Char: synclink, remove unnecessary checks
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / hugetlb.c
blob3a5aeb37c1107ed9735b2e0e943db3212bacaf3c
1 /*
2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
4 */
5 #include <linux/gfp.h>
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/module.h>
9 #include <linux/mm.h>
10 #include <linux/seq_file.h>
11 #include <linux/sysctl.h>
12 #include <linux/highmem.h>
13 #include <linux/mmu_notifier.h>
14 #include <linux/nodemask.h>
15 #include <linux/pagemap.h>
16 #include <linux/mempolicy.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
22 #include <asm/page.h>
23 #include <asm/pgtable.h>
24 #include <asm/io.h>
26 #include <linux/hugetlb.h>
27 #include <linux/node.h>
28 #include "internal.h"
30 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
31 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
32 unsigned long hugepages_treat_as_movable;
34 static int max_hstate;
35 unsigned int default_hstate_idx;
36 struct hstate hstates[HUGE_MAX_HSTATE];
38 __initdata LIST_HEAD(huge_boot_pages);
40 /* for command line parsing */
41 static struct hstate * __initdata parsed_hstate;
42 static unsigned long __initdata default_hstate_max_huge_pages;
43 static unsigned long __initdata default_hstate_size;
45 #define for_each_hstate(h) \
46 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
49 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
51 static DEFINE_SPINLOCK(hugetlb_lock);
54 * Region tracking -- allows tracking of reservations and instantiated pages
55 * across the pages in a mapping.
57 * The region data structures are protected by a combination of the mmap_sem
58 * and the hugetlb_instantion_mutex. To access or modify a region the caller
59 * must either hold the mmap_sem for write, or the mmap_sem for read and
60 * the hugetlb_instantiation mutex:
62 * down_write(&mm->mmap_sem);
63 * or
64 * down_read(&mm->mmap_sem);
65 * mutex_lock(&hugetlb_instantiation_mutex);
67 struct file_region {
68 struct list_head link;
69 long from;
70 long to;
73 static long region_add(struct list_head *head, long f, long t)
75 struct file_region *rg, *nrg, *trg;
77 /* Locate the region we are either in or before. */
78 list_for_each_entry(rg, head, link)
79 if (f <= rg->to)
80 break;
82 /* Round our left edge to the current segment if it encloses us. */
83 if (f > rg->from)
84 f = rg->from;
86 /* Check for and consume any regions we now overlap with. */
87 nrg = rg;
88 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
89 if (&rg->link == head)
90 break;
91 if (rg->from > t)
92 break;
94 /* If this area reaches higher then extend our area to
95 * include it completely. If this is not the first area
96 * which we intend to reuse, free it. */
97 if (rg->to > t)
98 t = rg->to;
99 if (rg != nrg) {
100 list_del(&rg->link);
101 kfree(rg);
104 nrg->from = f;
105 nrg->to = t;
106 return 0;
109 static long region_chg(struct list_head *head, long f, long t)
111 struct file_region *rg, *nrg;
112 long chg = 0;
114 /* Locate the region we are before or in. */
115 list_for_each_entry(rg, head, link)
116 if (f <= rg->to)
117 break;
119 /* If we are below the current region then a new region is required.
120 * Subtle, allocate a new region at the position but make it zero
121 * size such that we can guarantee to record the reservation. */
122 if (&rg->link == head || t < rg->from) {
123 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
124 if (!nrg)
125 return -ENOMEM;
126 nrg->from = f;
127 nrg->to = f;
128 INIT_LIST_HEAD(&nrg->link);
129 list_add(&nrg->link, rg->link.prev);
131 return t - f;
134 /* Round our left edge to the current segment if it encloses us. */
135 if (f > rg->from)
136 f = rg->from;
137 chg = t - f;
139 /* Check for and consume any regions we now overlap with. */
140 list_for_each_entry(rg, rg->link.prev, link) {
141 if (&rg->link == head)
142 break;
143 if (rg->from > t)
144 return chg;
146 /* We overlap with this area, if it extends futher than
147 * us then we must extend ourselves. Account for its
148 * existing reservation. */
149 if (rg->to > t) {
150 chg += rg->to - t;
151 t = rg->to;
153 chg -= rg->to - rg->from;
155 return chg;
158 static long region_truncate(struct list_head *head, long end)
160 struct file_region *rg, *trg;
161 long chg = 0;
163 /* Locate the region we are either in or before. */
164 list_for_each_entry(rg, head, link)
165 if (end <= rg->to)
166 break;
167 if (&rg->link == head)
168 return 0;
170 /* If we are in the middle of a region then adjust it. */
171 if (end > rg->from) {
172 chg = rg->to - end;
173 rg->to = end;
174 rg = list_entry(rg->link.next, typeof(*rg), link);
177 /* Drop any remaining regions. */
178 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
179 if (&rg->link == head)
180 break;
181 chg += rg->to - rg->from;
182 list_del(&rg->link);
183 kfree(rg);
185 return chg;
188 static long region_count(struct list_head *head, long f, long t)
190 struct file_region *rg;
191 long chg = 0;
193 /* Locate each segment we overlap with, and count that overlap. */
194 list_for_each_entry(rg, head, link) {
195 int seg_from;
196 int seg_to;
198 if (rg->to <= f)
199 continue;
200 if (rg->from >= t)
201 break;
203 seg_from = max(rg->from, f);
204 seg_to = min(rg->to, t);
206 chg += seg_to - seg_from;
209 return chg;
213 * Convert the address within this vma to the page offset within
214 * the mapping, in pagecache page units; huge pages here.
216 static pgoff_t vma_hugecache_offset(struct hstate *h,
217 struct vm_area_struct *vma, unsigned long address)
219 return ((address - vma->vm_start) >> huge_page_shift(h)) +
220 (vma->vm_pgoff >> huge_page_order(h));
224 * Return the size of the pages allocated when backing a VMA. In the majority
225 * cases this will be same size as used by the page table entries.
227 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
229 struct hstate *hstate;
231 if (!is_vm_hugetlb_page(vma))
232 return PAGE_SIZE;
234 hstate = hstate_vma(vma);
236 return 1UL << (hstate->order + PAGE_SHIFT);
238 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
241 * Return the page size being used by the MMU to back a VMA. In the majority
242 * of cases, the page size used by the kernel matches the MMU size. On
243 * architectures where it differs, an architecture-specific version of this
244 * function is required.
246 #ifndef vma_mmu_pagesize
247 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
249 return vma_kernel_pagesize(vma);
251 #endif
254 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
255 * bits of the reservation map pointer, which are always clear due to
256 * alignment.
258 #define HPAGE_RESV_OWNER (1UL << 0)
259 #define HPAGE_RESV_UNMAPPED (1UL << 1)
260 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
263 * These helpers are used to track how many pages are reserved for
264 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
265 * is guaranteed to have their future faults succeed.
267 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
268 * the reserve counters are updated with the hugetlb_lock held. It is safe
269 * to reset the VMA at fork() time as it is not in use yet and there is no
270 * chance of the global counters getting corrupted as a result of the values.
272 * The private mapping reservation is represented in a subtly different
273 * manner to a shared mapping. A shared mapping has a region map associated
274 * with the underlying file, this region map represents the backing file
275 * pages which have ever had a reservation assigned which this persists even
276 * after the page is instantiated. A private mapping has a region map
277 * associated with the original mmap which is attached to all VMAs which
278 * reference it, this region map represents those offsets which have consumed
279 * reservation ie. where pages have been instantiated.
281 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
283 return (unsigned long)vma->vm_private_data;
286 static void set_vma_private_data(struct vm_area_struct *vma,
287 unsigned long value)
289 vma->vm_private_data = (void *)value;
292 struct resv_map {
293 struct kref refs;
294 struct list_head regions;
297 static struct resv_map *resv_map_alloc(void)
299 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
300 if (!resv_map)
301 return NULL;
303 kref_init(&resv_map->refs);
304 INIT_LIST_HEAD(&resv_map->regions);
306 return resv_map;
309 static void resv_map_release(struct kref *ref)
311 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
313 /* Clear out any active regions before we release the map. */
314 region_truncate(&resv_map->regions, 0);
315 kfree(resv_map);
318 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
320 VM_BUG_ON(!is_vm_hugetlb_page(vma));
321 if (!(vma->vm_flags & VM_MAYSHARE))
322 return (struct resv_map *)(get_vma_private_data(vma) &
323 ~HPAGE_RESV_MASK);
324 return NULL;
327 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
329 VM_BUG_ON(!is_vm_hugetlb_page(vma));
330 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
332 set_vma_private_data(vma, (get_vma_private_data(vma) &
333 HPAGE_RESV_MASK) | (unsigned long)map);
336 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
338 VM_BUG_ON(!is_vm_hugetlb_page(vma));
339 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
341 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
344 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
346 VM_BUG_ON(!is_vm_hugetlb_page(vma));
348 return (get_vma_private_data(vma) & flag) != 0;
351 /* Decrement the reserved pages in the hugepage pool by one */
352 static void decrement_hugepage_resv_vma(struct hstate *h,
353 struct vm_area_struct *vma)
355 if (vma->vm_flags & VM_NORESERVE)
356 return;
358 if (vma->vm_flags & VM_MAYSHARE) {
359 /* Shared mappings always use reserves */
360 h->resv_huge_pages--;
361 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
363 * Only the process that called mmap() has reserves for
364 * private mappings.
