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/module.h>
8#include <linux/mm.h>
9#include <linux/seq_file.h>
10#include <linux/sysctl.h>
11#include <linux/highmem.h>
12#include <linux/mmu_notifier.h>
13#include <linux/nodemask.h>
14#include <linux/pagemap.h>
15#include <linux/mempolicy.h>
16#include <linux/compiler.h>
17#include <linux/cpuset.h>
18#include <linux/mutex.h>
19#include <linux/bootmem.h>
20#include <linux/sysfs.h>
21#include <linux/slab.h>
22#include <linux/rmap.h>
23#include <linux/swap.h>
24#include <linux/swapops.h>
25#include <linux/page-isolation.h>
26#include <linux/jhash.h>
27
28#include <asm/page.h>
29#include <asm/pgtable.h>
30#include <asm/tlb.h>
31
32#include <linux/io.h>
33#include <linux/hugetlb.h>
34#include <linux/hugetlb_cgroup.h>
35#include <linux/node.h>
36#include "internal.h"
37
38const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
39unsigned long hugepages_treat_as_movable;
40
41int hugetlb_max_hstate __read_mostly;
42unsigned int default_hstate_idx;
43struct hstate hstates[HUGE_MAX_HSTATE];
44
45__initdata LIST_HEAD(huge_boot_pages);
46
47/* for command line parsing */
48static struct hstate * __initdata parsed_hstate;
49static unsigned long __initdata default_hstate_max_huge_pages;
50static unsigned long __initdata default_hstate_size;
51
52/*
53 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
54 * free_huge_pages, and surplus_huge_pages.
55 */
56DEFINE_SPINLOCK(hugetlb_lock);
57
58/*
59 * Serializes faults on the same logical page. This is used to
60 * prevent spurious OOMs when the hugepage pool is fully utilized.
61 */
62static int num_fault_mutexes;
63static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp;
64
65static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
66{
67 bool free = (spool->count == 0) && (spool->used_hpages == 0);
68
69 spin_unlock(&spool->lock);
70
71 /* If no pages are used, and no other handles to the subpool
72 * remain, free the subpool the subpool remain */
73 if (free)
74 kfree(spool);
75}
76
77struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
78{
79 struct hugepage_subpool *spool;
80
81 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
82 if (!spool)
83 return NULL;
84
85 spin_lock_init(&spool->lock);
86 spool->count = 1;
87 spool->max_hpages = nr_blocks;
88 spool->used_hpages = 0;
89
90 return spool;
91}
92
93void hugepage_put_subpool(struct hugepage_subpool *spool)
94{
95 spin_lock(&spool->lock);
96 BUG_ON(!spool->count);
97 spool->count--;
98 unlock_or_release_subpool(spool);
99}
100
101static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
102 long delta)
103{
104 int ret = 0;
105
106 if (!spool)
107 return 0;
108
109 spin_lock(&spool->lock);
110 if ((spool->used_hpages + delta) <= spool->max_hpages) {
111 spool->used_hpages += delta;
112 } else {
113 ret = -ENOMEM;
114 }
115 spin_unlock(&spool->lock);
116
117 return ret;
118}
119
120static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
121 long delta)
122{
123 if (!spool)
124 return;
125
126 spin_lock(&spool->lock);
127 spool->used_hpages -= delta;
128 /* If hugetlbfs_put_super couldn't free spool due to
129 * an outstanding quota reference, free it now. */
130 unlock_or_release_subpool(spool);
131}
132
133static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
134{
135 return HUGETLBFS_SB(inode->i_sb)->spool;
136}
137
138static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
139{
140 return subpool_inode(file_inode(vma->vm_file));
141}
142
143/*
144 * Region tracking -- allows tracking of reservations and instantiated pages
145 * across the pages in a mapping.
146 *
147 * The region data structures are embedded into a resv_map and
148 * protected by a resv_map's lock
149 */
150struct file_region {
151 struct list_head link;
152 long from;
153 long to;
154};
155
156static long region_add(struct resv_map *resv, long f, long t)
157{
158 struct list_head *head = &resv->regions;
159 struct file_region *rg, *nrg, *trg;
160
161 spin_lock(&resv->lock);
162 /* Locate the region we are either in or before. */
163 list_for_each_entry(rg, head, link)
164 if (f <= rg->to)
165 break;
166
167 /* Round our left edge to the current segment if it encloses us. */
168 if (f > rg->from)
169 f = rg->from;
170
171 /* Check for and consume any regions we now overlap with. */
172 nrg = rg;
173 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
174 if (&rg->link == head)
175 break;
176 if (rg->from > t)
177 break;
178
179 /* If this area reaches higher then extend our area to
180 * include it completely. If this is not the first area
181 * which we intend to reuse, free it. */
182 if (rg->to > t)
183 t = rg->to;
184 if (rg != nrg) {
185 list_del(&rg->link);
186 kfree(rg);
187 }
188 }
189 nrg->from = f;
190 nrg->to = t;
191 spin_unlock(&resv->lock);
192 return 0;
193}
194
195static long region_chg(struct resv_map *resv, long f, long t)
196{
197 struct list_head *head = &resv->regions;
198 struct file_region *rg, *nrg = NULL;
199 long chg = 0;
200
201retry:
202 spin_lock(&resv->lock);
203 /* Locate the region we are before or in. */
204 list_for_each_entry(rg, head, link)
205 if (f <= rg->to)
206 break;
207
208 /* If we are below the current region then a new region is required.
209 * Subtle, allocate a new region at the position but make it zero
210 * size such that we can guarantee to record the reservation. */
211 if (&rg->link == head || t < rg->from) {
212 if (!nrg) {
213 spin_unlock(&resv->lock);
214 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
215 if (!nrg)
216 return -ENOMEM;
217
218 nrg->from = f;
219 nrg->to = f;
220 INIT_LIST_HEAD(&nrg->link);
221 goto retry;
222 }
223
224 list_add(&nrg->link, rg->link.prev);
225 chg = t - f;
226 goto out_nrg;
227 }
228
229 /* Round our left edge to the current segment if it encloses us. */
230 if (f > rg->from)
231 f = rg->from;
232 chg = t - f;
233
234 /* Check for and consume any regions we now overlap with. */
235 list_for_each_entry(rg, rg->link.prev, link) {
236 if (&rg->link == head)
237 break;
238 if (rg->from > t)
239 goto out;
240
241 /* We overlap with this area, if it extends further than
242 * us then we must extend ourselves. Account for its
243 * existing reservation. */
244 if (rg->to > t) {
245 chg += rg->to - t;
246 t = rg->to;
247 }
248 chg -= rg->to - rg->from;
249 }
250
251out:
252 spin_unlock(&resv->lock);
253 /* We already know we raced and no longer need the new region */
254 kfree(nrg);
255 return chg;
256out_nrg:
257 spin_unlock(&resv->lock);
258 return chg;
259}
260
261static long region_truncate(struct resv_map *resv, long end)
262{
263 struct list_head *head = &resv->regions;
264 struct file_region *rg, *trg;
265 long chg = 0;
266
267 spin_lock(&resv->lock);
268 /* Locate the region we are either in or before. */
269 list_for_each_entry(rg, head, link)
270 if (end <= rg->to)
271 break;
272 if (&rg->link == head)
273 goto out;
274
275 /* If we are in the middle of a region then adjust it. */
276 if (end > rg->from) {
277 chg = rg->to - end;
278 rg->to = end;
279 rg = list_entry(rg->link.next, typeof(*rg), link);
280 }
281
282 /* Drop any remaining regions. */
283 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
284 if (&rg->link == head)
285 break;
286 chg += rg->to - rg->from;
287 list_del(&rg->link);
288 kfree(rg);
289 }
290
291out:
292 spin_unlock(&resv->lock);
293 return chg;
294}
295
296static long region_count(struct resv_map *resv, long f, long t)
297{
298 struct list_head *head = &resv->regions;
299 struct file_region *rg;
300 long chg = 0;
301
302 spin_lock(&resv->lock);
303 /* Locate each segment we overlap with, and count that overlap. */
304 list_for_each_entry(rg, head, link) {
305 long seg_from;
306 long seg_to;
307
308 if (rg->to <= f)
309 continue;
310 if (rg->from >= t)
311 break;
312
313 seg_from = max(rg->from, f);
314 seg_to = min(rg->to, t);
315
316 chg += seg_to - seg_from;
317 }
318 spin_unlock(&resv->lock);
319
320 return chg;
321}
322
323/*
324 * Convert the address within this vma to the page offset within
325 * the mapping, in pagecache page units; huge pages here.
326 */
327static pgoff_t vma_hugecache_offset(struct hstate *h,
328 struct vm_area_struct *vma, unsigned long address)
329{
330 return ((address - vma->vm_start) >> huge_page_shift(h)) +
331 (vma->vm_pgoff >> huge_page_order(h));
332}
333
334pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
335 unsigned long address)
336{
337 return vma_hugecache_offset(hstate_vma(vma), vma, address);
338}
339
340/*
341 * Return the size of the pages allocated when backing a VMA. In the majority
342 * cases this will be same size as used by the page table entries.
343 */
344unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
345{
346 struct hstate *hstate;
347
348 if (!is_vm_hugetlb_page(vma))
349 return PAGE_SIZE;
350
351 hstate = hstate_vma(vma);
352
353 return 1UL << huge_page_shift(hstate);
354}
355EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
356
357/*
358 * Return the page size being used by the MMU to back a VMA. In the majority
359 * of cases, the page size used by the kernel matches the MMU size. On
360 * architectures where it differs, an architecture-specific version of this
361 * function is required.
362 */
363#ifndef vma_mmu_pagesize
364unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
365{
366 return vma_kernel_pagesize(vma);
367}
368#endif
369
370/*
371 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
372 * bits of the reservation map pointer, which are always clear due to
373 * alignment.
374 */
375#define HPAGE_RESV_OWNER (1UL << 0)
376#define HPAGE_RESV_UNMAPPED (1UL << 1)
377#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
378
379/*
380 * These helpers are used to track how many pages are reserved for
381 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
382 * is guaranteed to have their future faults succeed.
383 *
384 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
385 * the reserve counters are updated with the hugetlb_lock held. It is safe
386 * to reset the VMA at fork() time as it is not in use yet and there is no
387 * chance of the global counters getting corrupted as a result of the values.
388 *
389 * The private mapping reservation is represented in a subtly different
390 * manner to a shared mapping. A shared mapping has a region map associated
391 * with the underlying file, this region map represents the backing file
392 * pages which have ever had a reservation assigned which this persists even
393 * after the page is instantiated. A private mapping has a region map
394 * associated with the original mmap which is attached to all VMAs which
395 * reference it, this region map represents those offsets which have consumed
396 * reservation ie. where pages have been instantiated.
397 */
398static unsigned long get_vma_private_data(struct vm_area_struct *vma)
399{
400 return (unsigned long)vma->vm_private_data;
401}
402
403static void set_vma_private_data(struct vm_area_struct *vma,
404 unsigned long value)
405{
406 vma->vm_private_data = (void *)value;
407}
408
409struct resv_map *resv_map_alloc(void)
410{
411 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
412 if (!resv_map)
413 return NULL;
414
415 kref_init(&resv_map->refs);
416 spin_lock_init(&resv_map->lock);
417 INIT_LIST_HEAD(&resv_map->regions);
418
419 return resv_map;
420}
421
422void resv_map_release(struct kref *ref)
423{
424 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
425
426 /* Clear out any active regions before we release the map. */
427 region_truncate(resv_map, 0);
428 kfree(resv_map);
429}
430
431static inline struct resv_map *inode_resv_map(struct inode *inode)
432{
433 return inode->i_mapping->private_data;
434}
435
436static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
437{
438 VM_BUG_ON(!is_vm_hugetlb_page(vma));
439 if (vma->vm_flags & VM_MAYSHARE) {
440 struct address_space *mapping = vma->vm_file->f_mapping;
441 struct inode *inode = mapping->host;
442
443 return inode_resv_map(inode);
444
445 } else {
446 return (struct resv_map *)(get_vma_private_data(vma) &
447 ~HPAGE_RESV_MASK);
448 }
449}
450
451static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
452{
453 VM_BUG_ON(!is_vm_hugetlb_page(vma));
454 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
455
456 set_vma_private_data(vma, (get_vma_private_data(vma) &
457 HPAGE_RESV_MASK) | (unsigned long)map);
458}
459
460static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
461{
462 VM_BUG_ON(!is_vm_hugetlb_page(vma));
463 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
464
465 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
466}
467
468static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
469{
470 VM_BUG_ON(!is_vm_hugetlb_page(vma));
471
472 return (get_vma_private_data(vma) & flag) != 0;
473}
474
475/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
476void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
477{
478 VM_BUG_ON(!is_vm_hugetlb_page(vma));
479 if (!(vma->vm_flags & VM_MAYSHARE))
480 vma->vm_private_data = (void *)0;
481}
482
483/* Returns true if the VMA has associated reserve pages */
484static int vma_has_reserves(struct vm_area_struct *vma, long chg)
485{
486 if (vma->vm_flags & VM_NORESERVE) {
487 /*
488 * This address is already reserved by other process(chg == 0),
489 * so, we should decrement reserved count. Without decrementing,
490 * reserve count remains after releasing inode, because this
491 * allocated page will go into page cache and is regarded as
492 * coming from reserved pool in releasing step. Currently, we
493 * don't have any other solution to deal with this situation
494 * properly, so add work-around here.
495 */
496 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
497 return 1;
498 else
499 return 0;
500 }
501
502 /* Shared mappings always use reserves */
503 if (vma->vm_flags & VM_MAYSHARE)
504 return 1;
505
506 /*
507 * Only the process that called mmap() has reserves for
508 * private mappings.
