blob: 74c1b6b0b37b82dce75e06533c989ab73001afeb [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;
Nishanth Aravamudan064d9ef2008-02-13 15:03:19 -080027static unsigned long nr_overcommit_huge_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -070028unsigned long max_huge_pages;
Nishanth Aravamudan064d9ef2008-02-13 15:03:19 -080029unsigned long sysctl_overcommit_huge_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -070030static struct list_head hugepage_freelists[MAX_NUMNODES];
31static unsigned int nr_huge_pages_node[MAX_NUMNODES];
32static unsigned int free_huge_pages_node[MAX_NUMNODES];
Adam Litke7893d1d2007-10-16 01:26:18 -070033static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
Mel Gorman396faf02007-07-17 04:03:13 -070034static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35unsigned long hugepages_treat_as_movable;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -070036static int hugetlb_next_nid;
Mel Gorman396faf02007-07-17 04:03:13 -070037
David Gibson3935baa2006-03-22 00:08:53 -080038/*
39 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
40 */
41static DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080042
David Gibson79ac6ba2006-03-22 00:08:51 -080043static void clear_huge_page(struct page *page, unsigned long addr)
44{
45 int i;
46
47 might_sleep();
48 for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
49 cond_resched();
Ralf Baechle281e0e32007-10-01 01:20:10 -070050 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
David Gibson79ac6ba2006-03-22 00:08:51 -080051 }
52}
53
54static void copy_huge_page(struct page *dst, struct page *src,
Atsushi Nemoto9de455b2006-12-12 17:14:55 +000055 unsigned long addr, struct vm_area_struct *vma)
David Gibson79ac6ba2006-03-22 00:08:51 -080056{
57 int i;
58
59 might_sleep();
60 for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
61 cond_resched();
Atsushi Nemoto9de455b2006-12-12 17:14:55 +000062 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
David Gibson79ac6ba2006-03-22 00:08:51 -080063 }
64}
65
Linus Torvalds1da177e2005-04-16 15:20:36 -070066static void enqueue_huge_page(struct page *page)
67{
68 int nid = page_to_nid(page);
69 list_add(&page->lru, &hugepage_freelists[nid]);
70 free_huge_pages++;
71 free_huge_pages_node[nid]++;
72}
73
Nishanth Aravamudan348e1e02008-03-04 14:29:42 -080074static struct page *dequeue_huge_page(void)
75{
76 int nid;
77 struct page *page = NULL;
78
79 for (nid = 0; nid < MAX_NUMNODES; ++nid) {
80 if (!list_empty(&hugepage_freelists[nid])) {
81 page = list_entry(hugepage_freelists[nid].next,
82 struct page, lru);
83 list_del(&page->lru);
84 free_huge_pages--;
85 free_huge_pages_node[nid]--;
86 break;
87 }
88 }
89 return page;
90}
91
92static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
Christoph Lameter5da7ca82006-01-06 00:10:46 -080093 unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -070094{
Nishanth Aravamudan31a5c6e2007-07-15 23:38:02 -070095 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -070096 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -070097 struct mempolicy *mpol;
Mel Gorman396faf02007-07-17 04:03:13 -070098 struct zonelist *zonelist = huge_zonelist(vma, address,
Lee Schermerhorn480eccf2007-09-18 22:46:47 -070099 htlb_alloc_mask, &mpol);
Christoph Lameter96df9332006-01-06 00:10:45 -0800100 struct zone **z;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101
Christoph Lameter96df9332006-01-06 00:10:45 -0800102 for (z = zonelist->zones; *z; z++) {
Christoph Lameter89fa3022006-09-25 23:31:55 -0700103 nid = zone_to_nid(*z);
Mel Gorman396faf02007-07-17 04:03:13 -0700104 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
Andrew Morton3abf7af2007-07-19 01:49:08 -0700105 !list_empty(&hugepage_freelists[nid])) {
106 page = list_entry(hugepage_freelists[nid].next,
107 struct page, lru);
108 list_del(&page->lru);
109 free_huge_pages--;
110 free_huge_pages_node[nid]--;
Adam Litkee4e574b2007-10-16 01:26:19 -0700111 if (vma && vma->vm_flags & VM_MAYSHARE)
112 resv_huge_pages--;
Ken Chen5ab3ee72007-07-23 18:44:00 -0700113 break;
Andrew Morton3abf7af2007-07-19 01:49:08 -0700114 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115 }
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700116 mpol_free(mpol); /* unref if mpol !NULL */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117 return page;
118}
119
Adam Litke6af2acb2007-10-16 01:26:16 -0700120static void update_and_free_page(struct page *page)
121{
122 int i;
123 nr_huge_pages--;
124 nr_huge_pages_node[page_to_nid(page)]--;
125 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
126 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
127 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
128 1 << PG_private | 1<< PG_writeback);
129 }
130 set_compound_page_dtor(page, NULL);
131 set_page_refcounted(page);
132 __free_pages(page, HUGETLB_PAGE_ORDER);
133}
134
David Gibson27a85ef2006-03-22 00:08:56 -0800135static void free_huge_page(struct page *page)
136{
Adam Litke7893d1d2007-10-16 01:26:18 -0700137 int nid = page_to_nid(page);
Adam Litkec79fb752007-11-14 16:59:38 -0800138 struct address_space *mapping;
David Gibson27a85ef2006-03-22 00:08:56 -0800139
Adam Litkec79fb752007-11-14 16:59:38 -0800140 mapping = (struct address_space *) page_private(page);
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800141 set_page_private(page, 0);
Adam Litke7893d1d2007-10-16 01:26:18 -0700142 BUG_ON(page_count(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800143 INIT_LIST_HEAD(&page->lru);
144
145 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700146 if (surplus_huge_pages_node[nid]) {
147 update_and_free_page(page);
148 surplus_huge_pages--;
149 surplus_huge_pages_node[nid]--;
150 } else {
151 enqueue_huge_page(page);
152 }
David Gibson27a85ef2006-03-22 00:08:56 -0800153 spin_unlock(&hugetlb_lock);
Adam Litkec79fb752007-11-14 16:59:38 -0800154 if (mapping)
Adam Litke9a119c02007-11-14 16:59:41 -0800155 hugetlb_put_quota(mapping, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800156}
157
Adam Litke7893d1d2007-10-16 01:26:18 -0700158/*
159 * Increment or decrement surplus_huge_pages. Keep node-specific counters
160 * balanced by operating on them in a round-robin fashion.
