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
);
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 zonelist
*zonelist
= huge_zonelist(vma
, address
,
78 for (z
= zonelist
->zones
; *z
; z
++) {
79 nid
= zone_to_nid(*z
);
80 if (cpuset_zone_allowed_softwall(*z
, htlb_alloc_mask
) &&
81 !list_empty(&hugepage_freelists
[nid
]))
86 page
= list_entry(hugepage_freelists
[nid
].next
,
90 free_huge_pages_node
[nid
]--;
95 static void free_huge_page(struct page
*page
)
97 BUG_ON(page_count(page
));
99 INIT_LIST_HEAD(&page
->lru
);
101 spin_lock(&hugetlb_lock
);
102 enqueue_huge_page(page
);
103 spin_unlock(&hugetlb_lock
);
106 static int alloc_fresh_huge_page(void)
110 static DEFINE_SPINLOCK(nid_lock
);
113 spin_lock(&nid_lock
);
114 nid
= next_node(prev_nid
, node_online_map
);
115 if (nid
== MAX_NUMNODES
)
116 nid
= first_node(node_online_map
);
118 spin_unlock(&nid_lock
);
120 page
= alloc_pages_node(nid
, htlb_alloc_mask
|__GFP_COMP
|__GFP_NOWARN
,
123 set_compound_page_dtor(page
, free_huge_page
);
124 spin_lock(&hugetlb_lock
);
126 nr_huge_pages_node
[page_to_nid(page
)]++;
127 spin_unlock(&hugetlb_lock
);
128 put_page(page
); /* free it into the hugepage allocator */
134 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
139 spin_lock(&hugetlb_lock
);
140 if (vma
->vm_flags
& VM_MAYSHARE
)
142 else if (free_huge_pages
<= resv_huge_pages
)
145 page
= dequeue_huge_page(vma
, addr
);
149 spin_unlock(&hugetlb_lock
);
150 set_page_refcounted(page
);
154 if (vma
->vm_flags
& VM_MAYSHARE
)
156 spin_unlock(&hugetlb_lock
);
160 static int __init
hugetlb_init(void)
164 if (HPAGE_SHIFT
== 0)
167 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
168 INIT_LIST_HEAD(&hugepage_freelists
[i
]);
170 for (i
= 0; i
< max_huge_pages
; ++i
) {
171 if (!alloc_fresh_huge_page())
174 max_huge_pages
= free_huge_pages
= nr_huge_pages
= i
;
175 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages
);
178 module_init(hugetlb_init
);
180 static int __init
hugetlb_setup(char *s
)
182 if (sscanf(s
, "%lu", &max_huge_pages
) <= 0)
186 __setup("hugepages=", hugetlb_setup
);
188 static unsigned int cpuset_mems_nr(unsigned int *array
)
193 for_each_node_mask(node
, cpuset_current_mems_allowed
)
200 static void update_and_free_page(struct page
*page
)
204 nr_huge_pages_node
[page_to_nid(page
)]--;
205 for (i
= 0; i
< (HPAGE_SIZE
/ PAGE_SIZE
); i
++) {
206 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
| 1 << PG_referenced
|
207 1 << PG_dirty
| 1 << PG_active
| 1 << PG_reserved
|
208 1 << PG_private
| 1<< PG_writeback
);
210 page
[1].lru
.next
= NULL
;
211 set_page_refcounted(page
);
212 __free_pages(page
, HUGETLB_PAGE_ORDER
);
215 #ifdef CONFIG_HIGHMEM
216 static void try_to_free_low(unsigned long count
)
220 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
221 struct page
*page
, *next
;
222 list_for_each_entry_safe(page
, next
, &hugepage_freelists
[i
], lru
) {
223 if (PageHighMem(page
))
225 list_del(&page
->lru
);
226 update_and_free_page(page
);
228 free_huge_pages_node
[page_to_nid(page
)]--;
229 if (count
>= nr_huge_pages
)
235 static inline void try_to_free_low(unsigned long count
)
240 static unsigned long set_max_huge_pages(unsigned long count
)
242 while (count
> nr_huge_pages
) {
243 if (!alloc_fresh_huge_page())
244 return nr_huge_pages
;
246 if (count
>= nr_huge_pages
)
247 return nr_huge_pages
;
249 spin_lock(&hugetlb_lock
);
250 count
= max(count
, resv_huge_pages
);
251 try_to_free_low(count
);
252 while (count
< nr_huge_pages
) {
253 struct page
*page
= dequeue_huge_page(NULL
, 0);
256 update_and_free_page(page
);
258 spin_unlock(&hugetlb_lock
);
259 return nr_huge_pages
;
262 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
263 struct file
*file
, void __user
*buffer
,
264 size_t *length
, loff_t
*ppos
)
266 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
267 max_huge_pages
= set_max_huge_pages(max_huge_pages
