blob: 8b809ecefa39e4b54f4bfbed8830009a9e260e96 [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;
Adam Litke54f9f802007-10-16 01:26:20 -070034int hugetlb_dynamic_pool;
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);
David Gibson27a85ef2006-03-22 00:08:56 -0800119
Adam Litke7893d1d2007-10-16 01:26:18 -0700120 BUG_ON(page_count(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800121 INIT_LIST_HEAD(&page->lru);
122
123 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700124 if (surplus_huge_pages_node[nid]) {
125 update_and_free_page(page);
126 surplus_huge_pages--;
127 surplus_huge_pages_node[nid]--;
128 } else {
129 enqueue_huge_page(page);
130 }
David Gibson27a85ef2006-03-22 00:08:56 -0800131 spin_unlock(&hugetlb_lock);
132}
133
Adam Litke7893d1d2007-10-16 01:26:18 -0700134/*
135 * Increment or decrement surplus_huge_pages. Keep node-specific counters
136 * balanced by operating on them in a round-robin fashion.
137 * Returns 1 if an adjustment was made.
138 */
139static int adjust_pool_surplus(int delta)
140{
141 static int prev_nid;
142 int nid = prev_nid;
143 int ret = 0;
144
145 VM_BUG_ON(delta != -1 && delta != 1);
146 do {
147 nid = next_node(nid, node_online_map);
148 if (nid == MAX_NUMNODES)
149 nid = first_node(node_online_map);
150
151 /* To shrink on this node, there must be a surplus page */
152 if (delta < 0 && !surplus_huge_pages_node[nid])
153 continue;
154 /* Surplus cannot exceed the total number of pages */
155 if (delta > 0 && surplus_huge_pages_node[nid] >=
156 nr_huge_pages_node[nid])
157 continue;
158
159 surplus_huge_pages += delta;
160 surplus_huge_pages_node[nid] += delta;
161 ret = 1;
162 break;
163 } while (nid != prev_nid);
164
165 prev_nid = nid;
166 return ret;
167}
168
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700169static struct page *alloc_fresh_huge_page_node(int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700172
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700173 page = alloc_pages_node(nid,
174 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
175 HUGETLB_PAGE_ORDER);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 if (page) {
Andy Whitcroft33f2ef82006-12-06 20:33:32 -0800177 set_compound_page_dtor(page, free_huge_page);
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800178 spin_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179 nr_huge_pages++;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700180 nr_huge_pages_node[nid]++;
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800181 spin_unlock(&hugetlb_lock);
Nick Piggina4822892006-03-22 00:08:08 -0800182 put_page(page); /* free it into the hugepage allocator */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700184
185 return page;
186}
187
188static int alloc_fresh_huge_page(void)
189{
190 struct page *page;
191 int start_nid;
192 int next_nid;
193 int ret = 0;
194
195 start_nid = hugetlb_next_nid;
196
197 do {
198 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
199 if (page)
200 ret = 1;
201 /*
202 * Use a helper variable to find the next node and then
203 * copy it back to hugetlb_next_nid afterwards:
204 * otherwise there's a window in which a racer might
205 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
206 * But we don't need to use a spin_lock here: it really
207 * doesn't matter if occasionally a racer chooses the
208 * same nid as we do. Move nid forward in the mask even
209 * if we just successfully allocated a hugepage so that
210 * the next caller gets hugepages on the next node.
211 */
212 next_nid = next_node(hugetlb_next_nid, node_online_map);
213 if (next_nid == MAX_NUMNODES)
214 next_nid = first_node(node_online_map);
215 hugetlb_next_nid = next_nid;
216 } while (!page && hugetlb_next_nid != start_nid);
217
218 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219}
220
Adam Litke7893d1d2007-10-16 01:26:18 -0700221static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
222 unsigned long address)
223{
224 struct page *page;
225
Adam Litke54f9f802007-10-16 01:26:20 -0700226 /* Check if the dynamic pool is enabled */
227 if (!hugetlb_dynamic_pool)
228 return NULL;
229
Adam Litke7893d1d2007-10-16 01:26:18 -0700230 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
231 HUGETLB_PAGE_ORDER);
232 if (page) {
233 set_compound_page_dtor(page, free_huge_page);
234 spin_lock(&hugetlb_lock);
235 nr_huge_pages++;
236 nr_huge_pages_node[page_to_nid(page)]++;
237 surplus_huge_pages++;
238 surplus_huge_pages_node[page_to_nid(page)]++;
239 spin_unlock(&hugetlb_lock);
240 }
241
242 return page;
243}
244
Adam Litkee4e574b2007-10-16 01:26:19 -0700245/*
246 * Increase the hugetlb pool such that it can accomodate a reservation
247 * of size 'delta'.