366 h->resv_huge_pages--;
370 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
371 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
373 VM_BUG_ON(!is_vm_hugetlb_page(vma));
374 if (!(vma->vm_flags & VM_MAYSHARE))
375 vma->vm_private_data = (void *)0;
378 /* Returns true if the VMA has associated reserve pages */
379 static int vma_has_reserves(struct vm_area_struct *vma)
381 if (vma->vm_flags & VM_MAYSHARE)
382 return 1;
383 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
384 return 1;
385 return 0;
388 static void clear_gigantic_page(struct page *page,
389 unsigned long addr, unsigned long sz)
391 int i;
392 struct page *p = page;
394 might_sleep();
395 for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
396 cond_resched();
397 clear_user_highpage(p, addr + i * PAGE_SIZE);
400 static void clear_huge_page(struct page *page,
401 unsigned long addr, unsigned long sz)
403 int i;
405 if (unlikely(sz/PAGE_SIZE > MAX_ORDER_NR_PAGES)) {
406 clear_gigantic_page(page, addr, sz);
407 return;
410 might_sleep();
411 for (i = 0; i < sz/PAGE_SIZE; i++) {
412 cond_resched();
413 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
417 static void copy_gigantic_page(struct page *dst, struct page *src,
418 unsigned long addr, struct vm_area_struct *vma)
420 int i;
421 struct hstate *h = hstate_vma(vma);
422 struct page *dst_base = dst;
423 struct page *src_base = src;
424 might_sleep();
425 for (i = 0; i < pages_per_huge_page(h); ) {
426 cond_resched();
427 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
429 i++;
430 dst = mem_map_next(dst, dst_base, i);
431 src = mem_map_next(src, src_base, i);
434 static void copy_huge_page(struct page *dst, struct page *src,
435 unsigned long addr, struct vm_area_struct *vma)
437 int i;
438 struct hstate *h = hstate_vma(vma);
440 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
441 copy_gigantic_page(dst, src, addr, vma);
442 return;
445 might_sleep();
446 for (i = 0; i < pages_per_huge_page(h); i++) {
447 cond_resched();
448 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
452 static void enqueue_huge_page(struct hstate *h, struct page *page)
454 int nid = page_to_nid(page);
455 list_add(&page->lru, &h->hugepage_freelists[nid]);
456 h->free_huge_pages++;
457 h->free_huge_pages_node[nid]++;
460 static struct page *dequeue_huge_page_vma(struct hstate *h,
461 struct vm_area_struct *vma,
462 unsigned long address, int avoid_reserve)
464 int nid;
465 struct page *page = NULL;
466 struct mempolicy *mpol;
467 nodemask_t *nodemask;
468 struct zonelist *zonelist = huge_zonelist(vma, address,
469 htlb_alloc_mask, &mpol, &nodemask);
470 struct zone *zone;
471 struct zoneref *z;
474 * A child process with MAP_PRIVATE mappings created by their parent
475 * have no page reserves. This check ensures that reservations are
476 * not "stolen". The child may still get SIGKILLed
478 if (!vma_has_reserves(vma) &&
479 h->free_huge_pages - h->resv_huge_pages == 0)
480 return NULL;
482 /* If reserves cannot be used, ensure enough pages are in the pool */
483 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
484 return NULL;
486 for_each_zone_zonelist_nodemask(zone, z, zonelist,
487 MAX_NR_ZONES - 1, nodemask) {
488 nid = zone_to_nid(zone);
489 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
490 !list_empty(&h->hugepage_freelists[nid])) {
491 page = list_entry(h->hugepage_freelists[nid].next,
492 struct page, lru);
493 list_del(&page->lru);
494 h->free_huge_pages--;
495 h->free_huge_pages_node[nid]--;
497 if (!avoid_reserve)
498 decrement_hugepage_resv_vma(h, vma);
500 break;
503 mpol_cond_put(mpol);
504 return page;
507 static void update_and_free_page(struct hstate *h, struct page *page)
509 int i;
511 VM_BUG_ON(h->order >= MAX_ORDER);
513 h->nr_huge_pages--;
514 h->nr_huge_pages_node[page_to_nid(page)]--;
515 for (i = 0; i < pages_per_huge_page(h); i++) {
516 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
517 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
518 1 << PG_private | 1<< PG_writeback);
520 set_compound_page_dtor(page, NULL);
521 set_page_refcounted(page);
522 arch_release_hugepage(page);
523 __free_pages(page, huge_page_order(h));
526 struct hstate *size_to_hstate(unsigned long size)
528 struct hstate *h;
530 for_each_hstate(h) {
531 if (huge_page_size(h) == size)
532 return h;
534 return NULL;
537 static void free_huge_page(struct page *page)
540 * Can't pass hstate in here because it is called from the
541 * compound page destructor.
543 struct hstate *h = page_hstate(page);
544 int nid = page_to_nid(page);
545 struct address_space *mapping;
547 mapping = (struct address_space *) page_private(page);
548 set_page_private(page, 0);
549 BUG_ON(page_count(page));
550 INIT_LIST_HEAD(&page->lru);
552 spin_lock(&hugetlb_lock);
553 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
554 update_and_free_page(h, page);
555 h->surplus_huge_pages--;
556 h->surplus_huge_pages_node[nid]--;
557 } else {
558 enqueue_huge_page(h, page);
560 spin_unlock(&hugetlb_lock);
561 if (mapping)
562 hugetlb_put_quota(mapping, 1);
565 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
567 set_compound_page_dtor(page, free_huge_page);
568 spin_lock(&hugetlb_lock);
569 h->nr_huge_pages++;
570 h->nr_huge_pages_node[nid]++;
571 spin_unlock(&hugetlb_lock);
572 put_page(page); /* free it into the hugepage allocator */
575 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
577 int i;
578 int nr_pages = 1 << order;
579 struct page *p = page + 1;
581 /* we rely on prep_new_huge_page to set the destructor */
582 set_compound_order(page, order);
583 __SetPageHead(page);
584 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
585 __SetPageTail(p);
586 p->first_page = page;
590 int PageHuge(struct page *page)
592 compound_page_dtor *dtor;
594 if (!PageCompound(page))
595 return 0;
597 page = compound_head(page);
598 dtor = get_compound_page_dtor(page);
600 return dtor == free_huge_page;
603 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
605 struct page *page;
607 if (h->order >= MAX_ORDER)
608 return NULL;
610 page = alloc_pages_exact_node(nid,
611 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
612 __GFP_REPEAT|__GFP_NOWARN,
613 huge_page_order(h));
614 if (page) {
615 if (arch_prepare_hugepage(page)) {
616 __free_pages(page, huge_page_order(h));
617 return NULL;
619 prep_new_huge_page(h, page, nid);
622 return page;
626 * common helper functions for hstate_next_node_to_{alloc|free}.
627 * We may have allocated or freed a huge page based on a different
628 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
629 * be outside of *nodes_allowed. Ensure that we use an allowed
630 * node for alloc or free.
632 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
634 nid = next_node(nid, *nodes_allowed);
635 if (nid == MAX_NUMNODES)
636 nid = first_node(*nodes_allowed);
637 VM_BUG_ON(nid >= MAX_NUMNODES);
639 return nid;
642 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
644 if (!node_isset(nid, *nodes_allowed))
645 nid = next_node_allowed(nid, nodes_allowed);
646 return nid;
650 * returns the previously saved node ["this node"] from which to
651 * allocate a persistent huge page for the pool and advance the
652 * next node from which to allocate, handling wrap at end of node
653 * mask.
655 static int hstate_next_node_to_alloc(struct hstate *h,
656 nodemask_t *nodes_allowed)
658 int nid;
660 VM_BUG_ON(!nodes_allowed);
662 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
663 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
665 return nid;
668 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
670 struct page *page;
671 int start_nid;
672 int next_nid;
673 int ret = 0;
675 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
676 next_nid = start_nid;
678 do {
679 page = alloc_fresh_huge_page_node(h, next_nid);
680 if (page) {
681 ret = 1;
682 break;
684 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
685 } while (next_nid != start_nid);
687 if (ret)
688 count_vm_event(HTLB_BUDDY_PGALLOC);
689 else
690 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
692 return ret;
696 * helper for free_pool_huge_page() - return the previously saved
697 * node ["this node"] from which to free a huge page. Advance the
698 * next node id whether or not we find a free huge page to free so
699 * that the next attempt to free addresses the next node.
701 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
703 int nid;
705 VM_BUG_ON(!nodes_allowed);
707 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
708 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
710 return nid;
714 * Free huge page from pool from next node to free.
715 * Attempt to keep persistent huge pages more or less
716 * balanced over allowed nodes.
717 * Called with hugetlb_lock locked.
719 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
720 bool acct_surplus)
722 int start_nid;
723 int next_nid;
724 int ret = 0;
726 start_nid = hstate_next_node_to_free(h, nodes_allowed);
727 next_nid = start_nid;
729 do {
731 * If we're returning unused surplus pages, only examine
732 * nodes with surplus pages.
