blob: 1a5642074e342532f4a844ed0d2fdc86bd9b5de9 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
4 */
5#include <linux/gfp.h>
6#include <linux/list.h>
7#include <linux/init.h>
8#include <linux/module.h>
9#include <linux/mm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
12#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070013#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080014#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080015#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080016#include <linux/mutex.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080017
David Gibson63551ae2005-06-21 17:14:44 -070018#include <asm/page.h>
19#include <asm/pgtable.h>
20
21#include <linux/hugetlb.h>
Nick Piggin7835e982006-03-22 00:08:40 -080022#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070023
24const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -070025static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
Adam Litke7893d1d2007-10-16 01:26:18 -070026static unsigned long surplus_huge_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -070027unsigned long max_huge_pages;
28static struct list_head hugepage_freelists[MAX_NUMNODES];
29static unsigned int nr_huge_pages_node[MAX_NUMNODES];
30static unsigned int free_huge_pages_node[MAX_NUMNODES];
Adam Litke7893d1d2007-10-16 01:26:18 -070031static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
Mel Gorman396faf02007-07-17 04:03:13 -070032static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
33unsigned long hugepages_treat_as_movable;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -080034unsigned long nr_overcommit_huge_pages;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -070035static int hugetlb_next_nid;
Mel Gorman396faf02007-07-17 04:03:13 -070036
David Gibson3935baa2006-03-22 00:08:53 -080037/*
38 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
39 */
40static DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080041
David Gibson79ac6ba2006-03-22 00:08:51 -080042static void clear_huge_page(struct page *page, unsigned long addr)
43{
44 int i;
45
46 might_sleep();
47 for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
48 cond_resched();
Ralf Baechle281e0e32007-10-01 01:20:10 -070049 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
David Gibson79ac6ba2006-03-22 00:08:51 -080050 }
51}
52
53static void copy_huge_page(struct page *dst, struct page *src,
Atsushi Nemoto9de455b2006-12-12 17:14:55 +000054 unsigned long addr, struct vm_area_struct *vma)
David Gibson79ac6ba2006-03-22 00:08:51 -080055{
56 int i;
57
58 might_sleep();
59 for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
60 cond_resched();
Atsushi Nemoto9de455b2006-12-12 17:14:55 +000061 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
David Gibson79ac6ba2006-03-22 00:08:51 -080062 }
63}
64
Linus Torvalds1da177e2005-04-16 15:20:36 -070065static void enqueue_huge_page(struct page *page)
66{
67 int nid = page_to_nid(page);
68 list_add(&page->lru, &hugepage_freelists[nid]);
69 free_huge_pages++;
70 free_huge_pages_node[nid]++;
71}
72
Christoph Lameter5da7ca82006-01-06 00:10:46 -080073static struct page *dequeue_huge_page(struct vm_area_struct *vma,
74 unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -070075{
Nishanth Aravamudan31a5c6e2007-07-15 23:38:02 -070076 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -070077 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -070078 struct mempolicy *mpol;
Mel Gorman396faf02007-07-17 04:03:13 -070079 struct zonelist *zonelist = huge_zonelist(vma, address,
Lee Schermerhorn480eccf2007-09-18 22:46:47 -070080 htlb_alloc_mask, &mpol);
Christoph Lameter96df9332006-01-06 00:10:45 -080081 struct zone **z;
Linus Torvalds1da177e2005-04-16 15:20:36 -070082
Christoph Lameter96df9332006-01-06 00:10:45 -080083 for (z = zonelist->zones; *z; z++) {
Christoph Lameter89fa3022006-09-25 23:31:55 -070084 nid = zone_to_nid(*z);
Mel Gorman396faf02007-07-17 04:03:13 -070085 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
Andrew Morton3abf7af2007-07-19 01:49:08 -070086 !list_empty(&hugepage_freelists[nid])) {
87 page = list_entry(hugepage_freelists[nid].next,
88 struct page, lru);
89 list_del(&page->lru);
90 free_huge_pages--;
91 free_huge_pages_node[nid]--;
Adam Litkee4e574b2007-10-16 01:26:19 -070092 if (vma && vma->vm_flags & VM_MAYSHARE)
93 resv_huge_pages--;
Ken Chen5ab3ee72007-07-23 18:44:00 -070094 break;
Andrew Morton3abf7af2007-07-19 01:49:08 -070095 }
Linus Torvalds1da177e2005-04-16 15:20:36 -070096 }
Lee Schermerhorn480eccf2007-09-18 22:46:47 -070097 mpol_free(mpol); /* unref if mpol !NULL */
Linus Torvalds1da177e2005-04-16 15:20:36 -070098 return page;
99}
100
Adam Litke6af2acb2007-10-16 01:26:16 -0700101static void update_and_free_page(struct page *page)
102{
103 int i;
104 nr_huge_pages--;
105 nr_huge_pages_node[page_to_nid(page)]--;
106 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
107 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
108 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
109 1 << PG_private | 1<< PG_writeback);
110 }
111 set_compound_page_dtor(page, NULL);
112 set_page_refcounted(page);
113 __free_pages(page, HUGETLB_PAGE_ORDER);
114}
115
David Gibson27a85ef2006-03-22 00:08:56 -0800116static void free_huge_page(struct page *page)
117{
Adam Litke7893d1d2007-10-16 01:26:18 -0700118 int nid = page_to_nid(page);
Adam Litkec79fb752007-11-14 16:59:38 -0800119 struct address_space *mapping;
David Gibson27a85ef2006-03-22 00:08:56 -0800120
Adam Litkec79fb752007-11-14 16:59:38 -0800121 mapping = (struct address_space *) page_private(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700122 BUG_ON(page_count(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800123 INIT_LIST_HEAD(&page->lru);
124
125 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700126 if (surplus_huge_pages_node[nid]) {
127 update_and_free_page(page);
128 surplus_huge_pages--;
129 surplus_huge_pages_node[nid]--;
130 } else {
131 enqueue_huge_page(page);
132 }
David Gibson27a85ef2006-03-22 00:08:56 -0800133 spin_unlock(&hugetlb_lock);
Adam Litkec79fb752007-11-14 16:59:38 -0800134 if (mapping)
Adam Litke9a119c02007-11-14 16:59:41 -0800135 hugetlb_put_quota(mapping, 1);
Adam Litkec79fb752007-11-14 16:59:38 -0800136 set_page_private(page, 0);
David Gibson27a85ef2006-03-22 00:08:56 -0800137}
138
Adam Litke7893d1d2007-10-16 01:26:18 -0700139/*
140 * Increment or decrement surplus_huge_pages. Keep node-specific counters
141 * balanced by operating on them in a round-robin fashion.
