2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/module.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/cpuset.h>
16 #include <linux/mutex.h>
19 #include <asm/pgtable.h>
21 #include <linux/hugetlb.h>
24 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
25 static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
26 static unsigned long surplus_huge_pages;
27 static unsigned long nr_overcommit_huge_pages;
28 unsigned long max_huge_pages;
29 unsigned long sysctl_overcommit_huge_pages;
30 static struct list_head hugepage_freelists[MAX_NUMNODES];
31 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
32 static unsigned int free_huge_pages_node[MAX_NUMNODES];
33 static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35 unsigned long hugepages_treat_as_movable;
36 static int hugetlb_next_nid;
39 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
41 static DEFINE_SPINLOCK(hugetlb_lock);
44 * These helpers are used to track how many pages are reserved for
45 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
46 * is guaranteed to have their future faults succeed.
48 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
49 * the reserve counters are updated with the hugetlb_lock held. It is safe
50 * to reset the VMA at fork() time as it is not in use yet and there is no
51 * chance of the global counters getting corrupted as a result of the values.
53 static unsigned long vma_resv_huge_pages(struct vm_area_struct *vma)
55 VM_BUG_ON(!is_vm_hugetlb_page(vma));
56 if (!(vma->vm_flags & VM_SHARED))
57 return (unsigned long)vma->vm_private_data;
61 static void set_vma_resv_huge_pages(struct vm_area_struct *vma,
62 unsigned long reserve)
64 VM_BUG_ON(!is_vm_hugetlb_page(vma));
65 VM_BUG_ON(vma->vm_flags & VM_SHARED);
67 vma->vm_private_data = (void *)reserve;
70 /* Decrement the reserved pages in the hugepage pool by one */
71 static void decrement_hugepage_resv_vma(struct vm_area_struct *vma)
73 if (vma->vm_flags & VM_SHARED) {
74 /* Shared mappings always use reserves */
78 * Only the process that called mmap() has reserves for
81 if (vma_resv_huge_pages(vma)) {
83 reserve = (unsigned long)vma->vm_private_data - 1;
84 vma->vm_private_data = (void *)reserve;
89 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
91 VM_BUG_ON(!is_vm_hugetlb_page(vma));
92 if (!(vma->vm_flags & VM_SHARED))
93 vma->vm_private_data = (void *)0;
96 /* Returns true if the VMA has associated reserve pages */
97 static int vma_has_private_reserves(struct vm_area_struct *vma)
99 if (vma->vm_flags & VM_SHARED)
101 if (!vma_resv_huge_pages(vma))
106 static void clear_huge_page(struct page *page, unsigned long addr)
111 for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
113 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
117 static void copy_huge_page(struct page *dst, struct page *src,
118 unsigned long addr, struct vm_area_struct *vma)
123 for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
125 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
129 static void enqueue_huge_page(struct page *page)
131 int nid = page_to_nid(page);
132 list_add(&page->lru, &hugepage_freelists[nid]);
134 free_huge_pages_node[nid]++;
137 static struct page *dequeue_huge_page(void)
140 struct page *page = NULL;
142 for (nid = 0; nid < MAX_NUMNODES; ++nid) {
143 if (!list_empty(&hugepage_freelists[nid])) {
144 page = list_entry(hugepage_freelists[nid].next,
146 list_del(&page->lru);
148 free_huge_pages_node[nid]--;
155 static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
156 unsigned long address)
159 struct page *page = NULL;
160 struct mempolicy *mpol;