248 */
249static int gather_surplus_pages(int delta)
250{
251 struct list_head surplus_list;
252 struct page *page, *tmp;
253 int ret, i;
254 int needed, allocated;
255
256 needed = (resv_huge_pages + delta) - free_huge_pages;
257 if (needed <= 0)
258 return 0;
259
260 allocated = 0;
261 INIT_LIST_HEAD(&surplus_list);
262
263 ret = -ENOMEM;
264retry:
265 spin_unlock(&hugetlb_lock);
266 for (i = 0; i < needed; i++) {
267 page = alloc_buddy_huge_page(NULL, 0);
268 if (!page) {
269 /*
270 * We were not able to allocate enough pages to
271 * satisfy the entire reservation so we free what
272 * we've allocated so far.
273 */
274 spin_lock(&hugetlb_lock);
275 needed = 0;
276 goto free;
277 }
278
279 list_add(&page->lru, &surplus_list);
280 }
281 allocated += needed;
282
283 /*
284 * After retaking hugetlb_lock, we need to recalculate 'needed'
285 * because either resv_huge_pages or free_huge_pages may have changed.
286 */
287 spin_lock(&hugetlb_lock);
288 needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
289 if (needed > 0)
290 goto retry;
291
292 /*
293 * The surplus_list now contains _at_least_ the number of extra pages
294 * needed to accomodate the reservation. Add the appropriate number
295 * of pages to the hugetlb pool and free the extras back to the buddy
296 * allocator.
297 */
298 needed += allocated;
299 ret = 0;
300free:
301 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
302 list_del(&page->lru);
303 if ((--needed) >= 0)
304 enqueue_huge_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700305 else {
306 /*
307 * Decrement the refcount and free the page using its
308 * destructor. This must be done with hugetlb_lock
309 * unlocked which is safe because free_huge_page takes
310 * hugetlb_lock before deciding how to free the page.
311 */
312 spin_unlock(&hugetlb_lock);
313 put_page(page);
314 spin_lock(&hugetlb_lock);
315 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700316 }
317
318 return ret;
319}
320
321/*
322 * When releasing a hugetlb pool reservation, any surplus pages that were
323 * allocated to satisfy the reservation must be explicitly freed if they were
324 * never used.
325 */
326void return_unused_surplus_pages(unsigned long unused_resv_pages)
327{
328 static int nid = -1;
329 struct page *page;
330 unsigned long nr_pages;
331
332 nr_pages = min(unused_resv_pages, surplus_huge_pages);
333
334 while (nr_pages) {
335 nid = next_node(nid, node_online_map);
336 if (nid == MAX_NUMNODES)
337 nid = first_node(node_online_map);
338
339 if (!surplus_huge_pages_node[nid])
340 continue;
341
342 if (!list_empty(&hugepage_freelists[nid])) {
343 page = list_entry(hugepage_freelists[nid].next,
344 struct page, lru);
345 list_del(&page->lru);
346 update_and_free_page(page);
347 free_huge_pages--;
348 free_huge_pages_node[nid]--;
349 surplus_huge_pages--;
350 surplus_huge_pages_node[nid]--;
351 nr_pages--;
352 }
353 }
354}
355
David Gibson27a85ef2006-03-22 00:08:56 -0800356static struct page *alloc_huge_page(struct vm_area_struct *vma,
357 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358{
Adam Litke7893d1d2007-10-16 01:26:18 -0700359 struct page *page = NULL;
Adam Litkee4e574b2007-10-16 01:26:19 -0700360 int use_reserved_page = vma->vm_flags & VM_MAYSHARE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361
362 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -0700363 if (!use_reserved_page && (free_huge_pages <= resv_huge_pages))
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700364 goto fail;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800365
366 page = dequeue_huge_page(vma, addr);
367 if (!page)
368 goto fail;
369
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370 spin_unlock(&hugetlb_lock);
Nick Piggin7835e982006-03-22 00:08:40 -0800371 set_page_refcounted(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800373
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700374fail:
David Gibsonb45b5bd2006-03-22 00:08:55 -0800375 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700376
377 /*
378 * Private mappings do not use reserved huge pages so the allocation
379 * may have failed due to an undersized hugetlb pool. Try to grab a
380 * surplus huge page from the buddy allocator.
