blob: 54d42b009dbeb5feb15c2fb08ac09239e0bf9feb [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
4 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080016#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080017#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070018#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070019#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.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>
Lee Schermerhorn9a305232009-12-14 17:58:25 -080027#include <linux/node.h>
Nick Piggin7835e982006-03-22 00:08:40 -080028#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070029
30const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
Mel Gorman396faf02007-07-17 04:03:13 -070031static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
32unsigned long hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070033
Andi Kleene5ff2152008-07-23 21:27:42 -070034static int max_hstate;
35unsigned int default_hstate_idx;
36struct hstate hstates[HUGE_MAX_HSTATE];
37
Jon Tollefson53ba51d2008-07-23 21:27:52 -070038__initdata LIST_HEAD(huge_boot_pages);
39
Andi Kleene5ff2152008-07-23 21:27:42 -070040/* for command line parsing */
41static struct hstate * __initdata parsed_hstate;
42static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070043static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070044
45#define for_each_hstate(h) \
46 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
Mel Gorman396faf02007-07-17 04:03:13 -070047
David Gibson3935baa2006-03-22 00:08:53 -080048/*
49 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
50 */
51static DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080052
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -070053/*
Andy Whitcroft96822902008-07-23 21:27:29 -070054 * Region tracking -- allows tracking of reservations and instantiated pages
55 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -070056 *
57 * The region data structures are protected by a combination of the mmap_sem
58 * and the hugetlb_instantion_mutex. To access or modify a region the caller
59 * must either hold the mmap_sem for write, or the mmap_sem for read and
60 * the hugetlb_instantiation mutex:
61 *
62 * down_write(&mm->mmap_sem);
63 * or
64 * down_read(&mm->mmap_sem);
65 * mutex_lock(&hugetlb_instantiation_mutex);
Andy Whitcroft96822902008-07-23 21:27:29 -070066 */
67struct file_region {
68 struct list_head link;
69 long from;
70 long to;
71};
72
73static long region_add(struct list_head *head, long f, long t)
74{
75 struct file_region *rg, *nrg, *trg;
76
77 /* Locate the region we are either in or before. */
78 list_for_each_entry(rg, head, link)
79 if (f <= rg->to)
80 break;
81
82 /* Round our left edge to the current segment if it encloses us. */
83 if (f > rg->from)
84 f = rg->from;
85
86 /* Check for and consume any regions we now overlap with. */
87 nrg = rg;
88 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
89 if (&rg->link == head)
90 break;
91 if (rg->from > t)
92 break;
93
94 /* If this area reaches higher then extend our area to
95 * include it completely. If this is not the first area
96 * which we intend to reuse, free it. */
97 if (rg->to > t)
98 t = rg->to;
99 if (rg != nrg) {
100 list_del(&rg->link);
101 kfree(rg);
102 }
103 }
104 nrg->from = f;
105 nrg->to = t;
106 return 0;
107}
108
109static long region_chg(struct list_head *head, long f, long t)
110{
111 struct file_region *rg, *nrg;
112 long chg = 0;
113
114 /* Locate the region we are before or in. */
115 list_for_each_entry(rg, head, link)
116 if (f <= rg->to)
117 break;
118
119 /* If we are below the current region then a new region is required.
120 * Subtle, allocate a new region at the position but make it zero
121 * size such that we can guarantee to record the reservation. */
122 if (&rg->link == head || t < rg->from) {
123 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
124 if (!nrg)
125 return -ENOMEM;
126 nrg->from = f;
127 nrg->to = f;
128 INIT_LIST_HEAD(&nrg->link);
129 list_add(&nrg->link, rg->link.prev);
130
131 return t - f;
132 }
133
134 /* Round our left edge to the current segment if it encloses us. */
135 if (f > rg->from)
136 f = rg->from;
137 chg = t - f;
138
139 /* Check for and consume any regions we now overlap with. */
140 list_for_each_entry(rg, rg->link.prev, link) {
141 if (&rg->link == head)
142 break;
143 if (rg->from > t)
144 return chg;
145
146 /* We overlap with this area, if it extends futher than
147 * us then we must extend ourselves. Account for its
148 * existing reservation. */
149 if (rg->to > t) {
150 chg += rg->to - t;
151 t = rg->to;
152 }
153 chg -= rg->to - rg->from;
154 }
155 return chg;
156}
157
158static long region_truncate(struct list_head *head, long end)
159{
160 struct file_region *rg, *trg;
161 long chg = 0;
162
163 /* Locate the region we are either in or before. */
164 list_for_each_entry(rg, head, link)
165 if (end <= rg->to)
166 break;
167 if (&rg->link == head)
168 return 0;
169
170 /* If we are in the middle of a region then adjust it. */
171 if (end > rg->from) {
172 chg = rg->to - end;
173 rg->to = end;
174 rg = list_entry(rg->link.next, typeof(*rg), link);
175 }
176
177 /* Drop any remaining regions. */
178 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
179 if (&rg->link == head)
180 break;
181 chg += rg->to - rg->from;
182 list_del(&rg->link);
183 kfree(rg);
184 }
185 return chg;
186}
187
Andy Whitcroft84afd992008-07-23 21:27:32 -0700188static long region_count(struct list_head *head, long f, long t)
189{
190 struct file_region *rg;
191 long chg = 0;
192
193 /* Locate each segment we overlap with, and count that overlap. */
194 list_for_each_entry(rg, head, link) {
195 int seg_from;
196 int seg_to;
197
198 if (rg->to <= f)
199 continue;
200 if (rg->from >= t)
201 break;
202
203 seg_from = max(rg->from, f);
204 seg_to = min(rg->to, t);
205
206 chg += seg_to - seg_from;
207 }
208
209 return chg;
210}
211
Andy Whitcroft96822902008-07-23 21:27:29 -0700212/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700213 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700214 * the mapping, in pagecache page units; huge pages here.
215 */
Andi Kleena5516432008-07-23 21:27:41 -0700216static pgoff_t vma_hugecache_offset(struct hstate *h,
217 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700218{
Andi Kleena5516432008-07-23 21:27:41 -0700219 return ((address - vma->vm_start) >> huge_page_shift(h)) +
220 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700221}
222
Andy Whitcroft84afd992008-07-23 21:27:32 -0700223/*
Mel Gorman08fba692009-01-06 14:38:53 -0800224 * Return the size of the pages allocated when backing a VMA. In the majority
225 * cases this will be same size as used by the page table entries.
