blob: cb1b3a7ecdfcc5030ef0547f3544d697f3b6a30e [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
Christoph Lameter5da7ca82006-01-06 00:10:46 -080074static struct page *dequeue_huge_page(struct vm_area_struct *vma,
75 unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -070076{
Nishanth Aravamudan31a5c6e2007-07-15 23:38:02 -070077 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -070078 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -070079 struct mempolicy *mpol;
Mel Gorman396faf02007-07-17 04:03:13 -070080 struct zonelist *zonelist = huge_zonelist(vma, address,
Lee Schermerhorn480eccf2007-09-18 22:46:47 -070081 htlb_alloc_mask, &mpol);
Christoph Lameter96df9332006-01-06 00:10:45 -080082 struct zone **z;
Linus Torvalds1da177e2005-04-16 15:20:36 -070083
Christoph Lameter96df9332006-01-06 00:10:45 -080084 for (z = zonelist->zones; *z; z++) {
Christoph Lameter89fa3022006-09-25 23:31:55 -070085 nid = zone_to_nid(*z);
Mel Gorman396faf02007-07-17 04:03:13 -070086 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
Andrew Morton3abf7af2007-07-19 01:49:08 -070087 !list_empty(&hugepage_freelists[nid])) {
88 page = list_entry(hugepage_freelists[nid].next,
89 struct page, lru);
90 list_del(&page->lru);
91 free_huge_pages--;
92 free_huge_pages_node[nid]--;
Adam Litkee4e574b2007-10-16 01:26:19 -070093 if (vma && vma->vm_flags & VM_MAYSHARE)
94 resv_huge_pages--;
Ken Chen5ab3ee72007-07-23 18:44:00 -070095 break;
Andrew Morton3abf7af2007-07-19 01:49:08 -070096 }
Linus Torvalds1da177e2005-04-16 15:20:36 -070097 }
Lee Schermerhorn480eccf2007-09-18 22:46:47 -070098 mpol_free(mpol); /* unref if mpol !NULL */
Linus Torvalds1da177e2005-04-16 15:20:36 -070099 return page;
100}
101
Adam Litke6af2acb2007-10-16 01:26:16 -0700102static void update_and_free_page(struct page *page)
103{
104 int i;
105 nr_huge_pages--;
106 nr_huge_pages_node[page_to_nid(page)]--;
107 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
108 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
109 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
110 1 << PG_private | 1<< PG_writeback);
111 }
112 set_compound_page_dtor(page, NULL);
113 set_page_refcounted(page);
114 __free_pages(page, HUGETLB_PAGE_ORDER);
115}
116
David Gibson27a85ef2006-03-22 00:08:56 -0800117static void free_huge_page(struct page *page)
118{
Adam Litke7893d1d2007-10-16 01:26:18 -0700119 int nid = page_to_nid(page);
Adam Litkec79fb752007-11-14 16:59:38 -0800120 struct address_space *mapping;
David Gibson27a85ef2006-03-22 00:08:56 -0800121
Adam Litkec79fb752007-11-14 16:59:38 -0800122 mapping = (struct address_space *) page_private(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700123 BUG_ON(page_count(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800124 INIT_LIST_HEAD(&page->lru);
125
126 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700127 if (surplus_huge_pages_node[nid]) {
128 update_and_free_page(page);
129 surplus_huge_pages--;
130 surplus_huge_pages_node[nid]--;
131 } else {
132 enqueue_huge_page(page);
133 }
David Gibson27a85ef2006-03-22 00:08:56 -0800134 spin_unlock(&hugetlb_lock);
Adam Litkec79fb752007-11-14 16:59:38 -0800135 if (mapping)
Adam Litke9a119c02007-11-14 16:59:41 -0800136 hugetlb_put_quota(mapping, 1);
Adam Litkec79fb752007-11-14 16:59:38 -0800137 set_page_private(page, 0);
David Gibson27a85ef2006-03-22 00:08:56 -0800138}
139
Adam Litke7893d1d2007-10-16 01:26:18 -0700140/*
141 * Increment or decrement surplus_huge_pages. Keep node-specific counters
142 * balanced by operating on them in a round-robin fashion.
143 * Returns 1 if an adjustment was made.
144 */
145static int adjust_pool_surplus(int delta)
146{
147 static int prev_nid;
148 int nid = prev_nid;
149 int ret = 0;
150
151 VM_BUG_ON(delta != -1 && delta != 1);
152 do {
153 nid = next_node(nid, node_online_map);
154 if (nid == MAX_NUMNODES)
155 nid = first_node(node_online_map);
156
157 /* To shrink on this node, there must be a surplus page */
158 if (delta < 0 && !surplus_huge_pages_node[nid])
159 continue;
160 /* Surplus cannot exceed the total number of pages */
161 if (delta > 0 && surplus_huge_pages_node[nid] >=
162 nr_huge_pages_node[nid])
163 continue;
164
165 surplus_huge_pages += delta;
166 surplus_huge_pages_node[nid] += delta;
167 ret = 1;
168 break;
169 } while (nid != prev_nid);
170
171 prev_nid = nid;
172 return ret;
173}
174
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700175static struct page *alloc_fresh_huge_page_node(int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700178
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700179 page = alloc_pages_node(nid,
180 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
181 HUGETLB_PAGE_ORDER);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182 if (page) {
Andy Whitcroft33f2ef82006-12-06 20:33:32 -0800183 set_compound_page_dtor(page, free_huge_page);
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800184 spin_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185 nr_huge_pages++;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700186 nr_huge_pages_node[nid]++;
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800187 spin_unlock(&hugetlb_lock);
Nick Piggina4822892006-03-22 00:08:08 -0800188 put_page(page); /* free it into the hugepage allocator */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700190
191 return page;
192}
193
194static int alloc_fresh_huge_page(void)
195{
196 struct page *page;
197 int start_nid;
198 int next_nid;
199 int ret = 0;
200
201 start_nid = hugetlb_next_nid;
202
203 do {
204 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
205 if (page)
206 ret = 1;
207 /*
208 * Use a helper variable to find the next node and then
209 * copy it back to hugetlb_next_nid afterwards:
210 * otherwise there's a window in which a racer might
211 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
212 * But we don't need to use a spin_lock here: it really
213 * doesn't matter if occasionally a racer chooses the
214 * same nid as we do. Move nid forward in the mask even
215 * if we just successfully allocated a hugepage so that
216 * the next caller gets hugepages on the next node.
