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>
17 #include <asm/pgtable.h>
19 #include <linux/hugetlb.h>
21 const unsigned long hugetlb_zero
= 0, hugetlb_infinity
= ~0UL;
22 static unsigned long nr_huge_pages
, free_huge_pages
;
23 unsigned long max_huge_pages
;
24 static struct list_head hugepage_freelists
[MAX_NUMNODES
];
25 static unsigned int nr_huge_pages_node
[MAX_NUMNODES
];
26 static unsigned int free_huge_pages_node
[MAX_NUMNODES
];
29 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
31 static DEFINE_SPINLOCK(hugetlb_lock
);
33 static void enqueue_huge_page(struct page
*page
)
35 int nid
= page_to_nid(page
);
36 list_add(&page
->lru
, &hugepage_freelists
[nid
]);
38 free_huge_pages_node
[nid
]++;
41 static struct page
*dequeue_huge_page(struct vm_area_struct
*vma
,
42 unsigned long address
)
44 int nid
= numa_node_id();
45 struct page
*page
= NULL
;
46 struct zonelist
*zonelist
= huge_zonelist(vma
, address
);
49 for (z
= zonelist
->zones
; *z
; z
++) {
50 nid
= (*z
)->zone_pgdat
->node_id
;
51 if (!list_empty(&hugepage_freelists
[nid
]))
56 page
= list_entry(hugepage_freelists
[nid
].next
,
60 free_huge_pages_node
[nid
]--;
65 static struct page
*alloc_fresh_huge_page(void)
69 page
= alloc_pages_node(nid
, GFP_HIGHUSER
|__GFP_COMP
|__GFP_NOWARN
,
71 nid
= (nid
+ 1) % num_online_nodes();
73 spin_lock(&hugetlb_lock
);
75 nr_huge_pages_node
[page_to_nid(page
)]++;
76 spin_unlock(&hugetlb_lock
);
81 void free_huge_page(struct page
*page
)
83 BUG_ON(page_count(page
));
85 INIT_LIST_HEAD(&page
->lru
);
86 page
[1].mapping
= NULL
;
88 spin_lock(&hugetlb_lock
);
89 enqueue_huge_page(page
);
90 spin_unlock(&hugetlb_lock
);
93 struct page
*alloc_huge_page(struct vm_area_struct
*vma
, unsigned long addr
)
98 spin_lock(&hugetlb_lock
);
99 page
= dequeue_huge_page(vma
, addr
);
101 spin_unlock(&hugetlb_lock
);
104 spin_unlock(&hugetlb_lock
);
105 set_page_count(page
, 1);
106 page
[1].mapping
= (void *)free_huge_page
;
107 for (i
= 0; i
< (HPAGE_SIZE
/PAGE_SIZE
); ++i
)
108 clear_highpage(&page
[i
]);
112 static int __init
hugetlb_init(void)
117 if (HPAGE_SHIFT
== 0)
120 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
121 INIT_LIST_HEAD(&hugepage_freelists
[i
]);
123 for (i
= 0; i
< max_huge_pages
; ++i
) {
124 page
= alloc_fresh_huge_page();
127 spin_lock(&hugetlb_lock
);
128 enqueue_huge_page(page
);
129 spin_unlock(&hugetlb_lock
);
131 max_huge_pages
= free_huge_pages
= nr_huge_pages
= i
;
132 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages
);
135 module_init(hugetlb_init
);
137 static int __init
hugetlb_setup(char *s
)
139 if (sscanf(s
, "%lu", &max_huge_pages
) <= 0)
143 __setup("hugepages=", hugetlb_setup
);
146 static void update_and_free_page(struct page
*page
)
150 nr_huge_pages_node
[page_zone(page
)->zone_pgdat
->node_id
]--;
151 for (i
= 0; i
< (HPAGE_SIZE
/ PAGE_SIZE
); i
++) {
152 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
| 1 << PG_referenced
|
153 1 << PG_dirty
| 1 << PG_active
| 1 << PG_reserved
|
154 1 << PG_private
| 1<< PG_writeback
);
155 set_page_count(&page
[i
], 0);
157 set_page_count(page
, 1);
158 __free_pages(page
, HUGETLB_PAGE_ORDER
);
161 #ifdef CONFIG_HIGHMEM
162 static void try_to_free_low(unsigned long count
)
165 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
166 struct page
*page
, *next
;
167 list_for_each_entry_safe(page
, next
, &hugepage_freelists
[i
], lru
) {
168 if (PageHighMem(page
))
170 list_del(&page
->lru
);
171 update_and_free_page(page
);
172 nid
= page_zone(page
)->zone_pgdat
->node_id
;
174 free_huge_pages_node
[nid
]--;
175 if (count
>= nr_huge_pages
)
181 static inline void try_to_free_low(unsigned long count
)
186 static unsigned long set_max_huge_pages(unsigned long count
)
188 while (count
> nr_huge_pages
) {
189 struct page
*page
= alloc_fresh_huge_page();
191 return nr_huge_pages
;