161 nodemask_t *nodemask;
162 struct zonelist *zonelist = huge_zonelist(vma, address,
163 htlb_alloc_mask, &mpol, &nodemask);
168 * A child process with MAP_PRIVATE mappings created by their parent
169 * have no page reserves. This check ensures that reservations are
170 * not "stolen". The child may still get SIGKILLed
172 if (!vma_has_private_reserves(vma) &&
173 free_huge_pages - resv_huge_pages == 0)
176 for_each_zone_zonelist_nodemask(zone, z, zonelist,
177 MAX_NR_ZONES - 1, nodemask) {
178 nid = zone_to_nid(zone);
179 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
180 !list_empty(&hugepage_freelists[nid])) {
181 page = list_entry(hugepage_freelists[nid].next,
183 list_del(&page->lru);
185 free_huge_pages_node[nid]--;
186 decrement_hugepage_resv_vma(vma);
195 static void update_and_free_page(struct page *page)
199 nr_huge_pages_node[page_to_nid(page)]--;
200 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
201 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
202 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
203 1 << PG_private | 1<< PG_writeback);
205 set_compound_page_dtor(page, NULL);
206 set_page_refcounted(page);
207 arch_release_hugepage(page);
208 __free_pages(page, HUGETLB_PAGE_ORDER);
211 static void free_huge_page(struct page *page)
213 int nid = page_to_nid(page);
214 struct address_space *mapping;
216 mapping = (struct address_space *) page_private(page);
217 set_page_private(page, 0);
218 BUG_ON(page_count(page));
219 INIT_LIST_HEAD(&page->lru);
221 spin_lock(&hugetlb_lock);
222 if (surplus_huge_pages_node[nid]) {
223 update_and_free_page(page);
224 surplus_huge_pages--;
225 surplus_huge_pages_node[nid]--;
227 enqueue_huge_page(page);
229 spin_unlock(&hugetlb_lock);
231 hugetlb_put_quota(mapping, 1);
235 * Increment or decrement surplus_huge_pages. Keep node-specific counters
236 * balanced by operating on them in a round-robin fashion.
237 * Returns 1 if an adjustment was made.
239 static int adjust_pool_surplus(int delta)
245 VM_BUG_ON(delta != -1 && delta != 1);
247 nid = next_node(nid, node_online_map);
248 if (nid == MAX_NUMNODES)
249 nid = first_node(node_online_map);
251 /* To shrink on this node, there must be a surplus page */
252 if (delta < 0 && !surplus_huge_pages_node[nid])
254 /* Surplus cannot exceed the total number of pages */
255 if (delta > 0 && surplus_huge_pages_node[nid] >=
256 nr_huge_pages_node[nid])
259 surplus_huge_pages += delta;
260 surplus_huge_pages_node[nid] += delta;
263 } while (nid != prev_nid);
269 static struct page *alloc_fresh_huge_page_node(int nid)
273 page = alloc_pages_node(nid,
274 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
275 __GFP_REPEAT|__GFP_NOWARN,
278 if (arch_prepare_hugepage(page)) {
279 __free_pages(page, HUGETLB_PAGE_ORDER);
282 set_compound_page_dtor(page, free_huge_page);
283 spin_lock(&hugetlb_lock);
285 nr_huge_pages_node[nid]++;
286 spin_unlock(&hugetlb_lock);
287 put_page(page); /* free it into the hugepage allocator */
293 static int alloc_fresh_huge_page(void)
300 start_nid = hugetlb_next_nid;
303 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
307 * Use a helper variable to find the next node and then
308 * copy it back to hugetlb_next_nid afterwards:
309 * otherwise there's a window in which a racer might
310 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
311 * But we don't need to use a spin_lock here: it really
312 * doesn't matter if occasionally a racer chooses the
313 * same nid as we do. Move nid forward in the mask even
314 * if we just successfully allocated a hugepage so that
315 * the next caller gets hugepages on the next node.
