blob: 6058b53dcb8905bc33bc25371083b47167de7596 [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>
Alexey Dobriyane1759c22008-10-15 23:50:22 +040010#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070011#include <linux/sysctl.h>
12#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070013#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070015#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080016#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080017#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080018#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070019#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070020#include <linux/sysfs.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070021
David Gibson63551ae2005-06-21 17:14:44 -070022#include <asm/page.h>
23#include <asm/pgtable.h>
Adrian Bunk78a34ae2008-07-28 15:46:30 -070024#include <asm/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070025
26#include <linux/hugetlb.h>
Nick Piggin7835e982006-03-22 00:08:40 -080027#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070028
29const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
Mel Gorman396faf02007-07-17 04:03:13 -070030static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
31unsigned long hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070032
Andi Kleene5ff2152008-07-23 21:27:42 -070033static int max_hstate;
34unsigned int default_hstate_idx;
35struct hstate hstates[HUGE_MAX_HSTATE];
36
Jon Tollefson53ba51d2008-07-23 21:27:52 -070037__initdata LIST_HEAD(huge_boot_pages);
38
Andi Kleene5ff2152008-07-23 21:27:42 -070039/* for command line parsing */
40static struct hstate * __initdata parsed_hstate;
41static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070042static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070043
44#define for_each_hstate(h) \
45 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
Mel Gorman396faf02007-07-17 04:03:13 -070046
David Gibson3935baa2006-03-22 00:08:53 -080047/*
48 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
49 */
50static DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080051
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -070052/*
Andy Whitcroft96822902008-07-23 21:27:29 -070053 * Region tracking -- allows tracking of reservations and instantiated pages
54 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -070055 *
56 * The region data structures are protected by a combination of the mmap_sem
57 * and the hugetlb_instantion_mutex. To access or modify a region the caller
58 * must either hold the mmap_sem for write, or the mmap_sem for read and
59 * the hugetlb_instantiation mutex:
60 *
61 * down_write(&mm->mmap_sem);
62 * or
63 * down_read(&mm->mmap_sem);
64 * mutex_lock(&hugetlb_instantiation_mutex);
Andy Whitcroft96822902008-07-23 21:27:29 -070065 */
66struct file_region {
67 struct list_head link;
68 long from;
69 long to;
70};
71
72static long region_add(struct list_head *head, long f, long t)
73{
74 struct file_region *rg, *nrg, *trg;
75
76 /* Locate the region we are either in or before. */
77 list_for_each_entry(rg, head, link)
78 if (f <= rg->to)
79 break;
80
81 /* Round our left edge to the current segment if it encloses us. */
82 if (f > rg->from)
83 f = rg->from;
84
85 /* Check for and consume any regions we now overlap with. */
86 nrg = rg;
87 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
88 if (&rg->link == head)
89 break;
90 if (rg->from > t)
91 break;
92
93 /* If this area reaches higher then extend our area to
94 * include it completely. If this is not the first area
95 * which we intend to reuse, free it. */
96 if (rg->to > t)
97 t = rg->to;
98 if (rg != nrg) {
99 list_del(&rg->link);
100 kfree(rg);
101 }
102 }
103 nrg->from = f;
104 nrg->to = t;
105 return 0;
106}
107
108static long region_chg(struct list_head *head, long f, long t)
109{
110 struct file_region *rg, *nrg;
111 long chg = 0;
112
113 /* Locate the region we are before or in. */
114 list_for_each_entry(rg, head, link)
115 if (f <= rg->to)
116 break;
117
118 /* If we are below the current region then a new region is required.
119 * Subtle, allocate a new region at the position but make it zero
120 * size such that we can guarantee to record the reservation. */
121 if (&rg->link == head || t < rg->from) {
122 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
123 if (!nrg)
124 return -ENOMEM;
125 nrg->from = f;
126 nrg->to = f;
127 INIT_LIST_HEAD(&nrg->link);
128 list_add(&nrg->link, rg->link.prev);
129
130 return t - f;
131 }
132
133 /* Round our left edge to the current segment if it encloses us. */
134 if (f > rg->from)
135 f = rg->from;
136 chg = t - f;
137
138 /* Check for and consume any regions we now overlap with. */
139 list_for_each_entry(rg, rg->link.prev, link) {
140 if (&rg->link == head)
141 break;
142 if (rg->from > t)
143 return chg;
144
145 /* We overlap with this area, if it extends futher than
146 * us then we must extend ourselves. Account for its
147 * existing reservation. */
148 if (rg->to > t) {
149 chg += rg->to - t;
150 t = rg->to;
151 }
152 chg -= rg->to - rg->from;
153 }
154 return chg;
155}
156
157static long region_truncate(struct list_head *head, long end)
158{
159 struct file_region *rg, *trg;
160 long chg = 0;
161
162 /* Locate the region we are either in or before. */
163 list_for_each_entry(rg, head, link)
164 if (end <= rg->to)
165 break;
166 if (&rg->link == head)
167 return 0;
168
169 /* If we are in the middle of a region then adjust it. */
170 if (end > rg->from) {
171 chg = rg->to - end;
172 rg->to = end;
173 rg = list_entry(rg->link.next, typeof(*rg), link);
174 }
175
176 /* Drop any remaining regions. */
177 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
178 if (&rg->link == head)
179 break;
180 chg += rg->to - rg->from;
181 list_del(&rg->link);
182 kfree(rg);
183 }
184 return chg;
185}
186
Andy Whitcroft84afd992008-07-23 21:27:32 -0700187static long region_count(struct list_head *head, long f, long t)
188{
189 struct file_region *rg;
190 long chg = 0;
191
192 /* Locate each segment we overlap with, and count that overlap. */
193 list_for_each_entry(rg, head, link) {
194 int seg_from;
195 int seg_to;
196
197 if (rg->to <= f)
198 continue;
199 if (rg->from >= t)
200 break;
201
202 seg_from = max(rg->from, f);
203 seg_to = min(rg->to, t);
204
205 chg += seg_to - seg_from;
206 }
207
208 return chg;
209}
210
Andy Whitcroft96822902008-07-23 21:27:29 -0700211/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700212 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700213 * the mapping, in pagecache page units; huge pages here.
214 */
Andi Kleena5516432008-07-23 21:27:41 -0700215static pgoff_t vma_hugecache_offset(struct hstate *h,
216 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700217{
Andi Kleena5516432008-07-23 21:27:41 -0700218 return ((address - vma->vm_start) >> huge_page_shift(h)) +
219 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700220}
221
Andy Whitcroft84afd992008-07-23 21:27:32 -0700222/*
223 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
224 * bits of the reservation map pointer, which are always clear due to
225 * alignment.
226 */
227#define HPAGE_RESV_OWNER (1UL << 0)
228#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700229#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700230
Mel Gormana1e78772008-07-23 21:27:23 -0700231/*
232 * These helpers are used to track how many pages are reserved for
233 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
234 * is guaranteed to have their future faults succeed.
235 *
236 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
237 * the reserve counters are updated with the hugetlb_lock held. It is safe
238 * to reset the VMA at fork() time as it is not in use yet and there is no
239 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700240 *
241 * The private mapping reservation is represented in a subtly different
242 * manner to a shared mapping. A shared mapping has a region map associated
243 * with the underlying file, this region map represents the backing file
244 * pages which have ever had a reservation assigned which this persists even
245 * after the page is instantiated. A private mapping has a region map
246 * associated with the original mmap which is attached to all VMAs which
247 * reference it, this region map represents those offsets which have consumed
248 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700249 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700250static unsigned long get_vma_private_data(struct vm_area_struct *vma)
251{
252 return (unsigned long)vma->vm_private_data;
253}
254
255static void set_vma_private_data(struct vm_area_struct *vma,
256 unsigned long value)
257{
258 vma->vm_private_data = (void *)value;
259}
260
Andy Whitcroft84afd992008-07-23 21:27:32 -0700261struct resv_map {
262 struct kref refs;
263 struct list_head regions;
264};
265
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700266static struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700267{
268 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
269 if (!resv_map)
270 return NULL;
271
272 kref_init(&resv_map->refs);
273 INIT_LIST_HEAD(&resv_map->regions);
274
275 return resv_map;
276}
277
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700278static void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700279{
280 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
281
282 /* Clear out any active regions before we release the map. */
283 region_truncate(&resv_map->regions, 0);
284 kfree(resv_map);
285}
286
287static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700288{
289 VM_BUG_ON(!is_vm_hugetlb_page(vma));
290 if (!(vma->vm_flags & VM_SHARED))
Andy Whitcroft84afd992008-07-23 21:27:32 -0700291 return (struct resv_map *)(get_vma_private_data(vma) &
292 ~HPAGE_RESV_MASK);
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700293 return NULL;
Mel Gormana1e78772008-07-23 21:27:23 -0700294}
295
Andy Whitcroft84afd992008-07-23 21:27:32 -0700296static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700297{
298 VM_BUG_ON(!is_vm_hugetlb_page(vma));
299 VM_BUG_ON(vma->vm_flags & VM_SHARED);
300
Andy Whitcroft84afd992008-07-23 21:27:32 -0700301 set_vma_private_data(vma, (get_vma_private_data(vma) &
302 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700303}
304
305static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
306{
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700307 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700308 VM_BUG_ON(vma->vm_flags & VM_SHARED);
309
310 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700311}
312
313static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
314{
315 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700316
317 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700318}
319
320/* Decrement the reserved pages in the hugepage pool by one */
Andi Kleena5516432008-07-23 21:27:41 -0700321static void decrement_hugepage_resv_vma(struct hstate *h,
322 struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700323{
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700324 if (vma->vm_flags & VM_NORESERVE)
325 return;
326
Mel Gormana1e78772008-07-23 21:27:23 -0700327 if (vma->vm_flags & VM_SHARED) {
328 /* Shared mappings always use reserves */
Andi Kleena5516432008-07-23 21:27:41 -0700329 h->resv_huge_pages--;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700330 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Mel Gormana1e78772008-07-23 21:27:23 -0700331 /*
332 * Only the process that called mmap() has reserves for
333 * private mappings.