142 * Returns 1 if an adjustment was made.
143 */
144static int adjust_pool_surplus(int delta)
145{
146 static int prev_nid;
147 int nid = prev_nid;
148 int ret = 0;
149
150 VM_BUG_ON(delta != -1 && delta != 1);
151 do {
152 nid = next_node(nid, node_online_map);
153 if (nid == MAX_NUMNODES)
154 nid = first_node(node_online_map);
155
156 /* To shrink on this node, there must be a surplus page */
157 if (delta < 0 && !surplus_huge_pages_node[nid])
158 continue;
159 /* Surplus cannot exceed the total number of pages */
160 if (delta > 0 && surplus_huge_pages_node[nid] >=
161 nr_huge_pages_node[nid])
162 continue;
163
164 surplus_huge_pages += delta;
165 surplus_huge_pages_node[nid] += delta;
166 ret = 1;
167 break;
168 } while (nid != prev_nid);
169
170 prev_nid = nid;
171 return ret;
172}
173
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700174static struct page *alloc_fresh_huge_page_node(int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700177
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700178 page = alloc_pages_node(nid,
179 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
180 HUGETLB_PAGE_ORDER);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181 if (page) {
Andy Whitcroft33f2ef82006-12-06 20:33:32 -0800182 set_compound_page_dtor(page, free_huge_page);
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800183 spin_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 nr_huge_pages++;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700185 nr_huge_pages_node[nid]++;
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800186 spin_unlock(&hugetlb_lock);
Nick Piggina4822892006-03-22 00:08:08 -0800187 put_page(page); /* free it into the hugepage allocator */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700189
190 return page;
191}
192
193static int alloc_fresh_huge_page(void)
194{
195 struct page *page;
196 int start_nid;
197 int next_nid;
198 int ret = 0;
199
200 start_nid = hugetlb_next_nid;
201
202 do {
203 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
204 if (page)
205 ret = 1;
206 /*
207 * Use a helper variable to find the next node and then
208 * copy it back to hugetlb_next_nid afterwards:
209 * otherwise there's a window in which a racer might
210 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
211 * But we don't need to use a spin_lock here: it really
212 * doesn't matter if occasionally a racer chooses the
213 * same nid as we do. Move nid forward in the mask even
214 * if we just successfully allocated a hugepage so that
215 * the next caller gets hugepages on the next node.
216 */
217 next_nid = next_node(hugetlb_next_nid, node_online_map);
218 if (next_nid == MAX_NUMNODES)
219 next_nid = first_node(node_online_map);
220 hugetlb_next_nid = next_nid;
221 } while (!page && hugetlb_next_nid != start_nid);
222
223 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224}
225
Adam Litke7893d1d2007-10-16 01:26:18 -0700226static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
227 unsigned long address)
228{
229 struct page *page;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800230 unsigned int nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700231
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800232 /*
233 * Assume we will successfully allocate the surplus page to
234 * prevent racing processes from causing the surplus to exceed
235 * overcommit
236 *
237 * This however introduces a different race, where a process B
238 * tries to grow the static hugepage pool while alloc_pages() is
239 * called by process A. B will only examine the per-node
240 * counters in determining if surplus huge pages can be
241 * converted to normal huge pages in adjust_pool_surplus(). A
242 * won't be able to increment the per-node counter, until the
243 * lock is dropped by B, but B doesn't drop hugetlb_lock until
244 * no more huge pages can be converted from surplus to normal
245 * state (and doesn't try to convert again). Thus, we have a
246 * case where a surplus huge page exists, the pool is grown, and
247 * the surplus huge page still exists after, even though it
248 * should just have been converted to a normal huge page. This
249 * does not leak memory, though, as the hugepage will be freed
250 * once it is out of use. It also does not allow the counters to
251 * go out of whack in adjust_pool_surplus() as we don't modify
252 * the node values until we've gotten the hugepage and only the
253 * per-node value is checked there.
254 */
255 spin_lock(&hugetlb_lock);
256 if (surplus_huge_pages >= nr_overcommit_huge_pages) {
257 spin_unlock(&hugetlb_lock);
258 return NULL;
259 } else {
260 nr_huge_pages++;
261 surplus_huge_pages++;
262 }
263 spin_unlock(&hugetlb_lock);
264
Adam Litke7893d1d2007-10-16 01:26:18 -0700265 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
266 HUGETLB_PAGE_ORDER);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800267
268 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700269 if (page) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800270 nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700271 set_compound_page_dtor(page, free_huge_page);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800272 /*
273 * We incremented the global counters already
274 */
275 nr_huge_pages_node[nid]++;
276 surplus_huge_pages_node[nid]++;
277 } else {
278 nr_huge_pages--;
279 surplus_huge_pages--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700280 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800281 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700282
283 return page;
284}
285
Adam Litkee4e574b2007-10-16 01:26:19 -0700286/*
287 * Increase the hugetlb pool such that it can accomodate a reservation
288 * of size 'delta'.
