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 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
376 list_del(&page
->lru
);
378 enqueue_huge_page(page
);
381 * The page has a reference count of zero already, so
382 * call free_huge_page directly instead of using
383 * put_page. This must be done with hugetlb_lock
384 * unlocked which is safe because free_huge_page takes
385 * hugetlb_lock before deciding how to free the page.
387 spin_unlock(&hugetlb_lock
);
388 free_huge_page(page
);
389 spin_lock(&hugetlb_lock
);
397 * When releasing a hugetlb pool reservation, any surplus pages that were
398 * allocated to satisfy the reservation must be explicitly freed if they were
401 static void return_unused_surplus_pages(unsigned long unused_resv_pages
)
405 unsigned long nr_pages
;
408 * We want to release as many surplus pages as possible, spread
409 * evenly across all nodes. Iterate across all nodes until we
410 * can no longer free unreserved surplus pages. This occurs when
411 * the nodes with surplus pages have no free pages.
413 unsigned long remaining_iterations
= num_online_nodes();
415 /* Uncommit the reservation */
416 resv_huge_pages
-= unused_resv_pages
;
418 nr_pages
= min(unused_resv_pages
, surplus_huge_pages
);
420 while (remaining_iterations
-- && nr_pages
) {
421 nid
= next_node(nid
, node_online_map
);
422 if (nid
== MAX_NUMNODES
)
423 nid
= first_node(node_online_map
);
425 if (!surplus_huge_pages_node
[nid
])
428 if (!list_empty(&hugepage_freelists
[nid
])) {
429 page
= list_entry(hugepage_freelists
[nid
].next
,
431 list_del(&page
->lru
);
432 update_and_free_page(page
);
434 free_huge_pages_node
[nid
]--;
435 surplus_huge_pages
--;
436 surplus_huge_pages_node
[nid
]--;
438 remaining_iterations
= num_online_nodes();
444 static struct page
*alloc_huge_page_shared(struct vm_area_struct
*vma
,
449 spin_lock(&hugetlb_lock
);
450 page
= dequeue_huge_page_vma(vma
, addr
);
451 spin_unlock(&hugetlb_lock
);
452 return page
? page
: ERR_PTR(-VM_FAULT_OOM
);
455 static struct page
*alloc_huge_page_private(struct vm_area_struct
*vma
,
458 struct page
*page
= NULL
;
460 if (hugetlb_get_quota(vma
->vm_file
->f_mapping
, 1))
461 return ERR_PTR(-VM_FAULT_SIGBUS
);
463 spin_lock(&hugetlb_lock
);
464 if (free_huge_pages
> resv_huge_pages
)
465 page
= dequeue_huge_page_vma(vma
, addr
);
466 spin_unlock(&hugetlb_lock
);
468 page
= alloc_buddy_huge_page(vma
, addr
);
470 hugetlb_put_quota(vma
->vm_file
->f_mapping
, 1);
471 return ERR_PTR(-VM_FAULT_OOM
);
477 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
481 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
483 if (vma
->vm_flags
& VM_MAYSHARE
)
484 page
= alloc_huge_page_shared(vma
, addr
);
486 page
= alloc_huge_page_private(vma
, addr
);
489 set_page_refcounted(page
);
490 set_page_private(page
, (unsigned long) mapping
);
495 static int __init
hugetlb_init(void)
499 if (HPAGE_SHIFT
== 0)
502 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
503 INIT_LIST_HEAD(&hugepage_freelists
[i
]);
505 hugetlb_next_nid
= first_node(node_online_map
);
507 for (i
= 0; i
< max_huge_pages
; ++i
) {
508 if (!alloc_fresh_huge_page())
511 max_huge_pages
= free_huge_pages
= nr_huge_pages
= i
;
512 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages
);
515 module_init(hugetlb_init
);
517 static int __init
hugetlb_setup(char *s
)
519 if (sscanf(s
, "%lu", &max_huge_pages
) <= 0)
523 __setup("hugepages=", hugetlb_setup
);
525 static unsigned int cpuset_mems_nr(unsigned int *array
)
530 for_each_node_mask(node
, cpuset_current_mems_allowed
)
537 #ifdef CONFIG_HIGHMEM
538 static void try_to_free_low(unsigned long count
)
542 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
543 struct page
*page
, *next
;
544 list_for_each_entry_safe(page
, next
, &hugepage_freelists
[i
], lru
) {
545 if (count
>= nr_huge_pages
)
547 if (PageHighMem(page
))
549 list_del(&page
->lru
);
550 update_and_free_page(page
);
552 free_huge_pages_node
[page_to_nid(page
)]--;
557 static inline void try_to_free_low(unsigned long count
)
562 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
563 static unsigned long set_max_huge_pages(unsigned long count
)
565 unsigned long min_count
, ret
;
568 * Increase the pool size
569 * First take pages out of surplus state. Then make up the
570 * remaining difference by allocating fresh huge pages.