217 */
218 next_nid = next_node(hugetlb_next_nid, node_online_map);
219 if (next_nid == MAX_NUMNODES)
220 next_nid = first_node(node_online_map);
221 hugetlb_next_nid = next_nid;
222 } while (!page && hugetlb_next_nid != start_nid);
223
224 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225}
226
Adam Litke7893d1d2007-10-16 01:26:18 -0700227static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
228 unsigned long address)
229{
230 struct page *page;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800231 unsigned int nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700232
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800233 /*
234 * Assume we will successfully allocate the surplus page to
235 * prevent racing processes from causing the surplus to exceed
236 * overcommit
237 *
238 * This however introduces a different race, where a process B
239 * tries to grow the static hugepage pool while alloc_pages() is
240 * called by process A. B will only examine the per-node
241 * counters in determining if surplus huge pages can be
242 * converted to normal huge pages in adjust_pool_surplus(). A
243 * won't be able to increment the per-node counter, until the
244 * lock is dropped by B, but B doesn't drop hugetlb_lock until
245 * no more huge pages can be converted from surplus to normal
246 * state (and doesn't try to convert again). Thus, we have a
247 * case where a surplus huge page exists, the pool is grown, and
248 * the surplus huge page still exists after, even though it
249 * should just have been converted to a normal huge page. This
250 * does not leak memory, though, as the hugepage will be freed
251 * once it is out of use. It also does not allow the counters to
252 * go out of whack in adjust_pool_surplus() as we don't modify
253 * the node values until we've gotten the hugepage and only the
254 * per-node value is checked there.
255 */
256 spin_lock(&hugetlb_lock);
257 if (surplus_huge_pages >= nr_overcommit_huge_pages) {
258 spin_unlock(&hugetlb_lock);
259 return NULL;
260 } else {
261 nr_huge_pages++;
262 surplus_huge_pages++;
263 }
264 spin_unlock(&hugetlb_lock);
265
Adam Litke7893d1d2007-10-16 01:26:18 -0700266 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
267 HUGETLB_PAGE_ORDER);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800268
269 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700270 if (page) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800271 nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700272 set_compound_page_dtor(page, free_huge_page);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800273 /*
274 * We incremented the global counters already
275 */
276 nr_huge_pages_node[nid]++;
277 surplus_huge_pages_node[nid]++;
278 } else {
279 nr_huge_pages--;
280 surplus_huge_pages--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700281 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800282 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700283
284 return page;
285}
286
Adam Litkee4e574b2007-10-16 01:26:19 -0700287/*
288 * Increase the hugetlb pool such that it can accomodate a reservation
289 * of size 'delta'.
290 */
291static int gather_surplus_pages(int delta)
292{
293 struct list_head surplus_list;
294 struct page *page, *tmp;
295 int ret, i;
296 int needed, allocated;
297
298 needed = (resv_huge_pages + delta) - free_huge_pages;
299 if (needed <= 0)
300 return 0;
301
302 allocated = 0;
303 INIT_LIST_HEAD(&surplus_list);
304
305 ret = -ENOMEM;
306retry:
307 spin_unlock(&hugetlb_lock);
308 for (i = 0; i < needed; i++) {
309 page = alloc_buddy_huge_page(NULL, 0);
310 if (!page) {
311 /*
312 * We were not able to allocate enough pages to
313 * satisfy the entire reservation so we free what
314 * we've allocated so far.
315 */
316 spin_lock(&hugetlb_lock);
317 needed = 0;
318 goto free;
319 }
320
321 list_add(&page->lru, &surplus_list);
322 }
323 allocated += needed;
324
325 /*
326 * After retaking hugetlb_lock, we need to recalculate 'needed'
327 * because either resv_huge_pages or free_huge_pages may have changed.
328 */
329 spin_lock(&hugetlb_lock);
330 needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
331 if (needed > 0)
332 goto retry;
333
334 /*
335 * The surplus_list now contains _at_least_ the number of extra pages
336 * needed to accomodate the reservation. Add the appropriate number
337 * of pages to the hugetlb pool and free the extras back to the buddy
338 * allocator.
339 */
340 needed += allocated;
341 ret = 0;
342free:
343 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
344 list_del(&page->lru);
345 if ((--needed) >= 0)
346 enqueue_huge_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700347 else {
348 /*
349 * Decrement the refcount and free the page using its
350 * destructor. This must be done with hugetlb_lock
351 * unlocked which is safe because free_huge_page takes
352 * hugetlb_lock before deciding how to free the page.