509 */
510 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
511 return 1;
512
513 return 0;
514}
515
516static void enqueue_huge_page(struct hstate *h, struct page *page)
517{
518 int nid = page_to_nid(page);
519 list_move(&page->lru, &h->hugepage_freelists[nid]);
520 h->free_huge_pages++;
521 h->free_huge_pages_node[nid]++;
522}
523
524static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
525{
526 struct page *page;
527
528 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
529 if (!is_migrate_isolate_page(page))
530 break;
531 /*
532 * if 'non-isolated free hugepage' not found on the list,
533 * the allocation fails.
534 */
535 if (&h->hugepage_freelists[nid] == &page->lru)
536 return NULL;
537 list_move(&page->lru, &h->hugepage_activelist);
538 set_page_refcounted(page);
539 h->free_huge_pages--;
540 h->free_huge_pages_node[nid]--;
541 return page;
542}
543
544/* Movability of hugepages depends on migration support. */
545static inline gfp_t htlb_alloc_mask(struct hstate *h)
546{
547 if (hugepages_treat_as_movable || hugepage_migration_support(h))
548 return GFP_HIGHUSER_MOVABLE;
549 else
550 return GFP_HIGHUSER;
551}
552
553static struct page *dequeue_huge_page_vma(struct hstate *h,
554 struct vm_area_struct *vma,
555 unsigned long address, int avoid_reserve,
556 long chg)
557{
558 struct page *page = NULL;
559 struct mempolicy *mpol;
560 nodemask_t *nodemask;
561 struct zonelist *zonelist;
562 struct zone *zone;
563 struct zoneref *z;
564 unsigned int cpuset_mems_cookie;
565
566 /*
567 * A child process with MAP_PRIVATE mappings created by their parent
568 * have no page reserves. This check ensures that reservations are
569 * not "stolen". The child may still get SIGKILLed
570 */
571 if (!vma_has_reserves(vma, chg) &&
572 h->free_huge_pages - h->resv_huge_pages == 0)
573 goto err;
574
575 /* If reserves cannot be used, ensure enough pages are in the pool */
576 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
577 goto err;
578
579retry_cpuset:
580 cpuset_mems_cookie = read_mems_allowed_begin();
581 zonelist = huge_zonelist(vma, address,
582 htlb_alloc_mask(h), &mpol, &nodemask);
583
584 for_each_zone_zonelist_nodemask(zone, z, zonelist,
585 MAX_NR_ZONES - 1, nodemask) {
586 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) {
587 page = dequeue_huge_page_node(h, zone_to_nid(zone));
588 if (page) {
589 if (avoid_reserve)
590 break;
591 if (!vma_has_reserves(vma, chg))
592 break;
593
594 SetPagePrivate(page);
595 h->resv_huge_pages--;
596 break;
597 }
598 }
599 }
600
601 mpol_cond_put(mpol);
602 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
603 goto retry_cpuset;
604 return page;
605
606err:
607 return NULL;
608}
609
610static void update_and_free_page(struct hstate *h, struct page *page)
611{
612 int i;
613
614 VM_BUG_ON(h->order >= MAX_ORDER);
615
616 h->nr_huge_pages--;
617 h->nr_huge_pages_node[page_to_nid(page)]--;
618 for (i = 0; i < pages_per_huge_page(h); i++) {
619 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
620 1 << PG_referenced | 1 << PG_dirty |
621 1 << PG_active | 1 << PG_reserved |
622 1 << PG_private | 1 << PG_writeback);
623 }
624 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
625 set_compound_page_dtor(page, NULL);
626 set_page_refcounted(page);
627 arch_release_hugepage(page);
628 __free_pages(page, huge_page_order(h));
629}
630
631struct hstate *size_to_hstate(unsigned long size)
632{
633 struct hstate *h;
634
635 for_each_hstate(h) {
636 if (huge_page_size(h) == size)
637 return h;
638 }
639 return NULL;
640}
641
642static void free_huge_page(struct page *page)
643{
644 /*
645 * Can't pass hstate in here because it is called from the
646 * compound page destructor.
647 */
648 struct hstate *h = page_hstate(page);
649 int nid = page_to_nid(page);
650 struct hugepage_subpool *spool =
651 (struct hugepage_subpool *)page_private(page);
652 bool restore_reserve;
653
654 set_page_private(page, 0);
655 page->mapping = NULL;
656 BUG_ON(page_count(page));
657 BUG_ON(page_mapcount(page));
658 restore_reserve = PagePrivate(page);
659 ClearPagePrivate(page);
660
661 spin_lock(&hugetlb_lock);
662 hugetlb_cgroup_uncharge_page(hstate_index(h),
663 pages_per_huge_page(h), page);
664 if (restore_reserve)
665 h->resv_huge_pages++;
666
667 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
668 /* remove the page from active list */
669 list_del(&page->lru);
670 update_and_free_page(h, page);
671 h->surplus_huge_pages--;
672 h->surplus_huge_pages_node[nid]--;
673 } else {
674 arch_clear_hugepage_flags(page);
675 enqueue_huge_page(h, page);
676 }
677 spin_unlock(&hugetlb_lock);
678 hugepage_subpool_put_pages(spool, 1);
679}
680
681static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
682{
683 INIT_LIST_HEAD(&page->lru);
684 set_compound_page_dtor(page, free_huge_page);
685 spin_lock(&hugetlb_lock);
686 set_hugetlb_cgroup(page, NULL);
687 h->nr_huge_pages++;
688 h->nr_huge_pages_node[nid]++;
689 spin_unlock(&hugetlb_lock);
690 put_page(page); /* free it into the hugepage allocator */
691}
692
693static void __init prep_compound_gigantic_page(struct page *page,
694 unsigned long order)
695{
696 int i;
697 int nr_pages = 1 << order;
698 struct page *p = page + 1;
699
700 /* we rely on prep_new_huge_page to set the destructor */
701 set_compound_order(page, order);
702 __SetPageHead(page);
703 __ClearPageReserved(page);
704 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
705 __SetPageTail(p);
706 /*
707 * For gigantic hugepages allocated through bootmem at
708 * boot, it's safer to be consistent with the not-gigantic
709 * hugepages and clear the PG_reserved bit from all tail pages
710 * too. Otherwse drivers using get_user_pages() to access tail
711 * pages may get the reference counting wrong if they see
712 * PG_reserved set on a tail page (despite the head page not
713 * having PG_reserved set). Enforcing this consistency between
714 * head and tail pages allows drivers to optimize away a check
715 * on the head page when they need know if put_page() is needed
716 * after get_user_pages().
717 */
718 __ClearPageReserved(p);
719 set_page_count(p, 0);
720 p->first_page = page;
721 }
722}
723
724/*
725 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
726 * transparent huge pages. See the PageTransHuge() documentation for more
727 * details.
728 */
729int PageHuge(struct page *page)
730{
731 if (!PageCompound(page))
732 return 0;
733
734 page = compound_head(page);
735 return get_compound_page_dtor(page) == free_huge_page;
736}
737EXPORT_SYMBOL_GPL(PageHuge);
738
739/*
740 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
741 * normal or transparent huge pages.
742 */
743int PageHeadHuge(struct page *page_head)
744{
745 if (!PageHead(page_head))
746 return 0;
747
748 return get_compound_page_dtor(page_head) == free_huge_page;
749}
750
751pgoff_t __basepage_index(struct page *page)
752{
753 struct page *page_head = compound_head(page);
754 pgoff_t index = page_index(page_head);
755 unsigned long compound_idx;
756
757 if (!PageHuge(page_head))
758 return page_index(page);
759
760 if (compound_order(page_head) >= MAX_ORDER)
761 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
762 else
763 compound_idx = page - page_head;
764
765 return (index << compound_order(page_head)) + compound_idx;
766}
767
768static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
769{
770 struct page *page;
771
772 if (h->order >= MAX_ORDER)
773 return NULL;
774
775 page = alloc_pages_exact_node(nid,
776 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
777 __GFP_REPEAT|__GFP_NOWARN,
778 huge_page_order(h));
779 if (page) {
780 if (arch_prepare_hugepage(page)) {
781 __free_pages(page, huge_page_order(h));
782 return NULL;
783 }
784 prep_new_huge_page(h, page, nid);
785 }
786
787 return page;
788}
789
790/*
791 * common helper functions for hstate_next_node_to_{alloc|free}.
792 * We may have allocated or freed a huge page based on a different
793 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
794 * be outside of *nodes_allowed. Ensure that we use an allowed
795 * node for alloc or free.
796 */
797static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
798{
799 nid = next_node(nid, *nodes_allowed);
800 if (nid == MAX_NUMNODES)
801 nid = first_node(*nodes_allowed);
802 VM_BUG_ON(nid >= MAX_NUMNODES);
803
804 return nid;
805}
806
807static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
808{
809 if (!node_isset(nid, *nodes_allowed))
810 nid = next_node_allowed(nid, nodes_allowed);
811 return nid;
812}
813
814/*
815 * returns the previously saved node ["this node"] from which to
816 * allocate a persistent huge page for the pool and advance the
817 * next node from which to allocate, handling wrap at end of node
818 * mask.
819 */
820static int hstate_next_node_to_alloc(struct hstate *h,
821 nodemask_t *nodes_allowed)
822{
823 int nid;
824
825 VM_BUG_ON(!nodes_allowed);
826
827 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
828 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
829
830 return nid;
831}
832
833/*
834 * helper for free_pool_huge_page() - return the previously saved
835 * node ["this node"] from which to free a huge page. Advance the
836 * next node id whether or not we find a free huge page to free so
837 * that the next attempt to free addresses the next node.
838 */
839static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
840{
841 int nid;
842
843 VM_BUG_ON(!nodes_allowed);
844
845 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
846 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
847
848 return nid;
849}
850
851#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
852 for (nr_nodes = nodes_weight(*mask); \
853 nr_nodes > 0 && \
854 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
855 nr_nodes--)
856
857#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
858 for (nr_nodes = nodes_weight(*mask); \
859 nr_nodes > 0 && \
860 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
861 nr_nodes--)
862
863static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
864{
865 struct page *page;
866 int nr_nodes, node;
867 int ret = 0;
868
869 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
870 page = alloc_fresh_huge_page_node(h, node);
871 if (page) {
872 ret = 1;
873 break;
874 }
875 }
876
877 if (ret)
878 count_vm_event(HTLB_BUDDY_PGALLOC);
879 else
880 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
881
882 return ret;
883}
884
885/*
886 * Free huge page from pool from next node to free.
887 * Attempt to keep persistent huge pages more or less
888 * balanced over allowed nodes.
889 * Called with hugetlb_lock locked.
890 */
891static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
892 bool acct_surplus)
893{
894 int nr_nodes, node;
895 int ret = 0;
896
897 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
898 /*
899 * If we're returning unused surplus pages, only examine
900 * nodes with surplus pages.
901 */
902 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
903 !list_empty(&h->hugepage_freelists[node])) {
904 struct page *page =
905 list_entry(h->hugepage_freelists[node].next,
906 struct page, lru);
907 list_del(&page->lru);
908 h->free_huge_pages--;
909 h->free_huge_pages_node[node]--;
910 if (acct_surplus) {
911 h->surplus_huge_pages--;
912 h->surplus_huge_pages_node[node]--;
913 }
914 update_and_free_page(h, page);
915 ret = 1;
916 break;
917 }
918 }
919
920 return ret;
921}
922
923/*
924 * Dissolve a given free hugepage into free buddy pages. This function does
925 * nothing for in-use (including surplus) hugepages.
926 */
927static void dissolve_free_huge_page(struct page *page)
928{
929 spin_lock(&hugetlb_lock);
930 if (PageHuge(page) && !page_count(page)) {
931 struct hstate *h = page_hstate(page);
932 int nid = page_to_nid(page);
933 list_del(&page->lru);
934 h->free_huge_pages--;
935 h->free_huge_pages_node[nid]--;
936 update_and_free_page(h, page);
937 }
938 spin_unlock(&hugetlb_lock);
939}
940
941/*
942 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
943 * make specified memory blocks removable from the system.
944 * Note that start_pfn should aligned with (minimum) hugepage size.
945 */
946void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
947{
948 unsigned int order = 8 * sizeof(void *);
949 unsigned long pfn;
950 struct hstate *h;
951
952 /* Set scan step to minimum hugepage size */
953 for_each_hstate(h)
954 if (order > huge_page_order(h))
955 order = huge_page_order(h);
956 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
957 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
958 dissolve_free_huge_page(pfn_to_page(pfn));
959}
960
961static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
962{
963 struct page *page;
964 unsigned int r_nid;
965
966 if (h->order >= MAX_ORDER)
967 return NULL;
968
969 /*
970 * Assume we will successfully allocate the surplus page to
971 * prevent racing processes from causing the surplus to exceed
972 * overcommit
973 *
974 * This however introduces a different race, where a process B
975 * tries to grow the static hugepage pool while alloc_pages() is
976 * called by process A. B will only examine the per-node
977 * counters in determining if surplus huge pages can be
978 * converted to normal huge pages in adjust_pool_surplus(). A
979 * won't be able to increment the per-node counter, until the
980 * lock is dropped by B, but B doesn't drop hugetlb_lock until
981 * no more huge pages can be converted from surplus to normal
982 * state (and doesn't try to convert again). Thus, we have a
983 * case where a surplus huge page exists, the pool is grown, and
984 * the surplus huge page still exists after, even though it
985 * should just have been converted to a normal huge page. This
986 * does not leak memory, though, as the hugepage will be freed
987 * once it is out of use. It also does not allow the counters to
988 * go out of whack in adjust_pool_surplus() as we don't modify
989 * the node values until we've gotten the hugepage and only the
990 * per-node value is checked there.
