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
)
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 arch_release_hugepage(page
);
136 __free_pages(page
, HUGETLB_PAGE_ORDER
);
139 static void free_huge_page(struct page
*page
)
141 int nid
= page_to_nid(page
);
142 struct address_space
*mapping
;
144 mapping
= (struct address_space
*) page_private(page
);
145 set_page_private(page
, 0);
146 BUG_ON(page_count(page
));
147 INIT_LIST_HEAD(&page
->lru
);
149 spin_lock(&hugetlb_lock
);
150 if (surplus_huge_pages_node
[nid
]) {
151 update_and_free_page(page
);
152 surplus_huge_pages
--;
153 surplus_huge_pages_node
[nid
]--;
155 enqueue_huge_page(page
);
157 spin_unlock(&hugetlb_lock
);
159 hugetlb_put_quota(mapping
, 1);
163 * Increment or decrement surplus_huge_pages. Keep node-specific counters
164 * balanced by operating on them in a round-robin fashion.
165 * Returns 1 if an adjustment was made.
167 static int adjust_pool_surplus(int delta
)
173 VM_BUG_ON(delta
!= -1 && delta
!= 1);
175 nid
= next_node(nid
, node_online_map
);
176 if (nid
== MAX_NUMNODES
)
177 nid
= first_node(node_online_map
);
179 /* To shrink on this node, there must be a surplus page */
180 if (delta
< 0 && !surplus_huge_pages_node
[nid
])
182 /* Surplus cannot exceed the total number of pages */
183 if (delta
> 0 && surplus_huge_pages_node
[nid
] >=
184 nr_huge_pages_node
[nid
])
187 surplus_huge_pages
+= delta
;
188 surplus_huge_pages_node
[nid
] += delta
;
191 } while (nid
!= prev_nid
);
197 static struct page
*alloc_fresh_huge_page_node(int nid
)
201 page
= alloc_pages_node(nid
,
202 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|__GFP_NOWARN
,
205 if (arch_prepare_hugepage(page
)) {
206 __free_pages(page
, HUGETLB_PAGE_ORDER
);
209 set_compound_page_dtor(page
, free_huge_page
);
210 spin_lock(&hugetlb_lock
);
212 nr_huge_pages_node
[nid
]++;
213 spin_unlock(&hugetlb_lock
);
214 put_page(page
); /* free it into the hugepage allocator */
220 static int alloc_fresh_huge_page(void)
227 start_nid
= hugetlb_next_nid
;
230 page
= alloc_fresh_huge_page_node(hugetlb_next_nid
);
234 * Use a helper variable to find the next node and then
235 * copy it back to hugetlb_next_nid afterwards:
236 * otherwise there's a window in which a racer might
237 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
238 * But we don't need to use a spin_lock here: it really
239 * doesn't matter if occasionally a racer chooses the
240 * same nid as we do. Move nid forward in the mask even
241 * if we just successfully allocated a hugepage so that
242 * the next caller gets hugepages on the next node.
244 next_nid
= next_node(hugetlb_next_nid
, node_online_map
);
245 if (next_nid
== MAX_NUMNODES
)
246 next_nid
= first_node(node_online_map
);
247 hugetlb_next_nid
= next_nid
;
248 } while (!page
&& hugetlb_next_nid
!= start_nid
);
251 count_vm_event(HTLB_BUDDY_PGALLOC
);
253 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
258 static struct page
*alloc_buddy_huge_page(struct vm_area_struct
*vma
,
259 unsigned long address
)
265 * Assume we will successfully allocate the surplus page to
266 * prevent racing processes from causing the surplus to exceed
269 * This however introduces a different race, where a process B
270 * tries to grow the static hugepage pool while alloc_pages() is
271 * called by process A. B will only examine the per-node
272 * counters in determining if surplus huge pages can be
273 * converted to normal huge pages in adjust_pool_surplus(). A
274 * won't be able to increment the per-node counter, until the
275 * lock is dropped by B, but B doesn't drop hugetlb_lock until
276 * no more huge pages can be converted from surplus to normal
277 * state (and doesn't try to convert again). Thus, we have a
278 * case where a surplus huge page exists, the pool is grown, and
279 * the surplus huge page still exists after, even though it
280 * should just have been converted to a normal huge page. This
281 * does not leak memory, though, as the hugepage will be freed
282 * once it is out of use. It also does not allow the counters to
283 * go out of whack in adjust_pool_surplus() as we don't modify
284 * the node values until we've gotten the hugepage and only the
285 * per-node value is checked there.
