blob: 3be79dc18c5c65a0c3b251d4849e3c895eadefd3 [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>
Andi Kleenaa888a72008-07-23 21:27:47 -070017#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070018#include <linux/sysfs.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080019
David Gibson63551ae2005-06-21 17:14:44 -070020#include <asm/page.h>
21#include <asm/pgtable.h>
22
23#include <linux/hugetlb.h>
Nick Piggin7835e982006-03-22 00:08:40 -080024#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070025
26const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
Mel Gorman396faf02007-07-17 04:03:13 -070027static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
28unsigned long hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070029
Andi Kleene5ff2152008-07-23 21:27:42 -070030static int max_hstate;
31unsigned int default_hstate_idx;
32struct hstate hstates[HUGE_MAX_HSTATE];
33
Jon Tollefson53ba51d2008-07-23 21:27:52 -070034__initdata LIST_HEAD(huge_boot_pages);
35
Andi Kleene5ff2152008-07-23 21:27:42 -070036/* for command line parsing */
37static struct hstate * __initdata parsed_hstate;
38static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070039static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070040
41#define for_each_hstate(h) \
42 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
Mel Gorman396faf02007-07-17 04:03:13 -070043
David Gibson3935baa2006-03-22 00:08:53 -080044/*
45 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
46 */
47static DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080048
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -070049/*
Andy Whitcroft96822902008-07-23 21:27:29 -070050 * Region tracking -- allows tracking of reservations and instantiated pages
51 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -070052 *
53 * The region data structures are protected by a combination of the mmap_sem
54 * and the hugetlb_instantion_mutex. To access or modify a region the caller
55 * must either hold the mmap_sem for write, or the mmap_sem for read and
56 * the hugetlb_instantiation mutex:
57 *
58 * down_write(&mm->mmap_sem);
59 * or
60 * down_read(&mm->mmap_sem);
61 * mutex_lock(&hugetlb_instantiation_mutex);
Andy Whitcroft96822902008-07-23 21:27:29 -070062 */
63struct file_region {
64 struct list_head link;
65 long from;
66 long to;
67};
68
69static long region_add(struct list_head *head, long f, long t)
70{
71 struct file_region *rg, *nrg, *trg;
72
73 /* Locate the region we are either in or before. */
74 list_for_each_entry(rg, head, link)
75 if (f <= rg->to)
76 break;
77
78 /* Round our left edge to the current segment if it encloses us. */
79 if (f > rg->from)
80 f = rg->from;
81
82 /* Check for and consume any regions we now overlap with. */
83 nrg = rg;
84 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
85 if (&rg->link == head)
86 break;
87 if (rg->from > t)
88 break;
89
90 /* If this area reaches higher then extend our area to
91 * include it completely. If this is not the first area
92 * which we intend to reuse, free it. */
93 if (rg->to > t)
94 t = rg->to;
95 if (rg != nrg) {
96 list_del(&rg->link);
97 kfree(rg);
98 }
99 }
100 nrg->from = f;
101 nrg->to = t;
102 return 0;
103}
104
105static long region_chg(struct list_head *head, long f, long t)
106{
107 struct file_region *rg, *nrg;
108 long chg = 0;
109
110 /* Locate the region we are before or in. */
111 list_for_each_entry(rg, head, link)
112 if (f <= rg->to)
113 break;
114
115 /* If we are below the current region then a new region is required.
116 * Subtle, allocate a new region at the position but make it zero
117 * size such that we can guarantee to record the reservation. */
118 if (&rg->link == head || t < rg->from) {
119 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
120 if (!nrg)
121 return -ENOMEM;
122 nrg->from = f;
123 nrg->to = f;
124 INIT_LIST_HEAD(&nrg->link);
125 list_add(&nrg->link, rg->link.prev);
126
127 return t - f;
128 }
129
130 /* Round our left edge to the current segment if it encloses us. */
131 if (f > rg->from)
132 f = rg->from;
133 chg = t - f;
134
135 /* Check for and consume any regions we now overlap with. */
136 list_for_each_entry(rg, rg->link.prev, link) {
137 if (&rg->link == head)
138 break;
139 if (rg->from > t)
140 return chg;
141
142 /* We overlap with this area, if it extends futher than
143 * us then we must extend ourselves. Account for its
144 * existing reservation. */
145 if (rg->to > t) {
146 chg += rg->to - t;
147 t = rg->to;
148 }
149 chg -= rg->to - rg->from;
150 }
151 return chg;
152}
153
154static long region_truncate(struct list_head *head, long end)
155{
156 struct file_region *rg, *trg;
157 long chg = 0;
158
159 /* Locate the region we are either in or before. */
160 list_for_each_entry(rg, head, link)
161 if (end <= rg->to)
162 break;
163 if (&rg->link == head)
164 return 0;
165
166 /* If we are in the middle of a region then adjust it. */
167 if (end > rg->from) {
168 chg = rg->to - end;
169 rg->to = end;
170 rg = list_entry(rg->link.next, typeof(*rg), link);
171 }
172
173 /* Drop any remaining regions. */
174 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
175 if (&rg->link == head)
176 break;
177 chg += rg->to - rg->from;
178 list_del(&rg->link);
179 kfree(rg);
180 }
181 return chg;
182}
183
Andy Whitcroft84afd992008-07-23 21:27:32 -0700184static long region_count(struct list_head *head, long f, long t)
185{
186 struct file_region *rg;
187 long chg = 0;
188
189 /* Locate each segment we overlap with, and count that overlap. */
190 list_for_each_entry(rg, head, link) {
191 int seg_from;
192 int seg_to;
193
194 if (rg->to <= f)
195 continue;
196 if (rg->from >= t)
197 break;
198
199 seg_from = max(rg->from, f);
200 seg_to = min(rg->to, t);
201
202 chg += seg_to - seg_from;
203 }
204
205 return chg;
206}
207
Andy Whitcroft96822902008-07-23 21:27:29 -0700208/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700209 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700210 * the mapping, in pagecache page units; huge pages here.
211 */
Andi Kleena5516432008-07-23 21:27:41 -0700212static pgoff_t vma_hugecache_offset(struct hstate *h,
213 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700214{
Andi Kleena5516432008-07-23 21:27:41 -0700215 return ((address - vma->vm_start) >> huge_page_shift(h)) +
216 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700217}
218
Andy Whitcroft84afd992008-07-23 21:27:32 -0700219/*
220 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
221 * bits of the reservation map pointer, which are always clear due to
222 * alignment.
223 */
224#define HPAGE_RESV_OWNER (1UL << 0)
225#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700226#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700227
Mel Gormana1e78772008-07-23 21:27:23 -0700228/*
229 * These helpers are used to track how many pages are reserved for
230 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
231 * is guaranteed to have their future faults succeed.
232 *
233 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
234 * the reserve counters are updated with the hugetlb_lock held. It is safe
235 * to reset the VMA at fork() time as it is not in use yet and there is no
236 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700237 *
238 * The private mapping reservation is represented in a subtly different
239 * manner to a shared mapping. A shared mapping has a region map associated
240 * with the underlying file, this region map represents the backing file
241 * pages which have ever had a reservation assigned which this persists even
242 * after the page is instantiated. A private mapping has a region map
243 * associated with the original mmap which is attached to all VMAs which
244 * reference it, this region map represents those offsets which have consumed
245 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700246 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700247static unsigned long get_vma_private_data(struct vm_area_struct *vma)
248{
249 return (unsigned long)vma->vm_private_data;
250}
251
252static void set_vma_private_data(struct vm_area_struct *vma,
253 unsigned long value)
254{
255 vma->vm_private_data = (void *)value;
256}
257
Andy Whitcroft84afd992008-07-23 21:27:32 -0700258struct resv_map {
259 struct kref refs;
260 struct list_head regions;
261};
262
263struct resv_map *resv_map_alloc(void)
264{
265 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
266 if (!resv_map)
267 return NULL;
268
269 kref_init(&resv_map->refs);
270 INIT_LIST_HEAD(&resv_map->regions);
271
272 return resv_map;
273}
274
275void resv_map_release(struct kref *ref)
276{
277 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
278
279 /* Clear out any active regions before we release the map. */
280 region_truncate(&resv_map->regions, 0);
281 kfree(resv_map);
282}
283
284static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700285{
286 VM_BUG_ON(!is_vm_hugetlb_page(vma));
287 if (!(vma->vm_flags & VM_SHARED))
Andy Whitcroft84afd992008-07-23 21:27:32 -0700288 return (struct resv_map *)(get_vma_private_data(vma) &
289 ~HPAGE_RESV_MASK);
Mel Gormana1e78772008-07-23 21:27:23 -0700290 return 0;
291}
292
Andy Whitcroft84afd992008-07-23 21:27:32 -0700293static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700294{
295 VM_BUG_ON(!is_vm_hugetlb_page(vma));
296 VM_BUG_ON(vma->vm_flags & VM_SHARED);
297
Andy Whitcroft84afd992008-07-23 21:27:32 -0700298 set_vma_private_data(vma, (get_vma_private_data(vma) &
299 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700300}
301
302static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
303{
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700304 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700305 VM_BUG_ON(vma->vm_flags & VM_SHARED);
306
307 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700308}
309
310static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
311{
312 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700313
314 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700315}
316
317/* Decrement the reserved pages in the hugepage pool by one */
Andi Kleena5516432008-07-23 21:27:41 -0700318static void decrement_hugepage_resv_vma(struct hstate *h,
319 struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700320{
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700321 if (vma->vm_flags & VM_NORESERVE)
322 return;
323
Mel Gormana1e78772008-07-23 21:27:23 -0700324 if (vma->vm_flags & VM_SHARED) {
325 /* Shared mappings always use reserves */
Andi Kleena5516432008-07-23 21:27:41 -0700326 h->resv_huge_pages--;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700327 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Mel Gormana1e78772008-07-23 21:27:23 -0700328 /*
329 * Only the process that called mmap() has reserves for
330 * private mappings.