991 */
992 spin_lock(&hugetlb_lock);
993 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
994 spin_unlock(&hugetlb_lock);
995 return NULL;
996 } else {
997 h->nr_huge_pages++;
998 h->surplus_huge_pages++;
999 }
1000 spin_unlock(&hugetlb_lock);
1001
1002 if (nid == NUMA_NO_NODE)
1003 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
1004 __GFP_REPEAT|__GFP_NOWARN,
1005 huge_page_order(h));
1006 else
1007 page = alloc_pages_exact_node(nid,
1008 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1009 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
1010
1011 if (page && arch_prepare_hugepage(page)) {
1012 __free_pages(page, huge_page_order(h));
1013 page = NULL;
1014 }
1015
1016 spin_lock(&hugetlb_lock);
1017 if (page) {
1018 INIT_LIST_HEAD(&page->lru);
1019 r_nid = page_to_nid(page);
1020 set_compound_page_dtor(page, free_huge_page);
1021 set_hugetlb_cgroup(page, NULL);
1022 /*
1023 * We incremented the global counters already
1024 */
1025 h->nr_huge_pages_node[r_nid]++;
1026 h->surplus_huge_pages_node[r_nid]++;
1027 __count_vm_event(HTLB_BUDDY_PGALLOC);
1028 } else {
1029 h->nr_huge_pages--;
1030 h->surplus_huge_pages--;
1031 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1032 }
1033 spin_unlock(&hugetlb_lock);
1034
1035 return page;
1036}
1037
1038/*
1039 * This allocation function is useful in the context where vma is irrelevant.
1040 * E.g. soft-offlining uses this function because it only cares physical
1041 * address of error page.
1042 */
1043struct page *alloc_huge_page_node(struct hstate *h, int nid)
1044{
1045 struct page *page = NULL;
1046
1047 spin_lock(&hugetlb_lock);
1048 if (h->free_huge_pages - h->resv_huge_pages > 0)
1049 page = dequeue_huge_page_node(h, nid);
1050 spin_unlock(&hugetlb_lock);
1051
1052 if (!page)
1053 page = alloc_buddy_huge_page(h, nid);
1054
1055 return page;
1056}
1057
1058/*
1059 * Increase the hugetlb pool such that it can accommodate a reservation
1060 * of size 'delta'.
1061 */
1062static int gather_surplus_pages(struct hstate *h, int delta)
1063{
1064 struct list_head surplus_list;
1065 struct page *page, *tmp;
1066 int ret, i;
1067 int needed, allocated;
1068 bool alloc_ok = true;
1069
1070 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1071 if (needed <= 0) {
1072 h->resv_huge_pages += delta;
1073 return 0;
1074 }
1075
1076 allocated = 0;
1077 INIT_LIST_HEAD(&surplus_list);
1078
1079 ret = -ENOMEM;
1080retry:
1081 spin_unlock(&hugetlb_lock);
1082 for (i = 0; i < needed; i++) {
1083 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1084 if (!page) {
1085 alloc_ok = false;
1086 break;
1087 }
1088 list_add(&page->lru, &surplus_list);
1089 }
1090 allocated += i;
1091
1092 /*
1093 * After retaking hugetlb_lock, we need to recalculate 'needed'
1094 * because either resv_huge_pages or free_huge_pages may have changed.
1095 */
1096 spin_lock(&hugetlb_lock);
1097 needed = (h->resv_huge_pages + delta) -
1098 (h->free_huge_pages + allocated);
1099 if (needed > 0) {
1100 if (alloc_ok)
1101 goto retry;
1102 /*
1103 * We were not able to allocate enough pages to
1104 * satisfy the entire reservation so we free what
1105 * we've allocated so far.
1106 */
1107 goto free;
1108 }
1109 /*
1110 * The surplus_list now contains _at_least_ the number of extra pages
1111 * needed to accommodate the reservation. Add the appropriate number
1112 * of pages to the hugetlb pool and free the extras back to the buddy
1113 * allocator. Commit the entire reservation here to prevent another
1114 * process from stealing the pages as they are added to the pool but
1115 * before they are reserved.
1116 */
1117 needed += allocated;
1118 h->resv_huge_pages += delta;
1119 ret = 0;
1120
1121 /* Free the needed pages to the hugetlb pool */
1122 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1123 if ((--needed) < 0)
1124 break;
1125 /*
1126 * This page is now managed by the hugetlb allocator and has
1127 * no users -- drop the buddy allocator's reference.
1128 */
1129 put_page_testzero(page);
1130 VM_BUG_ON_PAGE(page_count(page), page);
1131 enqueue_huge_page(h, page);
1132 }
1133free:
1134 spin_unlock(&hugetlb_lock);
1135
1136 /* Free unnecessary surplus pages to the buddy allocator */
1137 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1138 put_page(page);
1139 spin_lock(&hugetlb_lock);
1140
1141 return ret;
1142}
1143
1144/*
1145 * When releasing a hugetlb pool reservation, any surplus pages that were
1146 * allocated to satisfy the reservation must be explicitly freed if they were
1147 * never used.
1148 * Called with hugetlb_lock held.
1149 */
1150static void return_unused_surplus_pages(struct hstate *h,
1151 unsigned long unused_resv_pages)
1152{
1153 unsigned long nr_pages;
1154
1155 /* Uncommit the reservation */
1156 h->resv_huge_pages -= unused_resv_pages;
1157
1158 /* Cannot return gigantic pages currently */
1159 if (h->order >= MAX_ORDER)
1160 return;
1161
1162 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1163
1164 /*
1165 * We want to release as many surplus pages as possible, spread
1166 * evenly across all nodes with memory. Iterate across these nodes
1167 * until we can no longer free unreserved surplus pages. This occurs
1168 * when the nodes with surplus pages have no free pages.
1169 * free_pool_huge_page() will balance the the freed pages across the
1170 * on-line nodes with memory and will handle the hstate accounting.
1171 */
1172 while (nr_pages--) {
1173 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1174 break;
1175 cond_resched_lock(&hugetlb_lock);
1176 }
1177}
1178
1179/*
1180 * Determine if the huge page at addr within the vma has an associated
1181 * reservation. Where it does not we will need to logically increase
1182 * reservation and actually increase subpool usage before an allocation
1183 * can occur. Where any new reservation would be required the
1184 * reservation change is prepared, but not committed. Once the page
1185 * has been allocated from the subpool and instantiated the change should
1186 * be committed via vma_commit_reservation. No action is required on
1187 * failure.
1188 */
1189static long vma_needs_reservation(struct hstate *h,
1190 struct vm_area_struct *vma, unsigned long addr)
1191{
1192 struct resv_map *resv;
1193 pgoff_t idx;
1194 long chg;
1195
1196 resv = vma_resv_map(vma);
1197 if (!resv)
1198 return 1;
1199
1200 idx = vma_hugecache_offset(h, vma, addr);
1201 chg = region_chg(resv, idx, idx + 1);
1202
1203 if (vma->vm_flags & VM_MAYSHARE)
1204 return chg;
1205 else
1206 return chg < 0 ? chg : 0;
1207}
1208static void vma_commit_reservation(struct hstate *h,
1209 struct vm_area_struct *vma, unsigned long addr)
1210{
1211 struct resv_map *resv;
1212 pgoff_t idx;
1213
1214 resv = vma_resv_map(vma);
1215 if (!resv)
1216 return;
1217
1218 idx = vma_hugecache_offset(h, vma, addr);
1219 region_add(resv, idx, idx + 1);
1220}
1221
1222static struct page *alloc_huge_page(struct vm_area_struct *vma,
1223 unsigned long addr, int avoid_reserve)
1224{
1225 struct hugepage_subpool *spool = subpool_vma(vma);
1226 struct hstate *h = hstate_vma(vma);
1227 struct page *page;
1228 long chg;
1229 int ret, idx;
1230 struct hugetlb_cgroup *h_cg;
1231
1232 idx = hstate_index(h);
1233 /*
1234 * Processes that did not create the mapping will have no
1235 * reserves and will not have accounted against subpool
1236 * limit. Check that the subpool limit can be made before
1237 * satisfying the allocation MAP_NORESERVE mappings may also
1238 * need pages and subpool limit allocated allocated if no reserve
1239 * mapping overlaps.
1240 */
1241 chg = vma_needs_reservation(h, vma, addr);
1242 if (chg < 0)
1243 return ERR_PTR(-ENOMEM);
1244 if (chg || avoid_reserve)
1245 if (hugepage_subpool_get_pages(spool, 1))
1246 return ERR_PTR(-ENOSPC);
1247
1248 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1249 if (ret) {
1250 if (chg || avoid_reserve)
1251 hugepage_subpool_put_pages(spool, 1);
1252 return ERR_PTR(-ENOSPC);
1253 }
1254 spin_lock(&hugetlb_lock);
1255 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1256 if (!page) {
1257 spin_unlock(&hugetlb_lock);
1258 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1259 if (!page) {
1260 hugetlb_cgroup_uncharge_cgroup(idx,
1261 pages_per_huge_page(h),
1262 h_cg);
1263 if (chg || avoid_reserve)
1264 hugepage_subpool_put_pages(spool, 1);
1265 return ERR_PTR(-ENOSPC);
1266 }
1267 spin_lock(&hugetlb_lock);
1268 list_move(&page->lru, &h->hugepage_activelist);
1269 /* Fall through */
1270 }
1271 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1272 spin_unlock(&hugetlb_lock);
1273
1274 set_page_private(page, (unsigned long)spool);
1275
1276 vma_commit_reservation(h, vma, addr);
1277 return page;
1278}
1279
1280/*
1281 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1282 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1283 * where no ERR_VALUE is expected to be returned.
1284 */
1285struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1286 unsigned long addr, int avoid_reserve)
1287{
1288 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1289 if (IS_ERR(page))
1290 page = NULL;
1291 return page;
1292}
1293
1294int __weak alloc_bootmem_huge_page(struct hstate *h)
1295{
1296 struct huge_bootmem_page *m;
1297 int nr_nodes, node;
1298
1299 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1300 void *addr;
1301
1302 addr = memblock_virt_alloc_try_nid_nopanic(
1303 huge_page_size(h), huge_page_size(h),
1304 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1305 if (addr) {
1306 /*
1307 * Use the beginning of the huge page to store the
1308 * huge_bootmem_page struct (until gather_bootmem
1309 * puts them into the mem_map).
1310 */
1311 m = addr;
1312 goto found;
1313 }
1314 }
1315 return 0;
1316
1317found:
1318 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1319 /* Put them into a private list first because mem_map is not up yet */
1320 list_add(&m->list, &huge_boot_pages);
1321 m->hstate = h;
1322 return 1;
1323}
1324
1325static void __init prep_compound_huge_page(struct page *page, int order)
1326{
1327 if (unlikely(order > (MAX_ORDER - 1)))
1328 prep_compound_gigantic_page(page, order);
1329 else
1330 prep_compound_page(page, order);
1331}
1332
1333/* Put bootmem huge pages into the standard lists after mem_map is up */
1334static void __init gather_bootmem_prealloc(void)
1335{
1336 struct huge_bootmem_page *m;
1337
1338 list_for_each_entry(m, &huge_boot_pages, list) {
1339 struct hstate *h = m->hstate;
1340 struct page *page;
1341
1342#ifdef CONFIG_HIGHMEM
1343 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1344 memblock_free_late(__pa(m),
1345 sizeof(struct huge_bootmem_page));
1346#else
1347 page = virt_to_page(m);
1348#endif
1349 WARN_ON(page_count(page) != 1);
1350 prep_compound_huge_page(page, h->order);
1351 WARN_ON(PageReserved(page));
1352 prep_new_huge_page(h, page, page_to_nid(page));
1353 /*
1354 * If we had gigantic hugepages allocated at boot time, we need
1355 * to restore the 'stolen' pages to totalram_pages in order to
1356 * fix confusing memory reports from free(1) and another
1357 * side-effects, like CommitLimit going negative.
1358 */
1359 if (h->order > (MAX_ORDER - 1))
1360 adjust_managed_page_count(page, 1 << h->order);
1361 }
1362}
1363
1364static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1365{
1366 unsigned long i;
1367
1368 for (i = 0; i < h->max_huge_pages; ++i) {
1369 if (h->order >= MAX_ORDER) {
1370 if (!alloc_bootmem_huge_page(h))
1371 break;
1372 } else if (!alloc_fresh_huge_page(h,
1373 &node_states[N_MEMORY]))
1374 break;
1375 }
1376 h->max_huge_pages = i;
1377}
1378
1379static void __init hugetlb_init_hstates(void)
1380{
1381 struct hstate *h;
1382
1383 for_each_hstate(h) {
1384 /* oversize hugepages were init'ed in early boot */
1385 if (h->order < MAX_ORDER)
1386 hugetlb_hstate_alloc_pages(h);
1387 }
1388}
1389
1390static char * __init memfmt(char *buf, unsigned long n)
1391{
1392 if (n >= (1UL << 30))
1393 sprintf(buf, "%lu GB", n >> 30);
1394 else if (n >= (1UL << 20))
1395 sprintf(buf, "%lu MB", n >> 20);
1396 else
1397 sprintf(buf, "%lu KB", n >> 10);
1398 return buf;
1399}
1400
1401static void __init report_hugepages(void)
1402{
1403 struct hstate *h;
1404
1405 for_each_hstate(h) {
1406 char buf[32];
1407 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1408 memfmt(buf, huge_page_size(h)),
1409 h->free_huge_pages);
1410 }
1411}
1412
1413#ifdef CONFIG_HIGHMEM
1414static void try_to_free_low(struct hstate *h, unsigned long count,
1415 nodemask_t *nodes_allowed)
1416{
1417 int i;
1418
1419 if (h->order >= MAX_ORDER)
1420 return;
1421
1422 for_each_node_mask(i, *nodes_allowed) {
1423 struct page *page, *next;
1424 struct list_head *freel = &h->hugepage_freelists[i];
1425 list_for_each_entry_safe(page, next, freel, lru) {
1426 if (count >= h->nr_huge_pages)
1427 return;
1428 if (PageHighMem(page))
1429 continue;
1430 list_del(&page->lru);
1431 update_and_free_page(h, page);
1432 h->free_huge_pages--;
1433 h->free_huge_pages_node[page_to_nid(page)]--;
1434 }
1435 }
1436}
1437#else
1438static inline void try_to_free_low(struct hstate *h, unsigned long count,
1439 nodemask_t *nodes_allowed)
1440{
1441}
1442#endif
1443
1444/*
1445 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1446 * balanced by operating on them in a round-robin fashion.
