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 static DEFINE_SPINLOCK(nid_lock
);
109 spin_lock(&nid_lock
);
110 nid
= next_node(prev_nid
, node_online_map
);
111 if (nid
== MAX_NUMNODES
)
112 nid
= first_node(node_online_map
);
114 spin_unlock(&nid_lock
);
116 page
= alloc_pages_node(nid
, GFP_HIGHUSER
|__GFP_COMP
|__GFP_NOWARN
,
119 set_compound_page_dtor(page
, free_huge_page
);
120 spin_lock(&hugetlb_lock
);
122 nr_huge_pages_node
[page_to_nid(page
)]++;
123 spin_unlock(&hugetlb_lock
);
124 put_page(page
); /* free it into the hugepage allocator */
130 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
135 spin_lock(&hugetlb_lock
);
136 if (vma
->vm_flags
& VM_MAYSHARE
)
138 else if (free_huge_pages
<= resv_huge_pages
)
141 page
= dequeue_huge_page(vma
, addr
);
145 spin_unlock(&hugetlb_lock
);
146 set_page_refcounted(page
);
150 if (vma
->vm_flags
& VM_MAYSHARE
)
152 spin_unlock(&hugetlb_lock
);
156 static int __init
hugetlb_init(void)
160 if (HPAGE_SHIFT
== 0)
163 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
164 INIT_LIST_HEAD(&hugepage_freelists
[i
]);
166 for (i
= 0; i
< max_huge_pages
; ++i
) {
167 if (!alloc_fresh_huge_page())
170 max_huge_pages
= free_huge_pages
= nr_huge_pages
= i
;
171 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages
);
174 module_init(hugetlb_init
);
176 static int __init
hugetlb_setup(char *s
)
178 if (sscanf(s
, "%lu", &max_huge_pages
) <= 0)
182 __setup("hugepages=", hugetlb_setup
);
184 static unsigned int cpuset_mems_nr(unsigned int *array
)
189 for_each_node_mask(node
, cpuset_current_mems_allowed
)
196 static void update_and_free_page(struct page
*page
)
200 nr_huge_pages_node
[page_to_nid(page
)]--;
201 for (i
= 0; i
< (HPAGE_SIZE
/ PAGE_SIZE
); i
++) {
202 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
| 1 << PG_referenced
|
203 1 << PG_dirty
| 1 << PG_active
| 1 << PG_reserved
|
204 1 << PG_private
| 1<< PG_writeback
);
206 page
[1].lru
.next
= NULL
;
207 set_page_refcounted(page
);
208 __free_pages(page
, HUGETLB_PAGE_ORDER
);
211 #ifdef CONFIG_HIGHMEM
212 static void try_to_free_low(unsigned long count
)
216 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
217 struct page
*page
, *next
;
218 list_for_each_entry_safe(page
, next
, &hugepage_freelists
[i
], lru
) {
219 if (PageHighMem(page
))
221 list_del(&page
->lru
);
222 update_and_free_page(page
);
224 free_huge_pages_node
[page_to_nid(page
)]--;
225 if (count
>= nr_huge_pages
)
231 static inline void try_to_free_low(unsigned long count
)
236 static unsigned long set_max_huge_pages(unsigned long count
)
238 while (count
> nr_huge_pages
) {
239 if (!alloc_fresh_huge_page())
240 return nr_huge_pages
;
242 if (count
>= nr_huge_pages
)
243 return nr_huge_pages
;
245 spin_lock(&hugetlb_lock
);
246 count
= max(count
, resv_huge_pages
);
247 try_to_free_low(count
);
248 while (count
< nr_huge_pages
) {
249 struct page
*page
= dequeue_huge_page(NULL
, 0);
252 update_and_free_page(page
);
254 spin_unlock(&hugetlb_lock
);
255 return nr_huge_pages
;
258 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
259 struct file
*file
, void __user
*buffer
,
260 size_t *length
, loff_t
*ppos
)
262 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
263 max_huge_pages
= set_max_huge_pages(max_huge_pages
);
266 #endif /* CONFIG_SYSCTL */
268 int hugetlb_report_meminfo(char *buf
)
271 "HugePages_Total: %5lu\n"
272 "HugePages_Free: %5lu\n"