331 */
Andi Kleena5516432008-07-23 21:27:41 -0700332 h->resv_huge_pages--;
Mel Gormana1e78772008-07-23 21:27:23 -0700333 }
334}
335
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700336/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700337void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
338{
339 VM_BUG_ON(!is_vm_hugetlb_page(vma));
340 if (!(vma->vm_flags & VM_SHARED))
341 vma->vm_private_data = (void *)0;
342}
343
344/* Returns true if the VMA has associated reserve pages */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700345static int vma_has_reserves(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700346{
347 if (vma->vm_flags & VM_SHARED)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700348 return 1;
349 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
350 return 1;
351 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700352}
353
Andi Kleena5516432008-07-23 21:27:41 -0700354static void clear_huge_page(struct page *page,
355 unsigned long addr, unsigned long sz)
David Gibson79ac6ba2006-03-22 00:08:51 -0800356{
357 int i;
358
359 might_sleep();
Andi Kleena5516432008-07-23 21:27:41 -0700360 for (i = 0; i < sz/PAGE_SIZE; i++) {
David Gibson79ac6ba2006-03-22 00:08:51 -0800361 cond_resched();
Ralf Baechle281e0e32007-10-01 01:20:10 -0700362 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
David Gibson79ac6ba2006-03-22 00:08:51 -0800363 }
364}
365
366static void copy_huge_page(struct page *dst, struct page *src,
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000367 unsigned long addr, struct vm_area_struct *vma)
David Gibson79ac6ba2006-03-22 00:08:51 -0800368{
369 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700370 struct hstate *h = hstate_vma(vma);
David Gibson79ac6ba2006-03-22 00:08:51 -0800371
372 might_sleep();
Andi Kleena5516432008-07-23 21:27:41 -0700373 for (i = 0; i < pages_per_huge_page(h); i++) {
David Gibson79ac6ba2006-03-22 00:08:51 -0800374 cond_resched();
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000375 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
David Gibson79ac6ba2006-03-22 00:08:51 -0800376 }
377}
378
Andi Kleena5516432008-07-23 21:27:41 -0700379static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380{
381 int nid = page_to_nid(page);
Andi Kleena5516432008-07-23 21:27:41 -0700382 list_add(&page->lru, &h->hugepage_freelists[nid]);
383 h->free_huge_pages++;
384 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385}
386
Andi Kleena5516432008-07-23 21:27:41 -0700387static struct page *dequeue_huge_page(struct hstate *h)
Nishanth Aravamudan348e1e02008-03-04 14:29:42 -0800388{
389 int nid;
390 struct page *page = NULL;
391
392 for (nid = 0; nid < MAX_NUMNODES; ++nid) {
Andi Kleena5516432008-07-23 21:27:41 -0700393 if (!list_empty(&h->hugepage_freelists[nid])) {
394 page = list_entry(h->hugepage_freelists[nid].next,
Nishanth Aravamudan348e1e02008-03-04 14:29:42 -0800395 struct page, lru);
396 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700397 h->free_huge_pages--;
398 h->free_huge_pages_node[nid]--;
Nishanth Aravamudan348e1e02008-03-04 14:29:42 -0800399 break;
400 }
401 }
402 return page;
403}
404
Andi Kleena5516432008-07-23 21:27:41 -0700405static struct page *dequeue_huge_page_vma(struct hstate *h,
406 struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700407 unsigned long address, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408{
Nishanth Aravamudan31a5c6e2007-07-15 23:38:02 -0700409 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700411 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700412 nodemask_t *nodemask;
Mel Gorman396faf02007-07-17 04:03:13 -0700413 struct zonelist *zonelist = huge_zonelist(vma, address,
Mel Gorman19770b32008-04-28 02:12:18 -0700414 htlb_alloc_mask, &mpol, &nodemask);
Mel Gormandd1a2392008-04-28 02:12:17 -0700415 struct zone *zone;
416 struct zoneref *z;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417
Mel Gormana1e78772008-07-23 21:27:23 -0700418 /*
419 * A child process with MAP_PRIVATE mappings created by their parent
420 * have no page reserves. This check ensures that reservations are
421 * not "stolen". The child may still get SIGKILLed
422 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700423 if (!vma_has_reserves(vma) &&
Andi Kleena5516432008-07-23 21:27:41 -0700424 h->free_huge_pages - h->resv_huge_pages == 0)
Mel Gormana1e78772008-07-23 21:27:23 -0700425 return NULL;
426
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700427 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700428 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700429 return NULL;
430
Mel Gorman19770b32008-04-28 02:12:18 -0700431 for_each_zone_zonelist_nodemask(zone, z, zonelist,
432 MAX_NR_ZONES - 1, nodemask) {
Mel Gorman54a6eb52008-04-28 02:12:16 -0700433 nid = zone_to_nid(zone);
434 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
Andi Kleena5516432008-07-23 21:27:41 -0700435 !list_empty(&h->hugepage_freelists[nid])) {
436 page = list_entry(h->hugepage_freelists[nid].next,
Andrew Morton3abf7af2007-07-19 01:49:08 -0700437 struct page, lru);
438 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700439 h->free_huge_pages--;
440 h->free_huge_pages_node[nid]--;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700441
442 if (!avoid_reserve)
Andi Kleena5516432008-07-23 21:27:41 -0700443 decrement_hugepage_resv_vma(h, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700444
Ken Chen5ab3ee72007-07-23 18:44:00 -0700445 break;
Andrew Morton3abf7af2007-07-19 01:49:08 -0700446 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 }
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700448 mpol_cond_put(mpol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 return page;
450}
451
Andi Kleena5516432008-07-23 21:27:41 -0700452static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700453{
454 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700455
456 h->nr_huge_pages--;
457 h->nr_huge_pages_node[page_to_nid(page)]--;
458 for (i = 0; i < pages_per_huge_page(h); i++) {
Adam Litke6af2acb2007-10-16 01:26:16 -0700459 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
460 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
461 1 << PG_private | 1<< PG_writeback);
462 }
463 set_compound_page_dtor(page, NULL);
464 set_page_refcounted(page);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700465 arch_release_hugepage(page);
Andi Kleena5516432008-07-23 21:27:41 -0700466 __free_pages(page, huge_page_order(h));
Adam Litke6af2acb2007-10-16 01:26:16 -0700467}
468
Andi Kleene5ff2152008-07-23 21:27:42 -0700469struct hstate *size_to_hstate(unsigned long size)
470{
471 struct hstate *h;
472
473 for_each_hstate(h) {
474 if (huge_page_size(h) == size)
475 return h;
476 }
477 return NULL;
478}
479
David Gibson27a85ef2006-03-22 00:08:56 -0800480static void free_huge_page(struct page *page)
481{
Andi Kleena5516432008-07-23 21:27:41 -0700482 /*
483 * Can't pass hstate in here because it is called from the
484 * compound page destructor.
485 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700486 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700487 int nid = page_to_nid(page);
Adam Litkec79fb752007-11-14 16:59:38 -0800488 struct address_space *mapping;
David Gibson27a85ef2006-03-22 00:08:56 -0800489
Adam Litkec79fb752007-11-14 16:59:38 -0800490 mapping = (struct address_space *) page_private(page);
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800491 set_page_private(page, 0);
Adam Litke7893d1d2007-10-16 01:26:18 -0700492 BUG_ON(page_count(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800493 INIT_LIST_HEAD(&page->lru);
494
495 spin_lock(&hugetlb_lock);
Andi Kleenaa888a72008-07-23 21:27:47 -0700496 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
Andi Kleena5516432008-07-23 21:27:41 -0700497 update_and_free_page(h, page);
498 h->surplus_huge_pages--;
499 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700500 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700501 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700502 }
David Gibson27a85ef2006-03-22 00:08:56 -0800503 spin_unlock(&hugetlb_lock);
Adam Litkec79fb752007-11-14 16:59:38 -0800504 if (mapping)
Adam Litke9a119c02007-11-14 16:59:41 -0800505 hugetlb_put_quota(mapping, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800506}
507
Adam Litke7893d1d2007-10-16 01:26:18 -0700508/*
509 * Increment or decrement surplus_huge_pages. Keep node-specific counters
510 * balanced by operating on them in a round-robin fashion.
511 * Returns 1 if an adjustment was made.
