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);
43 static void clear_huge_page(struct page *page, unsigned long addr)
48 for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
50 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
54 static void copy_huge_page(struct page *dst, struct page *src,
55 unsigned long addr, struct vm_area_struct *vma)
60 for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
62 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
66 static void enqueue_huge_page(struct page *page)
68 int nid = page_to_nid(page);
69 list_add(&page->lru, &hugepage_freelists[nid]);
71 free_huge_pages_node[nid]++;
74 static struct page *dequeue_huge_page(void)
77 struct page *page = NULL;
79 for (nid = 0; nid < MAX_NUMNODES; ++nid) {
80 if (!list_empty(&hugepage_freelists[nid])) {
81 page = list_entry(hugepage_freelists[nid].next,
85 free_huge_pages_node[nid]--;
92 static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
93 unsigned long address)
96 struct page *page = NULL;
97 struct mempolicy *mpol;
99 struct zonelist *zonelist = huge_zonelist(vma, address,
100 htlb_alloc_mask, &mpol, &nodemask);
104 for_each_zone_zonelist_nodemask(zone, z, zonelist,
105 MAX_NR_ZONES - 1, nodemask) {
106 nid = zone_to_nid(zone);
107 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
108 !list_empty(&hugepage_freelists[nid])) {
109 page = list_entry(hugepage_freelists[nid].next,
111 list_del(&page->lru);
113 free_huge_pages_node[nid]--;
114 if (vma && vma->vm_flags & VM_MAYSHARE)
119 mpol_free(mpol); /* unref if mpol !NULL */
123 static void update_and_free_page(struct page *page)
127 nr_huge_pages_node[page_to_nid(page)]--;
128 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
129 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
130 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
131 1 << PG_private | 1<< PG_writeback);
133 set_compound_page_dtor(page, NULL);
134 set_page_refcounted(page);
135 __free_pages(page, HUGETLB_PAGE_ORDER);
138 static void free_huge_page(struct page *page)
140 int nid = page_to_nid(page);
141 struct address_space *mapping;
143 mapping = (struct address_space *) page_private(page);
144 set_page_private(page, 0);
145 BUG_ON(page_count(page));
146 INIT_LIST_HEAD(&page->lru);
148 spin_lock(&hugetlb_lock);
149 if (surplus_huge_pages_node[nid]) {
150 update_and_free_page(page);
151 surplus_huge_pages--;
152 surplus_huge_pages_node[nid]--;
154 enqueue_huge_page(page);
156 spin_unlock(&hugetlb_lock);
158 hugetlb_put_quota(mapping, 1);
162 * Increment or decrement surplus_huge_pages. Keep node-specific counters
163 * balanced by operating on them in a round-robin fashion.
164 * Returns 1 if an adjustment was made.
166 static int adjust_pool_surplus(int delta)
172 VM_BUG_ON(delta != -1 && delta != 1);
174 nid = next_node(nid, node_online_map);
175 if (nid == MAX_NUMNODES)
176 nid = first_node(node_online_map);
178 /* To shrink on this node, there must be a surplus page */
179 if (delta < 0 && !surplus_huge_pages_node[nid])
181 /* Surplus cannot exceed the total number of pages */
182 if (delta > 0 && surplus_huge_pages_node[nid] >=
183 nr_huge_pages_node[nid])
186 surplus_huge_pages += delta;
187 surplus_huge_pages_node[nid] += delta;
190 } while (nid != prev_nid);
196 static struct page *alloc_fresh_huge_page_node(int nid)
200 page = alloc_pages_node(nid,
201 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
204 set_compound_page_dtor(page, free_huge_page);
205 spin_lock(&hugetlb_lock);
207 nr_huge_pages_node[nid]++;
208 spin_unlock(&hugetlb_lock);
209 put_page(page); /* free it into the hugepage allocator */
215 static int alloc_fresh_huge_page(void)
222 start_nid = hugetlb_next_nid;
225 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
229 * Use a helper variable to find the next node and then
230 * copy it back to hugetlb_next_nid afterwards:
231 * otherwise there's a window in which a racer might
232 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
233 * But we don't need to use a spin_lock here: it really
234 * doesn't matter if occasionally a racer chooses the
235 * same nid as we do. Move nid forward in the mask even
236 * if we just successfully allocated a hugepage so that
237 * the next caller gets hugepages on the next node.