161 * Returns 1 if an adjustment was made.
162 */
163static int adjust_pool_surplus(int delta)
164{
165 static int prev_nid;
166 int nid = prev_nid;
167 int ret = 0;
168
169 VM_BUG_ON(delta != -1 && delta != 1);
170 do {
171 nid = next_node(nid, node_online_map);
172 if (nid == MAX_NUMNODES)
173 nid = first_node(node_online_map);
174
175 /* To shrink on this node, there must be a surplus page */
176 if (delta < 0 && !surplus_huge_pages_node[nid])
177 continue;
178 /* Surplus cannot exceed the total number of pages */
179 if (delta > 0 && surplus_huge_pages_node[nid] >=
180 nr_huge_pages_node[nid])
181 continue;
182
183 surplus_huge_pages += delta;
184 surplus_huge_pages_node[nid] += delta;
185 ret = 1;
186 break;
187 } while (nid != prev_nid);
188
189 prev_nid = nid;
190 return ret;
191}
192
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700193static struct page *alloc_fresh_huge_page_node(int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700196
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700197 page = alloc_pages_node(nid,
198 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
199 HUGETLB_PAGE_ORDER);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200 if (page) {
Andy Whitcroft33f2ef82006-12-06 20:33:32 -0800201 set_compound_page_dtor(page, free_huge_page);
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800202 spin_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203 nr_huge_pages++;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700204 nr_huge_pages_node[nid]++;
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800205 spin_unlock(&hugetlb_lock);
Nick Piggina4822892006-03-22 00:08:08 -0800206 put_page(page); /* free it into the hugepage allocator */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700208
209 return page;
210}
211
212static int alloc_fresh_huge_page(void)
213{
214 struct page *page;
215 int start_nid;
216 int next_nid;
217 int ret = 0;
218
219 start_nid = hugetlb_next_nid;
220
221 do {
222 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
223 if (page)
224 ret = 1;
225 /*
226 * Use a helper variable to find the next node and then
227 * copy it back to hugetlb_next_nid afterwards:
228 * otherwise there's a window in which a racer might
229 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
230 * But we don't need to use a spin_lock here: it really
231 * doesn't matter if occasionally a racer chooses the
232 * same nid as we do. Move nid forward in the mask even
233 * if we just successfully allocated a hugepage so that
234 * the next caller gets hugepages on the next node.
235 */
236 next_nid = next_node(hugetlb_next_nid, node_online_map);
237 if (next_nid == MAX_NUMNODES)
238 next_nid = first_node(node_online_map);
239 hugetlb_next_nid = next_nid;
240 } while (!page && hugetlb_next_nid != start_nid);
241
242 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243}
244
Adam Litke7893d1d2007-10-16 01:26:18 -0700245static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
246 unsigned long address)
247{
248 struct page *page;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800249 unsigned int nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700250
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800251 /*
252 * Assume we will successfully allocate the surplus page to
253 * prevent racing processes from causing the surplus to exceed
254 * overcommit
255 *
256 * This however introduces a different race, where a process B
257 * tries to grow the static hugepage pool while alloc_pages() is
258 * called by process A. B will only examine the per-node
259 * counters in determining if surplus huge pages can be
260 * converted to normal huge pages in adjust_pool_surplus(). A
261 * won't be able to increment the per-node counter, until the
262 * lock is dropped by B, but B doesn't drop hugetlb_lock until
263 * no more huge pages can be converted from surplus to normal
264 * state (and doesn't try to convert again). Thus, we have a
265 * case where a surplus huge page exists, the pool is grown, and
266 * the surplus huge page still exists after, even though it
267 * should just have been converted to a normal huge page. This
268 * does not leak memory, though, as the hugepage will be freed
269 * once it is out of use. It also does not allow the counters to
270 * go out of whack in adjust_pool_surplus() as we don't modify
271 * the node values until we've gotten the hugepage and only the
272 * per-node value is checked there.
273 */
274 spin_lock(&hugetlb_lock);
275 if (surplus_huge_pages >= nr_overcommit_huge_pages) {
276 spin_unlock(&hugetlb_lock);
277 return NULL;
278 } else {
279 nr_huge_pages++;
280 surplus_huge_pages++;
281 }
282 spin_unlock(&hugetlb_lock);
283
Adam Litke7893d1d2007-10-16 01:26:18 -0700284 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
285 HUGETLB_PAGE_ORDER);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800286
287 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700288 if (page) {
Adam Litke2668db92008-03-10 11:43:50 -0700289 /*
290 * This page is now managed by the hugetlb allocator and has
291 * no users -- drop the buddy allocator's reference.
292 */
293 put_page_testzero(page);
294 VM_BUG_ON(page_count(page));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800295 nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700296 set_compound_page_dtor(page, free_huge_page);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800297 /*
298 * We incremented the global counters already
299 */
300 nr_huge_pages_node[nid]++;
301 surplus_huge_pages_node[nid]++;
302 } else {
303 nr_huge_pages--;
304 surplus_huge_pages--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700305 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800306 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700307
308 return page;
309}
310
Adam Litkee4e574b2007-10-16 01:26:19 -0700311/*
312 * Increase the hugetlb pool such that it can accomodate a reservation
313 * of size 'delta'.