512 */
Andi Kleena5516432008-07-23 21:27:41 -0700513static int adjust_pool_surplus(struct hstate *h, int delta)
Adam Litke7893d1d2007-10-16 01:26:18 -0700514{
515 static int prev_nid;
516 int nid = prev_nid;
517 int ret = 0;
518
519 VM_BUG_ON(delta != -1 && delta != 1);
520 do {
521 nid = next_node(nid, node_online_map);
522 if (nid == MAX_NUMNODES)
523 nid = first_node(node_online_map);
524
525 /* To shrink on this node, there must be a surplus page */
Andi Kleena5516432008-07-23 21:27:41 -0700526 if (delta < 0 && !h->surplus_huge_pages_node[nid])
Adam Litke7893d1d2007-10-16 01:26:18 -0700527 continue;
528 /* Surplus cannot exceed the total number of pages */
Andi Kleena5516432008-07-23 21:27:41 -0700529 if (delta > 0 && h->surplus_huge_pages_node[nid] >=
530 h->nr_huge_pages_node[nid])
Adam Litke7893d1d2007-10-16 01:26:18 -0700531 continue;
532
Andi Kleena5516432008-07-23 21:27:41 -0700533 h->surplus_huge_pages += delta;
534 h->surplus_huge_pages_node[nid] += delta;
Adam Litke7893d1d2007-10-16 01:26:18 -0700535 ret = 1;
536 break;
537 } while (nid != prev_nid);
538
539 prev_nid = nid;
540 return ret;
541}
542
Andi Kleena5516432008-07-23 21:27:41 -0700543static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700544{
545 set_compound_page_dtor(page, free_huge_page);
546 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700547 h->nr_huge_pages++;
548 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700549 spin_unlock(&hugetlb_lock);
550 put_page(page); /* free it into the hugepage allocator */
551}
552
Andi Kleena5516432008-07-23 21:27:41 -0700553static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700556
Andi Kleenaa888a72008-07-23 21:27:47 -0700557 if (h->order >= MAX_ORDER)
558 return NULL;
559
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700560 page = alloc_pages_node(nid,
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700561 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
562 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700563 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700565 if (arch_prepare_hugepage(page)) {
566 __free_pages(page, HUGETLB_PAGE_ORDER);
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700567 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700568 }
Andi Kleena5516432008-07-23 21:27:41 -0700569 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700571
572 return page;
573}
574
Andi Kleen5ced66c2008-07-23 21:27:45 -0700575/*
576 * Use a helper variable to find the next node and then
577 * copy it back to hugetlb_next_nid afterwards:
578 * otherwise there's a window in which a racer might
579 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
580 * But we don't need to use a spin_lock here: it really
581 * doesn't matter if occasionally a racer chooses the
582 * same nid as we do. Move nid forward in the mask even
583 * if we just successfully allocated a hugepage so that
584 * the next caller gets hugepages on the next node.
585 */
586static int hstate_next_node(struct hstate *h)
587{
588 int next_nid;
589 next_nid = next_node(h->hugetlb_next_nid, node_online_map);
590 if (next_nid == MAX_NUMNODES)
591 next_nid = first_node(node_online_map);
592 h->hugetlb_next_nid = next_nid;
593 return next_nid;
594}
595
Andi Kleena5516432008-07-23 21:27:41 -0700596static int alloc_fresh_huge_page(struct hstate *h)
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700597{
598 struct page *page;
599 int start_nid;
600 int next_nid;
601 int ret = 0;
602
Andi Kleena5516432008-07-23 21:27:41 -0700603 start_nid = h->hugetlb_next_nid;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700604
605 do {
Andi Kleena5516432008-07-23 21:27:41 -0700606 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700607 if (page)
608 ret = 1;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700609 next_nid = hstate_next_node(h);
Andi Kleena5516432008-07-23 21:27:41 -0700610 } while (!page && h->hugetlb_next_nid != start_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700611
Adam Litke3b116302008-04-28 02:13:06 -0700612 if (ret)
613 count_vm_event(HTLB_BUDDY_PGALLOC);
614 else
615 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
616
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700617 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618}
619
Andi Kleena5516432008-07-23 21:27:41 -0700620static struct page *alloc_buddy_huge_page(struct hstate *h,
621 struct vm_area_struct *vma, unsigned long address)
Adam Litke7893d1d2007-10-16 01:26:18 -0700622{
623 struct page *page;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800624 unsigned int nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700625
Andi Kleenaa888a72008-07-23 21:27:47 -0700626 if (h->order >= MAX_ORDER)
627 return NULL;
628
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800629 /*
630 * Assume we will successfully allocate the surplus page to
631 * prevent racing processes from causing the surplus to exceed
632 * overcommit
633 *
634 * This however introduces a different race, where a process B
635 * tries to grow the static hugepage pool while alloc_pages() is
636 * called by process A. B will only examine the per-node
637 * counters in determining if surplus huge pages can be
638 * converted to normal huge pages in adjust_pool_surplus(). A
639 * won't be able to increment the per-node counter, until the
640 * lock is dropped by B, but B doesn't drop hugetlb_lock until
641 * no more huge pages can be converted from surplus to normal
642 * state (and doesn't try to convert again). Thus, we have a
643 * case where a surplus huge page exists, the pool is grown, and
644 * the surplus huge page still exists after, even though it
645 * should just have been converted to a normal huge page. This
646 * does not leak memory, though, as the hugepage will be freed
647 * once it is out of use. It also does not allow the counters to
648 * go out of whack in adjust_pool_surplus() as we don't modify
649 * the node values until we've gotten the hugepage and only the
650 * per-node value is checked there.
651 */
652 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700653 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800654 spin_unlock(&hugetlb_lock);
655 return NULL;
656 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700657 h->nr_huge_pages++;
658 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800659 }
660 spin_unlock(&hugetlb_lock);
661
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700662 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
663 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700664 huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800665
666 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700667 if (page) {
Adam Litke2668db92008-03-10 11:43:50 -0700668 /*
669 * This page is now managed by the hugetlb allocator and has
670 * no users -- drop the buddy allocator's reference.
671 */
672 put_page_testzero(page);
673 VM_BUG_ON(page_count(page));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800674 nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700675 set_compound_page_dtor(page, free_huge_page);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800676 /*
677 * We incremented the global counters already
678 */
Andi Kleena5516432008-07-23 21:27:41 -0700679 h->nr_huge_pages_node[nid]++;
680 h->surplus_huge_pages_node[nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700681 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800682 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700683 h->nr_huge_pages--;
684 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700685 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700686 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800687 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700688
689 return page;
690}
691
Adam Litkee4e574b2007-10-16 01:26:19 -0700692/*
693 * Increase the hugetlb pool such that it can accomodate a reservation
694 * of size 'delta'.
695 */
Andi Kleena5516432008-07-23 21:27:41 -0700696static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -0700697{
698 struct list_head surplus_list;
699 struct page *page, *tmp;
700 int ret, i;
701 int needed, allocated;
702
Andi Kleena5516432008-07-23 21:27:41 -0700703 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800704 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -0700705 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700706 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800707 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700708
709 allocated = 0;
710 INIT_LIST_HEAD(&surplus_list);
711
712 ret = -ENOMEM;
713retry:
714 spin_unlock(&hugetlb_lock);
715 for (i = 0; i < needed; i++) {
Andi Kleena5516432008-07-23 21:27:41 -0700716 page = alloc_buddy_huge_page(h, NULL, 0);
Adam Litkee4e574b2007-10-16 01:26:19 -0700717 if (!page) {
718 /*
719 * We were not able to allocate enough pages to
720 * satisfy the entire reservation so we free what
721 * we've allocated so far.
722 */
723 spin_lock(&hugetlb_lock);
724 needed = 0;
725 goto free;
726 }
727
728 list_add(&page->lru, &surplus_list);
729 }
730 allocated += needed;
731
732 /*
733 * After retaking hugetlb_lock, we need to recalculate 'needed'
734 * because either resv_huge_pages or free_huge_pages may have changed.
735 */
736 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700737 needed = (h->resv_huge_pages + delta) -
738 (h->free_huge_pages + allocated);
Adam Litkee4e574b2007-10-16 01:26:19 -0700739 if (needed > 0)
740 goto retry;
741
742 /*
743 * The surplus_list now contains _at_least_ the number of extra pages
744 * needed to accomodate the reservation. Add the appropriate number
745 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -0800746 * allocator. Commit the entire reservation here to prevent another
747 * process from stealing the pages as they are added to the pool but
748 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -0700749 */
750 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -0700751 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700752 ret = 0;
753free:
Adam Litke19fc3f02008-04-28 02:12:20 -0700754 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -0700755 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -0700756 if ((--needed) < 0)
757 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700758 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700759 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -0700760 }
761
762 /* Free unnecessary surplus pages to the buddy allocator */
763 if (!list_empty(&surplus_list)) {
764 spin_unlock(&hugetlb_lock);
765 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
766 list_del(&page->lru);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700767 /*
Adam Litke2668db92008-03-10 11:43:50 -0700768 * The page has a reference count of zero already, so
769 * call free_huge_page directly instead of using
770 * put_page. This must be done with hugetlb_lock
Adam Litkeaf767cb2007-10-16 01:26:25 -0700771 * unlocked which is safe because free_huge_page takes
772 * hugetlb_lock before deciding how to free the page.
773 */
Adam Litke2668db92008-03-10 11:43:50 -0700774 free_huge_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700775 }
Adam Litke19fc3f02008-04-28 02:12:20 -0700776 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -0700777 }
778
779 return ret;
780}
781
782/*
783 * When releasing a hugetlb pool reservation, any surplus pages that were
784 * allocated to satisfy the reservation must be explicitly freed if they were
785 * never used.
786 */
Andi Kleena5516432008-07-23 21:27:41 -0700787static void return_unused_surplus_pages(struct hstate *h,
788 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -0700789{
790 static int nid = -1;
791 struct page *page;
792 unsigned long nr_pages;
793
Nishanth Aravamudan11320d12008-03-26 14:40:20 -0700794 /*
795 * We want to release as many surplus pages as possible, spread
796 * evenly across all nodes. Iterate across all nodes until we
797 * can no longer free unreserved surplus pages. This occurs when
798 * the nodes with surplus pages have no free pages.