381 */
Adam Litkee4e574b2007-10-16 01:26:19 -0700382 if (!use_reserved_page)
Adam Litke7893d1d2007-10-16 01:26:18 -0700383 page = alloc_buddy_huge_page(vma, addr);
384
385 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800386}
387
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388static int __init hugetlb_init(void)
389{
390 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391
Benjamin Herrenschmidt3c726f82005-11-07 11:06:55 +1100392 if (HPAGE_SHIFT == 0)
393 return 0;
394
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 for (i = 0; i < MAX_NUMNODES; ++i)
396 INIT_LIST_HEAD(&hugepage_freelists[i]);
397
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700398 hugetlb_next_nid = first_node(node_online_map);
399
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 for (i = 0; i < max_huge_pages; ++i) {
Nick Piggina4822892006-03-22 00:08:08 -0800401 if (!alloc_fresh_huge_page())
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 }
404 max_huge_pages = free_huge_pages = nr_huge_pages = i;
405 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
406 return 0;
407}
408module_init(hugetlb_init);
409
410static int __init hugetlb_setup(char *s)
411{
412 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
413 max_huge_pages = 0;
414 return 1;
415}
416__setup("hugepages=", hugetlb_setup);
417
Ken Chen8a630112007-05-09 02:33:34 -0700418static unsigned int cpuset_mems_nr(unsigned int *array)
419{
420 int node;
421 unsigned int nr = 0;
422
423 for_each_node_mask(node, cpuset_current_mems_allowed)
424 nr += array[node];
425
426 return nr;
427}
428
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429#ifdef CONFIG_SYSCTL
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430#ifdef CONFIG_HIGHMEM
431static void try_to_free_low(unsigned long count)
432{
Christoph Lameter4415cc82006-09-25 23:31:55 -0700433 int i;
434
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435 for (i = 0; i < MAX_NUMNODES; ++i) {
436 struct page *page, *next;
437 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
Adam Litke6b0c8802007-10-16 01:26:23 -0700438 if (count >= nr_huge_pages)
439 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440 if (PageHighMem(page))
441 continue;
442 list_del(&page->lru);
443 update_and_free_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 free_huge_pages--;
Christoph Lameter4415cc82006-09-25 23:31:55 -0700445 free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 }
447 }
448}
449#else
450static inline void try_to_free_low(unsigned long count)
451{
452}
453#endif
454
Adam Litke7893d1d2007-10-16 01:26:18 -0700455#define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456static unsigned long set_max_huge_pages(unsigned long count)
457{
Adam Litke7893d1d2007-10-16 01:26:18 -0700458 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459
Adam Litke7893d1d2007-10-16 01:26:18 -0700460 /*
461 * Increase the pool size
462 * First take pages out of surplus state. Then make up the
463 * remaining difference by allocating fresh huge pages.
464 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700466 while (surplus_huge_pages && count > persistent_huge_pages) {
467 if (!adjust_pool_surplus(-1))
468 break;
469 }
470
471 while (count > persistent_huge_pages) {
472 int ret;
473 /*
474 * If this allocation races such that we no longer need the
475 * page, free_huge_page will handle it by freeing the page
476 * and reducing the surplus.
477 */
478 spin_unlock(&hugetlb_lock);
479 ret = alloc_fresh_huge_page();
480 spin_lock(&hugetlb_lock);
481 if (!ret)
482 goto out;
483
484 }
Adam Litke7893d1d2007-10-16 01:26:18 -0700485
486 /*
487 * Decrease the pool size
488 * First return free pages to the buddy allocator (being careful
489 * to keep enough around to satisfy reservations). Then place
490 * pages into surplus state as needed so the pool will shrink
491 * to the desired size as pages become free.