289 */
290static int gather_surplus_pages(int delta)
291{
292 struct list_head surplus_list;
293 struct page *page, *tmp;
294 int ret, i;
295 int needed, allocated;
296
297 needed = (resv_huge_pages + delta) - free_huge_pages;
298 if (needed <= 0)
299 return 0;
300
301 allocated = 0;
302 INIT_LIST_HEAD(&surplus_list);
303
304 ret = -ENOMEM;
305retry:
306 spin_unlock(&hugetlb_lock);
307 for (i = 0; i < needed; i++) {
308 page = alloc_buddy_huge_page(NULL, 0);
309 if (!page) {
310 /*
311 * We were not able to allocate enough pages to
312 * satisfy the entire reservation so we free what
313 * we've allocated so far.
314 */
315 spin_lock(&hugetlb_lock);
316 needed = 0;
317 goto free;
318 }
319
320 list_add(&page->lru, &surplus_list);
321 }
322 allocated += needed;
323
324 /*
325 * After retaking hugetlb_lock, we need to recalculate 'needed'
326 * because either resv_huge_pages or free_huge_pages may have changed.
327 */
328 spin_lock(&hugetlb_lock);
329 needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
330 if (needed > 0)
331 goto retry;
332
333 /*
334 * The surplus_list now contains _at_least_ the number of extra pages
335 * needed to accomodate the reservation. Add the appropriate number
336 * of pages to the hugetlb pool and free the extras back to the buddy
337 * allocator.
338 */
339 needed += allocated;
340 ret = 0;
341free:
342 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
343 list_del(&page->lru);
344 if ((--needed) >= 0)
345 enqueue_huge_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700346 else {
347 /*
348 * Decrement the refcount and free the page using its
349 * destructor. This must be done with hugetlb_lock
350 * unlocked which is safe because free_huge_page takes
351 * hugetlb_lock before deciding how to free the page.
352 */
353 spin_unlock(&hugetlb_lock);
354 put_page(page);
355 spin_lock(&hugetlb_lock);
356 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700357 }
358
359 return ret;
360}
361
362/*
363 * When releasing a hugetlb pool reservation, any surplus pages that were
364 * allocated to satisfy the reservation must be explicitly freed if they were
365 * never used.
366 */
Adrian Bunk8cde0452007-11-14 16:59:43 -0800367static void return_unused_surplus_pages(unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -0700368{
369 static int nid = -1;
370 struct page *page;
371 unsigned long nr_pages;
372
373 nr_pages = min(unused_resv_pages, surplus_huge_pages);
374
375 while (nr_pages) {
376 nid = next_node(nid, node_online_map);
377 if (nid == MAX_NUMNODES)
378 nid = first_node(node_online_map);
379
380 if (!surplus_huge_pages_node[nid])
381 continue;
382
383 if (!list_empty(&hugepage_freelists[nid])) {
384 page = list_entry(hugepage_freelists[nid].next,
385 struct page, lru);
386 list_del(&page->lru);
387 update_and_free_page(page);
388 free_huge_pages--;
389 free_huge_pages_node[nid]--;
390 surplus_huge_pages--;
391 surplus_huge_pages_node[nid]--;
392 nr_pages--;
393 }
394 }
395}
396
Adam Litke348ea202007-11-14 16:59:37 -0800397
398static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
399 unsigned long addr)
400{
401 struct page *page;
402
403 spin_lock(&hugetlb_lock);
404 page = dequeue_huge_page(vma, addr);
405 spin_unlock(&hugetlb_lock);
Adam Litke90d8b7e2007-11-14 16:59:42 -0800406 return page ? page : ERR_PTR(-VM_FAULT_OOM);
Adam Litke348ea202007-11-14 16:59:37 -0800407}
408
409static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
410 unsigned long addr)
411{
412 struct page *page = NULL;
413
Adam Litke90d8b7e2007-11-14 16:59:42 -0800414 if (hugetlb_get_quota(vma->vm_file->f_mapping, 1))
415 return ERR_PTR(-VM_FAULT_SIGBUS);
416
Adam Litke348ea202007-11-14 16:59:37 -0800417 spin_lock(&hugetlb_lock);
418 if (free_huge_pages > resv_huge_pages)
419 page = dequeue_huge_page(vma, addr);
420 spin_unlock(&hugetlb_lock);
Ken Chen68842c92008-01-14 00:55:19 -0800421 if (!page) {
Adam Litke348ea202007-11-14 16:59:37 -0800422 page = alloc_buddy_huge_page(vma, addr);
Ken Chen68842c92008-01-14 00:55:19 -0800423 if (!page) {
424 hugetlb_put_quota(vma->vm_file->f_mapping, 1);
425 return ERR_PTR(-VM_FAULT_OOM);
426 }
427 }
428 return page;
Adam Litke348ea202007-11-14 16:59:37 -0800429}
430
David Gibson27a85ef2006-03-22 00:08:56 -0800431static struct page *alloc_huge_page(struct vm_area_struct *vma,
432 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433{
Adam Litke348ea202007-11-14 16:59:37 -0800434 struct page *page;
Adam Litke2fc39ce2007-11-14 16:59:39 -0800435 struct address_space *mapping = vma->vm_file->f_mapping;
436
Adam Litke348ea202007-11-14 16:59:37 -0800437 if (vma->vm_flags & VM_MAYSHARE)
438 page = alloc_huge_page_shared(vma, addr);
439 else
440 page = alloc_huge_page_private(vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -0800441
442 if (!IS_ERR(page)) {
Adam Litke348ea202007-11-14 16:59:37 -0800443 set_page_refcounted(page);
Adam Litke2fc39ce2007-11-14 16:59:39 -0800444 set_page_private(page, (unsigned long) mapping);
Adam Litke90d8b7e2007-11-14 16:59:42 -0800445 }
446 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800447}
448
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449static int __init hugetlb_init(void)
450{
451 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452
Benjamin Herrenschmidt3c726f82005-11-07 11:06:55 +1100453 if (HPAGE_SHIFT == 0)
454 return 0;
455
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 for (i = 0; i < MAX_NUMNODES; ++i)
457 INIT_LIST_HEAD(&hugepage_freelists[i]);
458
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700459 hugetlb_next_nid = first_node(node_online_map);
460
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 for (i = 0; i < max_huge_pages; ++i) {
Nick Piggina4822892006-03-22 00:08:08 -0800462 if (!alloc_fresh_huge_page())
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 }
465 max_huge_pages = free_huge_pages = nr_huge_pages = i;
466 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
467 return 0;
468}
469module_init(hugetlb_init);
470
471static int __init hugetlb_setup(char *s)
472{
473 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
474 max_huge_pages = 0;
475 return 1;
476}
477__setup("hugepages=", hugetlb_setup);
478