);
271 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
272 struct file
*file
, void __user
*buffer
,
273 size_t *length
, loff_t
*ppos
)
275 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
276 if (hugepages_treat_as_movable
)
277 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
279 htlb_alloc_mask
= GFP_HIGHUSER
;
283 #endif /* CONFIG_SYSCTL */
285 int hugetlb_report_meminfo(char *buf
)
288 "HugePages_Total: %5lu\n"
289 "HugePages_Free: %5lu\n"
290 "HugePages_Rsvd: %5lu\n"
291 "Hugepagesize: %5lu kB\n",
298 int hugetlb_report_node_meminfo(int nid
, char *buf
)
301 "Node %d HugePages_Total: %5u\n"
302 "Node %d HugePages_Free: %5u\n",
303 nid
, nr_huge_pages_node
[nid
],
304 nid
, free_huge_pages_node
[nid
]);
307 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
308 unsigned long hugetlb_total_pages(void)
310 return nr_huge_pages
* (HPAGE_SIZE
/ PAGE_SIZE
);
314 * We cannot handle pagefaults against hugetlb pages at all. They cause
315 * handle_mm_fault() to try to instantiate regular-sized pages in the
316 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
319 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
325 struct vm_operations_struct hugetlb_vm_ops
= {
326 .fault
= hugetlb_vm_op_fault
,
329 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
336 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
338 entry
= pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
340 entry
= pte_mkyoung(entry
);
341 entry
= pte_mkhuge(entry
);
346 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
347 unsigned long address
, pte_t
*ptep
)
351 entry
= pte_mkwrite(pte_mkdirty(*ptep
));
352 if (ptep_set_access_flags(vma
, address
, ptep
, entry
, 1)) {
353 update_mmu_cache(vma
, address
, entry
);
354 lazy_mmu_prot_update(entry
);
359 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
360 struct vm_area_struct
*vma
)
362 pte_t
*src_pte
, *dst_pte
, entry
;
363 struct page
*ptepage
;
367 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
369 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= HPAGE_SIZE
) {
370 src_pte
= huge_pte_offset(src
, addr
);
373 dst_pte
= huge_pte_alloc(dst
, addr
);
376 spin_lock(&dst
->page_table_lock
);
377 spin_lock(&src
->page_table_lock
);
378 if (!pte_none(*src_pte
)) {
380 ptep_set_wrprotect(src
, addr
, src_pte
);
382 ptepage
= pte_page(entry
);
384 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
386 spin_unlock(&src
->page_table_lock
);
387 spin_unlock(&dst
->page_table_lock
);
395 void __unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
398 struct mm_struct
*mm
= vma
->vm_mm
;
399 unsigned long address
;
405 * A page gathering list, protected by per file i_mmap_lock. The
406 * lock is used to avoid list corruption from multiple unmapping
407 * of the same page since we are using page->lru.
409 LIST_HEAD(page_list
);
411 WARN_ON(!is_vm_hugetlb_page(vma
));
412 BUG_ON(start
& ~HPAGE_MASK
);
413 BUG_ON(end
& ~HPAGE_MASK
);
415 spin_lock(&mm
->page_table_lock
);
416 for (address
= start
; address
< end
; address
+= HPAGE_SIZE
) {
417 ptep
= huge_pte_offset(mm
, address
);
421 if (huge_pmd_unshare(mm
, &address
, ptep
))
424 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
428 page
= pte_page(pte
);
430 set_page_dirty(page
);
431 list_add(&page
->lru
, &page_list
);
433 spin_unlock(&mm
->page_table_lock
);
434 flush_tlb_range(vma
, start
, end
);
435 list_for_each_entry_safe(page
, tmp
, &page_list
, lru
) {
436 list_del(&page
->lru
);
441 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
445 * It is undesirable to test vma->vm_file as it should be non-null
446 * for valid hugetlb area. However, vm_file will be NULL in the error
447 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
448 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
449 * to clean up. Since no pte has actually been setup, it is safe to
450 * do nothing in this case.