334 */
Andi Kleena5516432008-07-23 21:27:41 -0700335 h->resv_huge_pages--;
Mel Gormana1e78772008-07-23 21:27:23 -0700336 }
337}
338
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700339/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700340void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
341{
342 VM_BUG_ON(!is_vm_hugetlb_page(vma));
343 if (!(vma->vm_flags & VM_SHARED))
344 vma->vm_private_data = (void *)0;
345}
346
347/* Returns true if the VMA has associated reserve pages */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700348static int vma_has_reserves(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700349{
350 if (vma->vm_flags & VM_SHARED)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700351 return 1;
352 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
353 return 1;
354 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700355}
356
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800357static void clear_gigantic_page(struct page *page,
358 unsigned long addr, unsigned long sz)
359{
360 int i;
361 struct page *p = page;
362
363 might_sleep();
364 for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
365 cond_resched();
366 clear_user_highpage(p, addr + i * PAGE_SIZE);
367 }
368}
Andi Kleena5516432008-07-23 21:27:41 -0700369static void clear_huge_page(struct page *page,
370 unsigned long addr, unsigned long sz)
David Gibson79ac6ba2006-03-22 00:08:51 -0800371{
372 int i;
373
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800374 if (unlikely(sz > MAX_ORDER_NR_PAGES))
375 return clear_gigantic_page(page, addr, sz);
376
David Gibson79ac6ba2006-03-22 00:08:51 -0800377 might_sleep();
Andi Kleena5516432008-07-23 21:27:41 -0700378 for (i = 0; i < sz/PAGE_SIZE; i++) {
David Gibson79ac6ba2006-03-22 00:08:51 -0800379 cond_resched();
Ralf Baechle281e0e32007-10-01 01:20:10 -0700380 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
David Gibson79ac6ba2006-03-22 00:08:51 -0800381 }
382}
383
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800384static void copy_gigantic_page(struct page *dst, struct page *src,
385 unsigned long addr, struct vm_area_struct *vma)
386{
387 int i;
388 struct hstate *h = hstate_vma(vma);
389 struct page *dst_base = dst;
390 struct page *src_base = src;
391 might_sleep();
392 for (i = 0; i < pages_per_huge_page(h); ) {
393 cond_resched();
394 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
395
396 i++;
397 dst = mem_map_next(dst, dst_base, i);
398 src = mem_map_next(src, src_base, i);
399 }
400}
David Gibson79ac6ba2006-03-22 00:08:51 -0800401static void copy_huge_page(struct page *dst, struct page *src,
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000402 unsigned long addr, struct vm_area_struct *vma)
David Gibson79ac6ba2006-03-22 00:08:51 -0800403{
404 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700405 struct hstate *h = hstate_vma(vma);
David Gibson79ac6ba2006-03-22 00:08:51 -0800406
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800407 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES))
408 return copy_gigantic_page(dst, src, addr, vma);
409
David Gibson79ac6ba2006-03-22 00:08:51 -0800410 might_sleep();
Andi Kleena5516432008-07-23 21:27:41 -0700411 for (i = 0; i < pages_per_huge_page(h); i++) {
David Gibson79ac6ba2006-03-22 00:08:51 -0800412 cond_resched();
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000413 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
David Gibson79ac6ba2006-03-22 00:08:51 -0800414 }
415}
416
Andi Kleena5516432008-07-23 21:27:41 -0700417static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418{
419 int nid = page_to_nid(page);
Andi Kleena5516432008-07-23 21:27:41 -0700420 list_add(&page->lru, &h->hugepage_freelists[nid]);
421 h->free_huge_pages++;
422 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423}
424
Andi Kleena5516432008-07-23 21:27:41 -0700425static struct page *dequeue_huge_page(struct hstate *h)
Nishanth Aravamudan348e1e02008-03-04 14:29:42 -0800426{
427 int nid;
428 struct page *page = NULL;
429
430 for (nid = 0; nid < MAX_NUMNODES; ++nid) {
Andi Kleena5516432008-07-23 21:27:41 -0700431 if (!list_empty(&h->hugepage_freelists[nid])) {
432 page = list_entry(h->hugepage_freelists[nid].next,
Nishanth Aravamudan348e1e02008-03-04 14:29:42 -0800433 struct page, lru);
434 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700435 h->free_huge_pages--;
436 h->free_huge_pages_node[nid]--;
Nishanth Aravamudan348e1e02008-03-04 14:29:42 -0800437 break;
438 }
439 }
440 return page;
441}
442
Andi Kleena5516432008-07-23 21:27:41 -0700443static struct page *dequeue_huge_page_vma(struct hstate *h,
444 struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700445 unsigned long address, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446{
Nishanth Aravamudan31a5c6e2007-07-15 23:38:02 -0700447 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700449 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700450 nodemask_t *nodemask;
Mel Gorman396faf02007-07-17 04:03:13 -0700451 struct zonelist *zonelist = huge_zonelist(vma, address,
Mel Gorman19770b32008-04-28 02:12:18 -0700452 htlb_alloc_mask, &mpol, &nodemask);
Mel Gormandd1a2392008-04-28 02:12:17 -0700453 struct zone *zone;
454 struct zoneref *z;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455
Mel Gormana1e78772008-07-23 21:27:23 -0700456 /*
457 * A child process with MAP_PRIVATE mappings created by their parent
458 * have no page reserves. This check ensures that reservations are
459 * not "stolen". The child may still get SIGKILLed
460 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700461 if (!vma_has_reserves(vma) &&
Andi Kleena5516432008-07-23 21:27:41 -0700462 h->free_huge_pages - h->resv_huge_pages == 0)
Mel Gormana1e78772008-07-23 21:27:23 -0700463 return NULL;
464
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700465 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700466 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700467 return NULL;
468
Mel Gorman19770b32008-04-28 02:12:18 -0700469 for_each_zone_zonelist_nodemask(zone, z, zonelist,
470 MAX_NR_ZONES - 1, nodemask) {
Mel Gorman54a6eb52008-04-28 02:12:16 -0700471 nid = zone_to_nid(zone);
472 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
Andi Kleena5516432008-07-23 21:27:41 -0700473 !list_empty(&h->hugepage_freelists[nid])) {
474 page = list_entry(h->hugepage_freelists[nid].next,
Andrew Morton3abf7af2007-07-19 01:49:08 -0700475 struct page, lru);
476 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700477 h->free_huge_pages--;
478 h->free_huge_pages_node[nid]--;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700479
480 if (!avoid_reserve)
Andi Kleena5516432008-07-23 21:27:41 -0700481 decrement_hugepage_resv_vma(h, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700482
Ken Chen5ab3ee72007-07-23 18:44:00 -0700483 break;
Andrew Morton3abf7af2007-07-19 01:49:08 -0700484 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 }
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700486 mpol_cond_put(mpol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 return page;
488}
489
Andi Kleena5516432008-07-23 21:27:41 -0700490static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700491{
492 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700493
Andy Whitcroft18229df2008-11-06 12:53:27 -0800494 VM_BUG_ON(h->order >= MAX_ORDER);
495
Andi Kleena5516432008-07-23 21:27:41 -0700496 h->nr_huge_pages--;
497 h->nr_huge_pages_node[page_to_nid(page)]--;
498 for (i = 0; i < pages_per_huge_page(h); i++) {
Adam Litke6af2acb2007-10-16 01:26:16 -0700499 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
500 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
501 1 << PG_private | 1<< PG_writeback);
502 }
503 set_compound_page_dtor(page, NULL);
504 set_page_refcounted(page);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700505 arch_release_hugepage(page);
Andi Kleena5516432008-07-23 21:27:41 -0700506 __free_pages(page, huge_page_order(h));
Adam Litke6af2acb2007-10-16 01:26:16 -0700507}
508
Andi Kleene5ff2152008-07-23 21:27:42 -0700509struct hstate *size_to_hstate(unsigned long size)
510{
511 struct hstate *h;
512
513 for_each_hstate(h) {
514 if (huge_page_size(h) == size)
515 return h;
516 }
517 return NULL;
518}
519
David Gibson27a85ef2006-03-22 00:08:56 -0800520static void free_huge_page(struct page *page)
521{
Andi Kleena5516432008-07-23 21:27:41 -0700522 /*
523 * Can't pass hstate in here because it is called from the
524 * compound page destructor.
525 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700526 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700527 int nid = page_to_nid(page);
Adam Litkec79fb752007-11-14 16:59:38 -0800528 struct address_space *mapping;
David Gibson27a85ef2006-03-22 00:08:56 -0800529
Adam Litkec79fb752007-11-14 16:59:38 -0800530 mapping = (struct address_space *) page_private(page);
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800531 set_page_private(page, 0);
Adam Litke7893d1d2007-10-16 01:26:18 -0700532 BUG_ON(page_count(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800533 INIT_LIST_HEAD(&page->lru);
534
535 spin_lock(&hugetlb_lock);
Andi Kleenaa888a72008-07-23 21:27:47 -0700536 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
Andi Kleena5516432008-07-23 21:27:41 -0700537 update_and_free_page(h, page);
538 h->surplus_huge_pages--;
539 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700540 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700541 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700542 }
David Gibson27a85ef2006-03-22 00:08:56 -0800543 spin_unlock(&hugetlb_lock);
Adam Litkec79fb752007-11-14 16:59:38 -0800544 if (mapping)
Adam Litke9a119c02007-11-14 16:59:41 -0800545 hugetlb_put_quota(mapping, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800546}
547
Adam Litke7893d1d2007-10-16 01:26:18 -0700548/*
549 * Increment or decrement surplus_huge_pages. Keep node-specific counters
550 * balanced by operating on them in a round-robin fashion.
551 * Returns 1 if an adjustment was made.