Ken Chen8a630112007-05-09 02:33:34 -0700479static unsigned int cpuset_mems_nr(unsigned int *array)
480{
481 int node;
482 unsigned int nr = 0;
483
484 for_each_node_mask(node, cpuset_current_mems_allowed)
485 nr += array[node];
486
487 return nr;
488}
489
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490#ifdef CONFIG_SYSCTL
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491#ifdef CONFIG_HIGHMEM
492static void try_to_free_low(unsigned long count)
493{
Christoph Lameter4415cc82006-09-25 23:31:55 -0700494 int i;
495
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496 for (i = 0; i < MAX_NUMNODES; ++i) {
497 struct page *page, *next;
498 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
Adam Litke6b0c8802007-10-16 01:26:23 -0700499 if (count >= nr_huge_pages)
500 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 if (PageHighMem(page))
502 continue;
503 list_del(&page->lru);
504 update_and_free_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 free_huge_pages--;
Christoph Lameter4415cc82006-09-25 23:31:55 -0700506 free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 }
508 }
509}
510#else
511static inline void try_to_free_low(unsigned long count)
512{
513}
514#endif
515
Adam Litke7893d1d2007-10-16 01:26:18 -0700516#define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517static unsigned long set_max_huge_pages(unsigned long count)
518{
Adam Litke7893d1d2007-10-16 01:26:18 -0700519 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520
Adam Litke7893d1d2007-10-16 01:26:18 -0700521 /*
522 * Increase the pool size
523 * First take pages out of surplus state. Then make up the
524 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800525 *
526 * We might race with alloc_buddy_huge_page() here and be unable
527 * to convert a surplus huge page to a normal huge page. That is
528 * not critical, though, it just means the overall size of the
529 * pool might be one hugepage larger than it needs to be, but
530 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -0700531 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700533 while (surplus_huge_pages && count > persistent_huge_pages) {
534 if (!adjust_pool_surplus(-1))
535 break;
536 }
537
538 while (count > persistent_huge_pages) {
539 int ret;
540 /*
541 * If this allocation races such that we no longer need the
542 * page, free_huge_page will handle it by freeing the page
543 * and reducing the surplus.
544 */
545 spin_unlock(&hugetlb_lock);
546 ret = alloc_fresh_huge_page();
547 spin_lock(&hugetlb_lock);
548 if (!ret)
549 goto out;
550
551 }
Adam Litke7893d1d2007-10-16 01:26:18 -0700552
553 /*
554 * Decrease the pool size
555 * First return free pages to the buddy allocator (being careful
556 * to keep enough around to satisfy reservations). Then place
557 * pages into surplus state as needed so the pool will shrink
558 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800559 *
560 * By placing pages into the surplus state independent of the
561 * overcommit value, we are allowing the surplus pool size to
562 * exceed overcommit. There are few sane options here. Since
563 * alloc_buddy_huge_page() is checking the global counter,
564 * though, we'll note that we're not allowed to exceed surplus
565 * and won't grow the pool anywhere else. Not until one of the
566 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -0700567 */
Adam Litke6b0c8802007-10-16 01:26:23 -0700568 min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
569 min_count = max(count, min_count);
Adam Litke7893d1d2007-10-16 01:26:18 -0700570 try_to_free_low(min_count);
571 while (min_count < persistent_huge_pages) {
Christoph Lameter5da7ca82006-01-06 00:10:46 -0800572 struct page *page = dequeue_huge_page(NULL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 if (!page)
574 break;
575 update_and_free_page(page);
576 }
Adam Litke7893d1d2007-10-16 01:26:18 -0700577 while (count < persistent_huge_pages) {
578 if (!adjust_pool_surplus(1))
579 break;
580 }
581out:
582 ret = persistent_huge_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700584 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585}
586
587int hugetlb_sysctl_handler(struct ctl_table *table, int write,
588 struct file *file, void __user *buffer,
589 size_t *length, loff_t *ppos)
590{
591 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
592 max_huge_pages = set_max_huge_pages(max_huge_pages);
593 return 0;
594}
Mel Gorman396faf02007-07-17 04:03:13 -0700595
596int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
597 struct file *file, void __user *buffer,
598 size_t *length, loff_t *ppos)
599{
600 proc_dointvec(table, write, file, buffer, length, ppos);
601 if (hugepages_treat_as_movable)
602 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
603 else
604 htlb_alloc_mask = GFP_HIGHUSER;
605 return 0;
606}
607
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608#endif /* CONFIG_SYSCTL */
609
610int hugetlb_report_meminfo(char *buf)
611{
612 return sprintf(buf,
613 "HugePages_Total: %5lu\n"
614 "HugePages_Free: %5lu\n"
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700615 "HugePages_Rsvd: %5lu\n"
Adam Litke7893d1d2007-10-16 01:26:18 -0700616 "HugePages_Surp: %5lu\n"
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 "Hugepagesize: %5lu kB\n",
618 nr_huge_pages,
619 free_huge_pages,
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700620 resv_huge_pages,
Adam Litke7893d1d2007-10-16 01:26:18 -0700621 surplus_huge_pages,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 HPAGE_SIZE/1024);
623}
624
625int hugetlb_report_node_meminfo(int nid, char *buf)
626{
627 return sprintf(buf,
628 "Node %d HugePages_Total: %5u\n"
629 "Node %d HugePages_Free: %5u\n",
630 nid, nr_huge_pages_node[nid],
631 nid, free_huge_pages_node[nid]);
632}
633
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
635unsigned long hugetlb_total_pages(void)
636{
637 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
638}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639
640/*
641 * We cannot handle pagefaults against hugetlb pages at all. They cause
642 * handle_mm_fault() to try to instantiate regular-sized pages in the
643 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
644 * this far.