453 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
454 __unmap_hugepage_range(vma
, start
, end
);
455 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
459 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
460 unsigned long address
, pte_t
*ptep
, pte_t pte
)
462 struct page
*old_page
, *new_page
;
465 old_page
= pte_page(pte
);
467 /* If no-one else is actually using this page, avoid the copy
468 * and just make the page writable */
469 avoidcopy
= (page_count(old_page
) == 1);
471 set_huge_ptep_writable(vma
, address
, ptep
);
475 page_cache_get(old_page
);
476 new_page
= alloc_huge_page(vma
, address
);
479 page_cache_release(old_page
);
483 spin_unlock(&mm
->page_table_lock
);
484 copy_huge_page(new_page
, old_page
, address
, vma
);
485 spin_lock(&mm
->page_table_lock
);
487 ptep
= huge_pte_offset(mm
, address
& HPAGE_MASK
);
488 if (likely(pte_same(*ptep
, pte
))) {
490 set_huge_pte_at(mm
, address
, ptep
,
491 make_huge_pte(vma
, new_page
, 1));
492 /* Make the old page be freed below */
495 page_cache_release(new_page
);
496 page_cache_release(old_page
);
500 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
501 unsigned long address
, pte_t
*ptep
, int write_access
)
503 int ret
= VM_FAULT_SIGBUS
;
507 struct address_space
*mapping
;
510 mapping
= vma
->vm_file
->f_mapping
;
511 idx
= ((address
- vma
->vm_start
) >> HPAGE_SHIFT
)
512 + (vma
->vm_pgoff
>> (HPAGE_SHIFT
- PAGE_SHIFT
));
515 * Use page lock to guard against racing truncation
516 * before we get page_table_lock.
519 page
= find_lock_page(mapping
, idx
);
521 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
524 if (hugetlb_get_quota(mapping
))
526 page
= alloc_huge_page(vma
, address
);
528 hugetlb_put_quota(mapping
);
532 clear_huge_page(page
, address
);
534 if (vma
->vm_flags
& VM_SHARED
) {
537 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
540 hugetlb_put_quota(mapping
);
549 spin_lock(&mm
->page_table_lock
);
550 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
555 if (!pte_none(*ptep
))
558 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
559 && (vma
->vm_flags
& VM_SHARED
)));
560 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
562 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
563 /* Optimization, do the COW without a second fault */
564 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
);
567 spin_unlock(&mm
->page_table_lock
);
573 spin_unlock(&mm
->page_table_lock
);
574 hugetlb_put_quota(mapping
);
580 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
581 unsigned long address
, int write_access
)
586 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
588 ptep
= huge_pte_alloc(mm
, address
);
593 * Serialize hugepage allocation and instantiation, so that we don't
594 * get spurious allocation failures if two CPUs race to instantiate
595 * the same page in the page cache.
597 mutex_lock(&hugetlb_instantiation_mutex
);
599 if (pte_none(entry
)) {
600 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
601 mutex_unlock(&hugetlb_instantiation_mutex
);
607 spin_lock(&mm
->page_table_lock
);
608 /* Check for a racing update before calling hugetlb_cow */
609 if (likely(pte_same(entry
, *ptep
)))
610 if (write_access
&& !pte_write(entry
))
611 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
);
612 spin_unlock(&mm
->page_table_lock
);
613 mutex_unlock(&hugetlb_instantiation_mutex
);
618 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
619 struct page
**pages
, struct vm_area_struct
**vmas
,
620 unsigned long *position
, int *length
, int i
)
622 unsigned long pfn_offset
;
623 unsigned long vaddr
= *position
;
624 int remainder
= *length
;
626 spin_lock(&mm
->page_table_lock
);
627 while (vaddr
< vma
->vm_end
&& remainder
) {
632 * Some archs (sparc64, sh*) have multiple pte_ts to
633 * each hugepage. We have to make * sure we get the
634 * first, for the page indexing below to work.
636 pte
= huge_pte_offset(mm
, vaddr
& HPAGE_MASK
);
638 if (!pte
|| pte_none(*pte
)) {
641 spin_unlock(&mm
->page_table_lock
);
642 ret
= hugetlb_fault(mm
, vma
, vaddr
, 0);
643 spin_lock(&mm
->page_table_lock
);
644 if (!(ret
& VM_FAULT_MAJOR
))
653 pfn_offset
= (vaddr
& ~HPAGE_MASK
) >> PAGE_SHIFT
;
654 page
= pte_page(*pte
);
658 pages
[i
] = page
+ pfn_offset
;
668 if (vaddr
< vma
->vm_end
&& remainder
&&
669 pfn_offset
< HPAGE_SIZE
/PAGE_SIZE
) {
671 * We use pfn_offset to avoid touching the pageframes
672 * of this compound page.
