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
];
31 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
33 static DEFINE_SPINLOCK(hugetlb_lock
);
35 static void clear_huge_page(struct page
*page
, unsigned long addr
)
40 for (i
= 0; i
< (HPAGE_SIZE
/PAGE_SIZE
); i
++) {
42 clear_user_highpage(page
+ i
, addr
);
46 static void copy_huge_page(struct page
*dst
, struct page
*src
,
47 unsigned long addr
, struct vm_area_struct
*vma
)
52 for (i
= 0; i
< HPAGE_SIZE
/PAGE_SIZE
; i
++) {
54 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
58 static void enqueue_huge_page(struct page
*page
)
60 int nid
= page_to_nid(page
);
61 list_add(&page
->lru
, &hugepage_freelists
[nid
]);
63 free_huge_pages_node
[nid
]++;
66 static struct page
*dequeue_huge_page(struct vm_area_struct
*vma
,
67 unsigned long address
)
69 int nid
= numa_node_id();
70 struct page
*page
= NULL
;
71 struct zonelist
*zonelist
= huge_zonelist(vma
, address
);
74 for (z
= zonelist
->zones
; *z
; z
++) {
75 nid
= zone_to_nid(*z
);
76 if (cpuset_zone_allowed_softwall(*z
, GFP_HIGHUSER
) &&
77 !list_empty(&hugepage_freelists
[nid
]))
82 page
= list_entry(hugepage_freelists
[nid
].next
,
86 free_huge_pages_node
[nid
]--;
91 static void free_huge_page(struct page
*page
)
93 BUG_ON(page_count(page
));
95 INIT_LIST_HEAD(&page
->lru
);
97 spin_lock(&hugetlb_lock
);
98 enqueue_huge_page(page
);
99 spin_unlock(&hugetlb_lock
);
102 static int alloc_fresh_huge_page(void)
106 page
= alloc_pages_node(nid
, GFP_HIGHUSER
|__GFP_COMP
|__GFP_NOWARN
,
108 nid
= next_node(nid
, node_online_map
);
109 if (nid
== MAX_NUMNODES
)
110 nid
= first_node(node_online_map
);
112 set_compound_page_dtor(page
, free_huge_page
);
113 spin_lock(&hugetlb_lock
);
115 nr_huge_pages_node
[page_to_nid(page
)]++;
116 spin_unlock(&hugetlb_lock
);
117 put_page(page
); /* free it into the hugepage allocator */
123 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
128 spin_lock(&hugetlb_lock
);
129 if (vma
->vm_flags
& VM_MAYSHARE
)
131 else if (free_huge_pages
<= resv_huge_pages
)
134 page
= dequeue_huge_page(vma
, addr
);
138 spin_unlock(&hugetlb_lock
);
139 set_page_refcounted(page
);
143 if (vma
->vm_flags
& VM_MAYSHARE
)
145 spin_unlock(&hugetlb_lock
);
149 static int __init
hugetlb_init(void)
153 if (HPAGE_SHIFT
== 0)
156 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
157 INIT_LIST_HEAD(&hugepage_freelists
[i
]);
159 for (i
= 0; i
< max_huge_pages
; ++i
) {
160 if (!alloc_fresh_huge_page())
163 max_huge_pages
= free_huge_pages
= nr_huge_pages
= i
;
164 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages
);
167 module_init(hugetlb_init
);
169 static int __init
hugetlb_setup(char *s
)
171 if (sscanf(s
, "%lu", &max_huge_pages
) <= 0)
175 __setup("hugepages=", hugetlb_setup
);
178 static void update_and_free_page(struct page
*page
)
182 nr_huge_pages_node
[page_to_nid(page
)]--;
183 for (i
= 0; i
< (HPAGE_SIZE
/ PAGE_SIZE
); i
++) {
184 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
| 1 << PG_referenced
|
185 1 << PG_dirty
| 1 << PG_active
| 1 << PG_reserved
|
186 1 << PG_private
| 1<< PG_writeback
);
188 page
[1].lru
.next
= NULL
;
189 set_page_refcounted(page
);
190 __free_pages(page
, HUGETLB_PAGE_ORDER
);
193 #ifdef CONFIG_HIGHMEM
194 static void try_to_free_low(unsigned long count
)
198 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
199 struct page
*page
, *next
;
200 list_for_each_entry_safe(page
, next
, &hugepage_freelists
[i
], lru
) {
201 if (PageHighMem(page
))
203 list_del(&page
->lru
);
204 update_and_free_page(page
);
206 free_huge_pages_node
[page_to_nid(page
)]--;
207 if (count
>= nr_huge_pages
)
213 static inline void try_to_free_low(unsigned long count
)
218 static unsigned long set_max_huge_pages(unsigned long count
)
220 while (count
> nr_huge_pages
) {
221 if (!alloc_fresh_huge_page())
222 return nr_huge_pages
;
224 if (count
>= nr_huge_pages
)
225 return nr_huge_pages
;
227 spin_lock(&hugetlb_lock
);
228 count
= max(count
, resv_huge_pages
);
229 try_to_free_low(count
);
