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
, reserved_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
,
52 for (i
= 0; i
< HPAGE_SIZE
/PAGE_SIZE
; i
++) {
54 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
);
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
= (*z
)->zone_pgdat
->node_id
;
76 if (cpuset_zone_allowed(*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
= (nid
+ 1) % num_online_nodes();
110 page
[1].lru
.next
= (void *)free_huge_page
; /* dtor */
111 spin_lock(&hugetlb_lock
);
113 nr_huge_pages_node
[page_to_nid(page
)]++;
114 spin_unlock(&hugetlb_lock
);
115 put_page(page
); /* free it into the hugepage allocator */
121 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
124 struct inode
*inode
= vma
->vm_file
->f_dentry
->d_inode
;
129 spin_lock(&hugetlb_lock
);
131 if (vma
->vm_flags
& VM_MAYSHARE
) {
133 /* idx = radix tree index, i.e. offset into file in
134 * HPAGE_SIZE units */
135 idx
= ((addr
- vma
->vm_start
) >> HPAGE_SHIFT
)
136 + (vma
->vm_pgoff
>> (HPAGE_SHIFT
- PAGE_SHIFT
));
138 /* The hugetlbfs specific inode info stores the number
139 * of "guaranteed available" (huge) pages. That is,
140 * the first 'prereserved_hpages' pages of the inode
141 * are either already instantiated, or have been
142 * pre-reserved (by hugetlb_reserve_for_inode()). Here
143 * we're in the process of instantiating the page, so
144 * we use this to determine whether to draw from the
145 * pre-reserved pool or the truly free pool. */
146 if (idx
< HUGETLBFS_I(inode
)->prereserved_hpages
)
151 if (free_huge_pages
<= reserved_huge_pages
)
154 BUG_ON(reserved_huge_pages
== 0);
155 reserved_huge_pages
--;
158 page
= dequeue_huge_page(vma
, addr
);
162 spin_unlock(&hugetlb_lock
);
163 set_page_refcounted(page
);
167 WARN_ON(use_reserve
); /* reserved allocations shouldn't fail */
168 spin_unlock(&hugetlb_lock
);
172 /* hugetlb_extend_reservation()
174 * Ensure that at least 'atleast' hugepages are, and will remain,
175 * available to instantiate the first 'atleast' pages of the given
176 * inode. If the inode doesn't already have this many pages reserved
177 * or instantiated, set aside some hugepages in the reserved pool to
178 * satisfy later faults (or fail now if there aren't enough, rather
179 * than getting the SIGBUS later).
181 int hugetlb_extend_reservation(struct hugetlbfs_inode_info
*info
,
182 unsigned long atleast
)
184 struct inode
*inode
= &info
->vfs_inode
;
185 unsigned long change_in_reserve
= 0;
188 spin_lock(&hugetlb_lock
);
189 read_lock_irq(&inode
->i_mapping
->tree_lock
);
191 if (info
->prereserved_hpages
>= atleast
)
194 /* Because we always call this on shared mappings, none of the
195 * pages beyond info->prereserved_hpages can have been
196 * instantiated, so we need to reserve all of them now. */
197 change_in_reserve
= atleast
- info
->prereserved_hpages
;
199 if ((reserved_huge_pages
+ change_in_reserve
) > free_huge_pages
) {
204 reserved_huge_pages
+= change_in_reserve
;
205 info
->prereserved_hpages
= atleast
;
208 read_unlock_irq(&inode
->i_mapping
->tree_lock
);
209 spin_unlock(&hugetlb_lock
);
214 /* hugetlb_truncate_reservation()
216 * This returns pages reserved for the given inode to the general free
217 * hugepage pool. If the inode has any pages prereserved, but not
218 * instantiated, beyond offset (atmost << HPAGE_SIZE), then release
221 void hugetlb_truncate_reservation(struct hugetlbfs_inode_info
*info
,
222 unsigned long atmost
)
224 struct inode
*inode
= &info
->vfs_inode
;
225 struct address_space
*mapping
= inode
->i_mapping
;
227 unsigned long change_in_reserve
= 0;
230 spin_lock(&hugetlb_lock
);
231 read_lock_irq(&inode