353 */
354 spin_unlock(&hugetlb_lock);
355 put_page(page);
356 spin_lock(&hugetlb_lock);
357 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700358 }
359
360 return ret;
361}
362
363/*
364 * When releasing a hugetlb pool reservation, any surplus pages that were
365 * allocated to satisfy the reservation must be explicitly freed if they were
366 * never used.
367 */
Adrian Bunk8cde0452007-11-14 16:59:43 -0800368static void return_unused_surplus_pages(unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -0700369{
370 static int nid = -1;
371 struct page *page;
372 unsigned long nr_pages;
373
374 nr_pages = min(unused_resv_pages, surplus_huge_pages);
375
376 while (nr_pages) {
377 nid = next_node(nid, node_online_map);
378 if (nid == MAX_NUMNODES)
379 nid = first_node(node_online_map);
380
381 if (!surplus_huge_pages_node[nid])
382 continue;
383
384 if (!list_empty(&hugepage_freelists[nid])) {
385 page = list_entry(hugepage_freelists[nid].next,
386 struct page, lru);
387 list_del(&page->lru);
388 update_and_free_page(page);
389 free_huge_pages--;
390 free_huge_pages_node[nid]--;
391 surplus_huge_pages--;
392 surplus_huge_pages_node[nid]--;
393 nr_pages--;
394 }
395 }
396}
397
Adam Litke348ea202007-11-14 16:59:37 -0800398
399static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
400 unsigned long addr)
401{
402 struct page *page;
403
404 spin_lock(&hugetlb_lock);
405 page = dequeue_huge_page(vma, addr);
406 spin_unlock(&hugetlb_lock);
Adam Litke90d8b7e2007-11-14 16:59:42 -0800407 return page ? page : ERR_PTR(-VM_FAULT_OOM);
Adam Litke348ea202007-11-14 16:59:37 -0800408}
409
410static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
411 unsigned long addr)
412{
413 struct page *page = NULL;
414
Adam Litke90d8b7e2007-11-14 16:59:42 -0800415 if (hugetlb_get_quota(vma->vm_file->f_mapping, 1))
416 return ERR_PTR(-VM_FAULT_SIGBUS);
417
Adam Litke348ea202007-11-14 16:59:37 -0800418 spin_lock(&hugetlb_lock);
419 if (free_huge_pages > resv_huge_pages)
420 page = dequeue_huge_page(vma, addr);
421 spin_unlock(&hugetlb_lock);
Ken Chen68842c92008-01-14 00:55:19 -0800422 if (!page) {
Adam Litke348ea202007-11-14 16:59:37 -0800423 page = alloc_buddy_huge_page(vma, addr);
Ken Chen68842c92008-01-14 00:55:19 -0800424 if (!page) {
425 hugetlb_put_quota(vma->vm_file->f_mapping, 1);
426 return ERR_PTR(-VM_FAULT_OOM);
427 }
428 }
429 return page;
Adam Litke348ea202007-11-14 16:59:37 -0800430}
431
David Gibson27a85ef2006-03-22 00:08:56 -0800432static struct page *alloc_huge_page(struct vm_area_struct *vma,
433 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434{
Adam Litke348ea202007-11-14 16:59:37 -0800435 struct page *page;
Adam Litke2fc39ce2007-11-14 16:59:39 -0800436 struct address_space *mapping = vma->vm_file->f_mapping;
437
Adam Litke348ea202007-11-14 16:59:37 -0800438 if (vma->vm_flags & VM_MAYSHARE)
439 page = alloc_huge_page_shared(vma, addr);
440 else
441 page = alloc_huge_page_private(vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -0800442
443 if (!IS_ERR(page)) {
Adam Litke348ea202007-11-14 16:59:37 -0800444 set_page_refcounted(page);
Adam Litke2fc39ce2007-11-14 16:59:39 -0800445 set_page_private(page, (unsigned long) mapping);
Adam Litke90d8b7e2007-11-14 16:59:42 -0800446 }
447 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800448}
449
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450static int __init hugetlb_init(void)
451{
452 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453
Benjamin Herrenschmidt3c726f82005-11-07 11:06:55 +1100454 if (HPAGE_SHIFT == 0)
455 return 0;
456
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 for (i = 0; i < MAX_NUMNODES; ++i)
458 INIT_LIST_HEAD(&hugepage_freelists[i]);
459
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700460 hugetlb_next_nid = first_node(node_online_map);
461
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 for (i = 0; i < max_huge_pages; ++i) {
Nick Piggina4822892006-03-22 00:08:08 -0800463 if (!alloc_fresh_huge_page())
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 }
466 max_huge_pages = free_huge_pages = nr_huge_pages = i;
467 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
468 return 0;
469}
470module_init(hugetlb_init);
471
472static int __init hugetlb_setup(char *s)
473{
474 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
475 max_huge_pages = 0;
476 return 1;
477}
478__setup("hugepages=", hugetlb_setup);
479
Ken Chen8a630112007-05-09 02:33:34 -0700480static unsigned int cpuset_mems_nr(unsigned int *array)
481{
482 int node;
483 unsigned int nr = 0;
484
485 for_each_node_mask(node, cpuset_current_mems_allowed)
486 nr += array[node];
487
488 return nr;
489}
490
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491#ifdef CONFIG_SYSCTL
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492#ifdef CONFIG_HIGHMEM
493static void try_to_free_low(unsigned long count)
494{
Christoph Lameter4415cc82006-09-25 23:31:55 -0700495 int i;
496
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 for (i = 0; i < MAX_NUMNODES; ++i) {
498 struct page *page, *next;
499 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
Adam Litke6b0c8802007-10-16 01:26:23 -0700500 if (count >= nr_huge_pages)
501 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 if (PageHighMem(page))
503 continue;
504 list_del(&page->lru);
505 update_and_free_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 free_huge_pages--;
Christoph Lameter4415cc82006-09-25 23:31:55 -0700507 free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508 }
509 }
510}
511#else
512static inline void try_to_free_low(unsigned long count)