799 */
800 unsigned long remaining_iterations = num_online_nodes();
801
Adam Litkeac09b3a2008-03-04 14:29:38 -0800802 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -0700803 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800804
Andi Kleenaa888a72008-07-23 21:27:47 -0700805 /* Cannot return gigantic pages currently */
806 if (h->order >= MAX_ORDER)
807 return;
808
Andi Kleena5516432008-07-23 21:27:41 -0700809 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -0700810
Nishanth Aravamudan11320d12008-03-26 14:40:20 -0700811 while (remaining_iterations-- && nr_pages) {
Adam Litkee4e574b2007-10-16 01:26:19 -0700812 nid = next_node(nid, node_online_map);
813 if (nid == MAX_NUMNODES)
814 nid = first_node(node_online_map);
815
Andi Kleena5516432008-07-23 21:27:41 -0700816 if (!h->surplus_huge_pages_node[nid])
Adam Litkee4e574b2007-10-16 01:26:19 -0700817 continue;
818
Andi Kleena5516432008-07-23 21:27:41 -0700819 if (!list_empty(&h->hugepage_freelists[nid])) {
820 page = list_entry(h->hugepage_freelists[nid].next,
Adam Litkee4e574b2007-10-16 01:26:19 -0700821 struct page, lru);
822 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700823 update_and_free_page(h, page);
824 h->free_huge_pages--;
825 h->free_huge_pages_node[nid]--;
826 h->surplus_huge_pages--;
827 h->surplus_huge_pages_node[nid]--;
Adam Litkee4e574b2007-10-16 01:26:19 -0700828 nr_pages--;
Nishanth Aravamudan11320d12008-03-26 14:40:20 -0700829 remaining_iterations = num_online_nodes();
Adam Litkee4e574b2007-10-16 01:26:19 -0700830 }
831 }
832}
833
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700834/*
835 * Determine if the huge page at addr within the vma has an associated
836 * reservation. Where it does not we will need to logically increase
837 * reservation and actually increase quota before an allocation can occur.
838 * Where any new reservation would be required the reservation change is
839 * prepared, but not committed. Once the page has been quota'd allocated
840 * an instantiated the change should be committed via vma_commit_reservation.
841 * No action is required on failure.
842 */
Andi Kleena5516432008-07-23 21:27:41 -0700843static int vma_needs_reservation(struct hstate *h,
844 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700845{
846 struct address_space *mapping = vma->vm_file->f_mapping;
847 struct inode *inode = mapping->host;
848
849 if (vma->vm_flags & VM_SHARED) {
Andi Kleena5516432008-07-23 21:27:41 -0700850 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700851 return region_chg(&inode->i_mapping->private_list,
852 idx, idx + 1);
853
Andy Whitcroft84afd992008-07-23 21:27:32 -0700854 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
855 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700856
Andy Whitcroft84afd992008-07-23 21:27:32 -0700857 } else {
858 int err;
Andi Kleena5516432008-07-23 21:27:41 -0700859 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700860 struct resv_map *reservations = vma_resv_map(vma);
861
862 err = region_chg(&reservations->regions, idx, idx + 1);
863 if (err < 0)
864 return err;
865 return 0;
866 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700867}
Andi Kleena5516432008-07-23 21:27:41 -0700868static void vma_commit_reservation(struct hstate *h,
869 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700870{
871 struct address_space *mapping = vma->vm_file->f_mapping;
872 struct inode *inode = mapping->host;
873
874 if (vma->vm_flags & VM_SHARED) {
Andi Kleena5516432008-07-23 21:27:41 -0700875 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700876 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700877
878 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -0700879 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700880 struct resv_map *reservations = vma_resv_map(vma);
881
882 /* Mark this page used in the map. */
883 region_add(&reservations->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700884 }
885}
886
David Gibson27a85ef2006-03-22 00:08:56 -0800887static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700888 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889{
Andi Kleena5516432008-07-23 21:27:41 -0700890 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -0800891 struct page *page;
Adam Litke2fc39ce2007-11-14 16:59:39 -0800892 struct address_space *mapping = vma->vm_file->f_mapping;
Mel Gormana1e78772008-07-23 21:27:23 -0700893 struct inode *inode = mapping->host;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700894 unsigned int chg;
Adam Litke2fc39ce2007-11-14 16:59:39 -0800895
Mel Gormana1e78772008-07-23 21:27:23 -0700896 /*
897 * Processes that did not create the mapping will have no reserves and
898 * will not have accounted against quota. Check that the quota can be
899 * made before satisfying the allocation
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700900 * MAP_NORESERVE mappings may also need pages and quota allocated
901 * if no reserve mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -0700902 */
Andi Kleena5516432008-07-23 21:27:41 -0700903 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700904 if (chg < 0)
905 return ERR_PTR(chg);
906 if (chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700907 if (hugetlb_get_quota(inode->i_mapping, chg))
908 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -0700909
910 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700911 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
Mel Gormana1e78772008-07-23 21:27:23 -0700912 spin_unlock(&hugetlb_lock);
913
914 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -0700915 page = alloc_buddy_huge_page(h, vma, addr);
Mel Gormana1e78772008-07-23 21:27:23 -0700916 if (!page) {
917 hugetlb_put_quota(inode->i_mapping, chg);
918 return ERR_PTR(-VM_FAULT_OOM);
919 }
920 }
921
922 set_page_refcounted(page);
923 set_page_private(page, (unsigned long) mapping);
924
Andi Kleena5516432008-07-23 21:27:41 -0700925 vma_commit_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700926
Adam Litke90d8b7e2007-11-14 16:59:42 -0800927 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800928}
929
Jon Tollefson53ba51d2008-07-23 21:27:52 -0700930__attribute__((weak)) int alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -0700931{
932 struct huge_bootmem_page *m;
933 int nr_nodes = nodes_weight(node_online_map);
934
935 while (nr_nodes) {
936 void *addr;
937
938 addr = __alloc_bootmem_node_nopanic(
939 NODE_DATA(h->hugetlb_next_nid),
940 huge_page_size(h), huge_page_size(h), 0);
941
942 if (addr) {
943 /*
944 * Use the beginning of the huge page to store the
945 * huge_bootmem_page struct (until gather_bootmem
946 * puts them into the mem_map).
947 */
948 m = addr;
949 if (m)
950 goto found;
951 }
952 hstate_next_node(h);
953 nr_nodes--;
954 }
955 return 0;
956
957found:
958 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
959 /* Put them into a private list first because mem_map is not up yet */
960 list_add(&m->list, &huge_boot_pages);
961 m->hstate = h;
962 return 1;
963}
964
965/* Put bootmem huge pages into the standard lists after mem_map is up */
966static void __init gather_bootmem_prealloc(void)
967{
968 struct huge_bootmem_page *m;
969
970 list_for_each_entry(m, &huge_boot_pages, list) {
971 struct page *page = virt_to_page(m);
972 struct hstate *h = m->hstate;
973 __ClearPageReserved(page);
974 WARN_ON(page_count(page) != 1);
975 prep_compound_page(page, h->order);
976 prep_new_huge_page(h, page, page_to_nid(page));
977 }
978}
979
Andi Kleen8faa8b02008-07-23 21:27:48 -0700980static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981{
982 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983
Andi Kleene5ff2152008-07-23 21:27:42 -0700984 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -0700985 if (h->order >= MAX_ORDER) {
986 if (!alloc_bootmem_huge_page(h))
987 break;
988 } else if (!alloc_fresh_huge_page(h))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990 }
Andi Kleen8faa8b02008-07-23 21:27:48 -0700991 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -0700992}
993
994static void __init hugetlb_init_hstates(void)
995{
996 struct hstate *h;
997
998 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -0700999 /* oversize hugepages were init'ed in early boot */
1000 if (h->order < MAX_ORDER)
1001 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001002 }
1003}
1004
Andi Kleen4abd32d2008-07-23 21:27:49 -07001005static char * __init memfmt(char *buf, unsigned long n)
1006{
1007 if (n >= (1UL << 30))
1008 sprintf(buf, "%lu GB", n >> 30);
1009 else if (n >= (1UL << 20))
1010 sprintf(buf, "%lu MB", n >> 20);
1011 else
1012 sprintf(buf, "%lu KB", n >> 10);
1013 return buf;
1014}
1015
Andi Kleene5ff2152008-07-23 21:27:42 -07001016static void __init report_hugepages(void)
1017{
1018 struct hstate *h;
1019
1020 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001021 char buf[32];
1022 printk(KERN_INFO "HugeTLB registered %s page size, "
1023 "pre-allocated %ld pages\n",
1024 memfmt(buf, huge_page_size(h)),
1025 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001026 }
1027}
1028
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029#ifdef CONFIG_HIGHMEM
Andi Kleena5516432008-07-23 21:27:41 -07001030static void try_to_free_low(struct hstate *h, unsigned long count)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001032 int i;
1033
Andi Kleenaa888a72008-07-23 21:27:47 -07001034 if (h->order >= MAX_ORDER)
1035 return;
1036
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037 for (i = 0; i < MAX_NUMNODES; ++i) {
1038 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001039 struct list_head *freel = &h->hugepage_freelists[i];
1040 list_for_each_entry_safe(page, next, freel, lru) {
1041 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001042 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 if (PageHighMem(page))
1044 continue;
1045 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001046 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001047 h->free_huge_pages--;
1048 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049 }
1050 }
1051}
1052#else
Andi Kleena5516432008-07-23 21:27:41 -07001053static inline void try_to_free_low(struct hstate *h, unsigned long count)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054{
1055}
1056#endif
1057
Andi Kleena5516432008-07-23 21:27:41 -07001058#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Andi Kleene5ff2152008-07-23 21:27:42 -07001059static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060{
Adam Litke7893d1d2007-10-16 01:26:18 -07001061 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062
Andi Kleenaa888a72008-07-23 21:27:47 -07001063 if (h->order >= MAX_ORDER)
1064 return h->max_huge_pages;
1065
Adam Litke7893d1d2007-10-16 01:26:18 -07001066 /*
1067 * Increase the pool size
1068 * First take pages out of surplus state. Then make up the
1069 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001070 *
1071 * We might race with alloc_buddy_huge_page() here and be unable
1072 * to convert a surplus huge page to a normal huge page. That is
1073 * not critical, though, it just means the overall size of the
1074 * pool might be one hugepage larger than it needs to be, but
1075 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001076 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001078 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1079 if (!adjust_pool_surplus(h, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001080 break;
1081 }
1082
Andi Kleena5516432008-07-23 21:27:41 -07001083 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001084 /*
1085 * If this allocation races such that we no longer need the
1086 * page, free_huge_page will handle it by freeing the page
1087 * and reducing the surplus.