1447 * Returns 1 if an adjustment was made.
1448 */
1449static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1450 int delta)
1451{
1452 int nr_nodes, node;
1453
1454 VM_BUG_ON(delta != -1 && delta != 1);
1455
1456 if (delta < 0) {
1457 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1458 if (h->surplus_huge_pages_node[node])
1459 goto found;
1460 }
1461 } else {
1462 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1463 if (h->surplus_huge_pages_node[node] <
1464 h->nr_huge_pages_node[node])
1465 goto found;
1466 }
1467 }
1468 return 0;
1469
1470found:
1471 h->surplus_huge_pages += delta;
1472 h->surplus_huge_pages_node[node] += delta;
1473 return 1;
1474}
1475
1476#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1477static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1478 nodemask_t *nodes_allowed)
1479{
1480 unsigned long min_count, ret;
1481
1482 if (h->order >= MAX_ORDER)
1483 return h->max_huge_pages;
1484
1485 /*
1486 * Increase the pool size
1487 * First take pages out of surplus state. Then make up the
1488 * remaining difference by allocating fresh huge pages.
1489 *
1490 * We might race with alloc_buddy_huge_page() here and be unable
1491 * to convert a surplus huge page to a normal huge page. That is
1492 * not critical, though, it just means the overall size of the
1493 * pool might be one hugepage larger than it needs to be, but
1494 * within all the constraints specified by the sysctls.
1495 */
1496 spin_lock(&hugetlb_lock);
1497 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1498 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1499 break;
1500 }
1501
1502 while (count > persistent_huge_pages(h)) {
1503 /*
1504 * If this allocation races such that we no longer need the
1505 * page, free_huge_page will handle it by freeing the page
1506 * and reducing the surplus.
1507 */
1508 spin_unlock(&hugetlb_lock);
1509 ret = alloc_fresh_huge_page(h, nodes_allowed);
1510 spin_lock(&hugetlb_lock);
1511 if (!ret)
1512 goto out;
1513
1514 /* Bail for signals. Probably ctrl-c from user */
1515 if (signal_pending(current))
1516 goto out;
1517 }
1518
1519 /*
1520 * Decrease the pool size
1521 * First return free pages to the buddy allocator (being careful
1522 * to keep enough around to satisfy reservations). Then place
1523 * pages into surplus state as needed so the pool will shrink
1524 * to the desired size as pages become free.
1525 *
1526 * By placing pages into the surplus state independent of the
1527 * overcommit value, we are allowing the surplus pool size to
1528 * exceed overcommit. There are few sane options here. Since
1529 * alloc_buddy_huge_page() is checking the global counter,
1530 * though, we'll note that we're not allowed to exceed surplus
1531 * and won't grow the pool anywhere else. Not until one of the
1532 * sysctls are changed, or the surplus pages go out of use.
1533 */
1534 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1535 min_count = max(count, min_count);
1536 try_to_free_low(h, min_count, nodes_allowed);
1537 while (min_count < persistent_huge_pages(h)) {
1538 if (!free_pool_huge_page(h, nodes_allowed, 0))
1539 break;
1540 cond_resched_lock(&hugetlb_lock);
1541 }
1542 while (count < persistent_huge_pages(h)) {
1543 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1544 break;
1545 }
1546out:
1547 ret = persistent_huge_pages(h);
1548 spin_unlock(&hugetlb_lock);
1549 return ret;
1550}
1551
1552#define HSTATE_ATTR_RO(_name) \
1553 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1554
1555#define HSTATE_ATTR(_name) \
1556 static struct kobj_attribute _name##_attr = \
1557 __ATTR(_name, 0644, _name##_show, _name##_store)
1558
1559static struct kobject *hugepages_kobj;
1560static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1561
1562static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1563
1564static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1565{
1566 int i;
1567
1568 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1569 if (hstate_kobjs[i] == kobj) {
1570 if (nidp)
1571 *nidp = NUMA_NO_NODE;
1572 return &hstates[i];
1573 }
1574
1575 return kobj_to_node_hstate(kobj, nidp);
1576}
1577
1578static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1579 struct kobj_attribute *attr, char *buf)
1580{
1581 struct hstate *h;
1582 unsigned long nr_huge_pages;
1583 int nid;
1584
1585 h = kobj_to_hstate(kobj, &nid);
1586 if (nid == NUMA_NO_NODE)
1587 nr_huge_pages = h->nr_huge_pages;
1588 else
1589 nr_huge_pages = h->nr_huge_pages_node[nid];
1590
1591 return sprintf(buf, "%lu\n", nr_huge_pages);
1592}
1593
1594static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1595 struct kobject *kobj, struct kobj_attribute *attr,
1596 const char *buf, size_t len)
1597{
1598 int err;
1599 int nid;
1600 unsigned long count;
1601 struct hstate *h;
1602 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1603
1604 err = kstrtoul(buf, 10, &count);
1605 if (err)
1606 goto out;
1607
1608 h = kobj_to_hstate(kobj, &nid);
1609 if (h->order >= MAX_ORDER) {
1610 err = -EINVAL;
1611 goto out;
1612 }
1613
1614 if (nid == NUMA_NO_NODE) {
1615 /*
1616 * global hstate attribute
1617 */
1618 if (!(obey_mempolicy &&
1619 init_nodemask_of_mempolicy(nodes_allowed))) {
1620 NODEMASK_FREE(nodes_allowed);
1621 nodes_allowed = &node_states[N_MEMORY];
1622 }
1623 } else if (nodes_allowed) {
1624 /*
1625 * per node hstate attribute: adjust count to global,
1626 * but restrict alloc/free to the specified node.
1627 */
1628 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1629 init_nodemask_of_node(nodes_allowed, nid);
1630 } else
1631 nodes_allowed = &node_states[N_MEMORY];
1632
1633 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1634
1635 if (nodes_allowed != &node_states[N_MEMORY])
1636 NODEMASK_FREE(nodes_allowed);
1637
1638 return len;
1639out:
1640 NODEMASK_FREE(nodes_allowed);
1641 return err;
1642}
1643
1644static ssize_t nr_hugepages_show(struct kobject *kobj,
1645 struct kobj_attribute *attr, char *buf)
1646{
1647 return nr_hugepages_show_common(kobj, attr, buf);
1648}
1649
1650static ssize_t nr_hugepages_store(struct kobject *kobj,
1651 struct kobj_attribute *attr, const char *buf, size_t len)
1652{
1653 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1654}
1655HSTATE_ATTR(nr_hugepages);
1656
1657#ifdef CONFIG_NUMA
1658
1659/*
1660 * hstate attribute for optionally mempolicy-based constraint on persistent
1661 * huge page alloc/free.
1662 */
1663static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1664 struct kobj_attribute *attr, char *buf)
1665{
1666 return nr_hugepages_show_common(kobj, attr, buf);
1667}
1668
1669static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1670 struct kobj_attribute *attr, const char *buf, size_t len)
1671{
1672 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1673}
1674HSTATE_ATTR(nr_hugepages_mempolicy);
1675#endif
1676
1677
1678static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1679 struct kobj_attribute *attr, char *buf)
1680{
1681 struct hstate *h = kobj_to_hstate(kobj, NULL);
1682 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1683}
1684
1685static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1686 struct kobj_attribute *attr, const char *buf, size_t count)
1687{
1688 int err;
1689 unsigned long input;
1690 struct hstate *h = kobj_to_hstate(kobj, NULL);
1691
1692 if (h->order >= MAX_ORDER)
1693 return -EINVAL;
1694
1695 err = kstrtoul(buf, 10, &input);
1696 if (err)
1697 return err;
1698
1699 spin_lock(&hugetlb_lock);
1700 h->nr_overcommit_huge_pages = input;
1701 spin_unlock(&hugetlb_lock);
1702
1703 return count;
1704}
1705HSTATE_ATTR(nr_overcommit_hugepages);
1706
1707static ssize_t free_hugepages_show(struct kobject *kobj,
1708 struct kobj_attribute *attr, char *buf)
1709{
1710 struct hstate *h;
1711 unsigned long free_huge_pages;
1712 int nid;
1713
1714 h = kobj_to_hstate(kobj, &nid);
1715 if (nid == NUMA_NO_NODE)
1716 free_huge_pages = h->free_huge_pages;
1717 else
1718 free_huge_pages = h->free_huge_pages_node[nid];
1719
1720 return sprintf(buf, "%lu\n", free_huge_pages);
1721}
1722HSTATE_ATTR_RO(free_hugepages);
1723
1724static ssize_t resv_hugepages_show(struct kobject *kobj,
1725 struct kobj_attribute *attr, char *buf)
1726{
1727 struct hstate *h = kobj_to_hstate(kobj, NULL);
1728 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1729}
1730HSTATE_ATTR_RO(resv_hugepages);
1731
1732static ssize_t surplus_hugepages_show(struct kobject *kobj,
1733 struct kobj_attribute *attr, char *buf)
1734{
1735 struct hstate *h;
1736 unsigned long surplus_huge_pages;
1737 int nid;
1738
1739 h = kobj_to_hstate(kobj, &nid);
1740 if (nid == NUMA_NO_NODE)
1741 surplus_huge_pages = h->surplus_huge_pages;
1742 else
1743 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1744
1745 return sprintf(buf, "%lu\n", surplus_huge_pages);
1746}
1747HSTATE_ATTR_RO(surplus_hugepages);
1748
1749static struct attribute *hstate_attrs[] = {
1750 &nr_hugepages_attr.attr,
1751 &nr_overcommit_hugepages_attr.attr,
1752 &free_hugepages_attr.attr,
1753 &resv_hugepages_attr.attr,
1754 &surplus_hugepages_attr.attr,
1755#ifdef CONFIG_NUMA
1756 &nr_hugepages_mempolicy_attr.attr,
1757#endif
1758 NULL,
1759};
1760
1761static struct attribute_group hstate_attr_group = {
1762 .attrs = hstate_attrs,
1763};
1764
1765static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1766 struct kobject **hstate_kobjs,
1767 struct attribute_group *hstate_attr_group)
1768{
1769 int retval;
1770 int hi = hstate_index(h);
1771
1772 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1773 if (!hstate_kobjs[hi])
1774 return -ENOMEM;
1775
1776 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1777 if (retval)
1778 kobject_put(hstate_kobjs[hi]);
1779
1780 return retval;
1781}
1782
1783static void __init hugetlb_sysfs_init(void)
1784{
1785 struct hstate *h;
1786 int err;
1787
1788 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1789 if (!hugepages_kobj)
1790 return;
1791
1792 for_each_hstate(h) {
1793 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1794 hstate_kobjs, &hstate_attr_group);
1795 if (err)
1796 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1797 }
1798}
1799
1800#ifdef CONFIG_NUMA
1801
1802/*
1803 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1804 * with node devices in node_devices[] using a parallel array. The array
1805 * index of a node device or _hstate == node id.
1806 * This is here to avoid any static dependency of the node device driver, in
1807 * the base kernel, on the hugetlb module.
1808 */
1809struct node_hstate {
1810 struct kobject *hugepages_kobj;
1811 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1812};
1813struct node_hstate node_hstates[MAX_NUMNODES];
1814
1815/*
1816 * A subset of global hstate attributes for node devices
1817 */
1818static struct attribute *per_node_hstate_attrs[] = {
1819 &nr_hugepages_attr.attr,
1820 &free_hugepages_attr.attr,
1821 &surplus_hugepages_attr.attr,
1822 NULL,
1823};
1824
1825static struct attribute_group per_node_hstate_attr_group = {
1826 .attrs = per_node_hstate_attrs,
1827};
1828
1829/*
1830 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1831 * Returns node id via non-NULL nidp.
1832 */
1833static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1834{
1835 int nid;
1836
1837 for (nid = 0; nid < nr_node_ids; nid++) {
1838 struct node_hstate *nhs = &node_hstates[nid];
1839 int i;
1840 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1841 if (nhs->hstate_kobjs[i] == kobj) {
1842 if (nidp)
1843 *nidp = nid;
1844 return &hstates[i];
1845 }
1846 }
1847
1848 BUG();
1849 return NULL;
1850}
1851
1852/*
1853 * Unregister hstate attributes from a single node device.
1854 * No-op if no hstate attributes attached.
1855 */
1856static void hugetlb_unregister_node(struct node *node)
1857{
1858 struct hstate *h;
1859 struct node_hstate *nhs = &node_hstates[node->dev.id];
1860
1861 if (!nhs->hugepages_kobj)
1862 return; /* no hstate attributes */
1863
1864 for_each_hstate(h) {
1865 int idx = hstate_index(h);
1866 if (nhs->hstate_kobjs[idx]) {
1867 kobject_put(nhs->hstate_kobjs[idx]);
1868 nhs->hstate_kobjs[idx] = NULL;
1869 }
1870 }
1871
1872 kobject_put(nhs->hugepages_kobj);
1873 nhs->hugepages_kobj = NULL;
1874}
1875
1876/*
1877 * hugetlb module exit: unregister hstate attributes from node devices
1878 * that have them.