513{
514}
515#endif
516
Adam Litke7893d1d2007-10-16 01:26:18 -0700517#define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518static unsigned long set_max_huge_pages(unsigned long count)
519{
Adam Litke7893d1d2007-10-16 01:26:18 -0700520 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521
Adam Litke7893d1d2007-10-16 01:26:18 -0700522 /*
523 * Increase the pool size
524 * First take pages out of surplus state. Then make up the
525 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800526 *
527 * We might race with alloc_buddy_huge_page() here and be unable
528 * to convert a surplus huge page to a normal huge page. That is
529 * not critical, though, it just means the overall size of the
530 * pool might be one hugepage larger than it needs to be, but
531 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -0700532 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700534 while (surplus_huge_pages && count > persistent_huge_pages) {
535 if (!adjust_pool_surplus(-1))
536 break;
537 }
538
539 while (count > persistent_huge_pages) {
540 int ret;
541 /*
542 * If this allocation races such that we no longer need the
543 * page, free_huge_page will handle it by freeing the page
544 * and reducing the surplus.
545 */
546 spin_unlock(&hugetlb_lock);
547 ret = alloc_fresh_huge_page();
548 spin_lock(&hugetlb_lock);
549 if (!ret)
550 goto out;
551
552 }
Adam Litke7893d1d2007-10-16 01:26:18 -0700553
554 /*
555 * Decrease the pool size
556 * First return free pages to the buddy allocator (being careful
557 * to keep enough around to satisfy reservations). Then place
558 * pages into surplus state as needed so the pool will shrink
559 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800560 *
561 * By placing pages into the surplus state independent of the
562 * overcommit value, we are allowing the surplus pool size to
563 * exceed overcommit. There are few sane options here. Since
564 * alloc_buddy_huge_page() is checking the global counter,
565 * though, we'll note that we're not allowed to exceed surplus
566 * and won't grow the pool anywhere else. Not until one of the
567 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -0700568 */
Adam Litke6b0c8802007-10-16 01:26:23 -0700569 min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
570 min_count = max(count, min_count);
Adam Litke7893d1d2007-10-16 01:26:18 -0700571 try_to_free_low(min_count);
572 while (min_count < persistent_huge_pages) {
Christoph Lameter5da7ca82006-01-06 00:10:46 -0800573 struct page *page = dequeue_huge_page(NULL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 if (!page)
575 break;
576 update_and_free_page(page);
577 }
Adam Litke7893d1d2007-10-16 01:26:18 -0700578 while (count < persistent_huge_pages) {
579 if (!adjust_pool_surplus(1))
580 break;
581 }
582out:
583 ret = persistent_huge_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700585 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586}
587
588int hugetlb_sysctl_handler(struct ctl_table *table, int write,
589 struct file *file, void __user *buffer,
590 size_t *length, loff_t *ppos)
591{
592 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
593 max_huge_pages = set_max_huge_pages(max_huge_pages);
594 return 0;
595}
Mel Gorman396faf02007-07-17 04:03:13 -0700596
597int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
598 struct file *file, void __user *buffer,
599 size_t *length, loff_t *ppos)
600{
601 proc_dointvec(table, write, file, buffer, length, ppos);
602 if (hugepages_treat_as_movable)
603 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
604 else
605 htlb_alloc_mask = GFP_HIGHUSER;
606 return 0;
607}
608
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -0800609int hugetlb_overcommit_handler(struct ctl_table *table, int write,
610 struct file *file, void __user *buffer,
611 size_t *length, loff_t *ppos)
612{
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -0800613 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
Nishanth Aravamudan064d9ef2008-02-13 15:03:19 -0800614 spin_lock(&hugetlb_lock);
615 nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -0800616 spin_unlock(&hugetlb_lock);
617 return 0;
618}
619
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620#endif /* CONFIG_SYSCTL */
621
622int hugetlb_report_meminfo(char *buf)
623{
624 return sprintf(buf,
625 "HugePages_Total: %5lu\n"
626 "HugePages_Free: %5lu\n"
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700627 "HugePages_Rsvd: %5lu\n"
Adam Litke7893d1d2007-10-16 01:26:18 -0700628 "HugePages_Surp: %5lu\n"
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629 "Hugepagesize: %5lu kB\n",
630 nr_huge_pages,
631 free_huge_pages,
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700632 resv_huge_pages,
Adam Litke7893d1d2007-10-16 01:26:18 -0700633 surplus_huge_pages,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 HPAGE_SIZE/1024);
635}
636
637int hugetlb_report_node_meminfo(int nid, char *buf)
638{
639 return sprintf(buf,
640 "Node %d HugePages_Total: %5u\n"
641 "Node %d HugePages_Free: %5u\n",
642 nid, nr_huge_pages_node[nid],
643 nid, free_huge_pages_node[nid]);
644}
645
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
647unsigned long hugetlb_total_pages(void)
648{
649 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
650}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651
652/*
653 * We cannot handle pagefaults against hugetlb pages at all. They cause
654 * handle_mm_fault() to try to instantiate regular-sized pages in the
655 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
656 * this far.