1088 */
1089 spin_unlock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001090 ret = alloc_fresh_huge_page(h);
Adam Litke7893d1d2007-10-16 01:26:18 -07001091 spin_lock(&hugetlb_lock);
1092 if (!ret)
1093 goto out;
1094
1095 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001096
1097 /*
1098 * Decrease the pool size
1099 * First return free pages to the buddy allocator (being careful
1100 * to keep enough around to satisfy reservations). Then place
1101 * pages into surplus state as needed so the pool will shrink
1102 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001103 *
1104 * By placing pages into the surplus state independent of the
1105 * overcommit value, we are allowing the surplus pool size to
1106 * exceed overcommit. There are few sane options here. Since
1107 * alloc_buddy_huge_page() is checking the global counter,
1108 * though, we'll note that we're not allowed to exceed surplus
1109 * and won't grow the pool anywhere else. Not until one of the
1110 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001111 */
Andi Kleena5516432008-07-23 21:27:41 -07001112 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001113 min_count = max(count, min_count);
Andi Kleena5516432008-07-23 21:27:41 -07001114 try_to_free_low(h, min_count);
1115 while (min_count < persistent_huge_pages(h)) {
1116 struct page *page = dequeue_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117 if (!page)
1118 break;
Andi Kleena5516432008-07-23 21:27:41 -07001119 update_and_free_page(h, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 }
Andi Kleena5516432008-07-23 21:27:41 -07001121 while (count < persistent_huge_pages(h)) {
1122 if (!adjust_pool_surplus(h, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001123 break;
1124 }
1125out:
Andi Kleena5516432008-07-23 21:27:41 -07001126 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001128 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129}
1130
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001131#define HSTATE_ATTR_RO(_name) \
1132 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1133
1134#define HSTATE_ATTR(_name) \
1135 static struct kobj_attribute _name##_attr = \
1136 __ATTR(_name, 0644, _name##_show, _name##_store)
1137
1138static struct kobject *hugepages_kobj;
1139static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1140
1141static struct hstate *kobj_to_hstate(struct kobject *kobj)
1142{
1143 int i;
1144 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1145 if (hstate_kobjs[i] == kobj)
1146 return &hstates[i];
1147 BUG();
1148 return NULL;
1149}
1150
1151static ssize_t nr_hugepages_show(struct kobject *kobj,
1152 struct kobj_attribute *attr, char *buf)
1153{
1154 struct hstate *h = kobj_to_hstate(kobj);
1155 return sprintf(buf, "%lu\n", h->nr_huge_pages);
1156}
1157static ssize_t nr_hugepages_store(struct kobject *kobj,
1158 struct kobj_attribute *attr, const char *buf, size_t count)
1159{
1160 int err;
1161 unsigned long input;
1162 struct hstate *h = kobj_to_hstate(kobj);
1163
1164 err = strict_strtoul(buf, 10, &input);
1165 if (err)
1166 return 0;
1167
1168 h->max_huge_pages = set_max_huge_pages(h, input);
1169
1170 return count;
1171}
1172HSTATE_ATTR(nr_hugepages);
1173
1174static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1175 struct kobj_attribute *attr, char *buf)
1176{
1177 struct hstate *h = kobj_to_hstate(kobj);
1178 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1179}
1180static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1181 struct kobj_attribute *attr, const char *buf, size_t count)
1182{
1183 int err;
1184 unsigned long input;
1185 struct hstate *h = kobj_to_hstate(kobj);
1186
1187 err = strict_strtoul(buf, 10, &input);
1188 if (err)
1189 return 0;
1190
1191 spin_lock(&hugetlb_lock);
1192 h->nr_overcommit_huge_pages = input;
1193 spin_unlock(&hugetlb_lock);
1194
1195 return count;
1196}
1197HSTATE_ATTR(nr_overcommit_hugepages);
1198
1199static ssize_t free_hugepages_show(struct kobject *kobj,
1200 struct kobj_attribute *attr, char *buf)
1201{
1202 struct hstate *h = kobj_to_hstate(kobj);
1203 return sprintf(buf, "%lu\n", h->free_huge_pages);
1204}
1205HSTATE_ATTR_RO(free_hugepages);
1206
1207static ssize_t resv_hugepages_show(struct kobject *kobj,
1208 struct kobj_attribute *attr, char *buf)
1209{
1210 struct hstate *h = kobj_to_hstate(kobj);
1211 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1212}
1213HSTATE_ATTR_RO(resv_hugepages);
1214
1215static ssize_t surplus_hugepages_show(struct kobject *kobj,
1216 struct kobj_attribute *attr, char *buf)
1217{
1218 struct hstate *h = kobj_to_hstate(kobj);
1219 return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1220}
1221HSTATE_ATTR_RO(surplus_hugepages);
1222
1223static struct attribute *hstate_attrs[] = {
1224 &nr_hugepages_attr.attr,
1225 &nr_overcommit_hugepages_attr.attr,
1226 &free_hugepages_attr.attr,
1227 &resv_hugepages_attr.attr,
1228 &surplus_hugepages_attr.attr,
1229 NULL,
1230};
1231
1232static struct attribute_group hstate_attr_group = {
1233 .attrs = hstate_attrs,
1234};
1235
1236static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1237{
1238 int retval;
1239
1240 hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1241 hugepages_kobj);
1242 if (!hstate_kobjs[h - hstates])
1243 return -ENOMEM;
1244
1245 retval = sysfs_create_group(hstate_kobjs[h - hstates],
1246 &hstate_attr_group);
1247 if (retval)
1248 kobject_put(hstate_kobjs[h - hstates]);
1249
1250 return retval;
1251}
1252
1253static void __init hugetlb_sysfs_init(void)
1254{
1255 struct hstate *h;
1256 int err;
1257
1258 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1259 if (!hugepages_kobj)
1260 return;
1261
1262 for_each_hstate(h) {
1263 err = hugetlb_sysfs_add_hstate(h);
1264 if (err)
1265 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1266 h->name);
1267 }
1268}
1269
1270static void __exit hugetlb_exit(void)
1271{
1272 struct hstate *h;
1273
1274 for_each_hstate(h) {
1275 kobject_put(hstate_kobjs[h - hstates]);
1276 }
1277
1278 kobject_put(hugepages_kobj);
1279}
1280module_exit(hugetlb_exit);
1281
1282static int __init hugetlb_init(void)
1283{
1284 BUILD_BUG_ON(HPAGE_SHIFT == 0);
1285
Nick Piggine11bfbf2008-07-23 21:27:52 -07001286 if (!size_to_hstate(default_hstate_size)) {
1287 default_hstate_size = HPAGE_SIZE;
1288 if (!size_to_hstate(default_hstate_size))
1289 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001290 }
Nick Piggine11bfbf2008-07-23 21:27:52 -07001291 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1292 if (default_hstate_max_huge_pages)
1293 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001294
1295 hugetlb_init_hstates();
1296
Andi Kleenaa888a72008-07-23 21:27:47 -07001297 gather_bootmem_prealloc();
1298
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001299 report_hugepages();
1300
1301 hugetlb_sysfs_init();
1302
1303 return 0;
1304}
1305module_init(hugetlb_init);
1306
1307/* Should be called on processing a hugepagesz=... option */
1308void __init hugetlb_add_hstate(unsigned order)
1309{
1310 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001311 unsigned long i;
1312
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001313 if (size_to_hstate(PAGE_SIZE << order)) {
1314 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1315 return;
1316 }
1317 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1318 BUG_ON(order == 0);
1319 h = &hstates[max_hstate++];
1320 h->order = order;
1321 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001322 h->nr_huge_pages = 0;
1323 h->free_huge_pages = 0;
1324 for (i = 0; i < MAX_NUMNODES; ++i)
1325 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1326 h->hugetlb_next_nid = first_node(node_online_map);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001327 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1328 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001329
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001330 parsed_hstate = h;
1331}
1332
Nick Piggine11bfbf2008-07-23 21:27:52 -07001333static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001334{
1335 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001336 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001337
1338 /*
1339 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1340 * so this hugepages= parameter goes to the "default hstate".