314 */
315static int gather_surplus_pages(int delta)
316{
317 struct list_head surplus_list;
318 struct page *page, *tmp;
319 int ret, i;
320 int needed, allocated;
321
322 needed = (resv_huge_pages + delta) - free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800323 if (needed <= 0) {
324 resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700325 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800326 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700327
328 allocated = 0;
329 INIT_LIST_HEAD(&surplus_list);
330
331 ret = -ENOMEM;
332retry:
333 spin_unlock(&hugetlb_lock);
334 for (i = 0; i < needed; i++) {
335 page = alloc_buddy_huge_page(NULL, 0);
336 if (!page) {
337 /*
338 * We were not able to allocate enough pages to
339 * satisfy the entire reservation so we free what
340 * we've allocated so far.
341 */
342 spin_lock(&hugetlb_lock);
343 needed = 0;
344 goto free;
345 }
346
347 list_add(&page->lru, &surplus_list);
348 }
349 allocated += needed;
350
351 /*
352 * After retaking hugetlb_lock, we need to recalculate 'needed'
353 * because either resv_huge_pages or free_huge_pages may have changed.
354 */
355 spin_lock(&hugetlb_lock);
356 needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
357 if (needed > 0)
358 goto retry;
359
360 /*
361 * The surplus_list now contains _at_least_ the number of extra pages
362 * needed to accomodate the reservation. Add the appropriate number
363 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -0800364 * allocator. Commit the entire reservation here to prevent another
365 * process from stealing the pages as they are added to the pool but
366 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -0700367 */
368 needed += allocated;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800369 resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700370 ret = 0;
371free:
372 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
373 list_del(&page->lru);
374 if ((--needed) >= 0)
375 enqueue_huge_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700376 else {
377 /*
Adam Litke2668db92008-03-10 11:43:50 -0700378 * The page has a reference count of zero already, so
379 * call free_huge_page directly instead of using
380 * put_page. This must be done with hugetlb_lock
Adam Litkeaf767cb2007-10-16 01:26:25 -0700381 * unlocked which is safe because free_huge_page takes
382 * hugetlb_lock before deciding how to free the page.
383 */
384 spin_unlock(&hugetlb_lock);
Adam Litke2668db92008-03-10 11:43:50 -0700385 free_huge_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700386 spin_lock(&hugetlb_lock);
387 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700388 }
389
390 return ret;
391}
392
393/*
394 * When releasing a hugetlb pool reservation, any surplus pages that were
395 * allocated to satisfy the reservation must be explicitly freed if they were
396 * never used.
397 */
Adrian Bunk8cde0452007-11-14 16:59:43 -0800398static void return_unused_surplus_pages(unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -0700399{
400 static int nid = -1;
401 struct page *page;
402 unsigned long nr_pages;
403
Adam Litkeac09b3a2008-03-04 14:29:38 -0800404 /* Uncommit the reservation */
405 resv_huge_pages -= unused_resv_pages;
406
Adam Litkee4e574b2007-10-16 01:26:19 -0700407 nr_pages = min(unused_resv_pages, surplus_huge_pages);
408
409 while (nr_pages) {
410 nid = next_node(nid, node_online_map);
411 if (nid == MAX_NUMNODES)
412 nid = first_node(node_online_map);
413
414 if (!surplus_huge_pages_node[nid])
415 continue;
416
417 if (!list_empty(&hugepage_freelists[nid])) {
418 page = list_entry(hugepage_freelists[nid].next,
419 struct page, lru);
420 list_del(&page->lru);
421 update_and_free_page(page);
422 free_huge_pages--;
423 free_huge_pages_node[nid]--;
424 surplus_huge_pages--;
425 surplus_huge_pages_node[nid]--;
426 nr_pages--;
427 }
428 }
429}
430
Adam Litke348ea202007-11-14 16:59:37 -0800431
432static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
433 unsigned long addr)
434{
435 struct page *page;
436
437 spin_lock(&hugetlb_lock);
Nishanth Aravamudan348e1e02008-03-04 14:29:42 -0800438 page = dequeue_huge_page_vma(vma, addr);
Adam Litke348ea202007-11-14 16:59:37 -0800439 spin_unlock(&hugetlb_lock);
Adam Litke90d8b7e2007-11-14 16:59:42 -0800440 return page ? page : ERR_PTR(-VM_FAULT_OOM);
Adam Litke348ea202007-11-14 16:59:37 -0800441}
442
443static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
444 unsigned long addr)
445{
446 struct page *page = NULL;
447
Adam Litke90d8b7e2007-11-14 16:59:42 -0800448 if (hugetlb_get_quota(vma->vm_file->f_mapping, 1))
449 return ERR_PTR(-VM_FAULT_SIGBUS);
450
Adam Litke348ea202007-11-14 16:59:37 -0800451 spin_lock(&hugetlb_lock);
452 if (free_huge_pages > resv_huge_pages)
Nishanth Aravamudan348e1e02008-03-04 14:29:42 -0800453 page = dequeue_huge_page_vma(vma, addr);
Adam Litke348ea202007-11-14 16:59:37 -0800454 spin_unlock(&hugetlb_lock);
Ken Chen68842c92008-01-14 00:55:19 -0800455 if (!page) {
Adam Litke348ea202007-11-14 16:59:37 -0800456 page = alloc_buddy_huge_page(vma, addr);
Ken Chen68842c92008-01-14 00:55:19 -0800457 if (!page) {
458 hugetlb_put_quota(vma->vm_file->f_mapping, 1);
459 return ERR_PTR(-VM_FAULT_OOM);
460 }
461 }
462 return page;
Adam Litke348ea202007-11-14 16:59:37 -0800463}
464
David Gibson27a85ef2006-03-22 00:08:56 -0800465static struct page *alloc_huge_page(struct vm_area_struct *vma,
466 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467{
Adam Litke348ea202007-11-14 16:59:37 -0800468 struct page *page;
Adam Litke2fc39ce2007-11-14 16:59:39 -0800469 struct address_space *mapping = vma->vm_file->f_mapping;
470
Adam Litke348ea202007-11-14 16:59:37 -0800471 if (vma->vm_flags & VM_MAYSHARE)
472 page = alloc_huge_page_shared(vma, addr);
473 else
474 page = alloc_huge_page_private(vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -0800475
476 if (!IS_ERR(page)) {
Adam Litke348ea202007-11-14 16:59:37 -0800477 set_page_refcounted(page);