239 next_nid = next_node(hugetlb_next_nid, node_online_map);
240 if (next_nid == MAX_NUMNODES)
241 next_nid = first_node(node_online_map);
242 hugetlb_next_nid = next_nid;
243 } while (!page && hugetlb_next_nid != start_nid);
248 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
249 unsigned long address)
255 * Assume we will successfully allocate the surplus page to
256 * prevent racing processes from causing the surplus to exceed
259 * This however introduces a different race, where a process B
260 * tries to grow the static hugepage pool while alloc_pages() is
261 * called by process A. B will only examine the per-node
262 * counters in determining if surplus huge pages can be
263 * converted to normal huge pages in adjust_pool_surplus(). A
264 * won't be able to increment the per-node counter, until the
265 * lock is dropped by B, but B doesn't drop hugetlb_lock until
266 * no more huge pages can be converted from surplus to normal
267 * state (and doesn't try to convert again). Thus, we have a
268 * case where a surplus huge page exists, the pool is grown, and
269 * the surplus huge page still exists after, even though it
270 * should just have been converted to a normal huge page. This
271 * does not leak memory, though, as the hugepage will be freed
272 * once it is out of use. It also does not allow the counters to
273 * go out of whack in adjust_pool_surplus() as we don't modify
274 * the node values until we've gotten the hugepage and only the
275 * per-node value is checked there.
277 spin_lock(&hugetlb_lock);
278 if (surplus_huge_pages >= nr_overcommit_huge_pages) {
279 spin_unlock(&hugetlb_lock);
283 surplus_huge_pages++;
285 spin_unlock(&hugetlb_lock);
287 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
290 spin_lock(&hugetlb_lock);
293 * This page is now managed by the hugetlb allocator and has
294 * no users -- drop the buddy allocator's reference.
296 put_page_testzero(page);
297 VM_BUG_ON(page_count(page));
298 nid = page_to_nid(page);
299 set_compound_page_dtor(page, free_huge_page);
301 * We incremented the global counters already
303 nr_huge_pages_node[nid]++;
304 surplus_huge_pages_node[nid]++;
307 surplus_huge_pages--;
309 spin_unlock(&hugetlb_lock);
315 * Increase the hugetlb pool such that it can accomodate a reservation
318 static int gather_surplus_pages(int delta)
320 struct list_head surplus_list;
321 struct page *page, *tmp;
323 int needed, allocated;
325 needed = (resv_huge_pages + delta) - free_huge_pages;
327 resv_huge_pages += delta;
332 INIT_LIST_HEAD(&surplus_list);
336 spin_unlock(&hugetlb_lock);
337 for (i = 0; i < needed; i++) {
338 page = alloc_buddy_huge_page(NULL, 0);
341 * We were not able to allocate enough pages to
342 * satisfy the entire reservation so we free what
343 * we've allocated so far.
345 spin_lock(&hugetlb_lock);
350 list_add(&page->lru, &surplus_list);
355 * After retaking hugetlb_lock, we need to recalculate 'needed'
356 * because either resv_huge_pages or free_huge_pages may have changed.
358 spin_lock(&hugetlb_lock);
359 needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
364 * The surplus_list now contains _at_least_ the number of extra pages
365 * needed to accomodate the reservation. Add the appropriate number
366 * of pages to the hugetlb pool and free the extras back to the buddy
367 * allocator. Commit the entire reservation here to prevent another
368 * process from stealing the pages as they are added to the pool but
369 * before they are reserved.