734 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
735 !list_empty(&h->hugepage_freelists[next_nid])) {
736 struct page *page =
737 list_entry(h->hugepage_freelists[next_nid].next,
738 struct page, lru);
739 list_del(&page->lru);
740 h->free_huge_pages--;
741 h->free_huge_pages_node[next_nid]--;
742 if (acct_surplus) {
743 h->surplus_huge_pages--;
744 h->surplus_huge_pages_node[next_nid]--;
746 update_and_free_page(h, page);
747 ret = 1;
748 break;
750 next_nid = hstate_next_node_to_free(h, nodes_allowed);
751 } while (next_nid != start_nid);
753 return ret;
756 static struct page *alloc_buddy_huge_page(struct hstate *h,
757 struct vm_area_struct *vma, unsigned long address)
759 struct page *page;
760 unsigned int nid;
762 if (h->order >= MAX_ORDER)
763 return NULL;
766 * Assume we will successfully allocate the surplus page to
767 * prevent racing processes from causing the surplus to exceed
768 * overcommit
770 * This however introduces a different race, where a process B
771 * tries to grow the static hugepage pool while alloc_pages() is
772 * called by process A. B will only examine the per-node
773 * counters in determining if surplus huge pages can be
774 * converted to normal huge pages in adjust_pool_surplus(). A
775 * won't be able to increment the per-node counter, until the
776 * lock is dropped by B, but B doesn't drop hugetlb_lock until
777 * no more huge pages can be converted from surplus to normal
778 * state (and doesn't try to convert again). Thus, we have a
779 * case where a surplus huge page exists, the pool is grown, and
780 * the surplus huge page still exists after, even though it
781 * should just have been converted to a normal huge page. This
782 * does not leak memory, though, as the hugepage will be freed
783 * once it is out of use. It also does not allow the counters to
784 * go out of whack in adjust_pool_surplus() as we don't modify
785 * the node values until we've gotten the hugepage and only the
786 * per-node value is checked there.
788 spin_lock(&hugetlb_lock);
789 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
790 spin_unlock(&hugetlb_lock);
791 return NULL;
792 } else {
793 h->nr_huge_pages++;
794 h->surplus_huge_pages++;
796 spin_unlock(&hugetlb_lock);
798 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
799 __GFP_REPEAT|__GFP_NOWARN,
800 huge_page_order(h));
802 if (page && arch_prepare_hugepage(page)) {
803 __free_pages(page, huge_page_order(h));
804 return NULL;
807 spin_lock(&hugetlb_lock);
808 if (page) {
810 * This page is now managed by the hugetlb allocator and has
811 * no users -- drop the buddy allocator's reference.
813 put_page_testzero(page);
814 VM_BUG_ON(page_count(page));
815 nid = page_to_nid(page);
816 set_compound_page_dtor(page, free_huge_page);
818 * We incremented the global counters already
820 h->nr_huge_pages_node[nid]++;
821 h->surplus_huge_pages_node[nid]++;
822 __count_vm_event(HTLB_BUDDY_PGALLOC);
823 } else {
824 h->nr_huge_pages--;
825 h->surplus_huge_pages--;
826 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
828 spin_unlock(&hugetlb_lock);
830 return page;
834 * Increase the hugetlb pool such that it can accomodate a reservation
835 * of size 'delta'.
837 static int gather_surplus_pages(struct hstate *h, int delta)
839 struct list_head surplus_list;
840 struct page *page, *tmp;
841 int ret, i;
842 int needed, allocated;
844 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
845 if (needed <= 0) {
846 h->resv_huge_pages += delta;
847 return 0;
850 allocated = 0;
851 INIT_LIST_HEAD(&surplus_list);
853 ret = -ENOMEM;
854 retry:
855 spin_unlock(&hugetlb_lock);
856 for (i = 0; i < needed; i++) {
857 page = alloc_buddy_huge_page(h, NULL, 0);
858 if (!page) {
860 * We were not able to allocate enough pages to
861 * satisfy the entire reservation so we free what
862 * we've allocated so far.
864 spin_lock(&hugetlb_lock);
865 needed = 0;
866 goto free;
869 list_add(&page->lru, &surplus_list);
871 allocated += needed;
874 * After retaking hugetlb_lock, we need to recalculate 'needed'
875 * because either resv_huge_pages or free_huge_pages may have changed.
877 spin_lock(&hugetlb_lock);
878 needed = (h->resv_huge_pages + delta) -
879 (h->free_huge_pages + allocated);
880 if (needed > 0)
881 goto retry;
884 * The surplus_list now contains _at_least_ the number of extra pages
885 * needed to accomodate the reservation. Add the appropriate number
886 * of pages to the hugetlb pool and free the extras back to the buddy
887 * allocator. Commit the entire reservation here to prevent another
888 * process from stealing the pages as they are added to the pool but
889 * before they are reserved.
891 needed += allocated;
892 h->resv_huge_pages += delta;
893 ret = 0;
894 free:
895 /* Free the needed pages to the hugetlb pool */
896 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
897 if ((--needed) < 0)
898 break;
899 list_del(&page->lru);
900 enqueue_huge_page(h, page);
903 /* Free unnecessary surplus pages to the buddy allocator */
904 if (!list_empty(&surplus_list)) {
905 spin_unlock(&hugetlb_lock);
906 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
907 list_del(&page->lru);
909 * The page has a reference count of zero already, so
910 * call free_huge_page directly instead of using
911 * put_page. This must be done with hugetlb_lock
912 * unlocked which is safe because free_huge_page takes
913 * hugetlb_lock before deciding how to free the page.
915 free_huge_page(page);
917 spin_lock(&hugetlb_lock);
920 return ret;
924 * When releasing a hugetlb pool reservation, any surplus pages that were
925 * allocated to satisfy the reservation must be explicitly freed if they were
926 * never used.
927 * Called with hugetlb_lock held.
929 static void return_unused_surplus_pages(struct hstate *h,
930 unsigned long unused_resv_pages)
932 unsigned long nr_pages;
934 /* Uncommit the reservation */
935 h->resv_huge_pages -= unused_resv_pages;
937 /* Cannot return gigantic pages currently */
938 if (h->order >= MAX_ORDER)
939 return;
941 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
944 * We want to release as many surplus pages as possible, spread
945 * evenly across all nodes with memory. Iterate across these nodes
946 * until we can no longer free unreserved surplus pages. This occurs
947 * when the nodes with surplus pages have no free pages.
948 * free_pool_huge_page() will balance the the freed pages across the
949 * on-line nodes with memory and will handle the hstate accounting.
951 while (nr_pages--) {
952 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
953 break;
958 * Determine if the huge page at addr within the vma has an associated
959 * reservation. Where it does not we will need to logically increase
960 * reservation and actually increase quota before an allocation can occur.
961 * Where any new reservation would be required the reservation change is
962 * prepared, but not committed. Once the page has been quota'd allocated
963 * an instantiated the change should be committed via vma_commit_reservation.
964 * No action is required on failure.
966 static long vma_needs_reservation(struct hstate *h,
967 struct vm_area_struct *vma, unsigned long addr)
969 struct address_space *mapping = vma->vm_file->f_mapping;
970 struct inode *inode = mapping->host;
972 if (vma->vm_flags & VM_MAYSHARE) {
973 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
974 return region_chg(&inode->i_mapping->private_list,
975 idx, idx + 1);
977 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
978 return 1;
980 } else {
981 long err;
982 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
983 struct resv_map *reservations = vma_resv_map(vma);
985 err = region_chg(&reservations->regions, idx, idx + 1);
986 if (err < 0)
987 return err;
988 return 0;
991 static void vma_commit_reservation(struct hstate *h,
992 struct vm_area_struct *vma, unsigned long addr)
994 struct address_space *mapping = vma->vm_file->f_mapping;
995 struct inode *inode = mapping->host;
997 if (vma->vm_flags & VM_MAYSHARE) {
998 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
999 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1001 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1002 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1003 struct resv_map *reservations = vma_resv_map(vma);
1005 /* Mark this page used in the map. */
1006 region_add(&reservations->regions, idx, idx + 1);
1010 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1011 unsigned long addr, int avoid_reserve)
1013 struct hstate *h = hstate_vma(vma);
1014 struct page *page;
1015 struct address_space *mapping = vma->vm_file->f_mapping;
1016 struct inode *inode = mapping->host;
1017 long chg;
1020 * Processes that did not create the mapping will have no reserves and
1021 * will not have accounted against quota. Check that the quota can be
1022 * made before satisfying the allocation
1023 * MAP_NORESERVE mappings may also need pages and quota allocated
1024 * if no reserve mapping overlaps.
1026 chg = vma_needs_reservation(h, vma, addr);
1027 if (chg < 0)
1028 return ERR_PTR(chg);
1029 if (chg)
1030 if (hugetlb_get_quota(inode->i_mapping, chg))
1031 return ERR_PTR(-ENOSPC);
1033 spin_lock(&hugetlb_lock);
1034 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1035 spin_unlock(&hugetlb_lock);
1037 if (!page) {
1038 page = alloc_buddy_huge_page(h, vma, addr);
1039 if (!page) {
1040 hugetlb_put_quota(inode->i_mapping, chg);
1041 return ERR_PTR(-VM_FAULT_OOM);
1045 set_page_refcounted(page);
1046 set_page_private(page, (unsigned long) mapping);
1048 vma_commit_reservation(h, vma, addr);
1050 return page;
1053 int __weak alloc_bootmem_huge_page(struct hstate *h)
1055 struct huge_bootmem_page *m;
1056 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1058 while (nr_nodes) {
1059 void *addr;
1061 addr = __alloc_bootmem_node_nopanic(
1062 NODE_DATA(hstate_next_node_to_alloc(h,
1063 &node_states[N_HIGH_MEMORY])),
1064 huge_page_size(h), huge_page_size(h), 0);
1066 if (addr) {
1068 * Use the beginning of the huge page to store the
1069 * huge_bootmem_page struct (until gather_bootmem
1070 * puts them into the mem_map).