317 next_nid = next_node(hugetlb_next_nid, node_online_map);
318 if (next_nid == MAX_NUMNODES)
319 next_nid = first_node(node_online_map);
320 hugetlb_next_nid = next_nid;
321 } while (!page && hugetlb_next_nid != start_nid);
324 count_vm_event(HTLB_BUDDY_PGALLOC);
326 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
331 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
332 unsigned long address)
338 * Assume we will successfully allocate the surplus page to
339 * prevent racing processes from causing the surplus to exceed
342 * This however introduces a different race, where a process B
343 * tries to grow the static hugepage pool while alloc_pages() is
344 * called by process A. B will only examine the per-node
345 * counters in determining if surplus huge pages can be
346 * converted to normal huge pages in adjust_pool_surplus(). A
347 * won't be able to increment the per-node counter, until the
348 * lock is dropped by B, but B doesn't drop hugetlb_lock until
349 * no more huge pages can be converted from surplus to normal
350 * state (and doesn't try to convert again). Thus, we have a
351 * case where a surplus huge page exists, the pool is grown, and
352 * the surplus huge page still exists after, even though it
353 * should just have been converted to a normal huge page. This
354 * does not leak memory, though, as the hugepage will be freed
355 * once it is out of use. It also does not allow the counters to
356 * go out of whack in adjust_pool_surplus() as we don't modify
357 * the node values until we've gotten the hugepage and only the
358 * per-node value is checked there.
360 spin_lock(&hugetlb_lock);
361 if (surplus_huge_pages >= nr_overcommit_huge_pages) {
362 spin_unlock(&hugetlb_lock);
366 surplus_huge_pages++;
368 spin_unlock(&hugetlb_lock);
370 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
371 __GFP_REPEAT|__GFP_NOWARN,
374 spin_lock(&hugetlb_lock);
377 * This page is now managed by the hugetlb allocator and has
378 * no users -- drop the buddy allocator's reference.
380 put_page_testzero(page);
381 VM_BUG_ON(page_count(page));
382 nid = page_to_nid(page);
383 set_compound_page_dtor(page, free_huge_page);
385 * We incremented the global counters already
387 nr_huge_pages_node[nid]++;
388 surplus_huge_pages_node[nid]++;
389 __count_vm_event(HTLB_BUDDY_PGALLOC);
392 surplus_huge_pages--;
393 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
395 spin_unlock(&hugetlb_lock);
401 * Increase the hugetlb pool such that it can accomodate a reservation
404 static int gather_surplus_pages(int delta)
406 struct list_head surplus_list;
407 struct page *page, *tmp;
409 int needed, allocated;
411 needed = (resv_huge_pages + delta) - free_huge_pages;
413 resv_huge_pages += delta;
418 INIT_LIST_HEAD(&surplus_list);
422 spin_unlock(&hugetlb_lock);
423 for (i = 0; i < needed; i++) {
424 page = alloc_buddy_huge_page(NULL, 0);
427 * We were not able to allocate enough pages to
428 * satisfy the entire reservation so we free what
429 * we've allocated so far.
431 spin_lock(&hugetlb_lock);
436 list_add(&page->lru, &surplus_list);
441 * After retaking hugetlb_lock, we need to recalculate 'needed'
442 * because either resv_huge_pages or free_huge_pages may have changed.
444 spin_lock(&hugetlb_lock);
445 needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
450 * The surplus_list now contains _at_least_ the number of extra pages
451 * needed to accomodate the reservation. Add the appropriate number
452 * of pages to the hugetlb pool and free the extras back to the buddy
453 * allocator. Commit the entire reservation here to prevent another
454 * process from stealing the pages as they are added to the pool but
455 * before they are reserved.
458 resv_huge_pages += delta;
461 /* Free the needed pages to the hugetlb pool */
462 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
465 list_del(&page->lru);
466 enqueue_huge_page(page);
469 /* Free unnecessary surplus pages to the buddy allocator */
470 if (!list_empty(&surplus_list)) {
471 spin_unlock(&hugetlb_lock);
472 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
473 list_del(&page->lru);
475 * The page has a reference count of zero already, so
476 * call free_huge_page directly instead of using
477 * put_page. This must be done with hugetlb_lock
478 * unlocked which is safe because free_huge_page takes
479 * hugetlb_lock before deciding how to free the page.
481 free_huge_page(page);
483 spin_lock(&hugetlb_lock);
490 * When releasing a hugetlb pool reservation, any surplus pages that were
491 * allocated to satisfy the reservation must be explicitly freed if they were
494 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
498 unsigned long nr_pages;
501 * We want to release as many surplus pages as possible, spread
502 * evenly across all nodes. Iterate across all nodes until we
503 * can no longer free unreserved surplus pages. This occurs when
504 * the nodes with surplus pages have no free pages.