1879 */
1880static void hugetlb_unregister_all_nodes(void)
1881{
1882 int nid;
1883
1884 /*
1885 * disable node device registrations.
1886 */
1887 register_hugetlbfs_with_node(NULL, NULL);
1888
1889 /*
1890 * remove hstate attributes from any nodes that have them.
1891 */
1892 for (nid = 0; nid < nr_node_ids; nid++)
1893 hugetlb_unregister_node(node_devices[nid]);
1894}
1895
1896/*
1897 * Register hstate attributes for a single node device.
1898 * No-op if attributes already registered.
1899 */
1900static void hugetlb_register_node(struct node *node)
1901{
1902 struct hstate *h;
1903 struct node_hstate *nhs = &node_hstates[node->dev.id];
1904 int err;
1905
1906 if (nhs->hugepages_kobj)
1907 return; /* already allocated */
1908
1909 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1910 &node->dev.kobj);
1911 if (!nhs->hugepages_kobj)
1912 return;
1913
1914 for_each_hstate(h) {
1915 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1916 nhs->hstate_kobjs,
1917 &per_node_hstate_attr_group);
1918 if (err) {
1919 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1920 h->name, node->dev.id);
1921 hugetlb_unregister_node(node);
1922 break;
1923 }
1924 }
1925}
1926
1927/*
1928 * hugetlb init time: register hstate attributes for all registered node
1929 * devices of nodes that have memory. All on-line nodes should have
1930 * registered their associated device by this time.
1931 */
1932static void hugetlb_register_all_nodes(void)
1933{
1934 int nid;
1935
1936 for_each_node_state(nid, N_MEMORY) {
1937 struct node *node = node_devices[nid];
1938 if (node->dev.id == nid)
1939 hugetlb_register_node(node);
1940 }
1941
1942 /*
1943 * Let the node device driver know we're here so it can
1944 * [un]register hstate attributes on node hotplug.
1945 */
1946 register_hugetlbfs_with_node(hugetlb_register_node,
1947 hugetlb_unregister_node);
1948}
1949#else /* !CONFIG_NUMA */
1950
1951static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1952{
1953 BUG();
1954 if (nidp)
1955 *nidp = -1;
1956 return NULL;
1957}
1958
1959static void hugetlb_unregister_all_nodes(void) { }
1960
1961static void hugetlb_register_all_nodes(void) { }
1962
1963#endif
1964
1965static void __exit hugetlb_exit(void)
1966{
1967 struct hstate *h;
1968
1969 hugetlb_unregister_all_nodes();
1970
1971 for_each_hstate(h) {
1972 kobject_put(hstate_kobjs[hstate_index(h)]);
1973 }
1974
1975 kobject_put(hugepages_kobj);
1976 kfree(htlb_fault_mutex_table);
1977}
1978module_exit(hugetlb_exit);
1979
1980static int __init hugetlb_init(void)
1981{
1982 int i;
1983
1984 if (!hugepages_supported())
1985 return 0;
1986
1987 if (!size_to_hstate(default_hstate_size)) {
1988 default_hstate_size = HPAGE_SIZE;
1989 if (!size_to_hstate(default_hstate_size))
1990 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1991 }
1992 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1993 if (default_hstate_max_huge_pages)
1994 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1995
1996 hugetlb_init_hstates();
1997 gather_bootmem_prealloc();
1998 report_hugepages();
1999
2000 hugetlb_sysfs_init();
2001 hugetlb_register_all_nodes();
2002 hugetlb_cgroup_file_init();
2003
2004#ifdef CONFIG_SMP
2005 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2006#else
2007 num_fault_mutexes = 1;
2008#endif
2009 htlb_fault_mutex_table =
2010 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2011 BUG_ON(!htlb_fault_mutex_table);
2012
2013 for (i = 0; i < num_fault_mutexes; i++)
2014 mutex_init(&htlb_fault_mutex_table[i]);
2015 return 0;
2016}
2017module_init(hugetlb_init);
2018
2019/* Should be called on processing a hugepagesz=... option */
2020void __init hugetlb_add_hstate(unsigned order)
2021{
2022 struct hstate *h;
2023 unsigned long i;
2024
2025 if (size_to_hstate(PAGE_SIZE << order)) {
2026 pr_warning("hugepagesz= specified twice, ignoring\n");
2027 return;
2028 }
2029 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2030 BUG_ON(order == 0);
2031 h = &hstates[hugetlb_max_hstate++];
2032 h->order = order;
2033 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2034 h->nr_huge_pages = 0;
2035 h->free_huge_pages = 0;
2036 for (i = 0; i < MAX_NUMNODES; ++i)
2037 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2038 INIT_LIST_HEAD(&h->hugepage_activelist);
2039 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2040 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2041 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2042 huge_page_size(h)/1024);
2043
2044 parsed_hstate = h;
2045}
2046
2047static int __init hugetlb_nrpages_setup(char *s)
2048{
2049 unsigned long *mhp;
2050 static unsigned long *last_mhp;
2051
2052 /*
2053 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2054 * so this hugepages= parameter goes to the "default hstate".
2055 */
2056 if (!hugetlb_max_hstate)
2057 mhp = &default_hstate_max_huge_pages;
2058 else
2059 mhp = &parsed_hstate->max_huge_pages;
2060
2061 if (mhp == last_mhp) {
2062 pr_warning("hugepages= specified twice without "
2063 "interleaving hugepagesz=, ignoring\n");
2064 return 1;
2065 }
2066
2067 if (sscanf(s, "%lu", mhp) <= 0)
2068 *mhp = 0;
2069
2070 /*
2071 * Global state is always initialized later in hugetlb_init.
2072 * But we need to allocate >= MAX_ORDER hstates here early to still
2073 * use the bootmem allocator.
2074 */
2075 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2076 hugetlb_hstate_alloc_pages(parsed_hstate);
2077
2078 last_mhp = mhp;
2079
2080 return 1;
2081}
2082__setup("hugepages=", hugetlb_nrpages_setup);
2083
2084static int __init hugetlb_default_setup(char *s)
2085{
2086 default_hstate_size = memparse(s, &s);
2087 return 1;
2088}
2089__setup("default_hugepagesz=", hugetlb_default_setup);
2090
2091static unsigned int cpuset_mems_nr(unsigned int *array)
2092{
2093 int node;
2094 unsigned int nr = 0;
2095
2096 for_each_node_mask(node, cpuset_current_mems_allowed)
2097 nr += array[node];
2098
2099 return nr;
2100}
2101
2102#ifdef CONFIG_SYSCTL
2103static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2104 struct ctl_table *table, int write,
2105 void __user *buffer, size_t *length, loff_t *ppos)
2106{
2107 struct hstate *h = &default_hstate;
2108 unsigned long tmp;
2109 int ret;
2110
2111 if (!hugepages_supported())
2112 return -ENOTSUPP;
2113
2114 tmp = h->max_huge_pages;
2115
2116 if (write && h->order >= MAX_ORDER)
2117 return -EINVAL;
2118
2119 table->data = &tmp;
2120 table->maxlen = sizeof(unsigned long);
2121 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2122 if (ret)
2123 goto out;
2124
2125 if (write) {
2126 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2127 GFP_KERNEL | __GFP_NORETRY);
2128 if (!(obey_mempolicy &&
2129 init_nodemask_of_mempolicy(nodes_allowed))) {
2130 NODEMASK_FREE(nodes_allowed);
2131 nodes_allowed = &node_states[N_MEMORY];
2132 }
2133 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2134
2135 if (nodes_allowed != &node_states[N_MEMORY])
2136 NODEMASK_FREE(nodes_allowed);
2137 }
2138out:
2139 return ret;
2140}
2141
2142int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2143 void __user *buffer, size_t *length, loff_t *ppos)
2144{
2145
2146 return hugetlb_sysctl_handler_common(false, table, write,
2147 buffer, length, ppos);
2148}
2149
2150#ifdef CONFIG_NUMA
2151int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2152 void __user *buffer, size_t *length, loff_t *ppos)
2153{
2154 return hugetlb_sysctl_handler_common(true, table, write,
2155 buffer, length, ppos);
2156}
2157#endif /* CONFIG_NUMA */
2158
2159int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2160 void __user *buffer,
2161 size_t *length, loff_t *ppos)
2162{
2163 struct hstate *h = &default_hstate;
2164 unsigned long tmp;
2165 int ret;
2166
2167 if (!hugepages_supported())
2168 return -ENOTSUPP;
2169
2170 tmp = h->nr_overcommit_huge_pages;
2171
2172 if (write && h->order >= MAX_ORDER)
2173 return -EINVAL;
2174
2175 table->data = &tmp;
2176 table->maxlen = sizeof(unsigned long);
2177 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2178 if (ret)
2179 goto out;
2180
2181 if (write) {
2182 spin_lock(&hugetlb_lock);
2183 h->nr_overcommit_huge_pages = tmp;
2184 spin_unlock(&hugetlb_lock);
2185 }
2186out:
2187 return ret;
2188}
2189
2190#endif /* CONFIG_SYSCTL */
2191
2192void hugetlb_report_meminfo(struct seq_file *m)
2193{
2194 struct hstate *h = &default_hstate;
2195 if (!hugepages_supported())
2196 return;
2197 seq_printf(m,
2198 "HugePages_Total: %5lu\n"
2199 "HugePages_Free: %5lu\n"
2200 "HugePages_Rsvd: %5lu\n"
2201 "HugePages_Surp: %5lu\n"
2202 "Hugepagesize: %8lu kB\n",
2203 h->nr_huge_pages,
2204 h->free_huge_pages,
2205 h->resv_huge_pages,
2206 h->surplus_huge_pages,
2207 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2208}
2209
2210int hugetlb_report_node_meminfo(int nid, char *buf)
2211{
2212 struct hstate *h = &default_hstate;
2213 if (!hugepages_supported())
2214 return 0;
2215 return sprintf(buf,
2216 "Node %d HugePages_Total: %5u\n"
2217 "Node %d HugePages_Free: %5u\n"
2218 "Node %d HugePages_Surp: %5u\n",
2219 nid, h->nr_huge_pages_node[nid],
2220 nid, h->free_huge_pages_node[nid],
2221 nid, h->surplus_huge_pages_node[nid]);
2222}
2223
2224void hugetlb_show_meminfo(void)
2225{
2226 struct hstate *h;
2227 int nid;
2228
2229 if (!hugepages_supported())
2230 return;
2231
2232 for_each_node_state(nid, N_MEMORY)
2233 for_each_hstate(h)
2234 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2235 nid,
2236 h->nr_huge_pages_node[nid],
2237 h->free_huge_pages_node[nid],
2238 h->surplus_huge_pages_node[nid],
2239 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2240}
2241
2242/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2243unsigned long hugetlb_total_pages(void)
2244{
2245 struct hstate *h;
2246 unsigned long nr_total_pages = 0;
2247
2248 for_each_hstate(h)
2249 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2250 return nr_total_pages;
2251}
2252
2253static int hugetlb_acct_memory(struct hstate *h, long delta)
2254{
2255 int ret = -ENOMEM;
2256
2257 spin_lock(&hugetlb_lock);
2258 /*
2259 * When cpuset is configured, it breaks the strict hugetlb page
2260 * reservation as the accounting is done on a global variable. Such
2261 * reservation is completely rubbish in the presence of cpuset because
2262 * the reservation is not checked against page availability for the
2263 * current cpuset. Application can still potentially OOM'ed by kernel
2264 * with lack of free htlb page in cpuset that the task is in.
2265 * Attempt to enforce strict accounting with cpuset is almost
2266 * impossible (or too ugly) because cpuset is too fluid that
2267 * task or memory node can be dynamically moved between cpusets.
2268 *
2269 * The change of semantics for shared hugetlb mapping with cpuset is
2270 * undesirable. However, in order to preserve some of the semantics,
2271 * we fall back to check against current free page availability as
2272 * a best attempt and hopefully to minimize the impact of changing
2273 * semantics that cpuset has.
2274 */
2275 if (delta > 0) {
2276 if (gather_surplus_pages(h, delta) < 0)
2277 goto out;
2278
2279 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2280 return_unused_surplus_pages(h, delta);
2281 goto out;
2282 }
2283 }
2284
2285 ret = 0;
2286 if (delta < 0)
2287 return_unused_surplus_pages(h, (unsigned long) -delta);
2288
2289out:
2290 spin_unlock(&hugetlb_lock);
2291 return ret;
2292}
2293
2294static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2295{
2296 struct resv_map *resv = vma_resv_map(vma);
2297
2298 /*
2299 * This new VMA should share its siblings reservation map if present.
2300 * The VMA will only ever have a valid reservation map pointer where
2301 * it is being copied for another still existing VMA. As that VMA
2302 * has a reference to the reservation map it cannot disappear until
2303 * after this open call completes. It is therefore safe to take a
2304 * new reference here without additional locking.