226 */
227unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
228{
229 struct hstate *hstate;
230
231 if (!is_vm_hugetlb_page(vma))
232 return PAGE_SIZE;
233
234 hstate = hstate_vma(vma);
235
236 return 1UL << (hstate->order + PAGE_SHIFT);
237}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200238EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800239
240/*
Mel Gorman33402892009-01-06 14:38:54 -0800241 * Return the page size being used by the MMU to back a VMA. In the majority
242 * of cases, the page size used by the kernel matches the MMU size. On
243 * architectures where it differs, an architecture-specific version of this
244 * function is required.
245 */
246#ifndef vma_mmu_pagesize
247unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
248{
249 return vma_kernel_pagesize(vma);
250}
251#endif
252
253/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700254 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
255 * bits of the reservation map pointer, which are always clear due to
256 * alignment.
257 */
258#define HPAGE_RESV_OWNER (1UL << 0)
259#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700260#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700261
Mel Gormana1e78772008-07-23 21:27:23 -0700262/*
263 * These helpers are used to track how many pages are reserved for
264 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
265 * is guaranteed to have their future faults succeed.
266 *
267 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
268 * the reserve counters are updated with the hugetlb_lock held. It is safe
269 * to reset the VMA at fork() time as it is not in use yet and there is no
270 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700271 *
272 * The private mapping reservation is represented in a subtly different
273 * manner to a shared mapping. A shared mapping has a region map associated
274 * with the underlying file, this region map represents the backing file
275 * pages which have ever had a reservation assigned which this persists even
276 * after the page is instantiated. A private mapping has a region map
277 * associated with the original mmap which is attached to all VMAs which
278 * reference it, this region map represents those offsets which have consumed
279 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700280 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700281static unsigned long get_vma_private_data(struct vm_area_struct *vma)
282{
283 return (unsigned long)vma->vm_private_data;
284}
285
286static void set_vma_private_data(struct vm_area_struct *vma,
287 unsigned long value)
288{
289 vma->vm_private_data = (void *)value;
290}
291
Andy Whitcroft84afd992008-07-23 21:27:32 -0700292struct resv_map {
293 struct kref refs;
294 struct list_head regions;
295};
296
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700297static struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700298{
299 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
300 if (!resv_map)
301 return NULL;
302
303 kref_init(&resv_map->refs);
304 INIT_LIST_HEAD(&resv_map->regions);
305
306 return resv_map;
307}
308
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700309static void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700310{
311 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
312
313 /* Clear out any active regions before we release the map. */
314 region_truncate(&resv_map->regions, 0);
315 kfree(resv_map);
316}
317
318static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700319{
320 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700321 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft84afd992008-07-23 21:27:32 -0700322 return (struct resv_map *)(get_vma_private_data(vma) &
323 ~HPAGE_RESV_MASK);
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700324 return NULL;
Mel Gormana1e78772008-07-23 21:27:23 -0700325}
326
Andy Whitcroft84afd992008-07-23 21:27:32 -0700327static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700328{
329 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700330 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Mel Gormana1e78772008-07-23 21:27:23 -0700331
Andy Whitcroft84afd992008-07-23 21:27:32 -0700332 set_vma_private_data(vma, (get_vma_private_data(vma) &
333 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700334}
335
336static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
337{
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700338 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700339 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700340
341 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700342}
343
344static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
345{
346 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700347
348 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700349}
350
351/* Decrement the reserved pages in the hugepage pool by one */
Andi Kleena5516432008-07-23 21:27:41 -0700352static void decrement_hugepage_resv_vma(struct hstate *h,
353 struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700354{
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700355 if (vma->vm_flags & VM_NORESERVE)
356 return;
357
Mel Gormanf83a2752009-05-28 14:34:40 -0700358 if (vma->vm_flags & VM_MAYSHARE) {
Mel Gormana1e78772008-07-23 21:27:23 -0700359 /* Shared mappings always use reserves */
Andi Kleena5516432008-07-23 21:27:41 -0700360 h->resv_huge_pages--;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700361 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Mel Gormana1e78772008-07-23 21:27:23 -0700362 /*
363 * Only the process that called mmap() has reserves for
364 * private mappings.
365 */
Andi Kleena5516432008-07-23 21:27:41 -0700366 h->resv_huge_pages--;
Mel Gormana1e78772008-07-23 21:27:23 -0700367 }
368}
369
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700370/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700371void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
372{
373 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700374 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700375 vma->vm_private_data = (void *)0;
376}
377
378/* Returns true if the VMA has associated reserve pages */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700379static int vma_has_reserves(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700380{
Mel Gormanf83a2752009-05-28 14:34:40 -0700381 if (vma->vm_flags & VM_MAYSHARE)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700382 return 1;
383 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
384 return 1;
385 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700386}
387
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800388static void clear_gigantic_page(struct page *page,
389 unsigned long addr, unsigned long sz)
390{
391 int i;
392 struct page *p = page;
393
394 might_sleep();
395 for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
396 cond_resched();