506 unsigned long remaining_iterations = num_online_nodes();
508 /* Uncommit the reservation */
509 resv_huge_pages -= unused_resv_pages;
511 nr_pages = min(unused_resv_pages, surplus_huge_pages);
513 while (remaining_iterations-- && nr_pages) {
514 nid = next_node(nid, node_online_map);
515 if (nid == MAX_NUMNODES)
516 nid = first_node(node_online_map);
518 if (!surplus_huge_pages_node[nid])
521 if (!list_empty(&hugepage_freelists[nid])) {
522 page = list_entry(hugepage_freelists[nid].next,
524 list_del(&page->lru);
525 update_and_free_page(page);
527 free_huge_pages_node[nid]--;
528 surplus_huge_pages--;
529 surplus_huge_pages_node[nid]--;
531 remaining_iterations = num_online_nodes();
536 static struct page *alloc_huge_page(struct vm_area_struct *vma,
540 struct address_space *mapping = vma->vm_file->f_mapping;
541 struct inode *inode = mapping->host;
542 unsigned int chg = 0;
545 * Processes that did not create the mapping will have no reserves and
546 * will not have accounted against quota. Check that the quota can be
547 * made before satisfying the allocation
549 if (!vma_has_private_reserves(vma)) {
551 if (hugetlb_get_quota(inode->i_mapping, chg))
552 return ERR_PTR(-ENOSPC);
555 spin_lock(&hugetlb_lock);
556 page = dequeue_huge_page_vma(vma, addr);
557 spin_unlock(&hugetlb_lock);
560 page = alloc_buddy_huge_page(vma, addr);
562 hugetlb_put_quota(inode->i_mapping, chg);
563 return ERR_PTR(-VM_FAULT_OOM);
567 set_page_refcounted(page);
568 set_page_private(page, (unsigned long) mapping);
573 static int __init hugetlb_init(void)
577 if (HPAGE_SHIFT == 0)
580 for (i = 0; i < MAX_NUMNODES; ++i)
581 INIT_LIST_HEAD(&hugepage_freelists[i]);
583 hugetlb_next_nid = first_node(node_online_map);
585 for (i = 0; i < max_huge_pages; ++i) {
586 if (!alloc_fresh_huge_page())
589 max_huge_pages = free_huge_pages = nr_huge_pages = i;
590 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
593 module_init(hugetlb_init);
595 static int __init hugetlb_setup(char *s)
597 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
601 __setup("hugepages=", hugetlb_setup);
603 static unsigned int cpuset_mems_nr(unsigned int *array)
608 for_each_node_mask(node, cpuset_current_mems_allowed)
615 #ifdef CONFIG_HIGHMEM
616 static void try_to_free_low(unsigned long count)
620 for (i = 0; i < MAX_NUMNODES; ++i) {
621 struct page *page, *next;
622 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
623 if (count >= nr_huge_pages)
625 if (PageHighMem(page))
627 list_del(&page->lru);
628 update_and_free_page(page);
630 free_huge_pages_node[page_to_nid(page)]--;
635 static inline void try_to_free_low(unsigned long count)
640 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
641 static unsigned long set_max_huge_pages(unsigned long count)
643 unsigned long min_count, ret;
646 * Increase the pool size
647 * First take pages out of surplus state. Then make up the
648 * remaining difference by allocating fresh huge pages.
650 * We might race with alloc_buddy_huge_page() here and be unable
651 * to convert a surplus huge page to a normal huge page. That is
652 * not critical, though, it just means the overall size of the
653 * pool might be one hugepage larger than it needs to be, but
654 * within all the constraints specified by the sysctls.
656 spin_lock(&hugetlb_lock);
657 while (surplus_huge_pages && count > persistent_huge_pages) {
658 if (!adjust_pool_surplus(-1))
662 while (count > persistent_huge_pages) {
664 * If this allocation races such that we no longer need the
665 * page, free_huge_page will handle it by freeing the page
666 * and reducing the surplus.
