Loading...
1/*
2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
4 */
5#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/mm.h>
8#include <linux/seq_file.h>
9#include <linux/sysctl.h>
10#include <linux/highmem.h>
11#include <linux/mmu_notifier.h>
12#include <linux/nodemask.h>
13#include <linux/pagemap.h>
14#include <linux/mempolicy.h>
15#include <linux/compiler.h>
16#include <linux/cpuset.h>
17#include <linux/mutex.h>
18#include <linux/bootmem.h>
19#include <linux/sysfs.h>
20#include <linux/slab.h>
21#include <linux/mmdebug.h>
22#include <linux/sched/signal.h>
23#include <linux/rmap.h>
24#include <linux/string_helpers.h>
25#include <linux/swap.h>
26#include <linux/swapops.h>
27#include <linux/jhash.h>
28
29#include <asm/page.h>
30#include <asm/pgtable.h>
31#include <asm/tlb.h>
32
33#include <linux/io.h>
34#include <linux/hugetlb.h>
35#include <linux/hugetlb_cgroup.h>
36#include <linux/node.h>
37#include <linux/userfaultfd_k.h>
38#include <linux/page_owner.h>
39#include "internal.h"
40
41int hugetlb_max_hstate __read_mostly;
42unsigned int default_hstate_idx;
43struct hstate hstates[HUGE_MAX_HSTATE];
44/*
45 * Minimum page order among possible hugepage sizes, set to a proper value
46 * at boot time.
47 */
48static unsigned int minimum_order __read_mostly = UINT_MAX;
49
50__initdata LIST_HEAD(huge_boot_pages);
51
52/* for command line parsing */
53static struct hstate * __initdata parsed_hstate;
54static unsigned long __initdata default_hstate_max_huge_pages;
55static unsigned long __initdata default_hstate_size;
56static bool __initdata parsed_valid_hugepagesz = true;
57
58/*
59 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
60 * free_huge_pages, and surplus_huge_pages.
61 */
62DEFINE_SPINLOCK(hugetlb_lock);
63
64/*
65 * Serializes faults on the same logical page. This is used to
66 * prevent spurious OOMs when the hugepage pool is fully utilized.
67 */
68static int num_fault_mutexes;
69struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
70
71/* Forward declaration */
72static int hugetlb_acct_memory(struct hstate *h, long delta);
73
74static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
75{
76 bool free = (spool->count == 0) && (spool->used_hpages == 0);
77
78 spin_unlock(&spool->lock);
79
80 /* If no pages are used, and no other handles to the subpool
81 * remain, give up any reservations mased on minimum size and
82 * free the subpool */
83 if (free) {
84 if (spool->min_hpages != -1)
85 hugetlb_acct_memory(spool->hstate,
86 -spool->min_hpages);
87 kfree(spool);
88 }
89}
90
91struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
92 long min_hpages)
93{
94 struct hugepage_subpool *spool;
95
96 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
97 if (!spool)
98 return NULL;
99
100 spin_lock_init(&spool->lock);
101 spool->count = 1;
102 spool->max_hpages = max_hpages;
103 spool->hstate = h;
104 spool->min_hpages = min_hpages;
105
106 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
107 kfree(spool);
108 return NULL;
109 }
110 spool->rsv_hpages = min_hpages;
111
112 return spool;
113}
114
115void hugepage_put_subpool(struct hugepage_subpool *spool)
116{
117 spin_lock(&spool->lock);
118 BUG_ON(!spool->count);
119 spool->count--;
120 unlock_or_release_subpool(spool);
121}
122
123/*
124 * Subpool accounting for allocating and reserving pages.
125 * Return -ENOMEM if there are not enough resources to satisfy the
126 * the request. Otherwise, return the number of pages by which the
127 * global pools must be adjusted (upward). The returned value may
128 * only be different than the passed value (delta) in the case where
129 * a subpool minimum size must be manitained.
130 */
131static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
132 long delta)
133{
134 long ret = delta;
135
136 if (!spool)
137 return ret;
138
139 spin_lock(&spool->lock);
140
141 if (spool->max_hpages != -1) { /* maximum size accounting */
142 if ((spool->used_hpages + delta) <= spool->max_hpages)
143 spool->used_hpages += delta;
144 else {
145 ret = -ENOMEM;
146 goto unlock_ret;
147 }
148 }
149
150 /* minimum size accounting */
151 if (spool->min_hpages != -1 && spool->rsv_hpages) {
152 if (delta > spool->rsv_hpages) {
153 /*
154 * Asking for more reserves than those already taken on
155 * behalf of subpool. Return difference.
156 */
157 ret = delta - spool->rsv_hpages;
158 spool->rsv_hpages = 0;
159 } else {
160 ret = 0; /* reserves already accounted for */
161 spool->rsv_hpages -= delta;
162 }
163 }
164
165unlock_ret:
166 spin_unlock(&spool->lock);
167 return ret;
168}
169
170/*
171 * Subpool accounting for freeing and unreserving pages.
172 * Return the number of global page reservations that must be dropped.
173 * The return value may only be different than the passed value (delta)
174 * in the case where a subpool minimum size must be maintained.
175 */
176static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
177 long delta)
178{
179 long ret = delta;
180
181 if (!spool)
182 return delta;
183
184 spin_lock(&spool->lock);
185
186 if (spool->max_hpages != -1) /* maximum size accounting */
187 spool->used_hpages -= delta;
188
189 /* minimum size accounting */
190 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
191 if (spool->rsv_hpages + delta <= spool->min_hpages)
192 ret = 0;
193 else
194 ret = spool->rsv_hpages + delta - spool->min_hpages;
195
196 spool->rsv_hpages += delta;
197 if (spool->rsv_hpages > spool->min_hpages)
198 spool->rsv_hpages = spool->min_hpages;
199 }
200
201 /*
202 * If hugetlbfs_put_super couldn't free spool due to an outstanding
203 * quota reference, free it now.
204 */
205 unlock_or_release_subpool(spool);
206
207 return ret;
208}
209
210static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
211{
212 return HUGETLBFS_SB(inode->i_sb)->spool;
213}
214
215static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
216{
217 return subpool_inode(file_inode(vma->vm_file));
218}
219
220/*
221 * Region tracking -- allows tracking of reservations and instantiated pages
222 * across the pages in a mapping.
223 *
224 * The region data structures are embedded into a resv_map and protected
225 * by a resv_map's lock. The set of regions within the resv_map represent
226 * reservations for huge pages, or huge pages that have already been
227 * instantiated within the map. The from and to elements are huge page
228 * indicies into the associated mapping. from indicates the starting index
229 * of the region. to represents the first index past the end of the region.
230 *
231 * For example, a file region structure with from == 0 and to == 4 represents
232 * four huge pages in a mapping. It is important to note that the to element
233 * represents the first element past the end of the region. This is used in
234 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
235 *
236 * Interval notation of the form [from, to) will be used to indicate that
237 * the endpoint from is inclusive and to is exclusive.
238 */
239struct file_region {
240 struct list_head link;
241 long from;
242 long to;
243};
244
245/*
246 * Add the huge page range represented by [f, t) to the reserve
247 * map. In the normal case, existing regions will be expanded
248 * to accommodate the specified range. Sufficient regions should
249 * exist for expansion due to the previous call to region_chg
250 * with the same range. However, it is possible that region_del
251 * could have been called after region_chg and modifed the map
252 * in such a way that no region exists to be expanded. In this
253 * case, pull a region descriptor from the cache associated with
254 * the map and use that for the new range.
255 *
256 * Return the number of new huge pages added to the map. This
257 * number is greater than or equal to zero.
258 */
259static long region_add(struct resv_map *resv, long f, long t)
260{
261 struct list_head *head = &resv->regions;
262 struct file_region *rg, *nrg, *trg;
263 long add = 0;
264
265 spin_lock(&resv->lock);
266 /* Locate the region we are either in or before. */
267 list_for_each_entry(rg, head, link)
268 if (f <= rg->to)
269 break;
270
271 /*
272 * If no region exists which can be expanded to include the
273 * specified range, the list must have been modified by an
274 * interleving call to region_del(). Pull a region descriptor
275 * from the cache and use it for this range.
276 */
277 if (&rg->link == head || t < rg->from) {
278 VM_BUG_ON(resv->region_cache_count <= 0);
279
280 resv->region_cache_count--;
281 nrg = list_first_entry(&resv->region_cache, struct file_region,
282 link);
283 list_del(&nrg->link);
284
285 nrg->from = f;
286 nrg->to = t;
287 list_add(&nrg->link, rg->link.prev);
288
289 add += t - f;
290 goto out_locked;
291 }
292
293 /* Round our left edge to the current segment if it encloses us. */
294 if (f > rg->from)
295 f = rg->from;
296
297 /* Check for and consume any regions we now overlap with. */
298 nrg = rg;
299 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
300 if (&rg->link == head)
301 break;
302 if (rg->from > t)
303 break;
304
305 /* If this area reaches higher then extend our area to
306 * include it completely. If this is not the first area
307 * which we intend to reuse, free it. */
308 if (rg->to > t)
309 t = rg->to;
310 if (rg != nrg) {
311 /* Decrement return value by the deleted range.
312 * Another range will span this area so that by
313 * end of routine add will be >= zero
314 */
315 add -= (rg->to - rg->from);
316 list_del(&rg->link);
317 kfree(rg);
318 }
319 }
320
321 add += (nrg->from - f); /* Added to beginning of region */
322 nrg->from = f;
323 add += t - nrg->to; /* Added to end of region */
324 nrg->to = t;
325
326out_locked:
327 resv->adds_in_progress--;
328 spin_unlock(&resv->lock);
329 VM_BUG_ON(add < 0);
330 return add;
331}
332
333/*
334 * Examine the existing reserve map and determine how many
335 * huge pages in the specified range [f, t) are NOT currently
336 * represented. This routine is called before a subsequent
337 * call to region_add that will actually modify the reserve
338 * map to add the specified range [f, t). region_chg does
339 * not change the number of huge pages represented by the
340 * map. However, if the existing regions in the map can not
341 * be expanded to represent the new range, a new file_region
342 * structure is added to the map as a placeholder. This is
343 * so that the subsequent region_add call will have all the
344 * regions it needs and will not fail.
345 *
346 * Upon entry, region_chg will also examine the cache of region descriptors
347 * associated with the map. If there are not enough descriptors cached, one
348 * will be allocated for the in progress add operation.
349 *
350 * Returns the number of huge pages that need to be added to the existing
351 * reservation map for the range [f, t). This number is greater or equal to
352 * zero. -ENOMEM is returned if a new file_region structure or cache entry
353 * is needed and can not be allocated.
354 */
355static long region_chg(struct resv_map *resv, long f, long t)
356{
357 struct list_head *head = &resv->regions;
358 struct file_region *rg, *nrg = NULL;
359 long chg = 0;
360
361retry:
362 spin_lock(&resv->lock);
363retry_locked:
364 resv->adds_in_progress++;
365
366 /*
367 * Check for sufficient descriptors in the cache to accommodate
368 * the number of in progress add operations.
369 */
370 if (resv->adds_in_progress > resv->region_cache_count) {
371 struct file_region *trg;
372
373 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
374 /* Must drop lock to allocate a new descriptor. */
375 resv->adds_in_progress--;
376 spin_unlock(&resv->lock);
377
378 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
379 if (!trg) {
380 kfree(nrg);
381 return -ENOMEM;
382 }
383
384 spin_lock(&resv->lock);
385 list_add(&trg->link, &resv->region_cache);
386 resv->region_cache_count++;
387 goto retry_locked;
388 }
389
390 /* Locate the region we are before or in. */
391 list_for_each_entry(rg, head, link)
392 if (f <= rg->to)
393 break;
394
395 /* If we are below the current region then a new region is required.
396 * Subtle, allocate a new region at the position but make it zero
397 * size such that we can guarantee to record the reservation. */
398 if (&rg->link == head || t < rg->from) {
399 if (!nrg) {
400 resv->adds_in_progress--;
401 spin_unlock(&resv->lock);
402 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
403 if (!nrg)
404 return -ENOMEM;
405
406 nrg->from = f;
407 nrg->to = f;
408 INIT_LIST_HEAD(&nrg->link);
409 goto retry;
410 }
411
412 list_add(&nrg->link, rg->link.prev);
413 chg = t - f;
414 goto out_nrg;
415 }
416
417 /* Round our left edge to the current segment if it encloses us. */
418 if (f > rg->from)
419 f = rg->from;
420 chg = t - f;
421
422 /* Check for and consume any regions we now overlap with. */
423 list_for_each_entry(rg, rg->link.prev, link) {
424 if (&rg->link == head)
425 break;
426 if (rg->from > t)
427 goto out;
428
429 /* We overlap with this area, if it extends further than
430 * us then we must extend ourselves. Account for its
431 * existing reservation. */
432 if (rg->to > t) {
433 chg += rg->to - t;
434 t = rg->to;
435 }
436 chg -= rg->to - rg->from;
437 }
438
439out:
440 spin_unlock(&resv->lock);
441 /* We already know we raced and no longer need the new region */
442 kfree(nrg);
443 return chg;
444out_nrg:
445 spin_unlock(&resv->lock);
446 return chg;
447}
448
449/*
450 * Abort the in progress add operation. The adds_in_progress field
451 * of the resv_map keeps track of the operations in progress between
452 * calls to region_chg and region_add. Operations are sometimes
453 * aborted after the call to region_chg. In such cases, region_abort
454 * is called to decrement the adds_in_progress counter.
455 *
456 * NOTE: The range arguments [f, t) are not needed or used in this
457 * routine. They are kept to make reading the calling code easier as
458 * arguments will match the associated region_chg call.
459 */
460static void region_abort(struct resv_map *resv, long f, long t)
461{
462 spin_lock(&resv->lock);
463 VM_BUG_ON(!resv->region_cache_count);
464 resv->adds_in_progress--;
465 spin_unlock(&resv->lock);
466}
467
468/*
469 * Delete the specified range [f, t) from the reserve map. If the
470 * t parameter is LONG_MAX, this indicates that ALL regions after f
471 * should be deleted. Locate the regions which intersect [f, t)
472 * and either trim, delete or split the existing regions.
473 *
474 * Returns the number of huge pages deleted from the reserve map.
475 * In the normal case, the return value is zero or more. In the
476 * case where a region must be split, a new region descriptor must
477 * be allocated. If the allocation fails, -ENOMEM will be returned.
478 * NOTE: If the parameter t == LONG_MAX, then we will never split
479 * a region and possibly return -ENOMEM. Callers specifying
480 * t == LONG_MAX do not need to check for -ENOMEM error.
481 */
482static long region_del(struct resv_map *resv, long f, long t)
483{
484 struct list_head *head = &resv->regions;
485 struct file_region *rg, *trg;
486 struct file_region *nrg = NULL;
487 long del = 0;
488
489retry:
490 spin_lock(&resv->lock);
491 list_for_each_entry_safe(rg, trg, head, link) {
492 /*
493 * Skip regions before the range to be deleted. file_region
494 * ranges are normally of the form [from, to). However, there
495 * may be a "placeholder" entry in the map which is of the form
496 * (from, to) with from == to. Check for placeholder entries
497 * at the beginning of the range to be deleted.
498 */
499 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
500 continue;
501
502 if (rg->from >= t)
503 break;
504
505 if (f > rg->from && t < rg->to) { /* Must split region */
506 /*
507 * Check for an entry in the cache before dropping
508 * lock and attempting allocation.
509 */
510 if (!nrg &&
511 resv->region_cache_count > resv->adds_in_progress) {
512 nrg = list_first_entry(&resv->region_cache,
513 struct file_region,
514 link);
515 list_del(&nrg->link);
516 resv->region_cache_count--;
517 }
518
519 if (!nrg) {
520 spin_unlock(&resv->lock);
521 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
522 if (!nrg)
523 return -ENOMEM;
524 goto retry;
525 }
526
527 del += t - f;
528
529 /* New entry for end of split region */
530 nrg->from = t;
531 nrg->to = rg->to;
532 INIT_LIST_HEAD(&nrg->link);
533
534 /* Original entry is trimmed */
535 rg->to = f;
536
537 list_add(&nrg->link, &rg->link);
538 nrg = NULL;
539 break;
540 }
541
542 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
543 del += rg->to - rg->from;
544 list_del(&rg->link);
545 kfree(rg);
546 continue;
547 }
548
549 if (f <= rg->from) { /* Trim beginning of region */
550 del += t - rg->from;
551 rg->from = t;
552 } else { /* Trim end of region */
553 del += rg->to - f;
554 rg->to = f;
555 }
556 }
557
558 spin_unlock(&resv->lock);
559 kfree(nrg);
560 return del;
561}
562
563/*
564 * A rare out of memory error was encountered which prevented removal of
565 * the reserve map region for a page. The huge page itself was free'ed
566 * and removed from the page cache. This routine will adjust the subpool
567 * usage count, and the global reserve count if needed. By incrementing
568 * these counts, the reserve map entry which could not be deleted will
569 * appear as a "reserved" entry instead of simply dangling with incorrect
570 * counts.
571 */
572void hugetlb_fix_reserve_counts(struct inode *inode)
573{
574 struct hugepage_subpool *spool = subpool_inode(inode);
575 long rsv_adjust;
576
577 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
578 if (rsv_adjust) {
579 struct hstate *h = hstate_inode(inode);
580
581 hugetlb_acct_memory(h, 1);
582 }
583}
584
585/*
586 * Count and return the number of huge pages in the reserve map
587 * that intersect with the range [f, t).
588 */
589static long region_count(struct resv_map *resv, long f, long t)
590{
591 struct list_head *head = &resv->regions;
592 struct file_region *rg;
593 long chg = 0;
594
595 spin_lock(&resv->lock);
596 /* Locate each segment we overlap with, and count that overlap. */
597 list_for_each_entry(rg, head, link) {
598 long seg_from;
599 long seg_to;
600
601 if (rg->to <= f)
602 continue;
603 if (rg->from >= t)
604 break;
605
606 seg_from = max(rg->from, f);
607 seg_to = min(rg->to, t);
608
609 chg += seg_to - seg_from;
610 }
611 spin_unlock(&resv->lock);
612
613 return chg;
614}
615
616/*
617 * Convert the address within this vma to the page offset within
618 * the mapping, in pagecache page units; huge pages here.
619 */
620static pgoff_t vma_hugecache_offset(struct hstate *h,
621 struct vm_area_struct *vma, unsigned long address)
622{
623 return ((address - vma->vm_start) >> huge_page_shift(h)) +
624 (vma->vm_pgoff >> huge_page_order(h));
625}
626
627pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
628 unsigned long address)
629{
630 return vma_hugecache_offset(hstate_vma(vma), vma, address);
631}
632EXPORT_SYMBOL_GPL(linear_hugepage_index);
633
634/*
635 * Return the size of the pages allocated when backing a VMA. In the majority
636 * cases this will be same size as used by the page table entries.
637 */
638unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
639{
640 if (vma->vm_ops && vma->vm_ops->pagesize)
641 return vma->vm_ops->pagesize(vma);
642 return PAGE_SIZE;
643}
644EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
645
646/*
647 * Return the page size being used by the MMU to back a VMA. In the majority
648 * of cases, the page size used by the kernel matches the MMU size. On
649 * architectures where it differs, an architecture-specific 'strong'
650 * version of this symbol is required.
651 */
652__weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
653{
654 return vma_kernel_pagesize(vma);
655}
656
657/*
658 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
659 * bits of the reservation map pointer, which are always clear due to
660 * alignment.
661 */
662#define HPAGE_RESV_OWNER (1UL << 0)
663#define HPAGE_RESV_UNMAPPED (1UL << 1)
664#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
665
666/*
667 * These helpers are used to track how many pages are reserved for
668 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
669 * is guaranteed to have their future faults succeed.
670 *
671 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
672 * the reserve counters are updated with the hugetlb_lock held. It is safe
673 * to reset the VMA at fork() time as it is not in use yet and there is no
674 * chance of the global counters getting corrupted as a result of the values.
675 *
676 * The private mapping reservation is represented in a subtly different
677 * manner to a shared mapping. A shared mapping has a region map associated
678 * with the underlying file, this region map represents the backing file
679 * pages which have ever had a reservation assigned which this persists even
680 * after the page is instantiated. A private mapping has a region map
681 * associated with the original mmap which is attached to all VMAs which
682 * reference it, this region map represents those offsets which have consumed
683 * reservation ie. where pages have been instantiated.
684 */
685static unsigned long get_vma_private_data(struct vm_area_struct *vma)
686{
687 return (unsigned long)vma->vm_private_data;
688}
689
690static void set_vma_private_data(struct vm_area_struct *vma,
691 unsigned long value)
692{
693 vma->vm_private_data = (void *)value;
694}
695
696struct resv_map *resv_map_alloc(void)
697{
698 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
699 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
700
701 if (!resv_map || !rg) {
702 kfree(resv_map);
703 kfree(rg);
704 return NULL;
705 }
706
707 kref_init(&resv_map->refs);
708 spin_lock_init(&resv_map->lock);
709 INIT_LIST_HEAD(&resv_map->regions);
710
711 resv_map->adds_in_progress = 0;
712
713 INIT_LIST_HEAD(&resv_map->region_cache);
714 list_add(&rg->link, &resv_map->region_cache);
715 resv_map->region_cache_count = 1;
716
717 return resv_map;
718}
719
720void resv_map_release(struct kref *ref)
721{
722 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
723 struct list_head *head = &resv_map->region_cache;
724 struct file_region *rg, *trg;
725
726 /* Clear out any active regions before we release the map. */
727 region_del(resv_map, 0, LONG_MAX);
728
729 /* ... and any entries left in the cache */
730 list_for_each_entry_safe(rg, trg, head, link) {
731 list_del(&rg->link);
732 kfree(rg);
733 }
734
735 VM_BUG_ON(resv_map->adds_in_progress);
736
737 kfree(resv_map);
738}
739
740static inline struct resv_map *inode_resv_map(struct inode *inode)
741{
742 return inode->i_mapping->private_data;
743}
744
745static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
746{
747 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
748 if (vma->vm_flags & VM_MAYSHARE) {
749 struct address_space *mapping = vma->vm_file->f_mapping;
750 struct inode *inode = mapping->host;
751
752 return inode_resv_map(inode);
753
754 } else {
755 return (struct resv_map *)(get_vma_private_data(vma) &
756 ~HPAGE_RESV_MASK);
757 }
758}
759
760static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
761{
762 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
763 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
764
765 set_vma_private_data(vma, (get_vma_private_data(vma) &
766 HPAGE_RESV_MASK) | (unsigned long)map);
767}
768
769static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
770{
771 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
772 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
773
774 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
775}
776
777static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
778{
779 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
780
781 return (get_vma_private_data(vma) & flag) != 0;
782}
783
784/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
785void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
786{
787 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
788 if (!(vma->vm_flags & VM_MAYSHARE))
789 vma->vm_private_data = (void *)0;
790}
791
792/* Returns true if the VMA has associated reserve pages */
793static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
794{
795 if (vma->vm_flags & VM_NORESERVE) {
796 /*
797 * This address is already reserved by other process(chg == 0),
798 * so, we should decrement reserved count. Without decrementing,
799 * reserve count remains after releasing inode, because this
800 * allocated page will go into page cache and is regarded as
801 * coming from reserved pool in releasing step. Currently, we
802 * don't have any other solution to deal with this situation
803 * properly, so add work-around here.
804 */
805 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
806 return true;
807 else
808 return false;
809 }
810
811 /* Shared mappings always use reserves */
812 if (vma->vm_flags & VM_MAYSHARE) {
813 /*
814 * We know VM_NORESERVE is not set. Therefore, there SHOULD
815 * be a region map for all pages. The only situation where
816 * there is no region map is if a hole was punched via
817 * fallocate. In this case, there really are no reverves to
818 * use. This situation is indicated if chg != 0.
819 */
820 if (chg)
821 return false;
822 else
823 return true;
824 }
825
826 /*
827 * Only the process that called mmap() has reserves for
828 * private mappings.
829 */
830 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
831 /*
832 * Like the shared case above, a hole punch or truncate
833 * could have been performed on the private mapping.
834 * Examine the value of chg to determine if reserves
835 * actually exist or were previously consumed.
836 * Very Subtle - The value of chg comes from a previous
837 * call to vma_needs_reserves(). The reserve map for
838 * private mappings has different (opposite) semantics
839 * than that of shared mappings. vma_needs_reserves()
840 * has already taken this difference in semantics into
841 * account. Therefore, the meaning of chg is the same
842 * as in the shared case above. Code could easily be
843 * combined, but keeping it separate draws attention to
844 * subtle differences.
845 */
846 if (chg)
847 return false;
848 else
849 return true;
850 }
851
852 return false;
853}
854
855static void enqueue_huge_page(struct hstate *h, struct page *page)
856{
857 int nid = page_to_nid(page);
858 list_move(&page->lru, &h->hugepage_freelists[nid]);
859 h->free_huge_pages++;
860 h->free_huge_pages_node[nid]++;
861}
862
863static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
864{
865 struct page *page;
866
867 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
868 if (!PageHWPoison(page))
869 break;
870 /*
871 * if 'non-isolated free hugepage' not found on the list,
872 * the allocation fails.
873 */
874 if (&h->hugepage_freelists[nid] == &page->lru)
875 return NULL;
876 list_move(&page->lru, &h->hugepage_activelist);
877 set_page_refcounted(page);
878 h->free_huge_pages--;
879 h->free_huge_pages_node[nid]--;
880 return page;
881}
882
883static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
884 nodemask_t *nmask)
885{
886 unsigned int cpuset_mems_cookie;
887 struct zonelist *zonelist;
888 struct zone *zone;
889 struct zoneref *z;
890 int node = -1;
891
892 zonelist = node_zonelist(nid, gfp_mask);
893
894retry_cpuset:
895 cpuset_mems_cookie = read_mems_allowed_begin();
896 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
897 struct page *page;
898
899 if (!cpuset_zone_allowed(zone, gfp_mask))
900 continue;
901 /*
902 * no need to ask again on the same node. Pool is node rather than
903 * zone aware
904 */
905 if (zone_to_nid(zone) == node)
906 continue;
907 node = zone_to_nid(zone);
908
909 page = dequeue_huge_page_node_exact(h, node);
910 if (page)
911 return page;
912 }
913 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
914 goto retry_cpuset;
915
916 return NULL;
917}
918
919/* Movability of hugepages depends on migration support. */
920static inline gfp_t htlb_alloc_mask(struct hstate *h)
921{
922 if (hugepage_migration_supported(h))
923 return GFP_HIGHUSER_MOVABLE;
924 else
925 return GFP_HIGHUSER;
926}
927
928static struct page *dequeue_huge_page_vma(struct hstate *h,
929 struct vm_area_struct *vma,
930 unsigned long address, int avoid_reserve,
931 long chg)
932{
933 struct page *page;
934 struct mempolicy *mpol;
935 gfp_t gfp_mask;
936 nodemask_t *nodemask;
937 int nid;
938
939 /*
940 * A child process with MAP_PRIVATE mappings created by their parent
941 * have no page reserves. This check ensures that reservations are
942 * not "stolen". The child may still get SIGKILLed
943 */
944 if (!vma_has_reserves(vma, chg) &&
945 h->free_huge_pages - h->resv_huge_pages == 0)
946 goto err;
947
948 /* If reserves cannot be used, ensure enough pages are in the pool */
949 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
950 goto err;
951
952 gfp_mask = htlb_alloc_mask(h);
953 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
954 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
955 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
956 SetPagePrivate(page);
957 h->resv_huge_pages--;
958 }
959
960 mpol_cond_put(mpol);
961 return page;
962
963err:
964 return NULL;
965}
966
967/*
968 * common helper functions for hstate_next_node_to_{alloc|free}.
969 * We may have allocated or freed a huge page based on a different
970 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
971 * be outside of *nodes_allowed. Ensure that we use an allowed
972 * node for alloc or free.
973 */
974static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
975{
976 nid = next_node_in(nid, *nodes_allowed);
977 VM_BUG_ON(nid >= MAX_NUMNODES);
978
979 return nid;
980}
981
982static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
983{
984 if (!node_isset(nid, *nodes_allowed))
985 nid = next_node_allowed(nid, nodes_allowed);
986 return nid;
987}
988
989/*
990 * returns the previously saved node ["this node"] from which to
991 * allocate a persistent huge page for the pool and advance the
992 * next node from which to allocate, handling wrap at end of node
993 * mask.
994 */
995static int hstate_next_node_to_alloc(struct hstate *h,
996 nodemask_t *nodes_allowed)
997{
998 int nid;
999
1000 VM_BUG_ON(!nodes_allowed);
1001
1002 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1003 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1004
1005 return nid;
1006}
1007
1008/*
1009 * helper for free_pool_huge_page() - return the previously saved
1010 * node ["this node"] from which to free a huge page. Advance the
1011 * next node id whether or not we find a free huge page to free so
1012 * that the next attempt to free addresses the next node.
1013 */
1014static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1015{
1016 int nid;
1017
1018 VM_BUG_ON(!nodes_allowed);
1019
1020 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1021 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1022
1023 return nid;
1024}
1025
1026#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1027 for (nr_nodes = nodes_weight(*mask); \
1028 nr_nodes > 0 && \
1029 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1030 nr_nodes--)
1031
1032#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1033 for (nr_nodes = nodes_weight(*mask); \
1034 nr_nodes > 0 && \
1035 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1036 nr_nodes--)
1037
1038#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1039static void destroy_compound_gigantic_page(struct page *page,
1040 unsigned int order)
1041{
1042 int i;
1043 int nr_pages = 1 << order;
1044 struct page *p = page + 1;
1045
1046 atomic_set(compound_mapcount_ptr(page), 0);
1047 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1048 clear_compound_head(p);
1049 set_page_refcounted(p);
1050 }
1051
1052 set_compound_order(page, 0);
1053 __ClearPageHead(page);
1054}
1055
1056static void free_gigantic_page(struct page *page, unsigned int order)
1057{
1058 free_contig_range(page_to_pfn(page), 1 << order);
1059}
1060
1061static int __alloc_gigantic_page(unsigned long start_pfn,
1062 unsigned long nr_pages, gfp_t gfp_mask)
1063{
1064 unsigned long end_pfn = start_pfn + nr_pages;
1065 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1066 gfp_mask);
1067}
1068
1069static bool pfn_range_valid_gigantic(struct zone *z,
1070 unsigned long start_pfn, unsigned long nr_pages)
1071{
1072 unsigned long i, end_pfn = start_pfn + nr_pages;
1073 struct page *page;
1074
1075 for (i = start_pfn; i < end_pfn; i++) {
1076 if (!pfn_valid(i))
1077 return false;
1078
1079 page = pfn_to_page(i);
1080
1081 if (page_zone(page) != z)
1082 return false;
1083
1084 if (PageReserved(page))
1085 return false;
1086
1087 if (page_count(page) > 0)
1088 return false;
1089
1090 if (PageHuge(page))
1091 return false;
1092 }
1093
1094 return true;
1095}
1096
1097static bool zone_spans_last_pfn(const struct zone *zone,
1098 unsigned long start_pfn, unsigned long nr_pages)
1099{
1100 unsigned long last_pfn = start_pfn + nr_pages - 1;
1101 return zone_spans_pfn(zone, last_pfn);
1102}
1103
1104static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1105 int nid, nodemask_t *nodemask)
1106{
1107 unsigned int order = huge_page_order(h);
1108 unsigned long nr_pages = 1 << order;
1109 unsigned long ret, pfn, flags;
1110 struct zonelist *zonelist;
1111 struct zone *zone;
1112 struct zoneref *z;
1113
1114 zonelist = node_zonelist(nid, gfp_mask);
1115 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1116 spin_lock_irqsave(&zone->lock, flags);
1117
1118 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1119 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1120 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1121 /*
1122 * We release the zone lock here because
1123 * alloc_contig_range() will also lock the zone
1124 * at some point. If there's an allocation
1125 * spinning on this lock, it may win the race
1126 * and cause alloc_contig_range() to fail...
1127 */
1128 spin_unlock_irqrestore(&zone->lock, flags);
1129 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1130 if (!ret)
1131 return pfn_to_page(pfn);
1132 spin_lock_irqsave(&zone->lock, flags);
1133 }
1134 pfn += nr_pages;
1135 }
1136
1137 spin_unlock_irqrestore(&zone->lock, flags);
1138 }
1139
1140 return NULL;
1141}
1142
1143static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1144static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1145
1146#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1147static inline bool gigantic_page_supported(void) { return false; }
1148static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1149 int nid, nodemask_t *nodemask) { return NULL; }
1150static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1151static inline void destroy_compound_gigantic_page(struct page *page,
1152 unsigned int order) { }
1153#endif
1154
1155static void update_and_free_page(struct hstate *h, struct page *page)
1156{
1157 int i;
1158
1159 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1160 return;
1161
1162 h->nr_huge_pages--;
1163 h->nr_huge_pages_node[page_to_nid(page)]--;
1164 for (i = 0; i < pages_per_huge_page(h); i++) {
1165 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1166 1 << PG_referenced | 1 << PG_dirty |
1167 1 << PG_active | 1 << PG_private |
1168 1 << PG_writeback);
1169 }
1170 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1171 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1172 set_page_refcounted(page);
1173 if (hstate_is_gigantic(h)) {
1174 destroy_compound_gigantic_page(page, huge_page_order(h));
1175 free_gigantic_page(page, huge_page_order(h));
1176 } else {
1177 __free_pages(page, huge_page_order(h));
1178 }
1179}
1180
1181struct hstate *size_to_hstate(unsigned long size)
1182{
1183 struct hstate *h;
1184
1185 for_each_hstate(h) {
1186 if (huge_page_size(h) == size)
1187 return h;
1188 }
1189 return NULL;
1190}
1191
1192/*
1193 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1194 * to hstate->hugepage_activelist.)
1195 *
1196 * This function can be called for tail pages, but never returns true for them.
1197 */
1198bool page_huge_active(struct page *page)
1199{
1200 VM_BUG_ON_PAGE(!PageHuge(page), page);
1201 return PageHead(page) && PagePrivate(&page[1]);
1202}
1203
1204/* never called for tail page */
1205static void set_page_huge_active(struct page *page)
1206{
1207 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1208 SetPagePrivate(&page[1]);
1209}
1210
1211static void clear_page_huge_active(struct page *page)
1212{
1213 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1214 ClearPagePrivate(&page[1]);
1215}
1216
1217/*
1218 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1219 * code
1220 */
1221static inline bool PageHugeTemporary(struct page *page)
1222{
1223 if (!PageHuge(page))
1224 return false;
1225
1226 return (unsigned long)page[2].mapping == -1U;
1227}
1228
1229static inline void SetPageHugeTemporary(struct page *page)
1230{
1231 page[2].mapping = (void *)-1U;
1232}
1233
1234static inline void ClearPageHugeTemporary(struct page *page)
1235{
1236 page[2].mapping = NULL;
1237}
1238
1239void free_huge_page(struct page *page)
1240{
1241 /*
1242 * Can't pass hstate in here because it is called from the
1243 * compound page destructor.