492 */
Adam Litke6b0c8802007-10-16 01:26:23 -0700493 min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
494 min_count = max(count, min_count);
Adam Litke7893d1d2007-10-16 01:26:18 -0700495 try_to_free_low(min_count);
496 while (min_count < persistent_huge_pages) {
Christoph Lameter5da7ca82006-01-06 00:10:46 -0800497 struct page *page = dequeue_huge_page(NULL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 if (!page)
499 break;
500 update_and_free_page(page);
501 }
Adam Litke7893d1d2007-10-16 01:26:18 -0700502 while (count < persistent_huge_pages) {
503 if (!adjust_pool_surplus(1))
504 break;
505 }
506out:
507 ret = persistent_huge_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700509 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510}
511
512int hugetlb_sysctl_handler(struct ctl_table *table, int write,
513 struct file *file, void __user *buffer,
514 size_t *length, loff_t *ppos)
515{
516 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
517 max_huge_pages = set_max_huge_pages(max_huge_pages);
518 return 0;
519}
Mel Gorman396faf02007-07-17 04:03:13 -0700520
521int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
522 struct file *file, void __user *buffer,
523 size_t *length, loff_t *ppos)
524{
525 proc_dointvec(table, write, file, buffer, length, ppos);
526 if (hugepages_treat_as_movable)
527 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
528 else
529 htlb_alloc_mask = GFP_HIGHUSER;
530 return 0;
531}
532
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533#endif /* CONFIG_SYSCTL */
534
535int hugetlb_report_meminfo(char *buf)
536{
537 return sprintf(buf,
538 "HugePages_Total: %5lu\n"
539 "HugePages_Free: %5lu\n"
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700540 "HugePages_Rsvd: %5lu\n"
Adam Litke7893d1d2007-10-16 01:26:18 -0700541 "HugePages_Surp: %5lu\n"
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542 "Hugepagesize: %5lu kB\n",
543 nr_huge_pages,
544 free_huge_pages,
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700545 resv_huge_pages,
Adam Litke7893d1d2007-10-16 01:26:18 -0700546 surplus_huge_pages,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547 HPAGE_SIZE/1024);
548}
549
550int hugetlb_report_node_meminfo(int nid, char *buf)
551{
552 return sprintf(buf,
553 "Node %d HugePages_Total: %5u\n"
554 "Node %d HugePages_Free: %5u\n",
555 nid, nr_huge_pages_node[nid],
556 nid, free_huge_pages_node[nid]);
557}
558
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
560unsigned long hugetlb_total_pages(void)
561{
562 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
563}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564
565/*
566 * We cannot handle pagefaults against hugetlb pages at all. They cause
567 * handle_mm_fault() to try to instantiate regular-sized pages in the
568 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
569 * this far.
570 */
Nick Piggind0217ac2007-07-19 01:47:03 -0700571static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572{
573 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -0700574 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575}
576
577struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -0700578 .fault = hugetlb_vm_op_fault,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579};
580
David Gibson1e8f8892006-01-06 00:10:44 -0800581static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
582 int writable)
David Gibson63551ae2005-06-21 17:14:44 -0700583{
584 pte_t entry;
585
David Gibson1e8f8892006-01-06 00:10:44 -0800586 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -0700587 entry =
588 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
589 } else {
590 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
591 }
592 entry = pte_mkyoung(entry);
593 entry = pte_mkhuge(entry);
594
595 return entry;
596}
597
David Gibson1e8f8892006-01-06 00:10:44 -0800598static void set_huge_ptep_writable(struct vm_area_struct *vma,
599 unsigned long address, pte_t *ptep)
600{
601 pte_t entry;
602
603 entry = pte_mkwrite(pte_mkdirty(*ptep));
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -0700604 if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
605 update_mmu_cache(vma, address, entry);
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -0700606 }
David Gibson1e8f8892006-01-06 00:10:44 -0800607}
608
609
David Gibson63551ae2005-06-21 17:14:44 -0700610int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
611 struct vm_area_struct *vma)
612{
613 pte_t *src_pte, *dst_pte, entry;
614 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -0700615 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -0800616 int cow;
617
618 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -0700619
Hugh Dickins1c598272005-10-19 21:23:43 -0700620 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
Hugh Dickinsc74df322005-10-29 18:16:23 -0700621 src_pte = huge_pte_offset(src, addr);
622 if (!src_pte)
623 continue;
David Gibson63551ae2005-06-21 17:14:44 -0700624 dst_pte = huge_pte_alloc(dst, addr);
625 if (!dst_pte)
626 goto nomem;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700627 spin_lock(&dst->page_table_lock);
Hugh Dickins1c598272005-10-19 21:23:43 -0700628 spin_lock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700629 if (!pte_none(*src_pte)) {
David Gibson1e8f8892006-01-06 00:10:44 -0800630 if (cow)
631 ptep_set_wrprotect(src, addr, src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -0700632 entry = *src_pte;
633 ptepage = pte_page(entry);
634 get_page(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -0700635 set_huge_pte_at(dst, addr, dst_pte, entry);
636 }
637 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700638 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700639 }
640 return 0;
641
642nomem:
643 return -ENOMEM;
644}
645
Chen, Kenneth W502717f2006-10-11 01:20:46 -0700646void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
647 unsigned long end)
David Gibson63551ae2005-06-21 17:14:44 -0700648{
649 struct mm_struct *mm = vma->vm_mm;
650 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -0700651 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -0700652 pte_t pte;
653 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700654 struct page *tmp;
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -0800655 /*
656 * A page gathering list, protected by per file i_mmap_lock. The
657 * lock is used to avoid list corruption from multiple unmapping
658 * of the same page since we are using page->lru.