1072 m = addr;
1073 goto found;
1075 nr_nodes--;
1077 return 0;
1079 found:
1080 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1081 /* Put them into a private list first because mem_map is not up yet */
1082 list_add(&m->list, &huge_boot_pages);
1083 m->hstate = h;
1084 return 1;
1087 static void prep_compound_huge_page(struct page *page, int order)
1089 if (unlikely(order > (MAX_ORDER - 1)))
1090 prep_compound_gigantic_page(page, order);
1091 else
1092 prep_compound_page(page, order);
1095 /* Put bootmem huge pages into the standard lists after mem_map is up */
1096 static void __init gather_bootmem_prealloc(void)
1098 struct huge_bootmem_page *m;
1100 list_for_each_entry(m, &huge_boot_pages, list) {
1101 struct page *page = virt_to_page(m);
1102 struct hstate *h = m->hstate;
1103 __ClearPageReserved(page);
1104 WARN_ON(page_count(page) != 1);
1105 prep_compound_huge_page(page, h->order);
1106 prep_new_huge_page(h, page, page_to_nid(page));
1110 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1112 unsigned long i;
1114 for (i = 0; i < h->max_huge_pages; ++i) {
1115 if (h->order >= MAX_ORDER) {
1116 if (!alloc_bootmem_huge_page(h))
1117 break;
1118 } else if (!alloc_fresh_huge_page(h,
1119 &node_states[N_HIGH_MEMORY]))
1120 break;
1122 h->max_huge_pages = i;
1125 static void __init hugetlb_init_hstates(void)
1127 struct hstate *h;
1129 for_each_hstate(h) {
1130 /* oversize hugepages were init'ed in early boot */
1131 if (h->order < MAX_ORDER)
1132 hugetlb_hstate_alloc_pages(h);
1136 static char * __init memfmt(char *buf, unsigned long n)
1138 if (n >= (1UL << 30))
1139 sprintf(buf, "%lu GB", n >> 30);
1140 else if (n >= (1UL << 20))
1141 sprintf(buf, "%lu MB", n >> 20);
1142 else
1143 sprintf(buf, "%lu KB", n >> 10);
1144 return buf;
1147 static void __init report_hugepages(void)
1149 struct hstate *h;
1151 for_each_hstate(h) {
1152 char buf[32];
1153 printk(KERN_INFO "HugeTLB registered %s page size, "
1154 "pre-allocated %ld pages\n",
1155 memfmt(buf, huge_page_size(h)),
1156 h->free_huge_pages);
1160 #ifdef CONFIG_HIGHMEM
1161 static void try_to_free_low(struct hstate *h, unsigned long count,
1162 nodemask_t *nodes_allowed)
1164 int i;
1166 if (h->order >= MAX_ORDER)
1167 return;
1169 for_each_node_mask(i, *nodes_allowed) {
1170 struct page *page, *next;
1171 struct list_head *freel = &h->hugepage_freelists[i];
1172 list_for_each_entry_safe(page, next, freel, lru) {
1173 if (count >= h->nr_huge_pages)
1174 return;
1175 if (PageHighMem(page))
1176 continue;
1177 list_del(&page->lru);
1178 update_and_free_page(h, page);
1179 h->free_huge_pages--;
1180 h->free_huge_pages_node[page_to_nid(page)]--;
1184 #else
1185 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1186 nodemask_t *nodes_allowed)
1189 #endif
1192 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1193 * balanced by operating on them in a round-robin fashion.
1194 * Returns 1 if an adjustment was made.
1196 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1197 int delta)
1199 int start_nid, next_nid;
1200 int ret = 0;
1202 VM_BUG_ON(delta != -1 && delta != 1);
1204 if (delta < 0)
1205 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1206 else
1207 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1208 next_nid = start_nid;
1210 do {
1211 int nid = next_nid;
1212 if (delta < 0) {
1214 * To shrink on this node, there must be a surplus page
1216 if (!h->surplus_huge_pages_node[nid]) {
1217 next_nid = hstate_next_node_to_alloc(h,
1218 nodes_allowed);
1219 continue;
1222 if (delta > 0) {
1224 * Surplus cannot exceed the total number of pages
1226 if (h->surplus_huge_pages_node[nid] >=
1227 h->nr_huge_pages_node[nid]) {
1228 next_nid = hstate_next_node_to_free(h,
1229 nodes_allowed);
1230 continue;
1234 h->surplus_huge_pages += delta;
1235 h->surplus_huge_pages_node[nid] += delta;
1236 ret = 1;
1237 break;
1238 } while (next_nid != start_nid);
1240 return ret;
1243 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1244 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1245 nodemask_t *nodes_allowed)
1247 unsigned long min_count, ret;
1249 if (h->order >= MAX_ORDER)
1250 return h->max_huge_pages;
1253 * Increase the pool size
1254 * First take pages out of surplus state. Then make up the
1255 * remaining difference by allocating fresh huge pages.
1257 * We might race with alloc_buddy_huge_page() here and be unable
1258 * to convert a surplus huge page to a normal huge page. That is
1259 * not critical, though, it just means the overall size of the
1260 * pool might be one hugepage larger than it needs to be, but
1261 * within all the constraints specified by the sysctls.
1263 spin_lock(&hugetlb_lock);
1264 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1265 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1266 break;
1269 while (count > persistent_huge_pages(h)) {
1271 * If this allocation races such that we no longer need the
1272 * page, free_huge_page will handle it by freeing the page
1273 * and reducing the surplus.
1275 spin_unlock(&hugetlb_lock);
1276 ret = alloc_fresh_huge_page(h, nodes_allowed);
1277 spin_lock(&hugetlb_lock);
1278 if (!ret)
1279 goto out;
1281 /* Bail for signals. Probably ctrl-c from user */
1282 if (signal_pending(current))
1283 goto out;
1287 * Decrease the pool size
1288 * First return free pages to the buddy allocator (being careful
1289 * to keep enough around to satisfy reservations). Then place
1290 * pages into surplus state as needed so the pool will shrink
1291 * to the desired size as pages become free.
1293 * By placing pages into the surplus state independent of the
1294 * overcommit value, we are allowing the surplus pool size to
1295 * exceed overcommit. There are few sane options here. Since
1296 * alloc_buddy_huge_page() is checking the global counter,
1297 * though, we'll note that we're not allowed to exceed surplus
1298 * and won't grow the pool anywhere else. Not until one of the
1299 * sysctls are changed, or the surplus pages go out of use.
1301 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1302 min_count = max(count, min_count);
1303 try_to_free_low(h, min_count, nodes_allowed);
1304 while (min_count < persistent_huge_pages(h)) {
1305 if (!free_pool_huge_page(h, nodes_allowed, 0))
1306 break;
1308 while (count < persistent_huge_pages(h)) {
1309 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1310 break;
1312 out:
1313 ret = persistent_huge_pages(h);
1314 spin_unlock(&hugetlb_lock);
1315 return ret;
1318 #define HSTATE_ATTR_RO(_name) \
1319 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1321 #define HSTATE_ATTR(_name) \
1322 static struct kobj_attribute _name##_attr = \
1323 __ATTR(_name, 0644, _name##_show, _name##_store)
1325 static struct kobject *hugepages_kobj;
1326 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1328 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1330 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1332 int i;
1334 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1335 if (hstate_kobjs[i] == kobj) {
1336 if (nidp)
1337 *nidp = NUMA_NO_NODE;
1338 return &hstates[i];
1341 return kobj_to_node_hstate(kobj, nidp);
1344 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1345 struct kobj_attribute *attr, char *buf)
1347 struct hstate *h;
1348 unsigned long nr_huge_pages;
1349 int nid;
1351 h = kobj_to_hstate(kobj, &nid);
1352 if (nid == NUMA_NO_NODE)
1353 nr_huge_pages = h->nr_huge_pages;
1354 else
1355 nr_huge_pages = h->nr_huge_pages_node[nid];
1357 return sprintf(buf, "%lu\n", nr_huge_pages);
1359 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1360 struct kobject *kobj, struct kobj_attribute *attr,
1361 const char *buf, size_t len)
1363 int err;
1364 int nid;
1365 unsigned long count;
1366 struct hstate *h;
1367 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1369 err = strict_strtoul(buf, 10, &count);
1370 if (err)
1371 return 0;
1373 h = kobj_to_hstate(kobj, &nid);
1374 if (nid == NUMA_NO_NODE) {
1376 * global hstate attribute
1378 if (!(obey_mempolicy &&
1379 init_nodemask_of_mempolicy(nodes_allowed))) {
1380 NODEMASK_FREE(nodes_allowed);
1381 nodes_allowed = &node_states[N_HIGH_MEMORY];
1383 } else if (nodes_allowed) {
1385 * per node hstate attribute: adjust count to global,
1386 * but restrict alloc/free to the specified node.
1388 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1389 init_nodemask_of_node(nodes_allowed, nid);
1390 } else
1391 nodes_allowed = &node_states[N_HIGH_MEMORY];
1393 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1395 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1396 NODEMASK_FREE(nodes_allowed);
1398 return len;
1401 static ssize_t nr_hugepages_show(struct kobject *kobj,
1402 struct kobj_attribute *attr, char *buf)
1404 return nr_hugepages_show_common(kobj, attr, buf);
1407 static ssize_t nr_hugepages_store(struct kobject *kobj,
1408 struct kobj_attribute *attr, const char *buf, size_t len)
1410 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1412 HSTATE_ATTR(nr_hugepages);
1414 #ifdef CONFIG_NUMA
1417 * hstate attribute for optionally mempolicy-based constraint on persistent
1418 * huge page alloc/free.