Adam Litke2fc39ce2007-11-14 16:59:39 -0800478 set_page_private(page, (unsigned long) mapping);
Adam Litke90d8b7e2007-11-14 16:59:42 -0800479 }
480 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800481}
482
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483static int __init hugetlb_init(void)
484{
485 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486
Benjamin Herrenschmidt3c726f82005-11-07 11:06:55 +1100487 if (HPAGE_SHIFT == 0)
488 return 0;
489
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 for (i = 0; i < MAX_NUMNODES; ++i)
491 INIT_LIST_HEAD(&hugepage_freelists[i]);
492
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700493 hugetlb_next_nid = first_node(node_online_map);
494
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495 for (i = 0; i < max_huge_pages; ++i) {
Nick Piggina4822892006-03-22 00:08:08 -0800496 if (!alloc_fresh_huge_page())
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 }
499 max_huge_pages = free_huge_pages = nr_huge_pages = i;
500 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
501 return 0;
502}
503module_init(hugetlb_init);
504
505static int __init hugetlb_setup(char *s)
506{
507 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
508 max_huge_pages = 0;
509 return 1;
510}
511__setup("hugepages=", hugetlb_setup);
512
Ken Chen8a630112007-05-09 02:33:34 -0700513static unsigned int cpuset_mems_nr(unsigned int *array)
514{
515 int node;
516 unsigned int nr = 0;
517
518 for_each_node_mask(node, cpuset_current_mems_allowed)
519 nr += array[node];
520
521 return nr;
522}
523
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524#ifdef CONFIG_SYSCTL
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525#ifdef CONFIG_HIGHMEM
526static void try_to_free_low(unsigned long count)
527{
Christoph Lameter4415cc82006-09-25 23:31:55 -0700528 int i;
529
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 for (i = 0; i < MAX_NUMNODES; ++i) {
531 struct page *page, *next;
532 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
Adam Litke6b0c8802007-10-16 01:26:23 -0700533 if (count >= nr_huge_pages)
534 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 if (PageHighMem(page))
536 continue;
537 list_del(&page->lru);
538 update_and_free_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539 free_huge_pages--;
Christoph Lameter4415cc82006-09-25 23:31:55 -0700540 free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541 }
542 }
543}
544#else
545static inline void try_to_free_low(unsigned long count)
546{
547}
548#endif
549
Adam Litke7893d1d2007-10-16 01:26:18 -0700550#define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551static unsigned long set_max_huge_pages(unsigned long count)
552{
Adam Litke7893d1d2007-10-16 01:26:18 -0700553 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554
Adam Litke7893d1d2007-10-16 01:26:18 -0700555 /*
556 * Increase the pool size
557 * First take pages out of surplus state. Then make up the
558 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800559 *
560 * We might race with alloc_buddy_huge_page() here and be unable
561 * to convert a surplus huge page to a normal huge page. That is
562 * not critical, though, it just means the overall size of the
563 * pool might be one hugepage larger than it needs to be, but
564 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -0700565 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700567 while (surplus_huge_pages && count > persistent_huge_pages) {
568 if (!adjust_pool_surplus(-1))
569 break;
570 }
571
572 while (count > persistent_huge_pages) {
573 int ret;
574 /*
575 * If this allocation races such that we no longer need the
576 * page, free_huge_page will handle it by freeing the page
577 * and reducing the surplus.
578 */
579 spin_unlock(&hugetlb_lock);
580 ret = alloc_fresh_huge_page();
581 spin_lock(&hugetlb_lock);
582 if (!ret)
583 goto out;
584
585 }
Adam Litke7893d1d2007-10-16 01:26:18 -0700586
587 /*
588 * Decrease the pool size
589 * First return free pages to the buddy allocator (being careful
590 * to keep enough around to satisfy reservations). Then place
591 * pages into surplus state as needed so the pool will shrink
592 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800593 *
594 * By placing pages into the surplus state independent of the
595 * overcommit value, we are allowing the surplus pool size to
596 * exceed overcommit. There are few sane options here. Since
597 * alloc_buddy_huge_page() is checking the global counter,
598 * though, we'll note that we're not allowed to exceed surplus
599 * and won't grow the pool anywhere else. Not until one of the
600 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -0700601 */
Adam Litke6b0c8802007-10-16 01:26:23 -0700602 min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
603 min_count = max(count, min_count);
Adam Litke7893d1d2007-10-16 01:26:18 -0700604 try_to_free_low(min_count);
605 while (min_count < persistent_huge_pages) {
Nishanth Aravamudan348e1e02008-03-04 14:29:42 -0800606 struct page *page = dequeue_huge_page();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607 if (!page)
608 break;
609 update_and_free_page(page);
610 }
Adam Litke7893d1d2007-10-16 01:26:18 -0700611 while (count < persistent_huge_pages) {
612 if (!adjust_pool_surplus(1))
613 break;
614 }
615out:
616 ret = persistent_huge_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700618 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619}
620
621int hugetlb_sysctl_handler(struct ctl_table *table, int write,
622 struct file *file, void __user *buffer,
623 size_t *length, loff_t *ppos)
624{
625 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
626 max_huge_pages = set_max_huge_pages(max_huge_pages);
627 return 0;
628}
Mel Gorman396faf02007-07-17 04:03:13 -0700629
630int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
631 struct file *file, void __user *buffer,
632 size_t *length, loff_t *ppos)
