resv = vma_resv_map(vma);
84afd99b
AW
2620
2621 /*
2622 * This new VMA should share its siblings reservation map if present.
2623 * The VMA will only ever have a valid reservation map pointer where
2624 * it is being copied for another still existing VMA. As that VMA
25985edc 2625 * has a reference to the reservation map it cannot disappear until
84afd99b
AW
2626 * after this open call completes. It is therefore safe to take a
2627 * new reference here without additional locking.
2628 */
4e35f483 2629 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
f522c3ac 2630 kref_get(&resv->refs);
84afd99b
AW
2631}
2632
a1e78772
MG
2633static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2634{
a5516438 2635 struct hstate *h = hstate_vma(vma);
f522c3ac 2636 struct resv_map *resv = vma_resv_map(vma);
90481622 2637 struct hugepage_subpool *spool = subpool_vma(vma);
4e35f483 2638 unsigned long reserve, start, end;
1c5ecae3 2639 long gbl_reserve;
84afd99b 2640
4e35f483
JK
2641 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2642 return;
84afd99b 2643
4e35f483
JK
2644 start = vma_hugecache_offset(h, vma, vma->vm_start);
2645 end = vma_hugecache_offset(h, vma, vma->vm_end);
84afd99b 2646
4e35f483 2647 reserve = (end - start) - region_count(resv, start, end);
84afd99b 2648
4e35f483
JK
2649 kref_put(&resv->refs, resv_map_release);
2650
2651 if (reserve) {
1c5ecae3
MK
2652 /*
2653 * Decrement reserve counts. The global reserve count may be
2654 * adjusted if the subpool has a minimum size.
2655 */
2656 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2657 hugetlb_acct_memory(h, -gbl_reserve);
84afd99b 2658 }
a1e78772
MG
2659}
2660
1da177e4
LT
2661/*
2662 * We cannot handle pagefaults against hugetlb pages at all. They cause
2663 * handle_mm_fault() to try to instantiate regular-sized pages in the
2664 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2665 * this far.
2666 */
d0217ac0 2667static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1da177e4
LT
2668{
2669 BUG();
d0217ac0 2670 return 0;
1da177e4
LT
2671}
2672
f0f37e2f 2673const struct vm_operations_struct hugetlb_vm_ops = {
d0217ac0 2674 .fault = hugetlb_vm_op_fault,
84afd99b 2675 .open = hugetlb_vm_op_open,
a1e78772 2676 .close = hugetlb_vm_op_close,
1da177e4
LT
2677};
2678
1e8f889b
DG
2679static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2680 int writable)
63551ae0
DG
2681{
2682 pte_t entry;
2683
1e8f889b 2684 if (writable) {
106c992a
GS
2685 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2686 vma->vm_page_prot)));
63551ae0 2687 } else {
106c992a
GS
2688 entry = huge_pte_wrprotect(mk_huge_pte(page,
2689 vma->vm_page_prot));
63551ae0
DG
2690 }
2691 entry = pte_mkyoung(entry);
2692 entry = pte_mkhuge(entry);
d9ed9faa 2693 entry = arch_make_huge_pte(entry, vma, page, writable);
63551ae0
DG
2694
2695 return entry;
2696}
2697
1e8f889b
DG
2698static void set_huge_ptep_writable(struct vm_area_struct *vma,
2699 unsigned long address, pte_t *ptep)
2700{
2701 pte_t entry;
2702
106c992a 2703 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
32f84528 2704 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4b3073e1 2705 update_mmu_cache(vma, address, ptep);
1e8f889b
DG
2706}
2707
4a705fef
NH
2708static int is_hugetlb_entry_migration(pte_t pte)
2709{
2710 swp_entry_t swp;
2711
2712 if (huge_pte_none(pte) || pte_present(pte))
2713 return 0;
2714 swp = pte_to_swp_entry(pte);
2715 if (non_swap_entry(swp) && is_migration_entry(swp))
2716 return 1;
2717 else
2718 return 0;
2719}
2720
2721static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2722{
2723 swp_entry_t swp;
2724
2725 if (huge_pte_none(pte) || pte_present(pte))
2726 return 0;
2727 swp = pte_to_swp_entry(pte);
2728 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2729 return 1;
2730 else
2731 return 0;
2732}
1e8f889b 2733
63551ae0
DG
2734int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2735 struct vm_area_struct *vma)
2736{
2737 pte_t *src_pte, *dst_pte, entry;
2738 struct page *ptepage;
1c59827d 2739 unsigned long addr;
1e8f889b 2740 int cow;
a5516438
AK
2741 struct hstate *h = hstate_vma(vma);
2742 unsigned long sz = huge_page_size(h);
e8569dd2
AS
2743 unsigned long mmun_start; /* For mmu_notifiers */
2744 unsigned long mmun_end; /* For mmu_notifiers */
2745 int ret = 0;
1e8f889b
DG
2746
2747 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
63551ae0 2748
e8569dd2
AS
2749 mmun_start = vma->vm_start;
2750 mmun_end = vma->vm_end;
2751 if (cow)
2752 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2753
a5516438 2754 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
cb900f41 2755 spinlock_t *src_ptl, *dst_ptl;
c74df32c
HD
2756 src_pte = huge_pte_offset(src, addr);
2757 if (!src_pte)
2758 continue;
a5516438 2759 dst_pte = huge_pte_alloc(dst, addr, sz);
e8569dd2
AS
2760 if (!dst_pte) {
2761 ret = -ENOMEM;
2762 break;
2763 }
c5c99429
LW
2764
2765 /* If the pagetables are shared don't copy or take references */
2766 if (dst_pte == src_pte)
2767 continue;
2768
cb900f41
KS
2769 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2770 src_ptl = huge_pte_lockptr(h, src, src_pte);
2771 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4a705fef
NH
2772 entry = huge_ptep_get(src_pte);
2773 if (huge_pte_none(entry)) { /* skip none entry */
2774 ;
2775 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2776 is_hugetlb_entry_hwpoisoned(entry))) {
2777 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2778
2779 if (is_write_migration_entry(swp_entry) && cow) {
2780 /*
2781 * COW mappings require pages in both
2782 * parent and child to be set to read.