372 resv_huge_pages += delta;
375 /* Free the needed pages to the hugetlb pool */
376 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
379 list_del(&page->lru);
380 enqueue_huge_page(page);
383 /* Free unnecessary surplus pages to the buddy allocator */
384 if (!list_empty(&surplus_list)) {
385 spin_unlock(&hugetlb_lock);
386 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
387 list_del(&page->lru);
389 * The page has a reference count of zero already, so
390 * call free_huge_page directly instead of using
391 * put_page. This must be done with hugetlb_lock
392 * unlocked which is safe because free_huge_page takes
393 * hugetlb_lock before deciding how to free the page.
395 free_huge_page(page);
397 spin_lock(&hugetlb_lock);
404 * When releasing a hugetlb pool reservation, any surplus pages that were
405 * allocated to satisfy the reservation must be explicitly freed if they were
408 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
412 unsigned long nr_pages;
415 * We want to release as many surplus pages as possible, spread
416 * evenly across all nodes. Iterate across all nodes until we
417 * can no longer free unreserved surplus pages. This occurs when
418 * the nodes with surplus pages have no free pages.
420 unsigned long remaining_iterations = num_online_nodes();
422 /* Uncommit the reservation */
423 resv_huge_pages -= unused_resv_pages;
425 nr_pages = min(unused_resv_pages, surplus_huge_pages);
427 while (remaining_iterations-- && nr_pages) {
428 nid = next_node(nid, node_online_map);
429 if (nid == MAX_NUMNODES)
430 nid = first_node(node_online_map);
432 if (!surplus_huge_pages_node[nid])
435 if (!list_empty(&hugepage_freelists[nid])) {
436 page = list_entry(hugepage_freelists[nid].next,
438 list_del(&page->lru);
439 update_and_free_page(page);
441 free_huge_pages_node[nid]--;
442 surplus_huge_pages--;
443 surplus_huge_pages_node[nid]--;
445 remaining_iterations = num_online_nodes();
451 static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
456 spin_lock(&hugetlb_lock);
457 page = dequeue_huge_page_vma(vma, addr);
458 spin_unlock(&hugetlb_lock);
459 return page ? page : ERR_PTR(-VM_FAULT_OOM);
462 static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
465 struct page *page = NULL;
467 if (hugetlb_get_quota(vma->vm_file->f_mapping, 1))
468 return ERR_PTR(-VM_FAULT_SIGBUS);
470 spin_lock(&hugetlb_lock);
471 if (free_huge_pages > resv_huge_pages)
472 page = dequeue_huge_page_vma(vma, addr);
473 spin_unlock(&hugetlb_lock);
475 page = alloc_buddy_huge_page(vma, addr);
477 hugetlb_put_quota(vma->vm_file->f_mapping, 1);
478 return ERR_PTR(-VM_FAULT_OOM);
484 static struct page *alloc_huge_page(struct vm_area_struct *vma,
488 struct address_space *mapping = vma->vm_file->f_mapping;
490 if (vma->vm_flags & VM_MAYSHARE)
491 page = alloc_huge_page_shared(vma, addr);
493 page = alloc_huge_page_private(vma, addr);
496 set_page_refcounted(page);
497 set_page_private(page, (unsigned long) mapping);
502 static int __init hugetlb_init(void)
506 if (HPAGE_SHIFT == 0)
509 for (i = 0; i < MAX_NUMNODES; ++i)
510 INIT_LIST_HEAD(&hugepage_freelists[i]);
512 hugetlb_next_nid = first_node(node_online_map);
514 for (i = 0; i < max_huge_pages; ++i) {
515 if (!alloc_fresh_huge_page())
518 max_huge_pages = free_huge_pages = nr_huge_pages = i;
519 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
522 module_init(hugetlb_init);
524 static int __init hugetlb_setup(char *s)
526 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
530 __setup("hugepages=", hugetlb_setup);
532 static unsigned int cpuset_mems_nr(unsigned int *array)
537 for_each_node_mask(node, cpuset_current_mems_allowed)
544 #ifdef CONFIG_HIGHMEM
545 static void try_to_free_low(unsigned long count)
549 for (i = 0; i < MAX_NUMNODES; ++i) {
550 struct page *page, *next;
551 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
552 if (count >= nr_huge_pages)
554 if (PageHighMem(page))
556 list_del(&page->lru);
557 update_and_free_page(page);
559 free_huge_pages_node[page_to_nid(page)]--;
564 static inline void try_to_free_low(unsigned long count)
569 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
570 static unsigned long set_max_huge_pages(unsigned long count)
572 unsigned long min_count, ret;
575 * Increase the pool size
576 * First take pages out of surplus state. Then make up the
577 * remaining difference by allocating fresh huge pages.