668 spin_unlock(&hugetlb_lock);
669 ret = alloc_fresh_huge_page();
670 spin_lock(&hugetlb_lock);
677 * Decrease the pool size
678 * First return free pages to the buddy allocator (being careful
679 * to keep enough around to satisfy reservations). Then place
680 * pages into surplus state as needed so the pool will shrink
681 * to the desired size as pages become free.
683 * By placing pages into the surplus state independent of the
684 * overcommit value, we are allowing the surplus pool size to
685 * exceed overcommit. There are few sane options here. Since
686 * alloc_buddy_huge_page() is checking the global counter,
687 * though, we'll note that we're not allowed to exceed surplus
688 * and won't grow the pool anywhere else. Not until one of the
689 * sysctls are changed, or the surplus pages go out of use.
691 min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
692 min_count = max(count, min_count);
693 try_to_free_low(min_count);
694 while (min_count < persistent_huge_pages) {
695 struct page *page = dequeue_huge_page();
698 update_and_free_page(page);
700 while (count < persistent_huge_pages) {
701 if (!adjust_pool_surplus(1))
705 ret = persistent_huge_pages;
706 spin_unlock(&hugetlb_lock);
710 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
711 struct file *file, void __user *buffer,
712 size_t *length, loff_t *ppos)
714 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
715 max_huge_pages = set_max_huge_pages(max_huge_pages);
719 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
720 struct file *file, void __user *buffer,
721 size_t *length, loff_t *ppos)
723 proc_dointvec(table, write, file, buffer, length, ppos);
724 if (hugepages_treat_as_movable)
725 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
727 htlb_alloc_mask = GFP_HIGHUSER;
731 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
732 struct file *file, void __user *buffer,
733 size_t *length, loff_t *ppos)
735 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
736 spin_lock(&hugetlb_lock);
737 nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
738 spin_unlock(&hugetlb_lock);
742 #endif /* CONFIG_SYSCTL */
744 int hugetlb_report_meminfo(char *buf)
747 "HugePages_Total: %5lu\n"
748 "HugePages_Free: %5lu\n"
749 "HugePages_Rsvd: %5lu\n"
750 "HugePages_Surp: %5lu\n"
751 "Hugepagesize: %5lu kB\n",
759 int hugetlb_report_node_meminfo(int nid, char *buf)
762 "Node %d HugePages_Total: %5u\n"
763 "Node %d HugePages_Free: %5u\n"
764 "Node %d HugePages_Surp: %5u\n",
765 nid, nr_huge_pages_node[nid],
766 nid, free_huge_pages_node[nid],
767 nid, surplus_huge_pages_node[nid]);
770 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
771 unsigned long hugetlb_total_pages(void)
773 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
776 static int hugetlb_acct_memory(long delta)
780 spin_lock(&hugetlb_lock);
782 * When cpuset is configured, it breaks the strict hugetlb page
783 * reservation as the accounting is done on a global variable. Such
784 * reservation is completely rubbish in the presence of cpuset because
785 * the reservation is not checked against page availability for the
786 * current cpuset. Application can still potentially OOM'ed by kernel
787 * with lack of free htlb page in cpuset that the task is in.
788 * Attempt to enforce strict accounting with cpuset is almost
789 * impossible (or too ugly) because cpuset is too fluid that
790 * task or memory node can be dynamically moved between cpusets.
792 * The change of semantics for shared hugetlb mapping with cpuset is
793 * undesirable. However, in order to preserve some of the semantics,
794 * we fall back to check against current free page availability as
795 * a best attempt and hopefully to minimize the impact of changing
796 * semantics that cpuset has.