2783 */
2784 make_migration_entry_read(&swp_entry);
2785 entry = swp_entry_to_pte(swp_entry);
2786 set_huge_pte_at(src, addr, src_pte, entry);
2787 }
2788 set_huge_pte_at(dst, addr, dst_pte, entry);
2789 } else {
34ee645e 2790 if (cow) {
7f2e9525 2791 huge_ptep_set_wrprotect(src, addr, src_pte);
34ee645e
JR
2792 mmu_notifier_invalidate_range(src, mmun_start,
2793 mmun_end);
2794 }
0253d634 2795 entry = huge_ptep_get(src_pte);
1c59827d
HD
2796 ptepage = pte_page(entry);
2797 get_page(ptepage);
0fe6e20b 2798 page_dup_rmap(ptepage);
1c59827d
HD
2799 set_huge_pte_at(dst, addr, dst_pte, entry);
2800 }
cb900f41
KS
2801 spin_unlock(src_ptl);
2802 spin_unlock(dst_ptl);
63551ae0 2803 }
63551ae0 2804
e8569dd2
AS
2805 if (cow)
2806 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2807
2808 return ret;
63551ae0
DG
2809}
2810
24669e58
AK
2811void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2812 unsigned long start, unsigned long end,
2813 struct page *ref_page)
63551ae0 2814{
24669e58 2815 int force_flush = 0;
63551ae0
DG
2816 struct mm_struct *mm = vma->vm_mm;
2817 unsigned long address;
c7546f8f 2818 pte_t *ptep;
63551ae0 2819 pte_t pte;
cb900f41 2820 spinlock_t *ptl;
63551ae0 2821 struct page *page;
a5516438
AK
2822 struct hstate *h = hstate_vma(vma);
2823 unsigned long sz = huge_page_size(h);
2ec74c3e
SG
2824 const unsigned long mmun_start = start; /* For mmu_notifiers */
2825 const unsigned long mmun_end = end; /* For mmu_notifiers */
a5516438 2826
63551ae0 2827 WARN_ON(!is_vm_hugetlb_page(vma));
a5516438
AK
2828 BUG_ON(start & ~huge_page_mask(h));
2829 BUG_ON(end & ~huge_page_mask(h));
63551ae0 2830
24669e58 2831 tlb_start_vma(tlb, vma);
2ec74c3e 2832 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
569f48b8 2833 address = start;
24669e58 2834again:
569f48b8 2835 for (; address < end; address += sz) {
c7546f8f 2836 ptep = huge_pte_offset(mm, address);
4c887265 2837 if (!ptep)
c7546f8f
DG
2838 continue;
2839
cb900f41 2840 ptl = huge_pte_lock(h, mm, ptep);
39dde65c 2841 if (huge_pmd_unshare(mm, &address, ptep))
cb900f41 2842 goto unlock;
39dde65c 2843
6629326b
HD
2844 pte = huge_ptep_get(ptep);
2845 if (huge_pte_none(pte))
cb900f41 2846 goto unlock;
6629326b
HD
2847
2848 /*
9fbc1f63
NH
2849 * Migrating hugepage or HWPoisoned hugepage is already
2850 * unmapped and its refcount is dropped, so just clear pte here.
6629326b 2851 */
9fbc1f63 2852 if (unlikely(!pte_present(pte))) {
106c992a 2853 huge_pte_clear(mm, address, ptep);
cb900f41 2854 goto unlock;
8c4894c6 2855 }
6629326b
HD
2856
2857 page = pte_page(pte);
04f2cbe3
MG
2858 /*
2859 * If a reference page is supplied, it is because a specific
2860 * page is being unmapped, not a range. Ensure the page we
2861 * are about to unmap is the actual page of interest.
2862 */
2863 if (ref_page) {
04f2cbe3 2864 if (page != ref_page)
cb900f41 2865 goto unlock;
04f2cbe3
MG
2866
2867 /*
2868 * Mark the VMA as having unmapped its page so that
2869 * future faults in this VMA will fail rather than
2870 * looking like data was lost
2871 */
2872 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2873 }
2874
c7546f8f 2875 pte = huge_ptep_get_and_clear(mm, address, ptep);
24669e58 2876 tlb_remove_tlb_entry(tlb, ptep, address);
106c992a 2877 if (huge_pte_dirty(pte))
6649a386 2878 set_page_dirty(page);
9e81130b 2879
24669e58
AK
2880 page_remove_rmap(page);
2881 force_flush = !__tlb_remove_page(tlb, page);
cb900f41 2882 if (force_flush) {
569f48b8 2883 address += sz;
cb900f41 2884 spin_unlock(ptl);
24669e58 2885 break;
cb900f41 2886 }
9e81130b 2887 /* Bail out after unmapping reference page if supplied */
cb900f41
KS
2888 if (ref_page) {
2889 spin_unlock(ptl);
9e81130b 2890 break;
cb900f41
KS
2891 }
2892unlock:
2893 spin_unlock(ptl);
63551ae0 2894 }
24669e58
AK
2895 /*
2896 * mmu_gather ran out of room to batch pages, we break out of
2897 * the PTE lock to avoid doing the potential expensive TLB invalidate
2898 * and page-free while holding it.
2899 */
2900 if (force_flush) {
2901 force_flush = 0;
2902 tlb_flush_mmu(tlb);
2903 if (address < end && !ref_page)
2904 goto again;
fe1668ae 2905 }
2ec74c3e 2906 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
24669e58 2907 tlb_end_vma(tlb, vma);
1da177e4 2908}
63551ae0 2909
d833352a
MG
2910void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2911 struct vm_area_struct *vma, unsigned long start,
2912 unsigned long end, struct page *ref_page)
2913{
2914 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2915
2916 /*
2917 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2918 * test will fail on a vma being torn down, and not grab a page table
2919 * on its way out. We're lucky that the flag has such an appropriate
2920 * name, and can in fact be safely cleared here. We could clear it
2921 * before the __unmap_hugepage_range above, but all that's necessary
c8c06efa 2922 * is to clear it before releasing the i_mmap_rwsem. This works
d833352a 2923 * because in the context this is called, the VMA is about to be
c8c06efa 2924 * destroyed and the i_mmap_rwsem is held.
d833352a
MG
2925 */
2926 vma->vm_flags &= ~VM_MAYSHARE;
2927}
2928
502717f4 2929void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
04f2cbe3 2930 unsigned long end, struct page *ref_page)
502717f4 2931{
24669e58
AK
2932 struct mm_struct *mm;
2933 struct mmu_gather tlb;
2934
2935 mm = vma->vm_mm;
2936
2b047252 2937 tlb_gather_mmu(&tlb, mm, start, end);
24669e58
AK
2938 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2939 tlb_finish_mmu(&tlb, start, end);
502717f4
KC
2940}
2941
04f2cbe3
MG
2942/*
2943 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2944 * mappping it owns the reserve page for. The intention is to unmap the page
2945 * from other VMAs and let the children be SIGKILLed if they are faulting the
2946 * same region.
2947 */
2f4612af
DB
2948static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2949 struct page *page, unsigned long address)
04f2cbe3 2950{
7526674d 2951 struct hstate *h = hstate_vma(vma);
04f2cbe3
MG
2952 struct vm_area_struct *iter_vma;
2953 struct address_space *mapping;
04f2cbe3
MG
2954 pgoff_t pgoff;
2955
2956 /*
2957 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2958 * from page cache lookup which is in HPAGE_SIZE units.
2959 */
7526674d 2960 address = address & huge_page_mask(h);
36e4f20a
MH
2961 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2962 vma->vm_pgoff;
496ad9aa 2963 mapping = file_inode(vma->vm_file)->i_mapping;
04f2cbe3 2964
4eb2b1dc
MG
2965 /*
2966 * Take the mapping lock for the duration of the table walk. As
2967 * this mapping should be shared between all the VMAs,
2968 * __unmap_hugepage_range() is called as the lock is already held
2969 */
83cde9e8 2970 i_mmap_lock_write(mapping);
6b2dbba8 2971 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
04f2cbe3
MG
2972 /* Do not unmap the current VMA */
2973 if (iter_vma == vma)
2974 continue;
2975
2976 /*
2977 * Unmap the page from other VMAs without their own reserves.