677 spin_unlock(&mm
->page_table_lock
);
684 void hugetlb_change_protection(struct vm_area_struct
*vma
,
685 unsigned long address
, unsigned long end
, pgprot_t newprot
)
687 struct mm_struct
*mm
= vma
->vm_mm
;
688 unsigned long start
= address
;
692 BUG_ON(address
>= end
);
693 flush_cache_range(vma
, address
, end
);
695 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
696 spin_lock(&mm
->page_table_lock
);
697 for (; address
< end
; address
+= HPAGE_SIZE
) {
698 ptep
= huge_pte_offset(mm
, address
);
701 if (huge_pmd_unshare(mm
, &address
, ptep
))
703 if (!pte_none(*ptep
)) {
704 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
705 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
706 set_huge_pte_at(mm
, address
, ptep
, pte
);
707 lazy_mmu_prot_update(pte
);
710 spin_unlock(&mm
->page_table_lock
);
711 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
713 flush_tlb_range(vma
, start
, end
);
717 struct list_head link
;
722 static long region_add(struct list_head
*head
, long f
, long t
)
724 struct file_region
*rg
, *nrg
, *trg
;
726 /* Locate the region we are either in or before. */
727 list_for_each_entry(rg
, head
, link
)
731 /* Round our left edge to the current segment if it encloses us. */
735 /* Check for and consume any regions we now overlap with. */
737 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
738 if (&rg
->link
== head
)
743 /* If this area reaches higher then extend our area to
744 * include it completely. If this is not the first area
745 * which we intend to reuse, free it. */
758 static long region_chg(struct list_head
*head
, long f
, long t
)
760 struct file_region
*rg
, *nrg
;
763 /* Locate the region we are before or in. */
764 list_for_each_entry(rg
, head
, link
)
768 /* If we are below the current region then a new region is required.
769 * Subtle, allocate a new region at the position but make it zero
770 * size such that we can guarentee to record the reservation. */
771 if (&rg
->link
== head
|| t
< rg
->from
) {
772 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
777 INIT_LIST_HEAD(&nrg
->link
);
778 list_add(&nrg
->link
, rg
->link
.prev
);
783 /* Round our left edge to the current segment if it encloses us. */
788 /* Check for and consume any regions we now overlap with. */
789 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
790 if (&rg
->link
== head
)
795 /* We overlap with this area, if it extends futher than
796 * us then we must extend ourselves. Account for its
797 * existing reservation. */
802 chg
-= rg
->to
- rg
->from
;
807 static long region_truncate(struct list_head
*head
, long end
)
809 struct file_region
*rg
, *trg
;
812 /* Locate the region we are either in or before. */
813 list_for_each_entry(rg
, head
, link
)
816 if (&rg
->link
== head
)
819 /* If we are in the middle of a region then adjust it. */
820 if (end
> rg
->from
) {
823 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
826 /* Drop any remaining regions. */
827 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
828 if (&rg
->link
== head
)
830 chg
+= rg
->to
- rg
->from
;
837 static int hugetlb_acct_memory(long delta
)
841 spin_lock(&hugetlb_lock
);
842 if ((delta
+ resv_huge_pages
) <= free_huge_pages
) {
843 resv_huge_pages
+= delta
;
846 spin_unlock(&hugetlb_lock
);
850 int hugetlb_reserve_pages(struct inode
*inode
, long from
, long to
)
854 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
858 * When cpuset is configured, it breaks the strict hugetlb page
859 * reservation as the accounting is done on a global variable. Such
860 * reservation is completely rubbish in the presence of cpuset because
861 * the reservation is not checked against page availability for the
862 * current cpuset. Application can still potentially OOM'ed by kernel
863 * with lack of free htlb page in cpuset that the task is in.
864 * Attempt to enforce strict accounting with cpuset is almost
865 * impossible (or too ugly) because cpuset is too fluid that
866 * task or memory node can be dynamically moved between cpusets.
868 * The change of semantics for shared hugetlb mapping with cpuset is
869 * undesirable. However, in order to preserve some of the semantics,
870 * we fall back to check against current free page availability as
871 * a best attempt and hopefully to minimize the impact of changing
872 * semantics that cpuset has.
874 if (chg
> cpuset_mems_nr(free_huge_pages_node
))
877 ret
= hugetlb_acct_memory(chg
);
880 region_add(&inode
->i_mapping
->private_list
, from
, to
);
884 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
886 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
887 hugetlb_acct_memory(freed
- chg
);