->i_mapping
->tree_lock
);
233 if (info
->prereserved_hpages
<= atmost
)
236 /* Count pages which were reserved, but not instantiated, and
237 * which we can now release. */
238 for (idx
= atmost
; idx
< info
->prereserved_hpages
; idx
++) {
239 page
= radix_tree_lookup(&mapping
->page_tree
, idx
);
241 /* Pages which are already instantiated can't
242 * be unreserved (and in fact have already
243 * been removed from the reserved pool) */
247 BUG_ON(reserved_huge_pages
< change_in_reserve
);
248 reserved_huge_pages
-= change_in_reserve
;
249 info
->prereserved_hpages
= atmost
;
252 read_unlock_irq(&inode
->i_mapping
->tree_lock
);
253 spin_unlock(&hugetlb_lock
);
256 static int __init
hugetlb_init(void)
260 if (HPAGE_SHIFT
== 0)
263 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
264 INIT_LIST_HEAD(&hugepage_freelists
[i
]);
266 for (i
= 0; i
< max_huge_pages
; ++i
) {
267 if (!alloc_fresh_huge_page())
270 max_huge_pages
= free_huge_pages
= nr_huge_pages
= i
;
271 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages
);
274 module_init(hugetlb_init
);
276 static int __init
hugetlb_setup(char *s
)
278 if (sscanf(s
, "%lu", &max_huge_pages
) <= 0)
282 __setup("hugepages=", hugetlb_setup
);
285 static void update_and_free_page(struct page
*page
)
289 nr_huge_pages_node
[page_zone(page
)->zone_pgdat
->node_id
]--;
290 for (i
= 0; i
< (HPAGE_SIZE
/ PAGE_SIZE
); i
++) {
291 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
| 1 << PG_referenced
|
292 1 << PG_dirty
| 1 << PG_active
| 1 << PG_reserved
|
293 1 << PG_private
| 1<< PG_writeback
);
295 page
[1].lru
.next
= NULL
;
296 set_page_refcounted(page
);
297 __free_pages(page
, HUGETLB_PAGE_ORDER
);
300 #ifdef CONFIG_HIGHMEM
301 static void try_to_free_low(unsigned long count
)
304 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
305 struct page
*page
, *next
;
306 list_for_each_entry_safe(page
, next
, &hugepage_freelists
[i
], lru
) {
307 if (PageHighMem(page
))
309 list_del(&page
->lru
);
310 update_and_free_page(page
);
311 nid
= page_zone(page
)->zone_pgdat
->node_id
;
313 free_huge_pages_node
[nid
]--;
314 if (count
>= nr_huge_pages
)
320 static inline void try_to_free_low(unsigned long count
)
325 static unsigned long set_max_huge_pages(unsigned long count
)
327 while (count
> nr_huge_pages
) {
328 if (!alloc_fresh_huge_page())
329 return nr_huge_pages
;
331 if (count
>= nr_huge_pages
)
332 return nr_huge_pages
;
334 spin_lock(&hugetlb_lock
);
335 try_to_free_low(count
);
336 while (count
< nr_huge_pages
) {
337 struct page
*page
= dequeue_huge_page(NULL
, 0);
340 update_and_free_page(page
);
342 spin_unlock(&hugetlb_lock
);
343 return nr_huge_pages
;
346 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
347 struct file
*file
, void __user
*buffer
,
348 size_t *length
, loff_t
*ppos
)
350 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
351 max_huge_pages
= set_max_huge_pages(max_huge_pages
);
354 #endif /* CONFIG_SYSCTL */
356 int hugetlb_report_meminfo(char *buf
)
359 "HugePages_Total: %5lu\n"
360 "HugePages_Free: %5lu\n"
361 "HugePages_Rsvd: %5lu\n"
362 "Hugepagesize: %5lu kB\n",
369 int hugetlb_report_node_meminfo(int nid
, char *buf
)
372 "Node %d HugePages_Total: %5u\n"
373 "Node %d HugePages_Free: %5u\n",
374 nid
, nr_huge_pages_node
[nid
],
375 nid
, free_huge_pages_node
[nid
]);
378 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
379 unsigned long hugetlb_total_pages(void)
381 return nr_huge_pages
* (HPAGE_SIZE
/ PAGE_SIZE
);
385 * We cannot handle pagefaults against hugetlb pages at all. They cause
386 * handle_mm_fault() to try to instantiate regular-sized pages in the