273 "HugePages_Rsvd: %5lu\n"
274 "Hugepagesize: %5lu kB\n",
281 int hugetlb_report_node_meminfo(int nid
, char *buf
)
284 "Node %d HugePages_Total: %5u\n"
285 "Node %d HugePages_Free: %5u\n",
286 nid
, nr_huge_pages_node
[nid
],
287 nid
, free_huge_pages_node
[nid
]);
290 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
291 unsigned long hugetlb_total_pages(void)
293 return nr_huge_pages
* (HPAGE_SIZE
/ PAGE_SIZE
);
297 * We cannot handle pagefaults against hugetlb pages at all. They cause
298 * handle_mm_fault() to try to instantiate regular-sized pages in the
299 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
302 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
308 struct vm_operations_struct hugetlb_vm_ops
= {
309 .fault
= hugetlb_vm_op_fault
,
312 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
319 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
321 entry
= pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
323 entry
= pte_mkyoung(entry
);
324 entry
= pte_mkhuge(entry
);
329 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
330 unsigned long address
, pte_t
*ptep
)
334 entry
= pte_mkwrite(pte_mkdirty(*ptep
));
335 if (ptep_set_access_flags(vma
, address
, ptep
, entry
, 1)) {
336 update_mmu_cache(vma
, address
, entry
);
337 lazy_mmu_prot_update(entry
);
342 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
343 struct vm_area_struct
*vma
)
345 pte_t
*src_pte
, *dst_pte
, entry
;
346 struct page
*ptepage
;
350 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
352 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= HPAGE_SIZE
) {
353 src_pte
= huge_pte_offset(src
, addr
);
356 dst_pte
= huge_pte_alloc(dst
, addr
);
359 spin_lock(&dst
->page_table_lock
);
360 spin_lock(&src
->page_table_lock
);
361 if (!pte_none(*src_pte
)) {
363 ptep_set_wrprotect(src
, addr
, src_pte
);
365 ptepage
= pte_page(entry
);
367 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
369 spin_unlock(&src
->page_table_lock
);
370 spin_unlock(&dst
->page_table_lock
);
378 void __unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
381 struct mm_struct
*mm
= vma
->vm_mm
;
382 unsigned long address
;
388 * A page gathering list, protected by per file i_mmap_lock. The
389 * lock is used to avoid list corruption from multiple unmapping
390 * of the same page since we are using page->lru.
392 LIST_HEAD(page_list
);
394 WARN_ON(!is_vm_hugetlb_page(vma
));
395 BUG_ON(start
& ~HPAGE_MASK
);
396 BUG_ON(end
& ~HPAGE_MASK
);
398 spin_lock(&mm
->page_table_lock
);
399 for (address
= start
; address
< end
; address
+= HPAGE_SIZE
) {
400 ptep
= huge_pte_offset(mm
, address
);
404 if (huge_pmd_unshare(mm
, &address
, ptep
))
407 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
411 page
= pte_page(pte
);
413 set_page_dirty(page
);
414 list_add(&page
->lru
, &page_list
);
416 spin_unlock(&mm
->page_table_lock
);
417 flush_tlb_range(vma
, start
, end
);
418 list_for_each_entry_safe(page
, tmp
, &page_list
, lru
) {
419 list_del(&page
->lru
);
424 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
428 * It is undesirable to test vma->vm_file as it should be non-null
429 * for valid hugetlb area. However, vm_file will be NULL in the error
430 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
431 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
432 * to clean up. Since no pte has actually been setup, it is safe to
433 * do nothing in this case.