1244 */
1245 struct hstate *h = page_hstate(page);
1246 int nid = page_to_nid(page);
1247 struct hugepage_subpool *spool =
1248 (struct hugepage_subpool *)page_private(page);
1249 bool restore_reserve;
1250
1251 set_page_private(page, 0);
1252 page->mapping = NULL;
1253 VM_BUG_ON_PAGE(page_count(page), page);
1254 VM_BUG_ON_PAGE(page_mapcount(page), page);
1255 restore_reserve = PagePrivate(page);
1256 ClearPagePrivate(page);
1257
1258 /*
1259 * A return code of zero implies that the subpool will be under its
1260 * minimum size if the reservation is not restored after page is free.
1261 * Therefore, force restore_reserve operation.
1262 */
1263 if (hugepage_subpool_put_pages(spool, 1) == 0)
1264 restore_reserve = true;
1265
1266 spin_lock(&hugetlb_lock);
1267 clear_page_huge_active(page);
1268 hugetlb_cgroup_uncharge_page(hstate_index(h),
1269 pages_per_huge_page(h), page);
1270 if (restore_reserve)
1271 h->resv_huge_pages++;
1272
1273 if (PageHugeTemporary(page)) {
1274 list_del(&page->lru);
1275 ClearPageHugeTemporary(page);
1276 update_and_free_page(h, page);
1277 } else if (h->surplus_huge_pages_node[nid]) {
1278 /* remove the page from active list */
1279 list_del(&page->lru);
1280 update_and_free_page(h, page);
1281 h->surplus_huge_pages--;
1282 h->surplus_huge_pages_node[nid]--;
1283 } else {
1284 arch_clear_hugepage_flags(page);
1285 enqueue_huge_page(h, page);
1286 }
1287 spin_unlock(&hugetlb_lock);
1288}
1289
1290static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1291{
1292 INIT_LIST_HEAD(&page->lru);
1293 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1294 spin_lock(&hugetlb_lock);
1295 set_hugetlb_cgroup(page, NULL);
1296 h->nr_huge_pages++;
1297 h->nr_huge_pages_node[nid]++;
1298 spin_unlock(&hugetlb_lock);
1299}
1300
1301static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1302{
1303 int i;
1304 int nr_pages = 1 << order;
1305 struct page *p = page + 1;
1306
1307 /* we rely on prep_new_huge_page to set the destructor */
1308 set_compound_order(page, order);
1309 __ClearPageReserved(page);
1310 __SetPageHead(page);
1311 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1312 /*
1313 * For gigantic hugepages allocated through bootmem at
1314 * boot, it's safer to be consistent with the not-gigantic
1315 * hugepages and clear the PG_reserved bit from all tail pages
1316 * too. Otherwse drivers using get_user_pages() to access tail
1317 * pages may get the reference counting wrong if they see
1318 * PG_reserved set on a tail page (despite the head page not
1319 * having PG_reserved set). Enforcing this consistency between
1320 * head and tail pages allows drivers to optimize away a check
1321 * on the head page when they need know if put_page() is needed
1322 * after get_user_pages().
1323 */
1324 __ClearPageReserved(p);
1325 set_page_count(p, 0);
1326 set_compound_head(p, page);
1327 }
1328 atomic_set(compound_mapcount_ptr(page), -1);
1329}
1330
1331/*
1332 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1333 * transparent huge pages. See the PageTransHuge() documentation for more
1334 * details.
1335 */
1336int PageHuge(struct page *page)
1337{
1338 if (!PageCompound(page))
1339 return 0;
1340
1341 page = compound_head(page);
1342 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1343}
1344EXPORT_SYMBOL_GPL(PageHuge);
1345
1346/*
1347 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1348 * normal or transparent huge pages.
1349 */
1350int PageHeadHuge(struct page *page_head)
1351{
1352 if (!PageHead(page_head))
1353 return 0;
1354
1355 return get_compound_page_dtor(page_head) == free_huge_page;
1356}
1357
1358pgoff_t __basepage_index(struct page *page)
1359{
1360 struct page *page_head = compound_head(page);
1361 pgoff_t index = page_index(page_head);
1362 unsigned long compound_idx;
1363
1364 if (!PageHuge(page_head))
1365 return page_index(page);
1366
1367 if (compound_order(page_head) >= MAX_ORDER)
1368 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1369 else
1370 compound_idx = page - page_head;
1371
1372 return (index << compound_order(page_head)) + compound_idx;
1373}
1374
1375static struct page *alloc_buddy_huge_page(struct hstate *h,
1376 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1377{
1378 int order = huge_page_order(h);
1379 struct page *page;
1380
1381 gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1382 if (nid == NUMA_NO_NODE)
1383 nid = numa_mem_id();
1384 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1385 if (page)
1386 __count_vm_event(HTLB_BUDDY_PGALLOC);
1387 else
1388 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1389
1390 return page;
1391}
1392
1393/*
1394 * Common helper to allocate a fresh hugetlb page. All specific allocators
1395 * should use this function to get new hugetlb pages
1396 */
1397static struct page *alloc_fresh_huge_page(struct hstate *h,
1398 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1399{
1400 struct page *page;
1401
1402 if (hstate_is_gigantic(h))
1403 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1404 else
1405 page = alloc_buddy_huge_page(h, gfp_mask,
1406 nid, nmask);
1407 if (!page)
1408 return NULL;
1409
1410 if (hstate_is_gigantic(h))
1411 prep_compound_gigantic_page(page, huge_page_order(h));
1412 prep_new_huge_page(h, page, page_to_nid(page));
1413
1414 return page;
1415}
1416
1417/*
1418 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1419 * manner.
1420 */
1421static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1422{
1423 struct page *page;
1424 int nr_nodes, node;
1425 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1426
1427 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1428 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed);
1429 if (page)
1430 break;
1431 }
1432
1433 if (!page)
1434 return 0;
1435
1436 put_page(page); /* free it into the hugepage allocator */
1437
1438 return 1;
1439}
1440
1441/*
1442 * Free huge page from pool from next node to free.
1443 * Attempt to keep persistent huge pages more or less
1444 * balanced over allowed nodes.
1445 * Called with hugetlb_lock locked.
1446 */
1447static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1448 bool acct_surplus)
1449{
1450 int nr_nodes, node;
1451 int ret = 0;
1452
1453 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1454 /*
1455 * If we're returning unused surplus pages, only examine
1456 * nodes with surplus pages.
1457 */
1458 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1459 !list_empty(&h->hugepage_freelists[node])) {
1460 struct page *page =
1461 list_entry(h->hugepage_freelists[node].next,
1462 struct page, lru);
1463 list_del(&page->lru);
1464 h->free_huge_pages--;
1465 h->free_huge_pages_node[node]--;
1466 if (acct_surplus) {
1467 h->surplus_huge_pages--;
1468 h->surplus_huge_pages_node[node]--;
1469 }
1470 update_and_free_page(h, page);
1471 ret = 1;
1472 break;
1473 }
1474 }
1475
1476 return ret;
1477}
1478
1479/*
1480 * Dissolve a given free hugepage into free buddy pages. This function does
1481 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1482 * number of free hugepages would be reduced below the number of reserved
1483 * hugepages.
1484 */
1485int dissolve_free_huge_page(struct page *page)
1486{
1487 int rc = 0;
1488
1489 spin_lock(&hugetlb_lock);
1490 if (PageHuge(page) && !page_count(page)) {
1491 struct page *head = compound_head(page);
1492 struct hstate *h = page_hstate(head);
1493 int nid = page_to_nid(head);
1494 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1495 rc = -EBUSY;
1496 goto out;
1497 }
1498 /*
1499 * Move PageHWPoison flag from head page to the raw error page,
1500 * which makes any subpages rather than the error page reusable.
1501 */
1502 if (PageHWPoison(head) && page != head) {
1503 SetPageHWPoison(page);
1504 ClearPageHWPoison(head);
1505 }
1506 list_del(&head->lru);
1507 h->free_huge_pages--;
1508 h->free_huge_pages_node[nid]--;
1509 h->max_huge_pages--;
1510 update_and_free_page(h, head);
1511 }
1512out:
1513 spin_unlock(&hugetlb_lock);
1514 return rc;
1515}
1516
1517/*
1518 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1519 * make specified memory blocks removable from the system.
1520 * Note that this will dissolve a free gigantic hugepage completely, if any
1521 * part of it lies within the given range.
1522 * Also note that if dissolve_free_huge_page() returns with an error, all
1523 * free hugepages that were dissolved before that error are lost.
1524 */
1525int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1526{
1527 unsigned long pfn;
1528 struct page *page;
1529 int rc = 0;
1530
1531 if (!hugepages_supported())
1532 return rc;
1533
1534 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1535 page = pfn_to_page(pfn);
1536 if (PageHuge(page) && !page_count(page)) {
1537 rc = dissolve_free_huge_page(page);
1538 if (rc)
1539 break;
1540 }
1541 }
1542
1543 return rc;
1544}
1545
1546/*
1547 * Allocates a fresh surplus page from the page allocator.
1548 */
1549static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1550 int nid, nodemask_t *nmask)
1551{
1552 struct page *page = NULL;
1553
1554 if (hstate_is_gigantic(h))
1555 return NULL;
1556
1557 spin_lock(&hugetlb_lock);
1558 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1559 goto out_unlock;
1560 spin_unlock(&hugetlb_lock);
1561
1562 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1563 if (!page)
1564 return NULL;
1565
1566 spin_lock(&hugetlb_lock);
1567 /*
1568 * We could have raced with the pool size change.
1569 * Double check that and simply deallocate the new page
1570 * if we would end up overcommiting the surpluses. Abuse
1571 * temporary page to workaround the nasty free_huge_page
1572 * codeflow
1573 */
1574 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1575 SetPageHugeTemporary(page);
1576 put_page(page);
1577 page = NULL;
1578 } else {
1579 h->surplus_huge_pages++;
1580 h->surplus_huge_pages_node[page_to_nid(page)]++;
1581 }
1582
1583out_unlock:
1584 spin_unlock(&hugetlb_lock);
1585
1586 return page;
1587}
1588
1589static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1590 int nid, nodemask_t *nmask)
1591{
1592 struct page *page;
1593
1594 if (hstate_is_gigantic(h))
1595 return NULL;
1596
1597 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1598 if (!page)
1599 return NULL;
1600
1601 /*
1602 * We do not account these pages as surplus because they are only
1603 * temporary and will be released properly on the last reference
1604 */
1605 SetPageHugeTemporary(page);
1606
1607 return page;
1608}
1609
1610/*
1611 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1612 */
1613static
1614struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1615 struct vm_area_struct *vma, unsigned long addr)
1616{
1617 struct page *page;
1618 struct mempolicy *mpol;
1619 gfp_t gfp_mask = htlb_alloc_mask(h);
1620 int nid;
1621 nodemask_t *nodemask;
1622
1623 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1624 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1625 mpol_cond_put(mpol);
1626
1627 return page;
1628}
1629
1630/* page migration callback function */
1631struct page *alloc_huge_page_node(struct hstate *h, int nid)
1632{
1633 gfp_t gfp_mask = htlb_alloc_mask(h);
1634 struct page *page = NULL;
1635
1636 if (nid != NUMA_NO_NODE)
1637 gfp_mask |= __GFP_THISNODE;
1638
1639 spin_lock(&hugetlb_lock);
1640 if (h->free_huge_pages - h->resv_huge_pages > 0)
1641 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1642 spin_unlock(&hugetlb_lock);
1643
1644 if (!page)
1645 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1646
1647 return page;
1648}
1649
1650/* page migration callback function */
1651struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1652 nodemask_t *nmask)
1653{
1654 gfp_t gfp_mask = htlb_alloc_mask(h);
1655
1656 spin_lock(&hugetlb_lock);
1657 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1658 struct page *page;
1659
1660 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1661 if (page) {
1662 spin_unlock(&hugetlb_lock);
1663 return page;
1664 }
1665 }
1666 spin_unlock(&hugetlb_lock);
1667
1668 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1669}
1670
1671/* mempolicy aware migration callback */
1672struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1673 unsigned long address)
1674{
1675 struct mempolicy *mpol;
1676 nodemask_t *nodemask;
1677 struct page *page;
1678 gfp_t gfp_mask;
1679 int node;
1680
1681 gfp_mask = htlb_alloc_mask(h);
1682 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1683 page = alloc_huge_page_nodemask(h, node, nodemask);
1684 mpol_cond_put(mpol);
1685
1686 return page;
1687}
1688
1689/*
1690 * Increase the hugetlb pool such that it can accommodate a reservation
1691 * of size 'delta'.
1692 */
1693static int gather_surplus_pages(struct hstate *h, int delta)
1694{
1695 struct list_head surplus_list;
1696 struct page *page, *tmp;
1697 int ret, i;
1698 int needed, allocated;
1699 bool alloc_ok = true;
1700
1701 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1702 if (needed <= 0) {
1703 h->resv_huge_pages += delta;
1704 return 0;
1705 }
1706
1707 allocated = 0;
1708 INIT_LIST_HEAD(&surplus_list);
1709
1710 ret = -ENOMEM;
1711retry:
1712 spin_unlock(&hugetlb_lock);
1713 for (i = 0; i < needed; i++) {
1714 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1715 NUMA_NO_NODE, NULL);
1716 if (!page) {
1717 alloc_ok = false;
1718 break;
1719 }
1720 list_add(&page->lru, &surplus_list);
1721 cond_resched();
1722 }
1723 allocated += i;
1724
1725 /*
1726 * After retaking hugetlb_lock, we need to recalculate 'needed'
1727 * because either resv_huge_pages or free_huge_pages may have changed.
1728 */
1729 spin_lock(&hugetlb_lock);
1730 needed = (h->resv_huge_pages + delta) -
1731 (h->free_huge_pages + allocated);
1732 if (needed > 0) {
1733 if (alloc_ok)
1734 goto retry;
1735 /*
1736 * We were not able to allocate enough pages to
1737 * satisfy the entire reservation so we free what
1738 * we've allocated so far.
1739 */
1740 goto free;
1741 }
1742 /*
1743 * The surplus_list now contains _at_least_ the number of extra pages
1744 * needed to accommodate the reservation. Add the appropriate number
1745 * of pages to the hugetlb pool and free the extras back to the buddy
1746 * allocator. Commit the entire reservation here to prevent another
1747 * process from stealing the pages as they are added to the pool but
1748 * before they are reserved.
1749 */
1750 needed += allocated;
1751 h->resv_huge_pages += delta;
1752 ret = 0;
1753
1754 /* Free the needed pages to the hugetlb pool */
1755 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1756 if ((--needed) < 0)
1757 break;
1758 /*
1759 * This page is now managed by the hugetlb allocator and has
1760 * no users -- drop the buddy allocator's reference.
1761 */
1762 put_page_testzero(page);
1763 VM_BUG_ON_PAGE(page_count(page), page);
1764 enqueue_huge_page(h, page);
1765 }
1766free:
1767 spin_unlock(&hugetlb_lock);
1768
1769 /* Free unnecessary surplus pages to the buddy allocator */
1770 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1771 put_page(page);
1772 spin_lock(&hugetlb_lock);
1773
1774 return ret;
1775}
1776
1777/*
1778 * This routine has two main purposes:
1779 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1780 * in unused_resv_pages. This corresponds to the prior adjustments made
1781 * to the associated reservation map.
1782 * 2) Free any unused surplus pages that may have been allocated to satisfy
1783 * the reservation. As many as unused_resv_pages may be freed.
1784 *
1785 * Called with hugetlb_lock held. However, the lock could be dropped (and
1786 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1787 * we must make sure nobody else can claim pages we are in the process of
1788 * freeing. Do this by ensuring resv_huge_page always is greater than the
1789 * number of huge pages we plan to free when dropping the lock.
1790 */
1791static void return_unused_surplus_pages(struct hstate *h,
1792 unsigned long unused_resv_pages)
1793{
1794 unsigned long nr_pages;
1795
1796 /* Cannot return gigantic pages currently */
1797 if (hstate_is_gigantic(h))
1798 goto out;
1799
1800 /*
1801 * Part (or even all) of the reservation could have been backed
1802 * by pre-allocated pages. Only free surplus pages.
1803 */
1804 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1805
1806 /*
1807 * We want to release as many surplus pages as possible, spread
1808 * evenly across all nodes with memory. Iterate across these nodes
1809 * until we can no longer free unreserved surplus pages. This occurs
1810 * when the nodes with surplus pages have no free pages.
1811 * free_pool_huge_page() will balance the the freed pages across the
1812 * on-line nodes with memory and will handle the hstate accounting.
1813 *
1814 * Note that we decrement resv_huge_pages as we free the pages. If
1815 * we drop the lock, resv_huge_pages will still be sufficiently large
1816 * to cover subsequent pages we may free.
1817 */
1818 while (nr_pages--) {
1819 h->resv_huge_pages--;
1820 unused_resv_pages--;
1821 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1822 goto out;
1823 cond_resched_lock(&hugetlb_lock);
1824 }
1825
1826out:
1827 /* Fully uncommit the reservation */
1828 h->resv_huge_pages -= unused_resv_pages;
1829}
1830
1831
1832/*
1833 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1834 * are used by the huge page allocation routines to manage reservations.
1835 *
1836 * vma_needs_reservation is called to determine if the huge page at addr
1837 * within the vma has an associated reservation. If a reservation is
1838 * needed, the value 1 is returned. The caller is then responsible for
1839 * managing the global reservation and subpool usage counts. After
1840 * the huge page has been allocated, vma_commit_reservation is called
1841 * to add the page to the reservation map. If the page allocation fails,
1842 * the reservation must be ended instead of committed. vma_end_reservation
1843 * is called in such cases.
1844 *
1845 * In the normal case, vma_commit_reservation returns the same value
1846 * as the preceding vma_needs_reservation call. The only time this
1847 * is not the case is if a reserve map was changed between calls. It
1848 * is the responsibility of the caller to notice the difference and
1849 * take appropriate action.
1850 *
1851 * vma_add_reservation is used in error paths where a reservation must
1852 * be restored when a newly allocated huge page must be freed. It is
1853 * to be called after calling vma_needs_reservation to determine if a
1854 * reservation exists.
1855 */
1856enum vma_resv_mode {
1857 VMA_NEEDS_RESV,
1858 VMA_COMMIT_RESV,
1859 VMA_END_RESV,
1860 VMA_ADD_RESV,
1861};
1862static long __vma_reservation_common(struct hstate *h,
1863 struct vm_area_struct *vma, unsigned long addr,
1864 enum vma_resv_mode mode)
1865{
1866 struct resv_map *resv;
1867 pgoff_t idx;
1868 long ret;
1869
1870 resv = vma_resv_map(vma);
1871 if (!resv)
1872 return 1;
1873
1874 idx = vma_hugecache_offset(h, vma, addr);
1875 switch (mode) {
1876 case VMA_NEEDS_RESV:
1877 ret = region_chg(resv, idx, idx + 1);
1878 break;
1879 case VMA_COMMIT_RESV:
1880 ret = region_add(resv, idx, idx + 1);
1881 break;
1882 case VMA_END_RESV:
1883 region_abort(resv, idx, idx + 1);
1884 ret = 0;
1885 break;
1886 case VMA_ADD_RESV:
1887 if (vma->vm_flags & VM_MAYSHARE)
1888 ret = region_add(resv, idx, idx + 1);
1889 else {
1890 region_abort(resv, idx, idx + 1);
1891 ret = region_del(resv, idx, idx + 1);
1892 }
1893 break;
1894 default:
1895 BUG();
1896 }
1897
1898 if (vma->vm_flags & VM_MAYSHARE)
1899 return ret;
1900 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1901 /*
1902 * In most cases, reserves always exist for private mappings.
1903 * However, a file associated with mapping could have been
1904 * hole punched or truncated after reserves were consumed.
1905 * As subsequent fault on such a range will not use reserves.
1906 * Subtle - The reserve map for private mappings has the
1907 * opposite meaning than that of shared mappings. If NO
1908 * entry is in the reserve map, it means a reservation exists.
1909 * If an entry exists in the reserve map, it means the
1910 * reservation has already been consumed. As a result, the
1911 * return value of this routine is the opposite of the
1912 * value returned from reserve map manipulation routines above.
1913 */
1914 if (ret)
1915 return 0;
1916 else
1917 return 1;
1918 }
1919 else
1920 return ret < 0 ? ret : 0;
1921}
1922
1923static long vma_needs_reservation(struct hstate *h,
1924 struct vm_area_struct *vma, unsigned long addr)
1925{
1926 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1927}
1928
1929static long vma_commit_reservation(struct hstate *h,
1930 struct vm_area_struct *vma, unsigned long addr)
1931{
1932 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1933}
1934
1935static void vma_end_reservation(struct hstate *h,
1936 struct vm_area_struct *vma, unsigned long addr)
1937{
1938 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1939}
1940
1941static long vma_add_reservation(struct hstate *h,
1942 struct vm_area_struct *vma, unsigned long addr)
1943{
1944 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1945}
1946
1947/*
1948 * This routine is called to restore a reservation on error paths. In the
1949 * specific error paths, a huge page was allocated (via alloc_huge_page)
1950 * and is about to be freed. If a reservation for the page existed,
1951 * alloc_huge_page would have consumed the reservation and set PagePrivate
1952 * in the newly allocated page. When the page is freed via free_huge_page,
1953 * the global reservation count will be incremented if PagePrivate is set.
1954 * However, free_huge_page can not adjust the reserve map. Adjust the
1955 * reserve map here to be consistent with global reserve count adjustments
1956 * to be made by free_huge_page.
1957 */
1958static void restore_reserve_on_error(struct hstate *h,
1959 struct vm_area_struct *vma, unsigned long address,
1960 struct page *page)
1961{
1962 if (unlikely(PagePrivate(page))) {
1963 long rc = vma_needs_reservation(h, vma, address);
1964
1965 if (unlikely(rc < 0)) {
1966 /*
1967 * Rare out of memory condition in reserve map
1968 * manipulation. Clear PagePrivate so that
1969 * global reserve count will not be incremented
1970 * by free_huge_page. This will make it appear
1971 * as though the reservation for this page was
1972 * consumed. This may prevent the task from
1973 * faulting in the page at a later time. This
1974 * is better than inconsistent global huge page
1975 * accounting of reserve counts.
1976 */
1977 ClearPagePrivate(page);
1978 } else if (rc) {
1979 rc = vma_add_reservation(h, vma, address);
1980 if (unlikely(rc < 0))
1981 /*
1982 * See above comment about rare out of
1983 * memory condition.
1984 */
1985 ClearPagePrivate(page);
1986 } else
1987 vma_end_reservation(h, vma, address);
1988 }
1989}
1990
1991struct page *alloc_huge_page(struct vm_area_struct *vma,
1992 unsigned long addr, int avoid_reserve)
1993{
1994 struct hugepage_subpool *spool = subpool_vma(vma);
1995 struct hstate *h = hstate_vma(vma);
1996 struct page *page;
1997 long map_chg, map_commit;
1998 long gbl_chg;
1999 int ret, idx;
2000 struct hugetlb_cgroup *h_cg;
2001
2002 idx = hstate_index(h);
2003 /*
2004 * Examine the region/reserve map to determine if the process
2005 * has a reservation for the page to be allocated. A return
2006 * code of zero indicates a reservation exists (no change).
2007 */
2008 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2009 if (map_chg < 0)
2010 return ERR_PTR(-ENOMEM);
2011
2012 /*
2013 * Processes that did not create the mapping will have no
2014 * reserves as indicated by the region/reserve map. Check
2015 * that the allocation will not exceed the subpool limit.
2016 * Allocations for MAP_NORESERVE mappings also need to be
2017 * checked against any subpool limit.
2018 */
2019 if (map_chg || avoid_reserve) {
2020 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2021 if (gbl_chg < 0) {
2022 vma_end_reservation(h, vma, addr);
2023 return ERR_PTR(-ENOSPC);
2024 }
2025
2026 /*
2027 * Even though there was no reservation in the region/reserve
2028 * map, there could be reservations associated with the
2029 * subpool that can be used. This would be indicated if the
2030 * return value of hugepage_subpool_get_pages() is zero.
2031 * However, if avoid_reserve is specified we still avoid even
2032 * the subpool reservations.
2033 */
2034 if (avoid_reserve)
2035 gbl_chg = 1;
2036 }
2037
2038 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2039 if (ret)
2040 goto out_subpool_put;
2041
2042 spin_lock(&hugetlb_lock);
2043 /*
2044 * glb_chg is passed to indicate whether or not a page must be taken
2045 * from the global free pool (global change). gbl_chg == 0 indicates
2046 * a reservation exists for the allocation.
2047 */
2048 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2049 if (!page) {
2050 spin_unlock(&hugetlb_lock);
2051 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2052 if (!page)
2053 goto out_uncharge_cgroup;
2054 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2055 SetPagePrivate(page);
2056 h->resv_huge_pages--;
2057 }
2058 spin_lock(&hugetlb_lock);
2059 list_move(&page->lru, &h->hugepage_activelist);
2060 /* Fall through */
2061 }
2062 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2063 spin_unlock(&hugetlb_lock);
2064
2065 set_page_private(page, (unsigned long)spool);
2066
2067 map_commit = vma_commit_reservation(h, vma, addr);
2068 if (unlikely(map_chg > map_commit)) {
2069 /*
2070 * The page was added to the reservation map between
2071 * vma_needs_reservation and vma_commit_reservation.
2072 * This indicates a race with hugetlb_reserve_pages.
2073 * Adjust for the subpool count incremented above AND
2074 * in hugetlb_reserve_pages for the same page. Also,
2075 * the reservation count added in hugetlb_reserve_pages
2076 * no longer applies.
2077 */
2078 long rsv_adjust;
2079
2080 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2081 hugetlb_acct_memory(h, -rsv_adjust);
2082 }
2083 return page;
2084
2085out_uncharge_cgroup:
2086 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2087out_subpool_put:
2088 if (map_chg || avoid_reserve)
2089 hugepage_subpool_put_pages(spool, 1);
2090 vma_end_reservation(h, vma, addr);
2091 return ERR_PTR(-ENOSPC);
2092}
2093
2094int alloc_bootmem_huge_page(struct hstate *h)
2095 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2096int __alloc_bootmem_huge_page(struct hstate *h)
2097{
2098 struct huge_bootmem_page *m;
2099 int nr_nodes, node;
2100
2101 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2102 void *addr;
2103
2104 addr = memblock_virt_alloc_try_nid_nopanic(
2105 huge_page_size(h), huge_page_size(h),
2106 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2107 if (addr) {
2108 /*
2109 * Use the beginning of the huge page to store the
2110 * huge_bootmem_page struct (until gather_bootmem
2111 * puts them into the mem_map).
2112 */
2113 m = addr;
2114 goto found;
2115 }
2116 }
2117 return 0;
2118
2119found:
2120 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2121 /* Put them into a private list first because mem_map is not up yet */
2122 list_add(&m->list, &huge_boot_pages);
2123 m->hstate = h;
2124 return 1;
2125}
2126
2127static void __init prep_compound_huge_page(struct page *page,
2128 unsigned int order)
2129{
2130 if (unlikely(order > (MAX_ORDER - 1)))
2131 prep_compound_gigantic_page(page, order);
2132 else
2133 prep_compound_page(page, order);
2134}
2135
2136/* Put bootmem huge pages into the standard lists after mem_map is up */
2137static void __init gather_bootmem_prealloc(void)
2138{
2139 struct huge_bootmem_page *m;
2140
2141 list_for_each_entry(m, &huge_boot_pages, list) {
2142 struct hstate *h = m->hstate;
2143 struct page *page;
2144
2145#ifdef CONFIG_HIGHMEM
2146 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2147 memblock_free_late(__pa(m),
2148 sizeof(struct huge_bootmem_page));
2149#else
2150 page = virt_to_page(m);
2151#endif
2152 WARN_ON(page_count(page) != 1);
2153 prep_compound_huge_page(page, h->order);
2154 WARN_ON(PageReserved(page));
2155 prep_new_huge_page(h, page, page_to_nid(page));
2156 put_page(page); /* free it into the hugepage allocator */
2157
2158 /*
2159 * If we had gigantic hugepages allocated at boot time, we need
2160 * to restore the 'stolen' pages to totalram_pages in order to
2161 * fix confusing memory reports from free(1) and another
2162 * side-effects, like CommitLimit going negative.
2163 */
2164 if (hstate_is_gigantic(h))
2165 adjust_managed_page_count(page, 1 << h->order);
2166 }
2167}
2168
2169static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2170{
2171 unsigned long i;
2172
2173 for (i = 0; i < h->max_huge_pages; ++i) {
2174 if (hstate_is_gigantic(h)) {
2175 if (!alloc_bootmem_huge_page(h))
2176 break;
2177 } else if (!alloc_pool_huge_page(h,
2178 &node_states[N_MEMORY]))
2179 break;
2180 cond_resched();
2181 }
2182 if (i < h->max_huge_pages) {
2183 char buf[32];
2184
2185 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2186 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2187 h->max_huge_pages, buf, i);
2188 h->max_huge_pages = i;
2189 }
2190}
2191
2192static void __init hugetlb_init_hstates(void)
2193{
2194 struct hstate *h;
2195
2196 for_each_hstate(h) {
2197 if (minimum_order > huge_page_order(h))
2198 minimum_order = huge_page_order(h);
2199
2200 /* oversize hugepages were init'ed in early boot */
2201 if (!hstate_is_gigantic(h))
2202 hugetlb_hstate_alloc_pages(h);
2203 }
2204 VM_BUG_ON(minimum_order == UINT_MAX);
2205}
2206
2207static void __init report_hugepages(void)
2208{
2209 struct hstate *h;
2210
2211 for_each_hstate(h) {
2212 char buf[32];
2213
2214 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2215 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2216 buf, h->free_huge_pages);
2217 }
2218}
2219
2220#ifdef CONFIG_HIGHMEM
2221static void try_to_free_low(struct hstate *h, unsigned long count,
2222 nodemask_t *nodes_allowed)
2223{
2224 int i;
2225
2226 if (hstate_is_gigantic(h))
2227 return;
2228
2229 for_each_node_mask(i, *nodes_allowed) {
2230 struct page *page, *next;
2231 struct list_head *freel = &h->hugepage_freelists[i];
2232 list_for_each_entry_safe(page, next, freel, lru) {
2233 if (count >= h->nr_huge_pages)
2234 return;
2235 if (PageHighMem(page))
2236 continue;
2237 list_del(&page->lru);
2238 update_and_free_page(h, page);
2239 h->free_huge_pages--;
2240 h->free_huge_pages_node[page_to_nid(page)]--;
2241 }
2242 }
2243}
2244#else
2245static inline void try_to_free_low(struct hstate *h, unsigned long count,
2246 nodemask_t *nodes_allowed)
2247{
2248}
2249#endif
2250
2251/*
2252 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2253 * balanced by operating on them in a round-robin fashion.
2254 * Returns 1 if an adjustment was made.
2255 */
2256static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2257 int delta)
2258{
2259 int nr_nodes, node;
2260
2261 VM_BUG_ON(delta != -1 && delta != 1);
2262
2263 if (delta < 0) {
2264 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2265 if (h->surplus_huge_pages_node[node])
2266 goto found;
2267 }
2268 } else {
2269 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2270 if (h->surplus_huge_pages_node[node] <
2271 h->nr_huge_pages_node[node])
2272 goto found;
2273 }
2274 }
2275 return 0;
2276
2277found:
2278 h->surplus_huge_pages += delta;
2279 h->surplus_huge_pages_node[node] += delta;
2280 return 1;
2281}
2282
2283#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2284static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2285 nodemask_t *nodes_allowed)
2286{
2287 unsigned long min_count, ret;
2288
2289 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2290 return h->max_huge_pages;
2291
2292 /*
2293 * Increase the pool size
2294 * First take pages out of surplus state. Then make up the
2295 * remaining difference by allocating fresh huge pages.
2296 *
2297 * We might race with alloc_surplus_huge_page() here and be unable
2298 * to convert a surplus huge page to a normal huge page. That is
2299 * not critical, though, it just means the overall size of the
2300 * pool might be one hugepage larger than it needs to be, but
2301 * within all the constraints specified by the sysctls.
2302 */
2303 spin_lock(&hugetlb_lock);
2304 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2305 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2306 break;
2307 }
2308
2309 while (count > persistent_huge_pages(h)) {
2310 /*
2311 * If this allocation races such that we no longer need the
2312 * page, free_huge_page will handle it by freeing the page
2313 * and reducing the surplus.
2314 */
2315 spin_unlock(&hugetlb_lock);
2316
2317 /* yield cpu to avoid soft lockup */
2318 cond_resched();
2319
2320 ret = alloc_pool_huge_page(h, nodes_allowed);
2321 spin_lock(&hugetlb_lock);
2322 if (!ret)
2323 goto out;
2324
2325 /* Bail for signals. Probably ctrl-c from user */
2326 if (signal_pending(current))
2327 goto out;
2328 }
2329
2330 /*
2331 * Decrease the pool size
2332 * First return free pages to the buddy allocator (being careful
2333 * to keep enough around to satisfy reservations). Then place
2334 * pages into surplus state as needed so the pool will shrink
2335 * to the desired size as pages become free.
2336 *
2337 * By placing pages into the surplus state independent of the
2338 * overcommit value, we are allowing the surplus pool size to
2339 * exceed overcommit. There are few sane options here. Since
2340 * alloc_surplus_huge_page() is checking the global counter,
2341 * though, we'll note that we're not allowed to exceed surplus
2342 * and won't grow the pool anywhere else. Not until one of the
2343 * sysctls are changed, or the surplus pages go out of use.