230 while (count
< nr_huge_pages
) {
231 struct page
*page
= dequeue_huge_page(NULL
, 0);
234 update_and_free_page(page
);
236 spin_unlock(&hugetlb_lock
);
237 return nr_huge_pages
;
240 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
241 struct file
*file
, void __user
*buffer
,
242 size_t *length
, loff_t
*ppos
)
244 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
245 max_huge_pages
= set_max_huge_pages(max_huge_pages
);
248 #endif /* CONFIG_SYSCTL */
250 int hugetlb_report_meminfo(char *buf
)
253 "HugePages_Total: %5lu\n"
254 "HugePages_Free: %5lu\n"
255 "HugePages_Rsvd: %5lu\n"
256 "Hugepagesize: %5lu kB\n",
263 int hugetlb_report_node_meminfo(int nid
, char *buf
)
266 "Node %d HugePages_Total: %5u\n"
267 "Node %d HugePages_Free: %5u\n",
268 nid
, nr_huge_pages_node
[nid
],
269 nid
, free_huge_pages_node
[nid
]);
272 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
273 unsigned long hugetlb_total_pages(void)
275 return nr_huge_pages
* (HPAGE_SIZE
/ PAGE_SIZE
);
279 * We cannot handle pagefaults against hugetlb pages at all. They cause
280 * handle_mm_fault() to try to instantiate regular-sized pages in the
281 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
284 static struct page
*hugetlb_nopage(struct vm_area_struct
*vma
,
285 unsigned long address
, int *unused
)
291 struct vm_operations_struct hugetlb_vm_ops
= {
292 .nopage
= hugetlb_nopage
,
295 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
302 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
304 entry
= pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
306 entry
= pte_mkyoung(entry
);
307 entry
= pte_mkhuge(entry
);
312 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
313 unsigned long address
, pte_t
*ptep
)
317 entry
= pte_mkwrite(pte_mkdirty(*ptep
));
318 ptep_set_access_flags(vma
, address
, ptep
, entry
, 1);
319 update_mmu_cache(vma
, address
, entry
);
320 lazy_mmu_prot_update(entry
);
324 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
325 struct vm_area_struct
*vma
)
327 pte_t
*src_pte
, *dst_pte
, entry
;
328 struct page
*ptepage
;
332 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
334 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= HPAGE_SIZE
) {
335 src_pte
= huge_pte_offset(src
, addr
);
338 dst_pte
= huge_pte_alloc(dst
, addr
);
341 spin_lock(&dst
->page_table_lock
);
342 spin_lock(&src
->page_table_lock
);
343 if (!pte_none(*src_pte
)) {
345 ptep_set_wrprotect(src
, addr
, src_pte
);
347 ptepage
= pte_page(entry
);
349 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
351 spin_unlock(&src
->page_table_lock
);
352 spin_unlock(&dst
->page_table_lock
);
360 void __unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
363 struct mm_struct
*mm
= vma
->vm_mm
;
364 unsigned long address
;
370 * A page gathering list, protected by per file i_mmap_lock. The
371 * lock is used to avoid list corruption from multiple unmapping
372 * of the same page since we are using page->lru.
374 LIST_HEAD(page_list
);
376 WARN_ON(!is_vm_hugetlb_page(vma
));
377 BUG_ON(start
& ~HPAGE_MASK
);
378 BUG_ON(end
& ~HPAGE_MASK
);
380 spin_lock(&mm
->page_table_lock
);
381 for (address
= start
; address
< end
; address
+= HPAGE_SIZE
) {
382 ptep
= huge_pte_offset(mm
, address
);
386 if (huge_pmd_unshare(mm
, &address
, ptep
))
389 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
393 page
= pte_page(pte
);
395 set_page_dirty(page
);
396 list_add(&page
->lru
, &page_list
);
398 spin_unlock(&mm
->page_table_lock
);
399 flush_tlb_range(vma
, start
, end
);
400 list_for_each_entry_safe(page
, tmp
, &page_list
, lru
) {
401 list_del(&page
->lru
);
406 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
410 * It is undesirable to test vma->vm_file as it should be non-null
411 * for valid hugetlb area. However, vm_file will be NULL in the error
412 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
413 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
414 * to clean up. Since no pte has actually been setup, it is safe to
415 * do nothing in this case.