633{
634 proc_dointvec(table, write, file, buffer, length, ppos);
635 if (hugepages_treat_as_movable)
636 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
637 else
638 htlb_alloc_mask = GFP_HIGHUSER;
639 return 0;
640}
641
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -0800642int hugetlb_overcommit_handler(struct ctl_table *table, int write,
643 struct file *file, void __user *buffer,
644 size_t *length, loff_t *ppos)
645{
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -0800646 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
Nishanth Aravamudan064d9ef2008-02-13 15:03:19 -0800647 spin_lock(&hugetlb_lock);
648 nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -0800649 spin_unlock(&hugetlb_lock);
650 return 0;
651}
652
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653#endif /* CONFIG_SYSCTL */
654
655int hugetlb_report_meminfo(char *buf)
656{
657 return sprintf(buf,
658 "HugePages_Total: %5lu\n"
659 "HugePages_Free: %5lu\n"
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700660 "HugePages_Rsvd: %5lu\n"
Adam Litke7893d1d2007-10-16 01:26:18 -0700661 "HugePages_Surp: %5lu\n"
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662 "Hugepagesize: %5lu kB\n",
663 nr_huge_pages,
664 free_huge_pages,
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700665 resv_huge_pages,
Adam Litke7893d1d2007-10-16 01:26:18 -0700666 surplus_huge_pages,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 HPAGE_SIZE/1024);
668}
669
670int hugetlb_report_node_meminfo(int nid, char *buf)
671{
672 return sprintf(buf,
673 "Node %d HugePages_Total: %5u\n"
674 "Node %d HugePages_Free: %5u\n",
675 nid, nr_huge_pages_node[nid],
676 nid, free_huge_pages_node[nid]);
677}
678
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
680unsigned long hugetlb_total_pages(void)
681{
682 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
683}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684
685/*
686 * We cannot handle pagefaults against hugetlb pages at all. They cause
687 * handle_mm_fault() to try to instantiate regular-sized pages in the
688 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
689 * this far.
690 */
Nick Piggind0217ac2007-07-19 01:47:03 -0700691static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692{
693 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -0700694 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695}
696
697struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -0700698 .fault = hugetlb_vm_op_fault,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699};
700
David Gibson1e8f8892006-01-06 00:10:44 -0800701static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
702 int writable)
David Gibson63551ae2005-06-21 17:14:44 -0700703{
704 pte_t entry;
705
David Gibson1e8f8892006-01-06 00:10:44 -0800706 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -0700707 entry =
708 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
709 } else {
710 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
711 }
712 entry = pte_mkyoung(entry);
713 entry = pte_mkhuge(entry);
714
715 return entry;
716}
717
David Gibson1e8f8892006-01-06 00:10:44 -0800718static void set_huge_ptep_writable(struct vm_area_struct *vma,
719 unsigned long address, pte_t *ptep)
720{
721 pte_t entry;
722
723 entry = pte_mkwrite(pte_mkdirty(*ptep));
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -0700724 if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
725 update_mmu_cache(vma, address, entry);
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -0700726 }
David Gibson1e8f8892006-01-06 00:10:44 -0800727}
728
729
David Gibson63551ae2005-06-21 17:14:44 -0700730int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
731 struct vm_area_struct *vma)
732{
733 pte_t *src_pte, *dst_pte, entry;
734 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -0700735 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -0800736 int cow;
737
738 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -0700739
Hugh Dickins1c598272005-10-19 21:23:43 -0700740 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
Hugh Dickinsc74df322005-10-29 18:16:23 -0700741 src_pte = huge_pte_offset(src, addr);
742 if (!src_pte)
743 continue;
David Gibson63551ae2005-06-21 17:14:44 -0700744 dst_pte = huge_pte_alloc(dst, addr);
745 if (!dst_pte)
746 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -0800747
748 /* If the pagetables are shared don't copy or take references */
749 if (dst_pte == src_pte)
750 continue;
751
Hugh Dickinsc74df322005-10-29 18:16:23 -0700752 spin_lock(&dst->page_table_lock);
Hugh Dickins1c598272005-10-19 21:23:43 -0700753 spin_lock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700754 if (!pte_none(*src_pte)) {
David Gibson1e8f8892006-01-06 00:10:44 -0800755 if (cow)
756 ptep_set_wrprotect(src, addr, src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -0700757 entry = *src_pte;
758 ptepage = pte_page(entry);
759 get_page(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -0700760 set_huge_pte_at(dst, addr, dst_pte, entry);
761 }
762 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700763 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700764 }
765 return 0;
766
767nomem:
768 return -ENOMEM;
769}
770
Chen, Kenneth W502717f2006-10-11 01:20:46 -0700771void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
772 unsigned long end)
David Gibson63551ae2005-06-21 17:14:44 -0700773{
774 struct mm_struct *mm = vma->vm_mm;
775 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -0700776 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -0700777 pte_t pte;
778 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700779 struct page *tmp;
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -0800780 /*
781 * A page gathering list, protected by per file i_mmap_lock. The
782 * lock is used to avoid list corruption from multiple unmapping
783 * of the same page since we are using page->lru.