799 if (gather_surplus_pages(delta) < 0)
802 if (delta > cpuset_mems_nr(free_huge_pages_node)) {
803 return_unused_surplus_pages(delta);
810 return_unused_surplus_pages((unsigned long) -delta);
813 spin_unlock(&hugetlb_lock);
817 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
819 unsigned long reserve = vma_resv_huge_pages(vma);
821 hugetlb_acct_memory(-reserve);
825 * We cannot handle pagefaults against hugetlb pages at all. They cause
826 * handle_mm_fault() to try to instantiate regular-sized pages in the
827 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
830 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
836 struct vm_operations_struct hugetlb_vm_ops = {
837 .fault = hugetlb_vm_op_fault,
838 .close = hugetlb_vm_op_close,
841 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
848 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
850 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
852 entry = pte_mkyoung(entry);
853 entry = pte_mkhuge(entry);
858 static void set_huge_ptep_writable(struct vm_area_struct *vma,
859 unsigned long address, pte_t *ptep)
863 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
864 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
865 update_mmu_cache(vma, address, entry);
870 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
871 struct vm_area_struct *vma)
873 pte_t *src_pte, *dst_pte, entry;
874 struct page *ptepage;
878 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
880 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
881 src_pte = huge_pte_offset(src, addr);
884 dst_pte = huge_pte_alloc(dst, addr);
888 /* If the pagetables are shared don't copy or take references */
889 if (dst_pte == src_pte)
892 spin_lock(&dst->page_table_lock);
893 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
894 if (!huge_pte_none(huge_ptep_get(src_pte))) {
896 huge_ptep_set_wrprotect(src, addr, src_pte);
897 entry = huge_ptep_get(src_pte);
898 ptepage = pte_page(entry);
900 set_huge_pte_at(dst, addr, dst_pte, entry);
902 spin_unlock(&src->page_table_lock);
903 spin_unlock(&dst->page_table_lock);
911 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
914 struct mm_struct *mm = vma->vm_mm;
915 unsigned long address;
921 * A page gathering list, protected by per file i_mmap_lock. The
922 * lock is used to avoid list corruption from multiple unmapping
923 * of the same page since we are using page->lru.
925 LIST_HEAD(page_list);
927 WARN_ON(!is_vm_hugetlb_page(vma));
928 BUG_ON(start & ~HPAGE_MASK);
929 BUG_ON(end & ~HPAGE_MASK);
931 spin_lock(&mm->page_table_lock);
932 for (address = start; address < end; address += HPAGE_SIZE) {
933 ptep = huge_pte_offset(mm, address);
937 if (huge_pmd_unshare(mm, &address, ptep))
940 pte = huge_ptep_get_and_clear(mm, address, ptep);
941 if (huge_pte_none(pte))
944 page = pte_page(pte);
946 set_page_dirty(page);
947 list_add(&page->lru, &page_list);
949 spin_unlock(&mm->page_table_lock);
950 flush_tlb_range(vma, start, end);
951 list_for_each_entry_safe(page, tmp, &page_list, lru) {
952 list_del(&page->lru);
957 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
961 * It is undesirable to test vma->vm_file as it should be non-null
962 * for valid hugetlb area. However, vm_file will be NULL in the error
963 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
964 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
965 * to clean up. Since no pte has actually been setup, it is safe to
966 * do nothing in this case.
969 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
970 __unmap_hugepage_range(vma, start, end);
971 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
975 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
976 unsigned long address, pte_t *ptep, pte_t pte)
978 struct page *old_page, *new_page;
981 old_page = pte_page(pte);
983 /* If no-one else is actually using this page, avoid the copy
984 * and just make the page writable */
985 avoidcopy = (page_count(old_page) == 1);
987 set_huge_ptep_writable(vma, address, ptep);
991 page_cache_get(old_page);
992 new_page = alloc_huge_page(vma, address);
994 if (IS_ERR(new_page)) {
995 page_cache_release(old_page);
996 return -PTR_ERR(new_page);
999 spin_unlock(&mm->page_table_lock);
1000 copy_huge_page(new_page, old_page, address, vma);
1001 __SetPageUptodate(new_page);
1002 spin_lock(&mm->page_table_lock);
1004 ptep = huge_pte_offset(mm, address & HPAGE_MASK);
1005 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1007 huge_ptep_clear_flush(vma, address, ptep);
1008 set_huge_pte_at(mm, address, ptep,
1009 make_huge_pte(vma, new_page, 1));
1010 /* Make the old page be freed below */
1011 new_page = old_page;
1013 page_cache_release(new_page);
1014 page_cache_release(old_page);
1018 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1019 unsigned long address, pte_t *ptep, int write_access)
1021 int ret = VM_FAULT_SIGBUS;
1025 struct address_space *mapping;
1028 mapping = vma->vm_file->f_mapping;
1029 idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
1030 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
1033 * Use page lock to guard against racing truncation
1034 * before we get page_table_lock.