572 * We might race with alloc_buddy_huge_page() here and be unable
573 * to convert a surplus huge page to a normal huge page. That is
574 * not critical, though, it just means the overall size of the
575 * pool might be one hugepage larger than it needs to be, but
576 * within all the constraints specified by the sysctls.
578 spin_lock(&hugetlb_lock
);
579 while (surplus_huge_pages
&& count
> persistent_huge_pages
) {
580 if (!adjust_pool_surplus(-1))
584 while (count
> persistent_huge_pages
) {
587 * If this allocation races such that we no longer need the
588 * page, free_huge_page will handle it by freeing the page
589 * and reducing the surplus.
591 spin_unlock(&hugetlb_lock
);
592 ret
= alloc_fresh_huge_page();
593 spin_lock(&hugetlb_lock
);
600 * Decrease the pool size
601 * First return free pages to the buddy allocator (being careful
602 * to keep enough around to satisfy reservations). Then place
603 * pages into surplus state as needed so the pool will shrink
604 * to the desired size as pages become free.
606 * By placing pages into the surplus state independent of the
607 * overcommit value, we are allowing the surplus pool size to
608 * exceed overcommit. There are few sane options here. Since
609 * alloc_buddy_huge_page() is checking the global counter,
610 * though, we'll note that we're not allowed to exceed surplus
611 * and won't grow the pool anywhere else. Not until one of the
612 * sysctls are changed, or the surplus pages go out of use.
614 min_count
= resv_huge_pages
+ nr_huge_pages
- free_huge_pages
;
615 min_count
= max(count
, min_count
);
616 try_to_free_low(min_count
);
617 while (min_count
< persistent_huge_pages
) {
618 struct page
*page
= dequeue_huge_page();
621 update_and_free_page(page
);
623 while (count
< persistent_huge_pages
) {
624 if (!adjust_pool_surplus(1))
628 ret
= persistent_huge_pages
;
629 spin_unlock(&hugetlb_lock
);
633 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
634 struct file
*file
, void __user
*buffer
,
635 size_t *length
, loff_t
*ppos
)
637 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
638 max_huge_pages
= set_max_huge_pages(max_huge_pages
);
642 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
643 struct file
*file
, void __user
*buffer
,
644 size_t *length
, loff_t
*ppos
)
646 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
647 if (hugepages_treat_as_movable
)
648 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
650 htlb_alloc_mask
= GFP_HIGHUSER
;
654 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
655 struct file
*file
, void __user
*buffer
,
656 size_t *length
, loff_t
*ppos
)
658 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
659 spin_lock(&hugetlb_lock
);
660 nr_overcommit_huge_pages
= sysctl_overcommit_huge_pages
;
661 spin_unlock(&hugetlb_lock
);
665 #endif /* CONFIG_SYSCTL */
667 int hugetlb_report_meminfo(char *buf
)
670 "HugePages_Total: %5lu\n"
671 "HugePages_Free: %5lu\n"
672 "HugePages_Rsvd: %5lu\n"
673 "HugePages_Surp: %5lu\n"
674 "Hugepagesize: %5lu kB\n",
682 int hugetlb_report_node_meminfo(int nid
, char *buf
)
685 "Node %d HugePages_Total: %5u\n"
686 "Node %d HugePages_Free: %5u\n"
687 "Node %d HugePages_Surp: %5u\n",
688 nid
, nr_huge_pages_node
[nid
],
689 nid
, free_huge_pages_node
[nid
],
690 nid
, surplus_huge_pages_node
[nid
]);
693 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
694 unsigned long hugetlb_total_pages(void)
696 return nr_huge_pages
* (HPAGE_SIZE
/ PAGE_SIZE
);
700 * We cannot handle pagefaults against hugetlb pages at all. They cause
701 * handle_mm_fault() to try to instantiate regular-sized pages in the
702 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
705 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
711 struct vm_operations_struct hugetlb_vm_ops
= {
712 .fault
= hugetlb_vm_op_fault
,
715 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
722 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
724 entry
= pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
726 entry
= pte_mkyoung(entry
);
727 entry
= pte_mkhuge(entry
);