397 clear_user_highpage(p, addr + i * PAGE_SIZE);
398 }
399}
Andi Kleena5516432008-07-23 21:27:41 -0700400static void clear_huge_page(struct page *page,
401 unsigned long addr, unsigned long sz)
David Gibson79ac6ba2006-03-22 00:08:51 -0800402{
403 int i;
404
Andrea Arcangeli74dbdd22010-01-08 14:43:05 -0800405 if (unlikely(sz/PAGE_SIZE > MAX_ORDER_NR_PAGES)) {
Hannes Ederebdd4ae2009-01-06 14:39:58 -0800406 clear_gigantic_page(page, addr, sz);
407 return;
408 }
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800409
David Gibson79ac6ba2006-03-22 00:08:51 -0800410 might_sleep();
Andi Kleena5516432008-07-23 21:27:41 -0700411 for (i = 0; i < sz/PAGE_SIZE; i++) {
David Gibson79ac6ba2006-03-22 00:08:51 -0800412 cond_resched();
Ralf Baechle281e0e32007-10-01 01:20:10 -0700413 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
David Gibson79ac6ba2006-03-22 00:08:51 -0800414 }
415}
416
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800417static void copy_gigantic_page(struct page *dst, struct page *src,
418 unsigned long addr, struct vm_area_struct *vma)
419{
420 int i;
421 struct hstate *h = hstate_vma(vma);
422 struct page *dst_base = dst;
423 struct page *src_base = src;
424 might_sleep();
425 for (i = 0; i < pages_per_huge_page(h); ) {
426 cond_resched();
427 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
428
429 i++;
430 dst = mem_map_next(dst, dst_base, i);
431 src = mem_map_next(src, src_base, i);
432 }
433}
David Gibson79ac6ba2006-03-22 00:08:51 -0800434static void copy_huge_page(struct page *dst, struct page *src,
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000435 unsigned long addr, struct vm_area_struct *vma)
David Gibson79ac6ba2006-03-22 00:08:51 -0800436{
437 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700438 struct hstate *h = hstate_vma(vma);
David Gibson79ac6ba2006-03-22 00:08:51 -0800439
Hannes Ederebdd4ae2009-01-06 14:39:58 -0800440 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
441 copy_gigantic_page(dst, src, addr, vma);
442 return;
443 }
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800444
David Gibson79ac6ba2006-03-22 00:08:51 -0800445 might_sleep();
Andi Kleena5516432008-07-23 21:27:41 -0700446 for (i = 0; i < pages_per_huge_page(h); i++) {
David Gibson79ac6ba2006-03-22 00:08:51 -0800447 cond_resched();
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000448 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
David Gibson79ac6ba2006-03-22 00:08:51 -0800449 }
450}
451
Andi Kleena5516432008-07-23 21:27:41 -0700452static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453{
454 int nid = page_to_nid(page);
Andi Kleena5516432008-07-23 21:27:41 -0700455 list_add(&page->lru, &h->hugepage_freelists[nid]);
456 h->free_huge_pages++;
457 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458}
459
Andi Kleena5516432008-07-23 21:27:41 -0700460static struct page *dequeue_huge_page_vma(struct hstate *h,
461 struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700462 unsigned long address, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463{
Nishanth Aravamudan31a5c6e2007-07-15 23:38:02 -0700464 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700466 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700467 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700468 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700469 struct zone *zone;
470 struct zoneref *z;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471
Miao Xiec0ff7452010-05-24 14:32:08 -0700472 get_mems_allowed();
473 zonelist = huge_zonelist(vma, address,
474 htlb_alloc_mask, &mpol, &nodemask);
Mel Gormana1e78772008-07-23 21:27:23 -0700475 /*
476 * A child process with MAP_PRIVATE mappings created by their parent
477 * have no page reserves. This check ensures that reservations are
478 * not "stolen". The child may still get SIGKILLed
479 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700480 if (!vma_has_reserves(vma) &&
Andi Kleena5516432008-07-23 21:27:41 -0700481 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700482 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700483
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700484 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700485 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700486 goto err;;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700487
Mel Gorman19770b32008-04-28 02:12:18 -0700488 for_each_zone_zonelist_nodemask(zone, z, zonelist,
489 MAX_NR_ZONES - 1, nodemask) {
Mel Gorman54a6eb52008-04-28 02:12:16 -0700490 nid = zone_to_nid(zone);
491 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
Andi Kleena5516432008-07-23 21:27:41 -0700492 !list_empty(&h->hugepage_freelists[nid])) {
493 page = list_entry(h->hugepage_freelists[nid].next,
Andrew Morton3abf7af2007-07-19 01:49:08 -0700494 struct page, lru);
495 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700496 h->free_huge_pages--;
497 h->free_huge_pages_node[nid]--;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700498
499 if (!avoid_reserve)
Andi Kleena5516432008-07-23 21:27:41 -0700500 decrement_hugepage_resv_vma(h, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700501
Ken Chen5ab3ee72007-07-23 18:44:00 -0700502 break;
Andrew Morton3abf7af2007-07-19 01:49:08 -0700503 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 }
Miao Xiec0ff7452010-05-24 14:32:08 -0700505err:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700506 mpol_cond_put(mpol);
Miao Xiec0ff7452010-05-24 14:32:08 -0700507 put_mems_allowed();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508 return page;
509}
510
Andi Kleena5516432008-07-23 21:27:41 -0700511static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700512{
513 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700514
Andy Whitcroft18229df2008-11-06 12:53:27 -0800515 VM_BUG_ON(h->order >= MAX_ORDER);
516
Andi Kleena5516432008-07-23 21:27:41 -0700517 h->nr_huge_pages--;
518 h->nr_huge_pages_node[page_to_nid(page)]--;
519 for (i = 0; i < pages_per_huge_page(h); i++) {
Adam Litke6af2acb2007-10-16 01:26:16 -0700520 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
521 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
522 1 << PG_private | 1<< PG_writeback);
523 }
524 set_compound_page_dtor(page, NULL);
525 set_page_refcounted(page);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700526 arch_release_hugepage(page);
Andi Kleena5516432008-07-23 21:27:41 -0700527 __free_pages(page, huge_page_order(h));
Adam Litke6af2acb2007-10-16 01:26:16 -0700528}
529
Andi Kleene5ff2152008-07-23 21:27:42 -0700530struct hstate *size_to_hstate(unsigned long size)
531{
532 struct hstate *h;
533
534 for_each_hstate(h) {
535 if (huge_page_size(h) == size)
536 return h;
537 }
538 return NULL;
539}
540
David Gibson27a85ef2006-03-22 00:08:56 -0800541static void free_huge_page(struct page *page)
542{
Andi Kleena5516432008-07-23 21:27:41 -0700543 /*
544 * Can't pass hstate in here because it is called from the
545 * compound page destructor.