1341 */
1342 if (!max_hstate)
1343 mhp = &default_hstate_max_huge_pages;
1344 else
1345 mhp = &parsed_hstate->max_huge_pages;
1346
Andi Kleen8faa8b02008-07-23 21:27:48 -07001347 if (mhp == last_mhp) {
1348 printk(KERN_WARNING "hugepages= specified twice without "
1349 "interleaving hugepagesz=, ignoring\n");
1350 return 1;
1351 }
1352
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001353 if (sscanf(s, "%lu", mhp) <= 0)
1354 *mhp = 0;
1355
Andi Kleen8faa8b02008-07-23 21:27:48 -07001356 /*
1357 * Global state is always initialized later in hugetlb_init.
1358 * But we need to allocate >= MAX_ORDER hstates here early to still
1359 * use the bootmem allocator.
1360 */
1361 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1362 hugetlb_hstate_alloc_pages(parsed_hstate);
1363
1364 last_mhp = mhp;
1365
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001366 return 1;
1367}
Nick Piggine11bfbf2008-07-23 21:27:52 -07001368__setup("hugepages=", hugetlb_nrpages_setup);
1369
1370static int __init hugetlb_default_setup(char *s)
1371{
1372 default_hstate_size = memparse(s, &s);
1373 return 1;
1374}
1375__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001376
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07001377static unsigned int cpuset_mems_nr(unsigned int *array)
1378{
1379 int node;
1380 unsigned int nr = 0;
1381
1382 for_each_node_mask(node, cpuset_current_mems_allowed)
1383 nr += array[node];
1384
1385 return nr;
1386}
1387
1388#ifdef CONFIG_SYSCTL
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1390 struct file *file, void __user *buffer,
1391 size_t *length, loff_t *ppos)
1392{
Andi Kleene5ff2152008-07-23 21:27:42 -07001393 struct hstate *h = &default_hstate;
1394 unsigned long tmp;
1395
1396 if (!write)
1397 tmp = h->max_huge_pages;
1398
1399 table->data = &tmp;
1400 table->maxlen = sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001402
1403 if (write)
1404 h->max_huge_pages = set_max_huge_pages(h, tmp);
1405
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 return 0;
1407}
Mel Gorman396faf02007-07-17 04:03:13 -07001408
1409int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1410 struct file *file, void __user *buffer,
1411 size_t *length, loff_t *ppos)
1412{
1413 proc_dointvec(table, write, file, buffer, length, ppos);
1414 if (hugepages_treat_as_movable)
1415 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1416 else
1417 htlb_alloc_mask = GFP_HIGHUSER;
1418 return 0;
1419}
1420
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001421int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1422 struct file *file, void __user *buffer,
1423 size_t *length, loff_t *ppos)
1424{
Andi Kleena5516432008-07-23 21:27:41 -07001425 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07001426 unsigned long tmp;
1427
1428 if (!write)
1429 tmp = h->nr_overcommit_huge_pages;
1430
1431 table->data = &tmp;
1432 table->maxlen = sizeof(unsigned long);
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001433 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001434
1435 if (write) {
1436 spin_lock(&hugetlb_lock);
1437 h->nr_overcommit_huge_pages = tmp;
1438 spin_unlock(&hugetlb_lock);
1439 }
1440
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001441 return 0;
1442}
1443
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444#endif /* CONFIG_SYSCTL */
1445
1446int hugetlb_report_meminfo(char *buf)
1447{
Andi Kleena5516432008-07-23 21:27:41 -07001448 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449 return sprintf(buf,
1450 "HugePages_Total: %5lu\n"
1451 "HugePages_Free: %5lu\n"
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07001452 "HugePages_Rsvd: %5lu\n"
Adam Litke7893d1d2007-10-16 01:26:18 -07001453 "HugePages_Surp: %5lu\n"
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454 "Hugepagesize: %5lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07001455 h->nr_huge_pages,
1456 h->free_huge_pages,
1457 h->resv_huge_pages,
1458 h->surplus_huge_pages,
1459 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460}
1461
1462int hugetlb_report_node_meminfo(int nid, char *buf)
1463{
Andi Kleena5516432008-07-23 21:27:41 -07001464 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 return sprintf(buf,
1466 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07001467 "Node %d HugePages_Free: %5u\n"
1468 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07001469 nid, h->nr_huge_pages_node[nid],
1470 nid, h->free_huge_pages_node[nid],
1471 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472}
1473
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1475unsigned long hugetlb_total_pages(void)
1476{
Andi Kleena5516432008-07-23 21:27:41 -07001477 struct hstate *h = &default_hstate;
1478 return h->nr_huge_pages * pages_per_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480
Andi Kleena5516432008-07-23 21:27:41 -07001481static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001482{
1483 int ret = -ENOMEM;
1484
1485 spin_lock(&hugetlb_lock);
1486 /*
1487 * When cpuset is configured, it breaks the strict hugetlb page
1488 * reservation as the accounting is done on a global variable. Such
1489 * reservation is completely rubbish in the presence of cpuset because
1490 * the reservation is not checked against page availability for the
1491 * current cpuset. Application can still potentially OOM'ed by kernel
1492 * with lack of free htlb page in cpuset that the task is in.
1493 * Attempt to enforce strict accounting with cpuset is almost
1494 * impossible (or too ugly) because cpuset is too fluid that
1495 * task or memory node can be dynamically moved between cpusets.
1496 *
1497 * The change of semantics for shared hugetlb mapping with cpuset is
1498 * undesirable. However, in order to preserve some of the semantics,
1499 * we fall back to check against current free page availability as
1500 * a best attempt and hopefully to minimize the impact of changing
1501 * semantics that cpuset has.
1502 */
1503 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001504 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001505 goto out;
1506
Andi Kleena5516432008-07-23 21:27:41 -07001507 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1508 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001509 goto out;
1510 }
1511 }
1512
1513 ret = 0;
1514 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07001515 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001516
1517out:
1518 spin_unlock(&hugetlb_lock);
1519 return ret;
1520}
1521
Andy Whitcroft84afd992008-07-23 21:27:32 -07001522static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1523{
1524 struct resv_map *reservations = vma_resv_map(vma);
1525
1526 /*
1527 * This new VMA should share its siblings reservation map if present.
1528 * The VMA will only ever have a valid reservation map pointer where
1529 * it is being copied for another still existing VMA. As that VMA
1530 * has a reference to the reservation map it cannot dissappear until
1531 * after this open call completes. It is therefore safe to take a
1532 * new reference here without additional locking.
1533 */
1534 if (reservations)
1535 kref_get(&reservations->refs);
1536}
1537
Mel Gormana1e78772008-07-23 21:27:23 -07001538static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1539{
Andi Kleena5516432008-07-23 21:27:41 -07001540 struct hstate *h = hstate_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001541 struct resv_map *reservations = vma_resv_map(vma);
1542 unsigned long reserve;
1543 unsigned long start;
1544 unsigned long end;
1545
1546 if (reservations) {
Andi Kleena5516432008-07-23 21:27:41 -07001547 start = vma_hugecache_offset(h, vma, vma->vm_start);
1548 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001549
1550 reserve = (end - start) -
1551 region_count(&reservations->regions, start, end);
1552
1553 kref_put(&reservations->refs, resv_map_release);
1554
Adam Litke7251ff72008-07-23 21:27:59 -07001555 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07001556 hugetlb_acct_memory(h, -reserve);
Adam Litke7251ff72008-07-23 21:27:59 -07001557 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1558 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07001559 }
Mel Gormana1e78772008-07-23 21:27:23 -07001560}
1561
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562/*
1563 * We cannot handle pagefaults against hugetlb pages at all. They cause
1564 * handle_mm_fault() to try to instantiate regular-sized pages in the
1565 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1566 * this far.
1567 */
Nick Piggind0217ac2007-07-19 01:47:03 -07001568static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569{
1570 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07001571 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572}
1573
1574struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07001575 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07001576 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07001577 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578};
1579
David Gibson1e8f8892006-01-06 00:10:44 -08001580static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1581 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07001582{
1583 pte_t entry;
1584
David Gibson1e8f8892006-01-06 00:10:44 -08001585 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -07001586 entry =
1587 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1588 } else {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001589 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07001590 }
1591 entry = pte_mkyoung(entry);
1592 entry = pte_mkhuge(entry);
1593
1594 return entry;
1595}
1596
David Gibson1e8f8892006-01-06 00:10:44 -08001597static void set_huge_ptep_writable(struct vm_area_struct *vma,
1598 unsigned long address, pte_t *ptep)
1599{
1600 pte_t entry;
1601
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001602 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1603 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07001604 update_mmu_cache(vma, address, entry);
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07001605 }
David Gibson1e8f8892006-01-06 00:10:44 -08001606}
1607
1608
David Gibson63551ae2005-06-21 17:14:44 -07001609int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1610 struct vm_area_struct *vma)
1611{
1612 pte_t *src_pte, *dst_pte, entry;
1613 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07001614 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08001615 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07001616 struct hstate *h = hstate_vma(vma);
1617 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08001618
1619 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07001620
Andi Kleena5516432008-07-23 21:27:41 -07001621 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Hugh Dickinsc74df322005-10-29 18:16:23 -07001622 src_pte = huge_pte_offset(src, addr);
1623 if (!src_pte)
1624 continue;
Andi Kleena5516432008-07-23 21:27:41 -07001625 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07001626 if (!dst_pte)
1627 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08001628
1629 /* If the pagetables are shared don't copy or take references */
1630 if (dst_pte == src_pte)
1631 continue;
1632
Hugh Dickinsc74df322005-10-29 18:16:23 -07001633 spin_lock(&dst->page_table_lock);
Nick Piggin46478752008-06-05 22:45:57 -07001634 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001635 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08001636 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001637 huge_ptep_set_wrprotect(src, addr, src_pte);
1638 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07001639 ptepage = pte_page(entry);
1640 get_page(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07001641 set_huge_pte_at(dst, addr, dst_pte, entry);
1642 }
1643 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -07001644 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07001645 }
1646 return 0;
1647
1648nomem:
1649 return -ENOMEM;
1650}
1651
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001652void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001653 unsigned long end, struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07001654{
1655 struct mm_struct *mm = vma->vm_mm;
1656 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07001657 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07001658 pte_t pte;
1659 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001660 struct page *tmp;
Andi Kleena5516432008-07-23 21:27:41 -07001661 struct hstate *h = hstate_vma(vma);
1662 unsigned long sz = huge_page_size(h);
1663
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -08001664 /*
1665 * A page gathering list, protected by per file i_mmap_lock. The
1666 * lock is used to avoid list corruption from multiple unmapping
1667 * of the same page since we are using page->lru.