784 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700785 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -0700786
787 WARN_ON(!is_vm_hugetlb_page(vma));
788 BUG_ON(start & ~HPAGE_MASK);
789 BUG_ON(end & ~HPAGE_MASK);
790
Hugh Dickins508034a2005-10-29 18:16:30 -0700791 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700792 for (address = start; address < end; address += HPAGE_SIZE) {
David Gibsonc7546f82005-08-05 11:59:35 -0700793 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -0700794 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -0700795 continue;
796
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800797 if (huge_pmd_unshare(mm, &address, ptep))
798 continue;
799
David Gibsonc7546f82005-08-05 11:59:35 -0700800 pte = huge_ptep_get_and_clear(mm, address, ptep);
David Gibson63551ae2005-06-21 17:14:44 -0700801 if (pte_none(pte))
802 continue;
David Gibsonc7546f82005-08-05 11:59:35 -0700803
David Gibson63551ae2005-06-21 17:14:44 -0700804 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -0800805 if (pte_dirty(pte))
806 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700807 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -0700808 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -0700810 flush_tlb_range(vma, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700811 list_for_each_entry_safe(page, tmp, &page_list, lru) {
812 list_del(&page->lru);
813 put_page(page);
814 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815}
David Gibson63551ae2005-06-21 17:14:44 -0700816
Chen, Kenneth W502717f2006-10-11 01:20:46 -0700817void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
818 unsigned long end)
819{
820 /*
821 * It is undesirable to test vma->vm_file as it should be non-null
822 * for valid hugetlb area. However, vm_file will be NULL in the error
823 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
824 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
825 * to clean up. Since no pte has actually been setup, it is safe to
826 * do nothing in this case.
827 */
828 if (vma->vm_file) {
829 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
830 __unmap_hugepage_range(vma, start, end);
831 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
832 }
833}
834
David Gibson1e8f8892006-01-06 00:10:44 -0800835static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
836 unsigned long address, pte_t *ptep, pte_t pte)
837{
838 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -0800839 int avoidcopy;
David Gibson1e8f8892006-01-06 00:10:44 -0800840
841 old_page = pte_page(pte);
842
843 /* If no-one else is actually using this page, avoid the copy
844 * and just make the page writable */
845 avoidcopy = (page_count(old_page) == 1);
846 if (avoidcopy) {
847 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -0700848 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800849 }
850
851 page_cache_get(old_page);
Christoph Lameter5da7ca82006-01-06 00:10:46 -0800852 new_page = alloc_huge_page(vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -0800853
Adam Litke2fc39ce2007-11-14 16:59:39 -0800854 if (IS_ERR(new_page)) {
David Gibson1e8f8892006-01-06 00:10:44 -0800855 page_cache_release(old_page);
Adam Litke2fc39ce2007-11-14 16:59:39 -0800856 return -PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -0800857 }
858
859 spin_unlock(&mm->page_table_lock);
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000860 copy_huge_page(new_page, old_page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -0800861 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -0800862 spin_lock(&mm->page_table_lock);
863
864 ptep = huge_pte_offset(mm, address & HPAGE_MASK);
865 if (likely(pte_same(*ptep, pte))) {
866 /* Break COW */
867 set_huge_pte_at(mm, address, ptep,
868 make_huge_pte(vma, new_page, 1));
869 /* Make the old page be freed below */
870 new_page = old_page;
871 }
872 page_cache_release(new_page);
873 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -0700874 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800875}
876
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -0700877static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
David Gibson1e8f8892006-01-06 00:10:44 -0800878 unsigned long address, pte_t *ptep, int write_access)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100879{
880 int ret = VM_FAULT_SIGBUS;
Adam Litke4c887262005-10-29 18:16:46 -0700881 unsigned long idx;
882 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -0700883 struct page *page;
884 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -0800885 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -0700886
Adam Litke4c887262005-10-29 18:16:46 -0700887 mapping = vma->vm_file->f_mapping;
888 idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
889 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
890
891 /*
892 * Use page lock to guard against racing truncation
893 * before we get page_table_lock.
894 */
Christoph Lameter6bda6662006-01-06 00:10:49 -0800895retry:
896 page = find_lock_page(mapping, idx);
897 if (!page) {
Hugh Dickinsebed4bf2006-10-28 10:38:43 -0700898 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
899 if (idx >= size)
900 goto out;
Christoph Lameter6bda6662006-01-06 00:10:49 -0800901 page = alloc_huge_page(vma, address);
Adam Litke2fc39ce2007-11-14 16:59:39 -0800902 if (IS_ERR(page)) {
903 ret = -PTR_ERR(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -0800904 goto out;
905 }
David Gibson79ac6ba2006-03-22 00:08:51 -0800906 clear_huge_page(page, address);
Nick Piggin0ed361d2008-02-04 22:29:34 -0800907 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100908
Christoph Lameter6bda6662006-01-06 00:10:49 -0800909 if (vma->vm_flags & VM_SHARED) {
910 int err;
Ken Chen45c682a2007-11-14 16:59:44 -0800911 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -0800912
913 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
914 if (err) {
915 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -0800916 if (err == -EEXIST)
917 goto retry;
918 goto out;
919 }
Ken Chen45c682a2007-11-14 16:59:44 -0800920
921 spin_lock(&inode->i_lock);
922 inode->i_blocks += BLOCKS_PER_HUGEPAGE;
923 spin_unlock(&inode->i_lock);
Christoph Lameter6bda6662006-01-06 00:10:49 -0800924 } else
925 lock_page(page);
926 }
David Gibson1e8f8892006-01-06 00:10:44 -0800927
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100928 spin_lock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700929 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
930 if (idx >= size)
931 goto backout;
932
Nick Piggin83c54072007-07-19 01:47:05 -0700933 ret = 0;
Adam Litke86e52162006-01-06 00:10:43 -0800934 if (!pte_none(*ptep))
Adam Litke4c887262005-10-29 18:16:46 -0700935 goto backout;
936
David Gibson1e8f8892006-01-06 00:10:44 -0800937 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
938 && (vma->vm_flags & VM_SHARED)));
939 set_huge_pte_at(mm, address, ptep, new_pte);
940
941 if (write_access && !(vma->vm_flags & VM_SHARED)) {
942 /* Optimization, do the COW without a second fault */
943 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
944 }
945
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100946 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700947 unlock_page(page);
948out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100949 return ret;
Adam Litke4c887262005-10-29 18:16:46 -0700950
951backout:
952 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700953 unlock_page(page);
954 put_page(page);
955 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100956}
957
Adam Litke86e52162006-01-06 00:10:43 -0800958int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
959 unsigned long address, int write_access)
960{
961 pte_t *ptep;
962 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -0800963 int ret;
David Gibson3935baa2006-03-22 00:08:53 -0800964 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800965
966 ptep = huge_pte_alloc(mm, address);
967 if (!ptep)
968 return VM_FAULT_OOM;
969
David Gibson3935baa2006-03-22 00:08:53 -0800970 /*
971 * Serialize hugepage allocation and instantiation, so that we don't
972 * get spurious allocation failures if two CPUs race to instantiate
973 * the same page in the page cache.