546 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700547 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700548 int nid = page_to_nid(page);
Adam Litkec79fb752007-11-14 16:59:38 -0800549 struct address_space *mapping;
David Gibson27a85ef2006-03-22 00:08:56 -0800550
Adam Litkec79fb752007-11-14 16:59:38 -0800551 mapping = (struct address_space *) page_private(page);
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800552 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -0400553 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -0700554 BUG_ON(page_count(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800555 INIT_LIST_HEAD(&page->lru);
556
557 spin_lock(&hugetlb_lock);
Andi Kleenaa888a72008-07-23 21:27:47 -0700558 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
Andi Kleena5516432008-07-23 21:27:41 -0700559 update_and_free_page(h, page);
560 h->surplus_huge_pages--;
561 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700562 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700563 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700564 }
David Gibson27a85ef2006-03-22 00:08:56 -0800565 spin_unlock(&hugetlb_lock);
Adam Litkec79fb752007-11-14 16:59:38 -0800566 if (mapping)
Adam Litke9a119c02007-11-14 16:59:41 -0800567 hugetlb_put_quota(mapping, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800568}
569
Andi Kleena5516432008-07-23 21:27:41 -0700570static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700571{
572 set_compound_page_dtor(page, free_huge_page);
573 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700574 h->nr_huge_pages++;
575 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700576 spin_unlock(&hugetlb_lock);
577 put_page(page); /* free it into the hugepage allocator */
578}
579
Wu Fengguang20a03072009-06-16 15:32:22 -0700580static void prep_compound_gigantic_page(struct page *page, unsigned long order)
581{
582 int i;
583 int nr_pages = 1 << order;
584 struct page *p = page + 1;
585
586 /* we rely on prep_new_huge_page to set the destructor */
587 set_compound_order(page, order);
588 __SetPageHead(page);
589 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
590 __SetPageTail(p);
591 p->first_page = page;
592 }
593}
594
595int PageHuge(struct page *page)
596{
597 compound_page_dtor *dtor;
598
599 if (!PageCompound(page))
600 return 0;
601
602 page = compound_head(page);
603 dtor = get_compound_page_dtor(page);
604
605 return dtor == free_huge_page;
606}
607
Andi Kleena5516432008-07-23 21:27:41 -0700608static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700611
Andi Kleenaa888a72008-07-23 21:27:47 -0700612 if (h->order >= MAX_ORDER)
613 return NULL;
614
Mel Gorman6484eb32009-06-16 15:31:54 -0700615 page = alloc_pages_exact_node(nid,
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700616 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
617 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700618 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700620 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700621 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700622 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700623 }
Andi Kleena5516432008-07-23 21:27:41 -0700624 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700626
627 return page;
628}
629
Andi Kleen5ced66c2008-07-23 21:27:45 -0700630/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800631 * common helper functions for hstate_next_node_to_{alloc|free}.
632 * We may have allocated or freed a huge page based on a different
633 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
634 * be outside of *nodes_allowed. Ensure that we use an allowed
635 * node for alloc or free.
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800636 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800637static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800638{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800639 nid = next_node(nid, *nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800640 if (nid == MAX_NUMNODES)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800641 nid = first_node(*nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800642 VM_BUG_ON(nid >= MAX_NUMNODES);
643
644 return nid;
645}
646
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800647static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
Andi Kleen5ced66c2008-07-23 21:27:45 -0700648{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800649 if (!node_isset(nid, *nodes_allowed))
650 nid = next_node_allowed(nid, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800651 return nid;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700652}
653
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800654/*
655 * returns the previously saved node ["this node"] from which to
656 * allocate a persistent huge page for the pool and advance the
657 * next node from which to allocate, handling wrap at end of node
658 * mask.
659 */
660static int hstate_next_node_to_alloc(struct hstate *h,
661 nodemask_t *nodes_allowed)
662{
663 int nid;
664
665 VM_BUG_ON(!nodes_allowed);
666
667 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
668 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
669
670 return nid;
671}
672
673static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700674{
675 struct page *page;
676 int start_nid;
677 int next_nid;
678 int ret = 0;
679
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800680 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700681 next_nid = start_nid;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700682
683 do {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700684 page = alloc_fresh_huge_page_node(h, next_nid);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800685 if (page) {
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700686 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800687 break;
688 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800689 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800690 } while (next_nid != start_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700691
Adam Litke3b116302008-04-28 02:13:06 -0700692 if (ret)
693 count_vm_event(HTLB_BUDDY_PGALLOC);
694 else
695 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
696
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700697 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698}
699
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700700/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800701 * helper for free_pool_huge_page() - return the previously saved
702 * node ["this node"] from which to free a huge page. Advance the
703 * next node id whether or not we find a free huge page to free so
704 * that the next attempt to free addresses the next node.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700705 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800706static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700707{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800708 int nid;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800709
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800710 VM_BUG_ON(!nodes_allowed);
711
712 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
713 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
714
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800715 return nid;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700716}
717
718/*
719 * Free huge page from pool from next node to free.
720 * Attempt to keep persistent huge pages more or less
721 * balanced over allowed nodes.
722 * Called with hugetlb_lock locked.
723 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800724static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
725 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700726{
727 int start_nid;
728 int next_nid;
729 int ret = 0;
730
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800731 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700732 next_nid = start_nid;
733
734 do {
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700735 /*
736 * If we're returning unused surplus pages, only examine
737 * nodes with surplus pages.
738 */
739 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
740 !list_empty(&h->hugepage_freelists[next_nid])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700741 struct page *page =
742 list_entry(h->hugepage_freelists[next_nid].next,
743 struct page, lru);
744 list_del(&page->lru);
745 h->free_huge_pages--;
746 h->free_huge_pages_node[next_nid]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700747 if (acct_surplus) {
748 h->surplus_huge_pages--;
749 h->surplus_huge_pages_node[next_nid]--;
750 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700751 update_and_free_page(h, page);
752 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800753 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700754 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800755 next_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800756 } while (next_nid != start_nid);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700757
758 return ret;
759}
760
Andi Kleena5516432008-07-23 21:27:41 -0700761static struct page *alloc_buddy_huge_page(struct hstate *h,
762 struct vm_area_struct *vma, unsigned long address)
Adam Litke7893d1d2007-10-16 01:26:18 -0700763{
764 struct page *page;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800765 unsigned int nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700766
Andi Kleenaa888a72008-07-23 21:27:47 -0700767 if (h->order >= MAX_ORDER)
768 return NULL;
769
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800770 /*
771 * Assume we will successfully allocate the surplus page to
772 * prevent racing processes from causing the surplus to exceed
773 * overcommit
774 *
775 * This however introduces a different race, where a process B
776 * tries to grow the static hugepage pool while alloc_pages() is
777 * called by process A. B will only examine the per-node
778 * counters in determining if surplus huge pages can be
779 * converted to normal huge pages in adjust_pool_surplus(). A
780 * won't be able to increment the per-node counter, until the
781 * lock is dropped by B, but B doesn't drop hugetlb_lock until
782 * no more huge pages can be converted from surplus to normal
783 * state (and doesn't try to convert again). Thus, we have a
784 * case where a surplus huge page exists, the pool is grown, and
785 * the surplus huge page still exists after, even though it
786 * should just have been converted to a normal huge page. This
787 * does not leak memory, though, as the hugepage will be freed
788 * once it is out of use. It also does not allow the counters to
789 * go out of whack in adjust_pool_surplus() as we don't modify
790 * the node values until we've gotten the hugepage and only the
791 * per-node value is checked there.