552 */
Andi Kleena5516432008-07-23 21:27:41 -0700553static int adjust_pool_surplus(struct hstate *h, int delta)
Adam Litke7893d1d2007-10-16 01:26:18 -0700554{
555 static int prev_nid;
556 int nid = prev_nid;
557 int ret = 0;
558
559 VM_BUG_ON(delta != -1 && delta != 1);
560 do {
561 nid = next_node(nid, node_online_map);
562 if (nid == MAX_NUMNODES)
563 nid = first_node(node_online_map);
564
565 /* To shrink on this node, there must be a surplus page */
Andi Kleena5516432008-07-23 21:27:41 -0700566 if (delta < 0 && !h->surplus_huge_pages_node[nid])
Adam Litke7893d1d2007-10-16 01:26:18 -0700567 continue;
568 /* Surplus cannot exceed the total number of pages */
Andi Kleena5516432008-07-23 21:27:41 -0700569 if (delta > 0 && h->surplus_huge_pages_node[nid] >=
570 h->nr_huge_pages_node[nid])
Adam Litke7893d1d2007-10-16 01:26:18 -0700571 continue;
572
Andi Kleena5516432008-07-23 21:27:41 -0700573 h->surplus_huge_pages += delta;
574 h->surplus_huge_pages_node[nid] += delta;
Adam Litke7893d1d2007-10-16 01:26:18 -0700575 ret = 1;
576 break;
577 } while (nid != prev_nid);
578
579 prev_nid = nid;
580 return ret;
581}
582
Andi Kleena5516432008-07-23 21:27:41 -0700583static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700584{
585 set_compound_page_dtor(page, free_huge_page);
586 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700587 h->nr_huge_pages++;
588 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700589 spin_unlock(&hugetlb_lock);
590 put_page(page); /* free it into the hugepage allocator */
591}
592
Andi Kleena5516432008-07-23 21:27:41 -0700593static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700596
Andi Kleenaa888a72008-07-23 21:27:47 -0700597 if (h->order >= MAX_ORDER)
598 return NULL;
599
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700600 page = alloc_pages_node(nid,
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700601 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
602 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700603 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700605 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700606 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700607 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700608 }
Andi Kleena5516432008-07-23 21:27:41 -0700609 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700611
612 return page;
613}
614
Andi Kleen5ced66c2008-07-23 21:27:45 -0700615/*
616 * Use a helper variable to find the next node and then
617 * copy it back to hugetlb_next_nid afterwards:
618 * otherwise there's a window in which a racer might
619 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
620 * But we don't need to use a spin_lock here: it really
621 * doesn't matter if occasionally a racer chooses the
622 * same nid as we do. Move nid forward in the mask even
623 * if we just successfully allocated a hugepage so that
624 * the next caller gets hugepages on the next node.
625 */
626static int hstate_next_node(struct hstate *h)
627{
628 int next_nid;
629 next_nid = next_node(h->hugetlb_next_nid, node_online_map);
630 if (next_nid == MAX_NUMNODES)
631 next_nid = first_node(node_online_map);
632 h->hugetlb_next_nid = next_nid;
633 return next_nid;
634}
635
Andi Kleena5516432008-07-23 21:27:41 -0700636static int alloc_fresh_huge_page(struct hstate *h)
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700637{
638 struct page *page;
639 int start_nid;
640 int next_nid;
641 int ret = 0;
642
Andi Kleena5516432008-07-23 21:27:41 -0700643 start_nid = h->hugetlb_next_nid;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700644
645 do {
Andi Kleena5516432008-07-23 21:27:41 -0700646 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700647 if (page)
648 ret = 1;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700649 next_nid = hstate_next_node(h);
Andi Kleena5516432008-07-23 21:27:41 -0700650 } while (!page && h->hugetlb_next_nid != start_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700651
Adam Litke3b116302008-04-28 02:13:06 -0700652 if (ret)
653 count_vm_event(HTLB_BUDDY_PGALLOC);
654 else
655 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
656
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700657 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658}
659
Andi Kleena5516432008-07-23 21:27:41 -0700660static struct page *alloc_buddy_huge_page(struct hstate *h,
661 struct vm_area_struct *vma, unsigned long address)
Adam Litke7893d1d2007-10-16 01:26:18 -0700662{
663 struct page *page;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800664 unsigned int nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700665
Andi Kleenaa888a72008-07-23 21:27:47 -0700666 if (h->order >= MAX_ORDER)
667 return NULL;
668
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800669 /*
670 * Assume we will successfully allocate the surplus page to
671 * prevent racing processes from causing the surplus to exceed
672 * overcommit
673 *
674 * This however introduces a different race, where a process B
675 * tries to grow the static hugepage pool while alloc_pages() is
676 * called by process A. B will only examine the per-node
677 * counters in determining if surplus huge pages can be
678 * converted to normal huge pages in adjust_pool_surplus(). A
679 * won't be able to increment the per-node counter, until the
680 * lock is dropped by B, but B doesn't drop hugetlb_lock until
681 * no more huge pages can be converted from surplus to normal
682 * state (and doesn't try to convert again). Thus, we have a
683 * case where a surplus huge page exists, the pool is grown, and
684 * the surplus huge page still exists after, even though it
685 * should just have been converted to a normal huge page. This
686 * does not leak memory, though, as the hugepage will be freed
687 * once it is out of use. It also does not allow the counters to
688 * go out of whack in adjust_pool_surplus() as we don't modify
689 * the node values until we've gotten the hugepage and only the
690 * per-node value is checked there.
691 */
692 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700693 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800694 spin_unlock(&hugetlb_lock);
695 return NULL;
696 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700697 h->nr_huge_pages++;
698 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800699 }
700 spin_unlock(&hugetlb_lock);
701
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700702 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
703 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700704 huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800705
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700706 if (page && arch_prepare_hugepage(page)) {
707 __free_pages(page, huge_page_order(h));
708 return NULL;
709 }
710
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800711 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700712 if (page) {
Adam Litke2668db92008-03-10 11:43:50 -0700713 /*
714 * This page is now managed by the hugetlb allocator and has
715 * no users -- drop the buddy allocator's reference.
716 */
717 put_page_testzero(page);
718 VM_BUG_ON(page_count(page));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800719 nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700720 set_compound_page_dtor(page, free_huge_page);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800721 /*
722 * We incremented the global counters already
723 */
Andi Kleena5516432008-07-23 21:27:41 -0700724 h->nr_huge_pages_node[nid]++;
725 h->surplus_huge_pages_node[nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700726 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800727 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700728 h->nr_huge_pages--;
729 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700730 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700731 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800732 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700733
734 return page;
735}
736
Adam Litkee4e574b2007-10-16 01:26:19 -0700737/*
738 * Increase the hugetlb pool such that it can accomodate a reservation
739 * of size 'delta'.
740 */
Andi Kleena5516432008-07-23 21:27:41 -0700741static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -0700742{
743 struct list_head surplus_list;
744 struct page *page, *tmp;
745 int ret, i;
746 int needed, allocated;
747
Andi Kleena5516432008-07-23 21:27:41 -0700748 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800749 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -0700750 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700751 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800752 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700753
754 allocated = 0;
755 INIT_LIST_HEAD(&surplus_list);
756
757 ret = -ENOMEM;
758retry:
759 spin_unlock(&hugetlb_lock);
760 for (i = 0; i < needed; i++) {
Andi Kleena5516432008-07-23 21:27:41 -0700761 page = alloc_buddy_huge_page(h, NULL, 0);
Adam Litkee4e574b2007-10-16 01:26:19 -0700762 if (!page) {
763 /*
764 * We were not able to allocate enough pages to
765 * satisfy the entire reservation so we free what
766 * we've allocated so far.
767 */
768 spin_lock(&hugetlb_lock);
769 needed = 0;
770 goto free;
771 }
772
773 list_add(&page->lru, &surplus_list);
774 }
775 allocated += needed;
776
777 /*
778 * After retaking hugetlb_lock, we need to recalculate 'needed'
779 * because either resv_huge_pages or free_huge_pages may have changed.
780 */
781 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700782 needed = (h->resv_huge_pages + delta) -
783 (h->free_huge_pages + allocated);
Adam Litkee4e574b2007-10-16 01:26:19 -0700784 if (needed > 0)
785 goto retry;
786
787 /*
788 * The surplus_list now contains _at_least_ the number of extra pages
789 * needed to accomodate the reservation. Add the appropriate number
790 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -0800791 * allocator. Commit the entire reservation here to prevent another
792 * process from stealing the pages as they are added to the pool but
793 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -0700794 */
795 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -0700796 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700797 ret = 0;
798free:
Adam Litke19fc3f02008-04-28 02:12:20 -0700799 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -0700800 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -0700801 if ((--needed) < 0)
802 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700803 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700804 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -0700805 }
806
807 /* Free unnecessary surplus pages to the buddy allocator */
808 if (!list_empty(&surplus_list)) {
809 spin_unlock(&hugetlb_lock);
810 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
811 list_del(&page->lru);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700812 /*
Adam Litke2668db92008-03-10 11:43:50 -0700813 * The page has a reference count of zero already, so
814 * call free_huge_page directly instead of using
815 * put_page. This must be done with hugetlb_lock
Adam Litkeaf767cb2007-10-16 01:26:25 -0700816 * unlocked which is safe because free_huge_page takes
817 * hugetlb_lock before deciding how to free the page.
818 */
Adam Litke2668db92008-03-10 11:43:50 -0700819 free_huge_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700820 }
Adam Litke19fc3f02008-04-28 02:12:20 -0700821 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -0700822 }
823
824 return ret;
825}
826
827/*
828 * When releasing a hugetlb pool reservation, any surplus pages that were
829 * allocated to satisfy the reservation must be explicitly freed if they were
830 * never used.
831 */
Andi Kleena5516432008-07-23 21:27:41 -0700832static void return_unused_surplus_pages(struct hstate *h,
833 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -0700834{
835 static int nid = -1;
836 struct page *page;
837 unsigned long nr_pages;
838
Nishanth Aravamudan11320d12008-03-26 14:40:20 -0700839 /*
840 * We want to release as many surplus pages as possible, spread
841 * evenly across all nodes. Iterate across all nodes until we
842 * can no longer free unreserved surplus pages. This occurs when
843 * the nodes with surplus pages have no free pages.
