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 unsigned long max_huge_pages
;
27 static struct list_head hugepage_freelists
[MAX_NUMNODES
];
28 static unsigned int nr_huge_pages_node
[MAX_NUMNODES
];
29 static unsigned int free_huge_pages_node
[MAX_NUMNODES
];
30 static gfp_t htlb_alloc_mask
= GFP_HIGHUSER
;
31 unsigned long hugepages_treat_as_movable
;
34 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
36 static DEFINE_SPINLOCK(hugetlb_lock
);
38 static void clear_huge_page(struct page
*page
, unsigned long addr
)
43 for (i
= 0; i
< (HPAGE_SIZE
/PAGE_SIZE
); i
++) {
45 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
49 static void copy_huge_page(struct page
*dst
, struct page
*src
,
50 unsigned long addr
, struct vm_area_struct
*vma
)
55 for (i
= 0; i
< HPAGE_SIZE
/PAGE_SIZE
; i
++) {
57 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
61 static void enqueue_huge_page(struct page
*page
)
63 int nid
= page_to_nid(page
);
64 list_add(&page
->lru
, &hugepage_freelists
[nid
]);
66 free_huge_pages_node
[nid
]++;
69 static struct page
*dequeue_huge_page(struct vm_area_struct
*vma
,
70 unsigned long address
)
73 struct page
*page
= NULL
;
74 struct mempolicy
*mpol
;
75 struct zonelist
*zonelist
= huge_zonelist(vma
, address
,
76 htlb_alloc_mask
, &mpol
);
79 for (z
= zonelist
->zones
; *z
; z
++) {
80 nid
= zone_to_nid(*z
);
81 if (cpuset_zone_allowed_softwall(*z
, htlb_alloc_mask
) &&
82 !list_empty(&hugepage_freelists
[nid
])) {
83 page
= list_entry(hugepage_freelists
[nid
].next
,
87 free_huge_pages_node
[nid
]--;
91 mpol_free(mpol
); /* unref if mpol !NULL */
95 static void update_and_free_page(struct page
*page
)
99 nr_huge_pages_node
[page_to_nid(page
)]--;
100 for (i
= 0; i
< (HPAGE_SIZE
/ PAGE_SIZE
); i
++) {
101 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
| 1 << PG_referenced
|
102 1 << PG_dirty
| 1 << PG_active
| 1 << PG_reserved
|
103 1 << PG_private
| 1<< PG_writeback
);
105 set_compound_page_dtor(page
, NULL
);
106 set_page_refcounted(page
);
107 __free_pages(page
, HUGETLB_PAGE_ORDER
);
110 static void free_huge_page(struct page
*page
)
112 BUG_ON(page_count(page
));
114 INIT_LIST_HEAD(&page
->lru
);
116 spin_lock(&hugetlb_lock
);
117 enqueue_huge_page(page
);
118 spin_unlock(&hugetlb_lock
);
121 static int alloc_fresh_huge_page(void)
128 * Copy static prev_nid to local nid, work on that, then copy it
129 * back to prev_nid afterwards: otherwise there's a window in which
130 * a racer might pass invalid nid MAX_NUMNODES to alloc_pages_node.
131 * But we don't need to use a spin_lock here: it really doesn't
132 * matter if occasionally a racer chooses the same nid as we do.