1420 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1421 struct kobj_attribute *attr, char *buf)
1423 return nr_hugepages_show_common(kobj, attr, buf);
1426 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1427 struct kobj_attribute *attr, const char *buf, size_t len)
1429 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1431 HSTATE_ATTR(nr_hugepages_mempolicy);
1432 #endif
1435 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1436 struct kobj_attribute *attr, char *buf)
1438 struct hstate *h = kobj_to_hstate(kobj, NULL);
1439 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1441 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1442 struct kobj_attribute *attr, const char *buf, size_t count)
1444 int err;
1445 unsigned long input;
1446 struct hstate *h = kobj_to_hstate(kobj, NULL);
1448 err = strict_strtoul(buf, 10, &input);
1449 if (err)
1450 return 0;
1452 spin_lock(&hugetlb_lock);
1453 h->nr_overcommit_huge_pages = input;
1454 spin_unlock(&hugetlb_lock);
1456 return count;
1458 HSTATE_ATTR(nr_overcommit_hugepages);
1460 static ssize_t free_hugepages_show(struct kobject *kobj,
1461 struct kobj_attribute *attr, char *buf)
1463 struct hstate *h;
1464 unsigned long free_huge_pages;
1465 int nid;
1467 h = kobj_to_hstate(kobj, &nid);
1468 if (nid == NUMA_NO_NODE)
1469 free_huge_pages = h->free_huge_pages;
1470 else
1471 free_huge_pages = h->free_huge_pages_node[nid];
1473 return sprintf(buf, "%lu\n", free_huge_pages);
1475 HSTATE_ATTR_RO(free_hugepages);
1477 static ssize_t resv_hugepages_show(struct kobject *kobj,
1478 struct kobj_attribute *attr, char *buf)
1480 struct hstate *h = kobj_to_hstate(kobj, NULL);
1481 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1483 HSTATE_ATTR_RO(resv_hugepages);
1485 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1486 struct kobj_attribute *attr, char *buf)
1488 struct hstate *h;
1489 unsigned long surplus_huge_pages;
1490 int nid;
1492 h = kobj_to_hstate(kobj, &nid);
1493 if (nid == NUMA_NO_NODE)
1494 surplus_huge_pages = h->surplus_huge_pages;
1495 else
1496 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1498 return sprintf(buf, "%lu\n", surplus_huge_pages);
1500 HSTATE_ATTR_RO(surplus_hugepages);
1502 static struct attribute *hstate_attrs[] = {
1503 &nr_hugepages_attr.attr,
1504 &nr_overcommit_hugepages_attr.attr,
1505 &free_hugepages_attr.attr,
1506 &resv_hugepages_attr.attr,
1507 &surplus_hugepages_attr.attr,
1508 #ifdef CONFIG_NUMA
1509 &nr_hugepages_mempolicy_attr.attr,
1510 #endif
1511 NULL,
1514 static struct attribute_group hstate_attr_group = {
1515 .attrs = hstate_attrs,
1518 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1519 struct kobject **hstate_kobjs,
1520 struct attribute_group *hstate_attr_group)
1522 int retval;
1523 int hi = h - hstates;
1525 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1526 if (!hstate_kobjs[hi])
1527 return -ENOMEM;
1529 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1530 if (retval)
1531 kobject_put(hstate_kobjs[hi]);
1533 return retval;
1536 static void __init hugetlb_sysfs_init(void)
1538 struct hstate *h;
1539 int err;
1541 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1542 if (!hugepages_kobj)
1543 return;
1545 for_each_hstate(h) {
1546 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1547 hstate_kobjs, &hstate_attr_group);
1548 if (err)
1549 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1550 h->name);
1554 #ifdef CONFIG_NUMA
1557 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1558 * with node sysdevs in node_devices[] using a parallel array. The array
1559 * index of a node sysdev or _hstate == node id.
1560 * This is here to avoid any static dependency of the node sysdev driver, in
1561 * the base kernel, on the hugetlb module.
1563 struct node_hstate {
1564 struct kobject *hugepages_kobj;
1565 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1567 struct node_hstate node_hstates[MAX_NUMNODES];
1570 * A subset of global hstate attributes for node sysdevs
1572 static struct attribute *per_node_hstate_attrs[] = {
1573 &nr_hugepages_attr.attr,
1574 &free_hugepages_attr.attr,
1575 &surplus_hugepages_attr.attr,
1576 NULL,
1579 static struct attribute_group per_node_hstate_attr_group = {
1580 .attrs = per_node_hstate_attrs,
1584 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1585 * Returns node id via non-NULL nidp.
1587 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1589 int nid;
1591 for (nid = 0; nid < nr_node_ids; nid++) {
1592 struct node_hstate *nhs = &node_hstates[nid];
1593 int i;
1594 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1595 if (nhs->hstate_kobjs[i] == kobj) {
1596 if (nidp)
1597 *nidp = nid;
1598 return &hstates[i];
1602 BUG();
1603 return NULL;
1607 * Unregister hstate attributes from a single node sysdev.
1608 * No-op if no hstate attributes attached.
1610 void hugetlb_unregister_node(struct node *node)
1612 struct hstate *h;
1613 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1615 if (!nhs->hugepages_kobj)
1616 return; /* no hstate attributes */
1618 for_each_hstate(h)
1619 if (nhs->hstate_kobjs[h - hstates]) {
1620 kobject_put(nhs->hstate_kobjs[h - hstates]);
1621 nhs->hstate_kobjs[h - hstates] = NULL;
1624 kobject_put(nhs->hugepages_kobj);
1625 nhs->hugepages_kobj = NULL;
1629 * hugetlb module exit: unregister hstate attributes from node sysdevs
1630 * that have them.
1632 static void hugetlb_unregister_all_nodes(void)
1634 int nid;
1637 * disable node sysdev registrations.
1639 register_hugetlbfs_with_node(NULL, NULL);
1642 * remove hstate attributes from any nodes that have them.
1644 for (nid = 0; nid < nr_node_ids; nid++)
1645 hugetlb_unregister_node(&node_devices[nid]);
1649 * Register hstate attributes for a single node sysdev.
1650 * No-op if attributes already registered.
1652 void hugetlb_register_node(struct node *node)
1654 struct hstate *h;
1655 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1656 int err;
1658 if (nhs->hugepages_kobj)
1659 return; /* already allocated */
1661 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1662 &node->sysdev.kobj);
1663 if (!nhs->hugepages_kobj)
1664 return;
1666 for_each_hstate(h) {
1667 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1668 nhs->hstate_kobjs,
1669 &per_node_hstate_attr_group);
1670 if (err) {
1671 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1672 " for node %d\n",
1673 h->name, node->sysdev.id);
1674 hugetlb_unregister_node(node);
1675 break;
1681 * hugetlb init time: register hstate attributes for all registered node
1682 * sysdevs of nodes that have memory. All on-line nodes should have
1683 * registered their associated sysdev by this time.
1685 static void hugetlb_register_all_nodes(void)
1687 int nid;
1689 for_each_node_state(nid, N_HIGH_MEMORY) {
1690 struct node *node = &node_devices[nid];
1691 if (node->sysdev.id == nid)
1692 hugetlb_register_node(node);
1696 * Let the node sysdev driver know we're here so it can
1697 * [un]register hstate attributes on node hotplug.
1699 register_hugetlbfs_with_node(hugetlb_register_node,
1700 hugetlb_unregister_node);
1702 #else /* !CONFIG_NUMA */
1704 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1706 BUG();
1707 if (nidp)
1708 *nidp = -1;
1709 return NULL;
1712 static void hugetlb_unregister_all_nodes(void) { }
1714 static void hugetlb_register_all_nodes(void) { }
1716 #endif
1718 static void __exit hugetlb_exit(void)
1720 struct hstate *h;
1722 hugetlb_unregister_all_nodes();
1724 for_each_hstate(h) {
1725 kobject_put(hstate_kobjs[h - hstates]);
1728 kobject_put(hugepages_kobj);
1730 module_exit(hugetlb_exit);
1732 static int __init hugetlb_init(void)
1734 /* Some platform decide whether they support huge pages at boot
1735 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1736 * there is no such support
1738 if (HPAGE_SHIFT == 0)
1739 return 0;
1741 if (!size_to_hstate(default_hstate_size)) {
1742 default_hstate_size = HPAGE_SIZE;
1743 if (!size_to_hstate(default_hstate_size))
1744 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1746 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1747 if (default_hstate_max_huge_pages)
1748 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1750 hugetlb_init_hstates();
1752 gather_bootmem_prealloc();
1754 report_hugepages();
1756 hugetlb_sysfs_init();
1758 hugetlb_register_all_nodes();
1760 return 0;
1762 module_init(hugetlb_init);
1764 /* Should be called on processing a hugepagesz=... option */
1765 void __init hugetlb_add_hstate(unsigned order)
1767 struct hstate *h;
1768 unsigned long i;
1770 if (size_to_hstate(PAGE_SIZE << order)) {
1771 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1772 return;
1774 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1775 BUG_ON(order == 0);
1776 h = &hstates[max_hstate++];
1777 h->order = order;
1778 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1779 h->nr_huge_pages = 0;
1780 h->free_huge_pages = 0;
1781 for (i = 0; i < MAX_NUMNODES; ++i)
1782 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1783 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1784 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1785 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1786 huge_page_size(h)/1024);
1788 parsed_hstate = h;
1791 static int __init hugetlb_nrpages_setup(char *s)
1793 unsigned long *mhp;
1794 static unsigned long *last_mhp;
1797 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1798 * so this hugepages= parameter goes to the "default hstate".