844 */
845 unsigned long remaining_iterations = num_online_nodes();
846
Adam Litkeac09b3a2008-03-04 14:29:38 -0800847 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -0700848 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800849
Andi Kleenaa888a72008-07-23 21:27:47 -0700850 /* Cannot return gigantic pages currently */
851 if (h->order >= MAX_ORDER)
852 return;
853
Andi Kleena5516432008-07-23 21:27:41 -0700854 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -0700855
Nishanth Aravamudan11320d12008-03-26 14:40:20 -0700856 while (remaining_iterations-- && nr_pages) {
Adam Litkee4e574b2007-10-16 01:26:19 -0700857 nid = next_node(nid, node_online_map);
858 if (nid == MAX_NUMNODES)
859 nid = first_node(node_online_map);
860
Andi Kleena5516432008-07-23 21:27:41 -0700861 if (!h->surplus_huge_pages_node[nid])
Adam Litkee4e574b2007-10-16 01:26:19 -0700862 continue;
863
Andi Kleena5516432008-07-23 21:27:41 -0700864 if (!list_empty(&h->hugepage_freelists[nid])) {
865 page = list_entry(h->hugepage_freelists[nid].next,
Adam Litkee4e574b2007-10-16 01:26:19 -0700866 struct page, lru);
867 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700868 update_and_free_page(h, page);
869 h->free_huge_pages--;
870 h->free_huge_pages_node[nid]--;
871 h->surplus_huge_pages--;
872 h->surplus_huge_pages_node[nid]--;
Adam Litkee4e574b2007-10-16 01:26:19 -0700873 nr_pages--;
Nishanth Aravamudan11320d12008-03-26 14:40:20 -0700874 remaining_iterations = num_online_nodes();
Adam Litkee4e574b2007-10-16 01:26:19 -0700875 }
876 }
877}
878
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700879/*
880 * Determine if the huge page at addr within the vma has an associated
881 * reservation. Where it does not we will need to logically increase
882 * reservation and actually increase quota before an allocation can occur.
883 * Where any new reservation would be required the reservation change is
884 * prepared, but not committed. Once the page has been quota'd allocated
885 * an instantiated the change should be committed via vma_commit_reservation.
886 * No action is required on failure.
887 */
Andi Kleena5516432008-07-23 21:27:41 -0700888static int vma_needs_reservation(struct hstate *h,
889 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700890{
891 struct address_space *mapping = vma->vm_file->f_mapping;
892 struct inode *inode = mapping->host;
893
894 if (vma->vm_flags & VM_SHARED) {
Andi Kleena5516432008-07-23 21:27:41 -0700895 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700896 return region_chg(&inode->i_mapping->private_list,
897 idx, idx + 1);
898
Andy Whitcroft84afd992008-07-23 21:27:32 -0700899 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
900 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700901
Andy Whitcroft84afd992008-07-23 21:27:32 -0700902 } else {
903 int err;
Andi Kleena5516432008-07-23 21:27:41 -0700904 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700905 struct resv_map *reservations = vma_resv_map(vma);
906
907 err = region_chg(&reservations->regions, idx, idx + 1);
908 if (err < 0)
909 return err;
910 return 0;
911 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700912}
Andi Kleena5516432008-07-23 21:27:41 -0700913static void vma_commit_reservation(struct hstate *h,
914 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700915{
916 struct address_space *mapping = vma->vm_file->f_mapping;
917 struct inode *inode = mapping->host;
918
919 if (vma->vm_flags & VM_SHARED) {
Andi Kleena5516432008-07-23 21:27:41 -0700920 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700921 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700922
923 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -0700924 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700925 struct resv_map *reservations = vma_resv_map(vma);
926
927 /* Mark this page used in the map. */
928 region_add(&reservations->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700929 }
930}
931
David Gibson27a85ef2006-03-22 00:08:56 -0800932static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700933 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934{
Andi Kleena5516432008-07-23 21:27:41 -0700935 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -0800936 struct page *page;
Adam Litke2fc39ce2007-11-14 16:59:39 -0800937 struct address_space *mapping = vma->vm_file->f_mapping;
Mel Gormana1e78772008-07-23 21:27:23 -0700938 struct inode *inode = mapping->host;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700939 unsigned int chg;
Adam Litke2fc39ce2007-11-14 16:59:39 -0800940
Mel Gormana1e78772008-07-23 21:27:23 -0700941 /*
942 * Processes that did not create the mapping will have no reserves and
943 * will not have accounted against quota. Check that the quota can be
944 * made before satisfying the allocation
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700945 * MAP_NORESERVE mappings may also need pages and quota allocated
946 * if no reserve mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -0700947 */
Andi Kleena5516432008-07-23 21:27:41 -0700948 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700949 if (chg < 0)
950 return ERR_PTR(chg);
951 if (chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700952 if (hugetlb_get_quota(inode->i_mapping, chg))
953 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -0700954
955 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700956 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
Mel Gormana1e78772008-07-23 21:27:23 -0700957 spin_unlock(&hugetlb_lock);
958
959 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -0700960 page = alloc_buddy_huge_page(h, vma, addr);
Mel Gormana1e78772008-07-23 21:27:23 -0700961 if (!page) {
962 hugetlb_put_quota(inode->i_mapping, chg);
963 return ERR_PTR(-VM_FAULT_OOM);
964 }
965 }
966
967 set_page_refcounted(page);
968 set_page_private(page, (unsigned long) mapping);
969
Andi Kleena5516432008-07-23 21:27:41 -0700970 vma_commit_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700971
Adam Litke90d8b7e2007-11-14 16:59:42 -0800972 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800973}
974
Jon Tollefson53ba51d2008-07-23 21:27:52 -0700975__attribute__((weak)) int alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -0700976{
977 struct huge_bootmem_page *m;
978 int nr_nodes = nodes_weight(node_online_map);
979
980 while (nr_nodes) {
981 void *addr;
982
983 addr = __alloc_bootmem_node_nopanic(
984 NODE_DATA(h->hugetlb_next_nid),
985 huge_page_size(h), huge_page_size(h), 0);
986
987 if (addr) {
988 /*
989 * Use the beginning of the huge page to store the
990 * huge_bootmem_page struct (until gather_bootmem
991 * puts them into the mem_map).
992 */
993 m = addr;
994 if (m)
995 goto found;
996 }
997 hstate_next_node(h);
998 nr_nodes--;
999 }
1000 return 0;
1001
1002found:
1003 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1004 /* Put them into a private list first because mem_map is not up yet */
1005 list_add(&m->list, &huge_boot_pages);
1006 m->hstate = h;
1007 return 1;
1008}
1009
Andy Whitcroft18229df2008-11-06 12:53:27 -08001010static void prep_compound_huge_page(struct page *page, int order)
1011{
1012 if (unlikely(order > (MAX_ORDER - 1)))
1013 prep_compound_gigantic_page(page, order);
1014 else
1015 prep_compound_page(page, order);
1016}
1017
Andi Kleenaa888a72008-07-23 21:27:47 -07001018/* Put bootmem huge pages into the standard lists after mem_map is up */
1019static void __init gather_bootmem_prealloc(void)
1020{
1021 struct huge_bootmem_page *m;
1022
1023 list_for_each_entry(m, &huge_boot_pages, list) {
1024 struct page *page = virt_to_page(m);
1025 struct hstate *h = m->hstate;
1026 __ClearPageReserved(page);
1027 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001028 prep_compound_huge_page(page, h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001029 prep_new_huge_page(h, page, page_to_nid(page));
1030 }
1031}
1032
Andi Kleen8faa8b02008-07-23 21:27:48 -07001033static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034{
1035 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036
Andi Kleene5ff2152008-07-23 21:27:42 -07001037 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001038 if (h->order >= MAX_ORDER) {
1039 if (!alloc_bootmem_huge_page(h))
1040 break;
1041 } else if (!alloc_fresh_huge_page(h))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001044 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001045}
1046
1047static void __init hugetlb_init_hstates(void)
1048{
1049 struct hstate *h;
1050
1051 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001052 /* oversize hugepages were init'ed in early boot */
1053 if (h->order < MAX_ORDER)
1054 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001055 }
1056}
1057
Andi Kleen4abd32d2008-07-23 21:27:49 -07001058static char * __init memfmt(char *buf, unsigned long n)
1059{
1060 if (n >= (1UL << 30))
1061 sprintf(buf, "%lu GB", n >> 30);
1062 else if (n >= (1UL << 20))
1063 sprintf(buf, "%lu MB", n >> 20);
1064 else
1065 sprintf(buf, "%lu KB", n >> 10);
1066 return buf;
1067}
1068
Andi Kleene5ff2152008-07-23 21:27:42 -07001069static void __init report_hugepages(void)
1070{
1071 struct hstate *h;
1072
1073 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001074 char buf[32];
1075 printk(KERN_INFO "HugeTLB registered %s page size, "
1076 "pre-allocated %ld pages\n",
1077 memfmt(buf, huge_page_size(h)),
1078 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001079 }
1080}
1081
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082#ifdef CONFIG_HIGHMEM
Andi Kleena5516432008-07-23 21:27:41 -07001083static void try_to_free_low(struct hstate *h, unsigned long count)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001085 int i;
1086
Andi Kleenaa888a72008-07-23 21:27:47 -07001087 if (h->order >= MAX_ORDER)
1088 return;
1089
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 for (i = 0; i < MAX_NUMNODES; ++i) {
1091 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001092 struct list_head *freel = &h->hugepage_freelists[i];
1093 list_for_each_entry_safe(page, next, freel, lru) {
1094 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001095 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 if (PageHighMem(page))
1097 continue;
1098 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001099 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001100 h->free_huge_pages--;
1101 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 }
1103 }
1104}
1105#else
Andi Kleena5516432008-07-23 21:27:41 -07001106static inline void try_to_free_low(struct hstate *h, unsigned long count)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107{
1108}
1109#endif
1110
Andi Kleena5516432008-07-23 21:27:41 -07001111#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Andi Kleene5ff2152008-07-23 21:27:42 -07001112static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113{
Adam Litke7893d1d2007-10-16 01:26:18 -07001114 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115
Andi Kleenaa888a72008-07-23 21:27:47 -07001116 if (h->order >= MAX_ORDER)
1117 return h->max_huge_pages;
1118
Adam Litke7893d1d2007-10-16 01:26:18 -07001119 /*
1120 * Increase the pool size
1121 * First take pages out of surplus state. Then make up the
1122 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001123 *
1124 * We might race with alloc_buddy_huge_page() here and be unable
1125 * to convert a surplus huge page to a normal huge page. That is
1126 * not critical, though, it just means the overall size of the
1127 * pool might be one hugepage larger than it needs to be, but
1128 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001129 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001131 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1132 if (!adjust_pool_surplus(h, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001133 break;
1134 }
1135
Andi Kleena5516432008-07-23 21:27:41 -07001136 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001137 /*
1138 * If this allocation races such that we no longer need the
1139 * page, free_huge_page will handle it by freeing the page
1140 * and reducing the surplus.