192 spin_lock(&hugetlb_lock
);
193 enqueue_huge_page(page
);
194 spin_unlock(&hugetlb_lock
);
196 if (count
>= nr_huge_pages
)
197 return nr_huge_pages
;
199 spin_lock(&hugetlb_lock
);
200 try_to_free_low(count
);
201 while (count
< nr_huge_pages
) {
202 struct page
*page
= dequeue_huge_page(NULL
, 0);
205 update_and_free_page(page
);
207 spin_unlock(&hugetlb_lock
);
208 return nr_huge_pages
;
211 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
212 struct file
*file
, void __user
*buffer
,
213 size_t *length
, loff_t
*ppos
)
215 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
216 max_huge_pages
= set_max_huge_pages(max_huge_pages
);
219 #endif /* CONFIG_SYSCTL */
221 int hugetlb_report_meminfo(char *buf
)
224 "HugePages_Total: %5lu\n"
225 "HugePages_Free: %5lu\n"
226 "Hugepagesize: %5lu kB\n",
232 int hugetlb_report_node_meminfo(int nid
, char *buf
)
235 "Node %d HugePages_Total: %5u\n"
236 "Node %d HugePages_Free: %5u\n",
237 nid
, nr_huge_pages_node
[nid
],
238 nid
, free_huge_pages_node
[nid
]);
241 int is_hugepage_mem_enough(size_t size
)
243 return (size
+ ~HPAGE_MASK
)/HPAGE_SIZE
<= free_huge_pages
;
246 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
247 unsigned long hugetlb_total_pages(void)
249 return nr_huge_pages
* (HPAGE_SIZE
/ PAGE_SIZE
);
253 * We cannot handle pagefaults against hugetlb pages at all. They cause
254 * handle_mm_fault() to try to instantiate regular-sized pages in the
255 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
258 static struct page
*hugetlb_nopage(struct vm_area_struct
*vma
,
259 unsigned long address
, int *unused
)
265 struct vm_operations_struct hugetlb_vm_ops
= {
266 .nopage
= hugetlb_nopage
,
269 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
276 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
278 entry
= pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
280 entry
= pte_mkyoung(entry
);
281 entry
= pte_mkhuge(entry
);
286 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
287 unsigned long address
, pte_t
*ptep
)
291 entry
= pte_mkwrite(pte_mkdirty(*ptep
));
292 ptep_set_access_flags(vma
, address
, ptep
, entry
, 1);
293 update_mmu_cache(vma
, address
, entry
);
294 lazy_mmu_prot_update(entry
);
298 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
299 struct vm_area_struct
*vma
)
301 pte_t
*src_pte
, *dst_pte
, entry
;
302 struct page
*ptepage
;
306 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
308 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= HPAGE_SIZE
) {
309 src_pte
= huge_pte_offset(src
, addr
);
312 dst_pte
= huge_pte_alloc(dst
, addr
);
315 spin_lock(&dst
->page_table_lock
);
316 spin_lock(&src
->page_table_lock
);
317 if (!pte_none(*src_pte
)) {
319 ptep_set_wrprotect(src
, addr
, src_pte
);
321 ptepage
= pte_page(entry
);
323 add_mm_counter(dst
, file_rss
, HPAGE_SIZE
/ PAGE_SIZE
);
324 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
326 spin_unlock(&src
->page_table_lock
);
327 spin_unlock(&dst
->page_table_lock
);
335 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
338 struct mm_struct
*mm
= vma
->vm_mm
;
339 unsigned long address
;
344 WARN_ON(!is_vm_hugetlb_page(vma
));
345 BUG_ON(start
& ~HPAGE_MASK
);
346 BUG_ON(end
& ~HPAGE_MASK
);
348 spin_lock(&mm
->page_table_lock
);
350 /* Update high watermark before we lower rss */
351 update_hiwater_rss(mm
);
353 for (address
= start
; address
< end
; address
+= HPAGE_SIZE
) {
354 ptep
= huge_pte_offset(mm
, address
);
358 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
362 page
= pte_page(pte
);
364 add_mm_counter(mm
, file_rss
, (int) -(HPAGE_SIZE
/ PAGE_SIZE
));
367 spin_unlock(&mm
->page_table_lock
);
368 flush_tlb_range(vma
, start
, end
);
371 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
372 unsigned long address
, pte_t
*ptep
, pte_t pte
)
374 struct page
*old_page
, *new_page
;
377 old_page