2305 */
2306 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2307 kref_get(&resv->refs);
2308}
2309
2310static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2311{
2312 struct hstate *h = hstate_vma(vma);
2313 struct resv_map *resv = vma_resv_map(vma);
2314 struct hugepage_subpool *spool = subpool_vma(vma);
2315 unsigned long reserve, start, end;
2316
2317 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2318 return;
2319
2320 start = vma_hugecache_offset(h, vma, vma->vm_start);
2321 end = vma_hugecache_offset(h, vma, vma->vm_end);
2322
2323 reserve = (end - start) - region_count(resv, start, end);
2324
2325 kref_put(&resv->refs, resv_map_release);
2326
2327 if (reserve) {
2328 hugetlb_acct_memory(h, -reserve);
2329 hugepage_subpool_put_pages(spool, reserve);
2330 }
2331}
2332
2333/*
2334 * We cannot handle pagefaults against hugetlb pages at all. They cause
2335 * handle_mm_fault() to try to instantiate regular-sized pages in the
2336 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2337 * this far.
2338 */
2339static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2340{
2341 BUG();
2342 return 0;
2343}
2344
2345const struct vm_operations_struct hugetlb_vm_ops = {
2346 .fault = hugetlb_vm_op_fault,
2347 .open = hugetlb_vm_op_open,
2348 .close = hugetlb_vm_op_close,
2349};
2350
2351static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2352 int writable)
2353{
2354 pte_t entry;
2355
2356 if (writable) {
2357 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2358 vma->vm_page_prot)));
2359 } else {
2360 entry = huge_pte_wrprotect(mk_huge_pte(page,
2361 vma->vm_page_prot));
2362 }
2363 entry = pte_mkyoung(entry);
2364 entry = pte_mkhuge(entry);
2365 entry = arch_make_huge_pte(entry, vma, page, writable);
2366
2367 return entry;
2368}
2369
2370static void set_huge_ptep_writable(struct vm_area_struct *vma,
2371 unsigned long address, pte_t *ptep)
2372{
2373 pte_t entry;
2374
2375 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2376 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2377 update_mmu_cache(vma, address, ptep);
2378}
2379
2380
2381int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2382 struct vm_area_struct *vma)
2383{
2384 pte_t *src_pte, *dst_pte, entry;
2385 struct page *ptepage;
2386 unsigned long addr;
2387 int cow;
2388 struct hstate *h = hstate_vma(vma);
2389 unsigned long sz = huge_page_size(h);
2390 unsigned long mmun_start; /* For mmu_notifiers */
2391 unsigned long mmun_end; /* For mmu_notifiers */
2392 int ret = 0;
2393
2394 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2395
2396 mmun_start = vma->vm_start;
2397 mmun_end = vma->vm_end;
2398 if (cow)
2399 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2400
2401 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2402 spinlock_t *src_ptl, *dst_ptl;
2403 src_pte = huge_pte_offset(src, addr);
2404 if (!src_pte)
2405 continue;
2406 dst_pte = huge_pte_alloc(dst, addr, sz);
2407 if (!dst_pte) {
2408 ret = -ENOMEM;
2409 break;
2410 }
2411
2412 /* If the pagetables are shared don't copy or take references */
2413 if (dst_pte == src_pte)
2414 continue;
2415
2416 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2417 src_ptl = huge_pte_lockptr(h, src, src_pte);
2418 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2419 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2420 if (cow)
2421 huge_ptep_set_wrprotect(src, addr, src_pte);
2422 entry = huge_ptep_get(src_pte);
2423 ptepage = pte_page(entry);
2424 get_page(ptepage);
2425 page_dup_rmap(ptepage);
2426 set_huge_pte_at(dst, addr, dst_pte, entry);
2427 }
2428 spin_unlock(src_ptl);
2429 spin_unlock(dst_ptl);
2430 }
2431
2432 if (cow)
2433 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2434
2435 return ret;
2436}
2437
2438static int is_hugetlb_entry_migration(pte_t pte)
2439{
2440 swp_entry_t swp;
2441
2442 if (huge_pte_none(pte) || pte_present(pte))
2443 return 0;
2444 swp = pte_to_swp_entry(pte);
2445 if (non_swap_entry(swp) && is_migration_entry(swp))
2446 return 1;
2447 else
2448 return 0;
2449}
2450
2451static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2452{
2453 swp_entry_t swp;
2454
2455 if (huge_pte_none(pte) || pte_present(pte))
2456 return 0;
2457 swp = pte_to_swp_entry(pte);
2458 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2459 return 1;
2460 else
2461 return 0;
2462}
2463
2464void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2465 unsigned long start, unsigned long end,
2466 struct page *ref_page)
2467{
2468 int force_flush = 0;
2469 struct mm_struct *mm = vma->vm_mm;
2470 unsigned long address;
2471 pte_t *ptep;
2472 pte_t pte;
2473 spinlock_t *ptl;
2474 struct page *page;
2475 struct hstate *h = hstate_vma(vma);
2476 unsigned long sz = huge_page_size(h);
2477 const unsigned long mmun_start = start; /* For mmu_notifiers */
2478 const unsigned long mmun_end = end; /* For mmu_notifiers */
2479
2480 WARN_ON(!is_vm_hugetlb_page(vma));
2481 BUG_ON(start & ~huge_page_mask(h));
2482 BUG_ON(end & ~huge_page_mask(h));
2483
2484 tlb_start_vma(tlb, vma);
2485 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2486again:
2487 for (address = start; address < end; address += sz) {
2488 ptep = huge_pte_offset(mm, address);
2489 if (!ptep)
2490 continue;
2491
2492 ptl = huge_pte_lock(h, mm, ptep);
2493 if (huge_pmd_unshare(mm, &address, ptep))
2494 goto unlock;
2495
2496 pte = huge_ptep_get(ptep);
2497 if (huge_pte_none(pte))
2498 goto unlock;
2499
2500 /*
2501 * HWPoisoned hugepage is already unmapped and dropped reference
2502 */
2503 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2504 huge_pte_clear(mm, address, ptep);
2505 goto unlock;
2506 }
2507
2508 page = pte_page(pte);
2509 /*
2510 * If a reference page is supplied, it is because a specific
2511 * page is being unmapped, not a range. Ensure the page we
2512 * are about to unmap is the actual page of interest.
2513 */
2514 if (ref_page) {
2515 if (page != ref_page)
2516 goto unlock;
2517
2518 /*
2519 * Mark the VMA as having unmapped its page so that
2520 * future faults in this VMA will fail rather than
2521 * looking like data was lost
2522 */
2523 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2524 }
2525
2526 pte = huge_ptep_get_and_clear(mm, address, ptep);
2527 tlb_remove_tlb_entry(tlb, ptep, address);
2528 if (huge_pte_dirty(pte))
2529 set_page_dirty(page);
2530
2531 page_remove_rmap(page);
2532 force_flush = !__tlb_remove_page(tlb, page);
2533 if (force_flush) {
2534 spin_unlock(ptl);
2535 break;
2536 }
2537 /* Bail out after unmapping reference page if supplied */
2538 if (ref_page) {
2539 spin_unlock(ptl);
2540 break;
2541 }
2542unlock:
2543 spin_unlock(ptl);
2544 }
2545 /*
2546 * mmu_gather ran out of room to batch pages, we break out of
2547 * the PTE lock to avoid doing the potential expensive TLB invalidate
2548 * and page-free while holding it.
2549 */
2550 if (force_flush) {
2551 force_flush = 0;
2552 tlb_flush_mmu(tlb);
2553 if (address < end && !ref_page)
2554 goto again;
2555 }
2556 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2557 tlb_end_vma(tlb, vma);
2558}
2559
2560void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2561 struct vm_area_struct *vma, unsigned long start,
2562 unsigned long end, struct page *ref_page)
2563{
2564 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2565
2566 /*
2567 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2568 * test will fail on a vma being torn down, and not grab a page table
2569 * on its way out. We're lucky that the flag has such an appropriate
2570 * name, and can in fact be safely cleared here. We could clear it
2571 * before the __unmap_hugepage_range above, but all that's necessary
2572 * is to clear it before releasing the i_mmap_mutex. This works
2573 * because in the context this is called, the VMA is about to be
2574 * destroyed and the i_mmap_mutex is held.
2575 */
2576 vma->vm_flags &= ~VM_MAYSHARE;
2577}
2578
2579void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2580 unsigned long end, struct page *ref_page)
2581{
2582 struct mm_struct *mm;
2583 struct mmu_gather tlb;
2584
2585 mm = vma->vm_mm;
2586
2587 tlb_gather_mmu(&tlb, mm, start, end);
2588 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2589 tlb_finish_mmu(&tlb, start, end);
2590}
2591
2592/*
2593 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2594 * mappping it owns the reserve page for. The intention is to unmap the page
2595 * from other VMAs and let the children be SIGKILLed if they are faulting the
2596 * same region.
2597 */
2598static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2599 struct page *page, unsigned long address)
2600{
2601 struct hstate *h = hstate_vma(vma);
2602 struct vm_area_struct *iter_vma;
2603 struct address_space *mapping;
2604 pgoff_t pgoff;
2605
2606 /*
2607 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2608 * from page cache lookup which is in HPAGE_SIZE units.
2609 */
2610 address = address & huge_page_mask(h);
2611 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2612 vma->vm_pgoff;
2613 mapping = file_inode(vma->vm_file)->i_mapping;
2614
2615 /*
2616 * Take the mapping lock for the duration of the table walk. As
2617 * this mapping should be shared between all the VMAs,
2618 * __unmap_hugepage_range() is called as the lock is already held
2619 */
2620 mutex_lock(&mapping->i_mmap_mutex);
2621 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2622 /* Do not unmap the current VMA */
2623 if (iter_vma == vma)
2624 continue;
2625
2626 /*
2627 * Unmap the page from other VMAs without their own reserves.
2628 * They get marked to be SIGKILLed if they fault in these
2629 * areas. This is because a future no-page fault on this VMA
2630 * could insert a zeroed page instead of the data existing
2631 * from the time of fork. This would look like data corruption
2632 */
2633 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2634 unmap_hugepage_range(iter_vma, address,
2635 address + huge_page_size(h), page);
2636 }
2637 mutex_unlock(&mapping->i_mmap_mutex);
2638
2639 return 1;
2640}
2641
2642/*
2643 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2644 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2645 * cannot race with other handlers or page migration.
2646 * Keep the pte_same checks anyway to make transition from the mutex easier.
2647 */
2648static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2649 unsigned long address, pte_t *ptep, pte_t pte,
2650 struct page *pagecache_page, spinlock_t *ptl)
2651{
2652 struct hstate *h = hstate_vma(vma);
2653 struct page *old_page, *new_page;
2654 int outside_reserve = 0;
2655 unsigned long mmun_start; /* For mmu_notifiers */
2656 unsigned long mmun_end; /* For mmu_notifiers */
2657
2658 old_page = pte_page(pte);
2659
2660retry_avoidcopy:
2661 /* If no-one else is actually using this page, avoid the copy
2662 * and just make the page writable */
2663 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2664 page_move_anon_rmap(old_page, vma, address);
2665 set_huge_ptep_writable(vma, address, ptep);
2666 return 0;
2667 }
2668
2669 /*
2670 * If the process that created a MAP_PRIVATE mapping is about to
2671 * perform a COW due to a shared page count, attempt to satisfy
2672 * the allocation without using the existing reserves. The pagecache
2673 * page is used to determine if the reserve at this address was
2674 * consumed or not. If reserves were used, a partial faulted mapping
2675 * at the time of fork() could consume its reserves on COW instead
2676 * of the full address range.
2677 */
2678 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2679 old_page != pagecache_page)
2680 outside_reserve = 1;
2681
2682 page_cache_get(old_page);
2683
2684 /* Drop page table lock as buddy allocator may be called */
2685 spin_unlock(ptl);
2686 new_page = alloc_huge_page(vma, address, outside_reserve);
2687
2688 if (IS_ERR(new_page)) {
2689 long err = PTR_ERR(new_page);
2690 page_cache_release(old_page);
2691
2692 /*
2693 * If a process owning a MAP_PRIVATE mapping fails to COW,
2694 * it is due to references held by a child and an insufficient
2695 * huge page pool. To guarantee the original mappers
2696 * reliability, unmap the page from child processes. The child
2697 * may get SIGKILLed if it later faults.
2698 */
2699 if (outside_reserve) {
2700 BUG_ON(huge_pte_none(pte));
2701 if (unmap_ref_private(mm, vma, old_page, address)) {
2702 BUG_ON(huge_pte_none(pte));
2703 spin_lock(ptl);
2704 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2705 if (likely(ptep &&
2706 pte_same(huge_ptep_get(ptep), pte)))
2707 goto retry_avoidcopy;
2708 /*
2709 * race occurs while re-acquiring page table
2710 * lock, and our job is done.
2711 */
2712 return 0;
2713 }
2714 WARN_ON_ONCE(1);
2715 }
2716
2717 /* Caller expects lock to be held */
2718 spin_lock(ptl);
2719 if (err == -ENOMEM)
2720 return VM_FAULT_OOM;
2721 else
2722 return VM_FAULT_SIGBUS;
2723 }
2724
2725 /*
2726 * When the original hugepage is shared one, it does not have
2727 * anon_vma prepared.