1141 */
1142 spin_unlock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001143 ret = alloc_fresh_huge_page(h);
Adam Litke7893d1d2007-10-16 01:26:18 -07001144 spin_lock(&hugetlb_lock);
1145 if (!ret)
1146 goto out;
1147
1148 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001149
1150 /*
1151 * Decrease the pool size
1152 * First return free pages to the buddy allocator (being careful
1153 * to keep enough around to satisfy reservations). Then place
1154 * pages into surplus state as needed so the pool will shrink
1155 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001156 *
1157 * By placing pages into the surplus state independent of the
1158 * overcommit value, we are allowing the surplus pool size to
1159 * exceed overcommit. There are few sane options here. Since
1160 * alloc_buddy_huge_page() is checking the global counter,
1161 * though, we'll note that we're not allowed to exceed surplus
1162 * and won't grow the pool anywhere else. Not until one of the
1163 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001164 */
Andi Kleena5516432008-07-23 21:27:41 -07001165 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001166 min_count = max(count, min_count);
Andi Kleena5516432008-07-23 21:27:41 -07001167 try_to_free_low(h, min_count);
1168 while (min_count < persistent_huge_pages(h)) {
1169 struct page *page = dequeue_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 if (!page)
1171 break;
Andi Kleena5516432008-07-23 21:27:41 -07001172 update_and_free_page(h, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173 }
Andi Kleena5516432008-07-23 21:27:41 -07001174 while (count < persistent_huge_pages(h)) {
1175 if (!adjust_pool_surplus(h, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001176 break;
1177 }
1178out:
Andi Kleena5516432008-07-23 21:27:41 -07001179 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001181 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182}
1183
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001184#define HSTATE_ATTR_RO(_name) \
1185 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1186
1187#define HSTATE_ATTR(_name) \
1188 static struct kobj_attribute _name##_attr = \
1189 __ATTR(_name, 0644, _name##_show, _name##_store)
1190
1191static struct kobject *hugepages_kobj;
1192static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1193
1194static struct hstate *kobj_to_hstate(struct kobject *kobj)
1195{
1196 int i;
1197 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1198 if (hstate_kobjs[i] == kobj)
1199 return &hstates[i];
1200 BUG();
1201 return NULL;
1202}
1203
1204static ssize_t nr_hugepages_show(struct kobject *kobj,
1205 struct kobj_attribute *attr, char *buf)
1206{
1207 struct hstate *h = kobj_to_hstate(kobj);
1208 return sprintf(buf, "%lu\n", h->nr_huge_pages);
1209}
1210static ssize_t nr_hugepages_store(struct kobject *kobj,
1211 struct kobj_attribute *attr, const char *buf, size_t count)
1212{
1213 int err;
1214 unsigned long input;
1215 struct hstate *h = kobj_to_hstate(kobj);
1216
1217 err = strict_strtoul(buf, 10, &input);
1218 if (err)
1219 return 0;
1220
1221 h->max_huge_pages = set_max_huge_pages(h, input);
1222
1223 return count;
1224}
1225HSTATE_ATTR(nr_hugepages);
1226
1227static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1228 struct kobj_attribute *attr, char *buf)
1229{
1230 struct hstate *h = kobj_to_hstate(kobj);
1231 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1232}
1233static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1234 struct kobj_attribute *attr, const char *buf, size_t count)
1235{
1236 int err;
1237 unsigned long input;
1238 struct hstate *h = kobj_to_hstate(kobj);
1239
1240 err = strict_strtoul(buf, 10, &input);
1241 if (err)
1242 return 0;
1243
1244 spin_lock(&hugetlb_lock);
1245 h->nr_overcommit_huge_pages = input;
1246 spin_unlock(&hugetlb_lock);
1247
1248 return count;
1249}
1250HSTATE_ATTR(nr_overcommit_hugepages);
1251
1252static ssize_t free_hugepages_show(struct kobject *kobj,
1253 struct kobj_attribute *attr, char *buf)
1254{
1255 struct hstate *h = kobj_to_hstate(kobj);
1256 return sprintf(buf, "%lu\n", h->free_huge_pages);
1257}
1258HSTATE_ATTR_RO(free_hugepages);
1259
1260static ssize_t resv_hugepages_show(struct kobject *kobj,
1261 struct kobj_attribute *attr, char *buf)
1262{
1263 struct hstate *h = kobj_to_hstate(kobj);
1264 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1265}
1266HSTATE_ATTR_RO(resv_hugepages);
1267
1268static ssize_t surplus_hugepages_show(struct kobject *kobj,
1269 struct kobj_attribute *attr, char *buf)
1270{
1271 struct hstate *h = kobj_to_hstate(kobj);
1272 return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1273}
1274HSTATE_ATTR_RO(surplus_hugepages);
1275
1276static struct attribute *hstate_attrs[] = {
1277 &nr_hugepages_attr.attr,
1278 &nr_overcommit_hugepages_attr.attr,
1279 &free_hugepages_attr.attr,
1280 &resv_hugepages_attr.attr,
1281 &surplus_hugepages_attr.attr,
1282 NULL,
1283};
1284
1285static struct attribute_group hstate_attr_group = {
1286 .attrs = hstate_attrs,
1287};
1288
1289static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1290{
1291 int retval;
1292
1293 hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1294 hugepages_kobj);
1295 if (!hstate_kobjs[h - hstates])
1296 return -ENOMEM;
1297
1298 retval = sysfs_create_group(hstate_kobjs[h - hstates],
1299 &hstate_attr_group);
1300 if (retval)
1301 kobject_put(hstate_kobjs[h - hstates]);
1302
1303 return retval;
1304}
1305
1306static void __init hugetlb_sysfs_init(void)
1307{
1308 struct hstate *h;
1309 int err;
1310
1311 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1312 if (!hugepages_kobj)
1313 return;
1314
1315 for_each_hstate(h) {
1316 err = hugetlb_sysfs_add_hstate(h);
1317 if (err)
1318 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1319 h->name);
1320 }
1321}
1322
1323static void __exit hugetlb_exit(void)
1324{
1325 struct hstate *h;
1326
1327 for_each_hstate(h) {
1328 kobject_put(hstate_kobjs[h - hstates]);
1329 }
1330
1331 kobject_put(hugepages_kobj);
1332}
1333module_exit(hugetlb_exit);
1334
1335static int __init hugetlb_init(void)
1336{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001337 /* Some platform decide whether they support huge pages at boot
1338 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1339 * there is no such support
1340 */
1341 if (HPAGE_SHIFT == 0)
1342 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001343
Nick Piggine11bfbf2008-07-23 21:27:52 -07001344 if (!size_to_hstate(default_hstate_size)) {
1345 default_hstate_size = HPAGE_SIZE;
1346 if (!size_to_hstate(default_hstate_size))
1347 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001348 }
Nick Piggine11bfbf2008-07-23 21:27:52 -07001349 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1350 if (default_hstate_max_huge_pages)
1351 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001352
1353 hugetlb_init_hstates();
1354
Andi Kleenaa888a72008-07-23 21:27:47 -07001355 gather_bootmem_prealloc();
1356
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001357 report_hugepages();
1358
1359 hugetlb_sysfs_init();
1360
1361 return 0;
1362}
1363module_init(hugetlb_init);
1364
1365/* Should be called on processing a hugepagesz=... option */
1366void __init hugetlb_add_hstate(unsigned order)
1367{
1368 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001369 unsigned long i;
1370
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001371 if (size_to_hstate(PAGE_SIZE << order)) {
1372 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1373 return;
1374 }
1375 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1376 BUG_ON(order == 0);
1377 h = &hstates[max_hstate++];
1378 h->order = order;
1379 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001380 h->nr_huge_pages = 0;
1381 h->free_huge_pages = 0;
1382 for (i = 0; i < MAX_NUMNODES; ++i)
1383 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1384 h->hugetlb_next_nid = first_node(node_online_map);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001385 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1386 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001387
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001388 parsed_hstate = h;
1389}
1390
Nick Piggine11bfbf2008-07-23 21:27:52 -07001391static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001392{
1393 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001394 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001395
1396 /*
1397 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1398 * so this hugepages= parameter goes to the "default hstate".
1399 */
1400 if (!max_hstate)
1401 mhp = &default_hstate_max_huge_pages;
1402 else
1403 mhp = &parsed_hstate->max_huge_pages;
1404
Andi Kleen8faa8b02008-07-23 21:27:48 -07001405 if (mhp == last_mhp) {
1406 printk(KERN_WARNING "hugepages= specified twice without "
1407 "interleaving hugepagesz=, ignoring\n");
1408 return 1;
1409 }
1410
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001411 if (sscanf(s, "%lu", mhp) <= 0)
1412 *mhp = 0;
1413
Andi Kleen8faa8b02008-07-23 21:27:48 -07001414 /*
1415 * Global state is always initialized later in hugetlb_init.