657 */
Nick Piggind0217ac2007-07-19 01:47:03 -0700658static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659{
660 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -0700661 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662}
663
664struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -0700665 .fault = hugetlb_vm_op_fault,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666};
667
David Gibson1e8f8892006-01-06 00:10:44 -0800668static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
669 int writable)
David Gibson63551ae2005-06-21 17:14:44 -0700670{
671 pte_t entry;
672
David Gibson1e8f8892006-01-06 00:10:44 -0800673 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -0700674 entry =
675 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
676 } else {
677 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
678 }
679 entry = pte_mkyoung(entry);
680 entry = pte_mkhuge(entry);
681
682 return entry;
683}
684
David Gibson1e8f8892006-01-06 00:10:44 -0800685static void set_huge_ptep_writable(struct vm_area_struct *vma,
686 unsigned long address, pte_t *ptep)
687{
688 pte_t entry;
689
690 entry = pte_mkwrite(pte_mkdirty(*ptep));
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -0700691 if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
692 update_mmu_cache(vma, address, entry);
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -0700693 }
David Gibson1e8f8892006-01-06 00:10:44 -0800694}
695
696
David Gibson63551ae2005-06-21 17:14:44 -0700697int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
698 struct vm_area_struct *vma)
699{
700 pte_t *src_pte, *dst_pte, entry;
701 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -0700702 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -0800703 int cow;
704
705 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -0700706
Hugh Dickins1c598272005-10-19 21:23:43 -0700707 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
Hugh Dickinsc74df322005-10-29 18:16:23 -0700708 src_pte = huge_pte_offset(src, addr);
709 if (!src_pte)
710 continue;
David Gibson63551ae2005-06-21 17:14:44 -0700711 dst_pte = huge_pte_alloc(dst, addr);
712 if (!dst_pte)
713 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -0800714
715 /* If the pagetables are shared don't copy or take references */
716 if (dst_pte == src_pte)
717 continue;
718
Hugh Dickinsc74df322005-10-29 18:16:23 -0700719 spin_lock(&dst->page_table_lock);
Hugh Dickins1c598272005-10-19 21:23:43 -0700720 spin_lock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700721 if (!pte_none(*src_pte)) {
David Gibson1e8f8892006-01-06 00:10:44 -0800722 if (cow)
723 ptep_set_wrprotect(src, addr, src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -0700724 entry = *src_pte;
725 ptepage = pte_page(entry);
726 get_page(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -0700727 set_huge_pte_at(dst, addr, dst_pte, entry);
728 }
729 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700730 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700731 }
732 return 0;
733
734nomem:
735 return -ENOMEM;
736}
737
Chen, Kenneth W502717f2006-10-11 01:20:46 -0700738void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
739 unsigned long end)
David Gibson63551ae2005-06-21 17:14:44 -0700740{
741 struct mm_struct *mm = vma->vm_mm;
742 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -0700743 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -0700744 pte_t pte;
745 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700746 struct page *tmp;
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -0800747 /*
748 * A page gathering list, protected by per file i_mmap_lock. The
749 * lock is used to avoid list corruption from multiple unmapping
750 * of the same page since we are using page->lru.
751 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700752 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -0700753
754 WARN_ON(!is_vm_hugetlb_page(vma));
755 BUG_ON(start & ~HPAGE_MASK);
756 BUG_ON(end & ~HPAGE_MASK);
757
Hugh Dickins508034a2005-10-29 18:16:30 -0700758 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700759 for (address = start; address < end; address += HPAGE_SIZE) {
David Gibsonc7546f82005-08-05 11:59:35 -0700760 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -0700761 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -0700762 continue;
763
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800764 if (huge_pmd_unshare(mm, &address, ptep))
765 continue;
766
David Gibsonc7546f82005-08-05 11:59:35 -0700767 pte = huge_ptep_get_and_clear(mm, address, ptep);
David Gibson63551ae2005-06-21 17:14:44 -0700768 if (pte_none(pte))
769 continue;
David Gibsonc7546f82005-08-05 11:59:35 -0700770
David Gibson63551ae2005-06-21 17:14:44 -0700771 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -0800772 if (pte_dirty(pte))
773 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700774 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -0700775 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -0700777 flush_tlb_range(vma, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700778 list_for_each_entry_safe(page, tmp, &page_list, lru) {
779 list_del(&page->lru);
780 put_page(page);
781 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782}
David Gibson63551ae2005-06-21 17:14:44 -0700783
Chen, Kenneth W502717f2006-10-11 01:20:46 -0700784void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
785 unsigned long end)
786{
787 /*
788 * It is undesirable to test vma->vm_file as it should be non-null
789 * for valid hugetlb area. However, vm_file will be NULL in the error
790 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
791 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
792 * to clean up. Since no pte has actually been setup, it is safe to
793 * do nothing in this case.