2728 */
2729 if (unlikely(anon_vma_prepare(vma))) {
2730 page_cache_release(new_page);
2731 page_cache_release(old_page);
2732 /* Caller expects lock to be held */
2733 spin_lock(ptl);
2734 return VM_FAULT_OOM;
2735 }
2736
2737 copy_user_huge_page(new_page, old_page, address, vma,
2738 pages_per_huge_page(h));
2739 __SetPageUptodate(new_page);
2740
2741 mmun_start = address & huge_page_mask(h);
2742 mmun_end = mmun_start + huge_page_size(h);
2743 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2744 /*
2745 * Retake the page table lock to check for racing updates
2746 * before the page tables are altered
2747 */
2748 spin_lock(ptl);
2749 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2750 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
2751 ClearPagePrivate(new_page);
2752
2753 /* Break COW */
2754 huge_ptep_clear_flush(vma, address, ptep);
2755 set_huge_pte_at(mm, address, ptep,
2756 make_huge_pte(vma, new_page, 1));
2757 page_remove_rmap(old_page);
2758 hugepage_add_new_anon_rmap(new_page, vma, address);
2759 /* Make the old page be freed below */
2760 new_page = old_page;
2761 }
2762 spin_unlock(ptl);
2763 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2764 page_cache_release(new_page);
2765 page_cache_release(old_page);
2766
2767 /* Caller expects lock to be held */
2768 spin_lock(ptl);
2769 return 0;
2770}
2771
2772/* Return the pagecache page at a given address within a VMA */
2773static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2774 struct vm_area_struct *vma, unsigned long address)
2775{
2776 struct address_space *mapping;
2777 pgoff_t idx;
2778
2779 mapping = vma->vm_file->f_mapping;
2780 idx = vma_hugecache_offset(h, vma, address);
2781
2782 return find_lock_page(mapping, idx);
2783}
2784
2785/*
2786 * Return whether there is a pagecache page to back given address within VMA.
2787 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2788 */
2789static bool hugetlbfs_pagecache_present(struct hstate *h,
2790 struct vm_area_struct *vma, unsigned long address)
2791{
2792 struct address_space *mapping;
2793 pgoff_t idx;
2794 struct page *page;
2795
2796 mapping = vma->vm_file->f_mapping;
2797 idx = vma_hugecache_offset(h, vma, address);
2798
2799 page = find_get_page(mapping, idx);
2800 if (page)
2801 put_page(page);
2802 return page != NULL;
2803}
2804
2805static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2806 struct address_space *mapping, pgoff_t idx,
2807 unsigned long address, pte_t *ptep, unsigned int flags)
2808{
2809 struct hstate *h = hstate_vma(vma);
2810 int ret = VM_FAULT_SIGBUS;
2811 int anon_rmap = 0;
2812 unsigned long size;
2813 struct page *page;
2814 pte_t new_pte;
2815 spinlock_t *ptl;
2816
2817 /*
2818 * Currently, we are forced to kill the process in the event the
2819 * original mapper has unmapped pages from the child due to a failed
2820 * COW. Warn that such a situation has occurred as it may not be obvious
2821 */
2822 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2823 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2824 current->pid);
2825 return ret;
2826 }
2827
2828 /*
2829 * Use page lock to guard against racing truncation
2830 * before we get page_table_lock.
2831 */
2832retry:
2833 page = find_lock_page(mapping, idx);
2834 if (!page) {
2835 size = i_size_read(mapping->host) >> huge_page_shift(h);
2836 if (idx >= size)
2837 goto out;
2838 page = alloc_huge_page(vma, address, 0);
2839 if (IS_ERR(page)) {
2840 ret = PTR_ERR(page);
2841 if (ret == -ENOMEM)
2842 ret = VM_FAULT_OOM;
2843 else
2844 ret = VM_FAULT_SIGBUS;
2845 goto out;
2846 }
2847 clear_huge_page(page, address, pages_per_huge_page(h));
2848 __SetPageUptodate(page);
2849
2850 if (vma->vm_flags & VM_MAYSHARE) {
2851 int err;
2852 struct inode *inode = mapping->host;
2853
2854 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2855 if (err) {
2856 put_page(page);
2857 if (err == -EEXIST)
2858 goto retry;
2859 goto out;
2860 }
2861 ClearPagePrivate(page);
2862
2863 spin_lock(&inode->i_lock);
2864 inode->i_blocks += blocks_per_huge_page(h);
2865 spin_unlock(&inode->i_lock);
2866 } else {
2867 lock_page(page);
2868 if (unlikely(anon_vma_prepare(vma))) {
2869 ret = VM_FAULT_OOM;
2870 goto backout_unlocked;
2871 }
2872 anon_rmap = 1;
2873 }
2874 } else {
2875 /*
2876 * If memory error occurs between mmap() and fault, some process
2877 * don't have hwpoisoned swap entry for errored virtual address.
2878 * So we need to block hugepage fault by PG_hwpoison bit check.
2879 */
2880 if (unlikely(PageHWPoison(page))) {
2881 ret = VM_FAULT_HWPOISON |
2882 VM_FAULT_SET_HINDEX(hstate_index(h));
2883 goto backout_unlocked;
2884 }
2885 }
2886
2887 /*
2888 * If we are going to COW a private mapping later, we examine the
2889 * pending reservations for this page now. This will ensure that
2890 * any allocations necessary to record that reservation occur outside
2891 * the spinlock.
2892 */
2893 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2894 if (vma_needs_reservation(h, vma, address) < 0) {
2895 ret = VM_FAULT_OOM;
2896 goto backout_unlocked;
2897 }
2898
2899 ptl = huge_pte_lockptr(h, mm, ptep);
2900 spin_lock(ptl);
2901 size = i_size_read(mapping->host) >> huge_page_shift(h);
2902 if (idx >= size)
2903 goto backout;
2904
2905 ret = 0;
2906 if (!huge_pte_none(huge_ptep_get(ptep)))
2907 goto backout;
2908
2909 if (anon_rmap) {
2910 ClearPagePrivate(page);
2911 hugepage_add_new_anon_rmap(page, vma, address);
2912 } else
2913 page_dup_rmap(page);
2914 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2915 && (vma->vm_flags & VM_SHARED)));
2916 set_huge_pte_at(mm, address, ptep, new_pte);
2917
2918 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2919 /* Optimization, do the COW without a second fault */
2920 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
2921 }
2922
2923 spin_unlock(ptl);
2924 unlock_page(page);
2925out:
2926 return ret;
2927
2928backout:
2929 spin_unlock(ptl);
2930backout_unlocked:
2931 unlock_page(page);
2932 put_page(page);
2933 goto out;
2934}
2935
2936#ifdef CONFIG_SMP
2937static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
2938 struct vm_area_struct *vma,
2939 struct address_space *mapping,
2940 pgoff_t idx, unsigned long address)
2941{
2942 unsigned long key[2];
2943 u32 hash;
2944
2945 if (vma->vm_flags & VM_SHARED) {
2946 key[0] = (unsigned long) mapping;
2947 key[1] = idx;
2948 } else {
2949 key[0] = (unsigned long) mm;
2950 key[1] = address >> huge_page_shift(h);
2951 }
2952
2953 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
2954
2955 return hash & (num_fault_mutexes - 1);
2956}
2957#else
2958/*
2959 * For uniprocesor systems we always use a single mutex, so just
2960 * return 0 and avoid the hashing overhead.
2961 */
2962static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
2963 struct vm_area_struct *vma,
2964 struct address_space *mapping,
2965 pgoff_t idx, unsigned long address)
2966{
2967 return 0;
2968}
2969#endif
2970
2971int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2972 unsigned long address, unsigned int flags)
2973{
2974 pte_t *ptep, entry;
2975 spinlock_t *ptl;
2976 int ret;
2977 u32 hash;
2978 pgoff_t idx;
2979 struct page *page = NULL;
2980 struct page *pagecache_page = NULL;
2981 struct hstate *h = hstate_vma(vma);
2982 struct address_space *mapping;
2983
2984 address &= huge_page_mask(h);
2985
2986 ptep = huge_pte_offset(mm, address);
2987 if (ptep) {
2988 entry = huge_ptep_get(ptep);
2989 if (unlikely(is_hugetlb_entry_migration(entry))) {
2990 migration_entry_wait_huge(vma, mm, ptep);
2991 return 0;
2992 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2993 return VM_FAULT_HWPOISON_LARGE |
2994 VM_FAULT_SET_HINDEX(hstate_index(h));
2995 }
2996
2997 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2998 if (!ptep)
2999 return VM_FAULT_OOM;
3000
3001 mapping = vma->vm_file->f_mapping;
3002 idx = vma_hugecache_offset(h, vma, address);
3003
3004 /*
3005 * Serialize hugepage allocation and instantiation, so that we don't
3006 * get spurious allocation failures if two CPUs race to instantiate
3007 * the same page in the page cache.
3008 */
3009 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3010 mutex_lock(&htlb_fault_mutex_table[hash]);
3011
3012 entry = huge_ptep_get(ptep);
3013 if (huge_pte_none(entry)) {
3014 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3015 goto out_mutex;
3016 }
3017
3018 ret = 0;
3019
3020 /*
3021 * If we are going to COW the mapping later, we examine the pending
3022 * reservations for this page now. This will ensure that any
3023 * allocations necessary to record that reservation occur outside the
3024 * spinlock. For private mappings, we also lookup the pagecache
3025 * page now as it is used to determine if a reservation has been
3026 * consumed.
3027 */
3028 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3029 if (vma_needs_reservation(h, vma, address) < 0) {
3030 ret = VM_FAULT_OOM;
3031 goto out_mutex;
3032 }
3033
3034 if (!(vma->vm_flags & VM_MAYSHARE))
3035 pagecache_page = hugetlbfs_pagecache_page(h,
3036 vma, address);
3037 }
3038
3039 /*
3040 * hugetlb_cow() requires page locks of pte_page(entry) and
3041 * pagecache_page, so here we need take the former one
3042 * when page != pagecache_page or !pagecache_page.
3043 * Note that locking order is always pagecache_page -> page,
3044 * so no worry about deadlock.
3045 */
3046 page = pte_page(entry);
3047 get_page(page);
3048 if (page != pagecache_page)
3049 lock_page(page);
3050
3051 ptl = huge_pte_lockptr(h, mm, ptep);
3052 spin_lock(ptl);
3053 /* Check for a racing update before calling hugetlb_cow */
3054 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3055 goto out_ptl;
3056
3057
3058 if (flags & FAULT_FLAG_WRITE) {
3059 if (!huge_pte_write(entry)) {
3060 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3061 pagecache_page, ptl);
3062 goto out_ptl;
3063 }
3064 entry = huge_pte_mkdirty(entry);
3065 }
3066 entry = pte_mkyoung(entry);
3067 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3068 flags & FAULT_FLAG_WRITE))
3069 update_mmu_cache(vma, address, ptep);
3070
3071out_ptl:
3072 spin_unlock(ptl);
3073
3074 if (pagecache_page) {
3075 unlock_page(pagecache_page);
3076 put_page(pagecache_page);
3077 }
3078 if (page != pagecache_page)
3079 unlock_page(page);
3080 put_page(page);
3081
3082out_mutex:
3083 mutex_unlock(&htlb_fault_mutex_table[hash]);
3084 return ret;
3085}
3086
3087long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3088 struct page **pages, struct vm_area_struct **vmas,
3089 unsigned long *position, unsigned long *nr_pages,
3090 long i, unsigned int flags)
3091{
3092 unsigned long pfn_offset;
3093 unsigned long vaddr = *position;
3094 unsigned long remainder = *nr_pages;
3095 struct hstate *h = hstate_vma(vma);
3096
3097 while (vaddr < vma->vm_end && remainder) {
3098 pte_t *pte;
3099 spinlock_t *ptl = NULL;
3100 int absent;
3101 struct page *page;
3102
3103 /*
3104 * Some archs (sparc64, sh*) have multiple pte_ts to
3105 * each hugepage. We have to make sure we get the
3106 * first, for the page indexing below to work.
3107 *
3108 * Note that page table lock is not held when pte is null.
3109 */
3110 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3111 if (pte)
3112 ptl = huge_pte_lock(h, mm, pte);
3113 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3114
3115 /*
3116 * When coredumping, it suits get_dump_page if we just return
3117 * an error where there's an empty slot with no huge pagecache
3118 * to back it. This way, we avoid allocating a hugepage, and
3119 * the sparse dumpfile avoids allocating disk blocks, but its
3120 * huge holes still show up with zeroes where they need to be.
3121 */
3122 if (absent && (flags & FOLL_DUMP) &&
3123 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3124 if (pte)
3125 spin_unlock(ptl);
3126 remainder = 0;
3127 break;
3128 }
3129
3130 /*
3131 * We need call hugetlb_fault for both hugepages under migration
3132 * (in which case hugetlb_fault waits for the migration,) and
3133 * hwpoisoned hugepages (in which case we need to prevent the
3134 * caller from accessing to them.) In order to do this, we use
3135 * here is_swap_pte instead of is_hugetlb_entry_migration and
3136 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3137 * both cases, and because we can't follow correct pages
3138 * directly from any kind of swap entries.
3139 */
3140 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3141 ((flags & FOLL_WRITE) &&
3142 !huge_pte_write(huge_ptep_get(pte)))) {
3143 int ret;
3144
3145 if (pte)
3146 spin_unlock(ptl);
3147 ret = hugetlb_fault(mm, vma, vaddr,
3148 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3149 if (!(ret & VM_FAULT_ERROR))
3150 continue;
3151
3152 remainder = 0;
3153 break;
3154 }
3155
3156 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3157 page = pte_page(huge_ptep_get(pte));
3158same_page:
3159 if (pages) {
3160 pages[i] = mem_map_offset(page, pfn_offset);
3161 get_page_foll(pages[i]);
3162 }
3163
3164 if (vmas)
3165 vmas[i] = vma;
3166
3167 vaddr += PAGE_SIZE;
3168 ++pfn_offset;
3169 --remainder;
3170 ++i;
3171 if (vaddr < vma->vm_end && remainder &&
3172 pfn_offset < pages_per_huge_page(h)) {
3173 /*
3174 * We use pfn_offset to avoid touching the pageframes
3175 * of this compound page.