645 */
Nick Piggind0217ac2007-07-19 01:47:03 -0700646static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647{
648 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -0700649 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650}
651
652struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -0700653 .fault = hugetlb_vm_op_fault,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654};
655
David Gibson1e8f8892006-01-06 00:10:44 -0800656static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
657 int writable)
David Gibson63551ae2005-06-21 17:14:44 -0700658{
659 pte_t entry;
660
David Gibson1e8f8892006-01-06 00:10:44 -0800661 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -0700662 entry =
663 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
664 } else {
665 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
666 }
667 entry = pte_mkyoung(entry);
668 entry = pte_mkhuge(entry);
669
670 return entry;
671}
672
David Gibson1e8f8892006-01-06 00:10:44 -0800673static void set_huge_ptep_writable(struct vm_area_struct *vma,
674 unsigned long address, pte_t *ptep)
675{
676 pte_t entry;
677
678 entry = pte_mkwrite(pte_mkdirty(*ptep));
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -0700679 if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
680 update_mmu_cache(vma, address, entry);
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -0700681 }
David Gibson1e8f8892006-01-06 00:10:44 -0800682}
683
684
David Gibson63551ae2005-06-21 17:14:44 -0700685int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
686 struct vm_area_struct *vma)
687{
688 pte_t *src_pte, *dst_pte, entry;
689 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -0700690 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -0800691 int cow;
692
693 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -0700694
Hugh Dickins1c598272005-10-19 21:23:43 -0700695 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
Hugh Dickinsc74df322005-10-29 18:16:23 -0700696 src_pte = huge_pte_offset(src, addr);
697 if (!src_pte)
698 continue;
David Gibson63551ae2005-06-21 17:14:44 -0700699 dst_pte = huge_pte_alloc(dst, addr);
700 if (!dst_pte)
701 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -0800702
703 /* If the pagetables are shared don't copy or take references */
704 if (dst_pte == src_pte)
705 continue;
706
Hugh Dickinsc74df322005-10-29 18:16:23 -0700707 spin_lock(&dst->page_table_lock);
Hugh Dickins1c598272005-10-19 21:23:43 -0700708 spin_lock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700709 if (!pte_none(*src_pte)) {
David Gibson1e8f8892006-01-06 00:10:44 -0800710 if (cow)
711 ptep_set_wrprotect(src, addr, src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -0700712 entry = *src_pte;
713 ptepage = pte_page(entry);
714 get_page(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -0700715 set_huge_pte_at(dst, addr, dst_pte, entry);
716 }
717 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700718 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700719 }
720 return 0;
721
722nomem:
723 return -ENOMEM;
724}
725
Chen, Kenneth W502717f2006-10-11 01:20:46 -0700726void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
727 unsigned long end)
David Gibson63551ae2005-06-21 17:14:44 -0700728{
729 struct mm_struct *mm = vma->vm_mm;
730 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -0700731 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -0700732 pte_t pte;
733 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700734 struct page *tmp;
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -0800735 /*
736 * A page gathering list, protected by per file i_mmap_lock. The
737 * lock is used to avoid list corruption from multiple unmapping
738 * of the same page since we are using page->lru.
739 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700740 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -0700741
742 WARN_ON(!is_vm_hugetlb_page(vma));
743 BUG_ON(start & ~HPAGE_MASK);
744 BUG_ON(end & ~HPAGE_MASK);
745
Hugh Dickins508034a2005-10-29 18:16:30 -0700746 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700747 for (address = start; address < end; address += HPAGE_SIZE) {
David Gibsonc7546f82005-08-05 11:59:35 -0700748 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -0700749 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -0700750 continue;
751
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800752 if (huge_pmd_unshare(mm, &address, ptep))
753 continue;
754
David Gibsonc7546f82005-08-05 11:59:35 -0700755 pte = huge_ptep_get_and_clear(mm, address, ptep);
David Gibson63551ae2005-06-21 17:14:44 -0700756 if (pte_none(pte))
757 continue;
David Gibsonc7546f82005-08-05 11:59:35 -0700758
David Gibson63551ae2005-06-21 17:14:44 -0700759 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -0800760 if (pte_dirty(pte))
761 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700762 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -0700763 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -0700765 flush_tlb_range(vma, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700766 list_for_each_entry_safe(page, tmp, &page_list, lru) {
767 list_del(&page->lru);
768 put_page(page);
769 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770}
David Gibson63551ae2005-06-21 17:14:44 -0700771
Chen, Kenneth W502717f2006-10-11 01:20:46 -0700772void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
773 unsigned long end)
774{
775 /*
776 * It is undesirable to test vma->vm_file as it should be non-null
777 * for valid hugetlb area. However, vm_file will be NULL in the error
778 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
779 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
780 * to clean up. Since no pte has actually been setup, it is safe to
781 * do nothing in this case.