2978 * They get marked to be SIGKILLed if they fault in these
2979 * areas. This is because a future no-page fault on this VMA
2980 * could insert a zeroed page instead of the data existing
2981 * from the time of fork. This would look like data corruption
2982 */
2983 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
24669e58
AK
2984 unmap_hugepage_range(iter_vma, address,
2985 address + huge_page_size(h), page);
04f2cbe3 2986 }
83cde9e8 2987 i_mmap_unlock_write(mapping);
04f2cbe3
MG
2988}
2989
0fe6e20b
NH
2990/*
2991 * Hugetlb_cow() should be called with page lock of the original hugepage held.
ef009b25
MH
2992 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2993 * cannot race with other handlers or page migration.
2994 * Keep the pte_same checks anyway to make transition from the mutex easier.
0fe6e20b 2995 */
1e8f889b 2996static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
04f2cbe3 2997 unsigned long address, pte_t *ptep, pte_t pte,
cb900f41 2998 struct page *pagecache_page, spinlock_t *ptl)
1e8f889b 2999{
a5516438 3000 struct hstate *h = hstate_vma(vma);
1e8f889b 3001 struct page *old_page, *new_page;
ad4404a2 3002 int ret = 0, outside_reserve = 0;
2ec74c3e
SG
3003 unsigned long mmun_start; /* For mmu_notifiers */
3004 unsigned long mmun_end; /* For mmu_notifiers */
1e8f889b
DG
3005
3006 old_page = pte_page(pte);
3007
04f2cbe3 3008retry_avoidcopy:
1e8f889b
DG
3009 /* If no-one else is actually using this page, avoid the copy
3010 * and just make the page writable */
37a2140d
JK
3011 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3012 page_move_anon_rmap(old_page, vma, address);
1e8f889b 3013 set_huge_ptep_writable(vma, address, ptep);
83c54070 3014 return 0;
1e8f889b
DG
3015 }
3016
04f2cbe3
MG
3017 /*
3018 * If the process that created a MAP_PRIVATE mapping is about to
3019 * perform a COW due to a shared page count, attempt to satisfy
3020 * the allocation without using the existing reserves. The pagecache
3021 * page is used to determine if the reserve at this address was
3022 * consumed or not. If reserves were used, a partial faulted mapping
3023 * at the time of fork() could consume its reserves on COW instead
3024 * of the full address range.
3025 */
5944d011 3026 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
04f2cbe3
MG
3027 old_page != pagecache_page)
3028 outside_reserve = 1;
3029
1e8f889b 3030 page_cache_get(old_page);
b76c8cfb 3031
ad4404a2
DB
3032 /*
3033 * Drop page table lock as buddy allocator may be called. It will
3034 * be acquired again before returning to the caller, as expected.
3035 */
cb900f41 3036 spin_unlock(ptl);
04f2cbe3 3037 new_page = alloc_huge_page(vma, address, outside_reserve);
1e8f889b 3038
2fc39cec 3039 if (IS_ERR(new_page)) {
04f2cbe3
MG
3040 /*
3041 * If a process owning a MAP_PRIVATE mapping fails to COW,
3042 * it is due to references held by a child and an insufficient
3043 * huge page pool. To guarantee the original mappers
3044 * reliability, unmap the page from child processes. The child
3045 * may get SIGKILLed if it later faults.
3046 */
3047 if (outside_reserve) {
ad4404a2 3048 page_cache_release(old_page);
04f2cbe3 3049 BUG_ON(huge_pte_none(pte));
2f4612af
DB
3050 unmap_ref_private(mm, vma, old_page, address);
3051 BUG_ON(huge_pte_none(pte));
3052 spin_lock(ptl);
3053 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3054 if (likely(ptep &&
3055 pte_same(huge_ptep_get(ptep), pte)))
3056 goto retry_avoidcopy;
3057 /*
3058 * race occurs while re-acquiring page table
3059 * lock, and our job is done.
3060 */
3061 return 0;
04f2cbe3
MG
3062 }
3063
ad4404a2
DB
3064 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3065 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3066 goto out_release_old;
1e8f889b
DG
3067 }
3068
0fe6e20b
NH
3069 /*
3070 * When the original hugepage is shared one, it does not have
3071 * anon_vma prepared.
3072 */
44e2aa93 3073 if (unlikely(anon_vma_prepare(vma))) {
ad4404a2
DB
3074 ret = VM_FAULT_OOM;
3075 goto out_release_all;
44e2aa93 3076 }
0fe6e20b 3077
47ad8475
AA
3078 copy_user_huge_page(new_page, old_page, address, vma,
3079 pages_per_huge_page(h));
0ed361de 3080 __SetPageUptodate(new_page);
bcc54222 3081 set_page_huge_active(new_page);
1e8f889b 3082
2ec74c3e
SG
3083 mmun_start = address & huge_page_mask(h);
3084 mmun_end = mmun_start + huge_page_size(h);
3085 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
ad4404a2 3086
b76c8cfb 3087 /*
cb900f41 3088 * Retake the page table lock to check for racing updates
b76c8cfb
LW
3089 * before the page tables are altered
3090 */
cb900f41 3091 spin_lock(ptl);
a5516438 3092 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
a9af0c5d 3093 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
07443a85
JK
3094 ClearPagePrivate(new_page);
3095
1e8f889b 3096 /* Break COW */
8fe627ec 3097 huge_ptep_clear_flush(vma, address, ptep);
34ee645e 3098 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
1e8f889b
DG
3099 set_huge_pte_at(mm, address, ptep,
3100 make_huge_pte(vma, new_page, 1));
0fe6e20b 3101 page_remove_rmap(old_page);
cd67f0d2 3102 hugepage_add_new_anon_rmap(new_page, vma, address);
1e8f889b
DG
3103 /* Make the old page be freed below */
3104 new_page = old_page;
3105 }
cb900f41 3106 spin_unlock(ptl);
2ec74c3e 3107 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
ad4404a2 3108out_release_all:
1e8f889b 3109 page_cache_release(new_page);
ad4404a2 3110out_release_old:
1e8f889b 3111 page_cache_release(old_page);
8312034f 3112
ad4404a2
DB
3113 spin_lock(ptl); /* Caller expects lock to be held */
3114 return ret;
1e8f889b
DG
3115}
3116
04f2cbe3 3117/* Return the pagecache page at a given address within a VMA */
a5516438
AK
3118static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3119 struct vm_area_struct *vma, unsigned long address)
04f2cbe3
MG
3120{
3121 struct address_space *mapping;
e7c4b0bf 3122 pgoff_t idx;
04f2cbe3
MG
3123
3124 mapping = vma->vm_file->f_mapping;
a5516438 3125 idx = vma_hugecache_offset(h, vma, address);
04f2cbe3
MG
3126
3127 return find_lock_page(mapping, idx);
3128}
3129
3ae77f43
HD
3130/*
3131 * Return whether there is a pagecache page to back given address within VMA.