287 spin_lock(&hugetlb_lock
);
288 if (surplus_huge_pages
>= nr_overcommit_huge_pages
) {
289 spin_unlock(&hugetlb_lock
);
293 surplus_huge_pages
++;
295 spin_unlock(&hugetlb_lock
);
297 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|__GFP_NOWARN
,
300 spin_lock(&hugetlb_lock
);
303 * This page is now managed by the hugetlb allocator and has
304 * no users -- drop the buddy allocator's reference.
306 put_page_testzero(page
);
307 VM_BUG_ON(page_count(page
));
308 nid
= page_to_nid(page
);
309 set_compound_page_dtor(page
, free_huge_page
);
311 * We incremented the global counters already
313 nr_huge_pages_node
[nid
]++;
314 surplus_huge_pages_node
[nid
]++;
315 __count_vm_event(HTLB_BUDDY_PGALLOC
);
318 surplus_huge_pages
--;
319 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
321 spin_unlock(&hugetlb_lock
);
327 * Increase the hugetlb pool such that it can accomodate a reservation
330 static int gather_surplus_pages(int delta
)
332 struct list_head surplus_list
;
333 struct page
*page
, *tmp
;
335 int needed
, allocated
;
337 needed
= (resv_huge_pages
+ delta
) - free_huge_pages
;
339 resv_huge_pages
+= delta
;
344 INIT_LIST_HEAD(&surplus_list
);
348 spin_unlock(&hugetlb_lock
);
349 for (i
= 0; i
< needed
; i
++) {
350 page
= alloc_buddy_huge_page(NULL
, 0);
353 * We were not able to allocate enough pages to
354 * satisfy the entire reservation so we free what
355 * we've allocated so far.
357 spin_lock(&hugetlb_lock
);
362 list_add(&page
->lru
, &surplus_list
);
367 * After retaking hugetlb_lock, we need to recalculate 'needed'
368 * because either resv_huge_pages or free_huge_pages may have changed.
370 spin_lock(&hugetlb_lock
);
371 needed
= (resv_huge_pages
+ delta
) - (free_huge_pages
+ allocated
);
376 * The surplus_list now contains _at_least_ the number of extra pages
377 * needed to accomodate the reservation. Add the appropriate number
378 * of pages to the hugetlb pool and free the extras back to the buddy
379 * allocator. Commit the entire reservation here to prevent another
380 * process from stealing the pages as they are added to the pool but
381 * before they are reserved.
384 resv_huge_pages
+= delta
;
387 /* Free the needed pages to the hugetlb pool */
388 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
391 list_del(&page
->lru
);
392 enqueue_huge_page(page
);
395 /* Free unnecessary surplus pages to the buddy allocator */
396 if (!list_empty(&surplus_list
)) {
397 spin_unlock(&hugetlb_lock
);
398 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
399 list_del(&page
->lru
);
401 * The page has a reference count of zero already, so
402 * call free_huge_page directly instead of using
403 * put_page. This must be done with hugetlb_lock
404 * unlocked which is safe because free_huge_page takes
405 * hugetlb_lock before deciding how to free the page.
407 free_huge_page(page
);
409 spin_lock(&hugetlb_lock
);
416 * When releasing a hugetlb pool reservation, any surplus pages that were
417 * allocated to satisfy the reservation must be explicitly freed if they were
420 static void return_unused_surplus_pages(unsigned long unused_resv_pages
)
424 unsigned long nr_pages
;
427 * We want to release as many surplus pages as possible, spread
428 * evenly across all nodes. Iterate across all nodes until we
429 * can no longer free unreserved surplus pages. This occurs when
430 * the nodes with surplus pages have no free pages.