782 */
783 if (vma->vm_file) {
784 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
785 __unmap_hugepage_range(vma, start, end);
786 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
787 }
788}
789
David Gibson1e8f8892006-01-06 00:10:44 -0800790static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
791 unsigned long address, pte_t *ptep, pte_t pte)
792{
793 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -0800794 int avoidcopy;
David Gibson1e8f8892006-01-06 00:10:44 -0800795
796 old_page = pte_page(pte);
797
798 /* If no-one else is actually using this page, avoid the copy
799 * and just make the page writable */
800 avoidcopy = (page_count(old_page) == 1);
801 if (avoidcopy) {
802 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -0700803 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800804 }
805
806 page_cache_get(old_page);
Christoph Lameter5da7ca82006-01-06 00:10:46 -0800807 new_page = alloc_huge_page(vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -0800808
Adam Litke2fc39ce2007-11-14 16:59:39 -0800809 if (IS_ERR(new_page)) {
David Gibson1e8f8892006-01-06 00:10:44 -0800810 page_cache_release(old_page);
Adam Litke2fc39ce2007-11-14 16:59:39 -0800811 return -PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -0800812 }
813
814 spin_unlock(&mm->page_table_lock);
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000815 copy_huge_page(new_page, old_page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -0800816 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -0800817 spin_lock(&mm->page_table_lock);
818
819 ptep = huge_pte_offset(mm, address & HPAGE_MASK);
820 if (likely(pte_same(*ptep, pte))) {
821 /* Break COW */
822 set_huge_pte_at(mm, address, ptep,
823 make_huge_pte(vma, new_page, 1));
824 /* Make the old page be freed below */
825 new_page = old_page;
826 }
827 page_cache_release(new_page);
828 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -0700829 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800830}
831
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -0700832static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
David Gibson1e8f8892006-01-06 00:10:44 -0800833 unsigned long address, pte_t *ptep, int write_access)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100834{
835 int ret = VM_FAULT_SIGBUS;
Adam Litke4c887262005-10-29 18:16:46 -0700836 unsigned long idx;
837 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -0700838 struct page *page;
839 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -0800840 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -0700841
Adam Litke4c887262005-10-29 18:16:46 -0700842 mapping = vma->vm_file->f_mapping;
843 idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
844 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
845
846 /*
847 * Use page lock to guard against racing truncation
848 * before we get page_table_lock.
849 */
Christoph Lameter6bda6662006-01-06 00:10:49 -0800850retry:
851 page = find_lock_page(mapping, idx);
852 if (!page) {
Hugh Dickinsebed4bf2006-10-28 10:38:43 -0700853 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
854 if (idx >= size)
855 goto out;
Christoph Lameter6bda6662006-01-06 00:10:49 -0800856 page = alloc_huge_page(vma, address);
Adam Litke2fc39ce2007-11-14 16:59:39 -0800857 if (IS_ERR(page)) {
858 ret = -PTR_ERR(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -0800859 goto out;
860 }
David Gibson79ac6ba2006-03-22 00:08:51 -0800861 clear_huge_page(page, address);
Nick Piggin0ed361d2008-02-04 22:29:34 -0800862 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100863
Christoph Lameter6bda6662006-01-06 00:10:49 -0800864 if (vma->vm_flags & VM_SHARED) {
865 int err;
Ken Chen45c682a2007-11-14 16:59:44 -0800866 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -0800867
868 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
869 if (err) {
870 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -0800871 if (err == -EEXIST)
872 goto retry;
873 goto out;
874 }
Ken Chen45c682a2007-11-14 16:59:44 -0800875
876 spin_lock(&inode->i_lock);
877 inode->i_blocks += BLOCKS_PER_HUGEPAGE;
878 spin_unlock(&inode->i_lock);
Christoph Lameter6bda6662006-01-06 00:10:49 -0800879 } else
880 lock_page(page);
881 }
David Gibson1e8f8892006-01-06 00:10:44 -0800882
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100883 spin_lock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700884 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
885 if (idx >= size)
886 goto backout;
887
Nick Piggin83c54072007-07-19 01:47:05 -0700888 ret = 0;
Adam Litke86e52162006-01-06 00:10:43 -0800889 if (!pte_none(*ptep))
Adam Litke4c887262005-10-29 18:16:46 -0700890 goto backout;
891
David Gibson1e8f8892006-01-06 00:10:44 -0800892 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
893 && (vma->vm_flags & VM_SHARED)));
894 set_huge_pte_at(mm, address, ptep, new_pte);
895
896 if (write_access && !(vma->vm_flags & VM_SHARED)) {
897 /* Optimization, do the COW without a second fault */
898 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
899 }
900
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100901 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700902 unlock_page(page);
903out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100904 return ret;
Adam Litke4c887262005-10-29 18:16:46 -0700905
906backout:
907 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700908 unlock_page(page);
909 put_page(page);
910 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100911}
912
Adam Litke86e52162006-01-06 00:10:43 -0800913int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
914 unsigned long address, int write_access)
915{
916 pte_t *ptep;
917 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -0800918 int ret;
David Gibson3935baa2006-03-22 00:08:53 -0800919 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800920
921 ptep = huge_pte_alloc(mm, address);
922 if (!ptep)
923 return VM_FAULT_OOM;
924
David Gibson3935baa2006-03-22 00:08:53 -0800925 /*
926 * Serialize hugepage allocation and instantiation, so that we don't
927 * get spurious allocation failures if two CPUs race to instantiate
928 * the same page in the page cache.