3132 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3133 */
3134static bool hugetlbfs_pagecache_present(struct hstate *h,
2a15efc9
HD
3135 struct vm_area_struct *vma, unsigned long address)
3136{
3137 struct address_space *mapping;
3138 pgoff_t idx;
3139 struct page *page;
3140
3141 mapping = vma->vm_file->f_mapping;
3142 idx = vma_hugecache_offset(h, vma, address);
3143
3144 page = find_get_page(mapping, idx);
3145 if (page)
3146 put_page(page);
3147 return page != NULL;
3148}
3149
a1ed3dda 3150static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
8382d914
DB
3151 struct address_space *mapping, pgoff_t idx,
3152 unsigned long address, pte_t *ptep, unsigned int flags)
ac9b9c66 3153{
a5516438 3154 struct hstate *h = hstate_vma(vma);
ac9b9c66 3155 int ret = VM_FAULT_SIGBUS;
409eb8c2 3156 int anon_rmap = 0;
4c887265 3157 unsigned long size;
4c887265 3158 struct page *page;
1e8f889b 3159 pte_t new_pte;
cb900f41 3160 spinlock_t *ptl;
4c887265 3161
04f2cbe3
MG
3162 /*
3163 * Currently, we are forced to kill the process in the event the
3164 * original mapper has unmapped pages from the child due to a failed
25985edc 3165 * COW. Warn that such a situation has occurred as it may not be obvious
04f2cbe3
MG
3166 */
3167 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
ffb22af5
AM
3168 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3169 current->pid);
04f2cbe3
MG
3170 return ret;
3171 }
3172
4c887265
AL
3173 /*
3174 * Use page lock to guard against racing truncation
3175 * before we get page_table_lock.
3176 */
6bda666a
CL
3177retry:
3178 page = find_lock_page(mapping, idx);
3179 if (!page) {
a5516438 3180 size = i_size_read(mapping->host) >> huge_page_shift(h);
ebed4bfc
HD
3181 if (idx >= size)
3182 goto out;
04f2cbe3 3183 page = alloc_huge_page(vma, address, 0);
2fc39cec 3184 if (IS_ERR(page)) {
76dcee75
AK
3185 ret = PTR_ERR(page);
3186 if (ret == -ENOMEM)
3187 ret = VM_FAULT_OOM;
3188 else
3189 ret = VM_FAULT_SIGBUS;
6bda666a
CL
3190 goto out;
3191 }
47ad8475 3192 clear_huge_page(page, address, pages_per_huge_page(h));
0ed361de 3193 __SetPageUptodate(page);
bcc54222 3194 set_page_huge_active(page);
ac9b9c66 3195
f83a275d 3196 if (vma->vm_flags & VM_MAYSHARE) {
6bda666a 3197 int err;
45c682a6 3198 struct inode *inode = mapping->host;
6bda666a
CL
3199
3200 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3201 if (err) {
3202 put_page(page);
6bda666a
CL
3203 if (err == -EEXIST)
3204 goto retry;
3205 goto out;
3206 }
07443a85 3207 ClearPagePrivate(page);
45c682a6
KC
3208
3209 spin_lock(&inode->i_lock);
a5516438 3210 inode->i_blocks += blocks_per_huge_page(h);
45c682a6 3211 spin_unlock(&inode->i_lock);
23be7468 3212 } else {
6bda666a 3213 lock_page(page);
0fe6e20b
NH
3214 if (unlikely(anon_vma_prepare(vma))) {
3215 ret = VM_FAULT_OOM;
3216 goto backout_unlocked;
3217 }
409eb8c2 3218 anon_rmap = 1;
23be7468 3219 }
0fe6e20b 3220 } else {
998b4382
NH
3221 /*
3222 * If memory error occurs between mmap() and fault, some process
3223 * don't have hwpoisoned swap entry for errored virtual address.
3224 * So we need to block hugepage fault by PG_hwpoison bit check.
3225 */
3226 if (unlikely(PageHWPoison(page))) {
32f84528 3227 ret = VM_FAULT_HWPOISON |
972dc4de 3228 VM_FAULT_SET_HINDEX(hstate_index(h));
998b4382
NH
3229 goto backout_unlocked;
3230 }
6bda666a 3231 }
1e8f889b 3232
57303d80
AW
3233 /*
3234 * If we are going to COW a private mapping later, we examine the
3235 * pending reservations for this page now. This will ensure that
3236 * any allocations necessary to record that reservation occur outside
3237 * the spinlock.
3238 */
788c7df4 3239 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2b26736c
AW
3240 if (vma_needs_reservation(h, vma, address) < 0) {
3241 ret = VM_FAULT_OOM;
3242 goto backout_unlocked;
3243 }
57303d80 3244
cb900f41
KS
3245 ptl = huge_pte_lockptr(h, mm, ptep);
3246 spin_lock(ptl);
a5516438 3247 size = i_size_read(mapping->host) >> huge_page_shift(h);
4c887265
AL
3248 if (idx >= size)
3249 goto backout;
3250
83c54070 3251 ret = 0;
7f2e9525 3252 if (!huge_pte_none(huge_ptep_get(ptep)))
4c887265
AL
3253 goto backout;
3254
07443a85
JK
3255 if (anon_rmap) {
3256 ClearPagePrivate(page);
409eb8c2 3257 hugepage_add_new_anon_rmap(page, vma, address);
ac714904 3258 } else
409eb8c2 3259 page_dup_rmap(page);
1e8f889b
DG
3260 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3261 && (vma->vm_flags & VM_SHARED)));
3262 set_huge_pte_at(mm, address, ptep, new_pte);
3263
788c7df4 3264 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
1e8f889b 3265 /* Optimization, do the COW without a second fault */
cb900f41 3266 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
1e8f889b
DG
3267 }
3268
cb900f41 3269 spin_unlock(ptl);
4c887265
AL
3270 unlock_page(page);
3271out:
ac9b9c66 3272 return ret;
4c887265
AL
3273
3274backout:
cb900f41 3275 spin_unlock(ptl);
2b26736c 3276backout_unlocked:
4c887265
AL
3277 unlock_page(page);
3278 put_page(page);
3279 goto out;
ac9b9c66
HD
3280}
3281
8382d914
DB
3282#ifdef CONFIG_SMP
3283static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3284 struct vm_area_struct *vma,
3285 struct address_space *mapping,
3286 pgoff_t idx, unsigned long address)
3287{
3288 unsigned long key[2];
3289 u32 hash;
3290
3291 if (vma->vm_flags & VM_SHARED) {
3292 key[0] = (unsigned long) mapping;
3293 key[1] = idx;
3294 } else {
3295 key[0] = (unsigned long) mm;
3296 key[1] = address >> huge_page_shift(h);
3297 }
3298
3299 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3300
3301 return hash & (num_fault_mutexes - 1);
3302}
3303#else
3304/*
3305 * For uniprocesor systems we always use a single mutex, so just
3306 * return 0 and avoid the hashing overhead.