1668 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001669 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -07001670
1671 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07001672 BUG_ON(start & ~huge_page_mask(h));
1673 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07001674
Hugh Dickins508034a2005-10-29 18:16:30 -07001675 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001676 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07001677 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07001678 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07001679 continue;
1680
Chen, Kenneth W39dde652006-12-06 20:32:03 -08001681 if (huge_pmd_unshare(mm, &address, ptep))
1682 continue;
1683
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001684 /*
1685 * If a reference page is supplied, it is because a specific
1686 * page is being unmapped, not a range. Ensure the page we
1687 * are about to unmap is the actual page of interest.
1688 */
1689 if (ref_page) {
1690 pte = huge_ptep_get(ptep);
1691 if (huge_pte_none(pte))
1692 continue;
1693 page = pte_page(pte);
1694 if (page != ref_page)
1695 continue;
1696
1697 /*
1698 * Mark the VMA as having unmapped its page so that
1699 * future faults in this VMA will fail rather than
1700 * looking like data was lost
1701 */
1702 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1703 }
1704
David Gibsonc7546f82005-08-05 11:59:35 -07001705 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001706 if (huge_pte_none(pte))
David Gibson63551ae2005-06-21 17:14:44 -07001707 continue;
David Gibsonc7546f82005-08-05 11:59:35 -07001708
David Gibson63551ae2005-06-21 17:14:44 -07001709 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -08001710 if (pte_dirty(pte))
1711 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001712 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -07001713 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -07001715 flush_tlb_range(vma, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001716 list_for_each_entry_safe(page, tmp, &page_list, lru) {
1717 list_del(&page->lru);
1718 put_page(page);
1719 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720}
David Gibson63551ae2005-06-21 17:14:44 -07001721
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001722void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001723 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001724{
Andi Kleena137e1c2008-07-23 21:27:43 -07001725 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1726 __unmap_hugepage_range(vma, start, end, ref_page);
1727 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001728}
1729
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001730/*
1731 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1732 * mappping it owns the reserve page for. The intention is to unmap the page
1733 * from other VMAs and let the children be SIGKILLed if they are faulting the
1734 * same region.
1735 */
1736int unmap_ref_private(struct mm_struct *mm,
1737 struct vm_area_struct *vma,
1738 struct page *page,
1739 unsigned long address)
1740{
1741 struct vm_area_struct *iter_vma;
1742 struct address_space *mapping;
1743 struct prio_tree_iter iter;
1744 pgoff_t pgoff;
1745
1746 /*
1747 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1748 * from page cache lookup which is in HPAGE_SIZE units.
1749 */
1750 address = address & huge_page_mask(hstate_vma(vma));
1751 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1752 + (vma->vm_pgoff >> PAGE_SHIFT);
1753 mapping = (struct address_space *)page_private(page);
1754
1755 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1756 /* Do not unmap the current VMA */
1757 if (iter_vma == vma)
1758 continue;
1759
1760 /*
1761 * Unmap the page from other VMAs without their own reserves.
1762 * They get marked to be SIGKILLed if they fault in these
1763 * areas. This is because a future no-page fault on this VMA
1764 * could insert a zeroed page instead of the data existing
1765 * from the time of fork. This would look like data corruption
1766 */
1767 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1768 unmap_hugepage_range(iter_vma,
1769 address, address + HPAGE_SIZE,
1770 page);
1771 }
1772
1773 return 1;
1774}
1775
David Gibson1e8f8892006-01-06 00:10:44 -08001776static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001777 unsigned long address, pte_t *ptep, pte_t pte,
1778 struct page *pagecache_page)
David Gibson1e8f8892006-01-06 00:10:44 -08001779{
Andi Kleena5516432008-07-23 21:27:41 -07001780 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08001781 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -08001782 int avoidcopy;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001783 int outside_reserve = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08001784
1785 old_page = pte_page(pte);
1786
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001787retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08001788 /* If no-one else is actually using this page, avoid the copy
1789 * and just make the page writable */
1790 avoidcopy = (page_count(old_page) == 1);
1791 if (avoidcopy) {
1792 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07001793 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08001794 }
1795
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001796 /*
1797 * If the process that created a MAP_PRIVATE mapping is about to
1798 * perform a COW due to a shared page count, attempt to satisfy
1799 * the allocation without using the existing reserves. The pagecache
1800 * page is used to determine if the reserve at this address was
1801 * consumed or not. If reserves were used, a partial faulted mapping
1802 * at the time of fork() could consume its reserves on COW instead
1803 * of the full address range.
1804 */
1805 if (!(vma->vm_flags & VM_SHARED) &&
1806 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1807 old_page != pagecache_page)
1808 outside_reserve = 1;
1809
David Gibson1e8f8892006-01-06 00:10:44 -08001810 page_cache_get(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001811 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08001812
Adam Litke2fc39ce2007-11-14 16:59:39 -08001813 if (IS_ERR(new_page)) {
David Gibson1e8f8892006-01-06 00:10:44 -08001814 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001815
1816 /*
1817 * If a process owning a MAP_PRIVATE mapping fails to COW,
1818 * it is due to references held by a child and an insufficient
1819 * huge page pool. To guarantee the original mappers
1820 * reliability, unmap the page from child processes. The child
1821 * may get SIGKILLed if it later faults.
1822 */
1823 if (outside_reserve) {
1824 BUG_ON(huge_pte_none(pte));
1825 if (unmap_ref_private(mm, vma, old_page, address)) {
1826 BUG_ON(page_count(old_page) != 1);
1827 BUG_ON(huge_pte_none(pte));
1828 goto retry_avoidcopy;
1829 }
1830 WARN_ON_ONCE(1);
1831 }
1832
Adam Litke2fc39ce2007-11-14 16:59:39 -08001833 return -PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08001834 }
1835
1836 spin_unlock(&mm->page_table_lock);
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00001837 copy_huge_page(new_page, old_page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08001838 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08001839 spin_lock(&mm->page_table_lock);
1840
Andi Kleena5516432008-07-23 21:27:41 -07001841 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001842 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08001843 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07001844 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08001845 set_huge_pte_at(mm, address, ptep,
1846 make_huge_pte(vma, new_page, 1));
1847 /* Make the old page be freed below */
1848 new_page = old_page;
1849 }
1850 page_cache_release(new_page);
1851 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -07001852 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08001853}
1854
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001855/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07001856static struct page *hugetlbfs_pagecache_page(struct hstate *h,
1857 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001858{
1859 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07001860 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001861
1862 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07001863 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001864
1865 return find_lock_page(mapping, idx);
1866}
1867
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07001868static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
David Gibson1e8f8892006-01-06 00:10:44 -08001869 unsigned long address, pte_t *ptep, int write_access)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01001870{
Andi Kleena5516432008-07-23 21:27:41 -07001871 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01001872 int ret = VM_FAULT_SIGBUS;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07001873 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07001874 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07001875 struct page *page;
1876 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08001877 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -07001878
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001879 /*
1880 * Currently, we are forced to kill the process in the event the
1881 * original mapper has unmapped pages from the child due to a failed
1882 * COW. Warn that such a situation has occured as it may not be obvious
1883 */
1884 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1885 printk(KERN_WARNING
1886 "PID %d killed due to inadequate hugepage pool\n",
1887 current->pid);
1888 return ret;
1889 }
1890
Adam Litke4c887262005-10-29 18:16:46 -07001891 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07001892 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07001893
1894 /*
1895 * Use page lock to guard against racing truncation
1896 * before we get page_table_lock.