= pte_page(pte
);
379 /* If no-one else is actually using this page, avoid the copy
380 * and just make the page writable */
381 avoidcopy
= (page_count(old_page
) == 1);
383 set_huge_ptep_writable(vma
, address
, ptep
);
384 return VM_FAULT_MINOR
;
387 page_cache_get(old_page
);
388 new_page
= alloc_huge_page(vma
, address
);
391 page_cache_release(old_page
);
393 /* Logically this is OOM, not a SIGBUS, but an OOM
394 * could cause the kernel to go killing other
395 * processes which won't help the hugepage situation
397 return VM_FAULT_SIGBUS
;
400 spin_unlock(&mm
->page_table_lock
);
401 for (i
= 0; i
< HPAGE_SIZE
/PAGE_SIZE
; i
++)
402 copy_user_highpage(new_page
+ i
, old_page
+ i
,
403 address
+ i
*PAGE_SIZE
);
404 spin_lock(&mm
->page_table_lock
);
406 ptep
= huge_pte_offset(mm
, address
& HPAGE_MASK
);
407 if (likely(pte_same(*ptep
, pte
))) {
409 set_huge_pte_at(mm
, address
, ptep
,
410 make_huge_pte(vma
, new_page
, 1));
411 /* Make the old page be freed below */
414 page_cache_release(new_page
);
415 page_cache_release(old_page
);
416 return VM_FAULT_MINOR
;
419 int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
420 unsigned long address
, pte_t
*ptep
, int write_access
)
422 int ret
= VM_FAULT_SIGBUS
;
426 struct address_space
*mapping
;
429 mapping
= vma
->vm_file
->f_mapping
;
430 idx
= ((address
- vma
->vm_start
) >> HPAGE_SHIFT
)
431 + (vma
->vm_pgoff
>> (HPAGE_SHIFT
- PAGE_SHIFT
));
434 * Use page lock to guard against racing truncation
435 * before we get page_table_lock.
438 page
= find_lock_page(mapping
, idx
);
440 if (hugetlb_get_quota(mapping
))
442 page
= alloc_huge_page(vma
, address
);
444 hugetlb_put_quota(mapping
);
448 if (vma
->vm_flags
& VM_SHARED
) {
451 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
454 hugetlb_put_quota(mapping
);
463 spin_lock(&mm
->page_table_lock
);
464 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
468 ret
= VM_FAULT_MINOR
;
469 if (!pte_none(*ptep
))
472 add_mm_counter(mm
, file_rss
, HPAGE_SIZE
/ PAGE_SIZE
);
473 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
474 && (vma
->vm_flags
& VM_SHARED
)));
475 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
477 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
478 /* Optimization, do the COW without a second fault */
479 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
);
482 spin_unlock(&mm
->page_table_lock
);
488 spin_unlock(&mm
->page_table_lock
);
489 hugetlb_put_quota(mapping
);
495 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
496 unsigned long address
, int write_access
)
502 ptep
= huge_pte_alloc(mm
, address
);
508 return hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
510 ret
= VM_FAULT_MINOR
;
512 spin_lock(&mm
->page_table_lock
);
513 /* Check for a racing update before calling hugetlb_cow */
514 if (likely(pte_same(entry
, *ptep
)))
515 if (write_access
&& !pte_write(entry
))
516 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
);
517 spin_unlock(&mm
->page_table_lock
);
522 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
523 struct page
**pages
, struct vm_area_struct
**vmas
,
524 unsigned long *position
, int *length
, int i
)
526 unsigned long vpfn
, vaddr
= *position
;
527 int remainder
= *length
;
529 vpfn
= vaddr
/PAGE_SIZE
;
530 spin_lock(&mm
->page_table_lock
);
531 while (vaddr
< vma
->vm_end
&& remainder
) {
536 * Some archs (sparc64, sh*) have multiple pte_ts to
537 * each hugepage. We have to make * sure we get the
538 * first, for the page indexing below to work.
540 pte
= huge_pte_offset(mm
, vaddr
& HPAGE_MASK
);
542 if (!pte
|| pte_none(*pte
)) {
545 spin_unlock(&mm
->page_table_lock
);
546 ret
= hugetlb_fault(mm
, vma
, vaddr
, 0);
547 spin_lock(&mm
->page_table_lock
);
548 if (ret
== VM_FAULT_MINOR
)
558 page
= &pte_page(*pte
)[vpfn
% (HPAGE_SIZE
/PAGE_SIZE
)];
571 spin_unlock(&mm
->page_table_lock
);