134 nid
= next_node(prev_nid
, node_online_map
);
135 if (nid
== MAX_NUMNODES
)
136 nid
= first_node(node_online_map
);
139 page
= alloc_pages_node(nid
, htlb_alloc_mask
|__GFP_COMP
|__GFP_NOWARN
,
142 set_compound_page_dtor(page
, free_huge_page
);
143 spin_lock(&hugetlb_lock
);
145 nr_huge_pages_node
[page_to_nid(page
)]++;
146 spin_unlock(&hugetlb_lock
);
147 put_page(page
); /* free it into the hugepage allocator */
153 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
158 spin_lock(&hugetlb_lock
);
159 if (vma
->vm_flags
& VM_MAYSHARE
)
161 else if (free_huge_pages
<= resv_huge_pages
)
164 page
= dequeue_huge_page(vma
, addr
);
168 spin_unlock(&hugetlb_lock
);
169 set_page_refcounted(page
);
173 if (vma
->vm_flags
& VM_MAYSHARE
)
175 spin_unlock(&hugetlb_lock
);
179 static int __init
hugetlb_init(void)
183 if (HPAGE_SHIFT
== 0)
186 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
187 INIT_LIST_HEAD(&hugepage_freelists
[i
]);
189 for (i
= 0; i
< max_huge_pages
; ++i
) {
190 if (!alloc_fresh_huge_page())
193 max_huge_pages
= free_huge_pages
= nr_huge_pages
= i
;
194 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages
);
197 module_init(hugetlb_init
);
199 static int __init
hugetlb_setup(char *s
)
201 if (sscanf(s
, "%lu", &max_huge_pages
) <= 0)
205 __setup("hugepages=", hugetlb_setup
);
207 static unsigned int cpuset_mems_nr(unsigned int *array
)
212 for_each_node_mask(node
, cpuset_current_mems_allowed
)
219 #ifdef CONFIG_HIGHMEM
220 static void try_to_free_low(unsigned long count
)
224 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
225 struct page
*page
, *next
;
226 list_for_each_entry_safe(page
, next
, &hugepage_freelists
[i
], lru
) {
227 if (PageHighMem(page
))
229 list_del(&page
->lru
);
230 update_and_free_page(page
);
232 free_huge_pages_node
[page_to_nid(page
)]--;
233 if (count
>= nr_huge_pages
)
239 static inline void try_to_free_low(unsigned long count
)
244 static unsigned long set_max_huge_pages(unsigned long count
)
246 while (count
> nr_huge_pages
) {
247 if (!alloc_fresh_huge_page())
248 return nr_huge_pages
;
250 if (count
>= nr_huge_pages
)
251 return nr_huge_pages
;
253 spin_lock(&hugetlb_lock
);
254 count
= max(count
, resv_huge_pages
);
255 try_to_free_low(count
);
256 while (count
< nr_huge_pages
) {
257 struct page
*page
= dequeue_huge_page(NULL
, 0);
260 update_and_free_page(page
);
262 spin_unlock(&hugetlb_lock
);
263 return nr_huge_pages
;
266 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
267 struct file
*file
, void __user
*buffer
,
268 size_t *length
, loff_t
*ppos
)
270 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
271 max_huge_pages
= set_max_huge_pages(max_huge_pages
);
275 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
276 struct file
*file
, void __user
*buffer
,
277 size_t *length
, loff_t
*ppos
)
279 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
280 if (hugepages_treat_as_movable
)
281 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
283 htlb_alloc_mask
= GFP_HIGHUSER
;
287 #endif /* CONFIG_SYSCTL */
289 int hugetlb_report_meminfo(char *buf
)
292 "HugePages_Total: %5lu\n"
293 "HugePages_Free: %5lu\n"
294 "HugePages_Rsvd: %5lu\n"
295 "Hugepagesize: %5lu kB\n",
302 int hugetlb_report_node_meminfo(int nid
, char *buf
)
305 "Node %d HugePages_Total: %5u\n"
306 "Node %d HugePages_Free: %5u\n",
307 nid
, nr_huge_pages_node
[nid
],
308 nid
, free_huge_pages_node
[nid
]);
311 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
312 unsigned long hugetlb_total_pages(void)
314 return nr_huge_pages
* (HPAGE_SIZE
/ PAGE_SIZE
);
318 * We cannot handle pagefaults against hugetlb pages at all. They cause
319 * handle_mm_fault() to try to instantiate regular-sized pages in the
320 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
323 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
329 struct vm_operations_struct hugetlb_vm_ops
= {
330 .fault
= hugetlb_vm_op_fault
,
333 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
340 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