1037 page = find_lock_page(mapping, idx);
1039 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
1042 page = alloc_huge_page(vma, address);
1044 ret = -PTR_ERR(page);
1047 clear_huge_page(page, address);
1048 __SetPageUptodate(page);
1050 if (vma->vm_flags & VM_SHARED) {
1052 struct inode *inode = mapping->host;
1054 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1062 spin_lock(&inode->i_lock);
1063 inode->i_blocks += BLOCKS_PER_HUGEPAGE;
1064 spin_unlock(&inode->i_lock);
1069 spin_lock(&mm->page_table_lock);
1070 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
1075 if (!huge_pte_none(huge_ptep_get(ptep)))
1078 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
1079 && (vma->vm_flags & VM_SHARED)));
1080 set_huge_pte_at(mm, address, ptep, new_pte);
1082 if (write_access && !(vma->vm_flags & VM_SHARED)) {
1083 /* Optimization, do the COW without a second fault */
1084 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
1087 spin_unlock(&mm->page_table_lock);
1093 spin_unlock(&mm->page_table_lock);
1099 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
1100 unsigned long address, int write_access)
1105 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
1107 ptep = huge_pte_alloc(mm, address);
1109 return VM_FAULT_OOM;
1112 * Serialize hugepage allocation and instantiation, so that we don't
1113 * get spurious allocation failures if two CPUs race to instantiate
1114 * the same page in the page cache.
1116 mutex_lock(&hugetlb_instantiation_mutex);
1117 entry = huge_ptep_get(ptep);
1118 if (huge_pte_none(entry)) {
1119 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
1120 mutex_unlock(&hugetlb_instantiation_mutex);
1126 spin_lock(&mm->page_table_lock);
1127 /* Check for a racing update before calling hugetlb_cow */
1128 if (likely(pte_same(entry, huge_ptep_get(ptep))))
1129 if (write_access && !pte_write(entry))
1130 ret = hugetlb_cow(mm, vma, address, ptep, entry);
1131 spin_unlock(&mm->page_table_lock);
1132 mutex_unlock(&hugetlb_instantiation_mutex);
1137 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
1138 struct page **pages, struct vm_area_struct **vmas,
1139 unsigned long *position, int *length, int i,
1142 unsigned long pfn_offset;
1143 unsigned long vaddr = *position;
1144 int remainder = *length;
1146 spin_lock(&mm->page_table_lock);
1147 while (vaddr < vma->vm_end && remainder) {
1152 * Some archs (sparc64, sh*) have multiple pte_ts to
1153 * each hugepage. We have to make * sure we get the
1154 * first, for the page indexing below to work.
1156 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
1158 if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
1159 (write && !pte_write(huge_ptep_get(pte)))) {
1162 spin_unlock(&mm->page_table_lock);
1163 ret = hugetlb_fault(mm, vma, vaddr, write);
1164 spin_lock(&mm->page_table_lock);
1165 if (!(ret & VM_FAULT_ERROR))
1174 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1175 page = pte_page(huge_ptep_get(pte));
1179 pages[i] = page + pfn_offset;
1189 if (vaddr < vma->vm_end && remainder &&
1190 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1192 * We use pfn_offset to avoid touching the pageframes
1193 * of this compound page.