792 */
793 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700794 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800795 spin_unlock(&hugetlb_lock);
796 return NULL;
797 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700798 h->nr_huge_pages++;
799 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800800 }
801 spin_unlock(&hugetlb_lock);
802
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700803 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
804 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700805 huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800806
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700807 if (page && arch_prepare_hugepage(page)) {
808 __free_pages(page, huge_page_order(h));
809 return NULL;
810 }
811
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800812 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700813 if (page) {
Adam Litke2668db92008-03-10 11:43:50 -0700814 /*
815 * This page is now managed by the hugetlb allocator and has
816 * no users -- drop the buddy allocator's reference.
817 */
818 put_page_testzero(page);
819 VM_BUG_ON(page_count(page));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800820 nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700821 set_compound_page_dtor(page, free_huge_page);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800822 /*
823 * We incremented the global counters already
824 */
Andi Kleena5516432008-07-23 21:27:41 -0700825 h->nr_huge_pages_node[nid]++;
826 h->surplus_huge_pages_node[nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700827 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800828 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700829 h->nr_huge_pages--;
830 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700831 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700832 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800833 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700834
835 return page;
836}
837
Adam Litkee4e574b2007-10-16 01:26:19 -0700838/*
839 * Increase the hugetlb pool such that it can accomodate a reservation
840 * of size 'delta'.
841 */
Andi Kleena5516432008-07-23 21:27:41 -0700842static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -0700843{
844 struct list_head surplus_list;
845 struct page *page, *tmp;
846 int ret, i;
847 int needed, allocated;
848
Andi Kleena5516432008-07-23 21:27:41 -0700849 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800850 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -0700851 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700852 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800853 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700854
855 allocated = 0;
856 INIT_LIST_HEAD(&surplus_list);
857
858 ret = -ENOMEM;
859retry:
860 spin_unlock(&hugetlb_lock);
861 for (i = 0; i < needed; i++) {
Andi Kleena5516432008-07-23 21:27:41 -0700862 page = alloc_buddy_huge_page(h, NULL, 0);
Adam Litkee4e574b2007-10-16 01:26:19 -0700863 if (!page) {
864 /*
865 * We were not able to allocate enough pages to
866 * satisfy the entire reservation so we free what
867 * we've allocated so far.
868 */
869 spin_lock(&hugetlb_lock);
870 needed = 0;
871 goto free;
872 }
873
874 list_add(&page->lru, &surplus_list);
875 }
876 allocated += needed;
877
878 /*
879 * After retaking hugetlb_lock, we need to recalculate 'needed'
880 * because either resv_huge_pages or free_huge_pages may have changed.
881 */
882 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700883 needed = (h->resv_huge_pages + delta) -
884 (h->free_huge_pages + allocated);
Adam Litkee4e574b2007-10-16 01:26:19 -0700885 if (needed > 0)
886 goto retry;
887
888 /*
889 * The surplus_list now contains _at_least_ the number of extra pages
890 * needed to accomodate the reservation. Add the appropriate number
891 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -0800892 * allocator. Commit the entire reservation here to prevent another
893 * process from stealing the pages as they are added to the pool but
894 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -0700895 */
896 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -0700897 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700898 ret = 0;
899free:
Adam Litke19fc3f02008-04-28 02:12:20 -0700900 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -0700901 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -0700902 if ((--needed) < 0)
903 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700904 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700905 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -0700906 }
907
908 /* Free unnecessary surplus pages to the buddy allocator */
909 if (!list_empty(&surplus_list)) {
910 spin_unlock(&hugetlb_lock);
911 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
912 list_del(&page->lru);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700913 /*
Adam Litke2668db92008-03-10 11:43:50 -0700914 * The page has a reference count of zero already, so
915 * call free_huge_page directly instead of using
916 * put_page. This must be done with hugetlb_lock
Adam Litkeaf767cb2007-10-16 01:26:25 -0700917 * unlocked which is safe because free_huge_page takes
918 * hugetlb_lock before deciding how to free the page.
919 */
Adam Litke2668db92008-03-10 11:43:50 -0700920 free_huge_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700921 }
Adam Litke19fc3f02008-04-28 02:12:20 -0700922 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -0700923 }
924
925 return ret;
926}
927
928/*
929 * When releasing a hugetlb pool reservation, any surplus pages that were
930 * allocated to satisfy the reservation must be explicitly freed if they were
931 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700932 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -0700933 */
Andi Kleena5516432008-07-23 21:27:41 -0700934static void return_unused_surplus_pages(struct hstate *h,
935 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -0700936{
Adam Litkee4e574b2007-10-16 01:26:19 -0700937 unsigned long nr_pages;
938
Adam Litkeac09b3a2008-03-04 14:29:38 -0800939 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -0700940 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800941
Andi Kleenaa888a72008-07-23 21:27:47 -0700942 /* Cannot return gigantic pages currently */
943 if (h->order >= MAX_ORDER)
944 return;
945
Andi Kleena5516432008-07-23 21:27:41 -0700946 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -0700947
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700948 /*
949 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -0800950 * evenly across all nodes with memory. Iterate across these nodes
951 * until we can no longer free unreserved surplus pages. This occurs
952 * when the nodes with surplus pages have no free pages.
953 * free_pool_huge_page() will balance the the freed pages across the
954 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700955 */
956 while (nr_pages--) {
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -0800957 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700958 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700959 }
960}
961
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700962/*
963 * Determine if the huge page at addr within the vma has an associated
964 * reservation. Where it does not we will need to logically increase
965 * reservation and actually increase quota before an allocation can occur.
966 * Where any new reservation would be required the reservation change is
967 * prepared, but not committed. Once the page has been quota'd allocated
968 * an instantiated the change should be committed via vma_commit_reservation.
969 * No action is required on failure.