3176 */
3177 goto same_page;
3178 }
3179 spin_unlock(ptl);
3180 }
3181 *nr_pages = remainder;
3182 *position = vaddr;
3183
3184 return i ? i : -EFAULT;
3185}
3186
3187unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3188 unsigned long address, unsigned long end, pgprot_t newprot)
3189{
3190 struct mm_struct *mm = vma->vm_mm;
3191 unsigned long start = address;
3192 pte_t *ptep;
3193 pte_t pte;
3194 struct hstate *h = hstate_vma(vma);
3195 unsigned long pages = 0;
3196
3197 BUG_ON(address >= end);
3198 flush_cache_range(vma, address, end);
3199
3200 mmu_notifier_invalidate_range_start(mm, start, end);
3201 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3202 for (; address < end; address += huge_page_size(h)) {
3203 spinlock_t *ptl;
3204 ptep = huge_pte_offset(mm, address);
3205 if (!ptep)
3206 continue;
3207 ptl = huge_pte_lock(h, mm, ptep);
3208 if (huge_pmd_unshare(mm, &address, ptep)) {
3209 pages++;
3210 spin_unlock(ptl);
3211 continue;
3212 }
3213 if (!huge_pte_none(huge_ptep_get(ptep))) {
3214 pte = huge_ptep_get_and_clear(mm, address, ptep);
3215 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3216 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3217 set_huge_pte_at(mm, address, ptep, pte);
3218 pages++;
3219 }
3220 spin_unlock(ptl);
3221 }
3222 /*
3223 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3224 * may have cleared our pud entry and done put_page on the page table:
3225 * once we release i_mmap_mutex, another task can do the final put_page
3226 * and that page table be reused and filled with junk.
3227 */
3228 flush_tlb_range(vma, start, end);
3229 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3230 mmu_notifier_invalidate_range_end(mm, start, end);
3231
3232 return pages << h->order;
3233}
3234
3235int hugetlb_reserve_pages(struct inode *inode,
3236 long from, long to,
3237 struct vm_area_struct *vma,
3238 vm_flags_t vm_flags)
3239{
3240 long ret, chg;
3241 struct hstate *h = hstate_inode(inode);
3242 struct hugepage_subpool *spool = subpool_inode(inode);
3243 struct resv_map *resv_map;
3244
3245 /*
3246 * Only apply hugepage reservation if asked. At fault time, an
3247 * attempt will be made for VM_NORESERVE to allocate a page
3248 * without using reserves
3249 */
3250 if (vm_flags & VM_NORESERVE)
3251 return 0;
3252
3253 /*
3254 * Shared mappings base their reservation on the number of pages that
3255 * are already allocated on behalf of the file. Private mappings need
3256 * to reserve the full area even if read-only as mprotect() may be
3257 * called to make the mapping read-write. Assume !vma is a shm mapping
3258 */
3259 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3260 resv_map = inode_resv_map(inode);
3261
3262 chg = region_chg(resv_map, from, to);
3263
3264 } else {
3265 resv_map = resv_map_alloc();
3266 if (!resv_map)
3267 return -ENOMEM;
3268
3269 chg = to - from;
3270
3271 set_vma_resv_map(vma, resv_map);
3272 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3273 }
3274
3275 if (chg < 0) {
3276 ret = chg;
3277 goto out_err;
3278 }
3279
3280 /* There must be enough pages in the subpool for the mapping */
3281 if (hugepage_subpool_get_pages(spool, chg)) {
3282 ret = -ENOSPC;
3283 goto out_err;
3284 }
3285
3286 /*
3287 * Check enough hugepages are available for the reservation.
3288 * Hand the pages back to the subpool if there are not
3289 */
3290 ret = hugetlb_acct_memory(h, chg);
3291 if (ret < 0) {
3292 hugepage_subpool_put_pages(spool, chg);
3293 goto out_err;
3294 }
3295
3296 /*
3297 * Account for the reservations made. Shared mappings record regions
3298 * that have reservations as they are shared by multiple VMAs.
3299 * When the last VMA disappears, the region map says how much
3300 * the reservation was and the page cache tells how much of
3301 * the reservation was consumed. Private mappings are per-VMA and
3302 * only the consumed reservations are tracked. When the VMA
3303 * disappears, the original reservation is the VMA size and the
3304 * consumed reservations are stored in the map. Hence, nothing
3305 * else has to be done for private mappings here
3306 */
3307 if (!vma || vma->vm_flags & VM_MAYSHARE)
3308 region_add(resv_map, from, to);
3309 return 0;
3310out_err:
3311 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3312 kref_put(&resv_map->refs, resv_map_release);
3313 return ret;
3314}
3315
3316void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3317{
3318 struct hstate *h = hstate_inode(inode);
3319 struct resv_map *resv_map = inode_resv_map(inode);
3320 long chg = 0;
3321 struct hugepage_subpool *spool = subpool_inode(inode);
3322
3323 if (resv_map)
3324 chg = region_truncate(resv_map, offset);
3325 spin_lock(&inode->i_lock);
3326 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3327 spin_unlock(&inode->i_lock);
3328
3329 hugepage_subpool_put_pages(spool, (chg - freed));
3330 hugetlb_acct_memory(h, -(chg - freed));
3331}
3332
3333#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3334static unsigned long page_table_shareable(struct vm_area_struct *svma,
3335 struct vm_area_struct *vma,
3336 unsigned long addr, pgoff_t idx)
3337{
3338 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3339 svma->vm_start;
3340 unsigned long sbase = saddr & PUD_MASK;
3341 unsigned long s_end = sbase + PUD_SIZE;
3342
3343 /* Allow segments to share if only one is marked locked */
3344 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3345 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3346
3347 /*
3348 * match the virtual addresses, permission and the alignment of the
3349 * page table page.
3350 */
3351 if (pmd_index(addr) != pmd_index(saddr) ||
3352 vm_flags != svm_flags ||
3353 sbase < svma->vm_start || svma->vm_end < s_end)
3354 return 0;
3355
3356 return saddr;
3357}
3358
3359static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3360{
3361 unsigned long base = addr & PUD_MASK;
3362 unsigned long end = base + PUD_SIZE;
3363
3364 /*
3365 * check on proper vm_flags and page table alignment
3366 */
3367 if (vma->vm_flags & VM_MAYSHARE &&
3368 vma->vm_start <= base && end <= vma->vm_end)
3369 return 1;
3370 return 0;
3371}
3372
3373/*
3374 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3375 * and returns the corresponding pte. While this is not necessary for the
3376 * !shared pmd case because we can allocate the pmd later as well, it makes the
3377 * code much cleaner. pmd allocation is essential for the shared case because
3378 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3379 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3380 * bad pmd for sharing.
3381 */
3382pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3383{
3384 struct vm_area_struct *vma = find_vma(mm, addr);
3385 struct address_space *mapping = vma->vm_file->f_mapping;
3386 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3387 vma->vm_pgoff;
3388 struct vm_area_struct *svma;
3389 unsigned long saddr;
3390 pte_t *spte = NULL;
3391 pte_t *pte;
3392 spinlock_t *ptl;
3393
3394 if (!vma_shareable(vma, addr))
3395 return (pte_t *)pmd_alloc(mm, pud, addr);
3396
3397 mutex_lock(&mapping->i_mmap_mutex);
3398 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3399 if (svma == vma)
3400 continue;
3401
3402 saddr = page_table_shareable(svma, vma, addr, idx);
3403 if (saddr) {
3404 spte = huge_pte_offset(svma->vm_mm, saddr);
3405 if (spte) {
3406 get_page(virt_to_page(spte));
3407 break;
3408 }
3409 }
3410 }
3411
3412 if (!spte)
3413 goto out;
3414
3415 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3416 spin_lock(ptl);
3417 if (pud_none(*pud))
3418 pud_populate(mm, pud,
3419 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3420 else
3421 put_page(virt_to_page(spte));
3422 spin_unlock(ptl);
3423out:
3424 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3425 mutex_unlock(&mapping->i_mmap_mutex);
3426 return pte;
3427}
3428
3429/*
3430 * unmap huge page backed by shared pte.
3431 *
3432 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3433 * indicated by page_count > 1, unmap is achieved by clearing pud and
3434 * decrementing the ref count. If count == 1, the pte page is not shared.
3435 *
3436 * called with page table lock held.
3437 *
3438 * returns: 1 successfully unmapped a shared pte page
3439 * 0 the underlying pte page is not shared, or it is the last user
3440 */
3441int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3442{
3443 pgd_t *pgd = pgd_offset(mm, *addr);
3444 pud_t *pud = pud_offset(pgd, *addr);
3445
3446 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3447 if (page_count(virt_to_page(ptep)) == 1)
3448 return 0;
3449
3450 pud_clear(pud);
3451 put_page(virt_to_page(ptep));
3452 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3453 return 1;
3454}
3455#define want_pmd_share() (1)
3456#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3457pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3458{
3459 return NULL;
3460}
3461#define want_pmd_share() (0)
3462#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3463
3464#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3465pte_t *huge_pte_alloc(struct mm_struct *mm,
3466 unsigned long addr, unsigned long sz)
3467{
3468 pgd_t *pgd;
3469 pud_t *pud;
3470 pte_t *pte = NULL;
3471
3472 pgd = pgd_offset(mm, addr);
3473 pud = pud_alloc(mm, pgd, addr);
3474 if (pud) {
3475 if (sz == PUD_SIZE) {
3476 pte = (pte_t *)pud;
3477 } else {
3478 BUG_ON(sz != PMD_SIZE);
3479 if (want_pmd_share() && pud_none(*pud))
3480 pte = huge_pmd_share(mm, addr, pud);
3481 else
3482 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3483 }
3484 }
3485 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3486
3487 return pte;
3488}
3489
3490pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3491{
3492 pgd_t *pgd;
3493 pud_t *pud;
3494 pmd_t *pmd = NULL;
3495
3496 pgd = pgd_offset(mm, addr);
3497 if (pgd_present(*pgd)) {
3498 pud = pud_offset(pgd, addr);
3499 if (pud_present(*pud)) {
3500 if (pud_huge(*pud))
3501 return (pte_t *)pud;
3502 pmd = pmd_offset(pud, addr);
3503 }
3504 }
3505 return (pte_t *) pmd;
3506}
3507
3508struct page *
3509follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3510 pmd_t *pmd, int write)
3511{
3512 struct page *page;
3513
3514 page = pte_page(*(pte_t *)pmd);
3515 if (page)
3516 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3517 return page;
3518}
3519
3520struct page *
3521follow_huge_pud(struct mm_struct *mm, unsigned long address,
3522 pud_t *pud, int write)
3523{
3524 struct page *page;
3525
3526 page = pte_page(*(pte_t *)pud);
3527 if (page)
3528 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3529 return page;
3530}
3531
3532#else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3533
3534/* Can be overriden by architectures */
3535struct page * __weak
3536follow_huge_pud(struct mm_struct *mm, unsigned long address,
3537 pud_t *pud, int write)
3538{
3539 BUG();
3540 return NULL;
3541}
3542
3543#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3544
3545#ifdef CONFIG_MEMORY_FAILURE
3546
3547/* Should be called in hugetlb_lock */
3548static int is_hugepage_on_freelist(struct page *hpage)
3549{
3550 struct page *page;
3551 struct page *tmp;
3552 struct hstate *h = page_hstate(hpage);
3553 int nid = page_to_nid(hpage);
3554
3555 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3556 if (page == hpage)
3557 return 1;
3558 return 0;
3559}
3560
3561/*
3562 * This function is called from memory failure code.
3563 * Assume the caller holds page lock of the head page.
3564 */
3565int dequeue_hwpoisoned_huge_page(struct page *hpage)
3566{
3567 struct hstate *h = page_hstate(hpage);
3568 int nid = page_to_nid(hpage);
3569 int ret = -EBUSY;
3570
3571 spin_lock(&hugetlb_lock);
3572 if (is_hugepage_on_freelist(hpage)) {
3573 /*
3574 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3575 * but dangling hpage->lru can trigger list-debug warnings
3576 * (this happens when we call unpoison_memory() on it),
3577 * so let it point to itself with list_del_init().
3578 */
3579 list_del_init(&hpage->lru);
3580 set_page_refcounted(hpage);
3581 h->free_huge_pages--;
3582 h->free_huge_pages_node[nid]--;
3583 ret = 0;
3584 }
3585 spin_unlock(&hugetlb_lock);
3586 return ret;
3587}
3588#endif
3589
3590bool isolate_huge_page(struct page *page, struct list_head *list)
3591{
3592 VM_BUG_ON_PAGE(!PageHead(page), page);
3593 if (!get_page_unless_zero(page))
3594 return false;
3595 spin_lock(&hugetlb_lock);
3596 list_move_tail(&page->lru, list);
3597 spin_unlock(&hugetlb_lock);
3598 return true;
3599}
3600
3601void putback_active_hugepage(struct page *page)
3602{
3603 VM_BUG_ON_PAGE(!PageHead(page), page);
3604 spin_lock(&hugetlb_lock);
3605 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3606 spin_unlock(&hugetlb_lock);
3607 put_page(page);
3608}
3609
3610bool is_hugepage_active(struct page *page)
3611{
3612 VM_BUG_ON_PAGE(!PageHuge(page), page);
3613 /*
3614 * This function can be called for a tail page because the caller,
3615 * scan_movable_pages, scans through a given pfn-range which typically
3616 * covers one memory block. In systems using gigantic hugepage (1GB
3617 * for x86_64,) a hugepage is larger than a memory block, and we don't
3618 * support migrating such large hugepages for now, so return false
3619 * when called for tail pages.
3620 */
3621 if (PageTail(page))
3622 return false;
3623 /*
3624 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3625 * so we should return false for them.
3626 */
3627 if (unlikely(PageHWPoison(page)))
3628 return false;
3629 return page_count(page) > 0;
3630}