432 unsigned long remaining_iterations
= num_online_nodes();
434 /* Uncommit the reservation */
435 resv_huge_pages
-= unused_resv_pages
;
437 nr_pages
= min(unused_resv_pages
, surplus_huge_pages
);
439 while (remaining_iterations
-- && nr_pages
) {
440 nid
= next_node(nid
, node_online_map
);
441 if (nid
== MAX_NUMNODES
)
442 nid
= first_node(node_online_map
);
444 if (!surplus_huge_pages_node
[nid
])
447 if (!list_empty(&hugepage_freelists
[nid
])) {
448 page
= list_entry(hugepage_freelists
[nid
].next
,
450 list_del(&page
->lru
);
451 update_and_free_page(page
);
453 free_huge_pages_node
[nid
]--;
454 surplus_huge_pages
--;
455 surplus_huge_pages_node
[nid
]--;
457 remaining_iterations
= num_online_nodes();
463 static struct page
*alloc_huge_page_shared(struct vm_area_struct
*vma
,
468 spin_lock(&hugetlb_lock
);
469 page
= dequeue_huge_page_vma(vma
, addr
);
470 spin_unlock(&hugetlb_lock
);
471 return page
? page
: ERR_PTR(-VM_FAULT_OOM
);
474 static struct page
*alloc_huge_page_private(struct vm_area_struct
*vma
,
477 struct page
*page
= NULL
;
479 if (hugetlb_get_quota(vma
->vm_file
->f_mapping
, 1))
480 return ERR_PTR(-VM_FAULT_SIGBUS
);
482 spin_lock(&hugetlb_lock
);
483 if (free_huge_pages
> resv_huge_pages
)
484 page
= dequeue_huge_page_vma(vma
, addr
);
485 spin_unlock(&hugetlb_lock
);
487 page
= alloc_buddy_huge_page(vma
, addr
);
489 hugetlb_put_quota(vma
->vm_file
->f_mapping
, 1);
490 return ERR_PTR(-VM_FAULT_OOM
);
496 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
500 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
502 if (vma
->vm_flags
& VM_MAYSHARE
)
503 page
= alloc_huge_page_shared(vma
, addr
);
505 page
= alloc_huge_page_private(vma
, addr
);
508 set_page_refcounted(page
);
509 set_page_private(page
, (unsigned long) mapping
);
514 static int __init
hugetlb_init(void)
518 if (HPAGE_SHIFT
== 0)
521 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
522 INIT_LIST_HEAD(&hugepage_freelists
[i
]);
524 hugetlb_next_nid
= first_node(node_online_map
);
526 for (i
= 0; i
< max_huge_pages
; ++i
) {
527 if (!alloc_fresh_huge_page())
530 max_huge_pages
= free_huge_pages
= nr_huge_pages
= i
;
531 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages
);
534 module_init(hugetlb_init
);
536 static int __init
hugetlb_setup(char *s
)
538 if (sscanf(s
, "%lu", &max_huge_pages
) <= 0)
542 __setup("hugepages=", hugetlb_setup
);
544 static unsigned int cpuset_mems_nr(unsigned int *array
)
549 for_each_node_mask(node
, cpuset_current_mems_allowed
)
556 #ifdef CONFIG_HIGHMEM
557 static void try_to_free_low(unsigned long count
)
561 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
562 struct page
*page
, *next
;
563 list_for_each_entry_safe(page
, next
, &hugepage_freelists
[i
], lru
) {
564 if (count
>= nr_huge_pages
)
566 if (PageHighMem(page
))
568 list_del(&page
->lru
);
569 update_and_free_page(page
);
571 free_huge_pages_node
[page_to_nid(page
)]--;
576 static inline void try_to_free_low(unsigned long count
)
581 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
582 static unsigned long set_max_huge_pages(unsigned long count
)
584 unsigned long min_count
, ret
;
587 * Increase the pool size
588 * First take pages out of surplus state. Then make up the
589 * remaining difference by allocating fresh huge pages.
591 * We might race with alloc_buddy_huge_page() here and be unable
592 * to convert a surplus huge page to a normal huge page. That is
593 * not critical, though, it just means the overall size of the
594 * pool might be one hugepage larger than it needs to be, but
595 * within all the constraints specified by the sysctls.