2344 */
2345 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2346 min_count = max(count, min_count);
2347 try_to_free_low(h, min_count, nodes_allowed);
2348 while (min_count < persistent_huge_pages(h)) {
2349 if (!free_pool_huge_page(h, nodes_allowed, 0))
2350 break;
2351 cond_resched_lock(&hugetlb_lock);
2352 }
2353 while (count < persistent_huge_pages(h)) {
2354 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2355 break;
2356 }
2357out:
2358 ret = persistent_huge_pages(h);
2359 spin_unlock(&hugetlb_lock);
2360 return ret;
2361}
2362
2363#define HSTATE_ATTR_RO(_name) \
2364 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2365
2366#define HSTATE_ATTR(_name) \
2367 static struct kobj_attribute _name##_attr = \
2368 __ATTR(_name, 0644, _name##_show, _name##_store)
2369
2370static struct kobject *hugepages_kobj;
2371static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2372
2373static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2374
2375static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2376{
2377 int i;
2378
2379 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2380 if (hstate_kobjs[i] == kobj) {
2381 if (nidp)
2382 *nidp = NUMA_NO_NODE;
2383 return &hstates[i];
2384 }
2385
2386 return kobj_to_node_hstate(kobj, nidp);
2387}
2388
2389static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2390 struct kobj_attribute *attr, char *buf)
2391{
2392 struct hstate *h;
2393 unsigned long nr_huge_pages;
2394 int nid;
2395
2396 h = kobj_to_hstate(kobj, &nid);
2397 if (nid == NUMA_NO_NODE)
2398 nr_huge_pages = h->nr_huge_pages;
2399 else
2400 nr_huge_pages = h->nr_huge_pages_node[nid];
2401
2402 return sprintf(buf, "%lu\n", nr_huge_pages);
2403}
2404
2405static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2406 struct hstate *h, int nid,
2407 unsigned long count, size_t len)
2408{
2409 int err;
2410 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2411
2412 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2413 err = -EINVAL;
2414 goto out;
2415 }
2416
2417 if (nid == NUMA_NO_NODE) {
2418 /*
2419 * global hstate attribute
2420 */
2421 if (!(obey_mempolicy &&
2422 init_nodemask_of_mempolicy(nodes_allowed))) {
2423 NODEMASK_FREE(nodes_allowed);
2424 nodes_allowed = &node_states[N_MEMORY];
2425 }
2426 } else if (nodes_allowed) {
2427 /*
2428 * per node hstate attribute: adjust count to global,
2429 * but restrict alloc/free to the specified node.
2430 */
2431 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2432 init_nodemask_of_node(nodes_allowed, nid);
2433 } else
2434 nodes_allowed = &node_states[N_MEMORY];
2435
2436 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2437
2438 if (nodes_allowed != &node_states[N_MEMORY])
2439 NODEMASK_FREE(nodes_allowed);
2440
2441 return len;
2442out:
2443 NODEMASK_FREE(nodes_allowed);
2444 return err;
2445}
2446
2447static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2448 struct kobject *kobj, const char *buf,
2449 size_t len)
2450{
2451 struct hstate *h;
2452 unsigned long count;
2453 int nid;
2454 int err;
2455
2456 err = kstrtoul(buf, 10, &count);
2457 if (err)
2458 return err;
2459
2460 h = kobj_to_hstate(kobj, &nid);
2461 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2462}
2463
2464static ssize_t nr_hugepages_show(struct kobject *kobj,
2465 struct kobj_attribute *attr, char *buf)
2466{
2467 return nr_hugepages_show_common(kobj, attr, buf);
2468}
2469
2470static ssize_t nr_hugepages_store(struct kobject *kobj,
2471 struct kobj_attribute *attr, const char *buf, size_t len)
2472{
2473 return nr_hugepages_store_common(false, kobj, buf, len);
2474}
2475HSTATE_ATTR(nr_hugepages);
2476
2477#ifdef CONFIG_NUMA
2478
2479/*
2480 * hstate attribute for optionally mempolicy-based constraint on persistent
2481 * huge page alloc/free.
2482 */
2483static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2484 struct kobj_attribute *attr, char *buf)
2485{
2486 return nr_hugepages_show_common(kobj, attr, buf);
2487}
2488
2489static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2490 struct kobj_attribute *attr, const char *buf, size_t len)
2491{
2492 return nr_hugepages_store_common(true, kobj, buf, len);
2493}
2494HSTATE_ATTR(nr_hugepages_mempolicy);
2495#endif
2496
2497
2498static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2499 struct kobj_attribute *attr, char *buf)
2500{
2501 struct hstate *h = kobj_to_hstate(kobj, NULL);
2502 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2503}
2504
2505static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2506 struct kobj_attribute *attr, const char *buf, size_t count)
2507{
2508 int err;
2509 unsigned long input;
2510 struct hstate *h = kobj_to_hstate(kobj, NULL);
2511
2512 if (hstate_is_gigantic(h))
2513 return -EINVAL;
2514
2515 err = kstrtoul(buf, 10, &input);
2516 if (err)
2517 return err;
2518
2519 spin_lock(&hugetlb_lock);
2520 h->nr_overcommit_huge_pages = input;
2521 spin_unlock(&hugetlb_lock);
2522
2523 return count;
2524}
2525HSTATE_ATTR(nr_overcommit_hugepages);
2526
2527static ssize_t free_hugepages_show(struct kobject *kobj,
2528 struct kobj_attribute *attr, char *buf)
2529{
2530 struct hstate *h;
2531 unsigned long free_huge_pages;
2532 int nid;
2533
2534 h = kobj_to_hstate(kobj, &nid);
2535 if (nid == NUMA_NO_NODE)
2536 free_huge_pages = h->free_huge_pages;
2537 else
2538 free_huge_pages = h->free_huge_pages_node[nid];
2539
2540 return sprintf(buf, "%lu\n", free_huge_pages);
2541}
2542HSTATE_ATTR_RO(free_hugepages);
2543
2544static ssize_t resv_hugepages_show(struct kobject *kobj,
2545 struct kobj_attribute *attr, char *buf)
2546{
2547 struct hstate *h = kobj_to_hstate(kobj, NULL);
2548 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2549}
2550HSTATE_ATTR_RO(resv_hugepages);
2551
2552static ssize_t surplus_hugepages_show(struct kobject *kobj,
2553 struct kobj_attribute *attr, char *buf)
2554{
2555 struct hstate *h;
2556 unsigned long surplus_huge_pages;
2557 int nid;
2558
2559 h = kobj_to_hstate(kobj, &nid);
2560 if (nid == NUMA_NO_NODE)
2561 surplus_huge_pages = h->surplus_huge_pages;
2562 else
2563 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2564
2565 return sprintf(buf, "%lu\n", surplus_huge_pages);
2566}
2567HSTATE_ATTR_RO(surplus_hugepages);
2568
2569static struct attribute *hstate_attrs[] = {
2570 &nr_hugepages_attr.attr,
2571 &nr_overcommit_hugepages_attr.attr,
2572 &free_hugepages_attr.attr,
2573 &resv_hugepages_attr.attr,
2574 &surplus_hugepages_attr.attr,
2575#ifdef CONFIG_NUMA
2576 &nr_hugepages_mempolicy_attr.attr,
2577#endif
2578 NULL,
2579};
2580
2581static const struct attribute_group hstate_attr_group = {
2582 .attrs = hstate_attrs,
2583};
2584
2585static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2586 struct kobject **hstate_kobjs,
2587 const struct attribute_group *hstate_attr_group)
2588{
2589 int retval;
2590 int hi = hstate_index(h);
2591
2592 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2593 if (!hstate_kobjs[hi])
2594 return -ENOMEM;
2595
2596 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2597 if (retval)
2598 kobject_put(hstate_kobjs[hi]);
2599
2600 return retval;
2601}
2602
2603static void __init hugetlb_sysfs_init(void)
2604{
2605 struct hstate *h;
2606 int err;
2607
2608 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2609 if (!hugepages_kobj)
2610 return;
2611
2612 for_each_hstate(h) {
2613 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2614 hstate_kobjs, &hstate_attr_group);
2615 if (err)
2616 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2617 }
2618}
2619
2620#ifdef CONFIG_NUMA
2621
2622/*
2623 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2624 * with node devices in node_devices[] using a parallel array. The array
2625 * index of a node device or _hstate == node id.
2626 * This is here to avoid any static dependency of the node device driver, in
2627 * the base kernel, on the hugetlb module.
2628 */
2629struct node_hstate {
2630 struct kobject *hugepages_kobj;
2631 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2632};
2633static struct node_hstate node_hstates[MAX_NUMNODES];
2634
2635/*
2636 * A subset of global hstate attributes for node devices
2637 */
2638static struct attribute *per_node_hstate_attrs[] = {
2639 &nr_hugepages_attr.attr,
2640 &free_hugepages_attr.attr,
2641 &surplus_hugepages_attr.attr,
2642 NULL,
2643};
2644
2645static const struct attribute_group per_node_hstate_attr_group = {
2646 .attrs = per_node_hstate_attrs,
2647};
2648
2649/*
2650 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2651 * Returns node id via non-NULL nidp.
2652 */
2653static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2654{
2655 int nid;
2656
2657 for (nid = 0; nid < nr_node_ids; nid++) {
2658 struct node_hstate *nhs = &node_hstates[nid];
2659 int i;
2660 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2661 if (nhs->hstate_kobjs[i] == kobj) {
2662 if (nidp)
2663 *nidp = nid;
2664 return &hstates[i];
2665 }
2666 }
2667
2668 BUG();
2669 return NULL;
2670}
2671
2672/*
2673 * Unregister hstate attributes from a single node device.
2674 * No-op if no hstate attributes attached.
2675 */
2676static void hugetlb_unregister_node(struct node *node)
2677{
2678 struct hstate *h;
2679 struct node_hstate *nhs = &node_hstates[node->dev.id];
2680
2681 if (!nhs->hugepages_kobj)
2682 return; /* no hstate attributes */
2683
2684 for_each_hstate(h) {
2685 int idx = hstate_index(h);
2686 if (nhs->hstate_kobjs[idx]) {
2687 kobject_put(nhs->hstate_kobjs[idx]);
2688 nhs->hstate_kobjs[idx] = NULL;
2689 }
2690 }
2691
2692 kobject_put(nhs->hugepages_kobj);
2693 nhs->hugepages_kobj = NULL;
2694}
2695
2696
2697/*
2698 * Register hstate attributes for a single node device.
2699 * No-op if attributes already registered.
2700 */
2701static void hugetlb_register_node(struct node *node)
2702{
2703 struct hstate *h;
2704 struct node_hstate *nhs = &node_hstates[node->dev.id];
2705 int err;
2706
2707 if (nhs->hugepages_kobj)
2708 return; /* already allocated */
2709
2710 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2711 &node->dev.kobj);
2712 if (!nhs->hugepages_kobj)
2713 return;
2714
2715 for_each_hstate(h) {
2716 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2717 nhs->hstate_kobjs,
2718 &per_node_hstate_attr_group);
2719 if (err) {
2720 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2721 h->name, node->dev.id);
2722 hugetlb_unregister_node(node);
2723 break;
2724 }
2725 }
2726}
2727
2728/*
2729 * hugetlb init time: register hstate attributes for all registered node
2730 * devices of nodes that have memory. All on-line nodes should have
2731 * registered their associated device by this time.
2732 */
2733static void __init hugetlb_register_all_nodes(void)
2734{
2735 int nid;
2736
2737 for_each_node_state(nid, N_MEMORY) {
2738 struct node *node = node_devices[nid];
2739 if (node->dev.id == nid)
2740 hugetlb_register_node(node);
2741 }
2742
2743 /*
2744 * Let the node device driver know we're here so it can
2745 * [un]register hstate attributes on node hotplug.
2746 */
2747 register_hugetlbfs_with_node(hugetlb_register_node,
2748 hugetlb_unregister_node);
2749}
2750#else /* !CONFIG_NUMA */
2751
2752static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2753{
2754 BUG();
2755 if (nidp)
2756 *nidp = -1;
2757 return NULL;
2758}
2759
2760static void hugetlb_register_all_nodes(void) { }
2761
2762#endif
2763
2764static int __init hugetlb_init(void)
2765{
2766 int i;
2767
2768 if (!hugepages_supported())
2769 return 0;
2770
2771 if (!size_to_hstate(default_hstate_size)) {
2772 if (default_hstate_size != 0) {
2773 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2774 default_hstate_size, HPAGE_SIZE);
2775 }
2776
2777 default_hstate_size = HPAGE_SIZE;
2778 if (!size_to_hstate(default_hstate_size))
2779 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2780 }
2781 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2782 if (default_hstate_max_huge_pages) {
2783 if (!default_hstate.max_huge_pages)
2784 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2785 }
2786
2787 hugetlb_init_hstates();
2788 gather_bootmem_prealloc();
2789 report_hugepages();
2790
2791 hugetlb_sysfs_init();
2792 hugetlb_register_all_nodes();
2793 hugetlb_cgroup_file_init();
2794
2795#ifdef CONFIG_SMP
2796 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2797#else
2798 num_fault_mutexes = 1;
2799#endif
2800 hugetlb_fault_mutex_table =
2801 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2802 BUG_ON(!hugetlb_fault_mutex_table);
2803
2804 for (i = 0; i < num_fault_mutexes; i++)
2805 mutex_init(&hugetlb_fault_mutex_table[i]);
2806 return 0;
2807}
2808subsys_initcall(hugetlb_init);
2809
2810/* Should be called on processing a hugepagesz=... option */
2811void __init hugetlb_bad_size(void)
2812{
2813 parsed_valid_hugepagesz = false;
2814}
2815
2816void __init hugetlb_add_hstate(unsigned int order)
2817{
2818 struct hstate *h;
2819 unsigned long i;
2820
2821 if (size_to_hstate(PAGE_SIZE << order)) {
2822 pr_warn("hugepagesz= specified twice, ignoring\n");
2823 return;
2824 }
2825 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2826 BUG_ON(order == 0);
2827 h = &hstates[hugetlb_max_hstate++];
2828 h->order = order;
2829 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2830 h->nr_huge_pages = 0;
2831 h->free_huge_pages = 0;
2832 for (i = 0; i < MAX_NUMNODES; ++i)
2833 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2834 INIT_LIST_HEAD(&h->hugepage_activelist);
2835 h->next_nid_to_alloc = first_memory_node;
2836 h->next_nid_to_free = first_memory_node;
2837 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2838 huge_page_size(h)/1024);
2839
2840 parsed_hstate = h;
2841}
2842
2843static int __init hugetlb_nrpages_setup(char *s)
2844{
2845 unsigned long *mhp;
2846 static unsigned long *last_mhp;
2847
2848 if (!parsed_valid_hugepagesz) {
2849 pr_warn("hugepages = %s preceded by "
2850 "an unsupported hugepagesz, ignoring\n", s);
2851 parsed_valid_hugepagesz = true;
2852 return 1;
2853 }
2854 /*
2855 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2856 * so this hugepages= parameter goes to the "default hstate".
2857 */
2858 else if (!hugetlb_max_hstate)
2859 mhp = &default_hstate_max_huge_pages;
2860 else
2861 mhp = &parsed_hstate->max_huge_pages;
2862
2863 if (mhp == last_mhp) {
2864 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2865 return 1;
2866 }
2867
2868 if (sscanf(s, "%lu", mhp) <= 0)
2869 *mhp = 0;
2870
2871 /*
2872 * Global state is always initialized later in hugetlb_init.
2873 * But we need to allocate >= MAX_ORDER hstates here early to still
2874 * use the bootmem allocator.
2875 */
2876 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2877 hugetlb_hstate_alloc_pages(parsed_hstate);
2878
2879 last_mhp = mhp;
2880
2881 return 1;
2882}
2883__setup("hugepages=", hugetlb_nrpages_setup);
2884
2885static int __init hugetlb_default_setup(char *s)
2886{
2887 default_hstate_size = memparse(s, &s);
2888 return 1;
2889}
2890__setup("default_hugepagesz=", hugetlb_default_setup);
2891
2892static unsigned int cpuset_mems_nr(unsigned int *array)
2893{
2894 int node;
2895 unsigned int nr = 0;
2896
2897 for_each_node_mask(node, cpuset_current_mems_allowed)
2898 nr += array[node];
2899
2900 return nr;
2901}
2902
2903#ifdef CONFIG_SYSCTL
2904static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2905 struct ctl_table *table, int write,
2906 void __user *buffer, size_t *length, loff_t *ppos)
2907{
2908 struct hstate *h = &default_hstate;
2909 unsigned long tmp = h->max_huge_pages;
2910 int ret;
2911
2912 if (!hugepages_supported())
2913 return -EOPNOTSUPP;
2914
2915 table->data = &tmp;
2916 table->maxlen = sizeof(unsigned long);
2917 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2918 if (ret)
2919 goto out;
2920
2921 if (write)
2922 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2923 NUMA_NO_NODE, tmp, *length);
2924out:
2925 return ret;
2926}
2927
2928int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2929 void __user *buffer, size_t *length, loff_t *ppos)
2930{
2931
2932 return hugetlb_sysctl_handler_common(false, table, write,
2933 buffer, length, ppos);
2934}
2935
2936#ifdef CONFIG_NUMA
2937int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2938 void __user *buffer, size_t *length, loff_t *ppos)
2939{
2940 return hugetlb_sysctl_handler_common(true, table, write,
2941 buffer, length, ppos);
2942}
2943#endif /* CONFIG_NUMA */
2944
2945int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2946 void __user *buffer,
2947 size_t *length, loff_t *ppos)
2948{
2949 struct hstate *h = &default_hstate;
2950 unsigned long tmp;
2951 int ret;
2952
2953 if (!hugepages_supported())
2954 return -EOPNOTSUPP;
2955
2956 tmp = h->nr_overcommit_huge_pages;
2957
2958 if (write && hstate_is_gigantic(h))
2959 return -EINVAL;
2960
2961 table->data = &tmp;
2962 table->maxlen = sizeof(unsigned long);
2963 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2964 if (ret)
2965 goto out;
2966
2967 if (write) {
2968 spin_lock(&hugetlb_lock);
2969 h->nr_overcommit_huge_pages = tmp;
2970 spin_unlock(&hugetlb_lock);
2971 }
2972out:
2973 return ret;
2974}
2975
2976#endif /* CONFIG_SYSCTL */
2977
2978void hugetlb_report_meminfo(struct seq_file *m)
2979{
2980 struct hstate *h;
2981 unsigned long total = 0;
2982
2983 if (!hugepages_supported())
2984 return;
2985
2986 for_each_hstate(h) {
2987 unsigned long count = h->nr_huge_pages;
2988
2989 total += (PAGE_SIZE << huge_page_order(h)) * count;
2990
2991 if (h == &default_hstate)
2992 seq_printf(m,
2993 "HugePages_Total: %5lu\n"
2994 "HugePages_Free: %5lu\n"
2995 "HugePages_Rsvd: %5lu\n"
2996 "HugePages_Surp: %5lu\n"
2997 "Hugepagesize: %8lu kB\n",
2998 count,
2999 h->free_huge_pages,
3000 h->resv_huge_pages,
3001 h->surplus_huge_pages,
3002 (PAGE_SIZE << huge_page_order(h)) / 1024);
3003 }
3004
3005 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3006}
3007
3008int hugetlb_report_node_meminfo(int nid, char *buf)
3009{
3010 struct hstate *h = &default_hstate;
3011 if (!hugepages_supported())
3012 return 0;
3013 return sprintf(buf,
3014 "Node %d HugePages_Total: %5u\n"
3015 "Node %d HugePages_Free: %5u\n"
3016 "Node %d HugePages_Surp: %5u\n",
3017 nid, h->nr_huge_pages_node[nid],
3018 nid, h->free_huge_pages_node[nid],
3019 nid, h->surplus_huge_pages_node[nid]);
3020}
3021
3022void hugetlb_show_meminfo(void)
3023{
3024 struct hstate *h;
3025 int nid;
3026
3027 if (!hugepages_supported())
3028 return;
3029
3030 for_each_node_state(nid, N_MEMORY)
3031 for_each_hstate(h)
3032 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3033 nid,
3034 h->nr_huge_pages_node[nid],
3035 h->free_huge_pages_node[nid],
3036 h->surplus_huge_pages_node[nid],
3037 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3038}
3039
3040void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3041{
3042 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3043 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3044}
3045
3046/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3047unsigned long hugetlb_total_pages(void)
3048{
3049 struct hstate *h;
3050 unsigned long nr_total_pages = 0;
3051
3052 for_each_hstate(h)
3053 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3054 return nr_total_pages;
3055}
3056
3057static int hugetlb_acct_memory(struct hstate *h, long delta)
3058{
3059 int ret = -ENOMEM;
3060
3061 spin_lock(&hugetlb_lock);
3062 /*
3063 * When cpuset is configured, it breaks the strict hugetlb page
3064 * reservation as the accounting is done on a global variable. Such
3065 * reservation is completely rubbish in the presence of cpuset because
3066 * the reservation is not checked against page availability for the
3067 * current cpuset. Application can still potentially OOM'ed by kernel
3068 * with lack of free htlb page in cpuset that the task is in.
3069 * Attempt to enforce strict accounting with cpuset is almost
3070 * impossible (or too ugly) because cpuset is too fluid that
3071 * task or memory node can be dynamically moved between cpusets.
3072 *
3073 * The change of semantics for shared hugetlb mapping with cpuset is
3074 * undesirable. However, in order to preserve some of the semantics,
3075 * we fall back to check against current free page availability as
3076 * a best attempt and hopefully to minimize the impact of changing
3077 * semantics that cpuset has.
3078 */
3079 if (delta > 0) {
3080 if (gather_surplus_pages(h, delta) < 0)
3081 goto out;
3082
3083 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3084 return_unused_surplus_pages(h, delta);
3085 goto out;
3086 }
3087 }
3088
3089 ret = 0;
3090 if (delta < 0)
3091 return_unused_surplus_pages(h, (unsigned long) -delta);
3092
3093out:
3094 spin_unlock(&hugetlb_lock);
3095 return ret;
3096}
3097
3098static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3099{
3100 struct resv_map *resv = vma_resv_map(vma);
3101
3102 /*
3103 * This new VMA should share its siblings reservation map if present.
3104 * The VMA will only ever have a valid reservation map pointer where
3105 * it is being copied for another still existing VMA. As that VMA
3106 * has a reference to the reservation map it cannot disappear until
3107 * after this open call completes. It is therefore safe to take a
3108 * new reference here without additional locking.
3109 */
3110 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3111 kref_get(&resv->refs);
3112}
3113
3114static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3115{
3116 struct hstate *h = hstate_vma(vma);
3117 struct resv_map *resv = vma_resv_map(vma);
3118 struct hugepage_subpool *spool = subpool_vma(vma);
3119 unsigned long reserve, start, end;
3120 long gbl_reserve;
3121
3122 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3123 return;
3124
3125 start = vma_hugecache_offset(h, vma, vma->vm_start);
3126 end = vma_hugecache_offset(h, vma, vma->vm_end);
3127
3128 reserve = (end - start) - region_count(resv, start, end);
3129
3130 kref_put(&resv->refs, resv_map_release);
3131
3132 if (reserve) {
3133 /*
3134 * Decrement reserve counts. The global reserve count may be
3135 * adjusted if the subpool has a minimum size.
3136 */
3137 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3138 hugetlb_acct_memory(h, -gbl_reserve);
3139 }
3140}
3141
3142static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3143{
3144 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3145 return -EINVAL;
3146 return 0;
3147}
3148
3149static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3150{
3151 struct hstate *hstate = hstate_vma(vma);
3152
3153 return 1UL << huge_page_shift(hstate);
3154}
3155
3156/*
3157 * We cannot handle pagefaults against hugetlb pages at all. They cause
3158 * handle_mm_fault() to try to instantiate regular-sized pages in the
3159 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3160 * this far.
3161 */
3162static int hugetlb_vm_op_fault(struct vm_fault *vmf)
3163{
3164 BUG();
3165 return 0;
3166}
3167
3168const struct vm_operations_struct hugetlb_vm_ops = {
3169 .fault = hugetlb_vm_op_fault,
3170 .open = hugetlb_vm_op_open,
3171 .close = hugetlb_vm_op_close,
3172 .split = hugetlb_vm_op_split,
3173 .pagesize = hugetlb_vm_op_pagesize,
3174};
3175
3176static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3177 int writable)
3178{
3179 pte_t entry;
3180
3181 if (writable) {
3182 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3183 vma->vm_page_prot)));
3184 } else {
3185 entry = huge_pte_wrprotect(mk_huge_pte(page,
3186 vma->vm_page_prot));
3187 }
3188 entry = pte_mkyoung(entry);
3189 entry = pte_mkhuge(entry);
3190 entry = arch_make_huge_pte(entry, vma, page, writable);
3191
3192 return entry;
3193}
3194
3195static void set_huge_ptep_writable(struct vm_area_struct *vma,
3196 unsigned long address, pte_t *ptep)
3197{
3198 pte_t entry;
3199
3200 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3201 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3202 update_mmu_cache(vma, address, ptep);
3203}
3204
3205bool is_hugetlb_entry_migration(pte_t pte)
3206{
3207 swp_entry_t swp;
3208
3209 if (huge_pte_none(pte) || pte_present(pte))
3210 return false;
3211 swp = pte_to_swp_entry(pte);
3212 if (non_swap_entry(swp) && is_migration_entry(swp))
3213 return true;
3214 else
3215 return false;
3216}
3217
3218static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3219{
3220 swp_entry_t swp;
3221
3222 if (huge_pte_none(pte) || pte_present(pte))
3223 return 0;
3224 swp = pte_to_swp_entry(pte);
3225 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3226 return 1;
3227 else
3228 return 0;
3229}
3230
3231int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3232 struct vm_area_struct *vma)
3233{
3234 pte_t *src_pte, *dst_pte, entry;
3235 struct page *ptepage;
3236 unsigned long addr;
3237 int cow;
3238 struct hstate *h = hstate_vma(vma);
3239 unsigned long sz = huge_page_size(h);
3240 unsigned long mmun_start; /* For mmu_notifiers */
3241 unsigned long mmun_end; /* For mmu_notifiers */
3242 int ret = 0;
3243
3244 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3245
3246 mmun_start = vma->vm_start;
3247 mmun_end = vma->vm_end;
3248 if (cow)
3249 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3250
3251 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3252 spinlock_t *src_ptl, *dst_ptl;
3253 src_pte = huge_pte_offset(src, addr, sz);
3254 if (!src_pte)
3255 continue;
3256 dst_pte = huge_pte_alloc(dst, addr, sz);
3257 if (!dst_pte) {
3258 ret = -ENOMEM;
3259 break;
3260 }
3261
3262 /* If the pagetables are shared don't copy or take references */
3263 if (dst_pte == src_pte)
3264 continue;
3265
3266 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3267 src_ptl = huge_pte_lockptr(h, src, src_pte);
3268 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3269 entry = huge_ptep_get(src_pte);
3270 if (huge_pte_none(entry)) { /* skip none entry */
3271 ;
3272 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3273 is_hugetlb_entry_hwpoisoned(entry))) {
3274 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3275
3276 if (is_write_migration_entry(swp_entry) && cow) {
3277 /*
3278 * COW mappings require pages in both
3279 * parent and child to be set to read.
3280 */
3281 make_migration_entry_read(&swp_entry);
3282 entry = swp_entry_to_pte(swp_entry);
3283 set_huge_swap_pte_at(src, addr, src_pte,
3284 entry, sz);
3285 }
3286 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3287 } else {
3288 if (cow) {
3289 /*
3290 * No need to notify as we are downgrading page
3291 * table protection not changing it to point
3292 * to a new page.
3293 *
3294 * See Documentation/vm/mmu_notifier.txt
3295 */
3296 huge_ptep_set_wrprotect(src, addr, src_pte);
3297 }
3298 entry = huge_ptep_get(src_pte);
3299 ptepage = pte_page(entry);
3300 get_page(ptepage);
3301 page_dup_rmap(ptepage, true);
3302 set_huge_pte_at(dst, addr, dst_pte, entry);
3303 hugetlb_count_add(pages_per_huge_page(h), dst);
3304 }
3305 spin_unlock(src_ptl);
3306 spin_unlock(dst_ptl);
3307 }
3308
3309 if (cow)
3310 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3311
3312 return ret;
3313}
3314
3315void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3316 unsigned long start, unsigned long end,
3317 struct page *ref_page)
3318{
3319 struct mm_struct *mm = vma->vm_mm;
3320 unsigned long address;
3321 pte_t *ptep;
3322 pte_t pte;
3323 spinlock_t *ptl;
3324 struct page *page;
3325 struct hstate *h = hstate_vma(vma);
3326 unsigned long sz = huge_page_size(h);
3327 const unsigned long mmun_start = start; /* For mmu_notifiers */
3328 const unsigned long mmun_end = end; /* For mmu_notifiers */
3329
3330 WARN_ON(!is_vm_hugetlb_page(vma));
3331 BUG_ON(start & ~huge_page_mask(h));
3332 BUG_ON(end & ~huge_page_mask(h));
3333
3334 /*
3335 * This is a hugetlb vma, all the pte entries should point
3336 * to huge page.
3337 */
3338 tlb_remove_check_page_size_change(tlb, sz);
3339 tlb_start_vma(tlb, vma);
3340 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3341 address = start;
3342 for (; address < end; address += sz) {
3343 ptep = huge_pte_offset(mm, address, sz);
3344 if (!ptep)
3345 continue;
3346
3347 ptl = huge_pte_lock(h, mm, ptep);
3348 if (huge_pmd_unshare(mm, &address, ptep)) {
3349 spin_unlock(ptl);
3350 continue;
3351 }
3352
3353 pte = huge_ptep_get(ptep);
3354 if (huge_pte_none(pte)) {
3355 spin_unlock(ptl);
3356 continue;
3357 }
3358
3359 /*
3360 * Migrating hugepage or HWPoisoned hugepage is already
3361 * unmapped and its refcount is dropped, so just clear pte here.
3362 */
3363 if (unlikely(!pte_present(pte))) {
3364 huge_pte_clear(mm, address, ptep, sz);
3365 spin_unlock(ptl);
3366 continue;
3367 }
3368
3369 page = pte_page(pte);
3370 /*
3371 * If a reference page is supplied, it is because a specific
3372 * page is being unmapped, not a range. Ensure the page we
3373 * are about to unmap is the actual page of interest.
3374 */
3375 if (ref_page) {
3376 if (page != ref_page) {
3377 spin_unlock(ptl);
3378 continue;
3379 }
3380 /*
3381 * Mark the VMA as having unmapped its page so that
3382 * future faults in this VMA will fail rather than
3383 * looking like data was lost
3384 */
3385 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3386 }
3387
3388 pte = huge_ptep_get_and_clear(mm, address, ptep);
3389 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3390 if (huge_pte_dirty(pte))
3391 set_page_dirty(page);
3392
3393 hugetlb_count_sub(pages_per_huge_page(h), mm);
3394 page_remove_rmap(page, true);
3395
3396 spin_unlock(ptl);
3397 tlb_remove_page_size(tlb, page, huge_page_size(h));
3398 /*
3399 * Bail out after unmapping reference page if supplied
3400 */
3401 if (ref_page)
3402 break;
3403 }
3404 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3405 tlb_end_vma(tlb, vma);
3406}
3407
3408void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3409 struct vm_area_struct *vma, unsigned long start,
3410 unsigned long end, struct page *ref_page)
3411{
3412 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3413
3414 /*
3415 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3416 * test will fail on a vma being torn down, and not grab a page table
3417 * on its way out. We're lucky that the flag has such an appropriate
3418 * name, and can in fact be safely cleared here. We could clear it
3419 * before the __unmap_hugepage_range above, but all that's necessary
3420 * is to clear it before releasing the i_mmap_rwsem. This works
3421 * because in the context this is called, the VMA is about to be
3422 * destroyed and the i_mmap_rwsem is held.
3423 */
3424 vma->vm_flags &= ~VM_MAYSHARE;
3425}
3426
3427void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3428 unsigned long end, struct page *ref_page)
3429{
3430 struct mm_struct *mm;
3431 struct mmu_gather tlb;
3432
3433 mm = vma->vm_mm;
3434
3435 tlb_gather_mmu(&tlb, mm, start, end);
3436 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3437 tlb_finish_mmu(&tlb, start, end);
3438}
3439
3440/*
3441 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3442 * mappping it owns the reserve page for. The intention is to unmap the page
3443 * from other VMAs and let the children be SIGKILLed if they are faulting the
3444 * same region.
3445 */
3446static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3447 struct page *page, unsigned long address)
3448{
3449 struct hstate *h = hstate_vma(vma);
3450 struct vm_area_struct *iter_vma;
3451 struct address_space *mapping;
3452 pgoff_t pgoff;
3453
3454 /*
3455 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3456 * from page cache lookup which is in HPAGE_SIZE units.
3457 */
3458 address = address & huge_page_mask(h);
3459 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3460 vma->vm_pgoff;
3461 mapping = vma->vm_file->f_mapping;
3462
3463 /*
3464 * Take the mapping lock for the duration of the table walk. As
3465 * this mapping should be shared between all the VMAs,
3466 * __unmap_hugepage_range() is called as the lock is already held
3467 */
3468 i_mmap_lock_write(mapping);
3469 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3470 /* Do not unmap the current VMA */
3471 if (iter_vma == vma)
3472 continue;
3473
3474 /*
3475 * Shared VMAs have their own reserves and do not affect
3476 * MAP_PRIVATE accounting but it is possible that a shared
3477 * VMA is using the same page so check and skip such VMAs.
3478 */
3479 if (iter_vma->vm_flags & VM_MAYSHARE)
3480 continue;
3481
3482 /*
3483 * Unmap the page from other VMAs without their own reserves.
3484 * They get marked to be SIGKILLed if they fault in these
3485 * areas. This is because a future no-page fault on this VMA
3486 * could insert a zeroed page instead of the data existing
3487 * from the time of fork. This would look like data corruption
3488 */
3489 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3490 unmap_hugepage_range(iter_vma, address,
3491 address + huge_page_size(h), page);
3492 }
3493 i_mmap_unlock_write(mapping);
3494}
3495
3496/*
3497 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3498 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3499 * cannot race with other handlers or page migration.
3500 * Keep the pte_same checks anyway to make transition from the mutex easier.
