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 unsigned long max_huge_pages
;
28 static struct list_head hugepage_freelists
[MAX_NUMNODES
];
29 static unsigned int nr_huge_pages_node
[MAX_NUMNODES
];
30 static unsigned int free_huge_pages_node
[MAX_NUMNODES
];
31 static unsigned int surplus_huge_pages_node
[MAX_NUMNODES
];
32 static gfp_t htlb_alloc_mask
= GFP_HIGHUSER
;
33 unsigned long hugepages_treat_as_movable
;
34 int hugetlb_dynamic_pool
;
35 static int hugetlb_next_nid
;
38 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
40 static DEFINE_SPINLOCK(hugetlb_lock
);
42 static void clear_huge_page(struct page
*page
, unsigned long addr
)
47 for (i
= 0; i
< (HPAGE_SIZE
/PAGE_SIZE
); i
++) {
49 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
53 static void copy_huge_page(struct page
*dst
, struct page
*src
,
54 unsigned long addr
, struct vm_area_struct
*vma
)
59 for (i
= 0; i
< HPAGE_SIZE
/PAGE_SIZE
; i
++) {
61 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
65 static void enqueue_huge_page(struct page
*page
)
67 int nid
= page_to_nid(page
);
68 list_add(&page
->lru
, &hugepage_freelists
[nid
]);
70 free_huge_pages_node
[nid
]++;
73 static struct page
*dequeue_huge_page(struct vm_area_struct
*vma
,
74 unsigned long address
)
77 struct page
*page
= NULL
;
78 struct mempolicy
*mpol
;
79 struct zonelist
*zonelist
= huge_zonelist(vma
, address
,
80 htlb_alloc_mask
, &mpol
);
83 for (z
= zonelist
->zones
; *z
; z
++) {
84 nid
= zone_to_nid(*z
);
85 if (cpuset_zone_allowed_softwall(*z
, htlb_alloc_mask
) &&
86 !list_empty(&hugepage_freelists
[nid
])) {
87 page
= list_entry(hugepage_freelists
[nid
].next
,
91 free_huge_pages_node
[nid
]--;
92 if (vma
&& vma
->vm_flags
& VM_MAYSHARE
)
97 mpol_free(mpol
); /* unref if mpol !NULL */
101 static void update_and_free_page(struct page
*page
)
105 nr_huge_pages_node
[page_to_nid(page
)]--;
106 for (i
= 0; i
< (HPAGE_SIZE
/ PAGE_SIZE
); i
++) {
107 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
| 1 << PG_referenced
|
108 1 << PG_dirty
| 1 << PG_active
| 1 << PG_reserved
|
109 1 << PG_private
| 1<< PG_writeback
);
111 set_compound_page_dtor(page
, NULL
);
112 set_page_refcounted(page
);
113 __free_pages(page
, HUGETLB_PAGE_ORDER
);
116 static void free_huge_page(struct page
*page
)
118 int nid
= page_to_nid(page
);
120 BUG_ON(page_count(page
));
121 INIT_LIST_HEAD(&page
->lru
);
123 spin_lock(&hugetlb_lock
);
124 if (surplus_huge_pages_node
[nid
]) {
125 update_and_free_page(page
);
126 surplus_huge_pages
--;
127 surplus_huge_pages_node
[nid
]--;
129 enqueue_huge_page(page
);
131 spin_unlock(&hugetlb_lock
);
135 * Increment or decrement surplus_huge_pages. Keep node-specific counters
136 * balanced by operating on them in a round-robin fashion.
137 * Returns 1 if an adjustment was made.