436 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
437 __unmap_hugepage_range(vma
, start
, end
);
438 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
442 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
443 unsigned long address
, pte_t
*ptep
, pte_t pte
)
445 struct page
*old_page
, *new_page
;
448 old_page
= pte_page(pte
);
450 /* If no-one else is actually using this page, avoid the copy
451 * and just make the page writable */
452 avoidcopy
= (page_count(old_page
) == 1);
454 set_huge_ptep_writable(vma
, address
, ptep
);
458 page_cache_get(old_page
);
459 new_page
= alloc_huge_page(vma
, address
);
462 page_cache_release(old_page
);
466 spin_unlock(&mm
->page_table_lock
);
467 copy_huge_page(new_page
, old_page
, address
, vma
);
468 spin_lock(&mm
->page_table_lock
);
470 ptep
= huge_pte_offset(mm
, address
& HPAGE_MASK
);
471 if (likely(pte_same(*ptep
, pte
))) {
473 set_huge_pte_at(mm
, address
, ptep
,
474 make_huge_pte(vma
, new_page
, 1));
475 /* Make the old page be freed below */
478 page_cache_release(new_page
);
479 page_cache_release(old_page
);
483 int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
484 unsigned long address
, pte_t
*ptep
, int write_access
)
486 int ret
= VM_FAULT_SIGBUS
;
490 struct address_space
*mapping
;
493 mapping
= vma
->vm_file
->f_mapping
;
494 idx
= ((address
- vma
->vm_start
) >> HPAGE_SHIFT
)
495 + (vma
->vm_pgoff
>> (HPAGE_SHIFT
- PAGE_SHIFT
));
498 * Use page lock to guard against racing truncation
499 * before we get page_table_lock.
502 page
= find_lock_page(mapping
, idx
);
504 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
507 if (hugetlb_get_quota(mapping
))
509 page
= alloc_huge_page(vma
, address
);
511 hugetlb_put_quota(mapping
);
515 clear_huge_page(page
, address
);
517 if (vma
->vm_flags
& VM_SHARED
) {
520 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
523 hugetlb_put_quota(mapping
);
532 spin_lock(&mm
->page_table_lock
);
533 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
538 if (!pte_none(*ptep
))
541 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
542 && (vma
->vm_flags
& VM_SHARED
)));
543 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
545 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
546 /* Optimization, do the COW without a second fault */
547 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
);
550 spin_unlock(&mm
->page_table_lock
);
556 spin_unlock(&mm
->page_table_lock
);
557 hugetlb_put_quota(mapping
);
563 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
564 unsigned long address
, int write_access
)
569 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
571 ptep
= huge_pte_alloc(mm
, address
);
576 * Serialize hugepage allocation and instantiation, so that we don't
577 * get spurious allocation failures if two CPUs race to instantiate
578 * the same page in the page cache.
580 mutex_lock(&hugetlb_instantiation_mutex
);
582 if (pte_none(entry
)) {
583 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
584 mutex_unlock(&hugetlb_instantiation_mutex
);
590 spin_lock(&mm
->page_table_lock
);
591 /* Check for a racing update before calling hugetlb_cow */
592 if (likely(pte_same(entry
, *ptep
)))
593 if (write_access
&& !pte_write(entry
))
594 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
);
595 spin_unlock(&mm
->page_table_lock
);
596 mutex_unlock(&hugetlb_instantiation_mutex
);
601 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
602 struct page
**pages
, struct vm_area_struct
**vmas
,
603 unsigned long *position
, int *length
, int i
)
605 unsigned long pfn_offset
;
606 unsigned long vaddr
= *position
;
607 int remainder
= *length
;
609 spin_lock(&mm
->page_table_lock
);
610 while (vaddr
< vma
->vm_end
&& remainder
) {
615 * Some archs (sparc64, sh*) have multiple pte_ts to
616 * each hugepage. We have to make * sure we get the
617 * first, for the page indexing below to work.
619 pte
= huge_pte_offset(mm
, vaddr
& HPAGE_MASK
);
621 if (!pte
|| pte_none(*pte
)) {
624 spin_unlock(&mm
->page_table_lock
);
625 ret
= hugetlb_fault(mm
, vma
, vaddr
, 0);
626 spin_lock(&mm
->page_table_lock
);
627 if (!(ret
& VM_FAULT_ERROR
))
636 pfn_offset
= (vaddr
& ~HPAGE_MASK
) >> PAGE_SHIFT
;
637 page
= pte_page(*pte
);
641 pages
[i
] = page
+ pfn_offset
;
651 if (vaddr
< vma
->vm_end
&& remainder
&&
652 pfn_offset
< HPAGE_SIZE
/PAGE_SIZE
) {
654 * We use pfn_offset to avoid touching the pageframes
655 * of this compound page.