659 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700660 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -0700661
662 WARN_ON(!is_vm_hugetlb_page(vma));
663 BUG_ON(start & ~HPAGE_MASK);
664 BUG_ON(end & ~HPAGE_MASK);
665
Hugh Dickins508034a2005-10-29 18:16:30 -0700666 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700667 for (address = start; address < end; address += HPAGE_SIZE) {
David Gibsonc7546f82005-08-05 11:59:35 -0700668 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -0700669 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -0700670 continue;
671
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800672 if (huge_pmd_unshare(mm, &address, ptep))
673 continue;
674
David Gibsonc7546f82005-08-05 11:59:35 -0700675 pte = huge_ptep_get_and_clear(mm, address, ptep);
David Gibson63551ae2005-06-21 17:14:44 -0700676 if (pte_none(pte))
677 continue;
David Gibsonc7546f82005-08-05 11:59:35 -0700678
David Gibson63551ae2005-06-21 17:14:44 -0700679 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -0800680 if (pte_dirty(pte))
681 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700682 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -0700683 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -0700685 flush_tlb_range(vma, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700686 list_for_each_entry_safe(page, tmp, &page_list, lru) {
687 list_del(&page->lru);
688 put_page(page);
689 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690}
David Gibson63551ae2005-06-21 17:14:44 -0700691
Chen, Kenneth W502717f2006-10-11 01:20:46 -0700692void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
693 unsigned long end)
694{
695 /*
696 * It is undesirable to test vma->vm_file as it should be non-null
697 * for valid hugetlb area. However, vm_file will be NULL in the error
698 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
699 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
700 * to clean up. Since no pte has actually been setup, it is safe to
701 * do nothing in this case.
702 */
703 if (vma->vm_file) {
704 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
705 __unmap_hugepage_range(vma, start, end);
706 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
707 }
708}
709
David Gibson1e8f8892006-01-06 00:10:44 -0800710static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
711 unsigned long address, pte_t *ptep, pte_t pte)
712{
713 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -0800714 int avoidcopy;
David Gibson1e8f8892006-01-06 00:10:44 -0800715
716 old_page = pte_page(pte);
717
718 /* If no-one else is actually using this page, avoid the copy
719 * and just make the page writable */
720 avoidcopy = (page_count(old_page) == 1);
721 if (avoidcopy) {
722 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -0700723 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800724 }
725
726 page_cache_get(old_page);
Christoph Lameter5da7ca82006-01-06 00:10:46 -0800727 new_page = alloc_huge_page(vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -0800728
729 if (!new_page) {
730 page_cache_release(old_page);
Christoph Lameter0df420d2006-02-07 12:58:30 -0800731 return VM_FAULT_OOM;
David Gibson1e8f8892006-01-06 00:10:44 -0800732 }
733
734 spin_unlock(&mm->page_table_lock);
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000735 copy_huge_page(new_page, old_page, address, vma);
David Gibson1e8f8892006-01-06 00:10:44 -0800736 spin_lock(&mm->page_table_lock);
737
738 ptep = huge_pte_offset(mm, address & HPAGE_MASK);
739 if (likely(pte_same(*ptep, pte))) {
740 /* Break COW */
741 set_huge_pte_at(mm, address, ptep,
742 make_huge_pte(vma, new_page, 1));
743 /* Make the old page be freed below */
744 new_page = old_page;
745 }
746 page_cache_release(new_page);
747 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -0700748 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800749}
750
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -0700751static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
David Gibson1e8f8892006-01-06 00:10:44 -0800752 unsigned long address, pte_t *ptep, int write_access)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100753{
754 int ret = VM_FAULT_SIGBUS;
Adam Litke4c887262005-10-29 18:16:46 -0700755 unsigned long idx;
756 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -0700757 struct page *page;
758 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -0800759 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -0700760
Adam Litke4c887262005-10-29 18:16:46 -0700761 mapping = vma->vm_file->f_mapping;
762 idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
763 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
764
765 /*
766 * Use page lock to guard against racing truncation
767 * before we get page_table_lock.