929 */
930 mutex_lock(&hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800931 entry = *ptep;
David Gibson3935baa2006-03-22 00:08:53 -0800932 if (pte_none(entry)) {
933 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
934 mutex_unlock(&hugetlb_instantiation_mutex);
935 return ret;
936 }
Adam Litke86e52162006-01-06 00:10:43 -0800937
Nick Piggin83c54072007-07-19 01:47:05 -0700938 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800939
940 spin_lock(&mm->page_table_lock);
941 /* Check for a racing update before calling hugetlb_cow */
942 if (likely(pte_same(entry, *ptep)))
943 if (write_access && !pte_write(entry))
944 ret = hugetlb_cow(mm, vma, address, ptep, entry);
945 spin_unlock(&mm->page_table_lock);
David Gibson3935baa2006-03-22 00:08:53 -0800946 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -0800947
948 return ret;
Adam Litke86e52162006-01-06 00:10:43 -0800949}
950
David Gibson63551ae2005-06-21 17:14:44 -0700951int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
952 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -0800953 unsigned long *position, int *length, int i,
954 int write)
David Gibson63551ae2005-06-21 17:14:44 -0700955{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800956 unsigned long pfn_offset;
957 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -0700958 int remainder = *length;
959
Hugh Dickins1c598272005-10-19 21:23:43 -0700960 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700961 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -0700962 pte_t *pte;
963 struct page *page;
964
965 /*
966 * Some archs (sparc64, sh*) have multiple pte_ts to
967 * each hugepage. We have to make * sure we get the
968 * first, for the page indexing below to work.
969 */
970 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
971
Adam Litke72fad712007-12-10 15:49:28 -0800972 if (!pte || pte_none(*pte) || (write && !pte_write(*pte))) {
Adam Litke4c887262005-10-29 18:16:46 -0700973 int ret;
974
975 spin_unlock(&mm->page_table_lock);
Adam Litke5b23dbe2007-11-14 16:59:33 -0800976 ret = hugetlb_fault(mm, vma, vaddr, write);
Adam Litke4c887262005-10-29 18:16:46 -0700977 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -0700978 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -0700979 continue;
980
981 remainder = 0;
982 if (!i)
983 i = -EFAULT;
984 break;
985 }
David Gibson63551ae2005-06-21 17:14:44 -0700986
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800987 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
988 page = pte_page(*pte);
989same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -0800990 if (pages) {
991 get_page(page);
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800992 pages[i] = page + pfn_offset;
Chen, Kenneth Wd6692182006-03-31 02:29:57 -0800993 }
David Gibson63551ae2005-06-21 17:14:44 -0700994
995 if (vmas)
996 vmas[i] = vma;
997
998 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800999 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07001000 --remainder;
1001 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08001002 if (vaddr < vma->vm_end && remainder &&
1003 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1004 /*
1005 * We use pfn_offset to avoid touching the pageframes
1006 * of this compound page.
1007 */
1008 goto same_page;
1009 }
David Gibson63551ae2005-06-21 17:14:44 -07001010 }
Hugh Dickins1c598272005-10-19 21:23:43 -07001011 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07001012 *length = remainder;
1013 *position = vaddr;
1014
1015 return i;
1016}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001017
1018void hugetlb_change_protection(struct vm_area_struct *vma,
1019 unsigned long address, unsigned long end, pgprot_t newprot)
1020{
1021 struct mm_struct *mm = vma->vm_mm;
1022 unsigned long start = address;
1023 pte_t *ptep;
1024 pte_t pte;
1025
1026 BUG_ON(address >= end);
1027 flush_cache_range(vma, address, end);
1028
Chen, Kenneth W39dde652006-12-06 20:32:03 -08001029 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001030 spin_lock(&mm->page_table_lock);
1031 for (; address < end; address += HPAGE_SIZE) {
1032 ptep = huge_pte_offset(mm, address);
1033 if (!ptep)
1034 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08001035 if (huge_pmd_unshare(mm, &address, ptep))
1036 continue;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001037 if (!pte_none(*ptep)) {
1038 pte = huge_ptep_get_and_clear(mm, address, ptep);
1039 pte = pte_mkhuge(pte_modify(pte, newprot));
1040 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001041 }
1042 }
1043 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08001044 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001045
1046 flush_tlb_range(vma, start, end);
1047}
1048
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001049struct file_region {
1050 struct list_head link;
1051 long from;
1052 long to;
1053};
1054
1055static long region_add(struct list_head *head, long f, long t)
1056{
1057 struct file_region *rg, *nrg, *trg;
1058
1059 /* Locate the region we are either in or before. */
1060 list_for_each_entry(rg, head, link)
1061 if (f <= rg->to)
1062 break;
1063
1064 /* Round our left edge to the current segment if it encloses us. */
1065 if (f > rg->from)
1066 f = rg->from;
1067
1068 /* Check for and consume any regions we now overlap with. */
1069 nrg = rg;
1070 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1071 if (&rg->link == head)
1072 break;
1073 if (rg->from > t)
1074 break;
1075
1076 /* If this area reaches higher then extend our area to
1077 * include it completely. If this is not the first area
1078 * which we intend to reuse, free it. */
1079 if (rg->to > t)
1080 t = rg->to;
1081 if (rg != nrg) {
1082 list_del(&rg->link);
1083 kfree(rg);
1084 }
1085 }
1086 nrg->from = f;
1087 nrg->to = t;
1088 return 0;
1089}
1090
1091static long region_chg(struct list_head *head, long f, long t)
1092{
1093 struct file_region *rg, *nrg;
1094 long chg = 0;
1095
1096 /* Locate the region we are before or in. */
1097 list_for_each_entry(rg, head, link)
1098 if (f <= rg->to)
1099 break;
1100
1101 /* If we are below the current region then a new region is required.