794 */
795 if (vma->vm_file) {
796 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
797 __unmap_hugepage_range(vma, start, end);
798 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
799 }
800}
801
David Gibson1e8f8892006-01-06 00:10:44 -0800802static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
803 unsigned long address, pte_t *ptep, pte_t pte)
804{
805 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -0800806 int avoidcopy;
David Gibson1e8f8892006-01-06 00:10:44 -0800807
808 old_page = pte_page(pte);
809
810 /* If no-one else is actually using this page, avoid the copy
811 * and just make the page writable */
812 avoidcopy = (page_count(old_page) == 1);
813 if (avoidcopy) {
814 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -0700815 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800816 }
817
818 page_cache_get(old_page);
Christoph Lameter5da7ca82006-01-06 00:10:46 -0800819 new_page = alloc_huge_page(vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -0800820
Adam Litke2fc39ce2007-11-14 16:59:39 -0800821 if (IS_ERR(new_page)) {
David Gibson1e8f8892006-01-06 00:10:44 -0800822 page_cache_release(old_page);
Adam Litke2fc39ce2007-11-14 16:59:39 -0800823 return -PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -0800824 }
825
826 spin_unlock(&mm->page_table_lock);
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000827 copy_huge_page(new_page, old_page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -0800828 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -0800829 spin_lock(&mm->page_table_lock);
830
831 ptep = huge_pte_offset(mm, address & HPAGE_MASK);
832 if (likely(pte_same(*ptep, pte))) {
833 /* Break COW */
834 set_huge_pte_at(mm, address, ptep,
835 make_huge_pte(vma, new_page, 1));
836 /* Make the old page be freed below */
837 new_page = old_page;
838 }
839 page_cache_release(new_page);
840 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -0700841 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800842}
843
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -0700844static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
David Gibson1e8f8892006-01-06 00:10:44 -0800845 unsigned long address, pte_t *ptep, int write_access)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100846{
847 int ret = VM_FAULT_SIGBUS;
Adam Litke4c887262005-10-29 18:16:46 -0700848 unsigned long idx;
849 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -0700850 struct page *page;
851 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -0800852 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -0700853
Adam Litke4c887262005-10-29 18:16:46 -0700854 mapping = vma->vm_file->f_mapping;
855 idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
856 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
857
858 /*
859 * Use page lock to guard against racing truncation
860 * before we get page_table_lock.
861 */
Christoph Lameter6bda6662006-01-06 00:10:49 -0800862retry:
863 page = find_lock_page(mapping, idx);
864 if (!page) {
Hugh Dickinsebed4bf2006-10-28 10:38:43 -0700865 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
866 if (idx >= size)
867 goto out;
Christoph Lameter6bda6662006-01-06 00:10:49 -0800868 page = alloc_huge_page(vma, address);
Adam Litke2fc39ce2007-11-14 16:59:39 -0800869 if (IS_ERR(page)) {
870 ret = -PTR_ERR(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -0800871 goto out;
872 }
David Gibson79ac6ba2006-03-22 00:08:51 -0800873 clear_huge_page(page, address);
Nick Piggin0ed361d2008-02-04 22:29:34 -0800874 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100875
Christoph Lameter6bda6662006-01-06 00:10:49 -0800876 if (vma->vm_flags & VM_SHARED) {
877 int err;
Ken Chen45c682a2007-11-14 16:59:44 -0800878 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -0800879
880 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
881 if (err) {
882 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -0800883 if (err == -EEXIST)
884 goto retry;
885 goto out;
886 }
Ken Chen45c682a2007-11-14 16:59:44 -0800887
888 spin_lock(&inode->i_lock);
889 inode->i_blocks += BLOCKS_PER_HUGEPAGE;
890 spin_unlock(&inode->i_lock);
Christoph Lameter6bda6662006-01-06 00:10:49 -0800891 } else
892 lock_page(page);
893 }
David Gibson1e8f8892006-01-06 00:10:44 -0800894
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100895 spin_lock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700896 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
897 if (idx >= size)
898 goto backout;
899
Nick Piggin83c54072007-07-19 01:47:05 -0700900 ret = 0;
Adam Litke86e52162006-01-06 00:10:43 -0800901 if (!pte_none(*ptep))
Adam Litke4c887262005-10-29 18:16:46 -0700902 goto backout;
903
David Gibson1e8f8892006-01-06 00:10:44 -0800904 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
905 && (vma->vm_flags & VM_SHARED)));
906 set_huge_pte_at(mm, address, ptep, new_pte);
907
908 if (write_access && !(vma->vm_flags & VM_SHARED)) {
909 /* Optimization, do the COW without a second fault */
910 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
911 }
912
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100913 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700914 unlock_page(page);
915out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100916 return ret;
Adam Litke4c887262005-10-29 18:16:46 -0700917
918backout:
919 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700920 unlock_page(page);
921 put_page(page);
922 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100923}
924
Adam Litke86e52162006-01-06 00:10:43 -0800925int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
926 unsigned long address, int write_access)
927{
928 pte_t *ptep;
929 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -0800930 int ret;
David Gibson3935baa2006-03-22 00:08:53 -0800931 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800932
933 ptep = huge_pte_alloc(mm, address);
934 if (!ptep)
935 return VM_FAULT_OOM;
936
David Gibson3935baa2006-03-22 00:08:53 -0800937 /*
938 * Serialize hugepage allocation and instantiation, so that we don't
939 * get spurious allocation failures if two CPUs race to instantiate
940 * the same page in the page cache.