1198 spin_unlock(&mm->page_table_lock);
1199 *length = remainder;
1205 void hugetlb_change_protection(struct vm_area_struct *vma,
1206 unsigned long address, unsigned long end, pgprot_t newprot)
1208 struct mm_struct *mm = vma->vm_mm;
1209 unsigned long start = address;
1213 BUG_ON(address >= end);
1214 flush_cache_range(vma, address, end);
1216 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1217 spin_lock(&mm->page_table_lock);
1218 for (; address < end; address += HPAGE_SIZE) {
1219 ptep = huge_pte_offset(mm, address);
1222 if (huge_pmd_unshare(mm, &address, ptep))
1224 if (!huge_pte_none(huge_ptep_get(ptep))) {
1225 pte = huge_ptep_get_and_clear(mm, address, ptep);
1226 pte = pte_mkhuge(pte_modify(pte, newprot));
1227 set_huge_pte_at(mm, address, ptep, pte);
1230 spin_unlock(&mm->page_table_lock);
1231 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1233 flush_tlb_range(vma, start, end);
1236 struct file_region {
1237 struct list_head link;
1242 static long region_add(struct list_head *head, long f, long t)
1244 struct file_region *rg, *nrg, *trg;
1246 /* Locate the region we are either in or before. */
1247 list_for_each_entry(rg, head, link)
1251 /* Round our left edge to the current segment if it encloses us. */
1255 /* Check for and consume any regions we now overlap with. */
1257 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1258 if (&rg->link == head)
1263 /* If this area reaches higher then extend our area to
1264 * include it completely. If this is not the first area
1265 * which we intend to reuse, free it. */
1269 list_del(&rg->link);
1278 static long region_chg(struct list_head *head, long f, long t)
1280 struct file_region *rg, *nrg;
1283 /* Locate the region we are before or in. */
1284 list_for_each_entry(rg, head, link)
1288 /* If we are below the current region then a new region is required.
1289 * Subtle, allocate a new region at the position but make it zero
1290 * size such that we can guarantee to record the reservation. */
1291 if (&rg->link == head || t < rg->from) {
1292 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1297 INIT_LIST_HEAD(&nrg->link);
1298 list_add(&nrg->link, rg->link.prev);
1303 /* Round our left edge to the current segment if it encloses us. */
1308 /* Check for and consume any regions we now overlap with. */
1309 list_for_each_entry(rg, rg->link.prev, link) {
1310 if (&rg->link == head)
1315 /* We overlap with this area, if it extends futher than
1316 * us then we must extend ourselves. Account for its
1317 * existing reservation. */
1322 chg -= rg->to - rg->from;
1327 static long region_truncate(struct list_head *head, long end)
1329 struct file_region *rg, *trg;
1332 /* Locate the region we are either in or before. */
1333 list_for_each_entry(rg, head, link)
1336 if (&rg->link == head)
1339 /* If we are in the middle of a region then adjust it. */
1340 if (end > rg->from) {
1343 rg = list_entry(rg->link.next, typeof(*rg), link);
1346 /* Drop any remaining regions. */
1347 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1348 if (&rg->link == head)
1350 chg += rg->to - rg->from;
1351 list_del(&rg->link);
1357 int hugetlb_reserve_pages(struct inode *inode,
1359 struct vm_area_struct *vma)
1364 * Shared mappings base their reservation on the number of pages that
1365 * are already allocated on behalf of the file. Private mappings need
1366 * to reserve the full area even if read-only as mprotect() may be
1367 * called to make the mapping read-write. Assume !vma is a shm mapping
1369 if (!vma || vma->vm_flags & VM_SHARED)
1370 chg = region_chg(&inode->i_mapping->private_list, from, to);
1373 set_vma_resv_huge_pages(vma, chg);
1379 if (hugetlb_get_quota(inode->i_mapping, chg))
1381 ret = hugetlb_acct_memory(chg);
1383 hugetlb_put_quota(inode->i_mapping, chg);
1386 if (!vma || vma->vm_flags & VM_SHARED)
1387 region_add(&inode->i_mapping->private_list, from, to);
1391 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1393 long chg = region_truncate(&inode->i_mapping->private_list, offset);
1395 spin_lock(&inode->i_lock);
1396 inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1397 spin_unlock(&inode->i_lock);
1399 hugetlb_put_quota(inode->i_mapping, (chg - freed));
1400 hugetlb_acct_memory(-(chg - freed));