139 static int adjust_pool_surplus(int delta
)
145 VM_BUG_ON(delta
!= -1 && delta
!= 1);
147 nid
= next_node(nid
, node_online_map
);
148 if (nid
== MAX_NUMNODES
)
149 nid
= first_node(node_online_map
);
151 /* To shrink on this node, there must be a surplus page */
152 if (delta
< 0 && !surplus_huge_pages_node
[nid
])
154 /* Surplus cannot exceed the total number of pages */
155 if (delta
> 0 && surplus_huge_pages_node
[nid
] >=
156 nr_huge_pages_node
[nid
])
159 surplus_huge_pages
+= delta
;
160 surplus_huge_pages_node
[nid
] += delta
;
163 } while (nid
!= prev_nid
);
169 static struct page
*alloc_fresh_huge_page_node(int nid
)
173 page
= alloc_pages_node(nid
,
174 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|__GFP_NOWARN
,
177 set_compound_page_dtor(page
, free_huge_page
);
178 spin_lock(&hugetlb_lock
);
180 nr_huge_pages_node
[nid
]++;
181 spin_unlock(&hugetlb_lock
);
182 put_page(page
); /* free it into the hugepage allocator */
188 static int alloc_fresh_huge_page(void)
195 start_nid
= hugetlb_next_nid
;
198 page
= alloc_fresh_huge_page_node(hugetlb_next_nid
);
202 * Use a helper variable to find the next node and then
203 * copy it back to hugetlb_next_nid afterwards:
204 * otherwise there's a window in which a racer might
205 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
206 * But we don't need to use a spin_lock here: it really
207 * doesn't matter if occasionally a racer chooses the
208 * same nid as we do. Move nid forward in the mask even
209 * if we just successfully allocated a hugepage so that
210 * the next caller gets hugepages on the next node.
212 next_nid
= next_node(hugetlb_next_nid
, node_online_map
);
213 if (next_nid
== MAX_NUMNODES
)
214 next_nid
= first_node(node_online_map
);
215 hugetlb_next_nid
= next_nid
;
216 } while (!page
&& hugetlb_next_nid
!= start_nid
);
221 static struct page
*alloc_buddy_huge_page(struct vm_area_struct
*vma
,
222 unsigned long address
)
226 /* Check if the dynamic pool is enabled */
227 if (!hugetlb_dynamic_pool
)
230 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|__GFP_NOWARN
,
233 set_compound_page_dtor(page
, free_huge_page
);
234 spin_lock(&hugetlb_lock
);
236 nr_huge_pages_node
[page_to_nid(page
)]++;
237 surplus_huge_pages
++;
238 surplus_huge_pages_node
[page_to_nid(page
)]++;
239 spin_unlock(&hugetlb_lock
);
246 * Increase the hugetlb pool such that it can accomodate a reservation
249 static int gather_surplus_pages(int delta
)
251 struct list_head surplus_list
;
252 struct page
*page
, *tmp
;
254 int needed
, allocated
;
256 needed
= (resv_huge_pages
+ delta
) - free_huge_pages
;
261 INIT_LIST_HEAD(&surplus_list
);
265 spin_unlock(&hugetlb_lock
);
266 for (i
= 0; i
< needed
; i
++) {
267 page
= alloc_buddy_huge_page(NULL
, 0);
270 * We were not able to allocate enough pages to
271 * satisfy the entire reservation so we free what
272 * we've allocated so far.
274 spin_lock(&hugetlb_lock
);
279 list_add(&page
->lru
, &surplus_list
);
284 * After retaking hugetlb_lock, we need to recalculate 'needed'
285 * because either resv_huge_pages or free_huge_pages may have changed.
287 spin_lock(&hugetlb_lock
);
288 needed
= (resv_huge_pages
+ delta
) - (free_huge_pages
+ allocated
);
293 * The surplus_list now contains _at_least_ the number of extra pages
294 * needed to accomodate the reservation. Add the appropriate number
295 * of pages to the hugetlb pool and free the extras back to the buddy
301 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
302 list_del(&page
->lru
);
304 enqueue_huge_page(page
);
307 * Decrement the refcount and free the page using its
308 * destructor. This must be done with hugetlb_lock
309 * unlocked which is safe because free_huge_page takes
310 * hugetlb_lock before deciding how to free the page.