418 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
419 __unmap_hugepage_range(vma
, start
, end
);
420 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
424 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
425 unsigned long address
, pte_t
*ptep
, pte_t pte
)
427 struct page
*old_page
, *new_page
;
430 old_page
= pte_page(pte
);
432 /* If no-one else is actually using this page, avoid the copy
433 * and just make the page writable */
434 avoidcopy
= (page_count(old_page
) == 1);
436 set_huge_ptep_writable(vma
, address
, ptep
);
437 return VM_FAULT_MINOR
;
440 page_cache_get(old_page
);
441 new_page
= alloc_huge_page(vma
, address
);
444 page_cache_release(old_page
);
448 spin_unlock(&mm
->page_table_lock
);
449 copy_huge_page(new_page
, old_page
, address
, vma
);
450 spin_lock(&mm
->page_table_lock
);
452 ptep
= huge_pte_offset(mm
, address
& HPAGE_MASK
);
453 if (likely(pte_same(*ptep
, pte
))) {
455 set_huge_pte_at(mm
, address
, ptep
,
456 make_huge_pte(vma
, new_page
, 1));
457 /* Make the old page be freed below */
460 page_cache_release(new_page
);
461 page_cache_release(old_page
);
462 return VM_FAULT_MINOR
;
465 int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
466 unsigned long address
, pte_t
*ptep
, int write_access
)
468 int ret
= VM_FAULT_SIGBUS
;
472 struct address_space
*mapping
;
475 mapping
= vma
->vm_file
->f_mapping
;
476 idx
= ((address
- vma
->vm_start
) >> HPAGE_SHIFT
)
477 + (vma
->vm_pgoff
>> (HPAGE_SHIFT
- PAGE_SHIFT
));
480 * Use page lock to guard against racing truncation
481 * before we get page_table_lock.
484 page
= find_lock_page(mapping
, idx
);
486 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
489 if (hugetlb_get_quota(mapping
))
491 page
= alloc_huge_page(vma
, address
);
493 hugetlb_put_quota(mapping
);
497 clear_huge_page(page
, address
);
499 if (vma
->vm_flags
& VM_SHARED
) {
502 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
505 hugetlb_put_quota(mapping
);
514 spin_lock(&mm
->page_table_lock
);
515 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
519 ret
= VM_FAULT_MINOR
;
520 if (!pte_none(*ptep
))
523 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
524 && (vma
->vm_flags
& VM_SHARED
)));
525 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
527 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
528 /* Optimization, do the COW without a second fault */
529 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
);
532 spin_unlock(&mm
->page_table_lock
);
538 spin_unlock(&mm
->page_table_lock
);
539 hugetlb_put_quota(mapping
);
545 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
546 unsigned long address
, int write_access
)
551 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
553 ptep
= huge_pte_alloc(mm
, address
);
558 * Serialize hugepage allocation and instantiation, so that we don't
559 * get spurious allocation failures if two CPUs race to instantiate
560 * the same page in the page cache.
562 mutex_lock(&hugetlb_instantiation_mutex
);
564 if (pte_none(entry
)) {
565 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
566 mutex_unlock(&hugetlb_instantiation_mutex
);
570 ret
= VM_FAULT_MINOR
;
572 spin_lock(&mm
->page_table_lock
);
573 /* Check for a racing update before calling hugetlb_cow */
574 if (likely(pte_same(entry
, *ptep
)))
575 if (write_access
&& !pte_write(entry
))
576 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
);
577 spin_unlock(&mm
->page_table_lock
);
578 mutex_unlock(&hugetlb_instantiation_mutex
);
583 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
584 struct page
**pages
, struct vm_area_struct
**vmas
,
585 unsigned long *position
, int *length
, int i
)
587 unsigned long pfn_offset
;
588 unsigned long vaddr
= *position
;
589 int remainder
= *length
;
591 spin_lock(&mm
->page_table_lock
);
592 while (vaddr
< vma
->vm_end
&& remainder
) {
597 * Some archs (sparc64, sh*) have multiple pte_ts to
598 * each hugepage. We have to make * sure we get the
599 * first, for the page indexing below to work.