1800 if (!max_hstate)
1801 mhp = &default_hstate_max_huge_pages;
1802 else
1803 mhp = &parsed_hstate->max_huge_pages;
1805 if (mhp == last_mhp) {
1806 printk(KERN_WARNING "hugepages= specified twice without "
1807 "interleaving hugepagesz=, ignoring\n");
1808 return 1;
1811 if (sscanf(s, "%lu", mhp) <= 0)
1812 *mhp = 0;
1815 * Global state is always initialized later in hugetlb_init.
1816 * But we need to allocate >= MAX_ORDER hstates here early to still
1817 * use the bootmem allocator.
1819 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1820 hugetlb_hstate_alloc_pages(parsed_hstate);
1822 last_mhp = mhp;
1824 return 1;
1826 __setup("hugepages=", hugetlb_nrpages_setup);
1828 static int __init hugetlb_default_setup(char *s)
1830 default_hstate_size = memparse(s, &s);
1831 return 1;
1833 __setup("default_hugepagesz=", hugetlb_default_setup);
1835 static unsigned int cpuset_mems_nr(unsigned int *array)
1837 int node;
1838 unsigned int nr = 0;
1840 for_each_node_mask(node, cpuset_current_mems_allowed)
1841 nr += array[node];
1843 return nr;
1846 #ifdef CONFIG_SYSCTL
1847 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1848 struct ctl_table *table, int write,
1849 void __user *buffer, size_t *length, loff_t *ppos)
1851 struct hstate *h = &default_hstate;
1852 unsigned long tmp;
1854 if (!write)
1855 tmp = h->max_huge_pages;
1857 table->data = &tmp;
1858 table->maxlen = sizeof(unsigned long);
1859 proc_doulongvec_minmax(table, write, buffer, length, ppos);
1861 if (write) {
1862 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1863 GFP_KERNEL | __GFP_NORETRY);
1864 if (!(obey_mempolicy &&
1865 init_nodemask_of_mempolicy(nodes_allowed))) {
1866 NODEMASK_FREE(nodes_allowed);
1867 nodes_allowed = &node_states[N_HIGH_MEMORY];
1869 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1871 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1872 NODEMASK_FREE(nodes_allowed);
1875 return 0;
1878 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1879 void __user *buffer, size_t *length, loff_t *ppos)
1882 return hugetlb_sysctl_handler_common(false, table, write,
1883 buffer, length, ppos);
1886 #ifdef CONFIG_NUMA
1887 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1888 void __user *buffer, size_t *length, loff_t *ppos)
1890 return hugetlb_sysctl_handler_common(true, table, write,
1891 buffer, length, ppos);
1893 #endif /* CONFIG_NUMA */
1895 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1896 void __user *buffer,
1897 size_t *length, loff_t *ppos)
1899 proc_dointvec(table, write, buffer, length, ppos);
1900 if (hugepages_treat_as_movable)
1901 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1902 else
1903 htlb_alloc_mask = GFP_HIGHUSER;
1904 return 0;
1907 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1908 void __user *buffer,
1909 size_t *length, loff_t *ppos)
1911 struct hstate *h = &default_hstate;
1912 unsigned long tmp;
1914 if (!write)
1915 tmp = h->nr_overcommit_huge_pages;
1917 table->data = &tmp;
1918 table->maxlen = sizeof(unsigned long);
1919 proc_doulongvec_minmax(table, write, buffer, length, ppos);
1921 if (write) {
1922 spin_lock(&hugetlb_lock);
1923 h->nr_overcommit_huge_pages = tmp;
1924 spin_unlock(&hugetlb_lock);
1927 return 0;
1930 #endif /* CONFIG_SYSCTL */
1932 void hugetlb_report_meminfo(struct seq_file *m)
1934 struct hstate *h = &default_hstate;
1935 seq_printf(m,
1936 "HugePages_Total: %5lu\n"
1937 "HugePages_Free: %5lu\n"
1938 "HugePages_Rsvd: %5lu\n"
1939 "HugePages_Surp: %5lu\n"
1940 "Hugepagesize: %8lu kB\n",
1941 h->nr_huge_pages,
1942 h->free_huge_pages,
1943 h->resv_huge_pages,
1944 h->surplus_huge_pages,
1945 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1948 int hugetlb_report_node_meminfo(int nid, char *buf)
1950 struct hstate *h = &default_hstate;
1951 return sprintf(buf,
1952 "Node %d HugePages_Total: %5u\n"
1953 "Node %d HugePages_Free: %5u\n"
1954 "Node %d HugePages_Surp: %5u\n",
1955 nid, h->nr_huge_pages_node[nid],
1956 nid, h->free_huge_pages_node[nid],
1957 nid, h->surplus_huge_pages_node[nid]);
1960 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1961 unsigned long hugetlb_total_pages(void)
1963 struct hstate *h = &default_hstate;
1964 return h->nr_huge_pages * pages_per_huge_page(h);
1967 static int hugetlb_acct_memory(struct hstate *h, long delta)
1969 int ret = -ENOMEM;
1971 spin_lock(&hugetlb_lock);
1973 * When cpuset is configured, it breaks the strict hugetlb page
1974 * reservation as the accounting is done on a global variable. Such
1975 * reservation is completely rubbish in the presence of cpuset because
1976 * the reservation is not checked against page availability for the
1977 * current cpuset. Application can still potentially OOM'ed by kernel
1978 * with lack of free htlb page in cpuset that the task is in.
1979 * Attempt to enforce strict accounting with cpuset is almost
1980 * impossible (or too ugly) because cpuset is too fluid that
1981 * task or memory node can be dynamically moved between cpusets.
1983 * The change of semantics for shared hugetlb mapping with cpuset is
1984 * undesirable. However, in order to preserve some of the semantics,
1985 * we fall back to check against current free page availability as
1986 * a best attempt and hopefully to minimize the impact of changing
1987 * semantics that cpuset has.
1989 if (delta > 0) {
1990 if (gather_surplus_pages(h, delta) < 0)
1991 goto out;
1993 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1994 return_unused_surplus_pages(h, delta);
1995 goto out;
1999 ret = 0;
2000 if (delta < 0)
2001 return_unused_surplus_pages(h, (unsigned long) -delta);
2003 out:
2004 spin_unlock(&hugetlb_lock);
2005 return ret;
2008 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2010 struct resv_map *reservations = vma_resv_map(vma);
2013 * This new VMA should share its siblings reservation map if present.
2014 * The VMA will only ever have a valid reservation map pointer where
2015 * it is being copied for another still existing VMA. As that VMA
2016 * has a reference to the reservation map it cannot dissappear until
2017 * after this open call completes. It is therefore safe to take a
2018 * new reference here without additional locking.
2020 if (reservations)
2021 kref_get(&reservations->refs);
2024 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2026 struct hstate *h = hstate_vma(vma);
2027 struct resv_map *reservations = vma_resv_map(vma);
2028 unsigned long reserve;
2029 unsigned long start;
2030 unsigned long end;
2032 if (reservations) {
2033 start = vma_hugecache_offset(h, vma, vma->vm_start);
2034 end = vma_hugecache_offset(h, vma, vma->vm_end);
2036 reserve = (end - start) -
2037 region_count(&reservations->regions, start, end);
2039 kref_put(&reservations->refs, resv_map_release);
2041 if (reserve) {
2042 hugetlb_acct_memory(h, -reserve);
2043 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2049 * We cannot handle pagefaults against hugetlb pages at all. They cause
2050 * handle_mm_fault() to try to instantiate regular-sized pages in the
2051 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2052 * this far.
2054 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2056 BUG();
2057 return 0;
2060 const struct vm_operations_struct hugetlb_vm_ops = {
2061 .fault = hugetlb_vm_op_fault,
2062 .open = hugetlb_vm_op_open,
2063 .close = hugetlb_vm_op_close,
2066 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2067 int writable)
2069 pte_t entry;
2071 if (writable) {
2072 entry =
2073 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2074 } else {
2075 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2077 entry = pte_mkyoung(entry);
2078 entry = pte_mkhuge(entry);
2080 return entry;
2083 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2084 unsigned long address, pte_t *ptep)
2086 pte_t entry;
2088 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2089 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
2090 update_mmu_cache(vma, address, ptep);
2095 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2096 struct vm_area_struct *vma)
2098 pte_t *src_pte, *dst_pte, entry;
2099 struct page *ptepage;
2100 unsigned long addr;
2101 int cow;
2102 struct hstate *h = hstate_vma(vma);
2103 unsigned long sz = huge_page_size(h);
2105 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2107 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2108 src_pte = huge_pte_offset(src, addr);
2109 if (!src_pte)
2110 continue;
2111 dst_pte = huge_pte_alloc(dst, addr, sz);
2112 if (!dst_pte)
2113 goto nomem;
2115 /* If the pagetables are shared don't copy or take references */
2116 if (dst_pte == src_pte)
2117 continue;
2119 spin_lock(&dst->page_table_lock);
2120 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2121 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2122 if (cow)
2123 huge_ptep_set_wrprotect(src, addr, src_pte);
2124 entry = huge_ptep_get(src_pte);
2125 ptepage = pte_page(entry);
2126 get_page(ptepage);
2127 set_huge_pte_at(dst, addr, dst_pte, entry);
2129 spin_unlock(&src->page_table_lock);
2130 spin_unlock(&dst->page_table_lock);
2132 return 0;
2134 nomem:
2135 return -ENOMEM;
2138 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2139 unsigned long end, struct page *ref_page)
2141 struct mm_struct *mm = vma->vm_mm;
2142 unsigned long address;
2143 pte_t *ptep;
2144 pte_t pte;
2145 struct page *page;
2146 struct page *tmp;
2147 struct hstate *h = hstate_vma(vma);
2148 unsigned long sz = huge_page_size(h);
2151 * A page gathering list, protected by per file i_mmap_lock. The
2152 * lock is used to avoid list corruption from multiple unmapping
2153 * of the same page since we are using page->lru.