579 * We might race with alloc_buddy_huge_page() here and be unable
580 * to convert a surplus huge page to a normal huge page. That is
581 * not critical, though, it just means the overall size of the
582 * pool might be one hugepage larger than it needs to be, but
583 * within all the constraints specified by the sysctls.
585 spin_lock(&hugetlb_lock);
586 while (surplus_huge_pages && count > persistent_huge_pages) {
587 if (!adjust_pool_surplus(-1))
591 while (count > persistent_huge_pages) {
594 * If this allocation races such that we no longer need the
595 * page, free_huge_page will handle it by freeing the page
596 * and reducing the surplus.
598 spin_unlock(&hugetlb_lock);
599 ret = alloc_fresh_huge_page();
600 spin_lock(&hugetlb_lock);
607 * Decrease the pool size
608 * First return free pages to the buddy allocator (being careful
609 * to keep enough around to satisfy reservations). Then place
610 * pages into surplus state as needed so the pool will shrink
611 * to the desired size as pages become free.
613 * By placing pages into the surplus state independent of the
614 * overcommit value, we are allowing the surplus pool size to
615 * exceed overcommit. There are few sane options here. Since
616 * alloc_buddy_huge_page() is checking the global counter,
617 * though, we'll note that we're not allowed to exceed surplus
618 * and won't grow the pool anywhere else. Not until one of the
619 * sysctls are changed, or the surplus pages go out of use.
621 min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
622 min_count = max(count, min_count);
623 try_to_free_low(min_count);
624 while (min_count < persistent_huge_pages) {
625 struct page *page = dequeue_huge_page();
628 update_and_free_page(page);
630 while (count < persistent_huge_pages) {
631 if (!adjust_pool_surplus(1))
635 ret = persistent_huge_pages;
636 spin_unlock(&hugetlb_lock);
640 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
641 struct file *file, void __user *buffer,
642 size_t *length, loff_t *ppos)
644 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
645 max_huge_pages = set_max_huge_pages(max_huge_pages);
649 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
650 struct file *file, void __user *buffer,
651 size_t *length, loff_t *ppos)
653 proc_dointvec(table, write, file, buffer, length, ppos);
654 if (hugepages_treat_as_movable)
655 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
657 htlb_alloc_mask = GFP_HIGHUSER;
661 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
662 struct file *file, void __user *buffer,
663 size_t *length, loff_t *ppos)
665 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
666 spin_lock(&hugetlb_lock);
667 nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
668 spin_unlock(&hugetlb_lock);
672 #endif /* CONFIG_SYSCTL */
674 int hugetlb_report_meminfo(char *buf)
677 "HugePages_Total: %5lu\n"
678 "HugePages_Free: %5lu\n"
679 "HugePages_Rsvd: %5lu\n"
680 "HugePages_Surp: %5lu\n"
681 "Hugepagesize: %5lu kB\n",
689 int hugetlb_report_node_meminfo(int nid, char *buf)
692 "Node %d HugePages_Total: %5u\n"
693 "Node %d HugePages_Free: %5u\n"
694 "Node %d HugePages_Surp: %5u\n",
695 nid, nr_huge_pages_node[nid],
696 nid, free_huge_pages_node[nid],
697 nid, surplus_huge_pages_node[nid]);
700 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
701 unsigned long hugetlb_total_pages(void)
703 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
707 * We cannot handle pagefaults against hugetlb pages at all. They cause
708 * handle_mm_fault() to try to instantiate regular-sized pages in the
709 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
712 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
718 struct vm_operations_struct hugetlb_vm_ops = {
719 .fault = hugetlb_vm_op_fault,
722 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
729 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
731 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
733 entry = pte_mkyoung(entry);