3307 */
3308static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3309 struct vm_area_struct *vma,
3310 struct address_space *mapping,
3311 pgoff_t idx, unsigned long address)
3312{
3313 return 0;
3314}
3315#endif
3316
86e5216f 3317int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
788c7df4 3318 unsigned long address, unsigned int flags)
86e5216f 3319{
8382d914 3320 pte_t *ptep, entry;
cb900f41 3321 spinlock_t *ptl;
1e8f889b 3322 int ret;
8382d914
DB
3323 u32 hash;
3324 pgoff_t idx;
0fe6e20b 3325 struct page *page = NULL;
57303d80 3326 struct page *pagecache_page = NULL;
a5516438 3327 struct hstate *h = hstate_vma(vma);
8382d914 3328 struct address_space *mapping;
0f792cf9 3329 int need_wait_lock = 0;
86e5216f 3330
1e16a539
KH
3331 address &= huge_page_mask(h);
3332
fd6a03ed
NH
3333 ptep = huge_pte_offset(mm, address);
3334 if (ptep) {
3335 entry = huge_ptep_get(ptep);
290408d4 3336 if (unlikely(is_hugetlb_entry_migration(entry))) {
cb900f41 3337 migration_entry_wait_huge(vma, mm, ptep);
290408d4
NH
3338 return 0;
3339 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
32f84528 3340 return VM_FAULT_HWPOISON_LARGE |
972dc4de 3341 VM_FAULT_SET_HINDEX(hstate_index(h));
fd6a03ed
NH
3342 }
3343
a5516438 3344 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
86e5216f
AL
3345 if (!ptep)
3346 return VM_FAULT_OOM;
3347
8382d914
DB
3348 mapping = vma->vm_file->f_mapping;
3349 idx = vma_hugecache_offset(h, vma, address);
3350
3935baa9
DG
3351 /*
3352 * Serialize hugepage allocation and instantiation, so that we don't
3353 * get spurious allocation failures if two CPUs race to instantiate
3354 * the same page in the page cache.
3355 */
8382d914
DB
3356 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3357 mutex_lock(&htlb_fault_mutex_table[hash]);
3358
7f2e9525
GS
3359 entry = huge_ptep_get(ptep);
3360 if (huge_pte_none(entry)) {
8382d914 3361 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
b4d1d99f 3362 goto out_mutex;
3935baa9 3363 }
86e5216f 3364
83c54070 3365 ret = 0;
1e8f889b 3366
0f792cf9
NH
3367 /*
3368 * entry could be a migration/hwpoison entry at this point, so this
3369 * check prevents the kernel from going below assuming that we have
3370 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3371 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3372 * handle it.
3373 */
3374 if (!pte_present(entry))
3375 goto out_mutex;
3376
57303d80
AW
3377 /*
3378 * If we are going to COW the mapping later, we examine the pending
3379 * reservations for this page now. This will ensure that any
3380 * allocations necessary to record that reservation occur outside the
3381 * spinlock. For private mappings, we also lookup the pagecache
3382 * page now as it is used to determine if a reservation has been
3383 * consumed.
3384 */
106c992a 3385 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2b26736c
AW
3386 if (vma_needs_reservation(h, vma, address) < 0) {
3387 ret = VM_FAULT_OOM;
b4d1d99f 3388 goto out_mutex;
2b26736c 3389 }
57303d80 3390
f83a275d 3391 if (!(vma->vm_flags & VM_MAYSHARE))
57303d80
AW
3392 pagecache_page = hugetlbfs_pagecache_page(h,
3393 vma, address);
3394 }
3395
0f792cf9
NH
3396 ptl = huge_pte_lock(h, mm, ptep);
3397
3398 /* Check for a racing update before calling hugetlb_cow */
3399 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3400 goto out_ptl;
3401
56c9cfb1
NH
3402 /*
3403 * hugetlb_cow() requires page locks of pte_page(entry) and
3404 * pagecache_page, so here we need take the former one
3405 * when page != pagecache_page or !pagecache_page.
56c9cfb1
NH
3406 */
3407 page = pte_page(entry);
3408 if (page != pagecache_page)
0f792cf9
NH
3409 if (!trylock_page(page)) {
3410 need_wait_lock = 1;
3411 goto out_ptl;
3412 }
b4d1d99f 3413
0f792cf9 3414 get_page(page);
b4d1d99f 3415
788c7df4 3416 if (flags & FAULT_FLAG_WRITE) {
106c992a 3417 if (!huge_pte_write(entry)) {
57303d80 3418 ret = hugetlb_cow(mm, vma, address, ptep, entry,
cb900f41 3419 pagecache_page, ptl);
0f792cf9 3420 goto out_put_page;
b4d1d99f 3421 }
106c992a 3422 entry = huge_pte_mkdirty(entry);
b4d1d99f
DG
3423 }
3424 entry = pte_mkyoung(entry);
788c7df4
HD
3425 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3426 flags & FAULT_FLAG_WRITE))
4b3073e1 3427 update_mmu_cache(vma, address, ptep);
0f792cf9
NH
3428out_put_page:
3429 if (page != pagecache_page)
3430 unlock_page(page);
3431 put_page(page);
cb900f41
KS
3432out_ptl:
3433 spin_unlock(ptl);
57303d80
AW
3434
3435 if (pagecache_page) {
3436 unlock_page(pagecache_page);
3437 put_page(pagecache_page);
3438 }
b4d1d99f 3439out_mutex:
8382d914 3440 mutex_unlock(&htlb_fault_mutex_table[hash]);
0f792cf9
NH
3441 /*
3442 * Generally it's safe to hold refcount during waiting page lock. But
3443 * here we just wait to defer the next page fault to avoid busy loop and
3444 * the page is not used after unlocked before returning from the current
3445 * page fault. So we are safe from accessing freed page, even if we wait
3446 * here without taking refcount.
3447 */
3448 if (need_wait_lock)
3449 wait_on_page_locked(page);
1e8f889b 3450 return ret;
86e5216f
AL
3451}
3452
28a35716
ML
3453long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3454 struct page **pages, struct vm_area_struct **vmas,
3455 unsigned long *position, unsigned long *nr_pages,
3456 long i, unsigned int flags)
63551ae0 3457{
d5d4b0aa
KC
3458 unsigned long pfn_offset;
3459 unsigned long vaddr = *position;
28a35716 3460 unsigned long remainder = *nr_pages;
a5516438 3461 struct hstate *h = hstate_vma(vma);
63551ae0 3462
63551ae0 3463 while (vaddr < vma->vm_end && remainder) {
4c887265 3464 pte_t *pte;
cb900f41 3465 spinlock_t *ptl = NULL;
2a15efc9 3466 int absent;
4c887265 3467 struct page *page;
63551ae0 3468
02057967
DR
3469 /*
3470 * If we have a pending SIGKILL, don't keep faulting pages and
3471 * potentially allocating memory.
3472 */
3473 if (unlikely(fatal_signal_pending(current))) {
3474 remainder = 0;
3475 break;
3476 }
3477
4c887265
AL
3478 /*
3479 * Some archs (sparc64, sh*) have multiple pte_ts to
2a15efc9 3480 * each hugepage. We have to make sure we get the
4c887265 3481 * first, for the page indexing below to work.
cb900f41
KS
3482 *
3483 * Note that page table lock is not held when pte is null.
4c887265 3484 */
a5516438 3485 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
cb900f41
KS
3486 if (pte)
3487 ptl = huge_pte_lock(h, mm, pte);
2a15efc9
HD
3488 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3489
3490 /*
3491 * When coredumping, it suits get_dump_page if we just return
3ae77f43
HD
3492 * an error where there's an empty slot with no huge pagecache
3493 * to back it. This way, we avoid allocating a hugepage, and
3494 * the sparse dumpfile avoids allocating disk blocks, but its
3495 * huge holes still show up with zeroes where they need to be.