342 entry
= pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
344 entry
= pte_mkyoung(entry
);
345 entry
= pte_mkhuge(entry
);
350 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
351 unsigned long address
, pte_t
*ptep
)
355 entry
= pte_mkwrite(pte_mkdirty(*ptep
));
356 if (ptep_set_access_flags(vma
, address
, ptep
, entry
, 1)) {
357 update_mmu_cache(vma
, address
, entry
);
362 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
363 struct vm_area_struct
*vma
)
365 pte_t
*src_pte
, *dst_pte
, entry
;
366 struct page
*ptepage
;
370 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
372 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= HPAGE_SIZE
) {
373 src_pte
= huge_pte_offset(src
, addr
);
376 dst_pte
= huge_pte_alloc(dst
, addr
);
379 spin_lock(&dst
->page_table_lock
);
380 spin_lock(&src
->page_table_lock
);
381 if (!pte_none(*src_pte
)) {
383 ptep_set_wrprotect(src
, addr
, src_pte
);
385 ptepage
= pte_page(entry
);
387 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
389 spin_unlock(&src
->page_table_lock
);
390 spin_unlock(&dst
->page_table_lock
);
398 void __unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
401 struct mm_struct
*mm
= vma
->vm_mm
;
402 unsigned long address
;
408 * A page gathering list, protected by per file i_mmap_lock. The
409 * lock is used to avoid list corruption from multiple unmapping
410 * of the same page since we are using page->lru.
412 LIST_HEAD(page_list
);
414 WARN_ON(!is_vm_hugetlb_page(vma
));
415 BUG_ON(start
& ~HPAGE_MASK
);
416 BUG_ON(end
& ~HPAGE_MASK
);
418 spin_lock(&mm
->page_table_lock
);
419 for (address
= start
; address
< end
; address
+= HPAGE_SIZE
) {
420 ptep
= huge_pte_offset(mm
, address
);
424 if (huge_pmd_unshare(mm
, &address
, ptep
))
427 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
431 page
= pte_page(pte
);
433 set_page_dirty(page
);
434 list_add(&page
->lru
, &page_list
);
436 spin_unlock(&mm
->page_table_lock
);
437 flush_tlb_range(vma
, start
, end
);
438 list_for_each_entry_safe(page
, tmp
, &page_list
, lru
) {
439 list_del(&page
->lru
);
444 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
448 * It is undesirable to test vma->vm_file as it should be non-null
449 * for valid hugetlb area. However, vm_file will be NULL in the error
450 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
451 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
452 * to clean up. Since no pte has actually been setup, it is safe to
453 * do nothing in this case.
456 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
457 __unmap_hugepage_range(vma
, start
, end
);
458 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
462 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
463 unsigned long address
, pte_t
*ptep
, pte_t pte
)
465 struct page
*old_page
, *new_page
;
468 old_page
= pte_page(pte
);
470 /* If no-one else is actually using this page, avoid the copy
471 * and just make the page writable */
472 avoidcopy
= (page_count(old_page
) == 1);
474 set_huge_ptep_writable(vma
, address
, ptep
);
478 page_cache_get(old_page
);
479 new_page
= alloc_huge_page(vma
, address
);
482 page_cache_release(old_page
);
486 spin_unlock(&mm
->page_table_lock
);
487 copy_huge_page(new_page
, old_page
, address
, vma
);
488 spin_lock(&mm
->page_table_lock
);
490 ptep
= huge_pte_offset(mm
, address
& HPAGE_MASK
);
491 if (likely(pte_same(*ptep
, pte
))) {
493 set_huge_pte_at(mm
, address
, ptep
,
494 make_huge_pte(vma
, new_page
, 1));
495 /* Make the old page be freed below */
498 page_cache_release(new_page
);
499 page_cache_release(old_page
);
503 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
504 unsigned long address
, pte_t
*ptep
, int write_access
)
506 int ret
= VM_FAULT_SIGBUS
;
510 struct address_space
*mapping
;
513 mapping
= vma
->vm_file
->f_mapping
;
514 idx
= ((address
- vma
->vm_start
) >> HPAGE_SHIFT
)
515 + (vma
->vm_pgoff
>> (HPAGE_SHIFT
- PAGE_SHIFT
));
518 * Use page lock to guard against racing truncation
519 * before we get page_table_lock.