1897 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08001898retry:
1899 page = find_lock_page(mapping, idx);
1900 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07001901 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07001902 if (idx >= size)
1903 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001904 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08001905 if (IS_ERR(page)) {
1906 ret = -PTR_ERR(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08001907 goto out;
1908 }
Andi Kleena5516432008-07-23 21:27:41 -07001909 clear_huge_page(page, address, huge_page_size(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08001910 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01001911
Christoph Lameter6bda6662006-01-06 00:10:49 -08001912 if (vma->vm_flags & VM_SHARED) {
1913 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08001914 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08001915
1916 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1917 if (err) {
1918 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08001919 if (err == -EEXIST)
1920 goto retry;
1921 goto out;
1922 }
Ken Chen45c682a2007-11-14 16:59:44 -08001923
1924 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001925 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08001926 spin_unlock(&inode->i_lock);
Christoph Lameter6bda6662006-01-06 00:10:49 -08001927 } else
1928 lock_page(page);
1929 }
David Gibson1e8f8892006-01-06 00:10:44 -08001930
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01001931 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001932 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07001933 if (idx >= size)
1934 goto backout;
1935
Nick Piggin83c54072007-07-19 01:47:05 -07001936 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001937 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07001938 goto backout;
1939
David Gibson1e8f8892006-01-06 00:10:44 -08001940 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
1941 && (vma->vm_flags & VM_SHARED)));
1942 set_huge_pte_at(mm, address, ptep, new_pte);
1943
1944 if (write_access && !(vma->vm_flags & VM_SHARED)) {
1945 /* Optimization, do the COW without a second fault */
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001946 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
David Gibson1e8f8892006-01-06 00:10:44 -08001947 }
1948
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01001949 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07001950 unlock_page(page);
1951out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01001952 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07001953
1954backout:
1955 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07001956 unlock_page(page);
1957 put_page(page);
1958 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01001959}
1960
Adam Litke86e52162006-01-06 00:10:43 -08001961int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
1962 unsigned long address, int write_access)
1963{
1964 pte_t *ptep;
1965 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -08001966 int ret;
David Gibson3935baa2006-03-22 00:08:53 -08001967 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07001968 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08001969
Andi Kleena5516432008-07-23 21:27:41 -07001970 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08001971 if (!ptep)
1972 return VM_FAULT_OOM;
1973
David Gibson3935baa2006-03-22 00:08:53 -08001974 /*
1975 * Serialize hugepage allocation and instantiation, so that we don't
1976 * get spurious allocation failures if two CPUs race to instantiate
1977 * the same page in the page cache.
1978 */
1979 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001980 entry = huge_ptep_get(ptep);
1981 if (huge_pte_none(entry)) {
David Gibson3935baa2006-03-22 00:08:53 -08001982 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
1983 mutex_unlock(&hugetlb_instantiation_mutex);
1984 return ret;
1985 }
Adam Litke86e52162006-01-06 00:10:43 -08001986
Nick Piggin83c54072007-07-19 01:47:05 -07001987 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08001988
1989 spin_lock(&mm->page_table_lock);
1990 /* Check for a racing update before calling hugetlb_cow */
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001991 if (likely(pte_same(entry, huge_ptep_get(ptep))))
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001992 if (write_access && !pte_write(entry)) {
1993 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07001994 page = hugetlbfs_pagecache_page(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001995 ret = hugetlb_cow(mm, vma, address, ptep, entry, page);
1996 if (page) {
1997 unlock_page(page);
1998 put_page(page);
1999 }
2000 }
David Gibson1e8f8892006-01-06 00:10:44 -08002001 spin_unlock(&mm->page_table_lock);
David Gibson3935baa2006-03-22 00:08:53 -08002002 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08002003
2004 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08002005}
2006
Andi Kleenceb86872008-07-23 21:27:50 -07002007/* Can be overriden by architectures */
2008__attribute__((weak)) struct page *
2009follow_huge_pud(struct mm_struct *mm, unsigned long address,
2010 pud_t *pud, int write)
2011{
2012 BUG();
2013 return NULL;
2014}
2015
David Gibson63551ae2005-06-21 17:14:44 -07002016int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2017 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -08002018 unsigned long *position, int *length, int i,
2019 int write)
David Gibson63551ae2005-06-21 17:14:44 -07002020{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002021 unsigned long pfn_offset;
2022 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -07002023 int remainder = *length;
Andi Kleena5516432008-07-23 21:27:41 -07002024 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07002025
Hugh Dickins1c598272005-10-19 21:23:43 -07002026 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002027 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07002028 pte_t *pte;
2029 struct page *page;
2030
2031 /*
2032 * Some archs (sparc64, sh*) have multiple pte_ts to
2033 * each hugepage. We have to make * sure we get the
2034 * first, for the page indexing below to work.
2035 */
Andi Kleena5516432008-07-23 21:27:41 -07002036 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Adam Litke4c887262005-10-29 18:16:46 -07002037
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002038 if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
2039 (write && !pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07002040 int ret;
2041
2042 spin_unlock(&mm->page_table_lock);
Adam Litke5b23dbe2007-11-14 16:59:33 -08002043 ret = hugetlb_fault(mm, vma, vaddr, write);
Adam Litke4c887262005-10-29 18:16:46 -07002044 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07002045 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07002046 continue;
2047
2048 remainder = 0;
2049 if (!i)
2050 i = -EFAULT;
2051 break;
2052 }
David Gibson63551ae2005-06-21 17:14:44 -07002053
Andi Kleena5516432008-07-23 21:27:41 -07002054 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002055 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002056same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002057 if (pages) {
2058 get_page(page);
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002059 pages[i] = page + pfn_offset;
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002060 }
David Gibson63551ae2005-06-21 17:14:44 -07002061
2062 if (vmas)
2063 vmas[i] = vma;
2064
2065 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002066 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07002067 --remainder;
2068 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002069 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07002070 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002071 /*
2072 * We use pfn_offset to avoid touching the pageframes
2073 * of this compound page.
2074 */
2075 goto same_page;
2076 }
David Gibson63551ae2005-06-21 17:14:44 -07002077 }
Hugh Dickins1c598272005-10-19 21:23:43 -07002078 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002079 *length = remainder;
2080 *position = vaddr;
2081
2082 return i;
2083}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002084
2085void hugetlb_change_protection(struct vm_area_struct *vma,
2086 unsigned long address, unsigned long end, pgprot_t newprot)
2087{
2088 struct mm_struct *mm = vma->vm_mm;
2089 unsigned long start = address;
2090 pte_t *ptep;
2091 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07002092 struct hstate *h = hstate_vma(vma);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002093
2094 BUG_ON(address >= end);
2095 flush_cache_range(vma, address, end);
2096
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002097 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002098 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002099 for (; address < end; address += huge_page_size(h)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002100 ptep = huge_pte_offset(mm, address);
2101 if (!ptep)
2102 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002103 if (huge_pmd_unshare(mm, &address, ptep))
2104 continue;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002105 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002106 pte = huge_ptep_get_and_clear(mm, address, ptep);
2107 pte = pte_mkhuge(pte_modify(pte, newprot));
2108 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002109 }
2110 }
2111 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002112 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002113
2114 flush_tlb_range(vma, start, end);
2115}
2116
Mel Gormana1e78772008-07-23 21:27:23 -07002117int hugetlb_reserve_pages(struct inode *inode,
2118 long from, long to,
2119 struct vm_area_struct *vma)
Adam Litkee4e574b2007-10-16 01:26:19 -07002120{
2121 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07002122 struct hstate *h = hstate_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07002123
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002124 if (vma && vma->vm_flags & VM_NORESERVE)
2125 return 0;
2126
Mel Gormana1e78772008-07-23 21:27:23 -07002127 /*
2128 * Shared mappings base their reservation on the number of pages that
2129 * are already allocated on behalf of the file. Private mappings need
2130 * to reserve the full area even if read-only as mprotect() may be
2131 * called to make the mapping read-write. Assume !vma is a shm mapping
2132 */
2133 if (!vma || vma->vm_flags & VM_SHARED)
2134 chg = region_chg(&inode->i_mapping->private_list, from, to);
2135 else {
Andy Whitcroft84afd992008-07-23 21:27:32 -07002136 struct resv_map *resv_map = resv_map_alloc();
2137 if (!resv_map)
2138 return -ENOMEM;
2139
Mel Gormana1e78772008-07-23 21:27:23 -07002140 chg = to - from;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002141
2142 set_vma_resv_map(vma, resv_map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002143 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
Mel Gormana1e78772008-07-23 21:27:23 -07002144 }
2145
Adam Litkee4e574b2007-10-16 01:26:19 -07002146 if (chg < 0)
2147 return chg;
Ken Chen8a630112007-05-09 02:33:34 -07002148
Adam Litke90d8b7e2007-11-14 16:59:42 -08002149 if (hugetlb_get_quota(inode->i_mapping, chg))
2150 return -ENOSPC;
Andi Kleena5516432008-07-23 21:27:41 -07002151 ret = hugetlb_acct_memory(h, chg);
Ken Chen68842c92008-01-14 00:55:19 -08002152 if (ret < 0) {
2153 hugetlb_put_quota(inode->i_mapping, chg);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002154 return ret;
Ken Chen68842c92008-01-14 00:55:19 -08002155 }
Mel Gormana1e78772008-07-23 21:27:23 -07002156 if (!vma || vma->vm_flags & VM_SHARED)
2157 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002158 return 0;
2159}
2160
2161void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2162{
Andi Kleena5516432008-07-23 21:27:41 -07002163 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002164 long chg = region_truncate(&inode->i_mapping->private_list, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08002165
2166 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002167 inode->i_blocks -= blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002168 spin_unlock(&inode->i_lock);
2169
Adam Litke90d8b7e2007-11-14 16:59:42 -08002170 hugetlb_put_quota(inode->i_mapping, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07002171 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002172}