597 spin_lock(&hugetlb_lock
);
598 while (surplus_huge_pages
&& count
> persistent_huge_pages
) {
599 if (!adjust_pool_surplus(-1))
603 while (count
> persistent_huge_pages
) {
606 * If this allocation races such that we no longer need the
607 * page, free_huge_page will handle it by freeing the page
608 * and reducing the surplus.
610 spin_unlock(&hugetlb_lock
);
611 ret
= alloc_fresh_huge_page();
612 spin_lock(&hugetlb_lock
);
619 * Decrease the pool size
620 * First return free pages to the buddy allocator (being careful
621 * to keep enough around to satisfy reservations). Then place
622 * pages into surplus state as needed so the pool will shrink
623 * to the desired size as pages become free.
625 * By placing pages into the surplus state independent of the
626 * overcommit value, we are allowing the surplus pool size to
627 * exceed overcommit. There are few sane options here. Since
628 * alloc_buddy_huge_page() is checking the global counter,
629 * though, we'll note that we're not allowed to exceed surplus
630 * and won't grow the pool anywhere else. Not until one of the
631 * sysctls are changed, or the surplus pages go out of use.
633 min_count
= resv_huge_pages
+ nr_huge_pages
- free_huge_pages
;
634 min_count
= max(count
, min_count
);
635 try_to_free_low(min_count
);
636 while (min_count
< persistent_huge_pages
) {
637 struct page
*page
= dequeue_huge_page();
640 update_and_free_page(page
);
642 while (count
< persistent_huge_pages
) {
643 if (!adjust_pool_surplus(1))
647 ret
= persistent_huge_pages
;
648 spin_unlock(&hugetlb_lock
);
652 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
653 struct file
*file
, void __user
*buffer
,
654 size_t *length
, loff_t
*ppos
)
656 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
657 max_huge_pages
= set_max_huge_pages(max_huge_pages
);
661 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
662 struct file
*file
, void __user
*buffer
,
663 size_t *length
, loff_t
*ppos
)
665 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
666 if (hugepages_treat_as_movable
)
667 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
669 htlb_alloc_mask
= GFP_HIGHUSER
;
673 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
674 struct file
*file
, void __user
*buffer
,
675 size_t *length
, loff_t
*ppos
)
677 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
678 spin_lock(&hugetlb_lock
);
679 nr_overcommit_huge_pages
= sysctl_overcommit_huge_pages
;
680 spin_unlock(&hugetlb_lock
);
684 #endif /* CONFIG_SYSCTL */
686 int hugetlb_report_meminfo(char *buf
)
689 "HugePages_Total: %5lu\n"
690 "HugePages_Free: %5lu\n"
691 "HugePages_Rsvd: %5lu\n"
692 "HugePages_Surp: %5lu\n"
693 "Hugepagesize: %5lu kB\n",
701 int hugetlb_report_node_meminfo(int nid
, char *buf
)
704 "Node %d HugePages_Total: %5u\n"
705 "Node %d HugePages_Free: %5u\n"
706 "Node %d HugePages_Surp: %5u\n",
707 nid
, nr_huge_pages_node
[nid
],
708 nid
, free_huge_pages_node
[nid
],
709 nid
, surplus_huge_pages_node
[nid
]);
712 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
713 unsigned long hugetlb_total_pages(void)
715 return nr_huge_pages
* (HPAGE_SIZE
/ PAGE_SIZE
);
719 * We cannot handle pagefaults against hugetlb pages at all. They cause
720 * handle_mm_fault() to try to instantiate regular-sized pages in the
721 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
724 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
730 struct vm_operations_struct hugetlb_vm_ops
= {
731 .fault
= hugetlb_vm_op_fault
,
734 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
741 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
743 entry
= huge_pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
745 entry
= pte_mkyoung(entry
);
746 entry
= pte_mkhuge(entry
);
751 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
752 unsigned long address
, pte_t
*ptep
)
756 entry
= pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep
)));
757 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1)) {
758 update_mmu_cache(vma
, address
, entry
);
763 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
764 struct vm_area_struct
*vma
)
766 pte_t
*src_pte
, *dst_pte
, entry
;
767 struct page
*ptepage
;
771 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
773 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= HPAGE_SIZE
) {
774 src_pte
= huge_pte_offset(src
, addr
);
777 dst_pte
= huge_pte_alloc(dst
, addr
);
781 /* If the pagetables are shared don't copy or take references */
782 if (dst_pte
== src_pte
)
785 spin_lock(&dst
->page_table_lock
);
786 spin_lock(&src
->page_table_lock
);
787 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
789 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
790 entry
= huge_ptep_get(src_pte
);
791 ptepage
= pte_page(entry
);
793 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
795 spin_unlock(&src
->page_table_lock
);
796 spin_unlock(&dst
->page_table_lock
);
804 void __unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
807 struct mm_struct
*mm
= vma
->vm_mm
;
808 unsigned long address
;
814 * A page gathering list, protected by per file i_mmap_lock. The
815 * lock is used to avoid list corruption from multiple unmapping
816 * of the same page since we are using page->lru.