522 page
= find_lock_page(mapping
, idx
);
524 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
527 if (hugetlb_get_quota(mapping
))
529 page
= alloc_huge_page(vma
, address
);
531 hugetlb_put_quota(mapping
);
535 clear_huge_page(page
, address
);
537 if (vma
->vm_flags
& VM_SHARED
) {
540 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
543 hugetlb_put_quota(mapping
);
552 spin_lock(&mm
->page_table_lock
);
553 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
558 if (!pte_none(*ptep
))
561 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
562 && (vma
->vm_flags
& VM_SHARED
)));
563 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
565 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
566 /* Optimization, do the COW without a second fault */
567 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
);
570 spin_unlock(&mm
->page_table_lock
);
576 spin_unlock(&mm
->page_table_lock
);
577 hugetlb_put_quota(mapping
);
583 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
584 unsigned long address
, int write_access
)
589 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
591 ptep
= huge_pte_alloc(mm
, address
);
596 * Serialize hugepage allocation and instantiation, so that we don't
597 * get spurious allocation failures if two CPUs race to instantiate
598 * the same page in the page cache.
600 mutex_lock(&hugetlb_instantiation_mutex
);
602 if (pte_none(entry
)) {
603 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
604 mutex_unlock(&hugetlb_instantiation_mutex
);
610 spin_lock(&mm
->page_table_lock
);
611 /* Check for a racing update before calling hugetlb_cow */
612 if (likely(pte_same(entry
, *ptep
)))
613 if (write_access
&& !pte_write(entry
))
614 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
);
615 spin_unlock(&mm
->page_table_lock
);
616 mutex_unlock(&hugetlb_instantiation_mutex
);
621 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
622 struct page
**pages
, struct vm_area_struct
**vmas
,
623 unsigned long *position
, int *length
, int i
)
625 unsigned long pfn_offset
;
626 unsigned long vaddr
= *position
;
627 int remainder
= *length
;
629 spin_lock(&mm
->page_table_lock
);
630 while (vaddr
< vma
->vm_end
&& remainder
) {
635 * Some archs (sparc64, sh*) have multiple pte_ts to
636 * each hugepage. We have to make * sure we get the
637 * first, for the page indexing below to work.
639 pte
= huge_pte_offset(mm
, vaddr
& HPAGE_MASK
);
641 if (!pte
|| pte_none(*pte
)) {
644 spin_unlock(&mm
->page_table_lock
);
645 ret
= hugetlb_fault(mm
, vma
, vaddr
, 0);
646 spin_lock(&mm
->page_table_lock
);
647 if (!(ret
& VM_FAULT_ERROR
))
656 pfn_offset
= (vaddr
& ~HPAGE_MASK
) >> PAGE_SHIFT
;
657 page
= pte_page(*pte
);
661 pages
[i
] = page
+ pfn_offset
;
671 if (vaddr
< vma
->vm_end
&& remainder
&&
672 pfn_offset
< HPAGE_SIZE
/PAGE_SIZE
) {
674 * We use pfn_offset to avoid touching the pageframes
675 * of this compound page.