2a15efc9 3496 */
3ae77f43
HD
3497 if (absent && (flags & FOLL_DUMP) &&
3498 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
cb900f41
KS
3499 if (pte)
3500 spin_unlock(ptl);
2a15efc9
HD
3501 remainder = 0;
3502 break;
3503 }
63551ae0 3504
9cc3a5bd
NH
3505 /*
3506 * We need call hugetlb_fault for both hugepages under migration
3507 * (in which case hugetlb_fault waits for the migration,) and
3508 * hwpoisoned hugepages (in which case we need to prevent the
3509 * caller from accessing to them.) In order to do this, we use
3510 * here is_swap_pte instead of is_hugetlb_entry_migration and
3511 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3512 * both cases, and because we can't follow correct pages
3513 * directly from any kind of swap entries.
3514 */
3515 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
106c992a
GS
3516 ((flags & FOLL_WRITE) &&
3517 !huge_pte_write(huge_ptep_get(pte)))) {
4c887265 3518 int ret;
63551ae0 3519
cb900f41
KS
3520 if (pte)
3521 spin_unlock(ptl);
2a15efc9
HD
3522 ret = hugetlb_fault(mm, vma, vaddr,
3523 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
a89182c7 3524 if (!(ret & VM_FAULT_ERROR))
4c887265 3525 continue;
63551ae0 3526
4c887265 3527 remainder = 0;
4c887265
AL
3528 break;
3529 }
3530
a5516438 3531 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
7f2e9525 3532 page = pte_page(huge_ptep_get(pte));
d5d4b0aa 3533same_page:
d6692183 3534 if (pages) {
2a15efc9 3535 pages[i] = mem_map_offset(page, pfn_offset);
a0368d4e 3536 get_page_foll(pages[i]);
d6692183 3537 }
63551ae0
DG
3538
3539 if (vmas)
3540 vmas[i] = vma;
3541
3542 vaddr += PAGE_SIZE;
d5d4b0aa 3543 ++pfn_offset;
63551ae0
DG
3544 --remainder;
3545 ++i;
d5d4b0aa 3546 if (vaddr < vma->vm_end && remainder &&
a5516438 3547 pfn_offset < pages_per_huge_page(h)) {
d5d4b0aa
KC
3548 /*
3549 * We use pfn_offset to avoid touching the pageframes
3550 * of this compound page.
3551 */
3552 goto same_page;
3553 }
cb900f41 3554 spin_unlock(ptl);
63551ae0 3555 }
28a35716 3556 *nr_pages = remainder;
63551ae0
DG
3557 *position = vaddr;
3558
2a15efc9 3559 return i ? i : -EFAULT;
63551ae0 3560}
8f860591 3561
7da4d641 3562unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
8f860591
ZY
3563 unsigned long address, unsigned long end, pgprot_t newprot)
3564{
3565 struct mm_struct *mm = vma->vm_mm;
3566 unsigned long start = address;
3567 pte_t *ptep;
3568 pte_t pte;
a5516438 3569 struct hstate *h = hstate_vma(vma);
7da4d641 3570 unsigned long pages = 0;
8f860591
ZY
3571
3572 BUG_ON(address >= end);
3573 flush_cache_range(vma, address, end);
3574
a5338093 3575 mmu_notifier_invalidate_range_start(mm, start, end);
83cde9e8 3576 i_mmap_lock_write(vma->vm_file->f_mapping);
a5516438 3577 for (; address < end; address += huge_page_size(h)) {
cb900f41 3578 spinlock_t *ptl;
8f860591
ZY
3579 ptep = huge_pte_offset(mm, address);
3580 if (!ptep)
3581 continue;
cb900f41 3582 ptl = huge_pte_lock(h, mm, ptep);
7da4d641
PZ
3583 if (huge_pmd_unshare(mm, &address, ptep)) {
3584 pages++;
cb900f41 3585 spin_unlock(ptl);
39dde65c 3586 continue;
7da4d641 3587 }
a8bda28d
NH
3588 pte = huge_ptep_get(ptep);
3589 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3590 spin_unlock(ptl);
3591 continue;
3592 }
3593 if (unlikely(is_hugetlb_entry_migration(pte))) {
3594 swp_entry_t entry = pte_to_swp_entry(pte);
3595
3596 if (is_write_migration_entry(entry)) {
3597 pte_t newpte;
3598
3599 make_migration_entry_read(&entry);
3600 newpte = swp_entry_to_pte(entry);
3601 set_huge_pte_at(mm, address, ptep, newpte);
3602 pages++;
3603 }
3604 spin_unlock(ptl);
3605 continue;
3606 }
3607 if (!huge_pte_none(pte)) {
8f860591 3608 pte = huge_ptep_get_and_clear(mm, address, ptep);
106c992a 3609 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
be7517d6 3610 pte = arch_make_huge_pte(pte, vma, NULL, 0);
8f860591 3611 set_huge_pte_at(mm, address, ptep, pte);
7da4d641 3612 pages++;
8f860591 3613 }
cb900f41 3614 spin_unlock(ptl);
8f860591 3615 }
d833352a 3616 /*
c8c06efa 3617 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
d833352a 3618 * may have cleared our pud entry and done put_page on the page table:
c8c06efa 3619 * once we release i_mmap_rwsem, another task can do the final put_page
d833352a
MG
3620 * and that page table be reused and filled with junk.
3621 */
8f860591 3622 flush_tlb_range(vma, start, end);
34ee645e 3623 mmu_notifier_invalidate_range(mm, start, end);
83cde9e8 3624 i_mmap_unlock_write(vma->vm_file->f_mapping);
a5338093 3625 mmu_notifier_invalidate_range_end(mm, start, end);
7da4d641
PZ
3626
3627 return pages << h->order;
8f860591
ZY
3628}
3629
a1e78772
MG
3630int hugetlb_reserve_pages(struct inode *inode,
3631 long from, long to,
5a6fe125 3632 struct vm_area_struct *vma,
ca16d140 3633 vm_flags_t vm_flags)
e4e574b7 3634{
17c9d12e 3635 long ret, chg;
a5516438 3636 struct hstate *h = hstate_inode(inode);
90481622 3637 struct hugepage_subpool *spool = subpool_inode(inode);
9119a41e 3638 struct resv_map *resv_map;
1c5ecae3 3639 long gbl_reserve;
e4e574b7 3640
17c9d12e
MG
3641 /*
3642 * Only apply hugepage reservation if asked. At fault time, an
3643 * attempt will be made for VM_NORESERVE to allocate a page
90481622 3644 * without using reserves
17c9d12e 3645 */
ca16d140 3646 if (vm_flags & VM_NORESERVE)
17c9d12e
MG
3647 return 0;
3648
a1e78772
MG
3649 /*
3650 * Shared mappings base their reservation on the number of pages that
3651 * are already allocated on behalf of the file. Private mappings need
3652 * to reserve the full area even if read-only as mprotect() may be
3653 * called to make the mapping read-write. Assume !vma is a shm mapping
3654 */
9119a41e 3655 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4e35f483 3656 resv_map = inode_resv_map(inode);
9119a41e 3657
1406ec9b 3658 chg = region_chg(resv_map, from, to);
9119a41e
JK
3659
3660 } else {
3661 resv_map = resv_map_alloc();
17c9d12e
MG
3662 if (!resv_map)
3663 return -ENOMEM;
3664
a1e78772 3665 chg = to - from;
84afd99b 3666
17c9d12e
MG
3667 set_vma_resv_map(vma, resv_map);
3668 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3669 }
3670
c50ac050
DH
3671 if (chg < 0) {
3672 ret = chg;
3673 goto out_err;
3674 }
8a630112 3675
1c5ecae3
MK
3676 /*
3677 * There must be enough pages in the subpool for the mapping. If
3678 * the subpool has a minimum size, there may be some global
3679 * reservations already in place (gbl_reserve).