768 */
Christoph Lameter6bda6662006-01-06 00:10:49 -0800769retry:
770 page = find_lock_page(mapping, idx);
771 if (!page) {
Hugh Dickinsebed4bf2006-10-28 10:38:43 -0700772 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
773 if (idx >= size)
774 goto out;
Christoph Lameter6bda6662006-01-06 00:10:49 -0800775 if (hugetlb_get_quota(mapping))
776 goto out;
777 page = alloc_huge_page(vma, address);
778 if (!page) {
779 hugetlb_put_quota(mapping);
Christoph Lameter0df420d2006-02-07 12:58:30 -0800780 ret = VM_FAULT_OOM;
Christoph Lameter6bda6662006-01-06 00:10:49 -0800781 goto out;
782 }
David Gibson79ac6ba2006-03-22 00:08:51 -0800783 clear_huge_page(page, address);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100784
Christoph Lameter6bda6662006-01-06 00:10:49 -0800785 if (vma->vm_flags & VM_SHARED) {
786 int err;
787
788 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
789 if (err) {
790 put_page(page);
791 hugetlb_put_quota(mapping);
792 if (err == -EEXIST)
793 goto retry;
794 goto out;
795 }
796 } else
797 lock_page(page);
798 }
David Gibson1e8f8892006-01-06 00:10:44 -0800799
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100800 spin_lock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700801 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
802 if (idx >= size)
803 goto backout;
804
Nick Piggin83c54072007-07-19 01:47:05 -0700805 ret = 0;
Adam Litke86e52162006-01-06 00:10:43 -0800806 if (!pte_none(*ptep))
Adam Litke4c887262005-10-29 18:16:46 -0700807 goto backout;
808
David Gibson1e8f8892006-01-06 00:10:44 -0800809 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
810 && (vma->vm_flags & VM_SHARED)));
811 set_huge_pte_at(mm, address, ptep, new_pte);
812
813 if (write_access && !(vma->vm_flags & VM_SHARED)) {
814 /* Optimization, do the COW without a second fault */
815 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
816 }
817
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100818 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700819 unlock_page(page);
820out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100821 return ret;
Adam Litke4c887262005-10-29 18:16:46 -0700822
823backout:
824 spin_unlock(&mm->page_table_lock);
825 hugetlb_put_quota(mapping);
826 unlock_page(page);
827 put_page(page);
828 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100829}
830
Adam Litke86e52162006-01-06 00:10:43 -0800831int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
832 unsigned long address, int write_access)
833{
834 pte_t *ptep;
835 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -0800836 int ret;
David Gibson3935baa2006-03-22 00:08:53 -0800837 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800838
839 ptep = huge_pte_alloc(mm, address);
840 if (!ptep)
841 return VM_FAULT_OOM;
842
David Gibson3935baa2006-03-22 00:08:53 -0800843 /*
844 * Serialize hugepage allocation and instantiation, so that we don't
845 * get spurious allocation failures if two CPUs race to instantiate
846 * the same page in the page cache.
847 */
848 mutex_lock(&hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800849 entry = *ptep;
David Gibson3935baa2006-03-22 00:08:53 -0800850 if (pte_none(entry)) {
851 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
852 mutex_unlock(&hugetlb_instantiation_mutex);
853 return ret;
854 }
Adam Litke86e52162006-01-06 00:10:43 -0800855
Nick Piggin83c54072007-07-19 01:47:05 -0700856 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800857
858 spin_lock(&mm->page_table_lock);
859 /* Check for a racing update before calling hugetlb_cow */
860 if (likely(pte_same(entry, *ptep)))
861 if (write_access && !pte_write(entry))
862 ret = hugetlb_cow(mm, vma, address, ptep, entry);
863 spin_unlock(&mm->page_table_lock);
David Gibson3935baa2006-03-22 00:08:53 -0800864 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -0800865
866 return ret;
Adam Litke86e52162006-01-06 00:10:43 -0800867}
868
David Gibson63551ae2005-06-21 17:14:44 -0700869int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
870 struct page **pages, struct vm_area_struct **vmas,
871 unsigned long *position, int *length, int i)
872{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800873 unsigned long pfn_offset;
874 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -0700875 int remainder = *length;
876
Hugh Dickins1c598272005-10-19 21:23:43 -0700877 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700878 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -0700879 pte_t *pte;
880 struct page *page;
881
882 /*
883 * Some archs (sparc64, sh*) have multiple pte_ts to
884 * each hugepage. We have to make * sure we get the
885 * first, for the page indexing below to work.