1416 * But we need to allocate >= MAX_ORDER hstates here early to still
1417 * use the bootmem allocator.
1418 */
1419 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1420 hugetlb_hstate_alloc_pages(parsed_hstate);
1421
1422 last_mhp = mhp;
1423
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001424 return 1;
1425}
Nick Piggine11bfbf2008-07-23 21:27:52 -07001426__setup("hugepages=", hugetlb_nrpages_setup);
1427
1428static int __init hugetlb_default_setup(char *s)
1429{
1430 default_hstate_size = memparse(s, &s);
1431 return 1;
1432}
1433__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001434
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07001435static unsigned int cpuset_mems_nr(unsigned int *array)
1436{
1437 int node;
1438 unsigned int nr = 0;
1439
1440 for_each_node_mask(node, cpuset_current_mems_allowed)
1441 nr += array[node];
1442
1443 return nr;
1444}
1445
1446#ifdef CONFIG_SYSCTL
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1448 struct file *file, void __user *buffer,
1449 size_t *length, loff_t *ppos)
1450{
Andi Kleene5ff2152008-07-23 21:27:42 -07001451 struct hstate *h = &default_hstate;
1452 unsigned long tmp;
1453
1454 if (!write)
1455 tmp = h->max_huge_pages;
1456
1457 table->data = &tmp;
1458 table->maxlen = sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001460
1461 if (write)
1462 h->max_huge_pages = set_max_huge_pages(h, tmp);
1463
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 return 0;
1465}
Mel Gorman396faf02007-07-17 04:03:13 -07001466
1467int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1468 struct file *file, void __user *buffer,
1469 size_t *length, loff_t *ppos)
1470{
1471 proc_dointvec(table, write, file, buffer, length, ppos);
1472 if (hugepages_treat_as_movable)
1473 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1474 else
1475 htlb_alloc_mask = GFP_HIGHUSER;
1476 return 0;
1477}
1478
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001479int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1480 struct file *file, void __user *buffer,
1481 size_t *length, loff_t *ppos)
1482{
Andi Kleena5516432008-07-23 21:27:41 -07001483 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07001484 unsigned long tmp;
1485
1486 if (!write)
1487 tmp = h->nr_overcommit_huge_pages;
1488
1489 table->data = &tmp;
1490 table->maxlen = sizeof(unsigned long);
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001491 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001492
1493 if (write) {
1494 spin_lock(&hugetlb_lock);
1495 h->nr_overcommit_huge_pages = tmp;
1496 spin_unlock(&hugetlb_lock);
1497 }
1498
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001499 return 0;
1500}
1501
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502#endif /* CONFIG_SYSCTL */
1503
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001504void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505{
Andi Kleena5516432008-07-23 21:27:41 -07001506 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001507 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001508 "HugePages_Total: %5lu\n"
1509 "HugePages_Free: %5lu\n"
1510 "HugePages_Rsvd: %5lu\n"
1511 "HugePages_Surp: %5lu\n"
1512 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07001513 h->nr_huge_pages,
1514 h->free_huge_pages,
1515 h->resv_huge_pages,
1516 h->surplus_huge_pages,
1517 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518}
1519
1520int hugetlb_report_node_meminfo(int nid, char *buf)
1521{
Andi Kleena5516432008-07-23 21:27:41 -07001522 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523 return sprintf(buf,
1524 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07001525 "Node %d HugePages_Free: %5u\n"
1526 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07001527 nid, h->nr_huge_pages_node[nid],
1528 nid, h->free_huge_pages_node[nid],
1529 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530}
1531
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1533unsigned long hugetlb_total_pages(void)
1534{
Andi Kleena5516432008-07-23 21:27:41 -07001535 struct hstate *h = &default_hstate;
1536 return h->nr_huge_pages * pages_per_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538
Andi Kleena5516432008-07-23 21:27:41 -07001539static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001540{
1541 int ret = -ENOMEM;
1542
1543 spin_lock(&hugetlb_lock);
1544 /*
1545 * When cpuset is configured, it breaks the strict hugetlb page
1546 * reservation as the accounting is done on a global variable. Such
1547 * reservation is completely rubbish in the presence of cpuset because
1548 * the reservation is not checked against page availability for the
1549 * current cpuset. Application can still potentially OOM'ed by kernel
1550 * with lack of free htlb page in cpuset that the task is in.
1551 * Attempt to enforce strict accounting with cpuset is almost
1552 * impossible (or too ugly) because cpuset is too fluid that
1553 * task or memory node can be dynamically moved between cpusets.
1554 *
1555 * The change of semantics for shared hugetlb mapping with cpuset is
1556 * undesirable. However, in order to preserve some of the semantics,
1557 * we fall back to check against current free page availability as
1558 * a best attempt and hopefully to minimize the impact of changing
1559 * semantics that cpuset has.
1560 */
1561 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001562 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001563 goto out;
1564
Andi Kleena5516432008-07-23 21:27:41 -07001565 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1566 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001567 goto out;
1568 }
1569 }
1570
1571 ret = 0;
1572 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07001573 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001574
1575out:
1576 spin_unlock(&hugetlb_lock);
1577 return ret;
1578}
1579
Andy Whitcroft84afd992008-07-23 21:27:32 -07001580static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1581{
1582 struct resv_map *reservations = vma_resv_map(vma);
1583
1584 /*
1585 * This new VMA should share its siblings reservation map if present.
1586 * The VMA will only ever have a valid reservation map pointer where
1587 * it is being copied for another still existing VMA. As that VMA
1588 * has a reference to the reservation map it cannot dissappear until
1589 * after this open call completes. It is therefore safe to take a
1590 * new reference here without additional locking.
1591 */
1592 if (reservations)
1593 kref_get(&reservations->refs);
1594}
1595
Mel Gormana1e78772008-07-23 21:27:23 -07001596static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1597{
Andi Kleena5516432008-07-23 21:27:41 -07001598 struct hstate *h = hstate_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001599 struct resv_map *reservations = vma_resv_map(vma);
1600 unsigned long reserve;
1601 unsigned long start;
1602 unsigned long end;
1603
1604 if (reservations) {
Andi Kleena5516432008-07-23 21:27:41 -07001605 start = vma_hugecache_offset(h, vma, vma->vm_start);
1606 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001607
1608 reserve = (end - start) -
1609 region_count(&reservations->regions, start, end);
1610
1611 kref_put(&reservations->refs, resv_map_release);
1612
Adam Litke7251ff782008-07-23 21:27:59 -07001613 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07001614 hugetlb_acct_memory(h, -reserve);
Adam Litke7251ff782008-07-23 21:27:59 -07001615 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1616 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07001617 }
Mel Gormana1e78772008-07-23 21:27:23 -07001618}
1619
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620/*
1621 * We cannot handle pagefaults against hugetlb pages at all. They cause
1622 * handle_mm_fault() to try to instantiate regular-sized pages in the
1623 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1624 * this far.
1625 */
Nick Piggind0217ac2007-07-19 01:47:03 -07001626static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627{
1628 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07001629 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630}
1631
1632struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07001633 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07001634 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07001635 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636};
1637
David Gibson1e8f8892006-01-06 00:10:44 -08001638static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1639 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07001640{
1641 pte_t entry;
1642
David Gibson1e8f8892006-01-06 00:10:44 -08001643 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -07001644 entry =
1645 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1646 } else {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001647 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07001648 }
1649 entry = pte_mkyoung(entry);
1650 entry = pte_mkhuge(entry);
1651
1652 return entry;
1653}
1654
David Gibson1e8f8892006-01-06 00:10:44 -08001655static void set_huge_ptep_writable(struct vm_area_struct *vma,
1656 unsigned long address, pte_t *ptep)
1657{
1658 pte_t entry;
1659
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001660 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1661 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07001662 update_mmu_cache(vma, address, entry);
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07001663 }
David Gibson1e8f8892006-01-06 00:10:44 -08001664}
1665
1666
David Gibson63551ae2005-06-21 17:14:44 -07001667int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1668 struct vm_area_struct *vma)
1669{
1670 pte_t *src_pte, *dst_pte, entry;
1671 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07001672 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08001673 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07001674 struct hstate *h = hstate_vma(vma);
1675 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08001676
1677 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07001678
Andi Kleena5516432008-07-23 21:27:41 -07001679 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Hugh Dickinsc74df322005-10-29 18:16:23 -07001680 src_pte = huge_pte_offset(src, addr);
1681 if (!src_pte)
1682 continue;
Andi Kleena5516432008-07-23 21:27:41 -07001683 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07001684 if (!dst_pte)
1685 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08001686
1687 /* If the pagetables are shared don't copy or take references */
1688 if (dst_pte == src_pte)
1689 continue;
1690
Hugh Dickinsc74df322005-10-29 18:16:23 -07001691 spin_lock(&dst->page_table_lock);
Nick Piggin46478752008-06-05 22:45:57 -07001692 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001693 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08001694 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001695 huge_ptep_set_wrprotect(src, addr, src_pte);
1696 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07001697 ptepage = pte_page(entry);
1698 get_page(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07001699 set_huge_pte_at(dst, addr, dst_pte, entry);
1700 }
1701 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -07001702 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07001703 }
1704 return 0;
1705
1706nomem:
1707 return -ENOMEM;
1708}
1709
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001710void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001711 unsigned long end, struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07001712{
1713 struct mm_struct *mm = vma->vm_mm;
1714 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07001715 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07001716 pte_t pte;
1717 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001718 struct page *tmp;
Andi Kleena5516432008-07-23 21:27:41 -07001719 struct hstate *h = hstate_vma(vma);
1720 unsigned long sz = huge_page_size(h);
1721
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -08001722 /*
1723 * A page gathering list, protected by per file i_mmap_lock. The
1724 * lock is used to avoid list corruption from multiple unmapping
1725 * of the same page since we are using page->lru.