3501 */
3502static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3503 unsigned long address, pte_t *ptep,
3504 struct page *pagecache_page, spinlock_t *ptl)
3505{
3506 pte_t pte;
3507 struct hstate *h = hstate_vma(vma);
3508 struct page *old_page, *new_page;
3509 int ret = 0, outside_reserve = 0;
3510 unsigned long mmun_start; /* For mmu_notifiers */
3511 unsigned long mmun_end; /* For mmu_notifiers */
3512
3513 pte = huge_ptep_get(ptep);
3514 old_page = pte_page(pte);
3515
3516retry_avoidcopy:
3517 /* If no-one else is actually using this page, avoid the copy
3518 * and just make the page writable */
3519 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3520 page_move_anon_rmap(old_page, vma);
3521 set_huge_ptep_writable(vma, address, ptep);
3522 return 0;
3523 }
3524
3525 /*
3526 * If the process that created a MAP_PRIVATE mapping is about to
3527 * perform a COW due to a shared page count, attempt to satisfy
3528 * the allocation without using the existing reserves. The pagecache
3529 * page is used to determine if the reserve at this address was
3530 * consumed or not. If reserves were used, a partial faulted mapping
3531 * at the time of fork() could consume its reserves on COW instead
3532 * of the full address range.
3533 */
3534 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3535 old_page != pagecache_page)
3536 outside_reserve = 1;
3537
3538 get_page(old_page);
3539
3540 /*
3541 * Drop page table lock as buddy allocator may be called. It will
3542 * be acquired again before returning to the caller, as expected.
3543 */
3544 spin_unlock(ptl);
3545 new_page = alloc_huge_page(vma, address, outside_reserve);
3546
3547 if (IS_ERR(new_page)) {
3548 /*
3549 * If a process owning a MAP_PRIVATE mapping fails to COW,
3550 * it is due to references held by a child and an insufficient
3551 * huge page pool. To guarantee the original mappers
3552 * reliability, unmap the page from child processes. The child
3553 * may get SIGKILLed if it later faults.
3554 */
3555 if (outside_reserve) {
3556 put_page(old_page);
3557 BUG_ON(huge_pte_none(pte));
3558 unmap_ref_private(mm, vma, old_page, address);
3559 BUG_ON(huge_pte_none(pte));
3560 spin_lock(ptl);
3561 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3562 huge_page_size(h));
3563 if (likely(ptep &&
3564 pte_same(huge_ptep_get(ptep), pte)))
3565 goto retry_avoidcopy;
3566 /*
3567 * race occurs while re-acquiring page table
3568 * lock, and our job is done.
3569 */
3570 return 0;
3571 }
3572
3573 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3574 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3575 goto out_release_old;
3576 }
3577
3578 /*
3579 * When the original hugepage is shared one, it does not have
3580 * anon_vma prepared.
3581 */
3582 if (unlikely(anon_vma_prepare(vma))) {
3583 ret = VM_FAULT_OOM;
3584 goto out_release_all;
3585 }
3586
3587 copy_user_huge_page(new_page, old_page, address, vma,
3588 pages_per_huge_page(h));
3589 __SetPageUptodate(new_page);
3590 set_page_huge_active(new_page);
3591
3592 mmun_start = address & huge_page_mask(h);
3593 mmun_end = mmun_start + huge_page_size(h);
3594 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3595
3596 /*
3597 * Retake the page table lock to check for racing updates
3598 * before the page tables are altered
3599 */
3600 spin_lock(ptl);
3601 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3602 huge_page_size(h));
3603 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3604 ClearPagePrivate(new_page);
3605
3606 /* Break COW */
3607 huge_ptep_clear_flush(vma, address, ptep);
3608 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3609 set_huge_pte_at(mm, address, ptep,
3610 make_huge_pte(vma, new_page, 1));
3611 page_remove_rmap(old_page, true);
3612 hugepage_add_new_anon_rmap(new_page, vma, address);
3613 /* Make the old page be freed below */
3614 new_page = old_page;
3615 }
3616 spin_unlock(ptl);
3617 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3618out_release_all:
3619 restore_reserve_on_error(h, vma, address, new_page);
3620 put_page(new_page);
3621out_release_old:
3622 put_page(old_page);
3623
3624 spin_lock(ptl); /* Caller expects lock to be held */
3625 return ret;
3626}
3627
3628/* Return the pagecache page at a given address within a VMA */
3629static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3630 struct vm_area_struct *vma, unsigned long address)
3631{
3632 struct address_space *mapping;
3633 pgoff_t idx;
3634
3635 mapping = vma->vm_file->f_mapping;
3636 idx = vma_hugecache_offset(h, vma, address);
3637
3638 return find_lock_page(mapping, idx);
3639}
3640
3641/*
3642 * Return whether there is a pagecache page to back given address within VMA.
3643 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3644 */
3645static bool hugetlbfs_pagecache_present(struct hstate *h,
3646 struct vm_area_struct *vma, unsigned long address)
3647{
3648 struct address_space *mapping;
3649 pgoff_t idx;
3650 struct page *page;
3651
3652 mapping = vma->vm_file->f_mapping;
3653 idx = vma_hugecache_offset(h, vma, address);
3654
3655 page = find_get_page(mapping, idx);
3656 if (page)
3657 put_page(page);
3658 return page != NULL;
3659}
3660
3661int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3662 pgoff_t idx)
3663{
3664 struct inode *inode = mapping->host;
3665 struct hstate *h = hstate_inode(inode);
3666 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3667
3668 if (err)
3669 return err;
3670 ClearPagePrivate(page);
3671
3672 spin_lock(&inode->i_lock);
3673 inode->i_blocks += blocks_per_huge_page(h);
3674 spin_unlock(&inode->i_lock);
3675 return 0;
3676}
3677
3678static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3679 struct address_space *mapping, pgoff_t idx,
3680 unsigned long address, pte_t *ptep, unsigned int flags)
3681{
3682 struct hstate *h = hstate_vma(vma);
3683 int ret = VM_FAULT_SIGBUS;
3684 int anon_rmap = 0;
3685 unsigned long size;
3686 struct page *page;
3687 pte_t new_pte;
3688 spinlock_t *ptl;
3689
3690 /*
3691 * Currently, we are forced to kill the process in the event the
3692 * original mapper has unmapped pages from the child due to a failed
3693 * COW. Warn that such a situation has occurred as it may not be obvious
3694 */
3695 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3696 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3697 current->pid);
3698 return ret;
3699 }
3700
3701 /*
3702 * Use page lock to guard against racing truncation
3703 * before we get page_table_lock.
3704 */
3705retry:
3706 page = find_lock_page(mapping, idx);
3707 if (!page) {
3708 size = i_size_read(mapping->host) >> huge_page_shift(h);
3709 if (idx >= size)
3710 goto out;
3711
3712 /*
3713 * Check for page in userfault range
3714 */
3715 if (userfaultfd_missing(vma)) {
3716 u32 hash;
3717 struct vm_fault vmf = {
3718 .vma = vma,
3719 .address = address,
3720 .flags = flags,
3721 /*
3722 * Hard to debug if it ends up being
3723 * used by a callee that assumes
3724 * something about the other
3725 * uninitialized fields... same as in
3726 * memory.c
3727 */
3728 };
3729
3730 /*
3731 * hugetlb_fault_mutex must be dropped before
3732 * handling userfault. Reacquire after handling
3733 * fault to make calling code simpler.
3734 */
3735 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3736 idx, address);
3737 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3738 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3739 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3740 goto out;
3741 }
3742
3743 page = alloc_huge_page(vma, address, 0);
3744 if (IS_ERR(page)) {
3745 ret = PTR_ERR(page);
3746 if (ret == -ENOMEM)
3747 ret = VM_FAULT_OOM;
3748 else
3749 ret = VM_FAULT_SIGBUS;
3750 goto out;
3751 }
3752 clear_huge_page(page, address, pages_per_huge_page(h));
3753 __SetPageUptodate(page);
3754 set_page_huge_active(page);
3755
3756 if (vma->vm_flags & VM_MAYSHARE) {
3757 int err = huge_add_to_page_cache(page, mapping, idx);
3758 if (err) {
3759 put_page(page);
3760 if (err == -EEXIST)
3761 goto retry;
3762 goto out;
3763 }
3764 } else {
3765 lock_page(page);
3766 if (unlikely(anon_vma_prepare(vma))) {
3767 ret = VM_FAULT_OOM;
3768 goto backout_unlocked;
3769 }
3770 anon_rmap = 1;
3771 }
3772 } else {
3773 /*
3774 * If memory error occurs between mmap() and fault, some process
3775 * don't have hwpoisoned swap entry for errored virtual address.
3776 * So we need to block hugepage fault by PG_hwpoison bit check.
3777 */
3778 if (unlikely(PageHWPoison(page))) {
3779 ret = VM_FAULT_HWPOISON |
3780 VM_FAULT_SET_HINDEX(hstate_index(h));
3781 goto backout_unlocked;
3782 }
3783 }
3784
3785 /*
3786 * If we are going to COW a private mapping later, we examine the
3787 * pending reservations for this page now. This will ensure that
3788 * any allocations necessary to record that reservation occur outside
3789 * the spinlock.
3790 */
3791 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3792 if (vma_needs_reservation(h, vma, address) < 0) {
3793 ret = VM_FAULT_OOM;
3794 goto backout_unlocked;
3795 }
3796 /* Just decrements count, does not deallocate */
3797 vma_end_reservation(h, vma, address);
3798 }
3799
3800 ptl = huge_pte_lock(h, mm, ptep);
3801 size = i_size_read(mapping->host) >> huge_page_shift(h);
3802 if (idx >= size)
3803 goto backout;
3804
3805 ret = 0;
3806 if (!huge_pte_none(huge_ptep_get(ptep)))
3807 goto backout;
3808
3809 if (anon_rmap) {
3810 ClearPagePrivate(page);
3811 hugepage_add_new_anon_rmap(page, vma, address);
3812 } else
3813 page_dup_rmap(page, true);
3814 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3815 && (vma->vm_flags & VM_SHARED)));
3816 set_huge_pte_at(mm, address, ptep, new_pte);
3817
3818 hugetlb_count_add(pages_per_huge_page(h), mm);
3819 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3820 /* Optimization, do the COW without a second fault */
3821 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3822 }
3823
3824 spin_unlock(ptl);
3825 unlock_page(page);
3826out:
3827 return ret;
3828
3829backout:
3830 spin_unlock(ptl);
3831backout_unlocked:
3832 unlock_page(page);
3833 restore_reserve_on_error(h, vma, address, page);
3834 put_page(page);
3835 goto out;
3836}
3837
3838#ifdef CONFIG_SMP
3839u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3840 struct vm_area_struct *vma,
3841 struct address_space *mapping,
3842 pgoff_t idx, unsigned long address)
3843{
3844 unsigned long key[2];
3845 u32 hash;
3846
3847 if (vma->vm_flags & VM_SHARED) {
3848 key[0] = (unsigned long) mapping;
3849 key[1] = idx;
3850 } else {
3851 key[0] = (unsigned long) mm;
3852 key[1] = address >> huge_page_shift(h);
3853 }
3854
3855 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3856
3857 return hash & (num_fault_mutexes - 1);
3858}
3859#else
3860/*
3861 * For uniprocesor systems we always use a single mutex, so just
3862 * return 0 and avoid the hashing overhead.
3863 */
3864u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3865 struct vm_area_struct *vma,
3866 struct address_space *mapping,
3867 pgoff_t idx, unsigned long address)
3868{
3869 return 0;
3870}
3871#endif
3872
3873int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3874 unsigned long address, unsigned int flags)
3875{
3876 pte_t *ptep, entry;
3877 spinlock_t *ptl;
3878 int ret;
3879 u32 hash;
3880 pgoff_t idx;
3881 struct page *page = NULL;
3882 struct page *pagecache_page = NULL;
3883 struct hstate *h = hstate_vma(vma);
3884 struct address_space *mapping;
3885 int need_wait_lock = 0;
3886
3887 address &= huge_page_mask(h);
3888
3889 ptep = huge_pte_offset(mm, address, huge_page_size(h));
3890 if (ptep) {
3891 entry = huge_ptep_get(ptep);
3892 if (unlikely(is_hugetlb_entry_migration(entry))) {
3893 migration_entry_wait_huge(vma, mm, ptep);
3894 return 0;
3895 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3896 return VM_FAULT_HWPOISON_LARGE |
3897 VM_FAULT_SET_HINDEX(hstate_index(h));
3898 } else {
3899 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3900 if (!ptep)
3901 return VM_FAULT_OOM;
3902 }
3903
3904 mapping = vma->vm_file->f_mapping;
3905 idx = vma_hugecache_offset(h, vma, address);
3906
3907 /*
3908 * Serialize hugepage allocation and instantiation, so that we don't
3909 * get spurious allocation failures if two CPUs race to instantiate
3910 * the same page in the page cache.
3911 */
3912 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3913 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3914
3915 entry = huge_ptep_get(ptep);
3916 if (huge_pte_none(entry)) {
3917 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3918 goto out_mutex;
3919 }
3920
3921 ret = 0;
3922
3923 /*
3924 * entry could be a migration/hwpoison entry at this point, so this
3925 * check prevents the kernel from going below assuming that we have
3926 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3927 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3928 * handle it.
3929 */
3930 if (!pte_present(entry))
3931 goto out_mutex;
3932
3933 /*
3934 * If we are going to COW the mapping later, we examine the pending
3935 * reservations for this page now. This will ensure that any
3936 * allocations necessary to record that reservation occur outside the
3937 * spinlock. For private mappings, we also lookup the pagecache
3938 * page now as it is used to determine if a reservation has been
3939 * consumed.
3940 */
3941 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3942 if (vma_needs_reservation(h, vma, address) < 0) {
3943 ret = VM_FAULT_OOM;
3944 goto out_mutex;
3945 }
3946 /* Just decrements count, does not deallocate */
3947 vma_end_reservation(h, vma, address);
3948
3949 if (!(vma->vm_flags & VM_MAYSHARE))
3950 pagecache_page = hugetlbfs_pagecache_page(h,
3951 vma, address);
3952 }
3953
3954 ptl = huge_pte_lock(h, mm, ptep);
3955
3956 /* Check for a racing update before calling hugetlb_cow */
3957 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3958 goto out_ptl;
3959
3960 /*
3961 * hugetlb_cow() requires page locks of pte_page(entry) and
3962 * pagecache_page, so here we need take the former one
3963 * when page != pagecache_page or !pagecache_page.
3964 */
3965 page = pte_page(entry);
3966 if (page != pagecache_page)
3967 if (!trylock_page(page)) {
3968 need_wait_lock = 1;
3969 goto out_ptl;
3970 }
3971
3972 get_page(page);
3973
3974 if (flags & FAULT_FLAG_WRITE) {
3975 if (!huge_pte_write(entry)) {
3976 ret = hugetlb_cow(mm, vma, address, ptep,
3977 pagecache_page, ptl);
3978 goto out_put_page;
3979 }
3980 entry = huge_pte_mkdirty(entry);
3981 }
3982 entry = pte_mkyoung(entry);
3983 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3984 flags & FAULT_FLAG_WRITE))
3985 update_mmu_cache(vma, address, ptep);
3986out_put_page:
3987 if (page != pagecache_page)
3988 unlock_page(page);
3989 put_page(page);
3990out_ptl:
3991 spin_unlock(ptl);
3992
3993 if (pagecache_page) {
3994 unlock_page(pagecache_page);
3995 put_page(pagecache_page);
3996 }
3997out_mutex:
3998 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3999 /*
4000 * Generally it's safe to hold refcount during waiting page lock. But
4001 * here we just wait to defer the next page fault to avoid busy loop and
4002 * the page is not used after unlocked before returning from the current
4003 * page fault. So we are safe from accessing freed page, even if we wait
4004 * here without taking refcount.
4005 */
4006 if (need_wait_lock)
4007 wait_on_page_locked(page);
4008 return ret;
4009}
4010
4011/*
4012 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4013 * modifications for huge pages.
4014 */
4015int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4016 pte_t *dst_pte,
4017 struct vm_area_struct *dst_vma,
4018 unsigned long dst_addr,
4019 unsigned long src_addr,
4020 struct page **pagep)
4021{
4022 struct address_space *mapping;
4023 pgoff_t idx;
4024 unsigned long size;
4025 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4026 struct hstate *h = hstate_vma(dst_vma);
4027 pte_t _dst_pte;
4028 spinlock_t *ptl;
4029 int ret;
4030 struct page *page;
4031
4032 if (!*pagep) {
4033 ret = -ENOMEM;
4034 page = alloc_huge_page(dst_vma, dst_addr, 0);
4035 if (IS_ERR(page))
4036 goto out;
4037
4038 ret = copy_huge_page_from_user(page,
4039 (const void __user *) src_addr,
4040 pages_per_huge_page(h), false);
4041
4042 /* fallback to copy_from_user outside mmap_sem */
4043 if (unlikely(ret)) {
4044 ret = -EFAULT;
4045 *pagep = page;
4046 /* don't free the page */
4047 goto out;
4048 }
4049 } else {
4050 page = *pagep;
4051 *pagep = NULL;
4052 }
4053
4054 /*
4055 * The memory barrier inside __SetPageUptodate makes sure that
4056 * preceding stores to the page contents become visible before
4057 * the set_pte_at() write.
4058 */
4059 __SetPageUptodate(page);
4060 set_page_huge_active(page);
4061
4062 mapping = dst_vma->vm_file->f_mapping;
4063 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4064
4065 /*
4066 * If shared, add to page cache
4067 */
4068 if (vm_shared) {
4069 size = i_size_read(mapping->host) >> huge_page_shift(h);
4070 ret = -EFAULT;
4071 if (idx >= size)
4072 goto out_release_nounlock;
4073
4074 /*
4075 * Serialization between remove_inode_hugepages() and
4076 * huge_add_to_page_cache() below happens through the
4077 * hugetlb_fault_mutex_table that here must be hold by
4078 * the caller.
4079 */
4080 ret = huge_add_to_page_cache(page, mapping, idx);
4081 if (ret)
4082 goto out_release_nounlock;
4083 }
4084
4085 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4086 spin_lock(ptl);
4087
4088 /*
4089 * Recheck the i_size after holding PT lock to make sure not
4090 * to leave any page mapped (as page_mapped()) beyond the end
4091 * of the i_size (remove_inode_hugepages() is strict about
4092 * enforcing that). If we bail out here, we'll also leave a
4093 * page in the radix tree in the vm_shared case beyond the end
4094 * of the i_size, but remove_inode_hugepages() will take care
4095 * of it as soon as we drop the hugetlb_fault_mutex_table.
4096 */
4097 size = i_size_read(mapping->host) >> huge_page_shift(h);
4098 ret = -EFAULT;
4099 if (idx >= size)
4100 goto out_release_unlock;
4101
4102 ret = -EEXIST;
4103 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4104 goto out_release_unlock;
4105
4106 if (vm_shared) {
4107 page_dup_rmap(page, true);
4108 } else {
4109 ClearPagePrivate(page);
4110 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4111 }
4112
4113 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4114 if (dst_vma->vm_flags & VM_WRITE)
4115 _dst_pte = huge_pte_mkdirty(_dst_pte);
4116 _dst_pte = pte_mkyoung(_dst_pte);
4117
4118 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4119
4120 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4121 dst_vma->vm_flags & VM_WRITE);
4122 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4123
4124 /* No need to invalidate - it was non-present before */
4125 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4126
4127 spin_unlock(ptl);
4128 if (vm_shared)
4129 unlock_page(page);
4130 ret = 0;
4131out:
4132 return ret;
4133out_release_unlock:
4134 spin_unlock(ptl);
4135 if (vm_shared)
4136 unlock_page(page);
4137out_release_nounlock:
4138 put_page(page);
4139 goto out;
4140}
4141
4142long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4143 struct page **pages, struct vm_area_struct **vmas,
4144 unsigned long *position, unsigned long *nr_pages,
4145 long i, unsigned int flags, int *nonblocking)
4146{
4147 unsigned long pfn_offset;
4148 unsigned long vaddr = *position;
4149 unsigned long remainder = *nr_pages;
4150 struct hstate *h = hstate_vma(vma);
4151 int err = -EFAULT;
4152
4153 while (vaddr < vma->vm_end && remainder) {
4154 pte_t *pte;
4155 spinlock_t *ptl = NULL;
4156 int absent;
4157 struct page *page;
4158
4159 /*
4160 * If we have a pending SIGKILL, don't keep faulting pages and
4161 * potentially allocating memory.
4162 */
4163 if (unlikely(fatal_signal_pending(current))) {
4164 remainder = 0;
4165 break;
4166 }
4167
4168 /*
4169 * Some archs (sparc64, sh*) have multiple pte_ts to
4170 * each hugepage. We have to make sure we get the
4171 * first, for the page indexing below to work.
4172 *
4173 * Note that page table lock is not held when pte is null.
4174 */
4175 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4176 huge_page_size(h));
4177 if (pte)
4178 ptl = huge_pte_lock(h, mm, pte);
4179 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4180
4181 /*
4182 * When coredumping, it suits get_dump_page if we just return
4183 * an error where there's an empty slot with no huge pagecache
4184 * to back it. This way, we avoid allocating a hugepage, and
4185 * the sparse dumpfile avoids allocating disk blocks, but its
4186 * huge holes still show up with zeroes where they need to be.
4187 */
4188 if (absent && (flags & FOLL_DUMP) &&
4189 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4190 if (pte)
4191 spin_unlock(ptl);
4192 remainder = 0;
4193 break;
4194 }
4195
4196 /*
4197 * We need call hugetlb_fault for both hugepages under migration
4198 * (in which case hugetlb_fault waits for the migration,) and
4199 * hwpoisoned hugepages (in which case we need to prevent the
4200 * caller from accessing to them.) In order to do this, we use
4201 * here is_swap_pte instead of is_hugetlb_entry_migration and
4202 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4203 * both cases, and because we can't follow correct pages
4204 * directly from any kind of swap entries.
4205 */
4206 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4207 ((flags & FOLL_WRITE) &&
4208 !huge_pte_write(huge_ptep_get(pte)))) {
4209 int ret;
4210 unsigned int fault_flags = 0;
4211
4212 if (pte)
4213 spin_unlock(ptl);
4214 if (flags & FOLL_WRITE)
4215 fault_flags |= FAULT_FLAG_WRITE;
4216 if (nonblocking)
4217 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4218 if (flags & FOLL_NOWAIT)
4219 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4220 FAULT_FLAG_RETRY_NOWAIT;
4221 if (flags & FOLL_TRIED) {
4222 VM_WARN_ON_ONCE(fault_flags &
4223 FAULT_FLAG_ALLOW_RETRY);
4224 fault_flags |= FAULT_FLAG_TRIED;
4225 }
4226 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4227 if (ret & VM_FAULT_ERROR) {
4228 err = vm_fault_to_errno(ret, flags);
4229 remainder = 0;
4230 break;
4231 }
4232 if (ret & VM_FAULT_RETRY) {
4233 if (nonblocking)
4234 *nonblocking = 0;
4235 *nr_pages = 0;
4236 /*
4237 * VM_FAULT_RETRY must not return an
4238 * error, it will return zero
4239 * instead.
4240 *
4241 * No need to update "position" as the
4242 * caller will not check it after
4243 * *nr_pages is set to 0.
4244 */
4245 return i;
4246 }
4247 continue;
4248 }
4249
4250 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4251 page = pte_page(huge_ptep_get(pte));
4252same_page:
4253 if (pages) {
4254 pages[i] = mem_map_offset(page, pfn_offset);
4255 get_page(pages[i]);
4256 }
4257
4258 if (vmas)
4259 vmas[i] = vma;
4260
4261 vaddr += PAGE_SIZE;
4262 ++pfn_offset;
4263 --remainder;
4264 ++i;
4265 if (vaddr < vma->vm_end && remainder &&
4266 pfn_offset < pages_per_huge_page(h)) {
4267 /*
4268 * We use pfn_offset to avoid touching the pageframes
4269 * of this compound page.
4270 */
4271 goto same_page;
4272 }
4273 spin_unlock(ptl);
4274 }
4275 *nr_pages = remainder;
4276 /*
4277 * setting position is actually required only if remainder is
4278 * not zero but it's faster not to add a "if (remainder)"
4279 * branch.
4280 */
4281 *position = vaddr;
4282
4283 return i ? i : err;
4284}
4285
4286#ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4287/*
4288 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4289 * implement this.
4290 */
4291#define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4292#endif
4293
4294unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4295 unsigned long address, unsigned long end, pgprot_t newprot)
4296{
4297 struct mm_struct *mm = vma->vm_mm;
4298 unsigned long start = address;
4299 pte_t *ptep;
4300 pte_t pte;
4301 struct hstate *h = hstate_vma(vma);
4302 unsigned long pages = 0;
4303
4304 BUG_ON(address >= end);
4305 flush_cache_range(vma, address, end);
4306
4307 mmu_notifier_invalidate_range_start(mm, start, end);
4308 i_mmap_lock_write(vma->vm_file->f_mapping);
4309 for (; address < end; address += huge_page_size(h)) {
4310 spinlock_t *ptl;
4311 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4312 if (!ptep)
4313 continue;
4314 ptl = huge_pte_lock(h, mm, ptep);
4315 if (huge_pmd_unshare(mm, &address, ptep)) {
4316 pages++;
4317 spin_unlock(ptl);
4318 continue;
4319 }
4320 pte = huge_ptep_get(ptep);
4321 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4322 spin_unlock(ptl);
4323 continue;
4324 }
4325 if (unlikely(is_hugetlb_entry_migration(pte))) {
4326 swp_entry_t entry = pte_to_swp_entry(pte);
4327
4328 if (is_write_migration_entry(entry)) {
4329 pte_t newpte;
4330
4331 make_migration_entry_read(&entry);
4332 newpte = swp_entry_to_pte(entry);
4333 set_huge_swap_pte_at(mm, address, ptep,
4334 newpte, huge_page_size(h));
4335 pages++;
4336 }
4337 spin_unlock(ptl);
4338 continue;
4339 }
4340 if (!huge_pte_none(pte)) {
4341 pte = huge_ptep_get_and_clear(mm, address, ptep);
4342 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4343 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4344 set_huge_pte_at(mm, address, ptep, pte);
4345 pages++;
4346 }
4347 spin_unlock(ptl);
4348 }
4349 /*
4350 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4351 * may have cleared our pud entry and done put_page on the page table:
4352 * once we release i_mmap_rwsem, another task can do the final put_page
4353 * and that page table be reused and filled with junk.
4354 */
4355 flush_hugetlb_tlb_range(vma, start, end);
4356 /*
4357 * No need to call mmu_notifier_invalidate_range() we are downgrading
4358 * page table protection not changing it to point to a new page.
4359 *
4360 * See Documentation/vm/mmu_notifier.txt
4361 */
4362 i_mmap_unlock_write(vma->vm_file->f_mapping);
4363 mmu_notifier_invalidate_range_end(mm, start, end);
4364
4365 return pages << h->order;
4366}
4367
4368int hugetlb_reserve_pages(struct inode *inode,
4369 long from, long to,
4370 struct vm_area_struct *vma,
4371 vm_flags_t vm_flags)
4372{
4373 long ret, chg;
4374 struct hstate *h = hstate_inode(inode);
4375 struct hugepage_subpool *spool = subpool_inode(inode);
4376 struct resv_map *resv_map;
4377 long gbl_reserve;
4378
4379 /* This should never happen */
4380 if (from > to) {
4381 VM_WARN(1, "%s called with a negative range\n", __func__);
4382 return -EINVAL;
4383 }
4384
4385 /*
4386 * Only apply hugepage reservation if asked. At fault time, an
4387 * attempt will be made for VM_NORESERVE to allocate a page
4388 * without using reserves
4389 */
4390 if (vm_flags & VM_NORESERVE)
4391 return 0;
4392
4393 /*
4394 * Shared mappings base their reservation on the number of pages that
4395 * are already allocated on behalf of the file. Private mappings need
4396 * to reserve the full area even if read-only as mprotect() may be
4397 * called to make the mapping read-write. Assume !vma is a shm mapping
4398 */
4399 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4400 resv_map = inode_resv_map(inode);
4401
4402 chg = region_chg(resv_map, from, to);
4403
4404 } else {
4405 resv_map = resv_map_alloc();
4406 if (!resv_map)
4407 return -ENOMEM;
4408
4409 chg = to - from;
4410
4411 set_vma_resv_map(vma, resv_map);
4412 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4413 }
4414
4415 if (chg < 0) {
4416 ret = chg;
4417 goto out_err;
4418 }
4419
4420 /*
4421 * There must be enough pages in the subpool for the mapping. If
4422 * the subpool has a minimum size, there may be some global
4423 * reservations already in place (gbl_reserve).
4424 */
4425 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4426 if (gbl_reserve < 0) {
4427 ret = -ENOSPC;
4428 goto out_err;
4429 }
4430
4431 /*
4432 * Check enough hugepages are available for the reservation.
4433 * Hand the pages back to the subpool if there are not
4434 */
4435 ret = hugetlb_acct_memory(h, gbl_reserve);
4436 if (ret < 0) {
4437 /* put back original number of pages, chg */
4438 (void)hugepage_subpool_put_pages(spool, chg);
4439 goto out_err;
4440 }
4441
4442 /*
4443 * Account for the reservations made. Shared mappings record regions
4444 * that have reservations as they are shared by multiple VMAs.
4445 * When the last VMA disappears, the region map says how much
4446 * the reservation was and the page cache tells how much of
4447 * the reservation was consumed. Private mappings are per-VMA and
4448 * only the consumed reservations are tracked. When the VMA
4449 * disappears, the original reservation is the VMA size and the
4450 * consumed reservations are stored in the map. Hence, nothing
4451 * else has to be done for private mappings here
4452 */
4453 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4454 long add = region_add(resv_map, from, to);
4455
4456 if (unlikely(chg > add)) {
4457 /*
4458 * pages in this range were added to the reserve
4459 * map between region_chg and region_add. This
4460 * indicates a race with alloc_huge_page. Adjust
4461 * the subpool and reserve counts modified above
4462 * based on the difference.
4463 */
4464 long rsv_adjust;
4465
4466 rsv_adjust = hugepage_subpool_put_pages(spool,
4467 chg - add);
4468 hugetlb_acct_memory(h, -rsv_adjust);
4469 }
4470 }
4471 return 0;
4472out_err:
4473 if (!vma || vma->vm_flags & VM_MAYSHARE)
4474 /* Don't call region_abort if region_chg failed */
4475 if (chg >= 0)
4476 region_abort(resv_map, from, to);
4477 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4478 kref_put(&resv_map->refs, resv_map_release);
4479 return ret;
4480}
4481
4482long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4483 long freed)
4484{
4485 struct hstate *h = hstate_inode(inode);
4486 struct resv_map *resv_map = inode_resv_map(inode);
4487 long chg = 0;
4488 struct hugepage_subpool *spool = subpool_inode(inode);
4489 long gbl_reserve;
4490
4491 if (resv_map) {
4492 chg = region_del(resv_map, start, end);
4493 /*
4494 * region_del() can fail in the rare case where a region
4495 * must be split and another region descriptor can not be
4496 * allocated. If end == LONG_MAX, it will not fail.
4497 */
4498 if (chg < 0)
4499 return chg;
4500 }
4501
4502 spin_lock(&inode->i_lock);
4503 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4504 spin_unlock(&inode->i_lock);
4505
4506 /*
4507 * If the subpool has a minimum size, the number of global
4508 * reservations to be released may be adjusted.
4509 */
4510 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4511 hugetlb_acct_memory(h, -gbl_reserve);
4512
4513 return 0;
4514}
4515
4516#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4517static unsigned long page_table_shareable(struct vm_area_struct *svma,
4518 struct vm_area_struct *vma,
4519 unsigned long addr, pgoff_t idx)
4520{
4521 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4522 svma->vm_start;
4523 unsigned long sbase = saddr & PUD_MASK;
4524 unsigned long s_end = sbase + PUD_SIZE;
4525
4526 /* Allow segments to share if only one is marked locked */
4527 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4528 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4529
4530 /*
4531 * match the virtual addresses, permission and the alignment of the
4532 * page table page.
4533 */
4534 if (pmd_index(addr) != pmd_index(saddr) ||
4535 vm_flags != svm_flags ||
4536 sbase < svma->vm_start || svma->vm_end < s_end)
4537 return 0;
4538
4539 return saddr;
4540}
4541
4542static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4543{
4544 unsigned long base = addr & PUD_MASK;
4545 unsigned long end = base + PUD_SIZE;
4546
4547 /*
4548 * check on proper vm_flags and page table alignment
4549 */
4550 if (vma->vm_flags & VM_MAYSHARE &&
4551 vma->vm_start <= base && end <= vma->vm_end)
4552 return true;
4553 return false;
4554}
4555
4556/*
4557 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4558 * and returns the corresponding pte. While this is not necessary for the
4559 * !shared pmd case because we can allocate the pmd later as well, it makes the
4560 * code much cleaner. pmd allocation is essential for the shared case because
4561 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4562 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4563 * bad pmd for sharing.
4564 */
4565pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4566{
4567 struct vm_area_struct *vma = find_vma(mm, addr);
4568 struct address_space *mapping = vma->vm_file->f_mapping;
4569 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4570 vma->vm_pgoff;
4571 struct vm_area_struct *svma;
4572 unsigned long saddr;
4573 pte_t *spte = NULL;
4574 pte_t *pte;
4575 spinlock_t *ptl;
4576
4577 if (!vma_shareable(vma, addr))
4578 return (pte_t *)pmd_alloc(mm, pud, addr);
4579
4580 i_mmap_lock_write(mapping);
4581 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4582 if (svma == vma)
4583 continue;
4584
4585 saddr = page_table_shareable(svma, vma, addr, idx);
4586 if (saddr) {
4587 spte = huge_pte_offset(svma->vm_mm, saddr,
4588 vma_mmu_pagesize(svma));
4589 if (spte) {
4590 get_page(virt_to_page(spte));
4591 break;
4592 }
4593 }
4594 }
4595
4596 if (!spte)
4597 goto out;
4598
4599 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4600 if (pud_none(*pud)) {
4601 pud_populate(mm, pud,
4602 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4603 mm_inc_nr_pmds(mm);
4604 } else {
4605 put_page(virt_to_page(spte));
4606 }
4607 spin_unlock(ptl);
4608out:
4609 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4610 i_mmap_unlock_write(mapping);
4611 return pte;
4612}
4613
4614/*
4615 * unmap huge page backed by shared pte.
4616 *
4617 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4618 * indicated by page_count > 1, unmap is achieved by clearing pud and
4619 * decrementing the ref count. If count == 1, the pte page is not shared.
4620 *
4621 * called with page table lock held.