941 */
942 mutex_lock(&hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800943 entry = *ptep;
David Gibson3935baa2006-03-22 00:08:53 -0800944 if (pte_none(entry)) {
945 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
946 mutex_unlock(&hugetlb_instantiation_mutex);
947 return ret;
948 }
Adam Litke86e52162006-01-06 00:10:43 -0800949
Nick Piggin83c54072007-07-19 01:47:05 -0700950 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800951
952 spin_lock(&mm->page_table_lock);
953 /* Check for a racing update before calling hugetlb_cow */
954 if (likely(pte_same(entry, *ptep)))
955 if (write_access && !pte_write(entry))
956 ret = hugetlb_cow(mm, vma, address, ptep, entry);
957 spin_unlock(&mm->page_table_lock);
David Gibson3935baa2006-03-22 00:08:53 -0800958 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -0800959
960 return ret;
Adam Litke86e52162006-01-06 00:10:43 -0800961}
962
David Gibson63551ae2005-06-21 17:14:44 -0700963int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
964 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -0800965 unsigned long *position, int *length, int i,
966 int write)
David Gibson63551ae2005-06-21 17:14:44 -0700967{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800968 unsigned long pfn_offset;
969 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -0700970 int remainder = *length;
971
Hugh Dickins1c598272005-10-19 21:23:43 -0700972 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700973 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -0700974 pte_t *pte;
975 struct page *page;
976
977 /*
978 * Some archs (sparc64, sh*) have multiple pte_ts to
979 * each hugepage. We have to make * sure we get the
980 * first, for the page indexing below to work.
981 */
982 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
983
Adam Litke72fad712007-12-10 15:49:28 -0800984 if (!pte || pte_none(*pte) || (write && !pte_write(*pte))) {
Adam Litke4c887262005-10-29 18:16:46 -0700985 int ret;
986
987 spin_unlock(&mm->page_table_lock);
Adam Litke5b23dbe2007-11-14 16:59:33 -0800988 ret = hugetlb_fault(mm, vma, vaddr, write);
Adam Litke4c887262005-10-29 18:16:46 -0700989 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -0700990 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -0700991 continue;
992
993 remainder = 0;
994 if (!i)
995 i = -EFAULT;
996 break;
997 }
David Gibson63551ae2005-06-21 17:14:44 -0700998
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800999 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1000 page = pte_page(*pte);
1001same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08001002 if (pages) {
1003 get_page(page);
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08001004 pages[i] = page + pfn_offset;
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08001005 }
David Gibson63551ae2005-06-21 17:14:44 -07001006
1007 if (vmas)
1008 vmas[i] = vma;
1009
1010 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08001011 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07001012 --remainder;
1013 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08001014 if (vaddr < vma->vm_end && remainder &&
1015 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1016 /*
1017 * We use pfn_offset to avoid touching the pageframes
1018 * of this compound page.
1019 */
1020 goto same_page;
1021 }
David Gibson63551ae2005-06-21 17:14:44 -07001022 }
Hugh Dickins1c598272005-10-19 21:23:43 -07001023 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07001024 *length = remainder;
1025 *position = vaddr;
1026
1027 return i;
1028}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001029
1030void hugetlb_change_protection(struct vm_area_struct *vma,
1031 unsigned long address, unsigned long end, pgprot_t newprot)
1032{
1033 struct mm_struct *mm = vma->vm_mm;
1034 unsigned long start = address;
1035 pte_t *ptep;
1036 pte_t pte;
1037
1038 BUG_ON(address >= end);
1039 flush_cache_range(vma, address, end);
1040
Chen, Kenneth W39dde652006-12-06 20:32:03 -08001041 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001042 spin_lock(&mm->page_table_lock);
1043 for (; address < end; address += HPAGE_SIZE) {
1044 ptep = huge_pte_offset(mm, address);
1045 if (!ptep)
1046 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08001047 if (huge_pmd_unshare(mm, &address, ptep))
1048 continue;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001049 if (!pte_none(*ptep)) {
1050 pte = huge_ptep_get_and_clear(mm, address, ptep);
1051 pte = pte_mkhuge(pte_modify(pte, newprot));
1052 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001053 }
1054 }
1055 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08001056 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08001057
1058 flush_tlb_range(vma, start, end);
1059}
1060
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001061struct file_region {
1062 struct list_head link;
1063 long from;
1064 long to;
1065};
1066
1067static long region_add(struct list_head *head, long f, long t)
1068{
1069 struct file_region *rg, *nrg, *trg;
1070
1071 /* Locate the region we are either in or before. */
1072 list_for_each_entry(rg, head, link)
1073 if (f <= rg->to)
1074 break;
1075
1076 /* Round our left edge to the current segment if it encloses us. */
1077 if (f > rg->from)
1078 f = rg->from;
1079
1080 /* Check for and consume any regions we now overlap with. */
1081 nrg = rg;
1082 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1083 if (&rg->link == head)
1084 break;
1085 if (rg->from > t)
1086 break;
1087
1088 /* If this area reaches higher then extend our area to
1089 * include it completely. If this is not the first area
1090 * which we intend to reuse, free it. */
1091 if (rg->to > t)
1092 t = rg->to;
1093 if (rg != nrg) {
1094 list_del(&rg->link);
1095 kfree(rg);
1096 }
1097 }
1098 nrg->from = f;
1099 nrg->to = t;
1100 return 0;
1101}
1102
1103static long region_chg(struct list_head *head, long f, long t)
1104{
1105 struct file_region *rg, *nrg;
1106 long chg = 0;
1107
1108 /* Locate the region we are before or in. */
1109 list_for_each_entry(rg, head, link)
1110 if (f <= rg->to)
1111 break;
1112
1113 /* If we are below the current region then a new region is required.