886 */
887 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
888
889 if (!pte || pte_none(*pte)) {
890 int ret;
891
892 spin_unlock(&mm->page_table_lock);
893 ret = hugetlb_fault(mm, vma, vaddr, 0);
894 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -0700895 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -0700896 continue;
897
898 remainder = 0;
899 if (!i)
900 i = -EFAULT;
901 break;
902 }
David Gibson63551ae2005-06-21 17:14:44 -0700903
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800904 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
905 page = pte_page(*pte);
906same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -0800907 if (pages) {
908 get_page(page);
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800909 pages[i] = page + pfn_offset;
Chen, Kenneth Wd6692182006-03-31 02:29:57 -0800910 }
David Gibson63551ae2005-06-21 17:14:44 -0700911
912 if (vmas)
913 vmas[i] = vma;
914
915 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800916 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -0700917 --remainder;
918 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800919 if (vaddr < vma->vm_end && remainder &&
920 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
921 /*
922 * We use pfn_offset to avoid touching the pageframes
923 * of this compound page.
924 */
925 goto same_page;
926 }
David Gibson63551ae2005-06-21 17:14:44 -0700927 }
Hugh Dickins1c598272005-10-19 21:23:43 -0700928 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700929 *length = remainder;
930 *position = vaddr;
931
932 return i;
933}
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800934
935void hugetlb_change_protection(struct vm_area_struct *vma,
936 unsigned long address, unsigned long end, pgprot_t newprot)
937{
938 struct mm_struct *mm = vma->vm_mm;
939 unsigned long start = address;
940 pte_t *ptep;
941 pte_t pte;
942
943 BUG_ON(address >= end);
944 flush_cache_range(vma, address, end);
945
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800946 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800947 spin_lock(&mm->page_table_lock);
948 for (; address < end; address += HPAGE_SIZE) {
949 ptep = huge_pte_offset(mm, address);
950 if (!ptep)
951 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800952 if (huge_pmd_unshare(mm, &address, ptep))
953 continue;
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800954 if (!pte_none(*ptep)) {
955 pte = huge_ptep_get_and_clear(mm, address, ptep);
956 pte = pte_mkhuge(pte_modify(pte, newprot));
957 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800958 }
959 }
960 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800961 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800962
963 flush_tlb_range(vma, start, end);
964}
965
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700966struct file_region {
967 struct list_head link;
968 long from;
969 long to;
970};
971
972static long region_add(struct list_head *head, long f, long t)
973{
974 struct file_region *rg, *nrg, *trg;
975
976 /* Locate the region we are either in or before. */
977 list_for_each_entry(rg, head, link)
978 if (f <= rg->to)
979 break;
980
981 /* Round our left edge to the current segment if it encloses us. */
982 if (f > rg->from)
983 f = rg->from;
984
985 /* Check for and consume any regions we now overlap with. */
986 nrg = rg;
987 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
988 if (&rg->link == head)
989 break;
990 if (rg->from > t)
991 break;
992
993 /* If this area reaches higher then extend our area to
994 * include it completely. If this is not the first area
995 * which we intend to reuse, free it. */
996 if (rg->to > t)
997 t = rg->to;
998 if (rg != nrg) {
999 list_del(&rg->link);
1000 kfree(rg);
1001 }
1002 }
1003 nrg->from = f;
1004 nrg->to = t;
1005 return 0;
1006}
1007
1008static long region_chg(struct list_head *head, long f, long t)
1009{
1010 struct file_region *rg, *nrg;
1011 long chg = 0;
1012
1013 /* Locate the region we are before or in. */
1014 list_for_each_entry(rg, head, link)
1015 if (f <= rg->to)
1016 break;
1017
1018 /* If we are below the current region then a new region is required.