818 LIST_HEAD(page_list
);
820 WARN_ON(!is_vm_hugetlb_page(vma
));
821 BUG_ON(start
& ~HPAGE_MASK
);
822 BUG_ON(end
& ~HPAGE_MASK
);
824 spin_lock(&mm
->page_table_lock
);
825 for (address
= start
; address
< end
; address
+= HPAGE_SIZE
) {
826 ptep
= huge_pte_offset(mm
, address
);
830 if (huge_pmd_unshare(mm
, &address
, ptep
))
833 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
834 if (huge_pte_none(pte
))
837 page
= pte_page(pte
);
839 set_page_dirty(page
);
840 list_add(&page
->lru
, &page_list
);
842 spin_unlock(&mm
->page_table_lock
);
843 flush_tlb_range(vma
, start
, end
);
844 list_for_each_entry_safe(page
, tmp
, &page_list
, lru
) {
845 list_del(&page
->lru
);
850 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
854 * It is undesirable to test vma->vm_file as it should be non-null
855 * for valid hugetlb area. However, vm_file will be NULL in the error
856 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
857 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
858 * to clean up. Since no pte has actually been setup, it is safe to
859 * do nothing in this case.
862 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
863 __unmap_hugepage_range(vma
, start
, end
);
864 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
868 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
869 unsigned long address
, pte_t
*ptep
, pte_t pte
)
871 struct page
*old_page
, *new_page
;
874 old_page
= pte_page(pte
);
876 /* If no-one else is actually using this page, avoid the copy
877 * and just make the page writable */
878 avoidcopy
= (page_count(old_page
) == 1);
880 set_huge_ptep_writable(vma
, address
, ptep
);
884 page_cache_get(old_page
);
885 new_page
= alloc_huge_page(vma
, address
);
887 if (IS_ERR(new_page
)) {
888 page_cache_release(old_page
);
889 return -PTR_ERR(new_page
);
892 spin_unlock(&mm
->page_table_lock
);
893 copy_huge_page(new_page
, old_page
, address
, vma
);
894 __SetPageUptodate(new_page
);
895 spin_lock(&mm
->page_table_lock
);
897 ptep
= huge_pte_offset(mm
, address
& HPAGE_MASK
);
898 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
900 huge_ptep_clear_flush(vma
, address
, ptep
);
901 set_huge_pte_at(mm
, address
, ptep
,
902 make_huge_pte(vma
, new_page
, 1));
903 /* Make the old page be freed below */
906 page_cache_release(new_page
);
907 page_cache_release(old_page
);
911 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
912 unsigned long address
, pte_t
*ptep
, int write_access
)
914 int ret
= VM_FAULT_SIGBUS
;
918 struct address_space
*mapping
;
921 mapping
= vma
->vm_file
->f_mapping
;
922 idx
= ((address
- vma
->vm_start
) >> HPAGE_SHIFT
)
923 + (vma
->vm_pgoff
>> (HPAGE_SHIFT
- PAGE_SHIFT
));
926 * Use page lock to guard against racing truncation
927 * before we get page_table_lock.