387 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
390 static struct page
*hugetlb_nopage(struct vm_area_struct
*vma
,
391 unsigned long address
, int *unused
)
397 struct vm_operations_struct hugetlb_vm_ops
= {
398 .nopage
= hugetlb_nopage
,
401 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
408 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
410 entry
= pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
412 entry
= pte_mkyoung(entry
);
413 entry
= pte_mkhuge(entry
);
418 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
419 unsigned long address
, pte_t
*ptep
)
423 entry
= pte_mkwrite(pte_mkdirty(*ptep
));
424 ptep_set_access_flags(vma
, address
, ptep
, entry
, 1);
425 update_mmu_cache(vma
, address
, entry
);
426 lazy_mmu_prot_update(entry
);
430 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
431 struct vm_area_struct
*vma
)
433 pte_t
*src_pte
, *dst_pte
, entry
;
434 struct page
*ptepage
;
438 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
440 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= HPAGE_SIZE
) {
441 src_pte
= huge_pte_offset(src
, addr
);
444 dst_pte
= huge_pte_alloc(dst
, addr
);
447 spin_lock(&dst
->page_table_lock
);
448 spin_lock(&src
->page_table_lock
);
449 if (!pte_none(*src_pte
)) {
451 ptep_set_wrprotect(src
, addr
, src_pte
);
453 ptepage
= pte_page(entry
);
455 add_mm_counter(dst
, file_rss
, HPAGE_SIZE
/ PAGE_SIZE
);
456 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
458 spin_unlock(&src
->page_table_lock
);
459 spin_unlock(&dst
->page_table_lock
);
467 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
470 struct mm_struct
*mm
= vma
->vm_mm
;
471 unsigned long address
;
476 WARN_ON(!is_vm_hugetlb_page(vma
));
477 BUG_ON(start
& ~HPAGE_MASK
);
478 BUG_ON(end
& ~HPAGE_MASK
);
480 spin_lock(&mm
->page_table_lock
);
482 /* Update high watermark before we lower rss */
483 update_hiwater_rss(mm
);
485 for (address
= start
; address
< end
; address
+= HPAGE_SIZE
) {
486 ptep
= huge_pte_offset(mm
, address
);
490 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
494 page
= pte_page(pte
);
496 add_mm_counter(mm
, file_rss
, (int) -(HPAGE_SIZE
/ PAGE_SIZE
));
499 spin_unlock(&mm
->page_table_lock
);
500 flush_tlb_range(vma
, start
, end
);
503 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
504 unsigned long address
, pte_t
*ptep
, pte_t pte
)
506 struct page
*old_page
, *new_page
;
509 old_page
= pte_page(pte
);
511 /* If no-one else is actually using this page, avoid the copy
512 * and just make the page writable */
513 avoidcopy
= (page_count(old_page
) == 1);
515 set_huge_ptep_writable(vma
, address
, ptep
);
516 return VM_FAULT_MINOR
;
519 page_cache_get(old_page
);
520 new_page
= alloc_huge_page(vma
, address
);
523 page_cache_release(old_page
);
527 spin_unlock(&mm
->page_table_lock
);
528 copy_huge_page(new_page
, old_page
, address
);
529 spin_lock(&mm
->page_table_lock
);
531 ptep
= huge_pte_offset(mm
, address
& HPAGE_MASK
);
532 if (likely(pte_same(*ptep
, pte
))) {
534 set_huge_pte_at(mm
, address
, ptep
,
535 make_huge_pte(vma
, new_page
, 1));
536 /* Make the old page be freed below */
539 page_cache_release(new_page
);
540 page_cache_release(old_page
);
541 return VM_FAULT_MINOR
;
544 int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
545 unsigned long address
, pte_t
*ptep
, int write_access
)
547 int ret
= VM_FAULT_SIGBUS
;
551 struct address_space
*mapping
;
554 mapping
= vma
->vm_file
->f_mapping
;
555 idx
= ((address
- vma
->vm_start
) >> HPAGE_SHIFT
)
556 + (vma
->vm_pgoff
>> (HPAGE_SHIFT
- PAGE_SHIFT
));
559 * Use page lock to guard against racing truncation
560 * before we get page_table_lock.