312 spin_unlock(&hugetlb_lock
);
314 spin_lock(&hugetlb_lock
);
322 * When releasing a hugetlb pool reservation, any surplus pages that were
323 * allocated to satisfy the reservation must be explicitly freed if they were
326 void return_unused_surplus_pages(unsigned long unused_resv_pages
)
330 unsigned long nr_pages
;
332 nr_pages
= min(unused_resv_pages
, surplus_huge_pages
);
335 nid
= next_node(nid
, node_online_map
);
336 if (nid
== MAX_NUMNODES
)
337 nid
= first_node(node_online_map
);
339 if (!surplus_huge_pages_node
[nid
])
342 if (!list_empty(&hugepage_freelists
[nid
])) {
343 page
= list_entry(hugepage_freelists
[nid
].next
,
345 list_del(&page
->lru
);
346 update_and_free_page(page
);
348 free_huge_pages_node
[nid
]--;
349 surplus_huge_pages
--;
350 surplus_huge_pages_node
[nid
]--;
357 static struct page
*alloc_huge_page_shared(struct vm_area_struct
*vma
,
362 spin_lock(&hugetlb_lock
);
363 page
= dequeue_huge_page(vma
, addr
);
364 spin_unlock(&hugetlb_lock
);
368 static struct page
*alloc_huge_page_private(struct vm_area_struct
*vma
,
371 struct page
*page
= NULL
;
373 spin_lock(&hugetlb_lock
);
374 if (free_huge_pages
> resv_huge_pages
)
375 page
= dequeue_huge_page(vma
, addr
);
376 spin_unlock(&hugetlb_lock
);
378 page
= alloc_buddy_huge_page(vma
, addr
);
382 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
387 if (vma
->vm_flags
& VM_MAYSHARE
)
388 page
= alloc_huge_page_shared(vma
, addr
);
390 page
= alloc_huge_page_private(vma
, addr
);
392 set_page_refcounted(page
);
396 static int __init
hugetlb_init(void)
400 if (HPAGE_SHIFT
== 0)
403 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
404 INIT_LIST_HEAD(&hugepage_freelists
[i
]);
406 hugetlb_next_nid
= first_node(node_online_map
);
408 for (i
= 0; i
< max_huge_pages
; ++i
) {
409 if (!alloc_fresh_huge_page())
412 max_huge_pages
= free_huge_pages
= nr_huge_pages
= i
;
413 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages
);
416 module_init(hugetlb_init
);
418 static int __init
hugetlb_setup(char *s
)
420 if (sscanf(s
, "%lu", &max_huge_pages
) <= 0)
424 __setup("hugepages=", hugetlb_setup
);
426 static unsigned int cpuset_mems_nr(unsigned int *array
)
431 for_each_node_mask(node
, cpuset_current_mems_allowed
)
438 #ifdef CONFIG_HIGHMEM
439 static void try_to_free_low(unsigned long count
)
443 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
444 struct page
*page
, *next
;
445 list_for_each_entry_safe(page
, next
, &hugepage_freelists
[i
], lru
) {
446 if (count
>= nr_huge_pages
)
448 if (PageHighMem(page
))
450 list_del(&page
->lru
);
451 update_and_free_page(page
);
453 free_huge_pages_node
[page_to_nid(page
)]--;
458 static inline void try_to_free_low(unsigned long count
)
463 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
464 static unsigned long set_max_huge_pages(unsigned long count
)
466 unsigned long min_count
, ret
;
469 * Increase the pool size
470 * First take pages out of surplus state. Then make up the
471 * remaining difference by allocating fresh huge pages.
473 spin_lock(&hugetlb_lock
);
474 while (surplus_huge_pages
&& count
> persistent_huge_pages
) {
475 if (!adjust_pool_surplus(-1))
479 while (count
> persistent_huge_pages
) {
482 * If this allocation races such that we no longer need the
483 * page, free_huge_page will handle it by freeing the page
484 * and reducing the surplus.
486 spin_unlock(&hugetlb_lock
);
487 ret
= alloc_fresh_huge_page();
488 spin_lock(&hugetlb_lock
);
495 * Decrease the pool size
496 * First return free pages to the buddy allocator (being careful
497 * to keep enough around to satisfy reservations). Then place
498 * pages into surplus state as needed so the pool will shrink
499 * to the desired size as pages become free.