4622 *
4623 * returns: 1 successfully unmapped a shared pte page
4624 * 0 the underlying pte page is not shared, or it is the last user
4625 */
4626int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4627{
4628 pgd_t *pgd = pgd_offset(mm, *addr);
4629 p4d_t *p4d = p4d_offset(pgd, *addr);
4630 pud_t *pud = pud_offset(p4d, *addr);
4631
4632 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4633 if (page_count(virt_to_page(ptep)) == 1)
4634 return 0;
4635
4636 pud_clear(pud);
4637 put_page(virt_to_page(ptep));
4638 mm_dec_nr_pmds(mm);
4639 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4640 return 1;
4641}
4642#define want_pmd_share() (1)
4643#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4644pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4645{
4646 return NULL;
4647}
4648
4649int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4650{
4651 return 0;
4652}
4653#define want_pmd_share() (0)
4654#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4655
4656#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4657pte_t *huge_pte_alloc(struct mm_struct *mm,
4658 unsigned long addr, unsigned long sz)
4659{
4660 pgd_t *pgd;
4661 p4d_t *p4d;
4662 pud_t *pud;
4663 pte_t *pte = NULL;
4664
4665 pgd = pgd_offset(mm, addr);
4666 p4d = p4d_alloc(mm, pgd, addr);
4667 if (!p4d)
4668 return NULL;
4669 pud = pud_alloc(mm, p4d, addr);
4670 if (pud) {
4671 if (sz == PUD_SIZE) {
4672 pte = (pte_t *)pud;
4673 } else {
4674 BUG_ON(sz != PMD_SIZE);
4675 if (want_pmd_share() && pud_none(*pud))
4676 pte = huge_pmd_share(mm, addr, pud);
4677 else
4678 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4679 }
4680 }
4681 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4682
4683 return pte;
4684}
4685
4686/*
4687 * huge_pte_offset() - Walk the page table to resolve the hugepage
4688 * entry at address @addr
4689 *
4690 * Return: Pointer to page table or swap entry (PUD or PMD) for
4691 * address @addr, or NULL if a p*d_none() entry is encountered and the
4692 * size @sz doesn't match the hugepage size at this level of the page
4693 * table.
4694 */
4695pte_t *huge_pte_offset(struct mm_struct *mm,
4696 unsigned long addr, unsigned long sz)
4697{
4698 pgd_t *pgd;
4699 p4d_t *p4d;
4700 pud_t *pud;
4701 pmd_t *pmd;
4702
4703 pgd = pgd_offset(mm, addr);
4704 if (!pgd_present(*pgd))
4705 return NULL;
4706 p4d = p4d_offset(pgd, addr);
4707 if (!p4d_present(*p4d))
4708 return NULL;
4709
4710 pud = pud_offset(p4d, addr);
4711 if (sz != PUD_SIZE && pud_none(*pud))
4712 return NULL;
4713 /* hugepage or swap? */
4714 if (pud_huge(*pud) || !pud_present(*pud))
4715 return (pte_t *)pud;
4716
4717 pmd = pmd_offset(pud, addr);
4718 if (sz != PMD_SIZE && pmd_none(*pmd))
4719 return NULL;
4720 /* hugepage or swap? */
4721 if (pmd_huge(*pmd) || !pmd_present(*pmd))
4722 return (pte_t *)pmd;
4723
4724 return NULL;
4725}
4726
4727#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4728
4729/*
4730 * These functions are overwritable if your architecture needs its own
4731 * behavior.
4732 */
4733struct page * __weak
4734follow_huge_addr(struct mm_struct *mm, unsigned long address,
4735 int write)
4736{
4737 return ERR_PTR(-EINVAL);
4738}
4739
4740struct page * __weak
4741follow_huge_pd(struct vm_area_struct *vma,
4742 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4743{
4744 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4745 return NULL;
4746}
4747
4748struct page * __weak
4749follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4750 pmd_t *pmd, int flags)
4751{
4752 struct page *page = NULL;
4753 spinlock_t *ptl;
4754 pte_t pte;
4755retry:
4756 ptl = pmd_lockptr(mm, pmd);
4757 spin_lock(ptl);
4758 /*
4759 * make sure that the address range covered by this pmd is not
4760 * unmapped from other threads.
4761 */
4762 if (!pmd_huge(*pmd))
4763 goto out;
4764 pte = huge_ptep_get((pte_t *)pmd);
4765 if (pte_present(pte)) {
4766 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4767 if (flags & FOLL_GET)
4768 get_page(page);
4769 } else {
4770 if (is_hugetlb_entry_migration(pte)) {
4771 spin_unlock(ptl);
4772 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4773 goto retry;
4774 }
4775 /*
4776 * hwpoisoned entry is treated as no_page_table in
4777 * follow_page_mask().
4778 */
4779 }
4780out:
4781 spin_unlock(ptl);
4782 return page;
4783}
4784
4785struct page * __weak
4786follow_huge_pud(struct mm_struct *mm, unsigned long address,
4787 pud_t *pud, int flags)
4788{
4789 if (flags & FOLL_GET)
4790 return NULL;
4791
4792 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4793}
4794
4795struct page * __weak
4796follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4797{
4798 if (flags & FOLL_GET)
4799 return NULL;
4800
4801 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4802}
4803
4804bool isolate_huge_page(struct page *page, struct list_head *list)
4805{
4806 bool ret = true;
4807
4808 VM_BUG_ON_PAGE(!PageHead(page), page);
4809 spin_lock(&hugetlb_lock);
4810 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4811 ret = false;
4812 goto unlock;
4813 }
4814 clear_page_huge_active(page);
4815 list_move_tail(&page->lru, list);
4816unlock:
4817 spin_unlock(&hugetlb_lock);
4818 return ret;
4819}
4820
4821void putback_active_hugepage(struct page *page)
4822{
4823 VM_BUG_ON_PAGE(!PageHead(page), page);
4824 spin_lock(&hugetlb_lock);
4825 set_page_huge_active(page);
4826 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4827 spin_unlock(&hugetlb_lock);
4828 put_page(page);
4829}
4830
4831void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
4832{
4833 struct hstate *h = page_hstate(oldpage);
4834
4835 hugetlb_cgroup_migrate(oldpage, newpage);
4836 set_page_owner_migrate_reason(newpage, reason);
4837
4838 /*
4839 * transfer temporary state of the new huge page. This is
4840 * reverse to other transitions because the newpage is going to
4841 * be final while the old one will be freed so it takes over
4842 * the temporary status.
4843 *
4844 * Also note that we have to transfer the per-node surplus state
4845 * here as well otherwise the global surplus count will not match
4846 * the per-node's.
4847 */
4848 if (PageHugeTemporary(newpage)) {
4849 int old_nid = page_to_nid(oldpage);
4850 int new_nid = page_to_nid(newpage);
4851
4852 SetPageHugeTemporary(oldpage);
4853 ClearPageHugeTemporary(newpage);
4854
4855 spin_lock(&hugetlb_lock);
4856 if (h->surplus_huge_pages_node[old_nid]) {
4857 h->surplus_huge_pages_node[old_nid]--;
4858 h->surplus_huge_pages_node[new_nid]++;
4859 }
4860 spin_unlock(&hugetlb_lock);
4861 }
4862}
1/*
2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
4 */
5#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/mm.h>
8#include <linux/seq_file.h>
9#include <linux/sysctl.h>
10#include <linux/highmem.h>
11#include <linux/mmu_notifier.h>
12#include <linux/nodemask.h>
13#include <linux/pagemap.h>
14#include <linux/mempolicy.h>
15#include <linux/compiler.h>
16#include <linux/cpuset.h>
17#include <linux/mutex.h>
18#include <linux/bootmem.h>
19#include <linux/sysfs.h>
20#include <linux/slab.h>
21#include <linux/rmap.h>
22#include <linux/swap.h>
23#include <linux/swapops.h>
24#include <linux/page-isolation.h>
25#include <linux/jhash.h>
26
27#include <asm/page.h>
28#include <asm/pgtable.h>
29#include <asm/tlb.h>
30
31#include <linux/io.h>
32#include <linux/hugetlb.h>
33#include <linux/hugetlb_cgroup.h>
34#include <linux/node.h>
35#include "internal.h"
36
37int hugepages_treat_as_movable;
38
39int hugetlb_max_hstate __read_mostly;
40unsigned int default_hstate_idx;
41struct hstate hstates[HUGE_MAX_HSTATE];
42/*
43 * Minimum page order among possible hugepage sizes, set to a proper value
44 * at boot time.
45 */
46static unsigned int minimum_order __read_mostly = UINT_MAX;
47
48__initdata LIST_HEAD(huge_boot_pages);
49
50/* for command line parsing */
51static struct hstate * __initdata parsed_hstate;
52static unsigned long __initdata default_hstate_max_huge_pages;
53static unsigned long __initdata default_hstate_size;
54
55/*
56 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
57 * free_huge_pages, and surplus_huge_pages.
58 */
59DEFINE_SPINLOCK(hugetlb_lock);
60
61/*
62 * Serializes faults on the same logical page. This is used to
63 * prevent spurious OOMs when the hugepage pool is fully utilized.
64 */
65static int num_fault_mutexes;
66struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
67
68/* Forward declaration */
69static int hugetlb_acct_memory(struct hstate *h, long delta);
70
71static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
72{
73 bool free = (spool->count == 0) && (spool->used_hpages == 0);
74
75 spin_unlock(&spool->lock);
76
77 /* If no pages are used, and no other handles to the subpool
78 * remain, give up any reservations mased on minimum size and
79 * free the subpool */
80 if (free) {
81 if (spool->min_hpages != -1)
82 hugetlb_acct_memory(spool->hstate,
83 -spool->min_hpages);
84 kfree(spool);
85 }
86}
87
88struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
89 long min_hpages)
90{
91 struct hugepage_subpool *spool;
92
93 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
94 if (!spool)
95 return NULL;
96
97 spin_lock_init(&spool->lock);
98 spool->count = 1;
99 spool->max_hpages = max_hpages;
100 spool->hstate = h;
101 spool->min_hpages = min_hpages;
102
103 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
104 kfree(spool);
105 return NULL;
106 }
107 spool->rsv_hpages = min_hpages;
108
109 return spool;
110}
111
112void hugepage_put_subpool(struct hugepage_subpool *spool)
113{
114 spin_lock(&spool->lock);
115 BUG_ON(!spool->count);
116 spool->count--;
117 unlock_or_release_subpool(spool);
118}
119
120/*
121 * Subpool accounting for allocating and reserving pages.
122 * Return -ENOMEM if there are not enough resources to satisfy the
123 * the request. Otherwise, return the number of pages by which the
124 * global pools must be adjusted (upward). The returned value may
125 * only be different than the passed value (delta) in the case where
126 * a subpool minimum size must be manitained.
127 */
128static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
129 long delta)
130{
131 long ret = delta;
132
133 if (!spool)
134 return ret;
135
136 spin_lock(&spool->lock);
137
138 if (spool->max_hpages != -1) { /* maximum size accounting */
139 if ((spool->used_hpages + delta) <= spool->max_hpages)
140 spool->used_hpages += delta;
141 else {
142 ret = -ENOMEM;
143 goto unlock_ret;
144 }
145 }
146
147 if (spool->min_hpages != -1) { /* minimum size accounting */
148 if (delta > spool->rsv_hpages) {
149 /*
150 * Asking for more reserves than those already taken on
151 * behalf of subpool. Return difference.
152 */
153 ret = delta - spool->rsv_hpages;
154 spool->rsv_hpages = 0;
155 } else {
156 ret = 0; /* reserves already accounted for */
157 spool->rsv_hpages -= delta;
158 }
159 }
160
161unlock_ret:
162 spin_unlock(&spool->lock);
163 return ret;
164}
165
166/*
167 * Subpool accounting for freeing and unreserving pages.
168 * Return the number of global page reservations that must be dropped.
169 * The return value may only be different than the passed value (delta)
170 * in the case where a subpool minimum size must be maintained.
171 */
172static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
173 long delta)
174{
175 long ret = delta;
176
177 if (!spool)
178 return delta;
179
180 spin_lock(&spool->lock);
181
182 if (spool->max_hpages != -1) /* maximum size accounting */
183 spool->used_hpages -= delta;
184
185 if (spool->min_hpages != -1) { /* minimum size accounting */
186 if (spool->rsv_hpages + delta <= spool->min_hpages)
187 ret = 0;
188 else
189 ret = spool->rsv_hpages + delta - spool->min_hpages;
190
191 spool->rsv_hpages += delta;
192 if (spool->rsv_hpages > spool->min_hpages)
193 spool->rsv_hpages = spool->min_hpages;
194 }
195
196 /*
197 * If hugetlbfs_put_super couldn't free spool due to an outstanding
198 * quota reference, free it now.
199 */
200 unlock_or_release_subpool(spool);
201
202 return ret;
203}
204
205static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
206{
207 return HUGETLBFS_SB(inode->i_sb)->spool;
208}
209
210static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
211{
212 return subpool_inode(file_inode(vma->vm_file));
213}
214
215/*
216 * Region tracking -- allows tracking of reservations and instantiated pages
217 * across the pages in a mapping.
218 *
219 * The region data structures are embedded into a resv_map and protected
220 * by a resv_map's lock. The set of regions within the resv_map represent
221 * reservations for huge pages, or huge pages that have already been
222 * instantiated within the map. The from and to elements are huge page
223 * indicies into the associated mapping. from indicates the starting index
224 * of the region. to represents the first index past the end of the region.
225 *
226 * For example, a file region structure with from == 0 and to == 4 represents
227 * four huge pages in a mapping. It is important to note that the to element
228 * represents the first element past the end of the region. This is used in
229 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
230 *
231 * Interval notation of the form [from, to) will be used to indicate that
232 * the endpoint from is inclusive and to is exclusive.
233 */
234struct file_region {
235 struct list_head link;
236 long from;
237 long to;
238};
239
240/*
241 * Add the huge page range represented by [f, t) to the reserve
242 * map. In the normal case, existing regions will be expanded
243 * to accommodate the specified range. Sufficient regions should
244 * exist for expansion due to the previous call to region_chg
245 * with the same range. However, it is possible that region_del
246 * could have been called after region_chg and modifed the map
247 * in such a way that no region exists to be expanded. In this
248 * case, pull a region descriptor from the cache associated with
249 * the map and use that for the new range.
250 *
251 * Return the number of new huge pages added to the map. This
252 * number is greater than or equal to zero.
253 */
254static long region_add(struct resv_map *resv, long f, long t)
255{
256 struct list_head *head = &resv->regions;
257 struct file_region *rg, *nrg, *trg;
258 long add = 0;
259
260 spin_lock(&resv->lock);
261 /* Locate the region we are either in or before. */
262 list_for_each_entry(rg, head, link)
263 if (f <= rg->to)
264 break;
265
266 /*
267 * If no region exists which can be expanded to include the
268 * specified range, the list must have been modified by an
269 * interleving call to region_del(). Pull a region descriptor
270 * from the cache and use it for this range.
271 */
272 if (&rg->link == head || t < rg->from) {
273 VM_BUG_ON(resv->region_cache_count <= 0);
274
275 resv->region_cache_count--;
276 nrg = list_first_entry(&resv->region_cache, struct file_region,
277 link);
278 list_del(&nrg->link);
279
280 nrg->from = f;
281 nrg->to = t;
282 list_add(&nrg->link, rg->link.prev);
283
284 add += t - f;
285 goto out_locked;
286 }
287
288 /* Round our left edge to the current segment if it encloses us. */
289 if (f > rg->from)
290 f = rg->from;
291
292 /* Check for and consume any regions we now overlap with. */
293 nrg = rg;
294 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
295 if (&rg->link == head)
296 break;
297 if (rg->from > t)
298 break;
299
300 /* If this area reaches higher then extend our area to
301 * include it completely. If this is not the first area
302 * which we intend to reuse, free it. */
303 if (rg->to > t)
304 t = rg->to;
305 if (rg != nrg) {
306 /* Decrement return value by the deleted range.
307 * Another range will span this area so that by
308 * end of routine add will be >= zero
309 */
310 add -= (rg->to - rg->from);
311 list_del(&rg->link);
312 kfree(rg);
313 }
314 }
315
316 add += (nrg->from - f); /* Added to beginning of region */
317 nrg->from = f;
318 add += t - nrg->to; /* Added to end of region */
319 nrg->to = t;
320
321out_locked:
322 resv->adds_in_progress--;
323 spin_unlock(&resv->lock);
324 VM_BUG_ON(add < 0);
325 return add;
326}
327
328/*
329 * Examine the existing reserve map and determine how many
330 * huge pages in the specified range [f, t) are NOT currently
331 * represented. This routine is called before a subsequent
332 * call to region_add that will actually modify the reserve
333 * map to add the specified range [f, t). region_chg does
334 * not change the number of huge pages represented by the
335 * map. However, if the existing regions in the map can not
336 * be expanded to represent the new range, a new file_region
337 * structure is added to the map as a placeholder. This is
338 * so that the subsequent region_add call will have all the
339 * regions it needs and will not fail.
340 *
341 * Upon entry, region_chg will also examine the cache of region descriptors
342 * associated with the map. If there are not enough descriptors cached, one
343 * will be allocated for the in progress add operation.
344 *
345 * Returns the number of huge pages that need to be added to the existing
346 * reservation map for the range [f, t). This number is greater or equal to
347 * zero. -ENOMEM is returned if a new file_region structure or cache entry
348 * is needed and can not be allocated.
349 */
350static long region_chg(struct resv_map *resv, long f, long t)
351{
352 struct list_head *head = &resv->regions;
353 struct file_region *rg, *nrg = NULL;
354 long chg = 0;
355
356retry:
357 spin_lock(&resv->lock);
358retry_locked:
359 resv->adds_in_progress++;
360
361 /*
362 * Check for sufficient descriptors in the cache to accommodate
363 * the number of in progress add operations.
364 */
365 if (resv->adds_in_progress > resv->region_cache_count) {
366 struct file_region *trg;
367
368 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
369 /* Must drop lock to allocate a new descriptor. */
370 resv->adds_in_progress--;
371 spin_unlock(&resv->lock);
372
373 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
374 if (!trg) {
375 kfree(nrg);
376 return -ENOMEM;
377 }
378
379 spin_lock(&resv->lock);
380 list_add(&trg->link, &resv->region_cache);
381 resv->region_cache_count++;
382 goto retry_locked;
383 }
384
385 /* Locate the region we are before or in. */
386 list_for_each_entry(rg, head, link)
387 if (f <= rg->to)
388 break;
389
390 /* If we are below the current region then a new region is required.
391 * Subtle, allocate a new region at the position but make it zero
392 * size such that we can guarantee to record the reservation. */
393 if (&rg->link == head || t < rg->from) {
394 if (!nrg) {
395 resv->adds_in_progress--;
396 spin_unlock(&resv->lock);
397 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
398 if (!nrg)
399 return -ENOMEM;
400
401 nrg->from = f;
402 nrg->to = f;
403 INIT_LIST_HEAD(&nrg->link);
404 goto retry;
405 }
406
407 list_add(&nrg->link, rg->link.prev);
408 chg = t - f;
409 goto out_nrg;
410 }
411
412 /* Round our left edge to the current segment if it encloses us. */
413 if (f > rg->from)
414 f = rg->from;
415 chg = t - f;
416
417 /* Check for and consume any regions we now overlap with. */
418 list_for_each_entry(rg, rg->link.prev, link) {
419 if (&rg->link == head)
420 break;
421 if (rg->from > t)
422 goto out;
423
424 /* We overlap with this area, if it extends further than
425 * us then we must extend ourselves. Account for its
426 * existing reservation. */
427 if (rg->to > t) {
428 chg += rg->to - t;
429 t = rg->to;
430 }
431 chg -= rg->to - rg->from;
432 }
433
434out:
435 spin_unlock(&resv->lock);
436 /* We already know we raced and no longer need the new region */
437 kfree(nrg);
438 return chg;
439out_nrg:
440 spin_unlock(&resv->lock);
441 return chg;
442}
443
444/*
445 * Abort the in progress add operation. The adds_in_progress field
446 * of the resv_map keeps track of the operations in progress between
447 * calls to region_chg and region_add. Operations are sometimes
448 * aborted after the call to region_chg. In such cases, region_abort
449 * is called to decrement the adds_in_progress counter.
450 *
451 * NOTE: The range arguments [f, t) are not needed or used in this
452 * routine. They are kept to make reading the calling code easier as
453 * arguments will match the associated region_chg call.
454 */
455static void region_abort(struct resv_map *resv, long f, long t)
456{
457 spin_lock(&resv->lock);
458 VM_BUG_ON(!resv->region_cache_count);
459 resv->adds_in_progress--;
460 spin_unlock(&resv->lock);
461}
462
463/*
464 * Delete the specified range [f, t) from the reserve map. If the
465 * t parameter is LONG_MAX, this indicates that ALL regions after f
466 * should be deleted. Locate the regions which intersect [f, t)
467 * and either trim, delete or split the existing regions.
468 *
469 * Returns the number of huge pages deleted from the reserve map.
470 * In the normal case, the return value is zero or more. In the
471 * case where a region must be split, a new region descriptor must
472 * be allocated. If the allocation fails, -ENOMEM will be returned.
473 * NOTE: If the parameter t == LONG_MAX, then we will never split
474 * a region and possibly return -ENOMEM. Callers specifying
475 * t == LONG_MAX do not need to check for -ENOMEM error.
476 */
477static long region_del(struct resv_map *resv, long f, long t)
478{
479 struct list_head *head = &resv->regions;
480 struct file_region *rg, *trg;
481 struct file_region *nrg = NULL;
482 long del = 0;
483
484retry:
485 spin_lock(&resv->lock);
486 list_for_each_entry_safe(rg, trg, head, link) {
487 /*
488 * Skip regions before the range to be deleted. file_region
489 * ranges are normally of the form [from, to). However, there
490 * may be a "placeholder" entry in the map which is of the form
491 * (from, to) with from == to. Check for placeholder entries
492 * at the beginning of the range to be deleted.
493 */
494 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
495 continue;
496
497 if (rg->from >= t)
498 break;
499
500 if (f > rg->from && t < rg->to) { /* Must split region */
501 /*
502 * Check for an entry in the cache before dropping
503 * lock and attempting allocation.
504 */
505 if (!nrg &&
506 resv->region_cache_count > resv->adds_in_progress) {
507 nrg = list_first_entry(&resv->region_cache,
508 struct file_region,
509 link);
510 list_del(&nrg->link);
511 resv->region_cache_count--;
512 }
513
514 if (!nrg) {
515 spin_unlock(&resv->lock);
516 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
517 if (!nrg)
518 return -ENOMEM;
519 goto retry;
520 }
521
522 del += t - f;
523
524 /* New entry for end of split region */
525 nrg->from = t;
526 nrg->to = rg->to;
527 INIT_LIST_HEAD(&nrg->link);
528
529 /* Original entry is trimmed */
530 rg->to = f;
531
532 list_add(&nrg->link, &rg->link);
533 nrg = NULL;
534 break;
535 }
536
537 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
538 del += rg->to - rg->from;
539 list_del(&rg->link);
540 kfree(rg);
541 continue;
542 }
543
544 if (f <= rg->from) { /* Trim beginning of region */
545 del += t - rg->from;
546 rg->from = t;
547 } else { /* Trim end of region */
548 del += rg->to - f;
549 rg->to = f;
550 }
551 }
552
553 spin_unlock(&resv->lock);
554 kfree(nrg);
555 return del;
556}
557
558/*
559 * A rare out of memory error was encountered which prevented removal of
560 * the reserve map region for a page. The huge page itself was free'ed
561 * and removed from the page cache. This routine will adjust the subpool
562 * usage count, and the global reserve count if needed. By incrementing
563 * these counts, the reserve map entry which could not be deleted will
564 * appear as a "reserved" entry instead of simply dangling with incorrect
565 * counts.
566 */
567void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
568{
569 struct hugepage_subpool *spool = subpool_inode(inode);
570 long rsv_adjust;
571
572 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
573 if (restore_reserve && rsv_adjust) {
574 struct hstate *h = hstate_inode(inode);
575
576 hugetlb_acct_memory(h, 1);
577 }
578}
579
580/*
581 * Count and return the number of huge pages in the reserve map
582 * that intersect with the range [f, t).
583 */
584static long region_count(struct resv_map *resv, long f, long t)
585{
586 struct list_head *head = &resv->regions;
587 struct file_region *rg;
588 long chg = 0;
589
590 spin_lock(&resv->lock);
591 /* Locate each segment we overlap with, and count that overlap. */
592 list_for_each_entry(rg, head, link) {
593 long seg_from;
594 long seg_to;
595
596 if (rg->to <= f)
597 continue;
598 if (rg->from >= t)
599 break;
600
601 seg_from = max(rg->from, f);
602 seg_to = min(rg->to, t);
603
604 chg += seg_to - seg_from;
605 }
606 spin_unlock(&resv->lock);
607
608 return chg;
609}
610
611/*
612 * Convert the address within this vma to the page offset within
613 * the mapping, in pagecache page units; huge pages here.
614 */
615static pgoff_t vma_hugecache_offset(struct hstate *h,
616 struct vm_area_struct *vma, unsigned long address)
617{
618 return ((address - vma->vm_start) >> huge_page_shift(h)) +
619 (vma->vm_pgoff >> huge_page_order(h));
620}
621
622pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
623 unsigned long address)
624{
625 return vma_hugecache_offset(hstate_vma(vma), vma, address);
626}
627
628/*
629 * Return the size of the pages allocated when backing a VMA. In the majority
630 * cases this will be same size as used by the page table entries.
631 */
632unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
633{
634 struct hstate *hstate;
635
636 if (!is_vm_hugetlb_page(vma))
637 return PAGE_SIZE;
638
639 hstate = hstate_vma(vma);
640
641 return 1UL << huge_page_shift(hstate);
642}
643EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
644
645/*
646 * Return the page size being used by the MMU to back a VMA. In the majority
647 * of cases, the page size used by the kernel matches the MMU size. On
648 * architectures where it differs, an architecture-specific version of this
649 * function is required.
650 */
651#ifndef vma_mmu_pagesize
652unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
653{
654 return vma_kernel_pagesize(vma);
655}
656#endif
657
658/*
659 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
660 * bits of the reservation map pointer, which are always clear due to
661 * alignment.
662 */
663#define HPAGE_RESV_OWNER (1UL << 0)
664#define HPAGE_RESV_UNMAPPED (1UL << 1)
665#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
666
667/*
668 * These helpers are used to track how many pages are reserved for
669 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
670 * is guaranteed to have their future faults succeed.
671 *
672 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
673 * the reserve counters are updated with the hugetlb_lock held. It is safe
674 * to reset the VMA at fork() time as it is not in use yet and there is no
675 * chance of the global counters getting corrupted as a result of the values.
676 *
677 * The private mapping reservation is represented in a subtly different
678 * manner to a shared mapping. A shared mapping has a region map associated
679 * with the underlying file, this region map represents the backing file
680 * pages which have ever had a reservation assigned which this persists even
681 * after the page is instantiated. A private mapping has a region map
682 * associated with the original mmap which is attached to all VMAs which
683 * reference it, this region map represents those offsets which have consumed
684 * reservation ie. where pages have been instantiated.
685 */
686static unsigned long get_vma_private_data(struct vm_area_struct *vma)
687{
688 return (unsigned long)vma->vm_private_data;
689}
690
691static void set_vma_private_data(struct vm_area_struct *vma,
692 unsigned long value)
693{
694 vma->vm_private_data = (void *)value;
695}
696
697struct resv_map *resv_map_alloc(void)
698{
699 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
700 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
701
702 if (!resv_map || !rg) {
703 kfree(resv_map);
704 kfree(rg);
705 return NULL;
706 }
707
708 kref_init(&resv_map->refs);
709 spin_lock_init(&resv_map->lock);
710 INIT_LIST_HEAD(&resv_map->regions);
711
712 resv_map->adds_in_progress = 0;
713
714 INIT_LIST_HEAD(&resv_map->region_cache);
715 list_add(&rg->link, &resv_map->region_cache);
716 resv_map->region_cache_count = 1;
717
718 return resv_map;
719}
720
721void resv_map_release(struct kref *ref)
722{
723 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
724 struct list_head *head = &resv_map->region_cache;
725 struct file_region *rg, *trg;
726
727 /* Clear out any active regions before we release the map. */
728 region_del(resv_map, 0, LONG_MAX);
729
730 /* ... and any entries left in the cache */
731 list_for_each_entry_safe(rg, trg, head, link) {
732 list_del(&rg->link);
733 kfree(rg);
734 }
735
736 VM_BUG_ON(resv_map->adds_in_progress);
737
738 kfree(resv_map);
739}
740
741static inline struct resv_map *inode_resv_map(struct inode *inode)
742{
743 return inode->i_mapping->private_data;
744}
745
746static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
747{
748 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
749 if (vma->vm_flags & VM_MAYSHARE) {
750 struct address_space *mapping = vma->vm_file->f_mapping;
751 struct inode *inode = mapping->host;
752
753 return inode_resv_map(inode);
754
755 } else {
756 return (struct resv_map *)(get_vma_private_data(vma) &
757 ~HPAGE_RESV_MASK);
758 }
759}
760
761static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
762{
763 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
764 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
765
766 set_vma_private_data(vma, (get_vma_private_data(vma) &
767 HPAGE_RESV_MASK) | (unsigned long)map);
768}
769
770static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
771{
772 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
773 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
774
775 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
776}
777
778static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
779{
780 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
781
782 return (get_vma_private_data(vma) & flag) != 0;
783}
784
785/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
786void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
787{
788 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
789 if (!(vma->vm_flags & VM_MAYSHARE))
790 vma->vm_private_data = (void *)0;
791}
792
793/* Returns true if the VMA has associated reserve pages */
794static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
795{
796 if (vma->vm_flags & VM_NORESERVE) {
797 /*
798 * This address is already reserved by other process(chg == 0),
799 * so, we should decrement reserved count. Without decrementing,
800 * reserve count remains after releasing inode, because this
801 * allocated page will go into page cache and is regarded as
802 * coming from reserved pool in releasing step. Currently, we
803 * don't have any other solution to deal with this situation
804 * properly, so add work-around here.
805 */
806 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
807 return true;
808 else
809 return false;
810 }
811
812 /* Shared mappings always use reserves */
813 if (vma->vm_flags & VM_MAYSHARE) {
814 /*
815 * We know VM_NORESERVE is not set. Therefore, there SHOULD
816 * be a region map for all pages. The only situation where
817 * there is no region map is if a hole was punched via
818 * fallocate. In this case, there really are no reverves to
819 * use. This situation is indicated if chg != 0.
820 */
821 if (chg)
822 return false;
823 else
824 return true;
825 }
826
827 /*
828 * Only the process that called mmap() has reserves for
829 * private mappings.
830 */
831 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
832 return true;
833
834 return false;
835}
836
837static void enqueue_huge_page(struct hstate *h, struct page *page)
838{
839 int nid = page_to_nid(page);
840 list_move(&page->lru, &h->hugepage_freelists[nid]);
841 h->free_huge_pages++;
842 h->free_huge_pages_node[nid]++;
843}
844
845static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
846{
847 struct page *page;
848
849 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
850 if (!is_migrate_isolate_page(page))
851 break;
852 /*
853 * if 'non-isolated free hugepage' not found on the list,
854 * the allocation fails.
855 */
856 if (&h->hugepage_freelists[nid] == &page->lru)
857 return NULL;
858 list_move(&page->lru, &h->hugepage_activelist);
859 set_page_refcounted(page);
860 h->free_huge_pages--;
861 h->free_huge_pages_node[nid]--;
862 return page;
863}
864
865/* Movability of hugepages depends on migration support. */
866static inline gfp_t htlb_alloc_mask(struct hstate *h)
867{
868 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
869 return GFP_HIGHUSER_MOVABLE;
870 else
871 return GFP_HIGHUSER;
872}
873
874static struct page *dequeue_huge_page_vma(struct hstate *h,
875 struct vm_area_struct *vma,
876 unsigned long address, int avoid_reserve,
877 long chg)
878{
879 struct page *page = NULL;
880 struct mempolicy *mpol;
881 nodemask_t *nodemask;
882 struct zonelist *zonelist;
883 struct zone *zone;
884 struct zoneref *z;
885 unsigned int cpuset_mems_cookie;
886
887 /*
888 * A child process with MAP_PRIVATE mappings created by their parent
889 * have no page reserves. This check ensures that reservations are
890 * not "stolen". The child may still get SIGKILLed
891 */
892 if (!vma_has_reserves(vma, chg) &&
893 h->free_huge_pages - h->resv_huge_pages == 0)
894 goto err;
895
896 /* If reserves cannot be used, ensure enough pages are in the pool */
897 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
898 goto err;
899
900retry_cpuset:
901 cpuset_mems_cookie = read_mems_allowed_begin();
902 zonelist = huge_zonelist(vma, address,
903 htlb_alloc_mask(h), &mpol, &nodemask);
904
905 for_each_zone_zonelist_nodemask(zone, z, zonelist,
906 MAX_NR_ZONES - 1, nodemask) {
907 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
908 page = dequeue_huge_page_node(h, zone_to_nid(zone));
909 if (page) {
910 if (avoid_reserve)
911 break;
912 if (!vma_has_reserves(vma, chg))
913 break;
914
915 SetPagePrivate(page);
916 h->resv_huge_pages--;
917 break;
918 }
919 }
920 }
921
922 mpol_cond_put(mpol);
923 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
924 goto retry_cpuset;
925 return page;
926
927err:
928 return NULL;
929}
930
931/*
932 * common helper functions for hstate_next_node_to_{alloc|free}.
933 * We may have allocated or freed a huge page based on a different
934 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
935 * be outside of *nodes_allowed. Ensure that we use an allowed
936 * node for alloc or free.
937 */
938static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
939{
940 nid = next_node(nid, *nodes_allowed);
941 if (nid == MAX_NUMNODES)
942 nid = first_node(*nodes_allowed);
943 VM_BUG_ON(nid >= MAX_NUMNODES);
944
945 return nid;
946}
947
948static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
949{
950 if (!node_isset(nid, *nodes_allowed))
951 nid = next_node_allowed(nid, nodes_allowed);
952 return nid;
953}
954
955/*
956 * returns the previously saved node ["this node"] from which to
957 * allocate a persistent huge page for the pool and advance the
958 * next node from which to allocate, handling wrap at end of node
959 * mask.
960 */
961static int hstate_next_node_to_alloc(struct hstate *h,
962 nodemask_t *nodes_allowed)
963{
964 int nid;
965
966 VM_BUG_ON(!nodes_allowed);
967
968 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
969 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
970
971 return nid;
972}
973
974/*
975 * helper for free_pool_huge_page() - return the previously saved
976 * node ["this node"] from which to free a huge page. Advance the
977 * next node id whether or not we find a free huge page to free so
978 * that the next attempt to free addresses the next node.