734 entry = pte_mkhuge(entry);
739 static void set_huge_ptep_writable(struct vm_area_struct *vma,
740 unsigned long address, pte_t *ptep)
744 entry = pte_mkwrite(pte_mkdirty(*ptep));
745 if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
746 update_mmu_cache(vma, address, entry);
751 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
752 struct vm_area_struct *vma)
754 pte_t *src_pte, *dst_pte, entry;
755 struct page *ptepage;
759 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
761 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
762 src_pte = huge_pte_offset(src, addr);
765 dst_pte = huge_pte_alloc(dst, addr);
769 /* If the pagetables are shared don't copy or take references */
770 if (dst_pte == src_pte)
773 spin_lock(&dst->page_table_lock);
774 spin_lock(&src->page_table_lock);
775 if (!pte_none(*src_pte)) {
777 ptep_set_wrprotect(src, addr, src_pte);
779 ptepage = pte_page(entry);
781 set_huge_pte_at(dst, addr, dst_pte, entry);
783 spin_unlock(&src->page_table_lock);
784 spin_unlock(&dst->page_table_lock);
792 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
795 struct mm_struct *mm = vma->vm_mm;
796 unsigned long address;
802 * A page gathering list, protected by per file i_mmap_lock. The
803 * lock is used to avoid list corruption from multiple unmapping
804 * of the same page since we are using page->lru.
806 LIST_HEAD(page_list);
808 WARN_ON(!is_vm_hugetlb_page(vma));
809 BUG_ON(start & ~HPAGE_MASK);
810 BUG_ON(end & ~HPAGE_MASK);
812 spin_lock(&mm->page_table_lock);
813 for (address = start; address < end; address += HPAGE_SIZE) {
814 ptep = huge_pte_offset(mm, address);
818 if (huge_pmd_unshare(mm, &address, ptep))
821 pte = huge_ptep_get_and_clear(mm, address, ptep);
825 page = pte_page(pte);
827 set_page_dirty(page);
828 list_add(&page->lru, &page_list);
830 spin_unlock(&mm->page_table_lock);
831 flush_tlb_range(vma, start, end);
832 list_for_each_entry_safe(page, tmp, &page_list, lru) {
833 list_del(&page->lru);
838 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
842 * It is undesirable to test vma->vm_file as it should be non-null
843 * for valid hugetlb area. However, vm_file will be NULL in the error
844 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
845 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
846 * to clean up. Since no pte has actually been setup, it is safe to
847 * do nothing in this case.
850 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
851 __unmap_hugepage_range(vma, start, end);
852 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
856 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
857 unsigned long address, pte_t *ptep, pte_t pte)
859 struct page *old_page, *new_page;
862 old_page = pte_page(pte);
864 /* If no-one else is actually using this page, avoid the copy
865 * and just make the page writable */
866 avoidcopy = (page_count(old_page) == 1);
868 set_huge_ptep_writable(vma, address, ptep);
872 page_cache_get(old_page);
873 new_page = alloc_huge_page(vma, address);
875 if (IS_ERR(new_page)) {
876 page_cache_release(old_page);
877 return -PTR_ERR(new_page);
880 spin_unlock(&mm->page_table_lock);
881 copy_huge_page(new_page, old_page, address, vma);
882 __SetPageUptodate(new_page);
883 spin_lock(&mm->page_table_lock);
885 ptep = huge_pte_offset(mm, address & HPAGE_MASK);
886 if (likely(pte_same(*ptep, pte))) {
888 set_huge_pte_at(mm, address, ptep,
889 make_huge_pte(vma, new_page, 1));
890 /* Make the old page be freed below */
893 page_cache_release(new_page);
894 page_cache_release(old_page);
898 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
899 unsigned long address, pte_t *ptep, int write_access)
901 int ret = VM_FAULT_SIGBUS;
905 struct address_space *mapping;
908 mapping = vma->vm_file->f_mapping;
909 idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
910 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
913 * Use page lock to guard against racing truncation
914 * before we get page_table_lock.