732 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
733 unsigned long address
, pte_t
*ptep
)
737 entry
= pte_mkwrite(pte_mkdirty(*ptep
));
738 if (ptep_set_access_flags(vma
, address
, ptep
, entry
, 1)) {
739 update_mmu_cache(vma
, address
, entry
);
744 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
745 struct vm_area_struct
*vma
)
747 pte_t
*src_pte
, *dst_pte
, entry
;
748 struct page
*ptepage
;
752 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
754 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= HPAGE_SIZE
) {
755 src_pte
= huge_pte_offset(src
, addr
);
758 dst_pte
= huge_pte_alloc(dst
, addr
);
762 /* If the pagetables are shared don't copy or take references */
763 if (dst_pte
== src_pte
)
766 spin_lock(&dst
->page_table_lock
);
767 spin_lock(&src
->page_table_lock
);
768 if (!pte_none(*src_pte
)) {
770 ptep_set_wrprotect(src
, addr
, src_pte
);
772 ptepage
= pte_page(entry
);
774 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
776 spin_unlock(&src
->page_table_lock
);
777 spin_unlock(&dst
->page_table_lock
);
785 void __unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
788 struct mm_struct
*mm
= vma
->vm_mm
;
789 unsigned long address
;
795 * A page gathering list, protected by per file i_mmap_lock. The
796 * lock is used to avoid list corruption from multiple unmapping
797 * of the same page since we are using page->lru.
799 LIST_HEAD(page_list
);
801 WARN_ON(!is_vm_hugetlb_page(vma
));
802 BUG_ON(start
& ~HPAGE_MASK
);
803 BUG_ON(end
& ~HPAGE_MASK
);
805 spin_lock(&mm
->page_table_lock
);
806 for (address
= start
; address
< end
; address
+= HPAGE_SIZE
) {
807 ptep
= huge_pte_offset(mm
, address
);
811 if (huge_pmd_unshare(mm
, &address
, ptep
))
814 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
818 page
= pte_page(pte
);
820 set_page_dirty(page
);
821 list_add(&page
->lru
, &page_list
);
823 spin_unlock(&mm
->page_table_lock
);
824 flush_tlb_range(vma
, start
, end
);
825 list_for_each_entry_safe(page
, tmp
, &page_list
, lru
) {
826 list_del(&page
->lru
);
831 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
835 * It is undesirable to test vma->vm_file as it should be non-null
836 * for valid hugetlb area. However, vm_file will be NULL in the error
837 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
838 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
839 * to clean up. Since no pte has actually been setup, it is safe to
840 * do nothing in this case.
843 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
844 __unmap_hugepage_range(vma
, start
, end
);
845 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
849 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
850 unsigned long address
, pte_t
*ptep
, pte_t pte
)
852 struct page
*old_page
, *new_page
;
855 old_page
= pte_page(pte
);
857 /* If no-one else is actually using this page, avoid the copy
858 * and just make the page writable */
859 avoidcopy
= (page_count(old_page
) == 1);
861 set_huge_ptep_writable(vma
, address
, ptep
);
865 page_cache_get(old_page
);
866 new_page
= alloc_huge_page(vma
, address
);
868 if (IS_ERR(new_page
)) {
869 page_cache_release(old_page
);
870 return -PTR_ERR(new_page
);
873 spin_unlock(&mm
->page_table_lock
);
874 copy_huge_page(new_page
, old_page
, address
, vma
);
875 __SetPageUptodate(new_page
);
876 spin_lock(&mm
->page_table_lock
);
878 ptep
= huge_pte_offset(mm
, address
& HPAGE_MASK
);
879 if (likely(pte_same(*ptep
, pte
))) {
881 set_huge_pte_at(mm
, address
, ptep
,
882 make_huge_pte(vma
, new_page
, 1));
883 /* Make the old page be freed below */
886 page_cache_release(new_page
);
887 page_cache_release(old_page
);
891 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
892 unsigned long address
, pte_t
*ptep
, int write_access
)
894 int ret
= VM_FAULT_SIGBUS
;
898 struct address_space
*mapping
;
901 mapping
= vma
->vm_file
->f_mapping
;
902 idx
= ((address
- vma
->vm_start
) >> HPAGE_SHIFT
)
903 + (vma
->vm_pgoff
>> (HPAGE_SHIFT
- PAGE_SHIFT
));
906 * Use page lock to guard against racing truncation
907 * before we get page_table_lock.