1726 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001727 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -07001728
1729 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07001730 BUG_ON(start & ~huge_page_mask(h));
1731 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07001732
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001733 mmu_notifier_invalidate_range_start(mm, start, end);
Hugh Dickins508034a2005-10-29 18:16:30 -07001734 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001735 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07001736 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07001737 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07001738 continue;
1739
Chen, Kenneth W39dde652006-12-06 20:32:03 -08001740 if (huge_pmd_unshare(mm, &address, ptep))
1741 continue;
1742
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001743 /*
1744 * If a reference page is supplied, it is because a specific
1745 * page is being unmapped, not a range. Ensure the page we
1746 * are about to unmap is the actual page of interest.
1747 */
1748 if (ref_page) {
1749 pte = huge_ptep_get(ptep);
1750 if (huge_pte_none(pte))
1751 continue;
1752 page = pte_page(pte);
1753 if (page != ref_page)
1754 continue;
1755
1756 /*
1757 * Mark the VMA as having unmapped its page so that
1758 * future faults in this VMA will fail rather than
1759 * looking like data was lost
1760 */
1761 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1762 }
1763
David Gibsonc7546f82005-08-05 11:59:35 -07001764 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001765 if (huge_pte_none(pte))
David Gibson63551ae2005-06-21 17:14:44 -07001766 continue;
David Gibsonc7546f82005-08-05 11:59:35 -07001767
David Gibson63551ae2005-06-21 17:14:44 -07001768 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -08001769 if (pte_dirty(pte))
1770 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001771 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -07001772 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -07001774 flush_tlb_range(vma, start, end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001775 mmu_notifier_invalidate_range_end(mm, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001776 list_for_each_entry_safe(page, tmp, &page_list, lru) {
1777 list_del(&page->lru);
1778 put_page(page);
1779 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780}
David Gibson63551ae2005-06-21 17:14:44 -07001781
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001782void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001783 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001784{
Andi Kleena137e1c2008-07-23 21:27:43 -07001785 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1786 __unmap_hugepage_range(vma, start, end, ref_page);
1787 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001788}
1789
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001790/*
1791 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1792 * mappping it owns the reserve page for. The intention is to unmap the page
1793 * from other VMAs and let the children be SIGKILLed if they are faulting the
1794 * same region.
1795 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07001796static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
1797 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001798{
Adam Litke75266742008-11-12 13:24:56 -08001799 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001800 struct vm_area_struct *iter_vma;
1801 struct address_space *mapping;
1802 struct prio_tree_iter iter;
1803 pgoff_t pgoff;
1804
1805 /*
1806 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1807 * from page cache lookup which is in HPAGE_SIZE units.
1808 */
Adam Litke75266742008-11-12 13:24:56 -08001809 address = address & huge_page_mask(h);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001810 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1811 + (vma->vm_pgoff >> PAGE_SHIFT);
1812 mapping = (struct address_space *)page_private(page);
1813
1814 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1815 /* Do not unmap the current VMA */
1816 if (iter_vma == vma)
1817 continue;
1818
1819 /*
1820 * Unmap the page from other VMAs without their own reserves.
1821 * They get marked to be SIGKILLed if they fault in these
1822 * areas. This is because a future no-page fault on this VMA
1823 * could insert a zeroed page instead of the data existing
1824 * from the time of fork. This would look like data corruption
1825 */
1826 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1827 unmap_hugepage_range(iter_vma,
Adam Litke75266742008-11-12 13:24:56 -08001828 address, address + huge_page_size(h),
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001829 page);
1830 }
1831
1832 return 1;
1833}
1834
David Gibson1e8f8892006-01-06 00:10:44 -08001835static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001836 unsigned long address, pte_t *ptep, pte_t pte,
1837 struct page *pagecache_page)
David Gibson1e8f8892006-01-06 00:10:44 -08001838{
Andi Kleena5516432008-07-23 21:27:41 -07001839 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08001840 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -08001841 int avoidcopy;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001842 int outside_reserve = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08001843
1844 old_page = pte_page(pte);
1845
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001846retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08001847 /* If no-one else is actually using this page, avoid the copy
1848 * and just make the page writable */
1849 avoidcopy = (page_count(old_page) == 1);
1850 if (avoidcopy) {
1851 set_huge_ptep_writable(vma, address, ptep);
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 /*
1856 * If the process that created a MAP_PRIVATE mapping is about to
1857 * perform a COW due to a shared page count, attempt to satisfy
1858 * the allocation without using the existing reserves. The pagecache
1859 * page is used to determine if the reserve at this address was
1860 * consumed or not. If reserves were used, a partial faulted mapping
1861 * at the time of fork() could consume its reserves on COW instead
1862 * of the full address range.
1863 */
1864 if (!(vma->vm_flags & VM_SHARED) &&
1865 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1866 old_page != pagecache_page)
1867 outside_reserve = 1;
1868
David Gibson1e8f8892006-01-06 00:10:44 -08001869 page_cache_get(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001870 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08001871
Adam Litke2fc39ce2007-11-14 16:59:39 -08001872 if (IS_ERR(new_page)) {
David Gibson1e8f8892006-01-06 00:10:44 -08001873 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001874
1875 /*
1876 * If a process owning a MAP_PRIVATE mapping fails to COW,
1877 * it is due to references held by a child and an insufficient
1878 * huge page pool. To guarantee the original mappers
1879 * reliability, unmap the page from child processes. The child
1880 * may get SIGKILLed if it later faults.
1881 */
1882 if (outside_reserve) {
1883 BUG_ON(huge_pte_none(pte));
1884 if (unmap_ref_private(mm, vma, old_page, address)) {
1885 BUG_ON(page_count(old_page) != 1);
1886 BUG_ON(huge_pte_none(pte));
1887 goto retry_avoidcopy;
1888 }
1889 WARN_ON_ONCE(1);
1890 }
1891
Adam Litke2fc39ce2007-11-14 16:59:39 -08001892 return -PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08001893 }
1894
1895 spin_unlock(&mm->page_table_lock);
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00001896 copy_huge_page(new_page, old_page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08001897 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08001898 spin_lock(&mm->page_table_lock);
1899
Andi Kleena5516432008-07-23 21:27:41 -07001900 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001901 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08001902 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07001903 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08001904 set_huge_pte_at(mm, address, ptep,
1905 make_huge_pte(vma, new_page, 1));
1906 /* Make the old page be freed below */
1907 new_page = old_page;
1908 }
1909 page_cache_release(new_page);
1910 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -07001911 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08001912}
1913
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001914/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07001915static struct page *hugetlbfs_pagecache_page(struct hstate *h,
1916 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001917{
1918 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07001919 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001920
1921 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07001922 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001923
1924 return find_lock_page(mapping, idx);
1925}
1926
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07001927static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
David Gibson1e8f8892006-01-06 00:10:44 -08001928 unsigned long address, pte_t *ptep, int write_access)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01001929{
Andi Kleena5516432008-07-23 21:27:41 -07001930 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01001931 int ret = VM_FAULT_SIGBUS;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07001932 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07001933 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07001934 struct page *page;
1935 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08001936 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -07001937
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001938 /*
1939 * Currently, we are forced to kill the process in the event the
1940 * original mapper has unmapped pages from the child due to a failed
1941 * COW. Warn that such a situation has occured as it may not be obvious
1942 */
1943 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1944 printk(KERN_WARNING
1945 "PID %d killed due to inadequate hugepage pool\n",
1946 current->pid);
1947 return ret;
1948 }
1949
Adam Litke4c887262005-10-29 18:16:46 -07001950 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07001951 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07001952
1953 /*
1954 * Use page lock to guard against racing truncation
1955 * before we get page_table_lock.
1956 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08001957retry:
1958 page = find_lock_page(mapping, idx);
1959 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07001960 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07001961 if (idx >= size)
1962 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001963 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08001964 if (IS_ERR(page)) {
1965 ret = -PTR_ERR(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08001966 goto out;
1967 }
Andi Kleena5516432008-07-23 21:27:41 -07001968 clear_huge_page(page, address, huge_page_size(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08001969 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01001970
Christoph Lameter6bda6662006-01-06 00:10:49 -08001971 if (vma->vm_flags & VM_SHARED) {
1972 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08001973 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08001974
1975 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1976 if (err) {
1977 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08001978 if (err == -EEXIST)
1979 goto retry;
1980 goto out;
1981 }
Ken Chen45c682a2007-11-14 16:59:44 -08001982
1983 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001984 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08001985 spin_unlock(&inode->i_lock);
Christoph Lameter6bda6662006-01-06 00:10:49 -08001986 } else
1987 lock_page(page);
1988 }
David Gibson1e8f8892006-01-06 00:10:44 -08001989
Andy Whitcroft57303d82008-08-12 15:08:47 -07001990 /*
1991 * If we are going to COW a private mapping later, we examine the
1992 * pending reservations for this page now. This will ensure that
1993 * any allocations necessary to record that reservation occur outside
1994 * the spinlock.
1995 */
1996 if (write_access && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07001997 if (vma_needs_reservation(h, vma, address) < 0) {
1998 ret = VM_FAULT_OOM;
1999 goto backout_unlocked;
2000 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002001
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002002 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002003 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002004 if (idx >= size)
2005 goto backout;
2006
Nick Piggin83c54072007-07-19 01:47:05 -07002007 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002008 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002009 goto backout;
2010
David Gibson1e8f8892006-01-06 00:10:44 -08002011 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2012 && (vma->vm_flags & VM_SHARED)));
2013 set_huge_pte_at(mm, address, ptep, new_pte);
2014
2015 if (write_access && !(vma->vm_flags & VM_SHARED)) {
2016 /* Optimization, do the COW without a second fault */
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002017 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
David Gibson1e8f8892006-01-06 00:10:44 -08002018 }
2019
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002020 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07002021 unlock_page(page);
2022out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002023 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002024
2025backout:
2026 spin_unlock(&mm->page_table_lock);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002027backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002028 unlock_page(page);
2029 put_page(page);
2030 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002031}
2032
Adam Litke86e52162006-01-06 00:10:43 -08002033int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2034 unsigned long address, int write_access)
2035{
2036 pte_t *ptep;
2037 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -08002038 int ret;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002039 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002040 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002041 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002042
Andi Kleena5516432008-07-23 21:27:41 -07002043 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002044 if (!ptep)
2045 return VM_FAULT_OOM;
2046
David Gibson3935baa2006-03-22 00:08:53 -08002047 /*
2048 * Serialize hugepage allocation and instantiation, so that we don't
2049 * get spurious allocation failures if two CPUs race to instantiate
2050 * the same page in the page cache.