3680 */
3681 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
3682 if (gbl_reserve < 0) {
c50ac050
DH
3683 ret = -ENOSPC;
3684 goto out_err;
3685 }
5a6fe125
MG
3686
3687 /*
17c9d12e 3688 * Check enough hugepages are available for the reservation.
90481622 3689 * Hand the pages back to the subpool if there are not
5a6fe125 3690 */
1c5ecae3 3691 ret = hugetlb_acct_memory(h, gbl_reserve);
68842c9b 3692 if (ret < 0) {
1c5ecae3
MK
3693 /* put back original number of pages, chg */
3694 (void)hugepage_subpool_put_pages(spool, chg);
c50ac050 3695 goto out_err;
68842c9b 3696 }
17c9d12e
MG
3697
3698 /*
3699 * Account for the reservations made. Shared mappings record regions
3700 * that have reservations as they are shared by multiple VMAs.
3701 * When the last VMA disappears, the region map says how much
3702 * the reservation was and the page cache tells how much of
3703 * the reservation was consumed. Private mappings are per-VMA and
3704 * only the consumed reservations are tracked. When the VMA
3705 * disappears, the original reservation is the VMA size and the
3706 * consumed reservations are stored in the map. Hence, nothing
3707 * else has to be done for private mappings here
3708 */
33039678
MK
3709 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3710 long add = region_add(resv_map, from, to);
3711
3712 if (unlikely(chg > add)) {
3713 /*
3714 * pages in this range were added to the reserve
3715 * map between region_chg and region_add. This
3716 * indicates a race with alloc_huge_page. Adjust
3717 * the subpool and reserve counts modified above
3718 * based on the difference.
3719 */
3720 long rsv_adjust;
3721
3722 rsv_adjust = hugepage_subpool_put_pages(spool,
3723 chg - add);
3724 hugetlb_acct_memory(h, -rsv_adjust);
3725 }
3726 }
a43a8c39 3727 return 0;
c50ac050 3728out_err:
f031dd27
JK
3729 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3730 kref_put(&resv_map->refs, resv_map_release);
c50ac050 3731 return ret;
a43a8c39
KC
3732}
3733
3734void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3735{
a5516438 3736 struct hstate *h = hstate_inode(inode);
4e35f483 3737 struct resv_map *resv_map = inode_resv_map(inode);
9119a41e 3738 long chg = 0;
90481622 3739 struct hugepage_subpool *spool = subpool_inode(inode);
1c5ecae3 3740 long gbl_reserve;
45c682a6 3741
9119a41e 3742 if (resv_map)
1406ec9b 3743 chg = region_truncate(resv_map, offset);
45c682a6 3744 spin_lock(&inode->i_lock);
e4c6f8be 3745 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
45c682a6
KC
3746 spin_unlock(&inode->i_lock);
3747
1c5ecae3
MK
3748 /*
3749 * If the subpool has a minimum size, the number of global
3750 * reservations to be released may be adjusted.
3751 */
3752 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
3753 hugetlb_acct_memory(h, -gbl_reserve);
a43a8c39 3754}
93f70f90 3755
3212b535
SC
3756#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3757static unsigned long page_table_shareable(struct vm_area_struct *svma,
3758 struct vm_area_struct *vma,
3759 unsigned long addr, pgoff_t idx)
3760{
3761 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3762 svma->vm_start;
3763 unsigned long sbase = saddr & PUD_MASK;
3764 unsigned long s_end = sbase + PUD_SIZE;
3765
3766 /* Allow segments to share if only one is marked locked */
3767 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3768 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3769
3770 /*
3771 * match the virtual addresses, permission and the alignment of the
3772 * page table page.
3773 */
3774 if (pmd_index(addr) != pmd_index(saddr) ||
3775 vm_flags != svm_flags ||
3776 sbase < svma->vm_start || svma->vm_end < s_end)
3777 return 0;
3778
3779 return saddr;
3780}
3781
3782static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3783{
3784 unsigned long base = addr & PUD_MASK;
3785 unsigned long end = base + PUD_SIZE;
3786
3787 /*
3788 * check on proper vm_flags and page table alignment
3789 */
3790 if (vma->vm_flags & VM_MAYSHARE &&
3791 vma->vm_start <= base && end <= vma->vm_end)
3792 return 1;
3793 return 0;
3794}
3795
3796/*
3797 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3798 * and returns the corresponding pte. While this is not necessary for the
3799 * !shared pmd case because we can allocate the pmd later as well, it makes the
3800 * code much cleaner. pmd allocation is essential for the shared case because
c8c06efa 3801 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
3212b535
SC
3802 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3803 * bad pmd for sharing.
3804 */
3805pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3806{
3807 struct vm_area_struct *vma = find_vma(mm, addr);
3808 struct address_space *mapping = vma->vm_file->f_mapping;
3809 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3810 vma->vm_pgoff;
3811 struct vm_area_struct *svma;
3812 unsigned long saddr;
3813 pte_t *spte = NULL;
3814 pte_t *pte;
cb900f41 3815 spinlock_t *ptl;
3212b535
SC
3816
3817 if (!vma_shareable(vma, addr))
3818 return (pte_t *)pmd_alloc(mm, pud, addr);
3819
83cde9e8 3820 i_mmap_lock_write(mapping);
3212b535
SC
3821 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3822 if (svma == vma)
3823 continue;
3824
3825 saddr = page_table_shareable(svma, vma, addr, idx);
3826 if (saddr) {
3827 spte = huge_pte_offset(svma->vm_mm, saddr);
3828 if (spte) {
dc6c9a35 3829 mm_inc_nr_pmds(mm);
3212b535
SC
3830 get_page(virt_to_page(spte));
3831 break;
3832 }
3833 }
3834 }
3835
3836 if (!spte)
3837 goto out;
3838
cb900f41
KS
3839 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3840 spin_lock(ptl);
dc6c9a35 3841 if (pud_none(*pud)) {
3212b535
SC
3842 pud_populate(mm, pud,
3843 (pmd_t *)((unsigned long)spte & PAGE_MASK));
dc6c9a35 3844 } else {
3212b535 3845 put_page(virt_to_page(spte));
dc6c9a35
KS
3846 mm_inc_nr_pmds(mm);
3847 }
cb900f41 3848 spin_unlock(ptl);
3212b535
SC
3849out:
3850 pte = (pte_t *)pmd_alloc(mm, pud, addr);
83cde9e8 3851 i_mmap_unlock_write(mapping);
3212b535
SC
3852 return pte;
3853}
3854
3855/*
3856 * unmap huge page backed by shared pte.