930 page
= find_lock_page(mapping
, idx
);
932 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
935 page
= alloc_huge_page(vma
, address
);
937 ret
= -PTR_ERR(page
);
940 clear_huge_page(page
, address
);
941 __SetPageUptodate(page
);
943 if (vma
->vm_flags
& VM_SHARED
) {
945 struct inode
*inode
= mapping
->host
;
947 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
955 spin_lock(&inode
->i_lock
);
956 inode
->i_blocks
+= BLOCKS_PER_HUGEPAGE
;
957 spin_unlock(&inode
->i_lock
);
962 spin_lock(&mm
->page_table_lock
);
963 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
968 if (!huge_pte_none(huge_ptep_get(ptep
)))
971 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
972 && (vma
->vm_flags
& VM_SHARED
)));
973 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
975 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
976 /* Optimization, do the COW without a second fault */
977 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
);
980 spin_unlock(&mm
->page_table_lock
);
986 spin_unlock(&mm
->page_table_lock
);
992 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
993 unsigned long address
, int write_access
)
998 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
1000 ptep
= huge_pte_alloc(mm
, address
);
1002 return VM_FAULT_OOM
;
1005 * Serialize hugepage allocation and instantiation, so that we don't
1006 * get spurious allocation failures if two CPUs race to instantiate
1007 * the same page in the page cache.
1009 mutex_lock(&hugetlb_instantiation_mutex
);
1010 entry
= huge_ptep_get(ptep
);
1011 if (huge_pte_none(entry
)) {
1012 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
1013 mutex_unlock(&hugetlb_instantiation_mutex
);
1019 spin_lock(&mm
->page_table_lock
);
1020 /* Check for a racing update before calling hugetlb_cow */
1021 if (likely(pte_same(entry
, huge_ptep_get(ptep
))))
1022 if (write_access
&& !pte_write(entry
))
1023 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
);
1024 spin_unlock(&mm
->page_table_lock
);
1025 mutex_unlock(&hugetlb_instantiation_mutex
);
1030 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1031 struct page
**pages
, struct vm_area_struct
**vmas
,
1032 unsigned long *position
, int *length
, int i
,
1035 unsigned long pfn_offset
;
1036 unsigned long vaddr
= *position
;
1037 int remainder
= *length
;
1039 spin_lock(&mm
->page_table_lock
);
1040 while (vaddr
< vma
->vm_end
&& remainder
) {
1045 * Some archs (sparc64, sh*) have multiple pte_ts to
1046 * each hugepage. We have to make * sure we get the
1047 * first, for the page indexing below to work.
1049 pte
= huge_pte_offset(mm
, vaddr
& HPAGE_MASK
);
1051 if (!pte
|| huge_pte_none(huge_ptep_get(pte
)) ||
1052 (write
&& !pte_write(huge_ptep_get(pte
)))) {
1055 spin_unlock(&mm
->page_table_lock
);
1056 ret
= hugetlb_fault(mm
, vma
, vaddr
, write
);
1057 spin_lock(&mm
->page_table_lock
);
1058 if (!(ret
& VM_FAULT_ERROR
))
1067 pfn_offset
= (vaddr
& ~HPAGE_MASK
) >> PAGE_SHIFT
;
1068 page
= pte_page(huge_ptep_get(pte
));
1072 pages
[i
] = page
+ pfn_offset
;
1082 if (vaddr
< vma
->vm_end
&& remainder
&&
1083 pfn_offset
< HPAGE_SIZE
/PAGE_SIZE
) {
1085 * We use pfn_offset to avoid touching the pageframes
1086 * of this compound page.
1091 spin_unlock(&mm
->page_table_lock
);
1092 *length
= remainder
;
1098 void hugetlb_change_protection(struct vm_area_struct
*vma
,
1099 unsigned long address
, unsigned long end
, pgprot_t newprot
)
1101 struct mm_struct
*mm
= vma
->vm_mm
;
1102 unsigned long start
= address
;
1106 BUG_ON(address
>= end
);
1107 flush_cache_range(vma
, address
, end
);
1109 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1110 spin_lock(&mm
->page_table_lock
);
1111 for (; address
< end
; address
+= HPAGE_SIZE
) {
1112 ptep
= huge_pte_offset(mm
, address
);
1115 if (huge_pmd_unshare(mm
, &address
, ptep
))
1117 if (!huge_pte_none(huge_ptep_get(ptep
))) {
1118 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
1119 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
1120 set_huge_pte_at(mm
, address
, ptep
, pte
);
1123 spin_unlock(&mm
->page_table_lock
);
1124 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1126 flush_tlb_range(vma
, start
, end
);
1129 struct file_region
{
1130 struct list_head link
;
1135 static long region_add(struct list_head
*head
, long f
, long t
)
1137 struct file_region
*rg
, *nrg
, *trg
;
1139 /* Locate the region we are either in or before. */
1140 list_for_each_entry(rg
, head
, link
)
1144 /* Round our left edge to the current segment if it encloses us. */
1148 /* Check for and consume any regions we now overlap with. */
1150 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
1151 if (&rg
->link
== head
)
1156 /* If this area reaches higher then extend our area to
1157 * include it completely. If this is not the first area
1158 * which we intend to reuse, free it. */
1162 list_del(&rg
->link
);
1171 static long region_chg(struct list_head
*head
, long f
, long t
)
1173 struct file_region
*rg
, *nrg
;
1176 /* Locate the region we are before or in. */
1177 list_for_each_entry(rg
, head
, link
)
1181 /* If we are below the current region then a new region is required.