910 page
= find_lock_page(mapping
, idx
);
912 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
915 page
= alloc_huge_page(vma
, address
);
917 ret
= -PTR_ERR(page
);
920 clear_huge_page(page
, address
);
921 __SetPageUptodate(page
);
923 if (vma
->vm_flags
& VM_SHARED
) {
925 struct inode
*inode
= mapping
->host
;
927 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
935 spin_lock(&inode
->i_lock
);
936 inode
->i_blocks
+= BLOCKS_PER_HUGEPAGE
;
937 spin_unlock(&inode
->i_lock
);
942 spin_lock(&mm
->page_table_lock
);
943 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
948 if (!pte_none(*ptep
))
951 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
952 && (vma
->vm_flags
& VM_SHARED
)));
953 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
955 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
956 /* Optimization, do the COW without a second fault */
957 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
);
960 spin_unlock(&mm
->page_table_lock
);
966 spin_unlock(&mm
->page_table_lock
);
972 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
973 unsigned long address
, int write_access
)
978 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
980 ptep
= huge_pte_alloc(mm
, address
);
985 * Serialize hugepage allocation and instantiation, so that we don't
986 * get spurious allocation failures if two CPUs race to instantiate
987 * the same page in the page cache.
989 mutex_lock(&hugetlb_instantiation_mutex
);
991 if (pte_none(entry
)) {
992 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
993 mutex_unlock(&hugetlb_instantiation_mutex
);
999 spin_lock(&mm
->page_table_lock
);
1000 /* Check for a racing update before calling hugetlb_cow */
1001 if (likely(pte_same(entry
, *ptep
)))
1002 if (write_access
&& !pte_write(entry
))
1003 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
);
1004 spin_unlock(&mm
->page_table_lock
);
1005 mutex_unlock(&hugetlb_instantiation_mutex
);
1010 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1011 struct page
**pages
, struct vm_area_struct
**vmas
,
1012 unsigned long *position
, int *length
, int i
,
1015 unsigned long pfn_offset
;
1016 unsigned long vaddr
= *position
;
1017 int remainder
= *length
;
1019 spin_lock(&mm
->page_table_lock
);
1020 while (vaddr
< vma
->vm_end
&& remainder
) {
1025 * Some archs (sparc64, sh*) have multiple pte_ts to
1026 * each hugepage. We have to make * sure we get the
1027 * first, for the page indexing below to work.
1029 pte
= huge_pte_offset(mm
, vaddr
& HPAGE_MASK
);
1031 if (!pte
|| pte_none(*pte
) || (write
&& !pte_write(*pte
))) {
1034 spin_unlock(&mm
->page_table_lock
);
1035 ret
= hugetlb_fault(mm
, vma
, vaddr
, write
);
1036 spin_lock(&mm
->page_table_lock
);
1037 if (!(ret
& VM_FAULT_ERROR
))
1046 pfn_offset
= (vaddr
& ~HPAGE_MASK
) >> PAGE_SHIFT
;
1047 page
= pte_page(*pte
);
1051 pages
[i
] = page
+ pfn_offset
;
1061 if (vaddr
< vma
->vm_end
&& remainder
&&
1062 pfn_offset
< HPAGE_SIZE
/PAGE_SIZE
) {
1064 * We use pfn_offset to avoid touching the pageframes
1065 * of this compound page.