970 */
Roel Kluine2f17d92009-03-31 15:23:15 -0700971static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -0700972 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700973{
974 struct address_space *mapping = vma->vm_file->f_mapping;
975 struct inode *inode = mapping->host;
976
Mel Gormanf83a2752009-05-28 14:34:40 -0700977 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -0700978 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700979 return region_chg(&inode->i_mapping->private_list,
980 idx, idx + 1);
981
Andy Whitcroft84afd992008-07-23 21:27:32 -0700982 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
983 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700984
Andy Whitcroft84afd992008-07-23 21:27:32 -0700985 } else {
Roel Kluine2f17d92009-03-31 15:23:15 -0700986 long err;
Andi Kleena5516432008-07-23 21:27:41 -0700987 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700988 struct resv_map *reservations = vma_resv_map(vma);
989
990 err = region_chg(&reservations->regions, idx, idx + 1);
991 if (err < 0)
992 return err;
993 return 0;
994 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700995}
Andi Kleena5516432008-07-23 21:27:41 -0700996static void vma_commit_reservation(struct hstate *h,
997 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700998{
999 struct address_space *mapping = vma->vm_file->f_mapping;
1000 struct inode *inode = mapping->host;
1001
Mel Gormanf83a2752009-05-28 14:34:40 -07001002 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001003 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001004 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001005
1006 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -07001007 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001008 struct resv_map *reservations = vma_resv_map(vma);
1009
1010 /* Mark this page used in the map. */
1011 region_add(&reservations->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001012 }
1013}
1014
David Gibson27a85ef2006-03-22 00:08:56 -08001015static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001016 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017{
Andi Kleena5516432008-07-23 21:27:41 -07001018 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001019 struct page *page;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001020 struct address_space *mapping = vma->vm_file->f_mapping;
Mel Gormana1e78772008-07-23 21:27:23 -07001021 struct inode *inode = mapping->host;
Roel Kluine2f17d92009-03-31 15:23:15 -07001022 long chg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001023
Mel Gormana1e78772008-07-23 21:27:23 -07001024 /*
1025 * Processes that did not create the mapping will have no reserves and
1026 * will not have accounted against quota. Check that the quota can be
1027 * made before satisfying the allocation
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001028 * MAP_NORESERVE mappings may also need pages and quota allocated
1029 * if no reserve mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001030 */
Andi Kleena5516432008-07-23 21:27:41 -07001031 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001032 if (chg < 0)
1033 return ERR_PTR(chg);
1034 if (chg)
Mel Gormana1e78772008-07-23 21:27:23 -07001035 if (hugetlb_get_quota(inode->i_mapping, chg))
1036 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001037
1038 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001039 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
Mel Gormana1e78772008-07-23 21:27:23 -07001040 spin_unlock(&hugetlb_lock);
1041
1042 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07001043 page = alloc_buddy_huge_page(h, vma, addr);
Mel Gormana1e78772008-07-23 21:27:23 -07001044 if (!page) {
1045 hugetlb_put_quota(inode->i_mapping, chg);
Mel Gorman4a6018f2010-05-11 14:06:53 -07001046 return ERR_PTR(-VM_FAULT_SIGBUS);
Mel Gormana1e78772008-07-23 21:27:23 -07001047 }
1048 }
1049
1050 set_page_refcounted(page);
1051 set_page_private(page, (unsigned long) mapping);
1052
Andi Kleena5516432008-07-23 21:27:41 -07001053 vma_commit_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001054
Adam Litke90d8b7e2007-11-14 16:59:42 -08001055 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001056}
1057
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001058int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001059{
1060 struct huge_bootmem_page *m;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001061 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
Andi Kleenaa888a72008-07-23 21:27:47 -07001062
1063 while (nr_nodes) {
1064 void *addr;
1065
1066 addr = __alloc_bootmem_node_nopanic(
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001067 NODE_DATA(hstate_next_node_to_alloc(h,
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001068 &node_states[N_HIGH_MEMORY])),
Andi Kleenaa888a72008-07-23 21:27:47 -07001069 huge_page_size(h), huge_page_size(h), 0);
1070
1071 if (addr) {
1072 /*
1073 * Use the beginning of the huge page to store the
1074 * huge_bootmem_page struct (until gather_bootmem
1075 * puts them into the mem_map).
1076 */
1077 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001078 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001079 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001080 nr_nodes--;
1081 }
1082 return 0;
1083
1084found:
1085 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1086 /* Put them into a private list first because mem_map is not up yet */
1087 list_add(&m->list, &huge_boot_pages);
1088 m->hstate = h;
1089 return 1;
1090}
1091
Andy Whitcroft18229df2008-11-06 12:53:27 -08001092static void prep_compound_huge_page(struct page *page, int order)
1093{
1094 if (unlikely(order > (MAX_ORDER - 1)))
1095 prep_compound_gigantic_page(page, order);
1096 else
1097 prep_compound_page(page, order);
1098}
1099
Andi Kleenaa888a72008-07-23 21:27:47 -07001100/* Put bootmem huge pages into the standard lists after mem_map is up */
1101static void __init gather_bootmem_prealloc(void)
1102{
1103 struct huge_bootmem_page *m;
1104
1105 list_for_each_entry(m, &huge_boot_pages, list) {
1106 struct page *page = virt_to_page(m);
1107 struct hstate *h = m->hstate;
1108 __ClearPageReserved(page);
1109 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001110 prep_compound_huge_page(page, h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001111 prep_new_huge_page(h, page, page_to_nid(page));
1112 }
1113}
1114
Andi Kleen8faa8b02008-07-23 21:27:48 -07001115static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116{
1117 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118
Andi Kleene5ff2152008-07-23 21:27:42 -07001119 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001120 if (h->order >= MAX_ORDER) {
1121 if (!alloc_bootmem_huge_page(h))
1122 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001123 } else if (!alloc_fresh_huge_page(h,
1124 &node_states[N_HIGH_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001127 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001128}
1129
1130static void __init hugetlb_init_hstates(void)
1131{
1132 struct hstate *h;
1133
1134 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001135 /* oversize hugepages were init'ed in early boot */
1136 if (h->order < MAX_ORDER)
1137 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001138 }
1139}
1140
Andi Kleen4abd32d2008-07-23 21:27:49 -07001141static char * __init memfmt(char *buf, unsigned long n)
1142{
1143 if (n >= (1UL << 30))
1144 sprintf(buf, "%lu GB", n >> 30);
1145 else if (n >= (1UL << 20))
1146 sprintf(buf, "%lu MB", n >> 20);
1147 else
1148 sprintf(buf, "%lu KB", n >> 10);
1149 return buf;
1150}
1151
Andi Kleene5ff2152008-07-23 21:27:42 -07001152static void __init report_hugepages(void)
1153{
1154 struct hstate *h;
1155
1156 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001157 char buf[32];
1158 printk(KERN_INFO "HugeTLB registered %s page size, "
1159 "pre-allocated %ld pages\n",
1160 memfmt(buf, huge_page_size(h)),
1161 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001162 }
1163}
1164
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001166static void try_to_free_low(struct hstate *h, unsigned long count,
1167 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001169 int i;
1170
Andi Kleenaa888a72008-07-23 21:27:47 -07001171 if (h->order >= MAX_ORDER)
1172 return;
1173
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001174 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001176 struct list_head *freel = &h->hugepage_freelists[i];
1177 list_for_each_entry_safe(page, next, freel, lru) {
1178 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001179 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 if (PageHighMem(page))
1181 continue;
1182 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001183 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001184 h->free_huge_pages--;
1185 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 }
1187 }
1188}
1189#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001190static inline void try_to_free_low(struct hstate *h, unsigned long count,
1191 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192{
1193}
1194#endif
1195
Wu Fengguang20a03072009-06-16 15:32:22 -07001196/*
1197 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1198 * balanced by operating on them in a round-robin fashion.