563 page
= find_lock_page(mapping
, idx
);
565 if (hugetlb_get_quota(mapping
))
567 page
= alloc_huge_page(vma
, address
);
569 hugetlb_put_quota(mapping
);
573 clear_huge_page(page
, address
);
575 if (vma
->vm_flags
& VM_SHARED
) {
578 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
581 hugetlb_put_quota(mapping
);
590 spin_lock(&mm
->page_table_lock
);
591 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
595 ret
= VM_FAULT_MINOR
;
596 if (!pte_none(*ptep
))
599 add_mm_counter(mm
, file_rss
, HPAGE_SIZE
/ PAGE_SIZE
);
600 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
601 && (vma
->vm_flags
& VM_SHARED
)));
602 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
604 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
605 /* Optimization, do the COW without a second fault */
606 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
);
609 spin_unlock(&mm
->page_table_lock
);
615 spin_unlock(&mm
->page_table_lock
);
616 hugetlb_put_quota(mapping
);
622 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
623 unsigned long address
, int write_access
)
628 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
630 ptep
= huge_pte_alloc(mm
, address
);
635 * Serialize hugepage allocation and instantiation, so that we don't
636 * get spurious allocation failures if two CPUs race to instantiate
637 * the same page in the page cache.
639 mutex_lock(&hugetlb_instantiation_mutex
);
641 if (pte_none(entry
)) {
642 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
643 mutex_unlock(&hugetlb_instantiation_mutex
);
647 ret
= VM_FAULT_MINOR
;
649 spin_lock(&mm
->page_table_lock
);
650 /* Check for a racing update before calling hugetlb_cow */
651 if (likely(pte_same(entry
, *ptep
)))
652 if (write_access
&& !pte_write(entry
))
653 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
);
654 spin_unlock(&mm
->page_table_lock
);
655 mutex_unlock(&hugetlb_instantiation_mutex
);
660 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
661 struct page
**pages
, struct vm_area_struct
**vmas
,
662 unsigned long *position
, int *length
, int i
)
664 unsigned long pfn_offset
;
665 unsigned long vaddr
= *position
;
666 int remainder
= *length
;
668 spin_lock(&mm
->page_table_lock
);
669 while (vaddr
< vma
->vm_end
&& remainder
) {
674 * Some archs (sparc64, sh*) have multiple pte_ts to
675 * each hugepage. We have to make * sure we get the
676 * first, for the page indexing below to work.
678 pte
= huge_pte_offset(mm
, vaddr
& HPAGE_MASK
);
680 if (!pte
|| pte_none(*pte
)) {
683 spin_unlock(&mm
->page_table_lock
);
684 ret
= hugetlb_fault(mm
, vma
, vaddr
, 0);
685 spin_lock(&mm
->page_table_lock
);
686 if (ret
== VM_FAULT_MINOR
)
695 pfn_offset
= (vaddr
& ~HPAGE_MASK
) >> PAGE_SHIFT
;
696 page
= pte_page(*pte
);
700 pages
[i
] = page
+ pfn_offset
;
709 if (vaddr
< vma
->vm_end
&& remainder
&&
710 pfn_offset
< HPAGE_SIZE
/PAGE_SIZE
) {
712 * We use pfn_offset to avoid touching the pageframes
713 * of this compound page.
718 spin_unlock(&mm
->page_table_lock
);
725 void hugetlb_change_protection(struct vm_area_struct
*vma
,
726 unsigned long address
, unsigned long end
, pgprot_t newprot
)
728 struct mm_struct
*mm
= vma
->vm_mm
;
729 unsigned long start
= address
;
733 BUG_ON(address
>= end
);
734 flush_cache_range(vma
, address
, end
);
736 spin_lock(&mm
->page_table_lock
);
737 for (; address
< end
; address
+= HPAGE_SIZE
) {
738 ptep
= huge_pte_offset(mm
, address
);
741 if (!pte_none(*ptep
)) {
742 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
743 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
744 set_huge_pte_at(mm
, address
, ptep
, pte
);
745 lazy_mmu_prot_update(pte
);
748 spin_unlock(&mm
->page_table_lock
);
750 flush_tlb_range(vma
, start
, end
);