501 min_count
= resv_huge_pages
+ nr_huge_pages
- free_huge_pages
;
502 min_count
= max(count
, min_count
);
503 try_to_free_low(min_count
);
504 while (min_count
< persistent_huge_pages
) {
505 struct page
*page
= dequeue_huge_page(NULL
, 0);
508 update_and_free_page(page
);
510 while (count
< persistent_huge_pages
) {
511 if (!adjust_pool_surplus(1))
515 ret
= persistent_huge_pages
;
516 spin_unlock(&hugetlb_lock
);
520 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
521 struct file
*file
, void __user
*buffer
,
522 size_t *length
, loff_t
*ppos
)
524 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
525 max_huge_pages
= set_max_huge_pages(max_huge_pages
);
529 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
530 struct file
*file
, void __user
*buffer
,
531 size_t *length
, loff_t
*ppos
)
533 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
534 if (hugepages_treat_as_movable
)
535 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
537 htlb_alloc_mask
= GFP_HIGHUSER
;
541 #endif /* CONFIG_SYSCTL */
543 int hugetlb_report_meminfo(char *buf
)
546 "HugePages_Total: %5lu\n"
547 "HugePages_Free: %5lu\n"
548 "HugePages_Rsvd: %5lu\n"
549 "HugePages_Surp: %5lu\n"
550 "Hugepagesize: %5lu kB\n",
558 int hugetlb_report_node_meminfo(int nid
, char *buf
)
561 "Node %d HugePages_Total: %5u\n"
562 "Node %d HugePages_Free: %5u\n",
563 nid
, nr_huge_pages_node
[nid
],
564 nid
, free_huge_pages_node
[nid
]);
567 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
568 unsigned long hugetlb_total_pages(void)
570 return nr_huge_pages
* (HPAGE_SIZE
/ PAGE_SIZE
);
574 * We cannot handle pagefaults against hugetlb pages at all. They cause
575 * handle_mm_fault() to try to instantiate regular-sized pages in the
576 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
579 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
585 struct vm_operations_struct hugetlb_vm_ops
= {
586 .fault
= hugetlb_vm_op_fault
,
589 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
596 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
598 entry
= pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
600 entry
= pte_mkyoung(entry
);
601 entry
= pte_mkhuge(entry
);
606 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
607 unsigned long address
, pte_t
*ptep
)
611 entry
= pte_mkwrite(pte_mkdirty(*ptep
));
612 if (ptep_set_access_flags(vma
, address
, ptep
, entry
, 1)) {
613 update_mmu_cache(vma
, address
, entry
);
618 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
619 struct vm_area_struct
*vma
)
621 pte_t
*src_pte
, *dst_pte
, entry
;
622 struct page
*ptepage
;
626 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
628 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= HPAGE_SIZE
) {
629 src_pte
= huge_pte_offset(src
, addr
);
632 dst_pte
= huge_pte_alloc(dst
, addr
);
635 spin_lock(&dst
->page_table_lock
);
636 spin_lock(&src
->page_table_lock
);
637 if (!pte_none(*src_pte
)) {
639 ptep_set_wrprotect(src
, addr
, src_pte
);
641 ptepage
= pte_page(entry
);
643 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
645 spin_unlock(&src
->page_table_lock
);
646 spin_unlock(&dst
->page_table_lock
);
654 void __unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
657 struct mm_struct
*mm
= vma
->vm_mm
;
658 unsigned long address
;
664 * A page gathering list, protected by per file i_mmap_lock. The
665 * lock is used to avoid list corruption from multiple unmapping
666 * of the same page since we are using page->lru.
668 LIST_HEAD(page_list
);
670 WARN_ON(!is_vm_hugetlb_page(vma
));
671 BUG_ON(start
& ~HPAGE_MASK
);
672 BUG_ON(end
& ~HPAGE_MASK
);
674 spin_lock(&mm
->page_table_lock
);
675 for (address
= start
; address
< end
; address
+= HPAGE_SIZE
) {
676 ptep
= huge_pte_offset(mm
, address
);
680 if (huge_pmd_unshare(mm
, &address
, ptep
))
683 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
687 page
= pte_page(pte
);
689 set_page_dirty(page
);
690 list_add(&page
->lru
, &page_list
);
692 spin_unlock(&mm
->page_table_lock
);
693 flush_tlb_range(vma
, start
, end
);
694 list_for_each_entry_safe(page
, tmp
, &page_list
, lru
) {
695 list_del(&page
->lru
);
700 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
704 * It is undesirable to test vma->vm_file as it should be non-null
705 * for valid hugetlb area. However, vm_file will be NULL in the error
706 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
707 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
708 * to clean up. Since no pte has actually been setup, it is safe to
709 * do nothing in this case.