601 pte
= huge_pte_offset(mm
, vaddr
& HPAGE_MASK
);
603 if (!pte
|| pte_none(*pte
)) {
606 spin_unlock(&mm
->page_table_lock
);
607 ret
= hugetlb_fault(mm
, vma
, vaddr
, 0);
608 spin_lock(&mm
->page_table_lock
);
609 if (ret
== VM_FAULT_MINOR
)
618 pfn_offset
= (vaddr
& ~HPAGE_MASK
) >> PAGE_SHIFT
;
619 page
= pte_page(*pte
);
623 pages
[i
] = page
+ pfn_offset
;
633 if (vaddr
< vma
->vm_end
&& remainder
&&
634 pfn_offset
< HPAGE_SIZE
/PAGE_SIZE
) {
636 * We use pfn_offset to avoid touching the pageframes
637 * of this compound page.
642 spin_unlock(&mm
->page_table_lock
);
649 void hugetlb_change_protection(struct vm_area_struct
*vma
,
650 unsigned long address
, unsigned long end
, pgprot_t newprot
)
652 struct mm_struct
*mm
= vma
->vm_mm
;
653 unsigned long start
= address
;
657 BUG_ON(address
>= end
);
658 flush_cache_range(vma
, address
, end
);
660 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
661 spin_lock(&mm
->page_table_lock
);
662 for (; address
< end
; address
+= HPAGE_SIZE
) {
663 ptep
= huge_pte_offset(mm
, address
);
666 if (huge_pmd_unshare(mm
, &address
, ptep
))
668 if (!pte_none(*ptep
)) {
669 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
670 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
671 set_huge_pte_at(mm
, address
, ptep
, pte
);
672 lazy_mmu_prot_update(pte
);
675 spin_unlock(&mm
->page_table_lock
);
676 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
678 flush_tlb_range(vma
, start
, end
);
682 struct list_head link
;
687 static long region_add(struct list_head
*head
, long f
, long t
)
689 struct file_region
*rg
, *nrg
, *trg
;
691 /* Locate the region we are either in or before. */
692 list_for_each_entry(rg
, head
, link
)
696 /* Round our left edge to the current segment if it encloses us. */
700 /* Check for and consume any regions we now overlap with. */
702 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
703 if (&rg
->link
== head
)
708 /* If this area reaches higher then extend our area to
709 * include it completely. If this is not the first area
710 * which we intend to reuse, free it. */
723 static long region_chg(struct list_head
*head
, long f
, long t
)
725 struct file_region
*rg
, *nrg
;
728 /* Locate the region we are before or in. */
729 list_for_each_entry(rg
, head
, link
)
733 /* If we are below the current region then a new region is required.
734 * Subtle, allocate a new region at the position but make it zero
735 * size such that we can guarentee to record the reservation. */
736 if (&rg
->link
== head
|| t
< rg
->from
) {
737 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
742 INIT_LIST_HEAD(&nrg
->link
);
743 list_add(&nrg
->link
, rg
->link
.prev
);
748 /* Round our left edge to the current segment if it encloses us. */
753 /* Check for and consume any regions we now overlap with. */
754 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
755 if (&rg
->link
== head
)
760 /* We overlap with this area, if it extends futher than
761 * us then we must extend ourselves. Account for its
762 * existing reservation. */
767 chg
-= rg
->to
- rg
->from
;
772 static long region_truncate(struct list_head
*head
, long end
)
774 struct file_region
*rg
, *trg
;
777 /* Locate the region we are either in or before. */
778 list_for_each_entry(rg
, head
, link
)
781 if (&rg
->link
== head
)
784 /* If we are in the middle of a region then adjust it. */
785 if (end
> rg
->from
) {
788 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
791 /* Drop any remaining regions. */
792 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
793 if (&rg
->link
== head
)
795 chg
+= rg
->to
- rg
->from
;
802 static int hugetlb_acct_memory(long delta
)
806 spin_lock(&hugetlb_lock
);
807 if ((delta
+ resv_huge_pages
) <= free_huge_pages
) {
808 resv_huge_pages
+= delta
;
811 spin_unlock(&hugetlb_lock
);
815 int hugetlb_reserve_pages(struct inode
*inode
, long from
, long to
)
819 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
822 ret
= hugetlb_acct_memory(chg
);
825 region_add(&inode
->i_mapping
->private_list
, from
, to
);
829 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
831 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
832 hugetlb_acct_memory(freed
- chg
);