1182 * Subtle, allocate a new region at the position but make it zero
1183 * size such that we can guarantee to record the reservation. */
1184 if (&rg
->link
== head
|| t
< rg
->from
) {
1185 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
1190 INIT_LIST_HEAD(&nrg
->link
);
1191 list_add(&nrg
->link
, rg
->link
.prev
);
1196 /* Round our left edge to the current segment if it encloses us. */
1201 /* Check for and consume any regions we now overlap with. */
1202 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
1203 if (&rg
->link
== head
)
1208 /* We overlap with this area, if it extends futher than
1209 * us then we must extend ourselves. Account for its
1210 * existing reservation. */
1215 chg
-= rg
->to
- rg
->from
;
1220 static long region_truncate(struct list_head
*head
, long end
)
1222 struct file_region
*rg
, *trg
;
1225 /* Locate the region we are either in or before. */
1226 list_for_each_entry(rg
, head
, link
)
1229 if (&rg
->link
== head
)
1232 /* If we are in the middle of a region then adjust it. */
1233 if (end
> rg
->from
) {
1236 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
1239 /* Drop any remaining regions. */
1240 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
1241 if (&rg
->link
== head
)
1243 chg
+= rg
->to
- rg
->from
;
1244 list_del(&rg
->link
);
1250 static int hugetlb_acct_memory(long delta
)
1254 spin_lock(&hugetlb_lock
);
1256 * When cpuset is configured, it breaks the strict hugetlb page
1257 * reservation as the accounting is done on a global variable. Such
1258 * reservation is completely rubbish in the presence of cpuset because
1259 * the reservation is not checked against page availability for the
1260 * current cpuset. Application can still potentially OOM'ed by kernel
1261 * with lack of free htlb page in cpuset that the task is in.
1262 * Attempt to enforce strict accounting with cpuset is almost
1263 * impossible (or too ugly) because cpuset is too fluid that
1264 * task or memory node can be dynamically moved between cpusets.
1266 * The change of semantics for shared hugetlb mapping with cpuset is
1267 * undesirable. However, in order to preserve some of the semantics,
1268 * we fall back to check against current free page availability as
1269 * a best attempt and hopefully to minimize the impact of changing
1270 * semantics that cpuset has.
1273 if (gather_surplus_pages(delta
) < 0)
1276 if (delta
> cpuset_mems_nr(free_huge_pages_node
)) {
1277 return_unused_surplus_pages(delta
);
1284 return_unused_surplus_pages((unsigned long) -delta
);
1287 spin_unlock(&hugetlb_lock
);
1291 int hugetlb_reserve_pages(struct inode
*inode
, long from
, long to
)
1295 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
1299 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
1301 ret
= hugetlb_acct_memory(chg
);
1303 hugetlb_put_quota(inode
->i_mapping
, chg
);
1306 region_add(&inode
->i_mapping
->private_list
, from
, to
);
1310 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
1312 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
1314 spin_lock(&inode
->i_lock
);
1315 inode
->i_blocks
-= BLOCKS_PER_HUGEPAGE
* freed
;
1316 spin_unlock(&inode
->i_lock
);
1318 hugetlb_put_quota(inode
->i_mapping
, (chg
- freed
));
1319 hugetlb_acct_memory(-(chg
- freed
));