660 spin_unlock(&mm
->page_table_lock
);
667 void hugetlb_change_protection(struct vm_area_struct
*vma
,
668 unsigned long address
, unsigned long end
, pgprot_t newprot
)
670 struct mm_struct
*mm
= vma
->vm_mm
;
671 unsigned long start
= address
;
675 BUG_ON(address
>= end
);
676 flush_cache_range(vma
, address
, end
);
678 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
679 spin_lock(&mm
->page_table_lock
);
680 for (; address
< end
; address
+= HPAGE_SIZE
) {
681 ptep
= huge_pte_offset(mm
, address
);
684 if (huge_pmd_unshare(mm
, &address
, ptep
))
686 if (!pte_none(*ptep
)) {
687 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
688 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
689 set_huge_pte_at(mm
, address
, ptep
, pte
);
690 lazy_mmu_prot_update(pte
);
693 spin_unlock(&mm
->page_table_lock
);
694 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
696 flush_tlb_range(vma
, start
, end
);
700 struct list_head link
;
705 static long region_add(struct list_head
*head
, long f
, long t
)
707 struct file_region
*rg
, *nrg
, *trg
;
709 /* Locate the region we are either in or before. */
710 list_for_each_entry(rg
, head
, link
)
714 /* Round our left edge to the current segment if it encloses us. */
718 /* Check for and consume any regions we now overlap with. */
720 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
721 if (&rg
->link
== head
)
726 /* If this area reaches higher then extend our area to
727 * include it completely. If this is not the first area
728 * which we intend to reuse, free it. */
741 static long region_chg(struct list_head
*head
, long f
, long t
)
743 struct file_region
*rg
, *nrg
;
746 /* Locate the region we are before or in. */
747 list_for_each_entry(rg
, head
, link
)
751 /* If we are below the current region then a new region is required.
752 * Subtle, allocate a new region at the position but make it zero
753 * size such that we can guarentee to record the reservation. */
754 if (&rg
->link
== head
|| t
< rg
->from
) {
755 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
760 INIT_LIST_HEAD(&nrg
->link
);
761 list_add(&nrg
->link
, rg
->link
.prev
);
766 /* Round our left edge to the current segment if it encloses us. */
771 /* Check for and consume any regions we now overlap with. */
772 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
773 if (&rg
->link
== head
)
778 /* We overlap with this area, if it extends futher than
779 * us then we must extend ourselves. Account for its
780 * existing reservation. */
785 chg
-= rg
->to
- rg
->from
;
790 static long region_truncate(struct list_head
*head
, long end
)
792 struct file_region
*rg
, *trg
;
795 /* Locate the region we are either in or before. */
796 list_for_each_entry(rg
, head
, link
)
799 if (&rg
->link
== head
)
802 /* If we are in the middle of a region then adjust it. */
803 if (end
> rg
->from
) {
806 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
809 /* Drop any remaining regions. */
810 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
811 if (&rg
->link
== head
)
813 chg
+= rg
->to
- rg
->from
;
820 static int hugetlb_acct_memory(long delta
)
824 spin_lock(&hugetlb_lock
);
825 if ((delta
+ resv_huge_pages
) <= free_huge_pages
) {
826 resv_huge_pages
+= delta
;
829 spin_unlock(&hugetlb_lock
);
833 int hugetlb_reserve_pages(struct inode
*inode
, long from
, long to
)
837 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
841 * When cpuset is configured, it breaks the strict hugetlb page
842 * reservation as the accounting is done on a global variable. Such
843 * reservation is completely rubbish in the presence of cpuset because
844 * the reservation is not checked against page availability for the
845 * current cpuset. Application can still potentially OOM'ed by kernel
846 * with lack of free htlb page in cpuset that the task is in.
847 * Attempt to enforce strict accounting with cpuset is almost
848 * impossible (or too ugly) because cpuset is too fluid that
849 * task or memory node can be dynamically moved between cpusets.
851 * The change of semantics for shared hugetlb mapping with cpuset is
852 * undesirable. However, in order to preserve some of the semantics,
853 * we fall back to check against current free page availability as
854 * a best attempt and hopefully to minimize the impact of changing
855 * semantics that cpuset has.
857 if (chg
> cpuset_mems_nr(free_huge_pages_node
))
860 ret
= hugetlb_acct_memory(chg
);
863 region_add(&inode
->i_mapping
->private_list
, from
, to
);
867 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
869 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
870 hugetlb_acct_memory(freed
- chg
);