917 page = find_lock_page(mapping, idx);
919 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
922 page = alloc_huge_page(vma, address);
924 ret = -PTR_ERR(page);
927 clear_huge_page(page, address);
928 __SetPageUptodate(page);
930 if (vma->vm_flags & VM_SHARED) {
932 struct inode *inode = mapping->host;
934 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
942 spin_lock(&inode->i_lock);
943 inode->i_blocks += BLOCKS_PER_HUGEPAGE;
944 spin_unlock(&inode->i_lock);
949 spin_lock(&mm->page_table_lock);
950 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
955 if (!pte_none(*ptep))
958 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
959 && (vma->vm_flags & VM_SHARED)));
960 set_huge_pte_at(mm, address, ptep, new_pte);
962 if (write_access && !(vma->vm_flags & VM_SHARED)) {
963 /* Optimization, do the COW without a second fault */
964 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
967 spin_unlock(&mm->page_table_lock);
973 spin_unlock(&mm->page_table_lock);
979 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
980 unsigned long address, int write_access)
985 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
987 ptep = huge_pte_alloc(mm, address);
992 * Serialize hugepage allocation and instantiation, so that we don't
993 * get spurious allocation failures if two CPUs race to instantiate
994 * the same page in the page cache.
996 mutex_lock(&hugetlb_instantiation_mutex);
998 if (pte_none(entry)) {
999 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
1000 mutex_unlock(&hugetlb_instantiation_mutex);
1006 spin_lock(&mm->page_table_lock);
1007 /* Check for a racing update before calling hugetlb_cow */
1008 if (likely(pte_same(entry, *ptep)))
1009 if (write_access && !pte_write(entry))
1010 ret = hugetlb_cow(mm, vma, address, ptep, entry);
1011 spin_unlock(&mm->page_table_lock);
1012 mutex_unlock(&hugetlb_instantiation_mutex);
1017 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
1018 struct page **pages, struct vm_area_struct **vmas,
1019 unsigned long *position, int *length, int i,
1022 unsigned long pfn_offset;
1023 unsigned long vaddr = *position;
1024 int remainder = *length;
1026 spin_lock(&mm->page_table_lock);
1027 while (vaddr < vma->vm_end && remainder) {
1032 * Some archs (sparc64, sh*) have multiple pte_ts to
1033 * each hugepage. We have to make * sure we get the
1034 * first, for the page indexing below to work.
1036 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
1038 if (!pte || pte_none(*pte) || (write && !pte_write(*pte))) {
1041 spin_unlock(&mm->page_table_lock);
1042 ret = hugetlb_fault(mm, vma, vaddr, write);
1043 spin_lock(&mm->page_table_lock);
1044 if (!(ret & VM_FAULT_ERROR))
1053 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1054 page = pte_page(*pte);
1058 pages[i] = page + pfn_offset;
1068 if (vaddr < vma->vm_end && remainder &&
1069 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1071 * We use pfn_offset to avoid touching the pageframes
1072 * of this compound page.
1077 spin_unlock(&mm->page_table_lock);
1078 *length = remainder;
1084 void hugetlb_change_protection(struct vm_area_struct *vma,
1085 unsigned long address, unsigned long end, pgprot_t newprot)
1087 struct mm_struct *mm = vma->vm_mm;
1088 unsigned long start = address;
1092 BUG_ON(address >= end);
1093 flush_cache_range(vma, address, end);
1095 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1096 spin_lock(&mm->page_table_lock);
1097 for (; address < end; address += HPAGE_SIZE) {
1098 ptep = huge_pte_offset(mm, address);
1101 if (huge_pmd_unshare(mm, &address, ptep))
1103 if (!pte_none(*ptep)) {
1104 pte = huge_ptep_get_and_clear(mm, address, ptep);
1105 pte = pte_mkhuge(pte_modify(pte, newprot));
1106 set_huge_pte_at(mm, address, ptep, pte);
1109 spin_unlock(&mm->page_table_lock);
1110 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1112 flush_tlb_range(vma, start, end);
1115 struct file_region {
1116 struct list_head link;
1121 static long region_add(struct list_head *head, long f, long t)
1123 struct file_region *rg, *nrg, *trg;
1125 /* Locate the region we are either in or before. */
1126 list_for_each_entry(rg, head, link)
1130 /* Round our left edge to the current segment if it encloses us. */
1134 /* Check for and consume any regions we now overlap with. */
1136 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1137 if (&rg->link == head)
1142 /* If this area reaches higher then extend our area to
1143 * include it completely. If this is not the first area
1144 * which we intend to reuse, free it. */
1148 list_del(&rg->link);
1157 static long region_chg(struct list_head *head, long f, long t)
1159 struct file_region *rg, *nrg;
1162 /* Locate the region we are before or in. */
1163 list_for_each_entry(rg, head, link)
1167 /* If we are below the current region then a new region is required.