979 */
980static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
981{
982 int nid;
983
984 VM_BUG_ON(!nodes_allowed);
985
986 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
987 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
988
989 return nid;
990}
991
992#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
993 for (nr_nodes = nodes_weight(*mask); \
994 nr_nodes > 0 && \
995 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
996 nr_nodes--)
997
998#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
999 for (nr_nodes = nodes_weight(*mask); \
1000 nr_nodes > 0 && \
1001 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1002 nr_nodes--)
1003
1004#if defined(CONFIG_X86_64) && ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA))
1005static void destroy_compound_gigantic_page(struct page *page,
1006 unsigned int order)
1007{
1008 int i;
1009 int nr_pages = 1 << order;
1010 struct page *p = page + 1;
1011
1012 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1013 clear_compound_head(p);
1014 set_page_refcounted(p);
1015 }
1016
1017 set_compound_order(page, 0);
1018 __ClearPageHead(page);
1019}
1020
1021static void free_gigantic_page(struct page *page, unsigned int order)
1022{
1023 free_contig_range(page_to_pfn(page), 1 << order);
1024}
1025
1026static int __alloc_gigantic_page(unsigned long start_pfn,
1027 unsigned long nr_pages)
1028{
1029 unsigned long end_pfn = start_pfn + nr_pages;
1030 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1031}
1032
1033static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1034 unsigned long nr_pages)
1035{
1036 unsigned long i, end_pfn = start_pfn + nr_pages;
1037 struct page *page;
1038
1039 for (i = start_pfn; i < end_pfn; i++) {
1040 if (!pfn_valid(i))
1041 return false;
1042
1043 page = pfn_to_page(i);
1044
1045 if (PageReserved(page))
1046 return false;
1047
1048 if (page_count(page) > 0)
1049 return false;
1050
1051 if (PageHuge(page))
1052 return false;
1053 }
1054
1055 return true;
1056}
1057
1058static bool zone_spans_last_pfn(const struct zone *zone,
1059 unsigned long start_pfn, unsigned long nr_pages)
1060{
1061 unsigned long last_pfn = start_pfn + nr_pages - 1;
1062 return zone_spans_pfn(zone, last_pfn);
1063}
1064
1065static struct page *alloc_gigantic_page(int nid, unsigned int order)
1066{
1067 unsigned long nr_pages = 1 << order;
1068 unsigned long ret, pfn, flags;
1069 struct zone *z;
1070
1071 z = NODE_DATA(nid)->node_zones;
1072 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1073 spin_lock_irqsave(&z->lock, flags);
1074
1075 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1076 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1077 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1078 /*
1079 * We release the zone lock here because
1080 * alloc_contig_range() will also lock the zone
1081 * at some point. If there's an allocation
1082 * spinning on this lock, it may win the race
1083 * and cause alloc_contig_range() to fail...
1084 */
1085 spin_unlock_irqrestore(&z->lock, flags);
1086 ret = __alloc_gigantic_page(pfn, nr_pages);
1087 if (!ret)
1088 return pfn_to_page(pfn);
1089 spin_lock_irqsave(&z->lock, flags);
1090 }
1091 pfn += nr_pages;
1092 }
1093
1094 spin_unlock_irqrestore(&z->lock, flags);
1095 }
1096
1097 return NULL;
1098}
1099
1100static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1101static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1102
1103static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1104{
1105 struct page *page;
1106
1107 page = alloc_gigantic_page(nid, huge_page_order(h));
1108 if (page) {
1109 prep_compound_gigantic_page(page, huge_page_order(h));
1110 prep_new_huge_page(h, page, nid);
1111 }
1112
1113 return page;
1114}
1115
1116static int alloc_fresh_gigantic_page(struct hstate *h,
1117 nodemask_t *nodes_allowed)
1118{
1119 struct page *page = NULL;
1120 int nr_nodes, node;
1121
1122 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1123 page = alloc_fresh_gigantic_page_node(h, node);
1124 if (page)
1125 return 1;
1126 }
1127
1128 return 0;
1129}
1130
1131static inline bool gigantic_page_supported(void) { return true; }
1132#else
1133static inline bool gigantic_page_supported(void) { return false; }
1134static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1135static inline void destroy_compound_gigantic_page(struct page *page,
1136 unsigned int order) { }
1137static inline int alloc_fresh_gigantic_page(struct hstate *h,
1138 nodemask_t *nodes_allowed) { return 0; }
1139#endif
1140
1141static void update_and_free_page(struct hstate *h, struct page *page)
1142{
1143 int i;
1144
1145 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1146 return;
1147
1148 h->nr_huge_pages--;
1149 h->nr_huge_pages_node[page_to_nid(page)]--;
1150 for (i = 0; i < pages_per_huge_page(h); i++) {
1151 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1152 1 << PG_referenced | 1 << PG_dirty |
1153 1 << PG_active | 1 << PG_private |
1154 1 << PG_writeback);
1155 }
1156 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1157 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1158 set_page_refcounted(page);
1159 if (hstate_is_gigantic(h)) {
1160 destroy_compound_gigantic_page(page, huge_page_order(h));
1161 free_gigantic_page(page, huge_page_order(h));
1162 } else {
1163 __free_pages(page, huge_page_order(h));
1164 }
1165}
1166
1167struct hstate *size_to_hstate(unsigned long size)
1168{
1169 struct hstate *h;
1170
1171 for_each_hstate(h) {
1172 if (huge_page_size(h) == size)
1173 return h;
1174 }
1175 return NULL;
1176}
1177
1178/*
1179 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1180 * to hstate->hugepage_activelist.)
1181 *
1182 * This function can be called for tail pages, but never returns true for them.
1183 */
1184bool page_huge_active(struct page *page)
1185{
1186 VM_BUG_ON_PAGE(!PageHuge(page), page);
1187 return PageHead(page) && PagePrivate(&page[1]);
1188}
1189
1190/* never called for tail page */
1191static void set_page_huge_active(struct page *page)
1192{
1193 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1194 SetPagePrivate(&page[1]);
1195}
1196
1197static void clear_page_huge_active(struct page *page)
1198{
1199 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1200 ClearPagePrivate(&page[1]);
1201}
1202
1203void free_huge_page(struct page *page)
1204{
1205 /*
1206 * Can't pass hstate in here because it is called from the
1207 * compound page destructor.
1208 */
1209 struct hstate *h = page_hstate(page);
1210 int nid = page_to_nid(page);
1211 struct hugepage_subpool *spool =
1212 (struct hugepage_subpool *)page_private(page);
1213 bool restore_reserve;
1214
1215 set_page_private(page, 0);
1216 page->mapping = NULL;
1217 VM_BUG_ON_PAGE(page_count(page), page);
1218 VM_BUG_ON_PAGE(page_mapcount(page), page);
1219 restore_reserve = PagePrivate(page);
1220 ClearPagePrivate(page);
1221
1222 /*
1223 * A return code of zero implies that the subpool will be under its
1224 * minimum size if the reservation is not restored after page is free.
1225 * Therefore, force restore_reserve operation.
1226 */
1227 if (hugepage_subpool_put_pages(spool, 1) == 0)
1228 restore_reserve = true;
1229
1230 spin_lock(&hugetlb_lock);
1231 clear_page_huge_active(page);
1232 hugetlb_cgroup_uncharge_page(hstate_index(h),
1233 pages_per_huge_page(h), page);
1234 if (restore_reserve)
1235 h->resv_huge_pages++;
1236
1237 if (h->surplus_huge_pages_node[nid]) {
1238 /* remove the page from active list */
1239 list_del(&page->lru);
1240 update_and_free_page(h, page);
1241 h->surplus_huge_pages--;
1242 h->surplus_huge_pages_node[nid]--;
1243 } else {
1244 arch_clear_hugepage_flags(page);
1245 enqueue_huge_page(h, page);
1246 }
1247 spin_unlock(&hugetlb_lock);
1248}
1249
1250static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1251{
1252 INIT_LIST_HEAD(&page->lru);
1253 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1254 spin_lock(&hugetlb_lock);
1255 set_hugetlb_cgroup(page, NULL);
1256 h->nr_huge_pages++;
1257 h->nr_huge_pages_node[nid]++;
1258 spin_unlock(&hugetlb_lock);
1259 put_page(page); /* free it into the hugepage allocator */
1260}
1261
1262static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1263{
1264 int i;
1265 int nr_pages = 1 << order;
1266 struct page *p = page + 1;
1267
1268 /* we rely on prep_new_huge_page to set the destructor */
1269 set_compound_order(page, order);
1270 __ClearPageReserved(page);
1271 __SetPageHead(page);
1272 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1273 /*
1274 * For gigantic hugepages allocated through bootmem at
1275 * boot, it's safer to be consistent with the not-gigantic
1276 * hugepages and clear the PG_reserved bit from all tail pages
1277 * too. Otherwse drivers using get_user_pages() to access tail
1278 * pages may get the reference counting wrong if they see
1279 * PG_reserved set on a tail page (despite the head page not
1280 * having PG_reserved set). Enforcing this consistency between
1281 * head and tail pages allows drivers to optimize away a check
1282 * on the head page when they need know if put_page() is needed
1283 * after get_user_pages().
1284 */
1285 __ClearPageReserved(p);
1286 set_page_count(p, 0);
1287 set_compound_head(p, page);
1288 }
1289 atomic_set(compound_mapcount_ptr(page), -1);
1290}
1291
1292/*
1293 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1294 * transparent huge pages. See the PageTransHuge() documentation for more
1295 * details.
1296 */
1297int PageHuge(struct page *page)
1298{
1299 if (!PageCompound(page))
1300 return 0;
1301
1302 page = compound_head(page);
1303 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1304}
1305EXPORT_SYMBOL_GPL(PageHuge);
1306
1307/*
1308 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1309 * normal or transparent huge pages.
1310 */
1311int PageHeadHuge(struct page *page_head)
1312{
1313 if (!PageHead(page_head))
1314 return 0;
1315
1316 return get_compound_page_dtor(page_head) == free_huge_page;
1317}
1318
1319pgoff_t __basepage_index(struct page *page)
1320{
1321 struct page *page_head = compound_head(page);
1322 pgoff_t index = page_index(page_head);
1323 unsigned long compound_idx;
1324
1325 if (!PageHuge(page_head))
1326 return page_index(page);
1327
1328 if (compound_order(page_head) >= MAX_ORDER)
1329 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1330 else
1331 compound_idx = page - page_head;
1332
1333 return (index << compound_order(page_head)) + compound_idx;
1334}
1335
1336static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1337{
1338 struct page *page;
1339
1340 page = __alloc_pages_node(nid,
1341 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1342 __GFP_REPEAT|__GFP_NOWARN,
1343 huge_page_order(h));
1344 if (page) {
1345 prep_new_huge_page(h, page, nid);
1346 }
1347
1348 return page;
1349}
1350
1351static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1352{
1353 struct page *page;
1354 int nr_nodes, node;
1355 int ret = 0;
1356
1357 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1358 page = alloc_fresh_huge_page_node(h, node);
1359 if (page) {
1360 ret = 1;
1361 break;
1362 }
1363 }
1364
1365 if (ret)
1366 count_vm_event(HTLB_BUDDY_PGALLOC);
1367 else
1368 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1369
1370 return ret;
1371}
1372
1373/*
1374 * Free huge page from pool from next node to free.
1375 * Attempt to keep persistent huge pages more or less
1376 * balanced over allowed nodes.
1377 * Called with hugetlb_lock locked.
1378 */
1379static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1380 bool acct_surplus)
1381{
1382 int nr_nodes, node;
1383 int ret = 0;
1384
1385 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1386 /*
1387 * If we're returning unused surplus pages, only examine
1388 * nodes with surplus pages.
1389 */
1390 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1391 !list_empty(&h->hugepage_freelists[node])) {
1392 struct page *page =
1393 list_entry(h->hugepage_freelists[node].next,
1394 struct page, lru);
1395 list_del(&page->lru);
1396 h->free_huge_pages--;
1397 h->free_huge_pages_node[node]--;
1398 if (acct_surplus) {
1399 h->surplus_huge_pages--;
1400 h->surplus_huge_pages_node[node]--;
1401 }
1402 update_and_free_page(h, page);
1403 ret = 1;
1404 break;
1405 }
1406 }
1407
1408 return ret;
1409}
1410
1411/*
1412 * Dissolve a given free hugepage into free buddy pages. This function does
1413 * nothing for in-use (including surplus) hugepages.
1414 */
1415static void dissolve_free_huge_page(struct page *page)
1416{
1417 spin_lock(&hugetlb_lock);
1418 if (PageHuge(page) && !page_count(page)) {
1419 struct hstate *h = page_hstate(page);
1420 int nid = page_to_nid(page);
1421 list_del(&page->lru);
1422 h->free_huge_pages--;
1423 h->free_huge_pages_node[nid]--;
1424 update_and_free_page(h, page);
1425 }
1426 spin_unlock(&hugetlb_lock);
1427}
1428
1429/*
1430 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1431 * make specified memory blocks removable from the system.
1432 * Note that start_pfn should aligned with (minimum) hugepage size.
1433 */
1434void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1435{
1436 unsigned long pfn;
1437
1438 if (!hugepages_supported())
1439 return;
1440
1441 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1442 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1443 dissolve_free_huge_page(pfn_to_page(pfn));
1444}
1445
1446/*
1447 * There are 3 ways this can get called:
1448 * 1. With vma+addr: we use the VMA's memory policy
1449 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1450 * page from any node, and let the buddy allocator itself figure
1451 * it out.
1452 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1453 * strictly from 'nid'
1454 */
1455static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1456 struct vm_area_struct *vma, unsigned long addr, int nid)
1457{
1458 int order = huge_page_order(h);
1459 gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1460 unsigned int cpuset_mems_cookie;
1461
1462 /*
1463 * We need a VMA to get a memory policy. If we do not
1464 * have one, we use the 'nid' argument.
1465 *
1466 * The mempolicy stuff below has some non-inlined bits
1467 * and calls ->vm_ops. That makes it hard to optimize at
1468 * compile-time, even when NUMA is off and it does
1469 * nothing. This helps the compiler optimize it out.
1470 */
1471 if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
1472 /*
1473 * If a specific node is requested, make sure to
1474 * get memory from there, but only when a node
1475 * is explicitly specified.
1476 */
1477 if (nid != NUMA_NO_NODE)
1478 gfp |= __GFP_THISNODE;
1479 /*
1480 * Make sure to call something that can handle
1481 * nid=NUMA_NO_NODE
1482 */
1483 return alloc_pages_node(nid, gfp, order);
1484 }
1485
1486 /*
1487 * OK, so we have a VMA. Fetch the mempolicy and try to
1488 * allocate a huge page with it. We will only reach this
1489 * when CONFIG_NUMA=y.
1490 */
1491 do {
1492 struct page *page;
1493 struct mempolicy *mpol;
1494 struct zonelist *zl;
1495 nodemask_t *nodemask;
1496
1497 cpuset_mems_cookie = read_mems_allowed_begin();
1498 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1499 mpol_cond_put(mpol);
1500 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1501 if (page)
1502 return page;
1503 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1504
1505 return NULL;
1506}
1507
1508/*
1509 * There are two ways to allocate a huge page:
1510 * 1. When you have a VMA and an address (like a fault)
1511 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1512 *
1513 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1514 * this case which signifies that the allocation should be done with
1515 * respect for the VMA's memory policy.
1516 *
1517 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1518 * implies that memory policies will not be taken in to account.
1519 */
1520static struct page *__alloc_buddy_huge_page(struct hstate *h,
1521 struct vm_area_struct *vma, unsigned long addr, int nid)
1522{
1523 struct page *page;
1524 unsigned int r_nid;
1525
1526 if (hstate_is_gigantic(h))
1527 return NULL;
1528
1529 /*
1530 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1531 * This makes sure the caller is picking _one_ of the modes with which
1532 * we can call this function, not both.
1533 */
1534 if (vma || (addr != -1)) {
1535 VM_WARN_ON_ONCE(addr == -1);
1536 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
1537 }
1538 /*
1539 * Assume we will successfully allocate the surplus page to
1540 * prevent racing processes from causing the surplus to exceed
1541 * overcommit
1542 *
1543 * This however introduces a different race, where a process B
1544 * tries to grow the static hugepage pool while alloc_pages() is
1545 * called by process A. B will only examine the per-node
1546 * counters in determining if surplus huge pages can be
1547 * converted to normal huge pages in adjust_pool_surplus(). A
1548 * won't be able to increment the per-node counter, until the
1549 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1550 * no more huge pages can be converted from surplus to normal
1551 * state (and doesn't try to convert again). Thus, we have a
1552 * case where a surplus huge page exists, the pool is grown, and
1553 * the surplus huge page still exists after, even though it
1554 * should just have been converted to a normal huge page. This
1555 * does not leak memory, though, as the hugepage will be freed
1556 * once it is out of use. It also does not allow the counters to
1557 * go out of whack in adjust_pool_surplus() as we don't modify
1558 * the node values until we've gotten the hugepage and only the
1559 * per-node value is checked there.
1560 */
1561 spin_lock(&hugetlb_lock);
1562 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1563 spin_unlock(&hugetlb_lock);
1564 return NULL;
1565 } else {
1566 h->nr_huge_pages++;
1567 h->surplus_huge_pages++;
1568 }
1569 spin_unlock(&hugetlb_lock);
1570
1571 page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
1572
1573 spin_lock(&hugetlb_lock);
1574 if (page) {
1575 INIT_LIST_HEAD(&page->lru);
1576 r_nid = page_to_nid(page);
1577 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1578 set_hugetlb_cgroup(page, NULL);
1579 /*
1580 * We incremented the global counters already
1581 */
1582 h->nr_huge_pages_node[r_nid]++;
1583 h->surplus_huge_pages_node[r_nid]++;
1584 __count_vm_event(HTLB_BUDDY_PGALLOC);
1585 } else {
1586 h->nr_huge_pages--;
1587 h->surplus_huge_pages--;
1588 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1589 }
1590 spin_unlock(&hugetlb_lock);
1591
1592 return page;
1593}
1594
1595/*
1596 * Allocate a huge page from 'nid'. Note, 'nid' may be
1597 * NUMA_NO_NODE, which means that it may be allocated
1598 * anywhere.
1599 */
1600static
1601struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1602{
1603 unsigned long addr = -1;
1604
1605 return __alloc_buddy_huge_page(h, NULL, addr, nid);
1606}
1607
1608/*
1609 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1610 */
1611static
1612struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1613 struct vm_area_struct *vma, unsigned long addr)
1614{
1615 return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1616}
1617
1618/*
1619 * This allocation function is useful in the context where vma is irrelevant.
1620 * E.g. soft-offlining uses this function because it only cares physical
1621 * address of error page.
1622 */
1623struct page *alloc_huge_page_node(struct hstate *h, int nid)
1624{
1625 struct page *page = NULL;
1626
1627 spin_lock(&hugetlb_lock);
1628 if (h->free_huge_pages - h->resv_huge_pages > 0)
1629 page = dequeue_huge_page_node(h, nid);
1630 spin_unlock(&hugetlb_lock);
1631
1632 if (!page)
1633 page = __alloc_buddy_huge_page_no_mpol(h, nid);
1634
1635 return page;
1636}
1637
1638/*
1639 * Increase the hugetlb pool such that it can accommodate a reservation
1640 * of size 'delta'.
1641 */
1642static int gather_surplus_pages(struct hstate *h, int delta)
1643{
1644 struct list_head surplus_list;
1645 struct page *page, *tmp;
1646 int ret, i;
1647 int needed, allocated;
1648 bool alloc_ok = true;
1649
1650 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1651 if (needed <= 0) {
1652 h->resv_huge_pages += delta;
1653 return 0;
1654 }
1655
1656 allocated = 0;
1657 INIT_LIST_HEAD(&surplus_list);
1658
1659 ret = -ENOMEM;
1660retry:
1661 spin_unlock(&hugetlb_lock);
1662 for (i = 0; i < needed; i++) {
1663 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
1664 if (!page) {
1665 alloc_ok = false;
1666 break;
1667 }
1668 list_add(&page->lru, &surplus_list);
1669 }
1670 allocated += i;
1671
1672 /*
1673 * After retaking hugetlb_lock, we need to recalculate 'needed'
1674 * because either resv_huge_pages or free_huge_pages may have changed.
1675 */
1676 spin_lock(&hugetlb_lock);
1677 needed = (h->resv_huge_pages + delta) -
1678 (h->free_huge_pages + allocated);
1679 if (needed > 0) {
1680 if (alloc_ok)
1681 goto retry;
1682 /*
1683 * We were not able to allocate enough pages to
1684 * satisfy the entire reservation so we free what
1685 * we've allocated so far.
1686 */
1687 goto free;
1688 }
1689 /*
1690 * The surplus_list now contains _at_least_ the number of extra pages
1691 * needed to accommodate the reservation. Add the appropriate number
1692 * of pages to the hugetlb pool and free the extras back to the buddy
1693 * allocator. Commit the entire reservation here to prevent another
1694 * process from stealing the pages as they are added to the pool but
1695 * before they are reserved.
1696 */
1697 needed += allocated;
1698 h->resv_huge_pages += delta;
1699 ret = 0;
1700
1701 /* Free the needed pages to the hugetlb pool */
1702 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1703 if ((--needed) < 0)
1704 break;
1705 /*
1706 * This page is now managed by the hugetlb allocator and has
1707 * no users -- drop the buddy allocator's reference.
1708 */
1709 put_page_testzero(page);
1710 VM_BUG_ON_PAGE(page_count(page), page);
1711 enqueue_huge_page(h, page);
1712 }
1713free:
1714 spin_unlock(&hugetlb_lock);
1715
1716 /* Free unnecessary surplus pages to the buddy allocator */
1717 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1718 put_page(page);
1719 spin_lock(&hugetlb_lock);
1720
1721 return ret;
1722}
1723
1724/*
1725 * When releasing a hugetlb pool reservation, any surplus pages that were
1726 * allocated to satisfy the reservation must be explicitly freed if they were
1727 * never used.
1728 * Called with hugetlb_lock held.
1729 */
1730static void return_unused_surplus_pages(struct hstate *h,
1731 unsigned long unused_resv_pages)
1732{
1733 unsigned long nr_pages;
1734
1735 /* Uncommit the reservation */
1736 h->resv_huge_pages -= unused_resv_pages;
1737
1738 /* Cannot return gigantic pages currently */
1739 if (hstate_is_gigantic(h))
1740 return;
1741
1742 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1743
1744 /*
1745 * We want to release as many surplus pages as possible, spread
1746 * evenly across all nodes with memory. Iterate across these nodes
1747 * until we can no longer free unreserved surplus pages. This occurs
1748 * when the nodes with surplus pages have no free pages.
1749 * free_pool_huge_page() will balance the the freed pages across the
1750 * on-line nodes with memory and will handle the hstate accounting.
1751 */
1752 while (nr_pages--) {
1753 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1754 break;
1755 cond_resched_lock(&hugetlb_lock);
1756 }
1757}
1758
1759
1760/*
1761 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1762 * are used by the huge page allocation routines to manage reservations.
1763 *
1764 * vma_needs_reservation is called to determine if the huge page at addr
1765 * within the vma has an associated reservation. If a reservation is
1766 * needed, the value 1 is returned. The caller is then responsible for
1767 * managing the global reservation and subpool usage counts. After
1768 * the huge page has been allocated, vma_commit_reservation is called
1769 * to add the page to the reservation map. If the page allocation fails,
1770 * the reservation must be ended instead of committed. vma_end_reservation
1771 * is called in such cases.
1772 *
1773 * In the normal case, vma_commit_reservation returns the same value
1774 * as the preceding vma_needs_reservation call. The only time this
1775 * is not the case is if a reserve map was changed between calls. It
1776 * is the responsibility of the caller to notice the difference and
1777 * take appropriate action.
1778 */
1779enum vma_resv_mode {
1780 VMA_NEEDS_RESV,
1781 VMA_COMMIT_RESV,
1782 VMA_END_RESV,
1783};
1784static long __vma_reservation_common(struct hstate *h,
1785 struct vm_area_struct *vma, unsigned long addr,
1786 enum vma_resv_mode mode)
1787{
1788 struct resv_map *resv;
1789 pgoff_t idx;
1790 long ret;
1791
1792 resv = vma_resv_map(vma);
1793 if (!resv)
1794 return 1;
1795
1796 idx = vma_hugecache_offset(h, vma, addr);
1797 switch (mode) {
1798 case VMA_NEEDS_RESV:
1799 ret = region_chg(resv, idx, idx + 1);
1800 break;
1801 case VMA_COMMIT_RESV:
1802 ret = region_add(resv, idx, idx + 1);
1803 break;
1804 case VMA_END_RESV:
1805 region_abort(resv, idx, idx + 1);
1806 ret = 0;
1807 break;
1808 default:
1809 BUG();
1810 }
1811
1812 if (vma->vm_flags & VM_MAYSHARE)
1813 return ret;
1814 else
1815 return ret < 0 ? ret : 0;
1816}
1817
1818static long vma_needs_reservation(struct hstate *h,
1819 struct vm_area_struct *vma, unsigned long addr)
1820{
1821 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1822}
1823
1824static long vma_commit_reservation(struct hstate *h,
1825 struct vm_area_struct *vma, unsigned long addr)
1826{
1827 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1828}
1829
1830static void vma_end_reservation(struct hstate *h,
1831 struct vm_area_struct *vma, unsigned long addr)
1832{
1833 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1834}
1835
1836struct page *alloc_huge_page(struct vm_area_struct *vma,
1837 unsigned long addr, int avoid_reserve)
1838{
1839 struct hugepage_subpool *spool = subpool_vma(vma);
1840 struct hstate *h = hstate_vma(vma);
1841 struct page *page;
1842 long map_chg, map_commit;
1843 long gbl_chg;
1844 int ret, idx;
1845 struct hugetlb_cgroup *h_cg;
1846
1847 idx = hstate_index(h);
1848 /*
1849 * Examine the region/reserve map to determine if the process
1850 * has a reservation for the page to be allocated. A return
1851 * code of zero indicates a reservation exists (no change).
1852 */
1853 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1854 if (map_chg < 0)
1855 return ERR_PTR(-ENOMEM);
1856
1857 /*
1858 * Processes that did not create the mapping will have no
1859 * reserves as indicated by the region/reserve map. Check
1860 * that the allocation will not exceed the subpool limit.
1861 * Allocations for MAP_NORESERVE mappings also need to be
1862 * checked against any subpool limit.
1863 */
1864 if (map_chg || avoid_reserve) {
1865 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1866 if (gbl_chg < 0) {
1867 vma_end_reservation(h, vma, addr);
1868 return ERR_PTR(-ENOSPC);
1869 }
1870
1871 /*
1872 * Even though there was no reservation in the region/reserve
1873 * map, there could be reservations associated with the
1874 * subpool that can be used. This would be indicated if the
1875 * return value of hugepage_subpool_get_pages() is zero.
1876 * However, if avoid_reserve is specified we still avoid even
1877 * the subpool reservations.
1878 */
1879 if (avoid_reserve)
1880 gbl_chg = 1;
1881 }
1882
1883 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1884 if (ret)
1885 goto out_subpool_put;
1886
1887 spin_lock(&hugetlb_lock);
1888 /*
1889 * glb_chg is passed to indicate whether or not a page must be taken
1890 * from the global free pool (global change). gbl_chg == 0 indicates
1891 * a reservation exists for the allocation.
1892 */
1893 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
1894 if (!page) {
1895 spin_unlock(&hugetlb_lock);
1896 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
1897 if (!page)
1898 goto out_uncharge_cgroup;
1899 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
1900 SetPagePrivate(page);
1901 h->resv_huge_pages--;
1902 }
1903 spin_lock(&hugetlb_lock);
1904 list_move(&page->lru, &h->hugepage_activelist);
1905 /* Fall through */
1906 }
1907 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1908 spin_unlock(&hugetlb_lock);
1909
1910 set_page_private(page, (unsigned long)spool);
1911
1912 map_commit = vma_commit_reservation(h, vma, addr);
1913 if (unlikely(map_chg > map_commit)) {
1914 /*
1915 * The page was added to the reservation map between
1916 * vma_needs_reservation and vma_commit_reservation.
1917 * This indicates a race with hugetlb_reserve_pages.
1918 * Adjust for the subpool count incremented above AND
1919 * in hugetlb_reserve_pages for the same page. Also,
1920 * the reservation count added in hugetlb_reserve_pages
1921 * no longer applies.
1922 */
1923 long rsv_adjust;
1924
1925 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1926 hugetlb_acct_memory(h, -rsv_adjust);
1927 }
1928 return page;
1929
1930out_uncharge_cgroup:
1931 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1932out_subpool_put:
1933 if (map_chg || avoid_reserve)
1934 hugepage_subpool_put_pages(spool, 1);
1935 vma_end_reservation(h, vma, addr);
1936 return ERR_PTR(-ENOSPC);
1937}
1938
1939/*
1940 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1941 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1942 * where no ERR_VALUE is expected to be returned.
1943 */
1944struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1945 unsigned long addr, int avoid_reserve)
1946{
1947 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1948 if (IS_ERR(page))
1949 page = NULL;
1950 return page;
1951}
1952
1953int __weak alloc_bootmem_huge_page(struct hstate *h)
1954{
1955 struct huge_bootmem_page *m;
1956 int nr_nodes, node;
1957
1958 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1959 void *addr;
1960
1961 addr = memblock_virt_alloc_try_nid_nopanic(
1962 huge_page_size(h), huge_page_size(h),
1963 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1964 if (addr) {
1965 /*
1966 * Use the beginning of the huge page to store the
1967 * huge_bootmem_page struct (until gather_bootmem
1968 * puts them into the mem_map).
1969 */
1970 m = addr;
1971 goto found;
1972 }
1973 }
1974 return 0;
1975
1976found:
1977 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1978 /* Put them into a private list first because mem_map is not up yet */
1979 list_add(&m->list, &huge_boot_pages);
1980 m->hstate = h;
1981 return 1;
1982}
1983
1984static void __init prep_compound_huge_page(struct page *page,
1985 unsigned int order)
1986{
1987 if (unlikely(order > (MAX_ORDER - 1)))
1988 prep_compound_gigantic_page(page, order);
1989 else
1990 prep_compound_page(page, order);
1991}
1992
1993/* Put bootmem huge pages into the standard lists after mem_map is up */
1994static void __init gather_bootmem_prealloc(void)
1995{
1996 struct huge_bootmem_page *m;
1997
1998 list_for_each_entry(m, &huge_boot_pages, list) {
1999 struct hstate *h = m->hstate;
2000 struct page *page;
2001
2002#ifdef CONFIG_HIGHMEM
2003 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2004 memblock_free_late(__pa(m),
2005 sizeof(struct huge_bootmem_page));
2006#else
2007 page = virt_to_page(m);
2008#endif
2009 WARN_ON(page_count(page) != 1);
2010 prep_compound_huge_page(page, h->order);
2011 WARN_ON(PageReserved(page));
2012 prep_new_huge_page(h, page, page_to_nid(page));
2013 /*
2014 * If we had gigantic hugepages allocated at boot time, we need
2015 * to restore the 'stolen' pages to totalram_pages in order to
2016 * fix confusing memory reports from free(1) and another
2017 * side-effects, like CommitLimit going negative.
2018 */
2019 if (hstate_is_gigantic(h))
2020 adjust_managed_page_count(page, 1 << h->order);
2021 }
2022}
2023
2024static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2025{
2026 unsigned long i;
2027
2028 for (i = 0; i < h->max_huge_pages; ++i) {
2029 if (hstate_is_gigantic(h)) {
2030 if (!alloc_bootmem_huge_page(h))
2031 break;
2032 } else if (!alloc_fresh_huge_page(h,
2033 &node_states[N_MEMORY]))
2034 break;
2035 }
2036 h->max_huge_pages = i;
2037}
2038
2039static void __init hugetlb_init_hstates(void)
2040{
2041 struct hstate *h;
2042
2043 for_each_hstate(h) {
2044 if (minimum_order > huge_page_order(h))
2045 minimum_order = huge_page_order(h);
2046
2047 /* oversize hugepages were init'ed in early boot */
2048 if (!hstate_is_gigantic(h))
2049 hugetlb_hstate_alloc_pages(h);
2050 }
2051 VM_BUG_ON(minimum_order == UINT_MAX);
2052}
2053
2054static char * __init memfmt(char *buf, unsigned long n)
2055{
2056 if (n >= (1UL << 30))
2057 sprintf(buf, "%lu GB", n >> 30);
2058 else if (n >= (1UL << 20))
2059 sprintf(buf, "%lu MB", n >> 20);
2060 else
2061 sprintf(buf, "%lu KB", n >> 10);
2062 return buf;
2063}
2064
2065static void __init report_hugepages(void)
2066{
2067 struct hstate *h;
2068
2069 for_each_hstate(h) {
2070 char buf[32];
2071 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2072 memfmt(buf, huge_page_size(h)),
2073 h->free_huge_pages);
2074 }
2075}
2076
2077#ifdef CONFIG_HIGHMEM
2078static void try_to_free_low(struct hstate *h, unsigned long count,
2079 nodemask_t *nodes_allowed)
2080{
2081 int i;
2082
2083 if (hstate_is_gigantic(h))
2084 return;
2085
2086 for_each_node_mask(i, *nodes_allowed) {
2087 struct page *page, *next;
2088 struct list_head *freel = &h->hugepage_freelists[i];
2089 list_for_each_entry_safe(page, next, freel, lru) {
2090 if (count >= h->nr_huge_pages)
2091 return;
2092 if (PageHighMem(page))
2093 continue;
2094 list_del(&page->lru);
2095 update_and_free_page(h, page);
2096 h->free_huge_pages--;
2097 h->free_huge_pages_node[page_to_nid(page)]--;
2098 }
2099 }
2100}
2101#else
2102static inline void try_to_free_low(struct hstate *h, unsigned long count,
2103 nodemask_t *nodes_allowed)
2104{
2105}
2106#endif
2107
2108/*
2109 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2110 * balanced by operating on them in a round-robin fashion.
2111 * Returns 1 if an adjustment was made.