974 */
975 mutex_lock(&hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800976 entry = *ptep;
David Gibson3935baa2006-03-22 00:08:53 -0800977 if (pte_none(entry)) {
978 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
979 mutex_unlock(&hugetlb_instantiation_mutex);
980 return ret;
981 }
Adam Litke86e52162006-01-06 00:10:43 -0800982
Nick Piggin83c54072007-07-19 01:47:05 -0700983 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800984
985 spin_lock(&mm->page_table_lock);
986 /* Check for a racing update before calling hugetlb_cow */
987 if (likely(pte_same(entry, *ptep)))
988 if (write_access && !pte_write(entry))
989 ret = hugetlb_cow(mm, vma, address, ptep, entry);
990 spin_unlock(&mm->page_table_lock);
David Gibson3935baa2006-03-22 00:08:53 -0800991 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -0800992
993 return ret;
Adam Litke86e52162006-01-06 00:10:43 -0800994}
995
David Gibson63551ae2005-06-21 17:14:44 -0700996int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
997 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -0800998 unsigned long *position, int *length, int i,
999 int write)
David Gibson63551ae2005-06-21 17:14:44 -07001000{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08001001 unsigned long pfn_offset;
1002 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -07001003 int remainder = *length;
1004
Hugh Dickins1c598272005-10-19 21:23:43 -07001005 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07001006 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07001007 pte_t *pte;
1008 struct page *page;
1009
1010 /*
1011 * Some archs (sparc64, sh*) have multiple pte_ts to
1012 * each hugepage. We have to make * sure we get the
1013 * first, for the page indexing below to work.
1014 */
1015 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
1016
Adam Litke72fad712007-12-10 15:49:28 -08001017 if (!pte || pte_none(*pte) || (write && !pte_write(*pte))) {
Adam Litke4c887262005-10-29 18:16:46 -07001018 int ret;
1019
1020 spin_unlock(&mm->page_table_lock);
Adam Litke5b23dbe2007-11-14 16:59:33 -08001021 ret = hugetlb_fault(mm, vma, vaddr, write);
Adam Litke4c887262005-10-29 18:16:46 -07001022 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07001023 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07001024 continue;
1025
1026 remainder = 0;
1027 if (!i)
1028 i = -EFAULT;
1029 break;
1030 }
David Gibson63551ae2005-06-21 17:14:44 -07001031
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08001032 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1033 page = pte_page(*pte);
1034same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08001035 if (pages) {
1036 get_page(page);
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08001037 pages[i] = page + pfn_offset;
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08001038 }
David Gibson63551ae2005-06-21 17:14:44 -07001039
1040 if (vmas)
1041 vmas[i] = vma;
1042
1043 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08001044 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07001045 --remainder;
1046 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08001047 if (vaddr < vma->vm_end && remainder &&
1048 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1049 /*
1050 * We use pfn_offset to avoid touching the pageframes
1051 * of this compound page.
1052 */
1053 goto same_page;
1054 }
David Gibson63551ae2005-06-21 17:14:44 -07001055 }
Hugh Dickins1c598272005-10-19 21:23:43 -07001056 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07001057 *length = remainder;
1058 *position = vaddr;
1059
1060 return i;
1061}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001062
1063void hugetlb_change_protection(struct vm_area_struct *vma,
1064 unsigned long address, unsigned long end, pgprot_t newprot)
1065{
1066 struct mm_struct *mm = vma->vm_mm;
1067 unsigned long start = address;
1068 pte_t *ptep;
1069 pte_t pte;
1070
1071 BUG_ON(address >= end);
1072 flush_cache_range(vma, address, end);
1073
Chen, Kenneth W39dde652006-12-06 20:32:03 -08001074 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001075 spin_lock(&mm->page_table_lock);
1076 for (; address < end; address += HPAGE_SIZE) {
1077 ptep = huge_pte_offset(mm, address);
1078 if (!ptep)
1079 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08001080 if (huge_pmd_unshare(mm, &address, ptep))
1081 continue;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001082 if (!pte_none(*ptep)) {
1083 pte = huge_ptep_get_and_clear(mm, address, ptep);
1084 pte = pte_mkhuge(pte_modify(pte, newprot));
1085 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001086 }
1087 }
1088 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08001089 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001090
1091 flush_tlb_range(vma, start, end);
1092}
1093
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001094struct file_region {
1095 struct list_head link;
1096 long from;
1097 long to;
1098};
1099
1100static long region_add(struct list_head *head, long f, long t)
1101{
1102 struct file_region *rg, *nrg, *trg;
1103
1104 /* Locate the region we are either in or before. */
1105 list_for_each_entry(rg, head, link)
1106 if (f <= rg->to)
1107 break;
1108
1109 /* Round our left edge to the current segment if it encloses us. */
1110 if (f > rg->from)
1111 f = rg->from;
1112
1113 /* Check for and consume any regions we now overlap with. */
1114 nrg = rg;
1115 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1116 if (&rg->link == head)
1117 break;
1118 if (rg->from > t)
1119 break;
1120
1121 /* If this area reaches higher then extend our area to
1122 * include it completely. If this is not the first area
1123 * which we intend to reuse, free it. */
1124 if (rg->to > t)
1125 t = rg->to;
1126 if (rg != nrg) {
1127 list_del(&rg->link);
1128 kfree(rg);
1129 }
1130 }
1131 nrg->from = f;
1132 nrg->to = t;
1133 return 0;
1134}
1135
1136static long region_chg(struct list_head *head, long f, long t)
1137{
1138 struct file_region *rg, *nrg;
1139 long chg = 0;
1140
1141 /* Locate the region we are before or in. */
1142 list_for_each_entry(rg, head, link)
1143 if (f <= rg->to)
1144 break;
1145
1146 /* If we are below the current region then a new region is required.