3857 *
3858 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3859 * indicated by page_count > 1, unmap is achieved by clearing pud and
3860 * decrementing the ref count. If count == 1, the pte page is not shared.
3861 *
cb900f41 3862 * called with page table lock held.
3212b535
SC
3863 *
3864 * returns: 1 successfully unmapped a shared pte page
3865 * 0 the underlying pte page is not shared, or it is the last user
3866 */
3867int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3868{
3869 pgd_t *pgd = pgd_offset(mm, *addr);
3870 pud_t *pud = pud_offset(pgd, *addr);
3871
3872 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3873 if (page_count(virt_to_page(ptep)) == 1)
3874 return 0;
3875
3876 pud_clear(pud);
3877 put_page(virt_to_page(ptep));
dc6c9a35 3878 mm_dec_nr_pmds(mm);
3212b535
SC
3879 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3880 return 1;
3881}
9e5fc74c
SC
3882#define want_pmd_share() (1)
3883#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3884pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3885{
3886 return NULL;
3887}
e81f2d22
ZZ
3888
3889int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3890{
3891 return 0;
3892}
9e5fc74c 3893#define want_pmd_share() (0)
3212b535
SC
3894#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3895
9e5fc74c
SC
3896#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3897pte_t *huge_pte_alloc(struct mm_struct *mm,
3898 unsigned long addr, unsigned long sz)
3899{
3900 pgd_t *pgd;
3901 pud_t *pud;
3902 pte_t *pte = NULL;
3903
3904 pgd = pgd_offset(mm, addr);
3905 pud = pud_alloc(mm, pgd, addr);
3906 if (pud) {
3907 if (sz == PUD_SIZE) {
3908 pte = (pte_t *)pud;
3909 } else {
3910 BUG_ON(sz != PMD_SIZE);
3911 if (want_pmd_share() && pud_none(*pud))
3912 pte = huge_pmd_share(mm, addr, pud);
3913 else
3914 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3915 }
3916 }
3917 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3918
3919 return pte;
3920}
3921
3922pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3923{
3924 pgd_t *pgd;
3925 pud_t *pud;
3926 pmd_t *pmd = NULL;
3927
3928 pgd = pgd_offset(mm, addr);
3929 if (pgd_present(*pgd)) {
3930 pud = pud_offset(pgd, addr);
3931 if (pud_present(*pud)) {
3932 if (pud_huge(*pud))
3933 return (pte_t *)pud;
3934 pmd = pmd_offset(pud, addr);
3935 }
3936 }
3937 return (pte_t *) pmd;
3938}
3939
61f77eda
NH
3940#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3941
3942/*
3943 * These functions are overwritable if your architecture needs its own
3944 * behavior.
3945 */
3946struct page * __weak
3947follow_huge_addr(struct mm_struct *mm, unsigned long address,
3948 int write)
3949{
3950 return ERR_PTR(-EINVAL);
3951}
3952
3953struct page * __weak
9e5fc74c 3954follow_huge_pmd(struct mm_struct *mm, unsigned long address,
e66f17ff 3955 pmd_t *pmd, int flags)
9e5fc74c 3956{
e66f17ff
NH
3957 struct page *page = NULL;
3958 spinlock_t *ptl;
3959retry:
3960 ptl = pmd_lockptr(mm, pmd);
3961 spin_lock(ptl);
3962 /*
3963 * make sure that the address range covered by this pmd is not
3964 * unmapped from other threads.
3965 */
3966 if (!pmd_huge(*pmd))
3967 goto out;
3968 if (pmd_present(*pmd)) {
97534127 3969 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
e66f17ff
NH
3970 if (flags & FOLL_GET)
3971 get_page(page);
3972 } else {
3973 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
3974 spin_unlock(ptl);
3975 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
3976 goto retry;
3977 }
3978 /*
3979 * hwpoisoned entry is treated as no_page_table in
3980 * follow_page_mask().
3981 */
3982 }
3983out:
3984 spin_unlock(ptl);
9e5fc74c
SC
3985 return page;
3986}
3987
61f77eda 3988struct page * __weak
9e5fc74c 3989follow_huge_pud(struct mm_struct *mm, unsigned long address,
e66f17ff 3990 pud_t *pud, int flags)
9e5fc74c 3991{
e66f17ff
NH
3992 if (flags & FOLL_GET)
3993 return NULL;
9e5fc74c 3994
e66f17ff 3995 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
9e5fc74c
SC
3996}
3997
d5bd9106
AK
3998#ifdef CONFIG_MEMORY_FAILURE
3999
93f70f90
NH
4000/*
4001 * This function is called from memory failure code.
4002 * Assume the caller holds page lock of the head page.
4003 */
6de2b1aa 4004int dequeue_hwpoisoned_huge_page(struct page *hpage)
93f70f90
NH
4005{
4006 struct hstate *h = page_hstate(hpage);
4007 int nid = page_to_nid(hpage);
6de2b1aa 4008 int ret = -EBUSY;
93f70f90
NH
4009
4010 spin_lock(&hugetlb_lock);
7e1f049e
NH
4011 /*
4012 * Just checking !page_huge_active is not enough, because that could be
4013 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4014 */
4015 if (!page_huge_active(hpage) && !page_count(hpage)) {
56f2fb14
NH
4016 /*
4017 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4018 * but dangling hpage->lru can trigger list-debug warnings
4019 * (this happens when we call unpoison_memory() on it),
4020 * so let it point to itself with list_del_init().
4021 */
4022 list_del_init(&hpage->lru);
8c6c2ecb 4023 set_page_refcounted(hpage);
6de2b1aa
NH
4024 h->free_huge_pages--;
4025 h->free_huge_pages_node[nid]--;
4026 ret = 0;
4027 }
93f70f90 4028 spin_unlock(&hugetlb_lock);
6de2b1aa 4029 return ret;
93f70f90 4030}
6de2b1aa 4031#endif
31caf665
NH
4032
4033bool isolate_huge_page(struct page *page, struct list_head *list)
4034{
bcc54222
NH
4035 bool ret = true;
4036
309381fe 4037 VM_BUG_ON_PAGE(!PageHead(page), page);
31caf665 4038 spin_lock(&hugetlb_lock);
bcc54222
NH
4039 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4040 ret = false;
4041 goto unlock;
4042 }
4043 clear_page_huge_active(page);
31caf665 4044 list_move_tail(&page->lru, list);
bcc54222 4045unlock:
31caf665 4046 spin_unlock(&hugetlb_lock);
bcc54222 4047 return ret;
31caf665
NH
4048}
4049
4050void putback_active_hugepage(struct page *page)
4051{
309381fe 4052 VM_BUG_ON_PAGE(!PageHead(page), page);
31caf665 4053 spin_lock(&hugetlb_lock);
bcc54222 4054 set_page_huge_active(page);
31caf665
NH
4055 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4056 spin_unlock(&hugetlb_lock);
4057 put_page(page);
4058}