2112 */
2113static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2114 int delta)
2115{
2116 int nr_nodes, node;
2117
2118 VM_BUG_ON(delta != -1 && delta != 1);
2119
2120 if (delta < 0) {
2121 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2122 if (h->surplus_huge_pages_node[node])
2123 goto found;
2124 }
2125 } else {
2126 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2127 if (h->surplus_huge_pages_node[node] <
2128 h->nr_huge_pages_node[node])
2129 goto found;
2130 }
2131 }
2132 return 0;
2133
2134found:
2135 h->surplus_huge_pages += delta;
2136 h->surplus_huge_pages_node[node] += delta;
2137 return 1;
2138}
2139
2140#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2141static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2142 nodemask_t *nodes_allowed)
2143{
2144 unsigned long min_count, ret;
2145
2146 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2147 return h->max_huge_pages;
2148
2149 /*
2150 * Increase the pool size
2151 * First take pages out of surplus state. Then make up the
2152 * remaining difference by allocating fresh huge pages.
2153 *
2154 * We might race with __alloc_buddy_huge_page() here and be unable
2155 * to convert a surplus huge page to a normal huge page. That is
2156 * not critical, though, it just means the overall size of the
2157 * pool might be one hugepage larger than it needs to be, but
2158 * within all the constraints specified by the sysctls.
2159 */
2160 spin_lock(&hugetlb_lock);
2161 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2162 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2163 break;
2164 }
2165
2166 while (count > persistent_huge_pages(h)) {
2167 /*
2168 * If this allocation races such that we no longer need the
2169 * page, free_huge_page will handle it by freeing the page
2170 * and reducing the surplus.
2171 */
2172 spin_unlock(&hugetlb_lock);
2173 if (hstate_is_gigantic(h))
2174 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2175 else
2176 ret = alloc_fresh_huge_page(h, nodes_allowed);
2177 spin_lock(&hugetlb_lock);
2178 if (!ret)
2179 goto out;
2180
2181 /* Bail for signals. Probably ctrl-c from user */
2182 if (signal_pending(current))
2183 goto out;
2184 }
2185
2186 /*
2187 * Decrease the pool size
2188 * First return free pages to the buddy allocator (being careful
2189 * to keep enough around to satisfy reservations). Then place
2190 * pages into surplus state as needed so the pool will shrink
2191 * to the desired size as pages become free.
2192 *
2193 * By placing pages into the surplus state independent of the
2194 * overcommit value, we are allowing the surplus pool size to
2195 * exceed overcommit. There are few sane options here. Since
2196 * __alloc_buddy_huge_page() is checking the global counter,
2197 * though, we'll note that we're not allowed to exceed surplus
2198 * and won't grow the pool anywhere else. Not until one of the
2199 * sysctls are changed, or the surplus pages go out of use.
2200 */
2201 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2202 min_count = max(count, min_count);
2203 try_to_free_low(h, min_count, nodes_allowed);
2204 while (min_count < persistent_huge_pages(h)) {
2205 if (!free_pool_huge_page(h, nodes_allowed, 0))
2206 break;
2207 cond_resched_lock(&hugetlb_lock);
2208 }
2209 while (count < persistent_huge_pages(h)) {
2210 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2211 break;
2212 }
2213out:
2214 ret = persistent_huge_pages(h);
2215 spin_unlock(&hugetlb_lock);
2216 return ret;
2217}
2218
2219#define HSTATE_ATTR_RO(_name) \
2220 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2221
2222#define HSTATE_ATTR(_name) \
2223 static struct kobj_attribute _name##_attr = \
2224 __ATTR(_name, 0644, _name##_show, _name##_store)
2225
2226static struct kobject *hugepages_kobj;
2227static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2228
2229static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2230
2231static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2232{
2233 int i;
2234
2235 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2236 if (hstate_kobjs[i] == kobj) {
2237 if (nidp)
2238 *nidp = NUMA_NO_NODE;
2239 return &hstates[i];
2240 }
2241
2242 return kobj_to_node_hstate(kobj, nidp);
2243}
2244
2245static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2246 struct kobj_attribute *attr, char *buf)
2247{
2248 struct hstate *h;
2249 unsigned long nr_huge_pages;
2250 int nid;
2251
2252 h = kobj_to_hstate(kobj, &nid);
2253 if (nid == NUMA_NO_NODE)
2254 nr_huge_pages = h->nr_huge_pages;
2255 else
2256 nr_huge_pages = h->nr_huge_pages_node[nid];
2257
2258 return sprintf(buf, "%lu\n", nr_huge_pages);
2259}
2260
2261static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2262 struct hstate *h, int nid,
2263 unsigned long count, size_t len)
2264{
2265 int err;
2266 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2267
2268 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2269 err = -EINVAL;
2270 goto out;
2271 }
2272
2273 if (nid == NUMA_NO_NODE) {
2274 /*
2275 * global hstate attribute
2276 */
2277 if (!(obey_mempolicy &&
2278 init_nodemask_of_mempolicy(nodes_allowed))) {
2279 NODEMASK_FREE(nodes_allowed);
2280 nodes_allowed = &node_states[N_MEMORY];
2281 }
2282 } else if (nodes_allowed) {
2283 /*
2284 * per node hstate attribute: adjust count to global,
2285 * but restrict alloc/free to the specified node.
2286 */
2287 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2288 init_nodemask_of_node(nodes_allowed, nid);
2289 } else
2290 nodes_allowed = &node_states[N_MEMORY];
2291
2292 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2293
2294 if (nodes_allowed != &node_states[N_MEMORY])
2295 NODEMASK_FREE(nodes_allowed);
2296
2297 return len;
2298out:
2299 NODEMASK_FREE(nodes_allowed);
2300 return err;
2301}
2302
2303static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2304 struct kobject *kobj, const char *buf,
2305 size_t len)
2306{
2307 struct hstate *h;
2308 unsigned long count;
2309 int nid;
2310 int err;
2311
2312 err = kstrtoul(buf, 10, &count);
2313 if (err)
2314 return err;
2315
2316 h = kobj_to_hstate(kobj, &nid);
2317 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2318}
2319
2320static ssize_t nr_hugepages_show(struct kobject *kobj,
2321 struct kobj_attribute *attr, char *buf)
2322{
2323 return nr_hugepages_show_common(kobj, attr, buf);
2324}
2325
2326static ssize_t nr_hugepages_store(struct kobject *kobj,
2327 struct kobj_attribute *attr, const char *buf, size_t len)
2328{
2329 return nr_hugepages_store_common(false, kobj, buf, len);
2330}
2331HSTATE_ATTR(nr_hugepages);
2332
2333#ifdef CONFIG_NUMA
2334
2335/*
2336 * hstate attribute for optionally mempolicy-based constraint on persistent
2337 * huge page alloc/free.
2338 */
2339static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2340 struct kobj_attribute *attr, char *buf)
2341{
2342 return nr_hugepages_show_common(kobj, attr, buf);
2343}
2344
2345static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2346 struct kobj_attribute *attr, const char *buf, size_t len)
2347{
2348 return nr_hugepages_store_common(true, kobj, buf, len);
2349}
2350HSTATE_ATTR(nr_hugepages_mempolicy);
2351#endif
2352
2353
2354static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2355 struct kobj_attribute *attr, char *buf)
2356{
2357 struct hstate *h = kobj_to_hstate(kobj, NULL);
2358 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2359}
2360
2361static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2362 struct kobj_attribute *attr, const char *buf, size_t count)
2363{
2364 int err;
2365 unsigned long input;
2366 struct hstate *h = kobj_to_hstate(kobj, NULL);
2367
2368 if (hstate_is_gigantic(h))
2369 return -EINVAL;
2370
2371 err = kstrtoul(buf, 10, &input);
2372 if (err)
2373 return err;
2374
2375 spin_lock(&hugetlb_lock);
2376 h->nr_overcommit_huge_pages = input;
2377 spin_unlock(&hugetlb_lock);
2378
2379 return count;
2380}
2381HSTATE_ATTR(nr_overcommit_hugepages);
2382
2383static ssize_t free_hugepages_show(struct kobject *kobj,
2384 struct kobj_attribute *attr, char *buf)
2385{
2386 struct hstate *h;
2387 unsigned long free_huge_pages;
2388 int nid;
2389
2390 h = kobj_to_hstate(kobj, &nid);
2391 if (nid == NUMA_NO_NODE)
2392 free_huge_pages = h->free_huge_pages;
2393 else
2394 free_huge_pages = h->free_huge_pages_node[nid];
2395
2396 return sprintf(buf, "%lu\n", free_huge_pages);
2397}
2398HSTATE_ATTR_RO(free_hugepages);
2399
2400static ssize_t resv_hugepages_show(struct kobject *kobj,
2401 struct kobj_attribute *attr, char *buf)
2402{
2403 struct hstate *h = kobj_to_hstate(kobj, NULL);
2404 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2405}
2406HSTATE_ATTR_RO(resv_hugepages);
2407
2408static ssize_t surplus_hugepages_show(struct kobject *kobj,
2409 struct kobj_attribute *attr, char *buf)
2410{
2411 struct hstate *h;
2412 unsigned long surplus_huge_pages;
2413 int nid;
2414
2415 h = kobj_to_hstate(kobj, &nid);
2416 if (nid == NUMA_NO_NODE)
2417 surplus_huge_pages = h->surplus_huge_pages;
2418 else
2419 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2420
2421 return sprintf(buf, "%lu\n", surplus_huge_pages);
2422}
2423HSTATE_ATTR_RO(surplus_hugepages);
2424
2425static struct attribute *hstate_attrs[] = {
2426 &nr_hugepages_attr.attr,
2427 &nr_overcommit_hugepages_attr.attr,
2428 &free_hugepages_attr.attr,
2429 &resv_hugepages_attr.attr,
2430 &surplus_hugepages_attr.attr,
2431#ifdef CONFIG_NUMA
2432 &nr_hugepages_mempolicy_attr.attr,
2433#endif
2434 NULL,
2435};
2436
2437static struct attribute_group hstate_attr_group = {
2438 .attrs = hstate_attrs,
2439};
2440
2441static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2442 struct kobject **hstate_kobjs,
2443 struct attribute_group *hstate_attr_group)
2444{
2445 int retval;
2446 int hi = hstate_index(h);
2447
2448 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2449 if (!hstate_kobjs[hi])
2450 return -ENOMEM;
2451
2452 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2453 if (retval)
2454 kobject_put(hstate_kobjs[hi]);
2455
2456 return retval;
2457}
2458
2459static void __init hugetlb_sysfs_init(void)
2460{
2461 struct hstate *h;
2462 int err;
2463
2464 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2465 if (!hugepages_kobj)
2466 return;
2467
2468 for_each_hstate(h) {
2469 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2470 hstate_kobjs, &hstate_attr_group);
2471 if (err)
2472 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2473 }
2474}
2475
2476#ifdef CONFIG_NUMA
2477
2478/*
2479 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2480 * with node devices in node_devices[] using a parallel array. The array
2481 * index of a node device or _hstate == node id.
2482 * This is here to avoid any static dependency of the node device driver, in
2483 * the base kernel, on the hugetlb module.
2484 */
2485struct node_hstate {
2486 struct kobject *hugepages_kobj;
2487 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2488};
2489static struct node_hstate node_hstates[MAX_NUMNODES];
2490
2491/*
2492 * A subset of global hstate attributes for node devices
2493 */
2494static struct attribute *per_node_hstate_attrs[] = {
2495 &nr_hugepages_attr.attr,
2496 &free_hugepages_attr.attr,
2497 &surplus_hugepages_attr.attr,
2498 NULL,
2499};
2500
2501static struct attribute_group per_node_hstate_attr_group = {
2502 .attrs = per_node_hstate_attrs,
2503};
2504
2505/*
2506 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2507 * Returns node id via non-NULL nidp.
2508 */
2509static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2510{
2511 int nid;
2512
2513 for (nid = 0; nid < nr_node_ids; nid++) {
2514 struct node_hstate *nhs = &node_hstates[nid];
2515 int i;
2516 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2517 if (nhs->hstate_kobjs[i] == kobj) {
2518 if (nidp)
2519 *nidp = nid;
2520 return &hstates[i];
2521 }
2522 }
2523
2524 BUG();
2525 return NULL;
2526}
2527
2528/*
2529 * Unregister hstate attributes from a single node device.
2530 * No-op if no hstate attributes attached.
2531 */
2532static void hugetlb_unregister_node(struct node *node)
2533{
2534 struct hstate *h;
2535 struct node_hstate *nhs = &node_hstates[node->dev.id];
2536
2537 if (!nhs->hugepages_kobj)
2538 return; /* no hstate attributes */
2539
2540 for_each_hstate(h) {
2541 int idx = hstate_index(h);
2542 if (nhs->hstate_kobjs[idx]) {
2543 kobject_put(nhs->hstate_kobjs[idx]);
2544 nhs->hstate_kobjs[idx] = NULL;
2545 }
2546 }
2547
2548 kobject_put(nhs->hugepages_kobj);
2549 nhs->hugepages_kobj = NULL;
2550}
2551
2552
2553/*
2554 * Register hstate attributes for a single node device.
2555 * No-op if attributes already registered.
2556 */
2557static void hugetlb_register_node(struct node *node)
2558{
2559 struct hstate *h;
2560 struct node_hstate *nhs = &node_hstates[node->dev.id];
2561 int err;
2562
2563 if (nhs->hugepages_kobj)
2564 return; /* already allocated */
2565
2566 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2567 &node->dev.kobj);
2568 if (!nhs->hugepages_kobj)
2569 return;
2570
2571 for_each_hstate(h) {
2572 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2573 nhs->hstate_kobjs,
2574 &per_node_hstate_attr_group);
2575 if (err) {
2576 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2577 h->name, node->dev.id);
2578 hugetlb_unregister_node(node);
2579 break;
2580 }
2581 }
2582}
2583
2584/*
2585 * hugetlb init time: register hstate attributes for all registered node
2586 * devices of nodes that have memory. All on-line nodes should have
2587 * registered their associated device by this time.
2588 */
2589static void __init hugetlb_register_all_nodes(void)
2590{
2591 int nid;
2592
2593 for_each_node_state(nid, N_MEMORY) {
2594 struct node *node = node_devices[nid];
2595 if (node->dev.id == nid)
2596 hugetlb_register_node(node);
2597 }
2598
2599 /*
2600 * Let the node device driver know we're here so it can
2601 * [un]register hstate attributes on node hotplug.
2602 */
2603 register_hugetlbfs_with_node(hugetlb_register_node,
2604 hugetlb_unregister_node);
2605}
2606#else /* !CONFIG_NUMA */
2607
2608static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2609{
2610 BUG();
2611 if (nidp)
2612 *nidp = -1;
2613 return NULL;
2614}
2615
2616static void hugetlb_register_all_nodes(void) { }
2617
2618#endif
2619
2620static int __init hugetlb_init(void)
2621{
2622 int i;
2623
2624 if (!hugepages_supported())
2625 return 0;
2626
2627 if (!size_to_hstate(default_hstate_size)) {
2628 default_hstate_size = HPAGE_SIZE;
2629 if (!size_to_hstate(default_hstate_size))
2630 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2631 }
2632 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2633 if (default_hstate_max_huge_pages) {
2634 if (!default_hstate.max_huge_pages)
2635 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2636 }
2637
2638 hugetlb_init_hstates();
2639 gather_bootmem_prealloc();
2640 report_hugepages();
2641
2642 hugetlb_sysfs_init();
2643 hugetlb_register_all_nodes();
2644 hugetlb_cgroup_file_init();
2645
2646#ifdef CONFIG_SMP
2647 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2648#else
2649 num_fault_mutexes = 1;
2650#endif
2651 hugetlb_fault_mutex_table =
2652 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2653 BUG_ON(!hugetlb_fault_mutex_table);
2654
2655 for (i = 0; i < num_fault_mutexes; i++)
2656 mutex_init(&hugetlb_fault_mutex_table[i]);
2657 return 0;
2658}
2659subsys_initcall(hugetlb_init);
2660
2661/* Should be called on processing a hugepagesz=... option */
2662void __init hugetlb_add_hstate(unsigned int order)
2663{
2664 struct hstate *h;
2665 unsigned long i;
2666
2667 if (size_to_hstate(PAGE_SIZE << order)) {
2668 pr_warn("hugepagesz= specified twice, ignoring\n");
2669 return;
2670 }
2671 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2672 BUG_ON(order == 0);
2673 h = &hstates[hugetlb_max_hstate++];
2674 h->order = order;
2675 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2676 h->nr_huge_pages = 0;
2677 h->free_huge_pages = 0;
2678 for (i = 0; i < MAX_NUMNODES; ++i)
2679 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2680 INIT_LIST_HEAD(&h->hugepage_activelist);
2681 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2682 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2683 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2684 huge_page_size(h)/1024);
2685
2686 parsed_hstate = h;
2687}
2688
2689static int __init hugetlb_nrpages_setup(char *s)
2690{
2691 unsigned long *mhp;
2692 static unsigned long *last_mhp;
2693
2694 /*
2695 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2696 * so this hugepages= parameter goes to the "default hstate".
2697 */
2698 if (!hugetlb_max_hstate)
2699 mhp = &default_hstate_max_huge_pages;
2700 else
2701 mhp = &parsed_hstate->max_huge_pages;
2702
2703 if (mhp == last_mhp) {
2704 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2705 return 1;
2706 }
2707
2708 if (sscanf(s, "%lu", mhp) <= 0)
2709 *mhp = 0;
2710
2711 /*
2712 * Global state is always initialized later in hugetlb_init.
2713 * But we need to allocate >= MAX_ORDER hstates here early to still
2714 * use the bootmem allocator.
2715 */
2716 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2717 hugetlb_hstate_alloc_pages(parsed_hstate);
2718
2719 last_mhp = mhp;
2720
2721 return 1;
2722}
2723__setup("hugepages=", hugetlb_nrpages_setup);
2724
2725static int __init hugetlb_default_setup(char *s)
2726{
2727 default_hstate_size = memparse(s, &s);
2728 return 1;
2729}
2730__setup("default_hugepagesz=", hugetlb_default_setup);
2731
2732static unsigned int cpuset_mems_nr(unsigned int *array)
2733{
2734 int node;
2735 unsigned int nr = 0;
2736
2737 for_each_node_mask(node, cpuset_current_mems_allowed)
2738 nr += array[node];
2739
2740 return nr;
2741}
2742
2743#ifdef CONFIG_SYSCTL
2744static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2745 struct ctl_table *table, int write,
2746 void __user *buffer, size_t *length, loff_t *ppos)
2747{
2748 struct hstate *h = &default_hstate;
2749 unsigned long tmp = h->max_huge_pages;
2750 int ret;
2751
2752 if (!hugepages_supported())
2753 return -EOPNOTSUPP;
2754
2755 table->data = &tmp;
2756 table->maxlen = sizeof(unsigned long);
2757 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2758 if (ret)
2759 goto out;
2760
2761 if (write)
2762 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2763 NUMA_NO_NODE, tmp, *length);
2764out:
2765 return ret;
2766}
2767
2768int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2769 void __user *buffer, size_t *length, loff_t *ppos)
2770{
2771
2772 return hugetlb_sysctl_handler_common(false, table, write,
2773 buffer, length, ppos);
2774}
2775
2776#ifdef CONFIG_NUMA
2777int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2778 void __user *buffer, size_t *length, loff_t *ppos)
2779{
2780 return hugetlb_sysctl_handler_common(true, table, write,
2781 buffer, length, ppos);
2782}
2783#endif /* CONFIG_NUMA */
2784
2785int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2786 void __user *buffer,
2787 size_t *length, loff_t *ppos)
2788{
2789 struct hstate *h = &default_hstate;
2790 unsigned long tmp;
2791 int ret;
2792
2793 if (!hugepages_supported())
2794 return -EOPNOTSUPP;
2795
2796 tmp = h->nr_overcommit_huge_pages;
2797
2798 if (write && hstate_is_gigantic(h))
2799 return -EINVAL;
2800
2801 table->data = &tmp;
2802 table->maxlen = sizeof(unsigned long);
2803 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2804 if (ret)
2805 goto out;
2806
2807 if (write) {
2808 spin_lock(&hugetlb_lock);
2809 h->nr_overcommit_huge_pages = tmp;
2810 spin_unlock(&hugetlb_lock);
2811 }
2812out:
2813 return ret;
2814}
2815
2816#endif /* CONFIG_SYSCTL */
2817
2818void hugetlb_report_meminfo(struct seq_file *m)
2819{
2820 struct hstate *h = &default_hstate;
2821 if (!hugepages_supported())
2822 return;
2823 seq_printf(m,
2824 "HugePages_Total: %5lu\n"
2825 "HugePages_Free: %5lu\n"
2826 "HugePages_Rsvd: %5lu\n"
2827 "HugePages_Surp: %5lu\n"
2828 "Hugepagesize: %8lu kB\n",
2829 h->nr_huge_pages,
2830 h->free_huge_pages,
2831 h->resv_huge_pages,
2832 h->surplus_huge_pages,
2833 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2834}
2835
2836int hugetlb_report_node_meminfo(int nid, char *buf)
2837{
2838 struct hstate *h = &default_hstate;
2839 if (!hugepages_supported())
2840 return 0;
2841 return sprintf(buf,
2842 "Node %d HugePages_Total: %5u\n"
2843 "Node %d HugePages_Free: %5u\n"
2844 "Node %d HugePages_Surp: %5u\n",
2845 nid, h->nr_huge_pages_node[nid],
2846 nid, h->free_huge_pages_node[nid],
2847 nid, h->surplus_huge_pages_node[nid]);
2848}
2849
2850void hugetlb_show_meminfo(void)
2851{
2852 struct hstate *h;
2853 int nid;
2854
2855 if (!hugepages_supported())
2856 return;
2857
2858 for_each_node_state(nid, N_MEMORY)
2859 for_each_hstate(h)
2860 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2861 nid,
2862 h->nr_huge_pages_node[nid],
2863 h->free_huge_pages_node[nid],
2864 h->surplus_huge_pages_node[nid],
2865 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2866}
2867
2868void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
2869{
2870 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
2871 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
2872}
2873
2874/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2875unsigned long hugetlb_total_pages(void)
2876{
2877 struct hstate *h;
2878 unsigned long nr_total_pages = 0;
2879
2880 for_each_hstate(h)
2881 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2882 return nr_total_pages;
2883}
2884
2885static int hugetlb_acct_memory(struct hstate *h, long delta)
2886{
2887 int ret = -ENOMEM;
2888
2889 spin_lock(&hugetlb_lock);
2890 /*
2891 * When cpuset is configured, it breaks the strict hugetlb page
2892 * reservation as the accounting is done on a global variable. Such
2893 * reservation is completely rubbish in the presence of cpuset because
2894 * the reservation is not checked against page availability for the
2895 * current cpuset. Application can still potentially OOM'ed by kernel
2896 * with lack of free htlb page in cpuset that the task is in.
2897 * Attempt to enforce strict accounting with cpuset is almost
2898 * impossible (or too ugly) because cpuset is too fluid that
2899 * task or memory node can be dynamically moved between cpusets.
2900 *
2901 * The change of semantics for shared hugetlb mapping with cpuset is
2902 * undesirable. However, in order to preserve some of the semantics,
2903 * we fall back to check against current free page availability as
2904 * a best attempt and hopefully to minimize the impact of changing
2905 * semantics that cpuset has.
2906 */
2907 if (delta > 0) {
2908 if (gather_surplus_pages(h, delta) < 0)
2909 goto out;
2910
2911 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2912 return_unused_surplus_pages(h, delta);
2913 goto out;
2914 }
2915 }
2916
2917 ret = 0;
2918 if (delta < 0)
2919 return_unused_surplus_pages(h, (unsigned long) -delta);
2920
2921out:
2922 spin_unlock(&hugetlb_lock);
2923 return ret;
2924}
2925
2926static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2927{
2928 struct resv_map *resv = vma_resv_map(vma);
2929
2930 /*
2931 * This new VMA should share its siblings reservation map if present.
2932 * The VMA will only ever have a valid reservation map pointer where
2933 * it is being copied for another still existing VMA. As that VMA
2934 * has a reference to the reservation map it cannot disappear until
2935 * after this open call completes. It is therefore safe to take a
2936 * new reference here without additional locking.
2937 */
2938 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2939 kref_get(&resv->refs);
2940}
2941
2942static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2943{
2944 struct hstate *h = hstate_vma(vma);
2945 struct resv_map *resv = vma_resv_map(vma);
2946 struct hugepage_subpool *spool = subpool_vma(vma);
2947 unsigned long reserve, start, end;
2948 long gbl_reserve;
2949
2950 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2951 return;
2952
2953 start = vma_hugecache_offset(h, vma, vma->vm_start);
2954 end = vma_hugecache_offset(h, vma, vma->vm_end);
2955
2956 reserve = (end - start) - region_count(resv, start, end);
2957
2958 kref_put(&resv->refs, resv_map_release);
2959
2960 if (reserve) {
2961 /*
2962 * Decrement reserve counts. The global reserve count may be
2963 * adjusted if the subpool has a minimum size.
2964 */
2965 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2966 hugetlb_acct_memory(h, -gbl_reserve);
2967 }
2968}
2969
2970/*
2971 * We cannot handle pagefaults against hugetlb pages at all. They cause
2972 * handle_mm_fault() to try to instantiate regular-sized pages in the
2973 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2974 * this far.
2975 */
2976static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2977{
2978 BUG();
2979 return 0;
2980}
2981
2982const struct vm_operations_struct hugetlb_vm_ops = {
2983 .fault = hugetlb_vm_op_fault,
2984 .open = hugetlb_vm_op_open,
2985 .close = hugetlb_vm_op_close,
2986};
2987
2988static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2989 int writable)
2990{
2991 pte_t entry;
2992
2993 if (writable) {
2994 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2995 vma->vm_page_prot)));
2996 } else {
2997 entry = huge_pte_wrprotect(mk_huge_pte(page,
2998 vma->vm_page_prot));
2999 }
3000 entry = pte_mkyoung(entry);
3001 entry = pte_mkhuge(entry);
3002 entry = arch_make_huge_pte(entry, vma, page, writable);
3003
3004 return entry;
3005}
3006
3007static void set_huge_ptep_writable(struct vm_area_struct *vma,
3008 unsigned long address, pte_t *ptep)
3009{
3010 pte_t entry;
3011
3012 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3013 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3014 update_mmu_cache(vma, address, ptep);
3015}
3016
3017static int is_hugetlb_entry_migration(pte_t pte)
3018{
3019 swp_entry_t swp;
3020
3021 if (huge_pte_none(pte) || pte_present(pte))
3022 return 0;
3023 swp = pte_to_swp_entry(pte);
3024 if (non_swap_entry(swp) && is_migration_entry(swp))
3025 return 1;
3026 else
3027 return 0;
3028}
3029
3030static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3031{
3032 swp_entry_t swp;
3033
3034 if (huge_pte_none(pte) || pte_present(pte))
3035 return 0;
3036 swp = pte_to_swp_entry(pte);
3037 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3038 return 1;
3039 else
3040 return 0;
3041}
3042
3043int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3044 struct vm_area_struct *vma)
3045{
3046 pte_t *src_pte, *dst_pte, entry;
3047 struct page *ptepage;
3048 unsigned long addr;
3049 int cow;
3050 struct hstate *h = hstate_vma(vma);
3051 unsigned long sz = huge_page_size(h);
3052 unsigned long mmun_start; /* For mmu_notifiers */
3053 unsigned long mmun_end; /* For mmu_notifiers */
3054 int ret = 0;
3055
3056 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3057
3058 mmun_start = vma->vm_start;
3059 mmun_end = vma->vm_end;
3060 if (cow)
3061 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3062
3063 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3064 spinlock_t *src_ptl, *dst_ptl;
3065 src_pte = huge_pte_offset(src, addr);
3066 if (!src_pte)
3067 continue;
3068 dst_pte = huge_pte_alloc(dst, addr, sz);
3069 if (!dst_pte) {
3070 ret = -ENOMEM;
3071 break;
3072 }
3073
3074 /* If the pagetables are shared don't copy or take references */
3075 if (dst_pte == src_pte)
3076 continue;
3077
3078 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3079 src_ptl = huge_pte_lockptr(h, src, src_pte);
3080 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3081 entry = huge_ptep_get(src_pte);
3082 if (huge_pte_none(entry)) { /* skip none entry */
3083 ;
3084 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3085 is_hugetlb_entry_hwpoisoned(entry))) {
3086 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3087
3088 if (is_write_migration_entry(swp_entry) && cow) {
3089 /*
3090 * COW mappings require pages in both
3091 * parent and child to be set to read.
3092 */
3093 make_migration_entry_read(&swp_entry);
3094 entry = swp_entry_to_pte(swp_entry);
3095 set_huge_pte_at(src, addr, src_pte, entry);
3096 }
3097 set_huge_pte_at(dst, addr, dst_pte, entry);
3098 } else {
3099 if (cow) {
3100 huge_ptep_set_wrprotect(src, addr, src_pte);
3101 mmu_notifier_invalidate_range(src, mmun_start,
3102 mmun_end);
3103 }
3104 entry = huge_ptep_get(src_pte);
3105 ptepage = pte_page(entry);
3106 get_page(ptepage);
3107 page_dup_rmap(ptepage, true);
3108 set_huge_pte_at(dst, addr, dst_pte, entry);
3109 hugetlb_count_add(pages_per_huge_page(h), dst);
3110 }
3111 spin_unlock(src_ptl);
3112 spin_unlock(dst_ptl);
3113 }
3114
3115 if (cow)
3116 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3117
3118 return ret;
3119}
3120
3121void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3122 unsigned long start, unsigned long end,
3123 struct page *ref_page)
3124{
3125 int force_flush = 0;
3126 struct mm_struct *mm = vma->vm_mm;
3127 unsigned long address;
3128 pte_t *ptep;
3129 pte_t pte;
3130 spinlock_t *ptl;
3131 struct page *page;
3132 struct hstate *h = hstate_vma(vma);
3133 unsigned long sz = huge_page_size(h);
3134 const unsigned long mmun_start = start; /* For mmu_notifiers */
3135 const unsigned long mmun_end = end; /* For mmu_notifiers */
3136
3137 WARN_ON(!is_vm_hugetlb_page(vma));
3138 BUG_ON(start & ~huge_page_mask(h));
3139 BUG_ON(end & ~huge_page_mask(h));
3140
3141 tlb_start_vma(tlb, vma);
3142 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3143 address = start;
3144again:
3145 for (; address < end; address += sz) {
3146 ptep = huge_pte_offset(mm, address);
3147 if (!ptep)
3148 continue;
3149
3150 ptl = huge_pte_lock(h, mm, ptep);
3151 if (huge_pmd_unshare(mm, &address, ptep))
3152 goto unlock;
3153
3154 pte = huge_ptep_get(ptep);
3155 if (huge_pte_none(pte))
3156 goto unlock;
3157
3158 /*
3159 * Migrating hugepage or HWPoisoned hugepage is already
3160 * unmapped and its refcount is dropped, so just clear pte here.
3161 */
3162 if (unlikely(!pte_present(pte))) {
3163 huge_pte_clear(mm, address, ptep);
3164 goto unlock;
3165 }
3166
3167 page = pte_page(pte);
3168 /*
3169 * If a reference page is supplied, it is because a specific
3170 * page is being unmapped, not a range. Ensure the page we
3171 * are about to unmap is the actual page of interest.
3172 */
3173 if (ref_page) {
3174 if (page != ref_page)
3175 goto unlock;
3176
3177 /*
3178 * Mark the VMA as having unmapped its page so that
3179 * future faults in this VMA will fail rather than
3180 * looking like data was lost
3181 */
3182 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3183 }
3184
3185 pte = huge_ptep_get_and_clear(mm, address, ptep);
3186 tlb_remove_tlb_entry(tlb, ptep, address);
3187 if (huge_pte_dirty(pte))
3188 set_page_dirty(page);
3189
3190 hugetlb_count_sub(pages_per_huge_page(h), mm);
3191 page_remove_rmap(page, true);
3192 force_flush = !__tlb_remove_page(tlb, page);
3193 if (force_flush) {
3194 address += sz;
3195 spin_unlock(ptl);
3196 break;
3197 }
3198 /* Bail out after unmapping reference page if supplied */
3199 if (ref_page) {
3200 spin_unlock(ptl);
3201 break;
3202 }
3203unlock:
3204 spin_unlock(ptl);
3205 }
3206 /*
3207 * mmu_gather ran out of room to batch pages, we break out of
3208 * the PTE lock to avoid doing the potential expensive TLB invalidate
3209 * and page-free while holding it.
3210 */
3211 if (force_flush) {
3212 force_flush = 0;
3213 tlb_flush_mmu(tlb);
3214 if (address < end && !ref_page)
3215 goto again;
3216 }
3217 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3218 tlb_end_vma(tlb, vma);
3219}
3220
3221void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3222 struct vm_area_struct *vma, unsigned long start,
3223 unsigned long end, struct page *ref_page)
3224{
3225 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3226
3227 /*
3228 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3229 * test will fail on a vma being torn down, and not grab a page table
3230 * on its way out. We're lucky that the flag has such an appropriate
3231 * name, and can in fact be safely cleared here. We could clear it
3232 * before the __unmap_hugepage_range above, but all that's necessary
3233 * is to clear it before releasing the i_mmap_rwsem. This works
3234 * because in the context this is called, the VMA is about to be
3235 * destroyed and the i_mmap_rwsem is held.
3236 */
3237 vma->vm_flags &= ~VM_MAYSHARE;
3238}
3239
3240void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3241 unsigned long end, struct page *ref_page)
3242{
3243 struct mm_struct *mm;
3244 struct mmu_gather tlb;
3245
3246 mm = vma->vm_mm;
3247
3248 tlb_gather_mmu(&tlb, mm, start, end);
3249 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3250 tlb_finish_mmu(&tlb, start, end);
3251}
3252
3253/*
3254 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3255 * mappping it owns the reserve page for. The intention is to unmap the page
3256 * from other VMAs and let the children be SIGKILLed if they are faulting the
3257 * same region.
3258 */
3259static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3260 struct page *page, unsigned long address)
3261{
3262 struct hstate *h = hstate_vma(vma);
3263 struct vm_area_struct *iter_vma;
3264 struct address_space *mapping;
3265 pgoff_t pgoff;
3266
3267 /*
3268 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3269 * from page cache lookup which is in HPAGE_SIZE units.
3270 */
3271 address = address & huge_page_mask(h);
3272 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3273 vma->vm_pgoff;
3274 mapping = file_inode(vma->vm_file)->i_mapping;
3275
3276 /*
3277 * Take the mapping lock for the duration of the table walk. As
3278 * this mapping should be shared between all the VMAs,
3279 * __unmap_hugepage_range() is called as the lock is already held
3280 */
3281 i_mmap_lock_write(mapping);
3282 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3283 /* Do not unmap the current VMA */
3284 if (iter_vma == vma)
3285 continue;
3286
3287 /*
3288 * Shared VMAs have their own reserves and do not affect
3289 * MAP_PRIVATE accounting but it is possible that a shared
3290 * VMA is using the same page so check and skip such VMAs.