2155 LIST_HEAD(page_list);
2157 WARN_ON(!is_vm_hugetlb_page(vma));
2158 BUG_ON(start & ~huge_page_mask(h));
2159 BUG_ON(end & ~huge_page_mask(h));
2161 mmu_notifier_invalidate_range_start(mm, start, end);
2162 spin_lock(&mm->page_table_lock);
2163 for (address = start; address < end; address += sz) {
2164 ptep = huge_pte_offset(mm, address);
2165 if (!ptep)
2166 continue;
2168 if (huge_pmd_unshare(mm, &address, ptep))
2169 continue;
2172 * If a reference page is supplied, it is because a specific
2173 * page is being unmapped, not a range. Ensure the page we
2174 * are about to unmap is the actual page of interest.
2176 if (ref_page) {
2177 pte = huge_ptep_get(ptep);
2178 if (huge_pte_none(pte))
2179 continue;
2180 page = pte_page(pte);
2181 if (page != ref_page)
2182 continue;
2185 * Mark the VMA as having unmapped its page so that
2186 * future faults in this VMA will fail rather than
2187 * looking like data was lost
2189 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2192 pte = huge_ptep_get_and_clear(mm, address, ptep);
2193 if (huge_pte_none(pte))
2194 continue;
2196 page = pte_page(pte);
2197 if (pte_dirty(pte))
2198 set_page_dirty(page);
2199 list_add(&page->lru, &page_list);
2201 spin_unlock(&mm->page_table_lock);
2202 flush_tlb_range(vma, start, end);
2203 mmu_notifier_invalidate_range_end(mm, start, end);
2204 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2205 list_del(&page->lru);
2206 put_page(page);
2210 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2211 unsigned long end, struct page *ref_page)
2213 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2214 __unmap_hugepage_range(vma, start, end, ref_page);
2215 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2219 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2220 * mappping it owns the reserve page for. The intention is to unmap the page
2221 * from other VMAs and let the children be SIGKILLed if they are faulting the
2222 * same region.
2224 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2225 struct page *page, unsigned long address)
2227 struct hstate *h = hstate_vma(vma);
2228 struct vm_area_struct *iter_vma;
2229 struct address_space *mapping;
2230 struct prio_tree_iter iter;
2231 pgoff_t pgoff;
2234 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2235 * from page cache lookup which is in HPAGE_SIZE units.
2237 address = address & huge_page_mask(h);
2238 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2239 + (vma->vm_pgoff >> PAGE_SHIFT);
2240 mapping = (struct address_space *)page_private(page);
2243 * Take the mapping lock for the duration of the table walk. As
2244 * this mapping should be shared between all the VMAs,
2245 * __unmap_hugepage_range() is called as the lock is already held
2247 spin_lock(&mapping->i_mmap_lock);
2248 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2249 /* Do not unmap the current VMA */
2250 if (iter_vma == vma)
2251 continue;
2254 * Unmap the page from other VMAs without their own reserves.
2255 * They get marked to be SIGKILLed if they fault in these
2256 * areas. This is because a future no-page fault on this VMA
2257 * could insert a zeroed page instead of the data existing
2258 * from the time of fork. This would look like data corruption
2260 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2261 __unmap_hugepage_range(iter_vma,
2262 address, address + huge_page_size(h),
2263 page);
2265 spin_unlock(&mapping->i_mmap_lock);
2267 return 1;
2270 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2271 unsigned long address, pte_t *ptep, pte_t pte,
2272 struct page *pagecache_page)
2274 struct hstate *h = hstate_vma(vma);
2275 struct page *old_page, *new_page;
2276 int avoidcopy;
2277 int outside_reserve = 0;
2279 old_page = pte_page(pte);
2281 retry_avoidcopy:
2282 /* If no-one else is actually using this page, avoid the copy
2283 * and just make the page writable */
2284 avoidcopy = (page_count(old_page) == 1);
2285 if (avoidcopy) {
2286 set_huge_ptep_writable(vma, address, ptep);
2287 return 0;
2291 * If the process that created a MAP_PRIVATE mapping is about to
2292 * perform a COW due to a shared page count, attempt to satisfy
2293 * the allocation without using the existing reserves. The pagecache
2294 * page is used to determine if the reserve at this address was
2295 * consumed or not. If reserves were used, a partial faulted mapping
2296 * at the time of fork() could consume its reserves on COW instead
2297 * of the full address range.
2299 if (!(vma->vm_flags & VM_MAYSHARE) &&
2300 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2301 old_page != pagecache_page)
2302 outside_reserve = 1;
2304 page_cache_get(old_page);
2306 /* Drop page_table_lock as buddy allocator may be called */
2307 spin_unlock(&mm->page_table_lock);
2308 new_page = alloc_huge_page(vma, address, outside_reserve);
2310 if (IS_ERR(new_page)) {
2311 page_cache_release(old_page);
2314 * If a process owning a MAP_PRIVATE mapping fails to COW,
2315 * it is due to references held by a child and an insufficient
2316 * huge page pool. To guarantee the original mappers
2317 * reliability, unmap the page from child processes. The child
2318 * may get SIGKILLed if it later faults.
2320 if (outside_reserve) {
2321 BUG_ON(huge_pte_none(pte));
2322 if (unmap_ref_private(mm, vma, old_page, address)) {
2323 BUG_ON(page_count(old_page) != 1);
2324 BUG_ON(huge_pte_none(pte));
2325 spin_lock(&mm->page_table_lock);
2326 goto retry_avoidcopy;
2328 WARN_ON_ONCE(1);
2331 /* Caller expects lock to be held */
2332 spin_lock(&mm->page_table_lock);
2333 return -PTR_ERR(new_page);
2336 copy_huge_page(new_page, old_page, address, vma);
2337 __SetPageUptodate(new_page);
2340 * Retake the page_table_lock to check for racing updates
2341 * before the page tables are altered
2343 spin_lock(&mm->page_table_lock);
2344 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2345 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2346 /* Break COW */
2347 huge_ptep_clear_flush(vma, address, ptep);
2348 set_huge_pte_at(mm, address, ptep,
2349 make_huge_pte(vma, new_page, 1));
2350 /* Make the old page be freed below */
2351 new_page = old_page;
2353 page_cache_release(new_page);
2354 page_cache_release(old_page);
2355 return 0;
2358 /* Return the pagecache page at a given address within a VMA */
2359 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2360 struct vm_area_struct *vma, unsigned long address)
2362 struct address_space *mapping;
2363 pgoff_t idx;
2365 mapping = vma->vm_file->f_mapping;
2366 idx = vma_hugecache_offset(h, vma, address);
2368 return find_lock_page(mapping, idx);
2372 * Return whether there is a pagecache page to back given address within VMA.
2373 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2375 static bool hugetlbfs_pagecache_present(struct hstate *h,
2376 struct vm_area_struct *vma, unsigned long address)
2378 struct address_space *mapping;
2379 pgoff_t idx;
2380 struct page *page;
2382 mapping = vma->vm_file->f_mapping;
2383 idx = vma_hugecache_offset(h, vma, address);
2385 page = find_get_page(mapping, idx);
2386 if (page)
2387 put_page(page);
2388 return page != NULL;
2391 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2392 unsigned long address, pte_t *ptep, unsigned int flags)
2394 struct hstate *h = hstate_vma(vma);
2395 int ret = VM_FAULT_SIGBUS;
2396 pgoff_t idx;
2397 unsigned long size;
2398 struct page *page;
2399 struct address_space *mapping;
2400 pte_t new_pte;
2403 * Currently, we are forced to kill the process in the event the
2404 * original mapper has unmapped pages from the child due to a failed
2405 * COW. Warn that such a situation has occured as it may not be obvious
2407 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2408 printk(KERN_WARNING
2409 "PID %d killed due to inadequate hugepage pool\n",
2410 current->pid);
2411 return ret;
2414 mapping = vma->vm_file->f_mapping;
2415 idx = vma_hugecache_offset(h, vma, address);
2418 * Use page lock to guard against racing truncation
2419 * before we get page_table_lock.
2421 retry:
2422 page = find_lock_page(mapping, idx);
2423 if (!page) {
2424 size = i_size_read(mapping->host) >> huge_page_shift(h);
2425 if (idx >= size)
2426 goto out;
2427 page = alloc_huge_page(vma, address, 0);
2428 if (IS_ERR(page)) {
2429 ret = -PTR_ERR(page);
2430 goto out;
2432 clear_huge_page(page, address, huge_page_size(h));
2433 __SetPageUptodate(page);
2435 if (vma->vm_flags & VM_MAYSHARE) {
2436 int err;
2437 struct inode *inode = mapping->host;
2439 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2440 if (err) {
2441 put_page(page);
2442 if (err == -EEXIST)
2443 goto retry;
2444 goto out;
2447 spin_lock(&inode->i_lock);
2448 inode->i_blocks += blocks_per_huge_page(h);
2449 spin_unlock(&inode->i_lock);
2450 } else
2451 lock_page(page);
2455 * If we are going to COW a private mapping later, we examine the
2456 * pending reservations for this page now. This will ensure that
2457 * any allocations necessary to record that reservation occur outside
2458 * the spinlock.