1114 * Subtle, allocate a new region at the position but make it zero
Simon Arlott183ff222007-10-20 01:27:18 +02001115 * size such that we can guarantee to record the reservation. */
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001116 if (&rg->link == head || t < rg->from) {
1117 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07001118 if (!nrg)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001119 return -ENOMEM;
1120 nrg->from = f;
1121 nrg->to = f;
1122 INIT_LIST_HEAD(&nrg->link);
1123 list_add(&nrg->link, rg->link.prev);
1124
1125 return t - f;
1126 }
1127
1128 /* Round our left edge to the current segment if it encloses us. */
1129 if (f > rg->from)
1130 f = rg->from;
1131 chg = t - f;
1132
1133 /* Check for and consume any regions we now overlap with. */
1134 list_for_each_entry(rg, rg->link.prev, link) {
1135 if (&rg->link == head)
1136 break;
1137 if (rg->from > t)
1138 return chg;
1139
1140 /* We overlap with this area, if it extends futher than
1141 * us then we must extend ourselves. Account for its
1142 * existing reservation. */
1143 if (rg->to > t) {
1144 chg += rg->to - t;
1145 t = rg->to;
1146 }
1147 chg -= rg->to - rg->from;
1148 }
1149 return chg;
1150}
1151
1152static long region_truncate(struct list_head *head, long end)
1153{
1154 struct file_region *rg, *trg;
1155 long chg = 0;
1156
1157 /* Locate the region we are either in or before. */
1158 list_for_each_entry(rg, head, link)
1159 if (end <= rg->to)
1160 break;
1161 if (&rg->link == head)
1162 return 0;
1163
1164 /* If we are in the middle of a region then adjust it. */
1165 if (end > rg->from) {
1166 chg = rg->to - end;
1167 rg->to = end;
1168 rg = list_entry(rg->link.next, typeof(*rg), link);
1169 }
1170
1171 /* Drop any remaining regions. */
1172 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1173 if (&rg->link == head)
1174 break;
1175 chg += rg->to - rg->from;
1176 list_del(&rg->link);
1177 kfree(rg);
1178 }
1179 return chg;
1180}
1181
1182static int hugetlb_acct_memory(long delta)
1183{
1184 int ret = -ENOMEM;
1185
1186 spin_lock(&hugetlb_lock);
Ken Chen8a630112007-05-09 02:33:34 -07001187 /*
1188 * When cpuset is configured, it breaks the strict hugetlb page
1189 * reservation as the accounting is done on a global variable. Such
1190 * reservation is completely rubbish in the presence of cpuset because
1191 * the reservation is not checked against page availability for the
1192 * current cpuset. Application can still potentially OOM'ed by kernel
1193 * with lack of free htlb page in cpuset that the task is in.
1194 * Attempt to enforce strict accounting with cpuset is almost
1195 * impossible (or too ugly) because cpuset is too fluid that
1196 * task or memory node can be dynamically moved between cpusets.
1197 *
1198 * The change of semantics for shared hugetlb mapping with cpuset is
1199 * undesirable. However, in order to preserve some of the semantics,
1200 * we fall back to check against current free page availability as
1201 * a best attempt and hopefully to minimize the impact of changing
1202 * semantics that cpuset has.
1203 */
Adam Litkee4e574b2007-10-16 01:26:19 -07001204 if (delta > 0) {
1205 if (gather_surplus_pages(delta) < 0)
1206 goto out;
1207
1208 if (delta > cpuset_mems_nr(free_huge_pages_node))
1209 goto out;
1210 }
1211
1212 ret = 0;
1213 resv_huge_pages += delta;
1214 if (delta < 0)
1215 return_unused_surplus_pages((unsigned long) -delta);
1216
1217out:
1218 spin_unlock(&hugetlb_lock);
1219 return ret;
1220}
1221
1222int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1223{
1224 long ret, chg;
1225
1226 chg = region_chg(&inode->i_mapping->private_list, from, to);
1227 if (chg < 0)
1228 return chg;
Ken Chen8a630112007-05-09 02:33:34 -07001229
Adam Litke90d8b7e2007-11-14 16:59:42 -08001230 if (hugetlb_get_quota(inode->i_mapping, chg))
1231 return -ENOSPC;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001232 ret = hugetlb_acct_memory(chg);
Ken Chen68842c92008-01-14 00:55:19 -08001233 if (ret < 0) {
1234 hugetlb_put_quota(inode->i_mapping, chg);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001235 return ret;
Ken Chen68842c92008-01-14 00:55:19 -08001236 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001237 region_add(&inode->i_mapping->private_list, from, to);
1238 return 0;
1239}
1240
1241void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1242{
1243 long chg = region_truncate(&inode->i_mapping->private_list, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08001244
1245 spin_lock(&inode->i_lock);
1246 inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1247 spin_unlock(&inode->i_lock);
1248
Adam Litke90d8b7e2007-11-14 16:59:42 -08001249 hugetlb_put_quota(inode->i_mapping, (chg - freed));
1250 hugetlb_acct_memory(-(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001251}