1019 * Subtle, allocate a new region at the position but make it zero
Simon Arlott183ff222007-10-20 01:27:18 +02001020 * size such that we can guarantee to record the reservation. */
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001021 if (&rg->link == head || t < rg->from) {
1022 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001023 if (!nrg)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001024 return -ENOMEM;
1025 nrg->from = f;
1026 nrg->to = f;
1027 INIT_LIST_HEAD(&nrg->link);
1028 list_add(&nrg->link, rg->link.prev);
1029
1030 return t - f;
1031 }
1032
1033 /* Round our left edge to the current segment if it encloses us. */
1034 if (f > rg->from)
1035 f = rg->from;
1036 chg = t - f;
1037
1038 /* Check for and consume any regions we now overlap with. */
1039 list_for_each_entry(rg, rg->link.prev, link) {
1040 if (&rg->link == head)
1041 break;
1042 if (rg->from > t)
1043 return chg;
1044
1045 /* We overlap with this area, if it extends futher than
1046 * us then we must extend ourselves. Account for its
1047 * existing reservation. */
1048 if (rg->to > t) {
1049 chg += rg->to - t;
1050 t = rg->to;
1051 }
1052 chg -= rg->to - rg->from;
1053 }
1054 return chg;
1055}
1056
1057static long region_truncate(struct list_head *head, long end)
1058{
1059 struct file_region *rg, *trg;
1060 long chg = 0;
1061
1062 /* Locate the region we are either in or before. */
1063 list_for_each_entry(rg, head, link)
1064 if (end <= rg->to)
1065 break;
1066 if (&rg->link == head)
1067 return 0;
1068
1069 /* If we are in the middle of a region then adjust it. */
1070 if (end > rg->from) {
1071 chg = rg->to - end;
1072 rg->to = end;
1073 rg = list_entry(rg->link.next, typeof(*rg), link);
1074 }
1075
1076 /* Drop any remaining regions. */
1077 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1078 if (&rg->link == head)
1079 break;
1080 chg += rg->to - rg->from;
1081 list_del(&rg->link);
1082 kfree(rg);
1083 }
1084 return chg;
1085}
1086
1087static int hugetlb_acct_memory(long delta)
1088{
1089 int ret = -ENOMEM;
1090
1091 spin_lock(&hugetlb_lock);
Ken Chen8a630112007-05-09 02:33:34 -07001092 /*
1093 * When cpuset is configured, it breaks the strict hugetlb page
1094 * reservation as the accounting is done on a global variable. Such
1095 * reservation is completely rubbish in the presence of cpuset because
1096 * the reservation is not checked against page availability for the
1097 * current cpuset. Application can still potentially OOM'ed by kernel
1098 * with lack of free htlb page in cpuset that the task is in.
1099 * Attempt to enforce strict accounting with cpuset is almost
1100 * impossible (or too ugly) because cpuset is too fluid that
1101 * task or memory node can be dynamically moved between cpusets.
1102 *
1103 * The change of semantics for shared hugetlb mapping with cpuset is
1104 * undesirable. However, in order to preserve some of the semantics,
1105 * we fall back to check against current free page availability as
1106 * a best attempt and hopefully to minimize the impact of changing
1107 * semantics that cpuset has.
1108 */
Adam Litkee4e574b2007-10-16 01:26:19 -07001109 if (delta > 0) {
1110 if (gather_surplus_pages(delta) < 0)
1111 goto out;
1112
1113 if (delta > cpuset_mems_nr(free_huge_pages_node))
1114 goto out;
1115 }
1116
1117 ret = 0;
1118 resv_huge_pages += delta;
1119 if (delta < 0)
1120 return_unused_surplus_pages((unsigned long) -delta);
1121
1122out:
1123 spin_unlock(&hugetlb_lock);
1124 return ret;
1125}
1126
1127int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1128{
1129 long ret, chg;
1130
1131 chg = region_chg(&inode->i_mapping->private_list, from, to);
1132 if (chg < 0)
1133 return chg;
Ken Chen8a630112007-05-09 02:33:34 -07001134
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001135 ret = hugetlb_acct_memory(chg);
1136 if (ret < 0)
1137 return ret;
1138 region_add(&inode->i_mapping->private_list, from, to);
1139 return 0;
1140}
1141
1142void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1143{
1144 long chg = region_truncate(&inode->i_mapping->private_list, offset);
1145 hugetlb_acct_memory(freed - chg);
1146}