3291 */
3292 if (iter_vma->vm_flags & VM_MAYSHARE)
3293 continue;
3294
3295 /*
3296 * Unmap the page from other VMAs without their own reserves.
3297 * They get marked to be SIGKILLed if they fault in these
3298 * areas. This is because a future no-page fault on this VMA
3299 * could insert a zeroed page instead of the data existing
3300 * from the time of fork. This would look like data corruption
3301 */
3302 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3303 unmap_hugepage_range(iter_vma, address,
3304 address + huge_page_size(h), page);
3305 }
3306 i_mmap_unlock_write(mapping);
3307}
3308
3309/*
3310 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3311 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3312 * cannot race with other handlers or page migration.
3313 * Keep the pte_same checks anyway to make transition from the mutex easier.
3314 */
3315static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3316 unsigned long address, pte_t *ptep, pte_t pte,
3317 struct page *pagecache_page, spinlock_t *ptl)
3318{
3319 struct hstate *h = hstate_vma(vma);
3320 struct page *old_page, *new_page;
3321 int ret = 0, outside_reserve = 0;
3322 unsigned long mmun_start; /* For mmu_notifiers */
3323 unsigned long mmun_end; /* For mmu_notifiers */
3324
3325 old_page = pte_page(pte);
3326
3327retry_avoidcopy:
3328 /* If no-one else is actually using this page, avoid the copy
3329 * and just make the page writable */
3330 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3331 page_move_anon_rmap(old_page, vma, address);
3332 set_huge_ptep_writable(vma, address, ptep);
3333 return 0;
3334 }
3335
3336 /*
3337 * If the process that created a MAP_PRIVATE mapping is about to
3338 * perform a COW due to a shared page count, attempt to satisfy
3339 * the allocation without using the existing reserves. The pagecache
3340 * page is used to determine if the reserve at this address was
3341 * consumed or not. If reserves were used, a partial faulted mapping
3342 * at the time of fork() could consume its reserves on COW instead
3343 * of the full address range.
3344 */
3345 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3346 old_page != pagecache_page)
3347 outside_reserve = 1;
3348
3349 get_page(old_page);
3350
3351 /*
3352 * Drop page table lock as buddy allocator may be called. It will
3353 * be acquired again before returning to the caller, as expected.
3354 */
3355 spin_unlock(ptl);
3356 new_page = alloc_huge_page(vma, address, outside_reserve);
3357
3358 if (IS_ERR(new_page)) {
3359 /*
3360 * If a process owning a MAP_PRIVATE mapping fails to COW,
3361 * it is due to references held by a child and an insufficient
3362 * huge page pool. To guarantee the original mappers
3363 * reliability, unmap the page from child processes. The child
3364 * may get SIGKILLed if it later faults.
3365 */
3366 if (outside_reserve) {
3367 put_page(old_page);
3368 BUG_ON(huge_pte_none(pte));
3369 unmap_ref_private(mm, vma, old_page, address);
3370 BUG_ON(huge_pte_none(pte));
3371 spin_lock(ptl);
3372 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3373 if (likely(ptep &&
3374 pte_same(huge_ptep_get(ptep), pte)))
3375 goto retry_avoidcopy;
3376 /*
3377 * race occurs while re-acquiring page table
3378 * lock, and our job is done.
3379 */
3380 return 0;
3381 }
3382
3383 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3384 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3385 goto out_release_old;
3386 }
3387
3388 /*
3389 * When the original hugepage is shared one, it does not have
3390 * anon_vma prepared.
3391 */
3392 if (unlikely(anon_vma_prepare(vma))) {
3393 ret = VM_FAULT_OOM;
3394 goto out_release_all;
3395 }
3396
3397 copy_user_huge_page(new_page, old_page, address, vma,
3398 pages_per_huge_page(h));
3399 __SetPageUptodate(new_page);
3400 set_page_huge_active(new_page);
3401
3402 mmun_start = address & huge_page_mask(h);
3403 mmun_end = mmun_start + huge_page_size(h);
3404 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3405
3406 /*
3407 * Retake the page table lock to check for racing updates
3408 * before the page tables are altered
3409 */
3410 spin_lock(ptl);
3411 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3412 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3413 ClearPagePrivate(new_page);
3414
3415 /* Break COW */
3416 huge_ptep_clear_flush(vma, address, ptep);
3417 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3418 set_huge_pte_at(mm, address, ptep,
3419 make_huge_pte(vma, new_page, 1));
3420 page_remove_rmap(old_page, true);
3421 hugepage_add_new_anon_rmap(new_page, vma, address);
3422 /* Make the old page be freed below */
3423 new_page = old_page;
3424 }
3425 spin_unlock(ptl);
3426 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3427out_release_all:
3428 put_page(new_page);
3429out_release_old:
3430 put_page(old_page);
3431
3432 spin_lock(ptl); /* Caller expects lock to be held */
3433 return ret;
3434}
3435
3436/* Return the pagecache page at a given address within a VMA */
3437static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3438 struct vm_area_struct *vma, unsigned long address)
3439{
3440 struct address_space *mapping;
3441 pgoff_t idx;
3442
3443 mapping = vma->vm_file->f_mapping;
3444 idx = vma_hugecache_offset(h, vma, address);
3445
3446 return find_lock_page(mapping, idx);
3447}
3448
3449/*
3450 * Return whether there is a pagecache page to back given address within VMA.
3451 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3452 */
3453static bool hugetlbfs_pagecache_present(struct hstate *h,
3454 struct vm_area_struct *vma, unsigned long address)
3455{
3456 struct address_space *mapping;
3457 pgoff_t idx;
3458 struct page *page;
3459
3460 mapping = vma->vm_file->f_mapping;
3461 idx = vma_hugecache_offset(h, vma, address);
3462
3463 page = find_get_page(mapping, idx);
3464 if (page)
3465 put_page(page);
3466 return page != NULL;
3467}
3468
3469int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3470 pgoff_t idx)
3471{
3472 struct inode *inode = mapping->host;
3473 struct hstate *h = hstate_inode(inode);
3474 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3475
3476 if (err)
3477 return err;
3478 ClearPagePrivate(page);
3479
3480 spin_lock(&inode->i_lock);
3481 inode->i_blocks += blocks_per_huge_page(h);
3482 spin_unlock(&inode->i_lock);
3483 return 0;
3484}
3485
3486static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3487 struct address_space *mapping, pgoff_t idx,
3488 unsigned long address, pte_t *ptep, unsigned int flags)
3489{
3490 struct hstate *h = hstate_vma(vma);
3491 int ret = VM_FAULT_SIGBUS;
3492 int anon_rmap = 0;
3493 unsigned long size;
3494 struct page *page;
3495 pte_t new_pte;
3496 spinlock_t *ptl;
3497
3498 /*
3499 * Currently, we are forced to kill the process in the event the
3500 * original mapper has unmapped pages from the child due to a failed
3501 * COW. Warn that such a situation has occurred as it may not be obvious
3502 */
3503 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3504 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3505 current->pid);
3506 return ret;
3507 }
3508
3509 /*
3510 * Use page lock to guard against racing truncation
3511 * before we get page_table_lock.
3512 */
3513retry:
3514 page = find_lock_page(mapping, idx);
3515 if (!page) {
3516 size = i_size_read(mapping->host) >> huge_page_shift(h);
3517 if (idx >= size)
3518 goto out;
3519 page = alloc_huge_page(vma, address, 0);
3520 if (IS_ERR(page)) {
3521 ret = PTR_ERR(page);
3522 if (ret == -ENOMEM)
3523 ret = VM_FAULT_OOM;
3524 else
3525 ret = VM_FAULT_SIGBUS;
3526 goto out;
3527 }
3528 clear_huge_page(page, address, pages_per_huge_page(h));
3529 __SetPageUptodate(page);
3530 set_page_huge_active(page);
3531
3532 if (vma->vm_flags & VM_MAYSHARE) {
3533 int err = huge_add_to_page_cache(page, mapping, idx);
3534 if (err) {
3535 put_page(page);
3536 if (err == -EEXIST)
3537 goto retry;
3538 goto out;
3539 }
3540 } else {
3541 lock_page(page);
3542 if (unlikely(anon_vma_prepare(vma))) {
3543 ret = VM_FAULT_OOM;
3544 goto backout_unlocked;
3545 }
3546 anon_rmap = 1;
3547 }
3548 } else {
3549 /*
3550 * If memory error occurs between mmap() and fault, some process
3551 * don't have hwpoisoned swap entry for errored virtual address.
3552 * So we need to block hugepage fault by PG_hwpoison bit check.
3553 */
3554 if (unlikely(PageHWPoison(page))) {
3555 ret = VM_FAULT_HWPOISON |
3556 VM_FAULT_SET_HINDEX(hstate_index(h));
3557 goto backout_unlocked;
3558 }
3559 }
3560
3561 /*
3562 * If we are going to COW a private mapping later, we examine the
3563 * pending reservations for this page now. This will ensure that
3564 * any allocations necessary to record that reservation occur outside
3565 * the spinlock.
3566 */
3567 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3568 if (vma_needs_reservation(h, vma, address) < 0) {
3569 ret = VM_FAULT_OOM;
3570 goto backout_unlocked;
3571 }
3572 /* Just decrements count, does not deallocate */
3573 vma_end_reservation(h, vma, address);
3574 }
3575
3576 ptl = huge_pte_lockptr(h, mm, ptep);
3577 spin_lock(ptl);
3578 size = i_size_read(mapping->host) >> huge_page_shift(h);
3579 if (idx >= size)
3580 goto backout;
3581
3582 ret = 0;
3583 if (!huge_pte_none(huge_ptep_get(ptep)))
3584 goto backout;
3585
3586 if (anon_rmap) {
3587 ClearPagePrivate(page);
3588 hugepage_add_new_anon_rmap(page, vma, address);
3589 } else
3590 page_dup_rmap(page, true);
3591 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3592 && (vma->vm_flags & VM_SHARED)));
3593 set_huge_pte_at(mm, address, ptep, new_pte);
3594
3595 hugetlb_count_add(pages_per_huge_page(h), mm);
3596 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3597 /* Optimization, do the COW without a second fault */
3598 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3599 }
3600
3601 spin_unlock(ptl);
3602 unlock_page(page);
3603out:
3604 return ret;
3605
3606backout:
3607 spin_unlock(ptl);
3608backout_unlocked:
3609 unlock_page(page);
3610 put_page(page);
3611 goto out;
3612}
3613
3614#ifdef CONFIG_SMP
3615u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3616 struct vm_area_struct *vma,
3617 struct address_space *mapping,
3618 pgoff_t idx, unsigned long address)
3619{
3620 unsigned long key[2];
3621 u32 hash;
3622
3623 if (vma->vm_flags & VM_SHARED) {
3624 key[0] = (unsigned long) mapping;
3625 key[1] = idx;
3626 } else {
3627 key[0] = (unsigned long) mm;
3628 key[1] = address >> huge_page_shift(h);
3629 }
3630
3631 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3632
3633 return hash & (num_fault_mutexes - 1);
3634}
3635#else
3636/*
3637 * For uniprocesor systems we always use a single mutex, so just
3638 * return 0 and avoid the hashing overhead.
3639 */
3640u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3641 struct vm_area_struct *vma,
3642 struct address_space *mapping,
3643 pgoff_t idx, unsigned long address)
3644{
3645 return 0;
3646}
3647#endif
3648
3649int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3650 unsigned long address, unsigned int flags)
3651{
3652 pte_t *ptep, entry;
3653 spinlock_t *ptl;
3654 int ret;
3655 u32 hash;
3656 pgoff_t idx;
3657 struct page *page = NULL;
3658 struct page *pagecache_page = NULL;
3659 struct hstate *h = hstate_vma(vma);
3660 struct address_space *mapping;
3661 int need_wait_lock = 0;
3662
3663 address &= huge_page_mask(h);
3664
3665 ptep = huge_pte_offset(mm, address);
3666 if (ptep) {
3667 entry = huge_ptep_get(ptep);
3668 if (unlikely(is_hugetlb_entry_migration(entry))) {
3669 migration_entry_wait_huge(vma, mm, ptep);
3670 return 0;
3671 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3672 return VM_FAULT_HWPOISON_LARGE |
3673 VM_FAULT_SET_HINDEX(hstate_index(h));
3674 } else {
3675 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3676 if (!ptep)
3677 return VM_FAULT_OOM;
3678 }
3679
3680 mapping = vma->vm_file->f_mapping;
3681 idx = vma_hugecache_offset(h, vma, address);
3682
3683 /*
3684 * Serialize hugepage allocation and instantiation, so that we don't
3685 * get spurious allocation failures if two CPUs race to instantiate
3686 * the same page in the page cache.
3687 */
3688 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3689 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3690
3691 entry = huge_ptep_get(ptep);
3692 if (huge_pte_none(entry)) {
3693 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3694 goto out_mutex;
3695 }
3696
3697 ret = 0;
3698
3699 /*
3700 * entry could be a migration/hwpoison entry at this point, so this
3701 * check prevents the kernel from going below assuming that we have
3702 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3703 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3704 * handle it.
3705 */
3706 if (!pte_present(entry))
3707 goto out_mutex;
3708
3709 /*
3710 * If we are going to COW the mapping later, we examine the pending
3711 * reservations for this page now. This will ensure that any
3712 * allocations necessary to record that reservation occur outside the
3713 * spinlock. For private mappings, we also lookup the pagecache
3714 * page now as it is used to determine if a reservation has been
3715 * consumed.
3716 */
3717 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3718 if (vma_needs_reservation(h, vma, address) < 0) {
3719 ret = VM_FAULT_OOM;
3720 goto out_mutex;
3721 }
3722 /* Just decrements count, does not deallocate */
3723 vma_end_reservation(h, vma, address);
3724
3725 if (!(vma->vm_flags & VM_MAYSHARE))
3726 pagecache_page = hugetlbfs_pagecache_page(h,
3727 vma, address);
3728 }
3729
3730 ptl = huge_pte_lock(h, mm, ptep);
3731
3732 /* Check for a racing update before calling hugetlb_cow */
3733 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3734 goto out_ptl;
3735
3736 /*
3737 * hugetlb_cow() requires page locks of pte_page(entry) and
3738 * pagecache_page, so here we need take the former one
3739 * when page != pagecache_page or !pagecache_page.
3740 */
3741 page = pte_page(entry);
3742 if (page != pagecache_page)
3743 if (!trylock_page(page)) {
3744 need_wait_lock = 1;
3745 goto out_ptl;
3746 }
3747
3748 get_page(page);
3749
3750 if (flags & FAULT_FLAG_WRITE) {
3751 if (!huge_pte_write(entry)) {
3752 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3753 pagecache_page, ptl);
3754 goto out_put_page;
3755 }
3756 entry = huge_pte_mkdirty(entry);
3757 }
3758 entry = pte_mkyoung(entry);
3759 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3760 flags & FAULT_FLAG_WRITE))
3761 update_mmu_cache(vma, address, ptep);
3762out_put_page:
3763 if (page != pagecache_page)
3764 unlock_page(page);
3765 put_page(page);
3766out_ptl:
3767 spin_unlock(ptl);
3768
3769 if (pagecache_page) {
3770 unlock_page(pagecache_page);
3771 put_page(pagecache_page);
3772 }
3773out_mutex:
3774 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3775 /*
3776 * Generally it's safe to hold refcount during waiting page lock. But
3777 * here we just wait to defer the next page fault to avoid busy loop and
3778 * the page is not used after unlocked before returning from the current
3779 * page fault. So we are safe from accessing freed page, even if we wait
3780 * here without taking refcount.
3781 */
3782 if (need_wait_lock)
3783 wait_on_page_locked(page);
3784 return ret;
3785}
3786
3787long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3788 struct page **pages, struct vm_area_struct **vmas,
3789 unsigned long *position, unsigned long *nr_pages,
3790 long i, unsigned int flags)
3791{
3792 unsigned long pfn_offset;
3793 unsigned long vaddr = *position;
3794 unsigned long remainder = *nr_pages;
3795 struct hstate *h = hstate_vma(vma);
3796
3797 while (vaddr < vma->vm_end && remainder) {
3798 pte_t *pte;
3799 spinlock_t *ptl = NULL;
3800 int absent;
3801 struct page *page;
3802
3803 /*
3804 * If we have a pending SIGKILL, don't keep faulting pages and
3805 * potentially allocating memory.
3806 */
3807 if (unlikely(fatal_signal_pending(current))) {
3808 remainder = 0;
3809 break;
3810 }
3811
3812 /*
3813 * Some archs (sparc64, sh*) have multiple pte_ts to
3814 * each hugepage. We have to make sure we get the
3815 * first, for the page indexing below to work.
3816 *
3817 * Note that page table lock is not held when pte is null.
3818 */
3819 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3820 if (pte)
3821 ptl = huge_pte_lock(h, mm, pte);
3822 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3823
3824 /*
3825 * When coredumping, it suits get_dump_page if we just return
3826 * an error where there's an empty slot with no huge pagecache
3827 * to back it. This way, we avoid allocating a hugepage, and
3828 * the sparse dumpfile avoids allocating disk blocks, but its
3829 * huge holes still show up with zeroes where they need to be.
3830 */
3831 if (absent && (flags & FOLL_DUMP) &&
3832 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3833 if (pte)
3834 spin_unlock(ptl);
3835 remainder = 0;
3836 break;
3837 }
3838
3839 /*
3840 * We need call hugetlb_fault for both hugepages under migration
3841 * (in which case hugetlb_fault waits for the migration,) and
3842 * hwpoisoned hugepages (in which case we need to prevent the
3843 * caller from accessing to them.) In order to do this, we use
3844 * here is_swap_pte instead of is_hugetlb_entry_migration and
3845 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3846 * both cases, and because we can't follow correct pages
3847 * directly from any kind of swap entries.
3848 */
3849 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3850 ((flags & FOLL_WRITE) &&
3851 !huge_pte_write(huge_ptep_get(pte)))) {
3852 int ret;
3853
3854 if (pte)
3855 spin_unlock(ptl);
3856 ret = hugetlb_fault(mm, vma, vaddr,
3857 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3858 if (!(ret & VM_FAULT_ERROR))
3859 continue;
3860
3861 remainder = 0;
3862 break;
3863 }
3864
3865 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3866 page = pte_page(huge_ptep_get(pte));
3867same_page:
3868 if (pages) {
3869 pages[i] = mem_map_offset(page, pfn_offset);
3870 get_page(pages[i]);
3871 }
3872
3873 if (vmas)
3874 vmas[i] = vma;
3875
3876 vaddr += PAGE_SIZE;
3877 ++pfn_offset;
3878 --remainder;
3879 ++i;
3880 if (vaddr < vma->vm_end && remainder &&
3881 pfn_offset < pages_per_huge_page(h)) {
3882 /*
3883 * We use pfn_offset to avoid touching the pageframes
3884 * of this compound page.
3885 */
3886 goto same_page;
3887 }
3888 spin_unlock(ptl);
3889 }
3890 *nr_pages = remainder;
3891 *position = vaddr;
3892
3893 return i ? i : -EFAULT;
3894}
3895
3896unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3897 unsigned long address, unsigned long end, pgprot_t newprot)
3898{
3899 struct mm_struct *mm = vma->vm_mm;
3900 unsigned long start = address;
3901 pte_t *ptep;
3902 pte_t pte;
3903 struct hstate *h = hstate_vma(vma);
3904 unsigned long pages = 0;
3905
3906 BUG_ON(address >= end);
3907 flush_cache_range(vma, address, end);
3908
3909 mmu_notifier_invalidate_range_start(mm, start, end);
3910 i_mmap_lock_write(vma->vm_file->f_mapping);
3911 for (; address < end; address += huge_page_size(h)) {
3912 spinlock_t *ptl;
3913 ptep = huge_pte_offset(mm, address);
3914 if (!ptep)
3915 continue;
3916 ptl = huge_pte_lock(h, mm, ptep);
3917 if (huge_pmd_unshare(mm, &address, ptep)) {
3918 pages++;
3919 spin_unlock(ptl);
3920 continue;
3921 }
3922 pte = huge_ptep_get(ptep);
3923 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3924 spin_unlock(ptl);
3925 continue;
3926 }
3927 if (unlikely(is_hugetlb_entry_migration(pte))) {
3928 swp_entry_t entry = pte_to_swp_entry(pte);
3929
3930 if (is_write_migration_entry(entry)) {
3931 pte_t newpte;
3932
3933 make_migration_entry_read(&entry);
3934 newpte = swp_entry_to_pte(entry);
3935 set_huge_pte_at(mm, address, ptep, newpte);
3936 pages++;
3937 }
3938 spin_unlock(ptl);
3939 continue;
3940 }
3941 if (!huge_pte_none(pte)) {
3942 pte = huge_ptep_get_and_clear(mm, address, ptep);
3943 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3944 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3945 set_huge_pte_at(mm, address, ptep, pte);
3946 pages++;
3947 }
3948 spin_unlock(ptl);
3949 }
3950 /*
3951 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3952 * may have cleared our pud entry and done put_page on the page table:
3953 * once we release i_mmap_rwsem, another task can do the final put_page
3954 * and that page table be reused and filled with junk.
3955 */
3956 flush_tlb_range(vma, start, end);
3957 mmu_notifier_invalidate_range(mm, start, end);
3958 i_mmap_unlock_write(vma->vm_file->f_mapping);
3959 mmu_notifier_invalidate_range_end(mm, start, end);
3960
3961 return pages << h->order;
3962}
3963
3964int hugetlb_reserve_pages(struct inode *inode,
3965 long from, long to,
3966 struct vm_area_struct *vma,
3967 vm_flags_t vm_flags)
3968{
3969 long ret, chg;
3970 struct hstate *h = hstate_inode(inode);
3971 struct hugepage_subpool *spool = subpool_inode(inode);
3972 struct resv_map *resv_map;
3973 long gbl_reserve;
3974
3975 /*
3976 * Only apply hugepage reservation if asked. At fault time, an
3977 * attempt will be made for VM_NORESERVE to allocate a page
3978 * without using reserves
3979 */
3980 if (vm_flags & VM_NORESERVE)
3981 return 0;
3982
3983 /*
3984 * Shared mappings base their reservation on the number of pages that
3985 * are already allocated on behalf of the file. Private mappings need
3986 * to reserve the full area even if read-only as mprotect() may be
3987 * called to make the mapping read-write. Assume !vma is a shm mapping
3988 */
3989 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3990 resv_map = inode_resv_map(inode);
3991
3992 chg = region_chg(resv_map, from, to);
3993
3994 } else {
3995 resv_map = resv_map_alloc();
3996 if (!resv_map)
3997 return -ENOMEM;
3998
3999 chg = to - from;
4000
4001 set_vma_resv_map(vma, resv_map);
4002 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4003 }
4004
4005 if (chg < 0) {
4006 ret = chg;
4007 goto out_err;
4008 }
4009
4010 /*
4011 * There must be enough pages in the subpool for the mapping. If
4012 * the subpool has a minimum size, there may be some global
4013 * reservations already in place (gbl_reserve).
4014 */
4015 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4016 if (gbl_reserve < 0) {
4017 ret = -ENOSPC;
4018 goto out_err;
4019 }
4020
4021 /*
4022 * Check enough hugepages are available for the reservation.
4023 * Hand the pages back to the subpool if there are not
4024 */
4025 ret = hugetlb_acct_memory(h, gbl_reserve);
4026 if (ret < 0) {
4027 /* put back original number of pages, chg */
4028 (void)hugepage_subpool_put_pages(spool, chg);
4029 goto out_err;
4030 }
4031
4032 /*
4033 * Account for the reservations made. Shared mappings record regions
4034 * that have reservations as they are shared by multiple VMAs.
4035 * When the last VMA disappears, the region map says how much
4036 * the reservation was and the page cache tells how much of
4037 * the reservation was consumed. Private mappings are per-VMA and
4038 * only the consumed reservations are tracked. When the VMA
4039 * disappears, the original reservation is the VMA size and the
4040 * consumed reservations are stored in the map. Hence, nothing
4041 * else has to be done for private mappings here
4042 */
4043 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4044 long add = region_add(resv_map, from, to);
4045
4046 if (unlikely(chg > add)) {
4047 /*
4048 * pages in this range were added to the reserve
4049 * map between region_chg and region_add. This
4050 * indicates a race with alloc_huge_page. Adjust
4051 * the subpool and reserve counts modified above
4052 * based on the difference.
4053 */
4054 long rsv_adjust;
4055
4056 rsv_adjust = hugepage_subpool_put_pages(spool,
4057 chg - add);
4058 hugetlb_acct_memory(h, -rsv_adjust);
4059 }
4060 }
4061 return 0;
4062out_err:
4063 if (!vma || vma->vm_flags & VM_MAYSHARE)
4064 region_abort(resv_map, from, to);
4065 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4066 kref_put(&resv_map->refs, resv_map_release);
4067 return ret;
4068}
4069
4070long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4071 long freed)
4072{
4073 struct hstate *h = hstate_inode(inode);
4074 struct resv_map *resv_map = inode_resv_map(inode);
4075 long chg = 0;
4076 struct hugepage_subpool *spool = subpool_inode(inode);
4077 long gbl_reserve;
4078
4079 if (resv_map) {
4080 chg = region_del(resv_map, start, end);
4081 /*
4082 * region_del() can fail in the rare case where a region
4083 * must be split and another region descriptor can not be
4084 * allocated. If end == LONG_MAX, it will not fail.
4085 */
4086 if (chg < 0)
4087 return chg;
4088 }
4089
4090 spin_lock(&inode->i_lock);
4091 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4092 spin_unlock(&inode->i_lock);
4093
4094 /*
4095 * If the subpool has a minimum size, the number of global
4096 * reservations to be released may be adjusted.
4097 */
4098 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4099 hugetlb_acct_memory(h, -gbl_reserve);
4100
4101 return 0;
4102}
4103
4104#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4105static unsigned long page_table_shareable(struct vm_area_struct *svma,
4106 struct vm_area_struct *vma,
4107 unsigned long addr, pgoff_t idx)
4108{
4109 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4110 svma->vm_start;
4111 unsigned long sbase = saddr & PUD_MASK;
4112 unsigned long s_end = sbase + PUD_SIZE;
4113
4114 /* Allow segments to share if only one is marked locked */
4115 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4116 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4117
4118 /*
4119 * match the virtual addresses, permission and the alignment of the
4120 * page table page.
4121 */
4122 if (pmd_index(addr) != pmd_index(saddr) ||
4123 vm_flags != svm_flags ||
4124 sbase < svma->vm_start || svma->vm_end < s_end)
4125 return 0;
4126
4127 return saddr;
4128}
4129
4130static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4131{
4132 unsigned long base = addr & PUD_MASK;
4133 unsigned long end = base + PUD_SIZE;
4134
4135 /*
4136 * check on proper vm_flags and page table alignment
4137 */
4138 if (vma->vm_flags & VM_MAYSHARE &&
4139 vma->vm_start <= base && end <= vma->vm_end)
4140 return true;
4141 return false;
4142}
4143
4144/*
4145 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4146 * and returns the corresponding pte. While this is not necessary for the
4147 * !shared pmd case because we can allocate the pmd later as well, it makes the
4148 * code much cleaner. pmd allocation is essential for the shared case because
4149 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4150 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4151 * bad pmd for sharing.
4152 */
4153pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4154{
4155 struct vm_area_struct *vma = find_vma(mm, addr);
4156 struct address_space *mapping = vma->vm_file->f_mapping;
4157 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4158 vma->vm_pgoff;
4159 struct vm_area_struct *svma;
4160 unsigned long saddr;
4161 pte_t *spte = NULL;
4162 pte_t *pte;
4163 spinlock_t *ptl;
4164
4165 if (!vma_shareable(vma, addr))
4166 return (pte_t *)pmd_alloc(mm, pud, addr);
4167
4168 i_mmap_lock_write(mapping);
4169 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4170 if (svma == vma)
4171 continue;
4172
4173 saddr = page_table_shareable(svma, vma, addr, idx);
4174 if (saddr) {
4175 spte = huge_pte_offset(svma->vm_mm, saddr);
4176 if (spte) {
4177 mm_inc_nr_pmds(mm);
4178 get_page(virt_to_page(spte));
4179 break;
4180 }
4181 }
4182 }
4183
4184 if (!spte)
4185 goto out;
4186
4187 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4188 spin_lock(ptl);
4189 if (pud_none(*pud)) {
4190 pud_populate(mm, pud,
4191 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4192 } else {
4193 put_page(virt_to_page(spte));
4194 mm_inc_nr_pmds(mm);
4195 }
4196 spin_unlock(ptl);
4197out:
4198 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4199 i_mmap_unlock_write(mapping);
4200 return pte;
4201}
4202
4203/*
4204 * unmap huge page backed by shared pte.
4205 *
4206 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4207 * indicated by page_count > 1, unmap is achieved by clearing pud and
4208 * decrementing the ref count. If count == 1, the pte page is not shared.
4209 *
4210 * called with page table lock held.
4211 *
4212 * returns: 1 successfully unmapped a shared pte page
4213 * 0 the underlying pte page is not shared, or it is the last user
4214 */
4215int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4216{
4217 pgd_t *pgd = pgd_offset(mm, *addr);
4218 pud_t *pud = pud_offset(pgd, *addr);
4219
4220 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4221 if (page_count(virt_to_page(ptep)) == 1)
4222 return 0;
4223
4224 pud_clear(pud);
4225 put_page(virt_to_page(ptep));
4226 mm_dec_nr_pmds(mm);
4227 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4228 return 1;
4229}
4230#define want_pmd_share() (1)
4231#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4232pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4233{
4234 return NULL;
4235}
4236
4237int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4238{
4239 return 0;
4240}
4241#define want_pmd_share() (0)
4242#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4243
4244#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4245pte_t *huge_pte_alloc(struct mm_struct *mm,
4246 unsigned long addr, unsigned long sz)
4247{
4248 pgd_t *pgd;
4249 pud_t *pud;
4250 pte_t *pte = NULL;
4251
4252 pgd = pgd_offset(mm, addr);
4253 pud = pud_alloc(mm, pgd, addr);
4254 if (pud) {
4255 if (sz == PUD_SIZE) {
4256 pte = (pte_t *)pud;
4257 } else {
4258 BUG_ON(sz != PMD_SIZE);
4259 if (want_pmd_share() && pud_none(*pud))
4260 pte = huge_pmd_share(mm, addr, pud);
4261 else
4262 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4263 }
4264 }
4265 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4266
4267 return pte;
4268}
4269
4270pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4271{
4272 pgd_t *pgd;
4273 pud_t *pud;
4274 pmd_t *pmd = NULL;
4275
4276 pgd = pgd_offset(mm, addr);
4277 if (pgd_present(*pgd)) {
4278 pud = pud_offset(pgd, addr);
4279 if (pud_present(*pud)) {
4280 if (pud_huge(*pud))
4281 return (pte_t *)pud;
4282 pmd = pmd_offset(pud, addr);
4283 }
4284 }
4285 return (pte_t *) pmd;
4286}
4287
4288#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4289
4290/*
4291 * These functions are overwritable if your architecture needs its own
4292 * behavior.
4293 */
4294struct page * __weak
4295follow_huge_addr(struct mm_struct *mm, unsigned long address,
4296 int write)
4297{
4298 return ERR_PTR(-EINVAL);
4299}
4300
4301struct page * __weak
4302follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4303 pmd_t *pmd, int flags)
4304{
4305 struct page *page = NULL;
4306 spinlock_t *ptl;
4307retry:
4308 ptl = pmd_lockptr(mm, pmd);
4309 spin_lock(ptl);
4310 /*
4311 * make sure that the address range covered by this pmd is not
4312 * unmapped from other threads.
4313 */
4314 if (!pmd_huge(*pmd))
4315 goto out;
4316 if (pmd_present(*pmd)) {
4317 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4318 if (flags & FOLL_GET)
4319 get_page(page);
4320 } else {
4321 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4322 spin_unlock(ptl);
4323 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4324 goto retry;
4325 }
4326 /*
4327 * hwpoisoned entry is treated as no_page_table in
4328 * follow_page_mask().
4329 */
4330 }
4331out:
4332 spin_unlock(ptl);
4333 return page;
4334}
4335
4336struct page * __weak
4337follow_huge_pud(struct mm_struct *mm, unsigned long address,
4338 pud_t *pud, int flags)
4339{
4340 if (flags & FOLL_GET)
4341 return NULL;
4342
4343 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4344}
4345
4346#ifdef CONFIG_MEMORY_FAILURE
4347
4348/*
4349 * This function is called from memory failure code.
4350 * Assume the caller holds page lock of the head page.
4351 */
4352int dequeue_hwpoisoned_huge_page(struct page *hpage)
4353{
4354 struct hstate *h = page_hstate(hpage);
4355 int nid = page_to_nid(hpage);
4356 int ret = -EBUSY;
4357
4358 spin_lock(&hugetlb_lock);
4359 /*
4360 * Just checking !page_huge_active is not enough, because that could be
4361 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4362 */
4363 if (!page_huge_active(hpage) && !page_count(hpage)) {
4364 /*
4365 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4366 * but dangling hpage->lru can trigger list-debug warnings
4367 * (this happens when we call unpoison_memory() on it),
4368 * so let it point to itself with list_del_init().
4369 */
4370 list_del_init(&hpage->lru);
4371 set_page_refcounted(hpage);
4372 h->free_huge_pages--;
4373 h->free_huge_pages_node[nid]--;
4374 ret = 0;
4375 }
4376 spin_unlock(&hugetlb_lock);
4377 return ret;
4378}
4379#endif
4380
4381bool isolate_huge_page(struct page *page, struct list_head *list)
4382{
4383 bool ret = true;
4384
4385 VM_BUG_ON_PAGE(!PageHead(page), page);
4386 spin_lock(&hugetlb_lock);
4387 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4388 ret = false;
4389 goto unlock;
4390 }
4391 clear_page_huge_active(page);
4392 list_move_tail(&page->lru, list);
4393unlock:
4394 spin_unlock(&hugetlb_lock);
4395 return ret;
4396}
4397
4398void putback_active_hugepage(struct page *page)
4399{
4400 VM_BUG_ON_PAGE(!PageHead(page), page);
4401 spin_lock(&hugetlb_lock);
4402 set_page_huge_active(page);
4403 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4404 spin_unlock(&hugetlb_lock);
4405 put_page(page);
4406}