1199 * Returns 1 if an adjustment was made.
1200 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001201static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1202 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001203{
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001204 int start_nid, next_nid;
Wu Fengguang20a03072009-06-16 15:32:22 -07001205 int ret = 0;
1206
1207 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001208
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001209 if (delta < 0)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001210 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001211 else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001212 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001213 next_nid = start_nid;
1214
1215 do {
1216 int nid = next_nid;
1217 if (delta < 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001218 /*
1219 * To shrink on this node, there must be a surplus page
1220 */
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001221 if (!h->surplus_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001222 next_nid = hstate_next_node_to_alloc(h,
1223 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001224 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001225 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001226 }
1227 if (delta > 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001228 /*
1229 * Surplus cannot exceed the total number of pages
1230 */
1231 if (h->surplus_huge_pages_node[nid] >=
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001232 h->nr_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001233 next_nid = hstate_next_node_to_free(h,
1234 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001235 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001236 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001237 }
Wu Fengguang20a03072009-06-16 15:32:22 -07001238
1239 h->surplus_huge_pages += delta;
1240 h->surplus_huge_pages_node[nid] += delta;
1241 ret = 1;
1242 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001243 } while (next_nid != start_nid);
Wu Fengguang20a03072009-06-16 15:32:22 -07001244
Wu Fengguang20a03072009-06-16 15:32:22 -07001245 return ret;
1246}
1247
Andi Kleena5516432008-07-23 21:27:41 -07001248#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001249static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1250 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251{
Adam Litke7893d1d2007-10-16 01:26:18 -07001252 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253
Andi Kleenaa888a72008-07-23 21:27:47 -07001254 if (h->order >= MAX_ORDER)
1255 return h->max_huge_pages;
1256
Adam Litke7893d1d2007-10-16 01:26:18 -07001257 /*
1258 * Increase the pool size
1259 * First take pages out of surplus state. Then make up the
1260 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001261 *
1262 * We might race with alloc_buddy_huge_page() here and be unable
1263 * to convert a surplus huge page to a normal huge page. That is
1264 * not critical, though, it just means the overall size of the
1265 * pool might be one hugepage larger than it needs to be, but
1266 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001267 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001269 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001270 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001271 break;
1272 }
1273
Andi Kleena5516432008-07-23 21:27:41 -07001274 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001275 /*
1276 * If this allocation races such that we no longer need the
1277 * page, free_huge_page will handle it by freeing the page
1278 * and reducing the surplus.
1279 */
1280 spin_unlock(&hugetlb_lock);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001281 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001282 spin_lock(&hugetlb_lock);
1283 if (!ret)
1284 goto out;
1285
Mel Gorman536240f2009-12-14 17:59:56 -08001286 /* Bail for signals. Probably ctrl-c from user */
1287 if (signal_pending(current))
1288 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001289 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001290
1291 /*
1292 * Decrease the pool size
1293 * First return free pages to the buddy allocator (being careful
1294 * to keep enough around to satisfy reservations). Then place
1295 * pages into surplus state as needed so the pool will shrink
1296 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001297 *
1298 * By placing pages into the surplus state independent of the
1299 * overcommit value, we are allowing the surplus pool size to
1300 * exceed overcommit. There are few sane options here. Since
1301 * alloc_buddy_huge_page() is checking the global counter,
1302 * though, we'll note that we're not allowed to exceed surplus
1303 * and won't grow the pool anywhere else. Not until one of the
1304 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001305 */
Andi Kleena5516432008-07-23 21:27:41 -07001306 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001307 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001308 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001309 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001310 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312 }
Andi Kleena5516432008-07-23 21:27:41 -07001313 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001314 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001315 break;
1316 }
1317out:
Andi Kleena5516432008-07-23 21:27:41 -07001318 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001320 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321}
1322
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001323#define HSTATE_ATTR_RO(_name) \
1324 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1325
1326#define HSTATE_ATTR(_name) \
1327 static struct kobj_attribute _name##_attr = \
1328 __ATTR(_name, 0644, _name##_show, _name##_store)
1329
1330static struct kobject *hugepages_kobj;
1331static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1332
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001333static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1334
1335static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001336{
1337 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001338
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001339 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001340 if (hstate_kobjs[i] == kobj) {
1341 if (nidp)
1342 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001343 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001344 }
1345
1346 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001347}
1348
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001349static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001350 struct kobj_attribute *attr, char *buf)
1351{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001352 struct hstate *h;
1353 unsigned long nr_huge_pages;
1354 int nid;
1355
1356 h = kobj_to_hstate(kobj, &nid);
1357 if (nid == NUMA_NO_NODE)
1358 nr_huge_pages = h->nr_huge_pages;
1359 else
1360 nr_huge_pages = h->nr_huge_pages_node[nid];
1361
1362 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001363}
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001364static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1365 struct kobject *kobj, struct kobj_attribute *attr,
1366 const char *buf, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001367{
1368 int err;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001369 int nid;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001370 unsigned long count;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001371 struct hstate *h;
David Rientjesbad44b52009-12-14 17:58:38 -08001372 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001373
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001374 err = strict_strtoul(buf, 10, &count);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001375 if (err)
1376 return 0;
1377
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001378 h = kobj_to_hstate(kobj, &nid);
1379 if (nid == NUMA_NO_NODE) {
1380 /*
1381 * global hstate attribute
1382 */
1383 if (!(obey_mempolicy &&
1384 init_nodemask_of_mempolicy(nodes_allowed))) {
1385 NODEMASK_FREE(nodes_allowed);
1386 nodes_allowed = &node_states[N_HIGH_MEMORY];
1387 }
1388 } else if (nodes_allowed) {
1389 /*
1390 * per node hstate attribute: adjust count to global,
1391 * but restrict alloc/free to the specified node.