712 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
713 __unmap_hugepage_range(vma
, start
, end
);
714 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
718 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
719 unsigned long address
, pte_t
*ptep
, pte_t pte
)
721 struct page
*old_page
, *new_page
;
724 old_page
= pte_page(pte
);
726 /* If no-one else is actually using this page, avoid the copy
727 * and just make the page writable */
728 avoidcopy
= (page_count(old_page
) == 1);
730 set_huge_ptep_writable(vma
, address
, ptep
);
734 page_cache_get(old_page
);
735 new_page
= alloc_huge_page(vma
, address
);
738 page_cache_release(old_page
);
742 spin_unlock(&mm
->page_table_lock
);
743 copy_huge_page(new_page
, old_page
, address
, vma
);
744 spin_lock(&mm
->page_table_lock
);
746 ptep
= huge_pte_offset(mm
, address
& HPAGE_MASK
);
747 if (likely(pte_same(*ptep
, pte
))) {
749 set_huge_pte_at(mm
, address
, ptep
,
750 make_huge_pte(vma
, new_page
, 1));
751 /* Make the old page be freed below */
754 page_cache_release(new_page
);
755 page_cache_release(old_page
);
759 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
760 unsigned long address
, pte_t
*ptep
, int write_access
)
762 int ret
= VM_FAULT_SIGBUS
;
766 struct address_space
*mapping
;
769 mapping
= vma
->vm_file
->f_mapping
;
770 idx
= ((address
- vma
->vm_start
) >> HPAGE_SHIFT
)
771 + (vma
->vm_pgoff
>> (HPAGE_SHIFT
- PAGE_SHIFT
));
774 * Use page lock to guard against racing truncation
775 * before we get page_table_lock.
778 page
= find_lock_page(mapping
, idx
);
780 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
783 if (hugetlb_get_quota(mapping
))
785 page
= alloc_huge_page(vma
, address
);
787 hugetlb_put_quota(mapping
);
791 clear_huge_page(page
, address
);
793 if (vma
->vm_flags
& VM_SHARED
) {
796 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
799 hugetlb_put_quota(mapping
);
808 spin_lock(&mm
->page_table_lock
);
809 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
814 if (!pte_none(*ptep
))
817 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
818 && (vma
->vm_flags
& VM_SHARED
)));
819 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
821 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
822 /* Optimization, do the COW without a second fault */
823 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
);
826 spin_unlock(&mm
->page_table_lock
);
832 spin_unlock(&mm
->page_table_lock
);
833 hugetlb_put_quota(mapping
);
839 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
840 unsigned long address
, int write_access
)
845 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
847 ptep
= huge_pte_alloc(mm
, address
);
852 * Serialize hugepage allocation and instantiation, so that we don't
853 * get spurious allocation failures if two CPUs race to instantiate
854 * the same page in the page cache.
856 mutex_lock(&hugetlb_instantiation_mutex
);
858 if (pte_none(entry
)) {
859 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
860 mutex_unlock(&hugetlb_instantiation_mutex
);
866 spin_lock(&mm
->page_table_lock
);
867 /* Check for a racing update before calling hugetlb_cow */
868 if (likely(pte_same(entry
, *ptep
)))
869 if (write_access
&& !pte_write(entry
))
870 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
);
871 spin_unlock(&mm
->page_table_lock
);
872 mutex_unlock(&hugetlb_instantiation_mutex
);
877 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
878 struct page
**pages
, struct vm_area_struct
**vmas
,
879 unsigned long *position
, int *length
, int i
,
882 unsigned long pfn_offset
;
883 unsigned long vaddr
= *position
;
884 int remainder
= *length
;
886 spin_lock(&mm
->page_table_lock
);
887 while (vaddr
< vma
->vm_end
&& remainder
) {
892 * Some archs (sparc64, sh*) have multiple pte_ts to
893 * each hugepage. We have to make * sure we get the
894 * first, for the page indexing below to work.
896 pte
= huge_pte_offset(mm
, vaddr
& HPAGE_MASK
);
898 if (!pte
|| pte_none(*pte
)) {
901 spin_unlock(&mm
->page_table_lock
);
902 ret
= hugetlb_fault(mm
, vma
, vaddr
, write
);
903 spin_lock(&mm
->page_table_lock
);
904 if (!(ret
& VM_FAULT_ERROR
))
913 pfn_offset
= (vaddr
& ~HPAGE_MASK
) >> PAGE_SHIFT
;
914 page
= pte_page(*pte
);
918 pages
[i
] = page
+ pfn_offset
;
928 if (vaddr
< vma
->vm_end
&& remainder
&&
929 pfn_offset
< HPAGE_SIZE
/PAGE_SIZE
) {
931 * We use pfn_offset to avoid touching the pageframes
932 * of this compound page.