1102 * Subtle, allocate a new region at the position but make it zero
Simon Arlott183ff222007-10-20 01:27:18 +02001103 * size such that we can guarantee to record the reservation. */
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001104 if (&rg->link == head || t < rg->from) {
1105 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001106 if (!nrg)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001107 return -ENOMEM;
1108 nrg->from = f;
1109 nrg->to = f;
1110 INIT_LIST_HEAD(&nrg->link);
1111 list_add(&nrg->link, rg->link.prev);
1112
1113 return t - f;
1114 }
1115
1116 /* Round our left edge to the current segment if it encloses us. */
1117 if (f > rg->from)
1118 f = rg->from;
1119 chg = t - f;
1120
1121 /* Check for and consume any regions we now overlap with. */
1122 list_for_each_entry(rg, rg->link.prev, link) {
1123 if (&rg->link == head)
1124 break;
1125 if (rg->from > t)
1126 return chg;
1127
1128 /* We overlap with this area, if it extends futher than
1129 * us then we must extend ourselves. Account for its
1130 * existing reservation. */
1131 if (rg->to > t) {
1132 chg += rg->to - t;
1133 t = rg->to;
1134 }
1135 chg -= rg->to - rg->from;
1136 }
1137 return chg;
1138}
1139
1140static long region_truncate(struct list_head *head, long end)
1141{
1142 struct file_region *rg, *trg;
1143 long chg = 0;
1144
1145 /* Locate the region we are either in or before. */
1146 list_for_each_entry(rg, head, link)
1147 if (end <= rg->to)
1148 break;
1149 if (&rg->link == head)
1150 return 0;
1151
1152 /* If we are in the middle of a region then adjust it. */
1153 if (end > rg->from) {
1154 chg = rg->to - end;
1155 rg->to = end;
1156 rg = list_entry(rg->link.next, typeof(*rg), link);
1157 }
1158
1159 /* Drop any remaining regions. */
1160 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1161 if (&rg->link == head)
1162 break;
1163 chg += rg->to - rg->from;
1164 list_del(&rg->link);
1165 kfree(rg);
1166 }
1167 return chg;
1168}
1169
1170static int hugetlb_acct_memory(long delta)
1171{
1172 int ret = -ENOMEM;
1173
1174 spin_lock(&hugetlb_lock);
Ken Chen8a630112007-05-09 02:33:34 -07001175 /*
1176 * When cpuset is configured, it breaks the strict hugetlb page
1177 * reservation as the accounting is done on a global variable. Such
1178 * reservation is completely rubbish in the presence of cpuset because
1179 * the reservation is not checked against page availability for the
1180 * current cpuset. Application can still potentially OOM'ed by kernel
1181 * with lack of free htlb page in cpuset that the task is in.
1182 * Attempt to enforce strict accounting with cpuset is almost
1183 * impossible (or too ugly) because cpuset is too fluid that
1184 * task or memory node can be dynamically moved between cpusets.
1185 *
1186 * The change of semantics for shared hugetlb mapping with cpuset is
1187 * undesirable. However, in order to preserve some of the semantics,
1188 * we fall back to check against current free page availability as
1189 * a best attempt and hopefully to minimize the impact of changing
1190 * semantics that cpuset has.
1191 */
Adam Litkee4e574b2007-10-16 01:26:19 -07001192 if (delta > 0) {
1193 if (gather_surplus_pages(delta) < 0)
1194 goto out;
1195
1196 if (delta > cpuset_mems_nr(free_huge_pages_node))
1197 goto out;
1198 }
1199
1200 ret = 0;
1201 resv_huge_pages += delta;
1202 if (delta < 0)
1203 return_unused_surplus_pages((unsigned long) -delta);
1204
1205out:
1206 spin_unlock(&hugetlb_lock);
1207 return ret;
1208}
1209
1210int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1211{
1212 long ret, chg;
1213
1214 chg = region_chg(&inode->i_mapping->private_list, from, to);
1215 if (chg < 0)
1216 return chg;
Ken Chen8a630112007-05-09 02:33:34 -07001217
Adam Litke90d8b7e2007-11-14 16:59:42 -08001218 if (hugetlb_get_quota(inode->i_mapping, chg))
1219 return -ENOSPC;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001220 ret = hugetlb_acct_memory(chg);
Ken Chen68842c92008-01-14 00:55:19 -08001221 if (ret < 0) {
1222 hugetlb_put_quota(inode->i_mapping, chg);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001223 return ret;
Ken Chen68842c92008-01-14 00:55:19 -08001224 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001225 region_add(&inode->i_mapping->private_list, from, to);
1226 return 0;
1227}
1228
1229void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1230{
1231 long chg = region_truncate(&inode->i_mapping->private_list, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08001232
1233 spin_lock(&inode->i_lock);
1234 inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1235 spin_unlock(&inode->i_lock);
1236
Adam Litke90d8b7e2007-11-14 16:59:42 -08001237 hugetlb_put_quota(inode->i_mapping, (chg - freed));
1238 hugetlb_acct_memory(-(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001239}