2460 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2461 if (vma_needs_reservation(h, vma, address) < 0) {
2462 ret = VM_FAULT_OOM;
2463 goto backout_unlocked;
2466 spin_lock(&mm->page_table_lock);
2467 size = i_size_read(mapping->host) >> huge_page_shift(h);
2468 if (idx >= size)
2469 goto backout;
2471 ret = 0;
2472 if (!huge_pte_none(huge_ptep_get(ptep)))
2473 goto backout;
2475 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2476 && (vma->vm_flags & VM_SHARED)));
2477 set_huge_pte_at(mm, address, ptep, new_pte);
2479 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2480 /* Optimization, do the COW without a second fault */
2481 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2484 spin_unlock(&mm->page_table_lock);
2485 unlock_page(page);
2486 out:
2487 return ret;
2489 backout:
2490 spin_unlock(&mm->page_table_lock);
2491 backout_unlocked:
2492 unlock_page(page);
2493 put_page(page);
2494 goto out;
2497 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2498 unsigned long address, unsigned int flags)
2500 pte_t *ptep;
2501 pte_t entry;
2502 int ret;
2503 struct page *pagecache_page = NULL;
2504 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2505 struct hstate *h = hstate_vma(vma);
2507 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2508 if (!ptep)
2509 return VM_FAULT_OOM;
2512 * Serialize hugepage allocation and instantiation, so that we don't
2513 * get spurious allocation failures if two CPUs race to instantiate
2514 * the same page in the page cache.
2516 mutex_lock(&hugetlb_instantiation_mutex);
2517 entry = huge_ptep_get(ptep);
2518 if (huge_pte_none(entry)) {
2519 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2520 goto out_mutex;
2523 ret = 0;
2526 * If we are going to COW the mapping later, we examine the pending
2527 * reservations for this page now. This will ensure that any
2528 * allocations necessary to record that reservation occur outside the
2529 * spinlock. For private mappings, we also lookup the pagecache
2530 * page now as it is used to determine if a reservation has been
2531 * consumed.
2533 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2534 if (vma_needs_reservation(h, vma, address) < 0) {
2535 ret = VM_FAULT_OOM;
2536 goto out_mutex;
2539 if (!(vma->vm_flags & VM_MAYSHARE))
2540 pagecache_page = hugetlbfs_pagecache_page(h,
2541 vma, address);
2544 spin_lock(&mm->page_table_lock);
2545 /* Check for a racing update before calling hugetlb_cow */
2546 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2547 goto out_page_table_lock;
2550 if (flags & FAULT_FLAG_WRITE) {
2551 if (!pte_write(entry)) {
2552 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2553 pagecache_page);
2554 goto out_page_table_lock;
2556 entry = pte_mkdirty(entry);
2558 entry = pte_mkyoung(entry);
2559 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2560 flags & FAULT_FLAG_WRITE))
2561 update_mmu_cache(vma, address, ptep);
2563 out_page_table_lock:
2564 spin_unlock(&mm->page_table_lock);
2566 if (pagecache_page) {
2567 unlock_page(pagecache_page);
2568 put_page(pagecache_page);
2571 out_mutex:
2572 mutex_unlock(&hugetlb_instantiation_mutex);
2574 return ret;
2577 /* Can be overriden by architectures */
2578 __attribute__((weak)) struct page *
2579 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2580 pud_t *pud, int write)
2582 BUG();
2583 return NULL;
2586 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2587 struct page **pages, struct vm_area_struct **vmas,
2588 unsigned long *position, int *length, int i,
2589 unsigned int flags)
2591 unsigned long pfn_offset;
2592 unsigned long vaddr = *position;
2593 int remainder = *length;
2594 struct hstate *h = hstate_vma(vma);
2596 spin_lock(&mm->page_table_lock);
2597 while (vaddr < vma->vm_end && remainder) {
2598 pte_t *pte;
2599 int absent;
2600 struct page *page;
2603 * Some archs (sparc64, sh*) have multiple pte_ts to
2604 * each hugepage. We have to make sure we get the
2605 * first, for the page indexing below to work.
2607 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2608 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2611 * When coredumping, it suits get_dump_page if we just return
2612 * an error where there's an empty slot with no huge pagecache
2613 * to back it. This way, we avoid allocating a hugepage, and
2614 * the sparse dumpfile avoids allocating disk blocks, but its
2615 * huge holes still show up with zeroes where they need to be.
2617 if (absent && (flags & FOLL_DUMP) &&
2618 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2619 remainder = 0;
2620 break;
2623 if (absent ||
2624 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2625 int ret;
2627 spin_unlock(&mm->page_table_lock);
2628 ret = hugetlb_fault(mm, vma, vaddr,
2629 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2630 spin_lock(&mm->page_table_lock);
2631 if (!(ret & VM_FAULT_ERROR))
2632 continue;
2634 remainder = 0;
2635 break;
2638 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2639 page = pte_page(huge_ptep_get(pte));
2640 same_page:
2641 if (pages) {
2642 pages[i] = mem_map_offset(page, pfn_offset);
2643 get_page(pages[i]);
2646 if (vmas)
2647 vmas[i] = vma;
2649 vaddr += PAGE_SIZE;
2650 ++pfn_offset;
2651 --remainder;
2652 ++i;
2653 if (vaddr < vma->vm_end && remainder &&
2654 pfn_offset < pages_per_huge_page(h)) {
2656 * We use pfn_offset to avoid touching the pageframes
2657 * of this compound page.
2659 goto same_page;
2662 spin_unlock(&mm->page_table_lock);
2663 *length = remainder;
2664 *position = vaddr;
2666 return i ? i : -EFAULT;
2669 void hugetlb_change_protection(struct vm_area_struct *vma,
2670 unsigned long address, unsigned long end, pgprot_t newprot)
2672 struct mm_struct *mm = vma->vm_mm;
2673 unsigned long start = address;
2674 pte_t *ptep;
2675 pte_t pte;
2676 struct hstate *h = hstate_vma(vma);
2678 BUG_ON(address >= end);
2679 flush_cache_range(vma, address, end);
2681 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2682 spin_lock(&mm->page_table_lock);
2683 for (; address < end; address += huge_page_size(h)) {
2684 ptep = huge_pte_offset(mm, address);
2685 if (!ptep)
2686 continue;
2687 if (huge_pmd_unshare(mm, &address, ptep))
2688 continue;
2689 if (!huge_pte_none(huge_ptep_get(ptep))) {
2690 pte = huge_ptep_get_and_clear(mm, address, ptep);
2691 pte = pte_mkhuge(pte_modify(pte, newprot));
2692 set_huge_pte_at(mm, address, ptep, pte);
2695 spin_unlock(&mm->page_table_lock);
2696 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2698 flush_tlb_range(vma, start, end);
2701 int hugetlb_reserve_pages(struct inode *inode,
2702 long from, long to,
2703 struct vm_area_struct *vma,
2704 int acctflag)
2706 long ret, chg;
2707 struct hstate *h = hstate_inode(inode);
2710 * Only apply hugepage reservation if asked. At fault time, an
2711 * attempt will be made for VM_NORESERVE to allocate a page
2712 * and filesystem quota without using reserves
2714 if (acctflag & VM_NORESERVE)
2715 return 0;
2718 * Shared mappings base their reservation on the number of pages that
2719 * are already allocated on behalf of the file. Private mappings need
2720 * to reserve the full area even if read-only as mprotect() may be
2721 * called to make the mapping read-write. Assume !vma is a shm mapping
2723 if (!vma || vma->vm_flags & VM_MAYSHARE)
2724 chg = region_chg(&inode->i_mapping->private_list, from, to);
2725 else {
2726 struct resv_map *resv_map = resv_map_alloc();
2727 if (!resv_map)
2728 return -ENOMEM;
2730 chg = to - from;
2732 set_vma_resv_map(vma, resv_map);
2733 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2736 if (chg < 0)
2737 return chg;
2739 /* There must be enough filesystem quota for the mapping */
2740 if (hugetlb_get_quota(inode->i_mapping, chg))
2741 return -ENOSPC;
2744 * Check enough hugepages are available for the reservation.
2745 * Hand back the quota if there are not
2747 ret = hugetlb_acct_memory(h, chg);
2748 if (ret < 0) {
2749 hugetlb_put_quota(inode->i_mapping, chg);
2750 return ret;
2754 * Account for the reservations made. Shared mappings record regions
2755 * that have reservations as they are shared by multiple VMAs.
2756 * When the last VMA disappears, the region map says how much
2757 * the reservation was and the page cache tells how much of
2758 * the reservation was consumed. Private mappings are per-VMA and
2759 * only the consumed reservations are tracked. When the VMA
2760 * disappears, the original reservation is the VMA size and the
2761 * consumed reservations are stored in the map. Hence, nothing
2762 * else has to be done for private mappings here
2764 if (!vma || vma->vm_flags & VM_MAYSHARE)
2765 region_add(&inode->i_mapping->private_list, from, to);
2766 return 0;
2769 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2771 struct hstate *h = hstate_inode(inode);
2772 long chg = region_truncate(&inode->i_mapping->private_list, offset);
2774 spin_lock(&inode->i_lock);
2775 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2776 spin_unlock(&inode->i_lock);
2778 hugetlb_put_quota(inode->i_mapping, (chg - freed));
2779 hugetlb_acct_memory(h, -(chg - freed));