937 spin_unlock(&mm
->page_table_lock
);
944 void hugetlb_change_protection(struct vm_area_struct
*vma
,
945 unsigned long address
, unsigned long end
, pgprot_t newprot
)
947 struct mm_struct
*mm
= vma
->vm_mm
;
948 unsigned long start
= address
;
952 BUG_ON(address
>= end
);
953 flush_cache_range(vma
, address
, end
);
955 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
956 spin_lock(&mm
->page_table_lock
);
957 for (; address
< end
; address
+= HPAGE_SIZE
) {
958 ptep
= huge_pte_offset(mm
, address
);
961 if (huge_pmd_unshare(mm
, &address
, ptep
))
963 if (!pte_none(*ptep
)) {
964 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
965 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
966 set_huge_pte_at(mm
, address
, ptep
, pte
);
969 spin_unlock(&mm
->page_table_lock
);
970 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
972 flush_tlb_range(vma
, start
, end
);
976 struct list_head link
;
981 static long region_add(struct list_head
*head
, long f
, long t
)
983 struct file_region
*rg
, *nrg
, *trg
;
985 /* Locate the region we are either in or before. */
986 list_for_each_entry(rg
, head
, link
)
990 /* Round our left edge to the current segment if it encloses us. */
994 /* Check for and consume any regions we now overlap with. */
996 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
997 if (&rg
->link
== head
)
1002 /* If this area reaches higher then extend our area to
1003 * include it completely. If this is not the first area
1004 * which we intend to reuse, free it. */
1008 list_del(&rg
->link
);
1017 static long region_chg(struct list_head
*head
, long f
, long t
)
1019 struct file_region
*rg
, *nrg
;
1022 /* Locate the region we are before or in. */
1023 list_for_each_entry(rg
, head
, link
)
1027 /* If we are below the current region then a new region is required.
1028 * Subtle, allocate a new region at the position but make it zero
1029 * size such that we can guarantee to record the reservation. */
1030 if (&rg
->link
== head
|| t
< rg
->from
) {
1031 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
1036 INIT_LIST_HEAD(&nrg
->link
);
1037 list_add(&nrg
->link
, rg
->link
.prev
);
1042 /* Round our left edge to the current segment if it encloses us. */
1047 /* Check for and consume any regions we now overlap with. */
1048 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
1049 if (&rg
->link
== head
)
1054 /* We overlap with this area, if it extends futher than
1055 * us then we must extend ourselves. Account for its
1056 * existing reservation. */
1061 chg
-= rg
->to
- rg
->from
;
1066 static long region_truncate(struct list_head
*head
, long end
)
1068 struct file_region
*rg
, *trg
;
1071 /* Locate the region we are either in or before. */
1072 list_for_each_entry(rg
, head
, link
)
1075 if (&rg
->link
== head
)
1078 /* If we are in the middle of a region then adjust it. */
1079 if (end
> rg
->from
) {
1082 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
1085 /* Drop any remaining regions. */
1086 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
1087 if (&rg
->link
== head
)
1089 chg
+= rg
->to
- rg
->from
;
1090 list_del(&rg
->link
);
1096 static int hugetlb_acct_memory(long delta
)
1100 spin_lock(&hugetlb_lock
);
1102 * When cpuset is configured, it breaks the strict hugetlb page
1103 * reservation as the accounting is done on a global variable. Such
1104 * reservation is completely rubbish in the presence of cpuset because
1105 * the reservation is not checked against page availability for the
1106 * current cpuset. Application can still potentially OOM'ed by kernel
1107 * with lack of free htlb page in cpuset that the task is in.
1108 * Attempt to enforce strict accounting with cpuset is almost
1109 * impossible (or too ugly) because cpuset is too fluid that
1110 * task or memory node can be dynamically moved between cpusets.
1112 * The change of semantics for shared hugetlb mapping with cpuset is
1113 * undesirable. However, in order to preserve some of the semantics,
1114 * we fall back to check against current free page availability as
1115 * a best attempt and hopefully to minimize the impact of changing
1116 * semantics that cpuset has.
1119 if (gather_surplus_pages(delta
) < 0)
1122 if (delta
> cpuset_mems_nr(free_huge_pages_node
))
1127 resv_huge_pages
+= delta
;
1129 return_unused_surplus_pages((unsigned long) -delta
);
1132 spin_unlock(&hugetlb_lock
);
1136 int hugetlb_reserve_pages(struct inode
*inode
, long from
, long to
)
1140 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
1144 ret
= hugetlb_acct_memory(chg
);
1147 region_add(&inode
->i_mapping
->private_list
, from
, to
);
1151 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
1153 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
1154 hugetlb_acct_memory(freed
- chg
);