1070 spin_unlock(&mm
->page_table_lock
);
1071 *length
= remainder
;
1077 void hugetlb_change_protection(struct vm_area_struct
*vma
,
1078 unsigned long address
, unsigned long end
, pgprot_t newprot
)
1080 struct mm_struct
*mm
= vma
->vm_mm
;
1081 unsigned long start
= address
;
1085 BUG_ON(address
>= end
);
1086 flush_cache_range(vma
, address
, end
);
1088 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1089 spin_lock(&mm
->page_table_lock
);
1090 for (; address
< end
; address
+= HPAGE_SIZE
) {
1091 ptep
= huge_pte_offset(mm
, address
);
1094 if (huge_pmd_unshare(mm
, &address
, ptep
))
1096 if (!pte_none(*ptep
)) {
1097 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
1098 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
1099 set_huge_pte_at(mm
, address
, ptep
, pte
);
1102 spin_unlock(&mm
->page_table_lock
);
1103 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1105 flush_tlb_range(vma
, start
, end
);
1108 struct file_region
{
1109 struct list_head link
;
1114 static long region_add(struct list_head
*head
, long f
, long t
)
1116 struct file_region
*rg
, *nrg
, *trg
;
1118 /* Locate the region we are either in or before. */
1119 list_for_each_entry(rg
, head
, link
)
1123 /* Round our left edge to the current segment if it encloses us. */
1127 /* Check for and consume any regions we now overlap with. */
1129 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
1130 if (&rg
->link
== head
)
1135 /* If this area reaches higher then extend our area to
1136 * include it completely. If this is not the first area
1137 * which we intend to reuse, free it. */
1141 list_del(&rg
->link
);
1150 static long region_chg(struct list_head
*head
, long f
, long t
)
1152 struct file_region
*rg
, *nrg
;
1155 /* Locate the region we are before or in. */
1156 list_for_each_entry(rg
, head
, link
)
1160 /* If we are below the current region then a new region is required.
1161 * Subtle, allocate a new region at the position but make it zero
1162 * size such that we can guarantee to record the reservation. */
1163 if (&rg
->link
== head
|| t
< rg
->from
) {
1164 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
1169 INIT_LIST_HEAD(&nrg
->link
);
1170 list_add(&nrg
->link
, rg
->link
.prev
);
1175 /* Round our left edge to the current segment if it encloses us. */
1180 /* Check for and consume any regions we now overlap with. */
1181 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
1182 if (&rg
->link
== head
)
1187 /* We overlap with this area, if it extends futher than
1188 * us then we must extend ourselves. Account for its
1189 * existing reservation. */
1194 chg
-= rg
->to
- rg
->from
;
1199 static long region_truncate(struct list_head
*head
, long end
)
1201 struct file_region
*rg
, *trg
;
1204 /* Locate the region we are either in or before. */
1205 list_for_each_entry(rg
, head
, link
)
1208 if (&rg
->link
== head
)
1211 /* If we are in the middle of a region then adjust it. */
1212 if (end
> rg
->from
) {
1215 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
1218 /* Drop any remaining regions. */
1219 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
1220 if (&rg
->link
== head
)
1222 chg
+= rg
->to
- rg
->from
;
1223 list_del(&rg
->link
);
1229 static int hugetlb_acct_memory(long delta
)
1233 spin_lock(&hugetlb_lock
);
1235 * When cpuset is configured, it breaks the strict hugetlb page
1236 * reservation as the accounting is done on a global variable. Such
1237 * reservation is completely rubbish in the presence of cpuset because
1238 * the reservation is not checked against page availability for the
1239 * current cpuset. Application can still potentially OOM'ed by kernel
1240 * with lack of free htlb page in cpuset that the task is in.
1241 * Attempt to enforce strict accounting with cpuset is almost
1242 * impossible (or too ugly) because cpuset is too fluid that
1243 * task or memory node can be dynamically moved between cpusets.
1245 * The change of semantics for shared hugetlb mapping with cpuset is
1246 * undesirable. However, in order to preserve some of the semantics,
1247 * we fall back to check against current free page availability as
1248 * a best attempt and hopefully to minimize the impact of changing
1249 * semantics that cpuset has.
1252 if (gather_surplus_pages(delta
) < 0)
1255 if (delta
> cpuset_mems_nr(free_huge_pages_node
)) {
1256 return_unused_surplus_pages(delta
);
1263 return_unused_surplus_pages((unsigned long) -delta
);
1266 spin_unlock(&hugetlb_lock
);
1270 int hugetlb_reserve_pages(struct inode
*inode
, long from
, long to
)
1274 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
1278 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
1280 ret
= hugetlb_acct_memory(chg
);
1282 hugetlb_put_quota(inode
->i_mapping
, chg
);
1285 region_add(&inode
->i_mapping
->private_list
, from
, to
);
1289 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
1291 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
1293 spin_lock(&inode
->i_lock
);
1294 inode
->i_blocks
-= BLOCKS_PER_HUGEPAGE
* freed
;
1295 spin_unlock(&inode
->i_lock
);
1297 hugetlb_put_quota(inode
->i_mapping
, (chg
- freed
));
1298 hugetlb_acct_memory(-(chg
- freed
));