2051 */
2052 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002053 entry = huge_ptep_get(ptep);
2054 if (huge_pte_none(entry)) {
David Gibson3935baa2006-03-22 00:08:53 -08002055 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
David Gibsonb4d1d992008-10-15 22:01:11 -07002056 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002057 }
Adam Litke86e52162006-01-06 00:10:43 -08002058
Nick Piggin83c54072007-07-19 01:47:05 -07002059 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002060
Andy Whitcroft57303d82008-08-12 15:08:47 -07002061 /*
2062 * If we are going to COW the mapping later, we examine the pending
2063 * reservations for this page now. This will ensure that any
2064 * allocations necessary to record that reservation occur outside the
2065 * spinlock. For private mappings, we also lookup the pagecache
2066 * page now as it is used to determine if a reservation has been
2067 * consumed.
2068 */
2069 if (write_access && !pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002070 if (vma_needs_reservation(h, vma, address) < 0) {
2071 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002072 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002073 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002074
2075 if (!(vma->vm_flags & VM_SHARED))
2076 pagecache_page = hugetlbfs_pagecache_page(h,
2077 vma, address);
2078 }
2079
David Gibson1e8f8892006-01-06 00:10:44 -08002080 spin_lock(&mm->page_table_lock);
2081 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002082 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2083 goto out_page_table_lock;
2084
2085
2086 if (write_access) {
2087 if (!pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002088 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2089 pagecache_page);
David Gibsonb4d1d992008-10-15 22:01:11 -07002090 goto out_page_table_lock;
2091 }
2092 entry = pte_mkdirty(entry);
2093 }
2094 entry = pte_mkyoung(entry);
2095 if (huge_ptep_set_access_flags(vma, address, ptep, entry, write_access))
2096 update_mmu_cache(vma, address, entry);
2097
2098out_page_table_lock:
David Gibson1e8f8892006-01-06 00:10:44 -08002099 spin_unlock(&mm->page_table_lock);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002100
2101 if (pagecache_page) {
2102 unlock_page(pagecache_page);
2103 put_page(pagecache_page);
2104 }
2105
David Gibsonb4d1d992008-10-15 22:01:11 -07002106out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08002107 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08002108
2109 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08002110}
2111
Andi Kleenceb86872008-07-23 21:27:50 -07002112/* Can be overriden by architectures */
2113__attribute__((weak)) struct page *
2114follow_huge_pud(struct mm_struct *mm, unsigned long address,
2115 pud_t *pud, int write)
2116{
2117 BUG();
2118 return NULL;
2119}
2120
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002121static int huge_zeropage_ok(pte_t *ptep, int write, int shared)
2122{
2123 if (!ptep || write || shared)
2124 return 0;
2125 else
2126 return huge_pte_none(huge_ptep_get(ptep));
2127}
2128
David Gibson63551ae2005-06-21 17:14:44 -07002129int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2130 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -08002131 unsigned long *position, int *length, int i,
2132 int write)
David Gibson63551ae2005-06-21 17:14:44 -07002133{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002134 unsigned long pfn_offset;
2135 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -07002136 int remainder = *length;
Andi Kleena5516432008-07-23 21:27:41 -07002137 struct hstate *h = hstate_vma(vma);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002138 int zeropage_ok = 0;
2139 int shared = vma->vm_flags & VM_SHARED;
David Gibson63551ae2005-06-21 17:14:44 -07002140
Hugh Dickins1c598272005-10-19 21:23:43 -07002141 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002142 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07002143 pte_t *pte;
2144 struct page *page;
2145
2146 /*
2147 * Some archs (sparc64, sh*) have multiple pte_ts to
2148 * each hugepage. We have to make * sure we get the
2149 * first, for the page indexing below to work.
2150 */
Andi Kleena5516432008-07-23 21:27:41 -07002151 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002152 if (huge_zeropage_ok(pte, write, shared))
2153 zeropage_ok = 1;
Adam Litke4c887262005-10-29 18:16:46 -07002154
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002155 if (!pte ||
2156 (huge_pte_none(huge_ptep_get(pte)) && !zeropage_ok) ||
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002157 (write && !pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07002158 int ret;
2159
2160 spin_unlock(&mm->page_table_lock);
Adam Litke5b23dbe2007-11-14 16:59:33 -08002161 ret = hugetlb_fault(mm, vma, vaddr, write);
Adam Litke4c887262005-10-29 18:16:46 -07002162 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07002163 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07002164 continue;
2165
2166 remainder = 0;
2167 if (!i)
2168 i = -EFAULT;
2169 break;
2170 }
David Gibson63551ae2005-06-21 17:14:44 -07002171
Andi Kleena5516432008-07-23 21:27:41 -07002172 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002173 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002174same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002175 if (pages) {
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002176 if (zeropage_ok)
2177 pages[i] = ZERO_PAGE(0);
2178 else
Andy Whitcroft69d177c2008-11-06 12:53:26 -08002179 pages[i] = mem_map_offset(page, pfn_offset);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002180 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002181 }
David Gibson63551ae2005-06-21 17:14:44 -07002182
2183 if (vmas)
2184 vmas[i] = vma;
2185
2186 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002187 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07002188 --remainder;
2189 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002190 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07002191 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002192 /*
2193 * We use pfn_offset to avoid touching the pageframes
2194 * of this compound page.
2195 */
2196 goto same_page;
2197 }
David Gibson63551ae2005-06-21 17:14:44 -07002198 }
Hugh Dickins1c598272005-10-19 21:23:43 -07002199 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002200 *length = remainder;
2201 *position = vaddr;
2202
2203 return i;
2204}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002205
2206void hugetlb_change_protection(struct vm_area_struct *vma,
2207 unsigned long address, unsigned long end, pgprot_t newprot)
2208{
2209 struct mm_struct *mm = vma->vm_mm;
2210 unsigned long start = address;
2211 pte_t *ptep;
2212 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07002213 struct hstate *h = hstate_vma(vma);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002214
2215 BUG_ON(address >= end);
2216 flush_cache_range(vma, address, end);
2217
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002218 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002219 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002220 for (; address < end; address += huge_page_size(h)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002221 ptep = huge_pte_offset(mm, address);
2222 if (!ptep)
2223 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002224 if (huge_pmd_unshare(mm, &address, ptep))
2225 continue;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002226 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002227 pte = huge_ptep_get_and_clear(mm, address, ptep);
2228 pte = pte_mkhuge(pte_modify(pte, newprot));
2229 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002230 }
2231 }
2232 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002233 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002234
2235 flush_tlb_range(vma, start, end);
2236}
2237
Mel Gormana1e78772008-07-23 21:27:23 -07002238int hugetlb_reserve_pages(struct inode *inode,
2239 long from, long to,
2240 struct vm_area_struct *vma)
Adam Litkee4e574b2007-10-16 01:26:19 -07002241{
2242 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07002243 struct hstate *h = hstate_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07002244
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002245 if (vma && vma->vm_flags & VM_NORESERVE)
2246 return 0;
2247
Mel Gormana1e78772008-07-23 21:27:23 -07002248 /*
2249 * Shared mappings base their reservation on the number of pages that
2250 * are already allocated on behalf of the file. Private mappings need
2251 * to reserve the full area even if read-only as mprotect() may be
2252 * called to make the mapping read-write. Assume !vma is a shm mapping
2253 */
2254 if (!vma || vma->vm_flags & VM_SHARED)
2255 chg = region_chg(&inode->i_mapping->private_list, from, to);
2256 else {
Andy Whitcroft84afd992008-07-23 21:27:32 -07002257 struct resv_map *resv_map = resv_map_alloc();
2258 if (!resv_map)
2259 return -ENOMEM;
2260
Mel Gormana1e78772008-07-23 21:27:23 -07002261 chg = to - from;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002262
2263 set_vma_resv_map(vma, resv_map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002264 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
Mel Gormana1e78772008-07-23 21:27:23 -07002265 }
2266
Adam Litkee4e574b2007-10-16 01:26:19 -07002267 if (chg < 0)
2268 return chg;
Ken Chen8a630112007-05-09 02:33:34 -07002269
Adam Litke90d8b7e2007-11-14 16:59:42 -08002270 if (hugetlb_get_quota(inode->i_mapping, chg))
2271 return -ENOSPC;
Andi Kleena5516432008-07-23 21:27:41 -07002272 ret = hugetlb_acct_memory(h, chg);
Ken Chen68842c92008-01-14 00:55:19 -08002273 if (ret < 0) {
2274 hugetlb_put_quota(inode->i_mapping, chg);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002275 return ret;
Ken Chen68842c92008-01-14 00:55:19 -08002276 }
Mel Gormana1e78772008-07-23 21:27:23 -07002277 if (!vma || vma->vm_flags & VM_SHARED)
2278 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002279 return 0;
2280}
2281
2282void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2283{
Andi Kleena5516432008-07-23 21:27:41 -07002284 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002285 long chg = region_truncate(&inode->i_mapping->private_list, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08002286
2287 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002288 inode->i_blocks -= blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002289 spin_unlock(&inode->i_lock);
2290
Adam Litke90d8b7e2007-11-14 16:59:42 -08002291 hugetlb_put_quota(inode->i_mapping, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07002292 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002293}