1147 * Subtle, allocate a new region at the position but make it zero
Simon Arlott183ff222007-10-20 01:27:18 +02001148 * size such that we can guarantee to record the reservation. */
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001149 if (&rg->link == head || t < rg->from) {
1150 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001151 if (!nrg)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001152 return -ENOMEM;
1153 nrg->from = f;
1154 nrg->to = f;
1155 INIT_LIST_HEAD(&nrg->link);
1156 list_add(&nrg->link, rg->link.prev);
1157
1158 return t - f;
1159 }
1160
1161 /* Round our left edge to the current segment if it encloses us. */
1162 if (f > rg->from)
1163 f = rg->from;
1164 chg = t - f;
1165
1166 /* Check for and consume any regions we now overlap with. */
1167 list_for_each_entry(rg, rg->link.prev, link) {
1168 if (&rg->link == head)
1169 break;
1170 if (rg->from > t)
1171 return chg;
1172
1173 /* We overlap with this area, if it extends futher than
1174 * us then we must extend ourselves. Account for its
1175 * existing reservation. */
1176 if (rg->to > t) {
1177 chg += rg->to - t;
1178 t = rg->to;
1179 }
1180 chg -= rg->to - rg->from;
1181 }
1182 return chg;
1183}
1184
1185static long region_truncate(struct list_head *head, long end)
1186{
1187 struct file_region *rg, *trg;
1188 long chg = 0;
1189
1190 /* Locate the region we are either in or before. */
1191 list_for_each_entry(rg, head, link)
1192 if (end <= rg->to)
1193 break;
1194 if (&rg->link == head)
1195 return 0;
1196
1197 /* If we are in the middle of a region then adjust it. */
1198 if (end > rg->from) {
1199 chg = rg->to - end;
1200 rg->to = end;
1201 rg = list_entry(rg->link.next, typeof(*rg), link);
1202 }
1203
1204 /* Drop any remaining regions. */
1205 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1206 if (&rg->link == head)
1207 break;
1208 chg += rg->to - rg->from;
1209 list_del(&rg->link);
1210 kfree(rg);
1211 }
1212 return chg;
1213}
1214
1215static int hugetlb_acct_memory(long delta)
1216{
1217 int ret = -ENOMEM;
1218
1219 spin_lock(&hugetlb_lock);
Ken Chen8a630112007-05-09 02:33:34 -07001220 /*
1221 * When cpuset is configured, it breaks the strict hugetlb page
1222 * reservation as the accounting is done on a global variable. Such
1223 * reservation is completely rubbish in the presence of cpuset because
1224 * the reservation is not checked against page availability for the
1225 * current cpuset. Application can still potentially OOM'ed by kernel
1226 * with lack of free htlb page in cpuset that the task is in.
1227 * Attempt to enforce strict accounting with cpuset is almost
1228 * impossible (or too ugly) because cpuset is too fluid that
1229 * task or memory node can be dynamically moved between cpusets.
1230 *
1231 * The change of semantics for shared hugetlb mapping with cpuset is
1232 * undesirable. However, in order to preserve some of the semantics,
1233 * we fall back to check against current free page availability as
1234 * a best attempt and hopefully to minimize the impact of changing
1235 * semantics that cpuset has.
1236 */
Adam Litkee4e574b2007-10-16 01:26:19 -07001237 if (delta > 0) {
1238 if (gather_surplus_pages(delta) < 0)
1239 goto out;
1240
Adam Litkeac09b3a2008-03-04 14:29:38 -08001241 if (delta > cpuset_mems_nr(free_huge_pages_node)) {
1242 return_unused_surplus_pages(delta);
Adam Litkee4e574b2007-10-16 01:26:19 -07001243 goto out;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001244 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001245 }
1246
1247 ret = 0;
Adam Litkee4e574b2007-10-16 01:26:19 -07001248 if (delta < 0)
1249 return_unused_surplus_pages((unsigned long) -delta);
1250
1251out:
1252 spin_unlock(&hugetlb_lock);
1253 return ret;
1254}
1255
1256int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1257{
1258 long ret, chg;
1259
1260 chg = region_chg(&inode->i_mapping->private_list, from, to);
1261 if (chg < 0)
1262 return chg;
Ken Chen8a630112007-05-09 02:33:34 -07001263
Adam Litke90d8b7e2007-11-14 16:59:42 -08001264 if (hugetlb_get_quota(inode->i_mapping, chg))
1265 return -ENOSPC;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001266 ret = hugetlb_acct_memory(chg);
Ken Chen68842c92008-01-14 00:55:19 -08001267 if (ret < 0) {
1268 hugetlb_put_quota(inode->i_mapping, chg);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001269 return ret;
Ken Chen68842c92008-01-14 00:55:19 -08001270 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001271 region_add(&inode->i_mapping->private_list, from, to);
1272 return 0;
1273}
1274
1275void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1276{
1277 long chg = region_truncate(&inode->i_mapping->private_list, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08001278
1279 spin_lock(&inode->i_lock);
1280 inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1281 spin_unlock(&inode->i_lock);
1282
Adam Litke90d8b7e2007-11-14 16:59:42 -08001283 hugetlb_put_quota(inode->i_mapping, (chg - freed));
1284 hugetlb_acct_memory(-(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001285}