1168 * Subtle, allocate a new region at the position but make it zero
1169 * size such that we can guarantee to record the reservation. */
1170 if (&rg->link == head || t < rg->from) {
1171 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1176 INIT_LIST_HEAD(&nrg->link);
1177 list_add(&nrg->link, rg->link.prev);
1182 /* Round our left edge to the current segment if it encloses us. */
1187 /* Check for and consume any regions we now overlap with. */
1188 list_for_each_entry(rg, rg->link.prev, link) {
1189 if (&rg->link == head)
1194 /* We overlap with this area, if it extends futher than
1195 * us then we must extend ourselves. Account for its
1196 * existing reservation. */
1201 chg -= rg->to - rg->from;
1206 static long region_truncate(struct list_head *head, long end)
1208 struct file_region *rg, *trg;
1211 /* Locate the region we are either in or before. */
1212 list_for_each_entry(rg, head, link)
1215 if (&rg->link == head)
1218 /* If we are in the middle of a region then adjust it. */
1219 if (end > rg->from) {
1222 rg = list_entry(rg->link.next, typeof(*rg), link);
1225 /* Drop any remaining regions. */
1226 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1227 if (&rg->link == head)
1229 chg += rg->to - rg->from;
1230 list_del(&rg->link);
1236 static int hugetlb_acct_memory(long delta)
1240 spin_lock(&hugetlb_lock);
1242 * When cpuset is configured, it breaks the strict hugetlb page
1243 * reservation as the accounting is done on a global variable. Such
1244 * reservation is completely rubbish in the presence of cpuset because
1245 * the reservation is not checked against page availability for the
1246 * current cpuset. Application can still potentially OOM'ed by kernel
1247 * with lack of free htlb page in cpuset that the task is in.
1248 * Attempt to enforce strict accounting with cpuset is almost
1249 * impossible (or too ugly) because cpuset is too fluid that
1250 * task or memory node can be dynamically moved between cpusets.
1252 * The change of semantics for shared hugetlb mapping with cpuset is
1253 * undesirable. However, in order to preserve some of the semantics,
1254 * we fall back to check against current free page availability as
1255 * a best attempt and hopefully to minimize the impact of changing
1256 * semantics that cpuset has.
1259 if (gather_surplus_pages(delta) < 0)
1262 if (delta > cpuset_mems_nr(free_huge_pages_node)) {
1263 return_unused_surplus_pages(delta);
1270 return_unused_surplus_pages((unsigned long) -delta);
1273 spin_unlock(&hugetlb_lock);
1277 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1281 chg = region_chg(&inode->i_mapping->private_list, from, to);
1285 if (hugetlb_get_quota(inode->i_mapping, chg))
1287 ret = hugetlb_acct_memory(chg);
1289 hugetlb_put_quota(inode->i_mapping, chg);
1292 region_add(&inode->i_mapping->private_list, from, to);
1296 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1298 long chg = region_truncate(&inode->i_mapping->private_list, offset);
1300 spin_lock(&inode->i_lock);
1301 inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1302 spin_unlock(&inode->i_lock);
1304 hugetlb_put_quota(inode->i_mapping, (chg - freed));
1305 hugetlb_acct_memory(-(chg - freed));