680 spin_unlock(&mm
->page_table_lock
);
687 void hugetlb_change_protection(struct vm_area_struct
*vma
,
688 unsigned long address
, unsigned long end
, pgprot_t newprot
)
690 struct mm_struct
*mm
= vma
->vm_mm
;
691 unsigned long start
= address
;
695 BUG_ON(address
>= end
);
696 flush_cache_range(vma
, address
, end
);
698 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
699 spin_lock(&mm
->page_table_lock
);
700 for (; address
< end
; address
+= HPAGE_SIZE
) {
701 ptep
= huge_pte_offset(mm
, address
);
704 if (huge_pmd_unshare(mm
, &address
, ptep
))
706 if (!pte_none(*ptep
)) {
707 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
708 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
709 set_huge_pte_at(mm
, address
, ptep
, pte
);
712 spin_unlock(&mm
->page_table_lock
);
713 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
715 flush_tlb_range(vma
, start
, end
);
719 struct list_head link
;
724 static long region_add(struct list_head
*head
, long f
, long t
)
726 struct file_region
*rg
, *nrg
, *trg
;
728 /* Locate the region we are either in or before. */
729 list_for_each_entry(rg
, head
, link
)
733 /* Round our left edge to the current segment if it encloses us. */
737 /* Check for and consume any regions we now overlap with. */
739 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
740 if (&rg
->link
== head
)
745 /* If this area reaches higher then extend our area to
746 * include it completely. If this is not the first area
747 * which we intend to reuse, free it. */
760 static long region_chg(struct list_head
*head
, long f
, long t
)
762 struct file_region
*rg
, *nrg
;
765 /* Locate the region we are before or in. */
766 list_for_each_entry(rg
, head
, link
)
770 /* If we are below the current region then a new region is required.
771 * Subtle, allocate a new region at the position but make it zero
772 * size such that we can guarentee to record the reservation. */
773 if (&rg
->link
== head
|| t
< rg
->from
) {
774 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
779 INIT_LIST_HEAD(&nrg
->link
);
780 list_add(&nrg
->link
, rg
->link
.prev
);
785 /* Round our left edge to the current segment if it encloses us. */
790 /* Check for and consume any regions we now overlap with. */
791 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
792 if (&rg
->link
== head
)
797 /* We overlap with this area, if it extends futher than
798 * us then we must extend ourselves. Account for its
799 * existing reservation. */
804 chg
-= rg
->to
- rg
->from
;
809 static long region_truncate(struct list_head
*head
, long end
)
811 struct file_region
*rg
, *trg
;
814 /* Locate the region we are either in or before. */
815 list_for_each_entry(rg
, head
, link
)
818 if (&rg
->link
== head
)
821 /* If we are in the middle of a region then adjust it. */
822 if (end
> rg
->from
) {
825 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
828 /* Drop any remaining regions. */
829 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
830 if (&rg
->link
== head
)
832 chg
+= rg
->to
- rg
->from
;
839 static int hugetlb_acct_memory(long delta
)
843 spin_lock(&hugetlb_lock
);
844 if ((delta
+ resv_huge_pages
) <= free_huge_pages
) {
845 resv_huge_pages
+= delta
;
848 spin_unlock(&hugetlb_lock
);
852 int hugetlb_reserve_pages(struct inode
*inode
, long from
, long to
)
856 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
860 * When cpuset is configured, it breaks the strict hugetlb page
861 * reservation as the accounting is done on a global variable. Such
862 * reservation is completely rubbish in the presence of cpuset because
863 * the reservation is not checked against page availability for the
864 * current cpuset. Application can still potentially OOM'ed by kernel
865 * with lack of free htlb page in cpuset that the task is in.
866 * Attempt to enforce strict accounting with cpuset is almost
867 * impossible (or too ugly) because cpuset is too fluid that
868 * task or memory node can be dynamically moved between cpusets.
870 * The change of semantics for shared hugetlb mapping with cpuset is
871 * undesirable. However, in order to preserve some of the semantics,
872 * we fall back to check against current free page availability as
873 * a best attempt and hopefully to minimize the impact of changing
874 * semantics that cpuset has.
876 if (chg
> cpuset_mems_nr(free_huge_pages_node
))
879 ret
= hugetlb_acct_memory(chg
);
882 region_add(&inode
->i_mapping
->private_list
, from
, to
);
886 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
888 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
889 hugetlb_acct_memory(freed
- chg
);