1392 */
1393 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1394 init_nodemask_of_node(nodes_allowed, nid);
1395 } else
1396 nodes_allowed = &node_states[N_HIGH_MEMORY];
1397
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001398 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001399
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001400 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001401 NODEMASK_FREE(nodes_allowed);
1402
1403 return len;
1404}
1405
1406static ssize_t nr_hugepages_show(struct kobject *kobj,
1407 struct kobj_attribute *attr, char *buf)
1408{
1409 return nr_hugepages_show_common(kobj, attr, buf);
1410}
1411
1412static ssize_t nr_hugepages_store(struct kobject *kobj,
1413 struct kobj_attribute *attr, const char *buf, size_t len)
1414{
1415 return nr_hugepages_store_common(false, kobj, attr, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001416}
1417HSTATE_ATTR(nr_hugepages);
1418
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001419#ifdef CONFIG_NUMA
1420
1421/*
1422 * hstate attribute for optionally mempolicy-based constraint on persistent
1423 * huge page alloc/free.
1424 */
1425static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1426 struct kobj_attribute *attr, char *buf)
1427{
1428 return nr_hugepages_show_common(kobj, attr, buf);
1429}
1430
1431static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1432 struct kobj_attribute *attr, const char *buf, size_t len)
1433{
1434 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1435}
1436HSTATE_ATTR(nr_hugepages_mempolicy);
1437#endif
1438
1439
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001440static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1441 struct kobj_attribute *attr, char *buf)
1442{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001443 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001444 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1445}
1446static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1447 struct kobj_attribute *attr, const char *buf, size_t count)
1448{
1449 int err;
1450 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001451 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001452
1453 err = strict_strtoul(buf, 10, &input);
1454 if (err)
1455 return 0;
1456
1457 spin_lock(&hugetlb_lock);
1458 h->nr_overcommit_huge_pages = input;
1459 spin_unlock(&hugetlb_lock);
1460
1461 return count;
1462}
1463HSTATE_ATTR(nr_overcommit_hugepages);
1464
1465static ssize_t free_hugepages_show(struct kobject *kobj,
1466 struct kobj_attribute *attr, char *buf)
1467{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001468 struct hstate *h;
1469 unsigned long free_huge_pages;
1470 int nid;
1471
1472 h = kobj_to_hstate(kobj, &nid);
1473 if (nid == NUMA_NO_NODE)
1474 free_huge_pages = h->free_huge_pages;
1475 else
1476 free_huge_pages = h->free_huge_pages_node[nid];
1477
1478 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001479}
1480HSTATE_ATTR_RO(free_hugepages);
1481
1482static ssize_t resv_hugepages_show(struct kobject *kobj,
1483 struct kobj_attribute *attr, char *buf)
1484{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001485 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001486 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1487}
1488HSTATE_ATTR_RO(resv_hugepages);
1489
1490static ssize_t surplus_hugepages_show(struct kobject *kobj,
1491 struct kobj_attribute *attr, char *buf)
1492{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001493 struct hstate *h;
1494 unsigned long surplus_huge_pages;
1495 int nid;
1496
1497 h = kobj_to_hstate(kobj, &nid);
1498 if (nid == NUMA_NO_NODE)
1499 surplus_huge_pages = h->surplus_huge_pages;
1500 else
1501 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1502
1503 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001504}
1505HSTATE_ATTR_RO(surplus_hugepages);
1506
1507static struct attribute *hstate_attrs[] = {
1508 &nr_hugepages_attr.attr,
1509 &nr_overcommit_hugepages_attr.attr,
1510 &free_hugepages_attr.attr,
1511 &resv_hugepages_attr.attr,
1512 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001513#ifdef CONFIG_NUMA
1514 &nr_hugepages_mempolicy_attr.attr,
1515#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001516 NULL,
1517};
1518
1519static struct attribute_group hstate_attr_group = {
1520 .attrs = hstate_attrs,
1521};
1522
Jeff Mahoney094e9532010-02-02 13:44:14 -08001523static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1524 struct kobject **hstate_kobjs,
1525 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001526{
1527 int retval;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001528 int hi = h - hstates;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001529
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001530 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1531 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001532 return -ENOMEM;
1533
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001534 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001535 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001536 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001537
1538 return retval;
1539}
1540
1541static void __init hugetlb_sysfs_init(void)
1542{
1543 struct hstate *h;
1544 int err;
1545
1546 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1547 if (!hugepages_kobj)
1548 return;
1549
1550 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001551 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1552 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001553 if (err)
1554 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1555 h->name);
1556 }
1557}
1558
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001559#ifdef CONFIG_NUMA
1560
1561/*
1562 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1563 * with node sysdevs in node_devices[] using a parallel array. The array
1564 * index of a node sysdev or _hstate == node id.
1565 * This is here to avoid any static dependency of the node sysdev driver, in
1566 * the base kernel, on the hugetlb module.
1567 */
1568struct node_hstate {
1569 struct kobject *hugepages_kobj;
1570 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1571};
1572struct node_hstate node_hstates[MAX_NUMNODES];
1573
1574/*
1575 * A subset of global hstate attributes for node sysdevs
1576 */
1577static struct attribute *per_node_hstate_attrs[] = {
1578 &nr_hugepages_attr.attr,
1579 &free_hugepages_attr.attr,
1580 &surplus_hugepages_attr.attr,
1581 NULL,
1582};
1583
1584static struct attribute_group per_node_hstate_attr_group = {
1585 .attrs = per_node_hstate_attrs,
1586};
1587
1588/*
1589 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1590 * Returns node id via non-NULL nidp.
1591 */
1592static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1593{
1594 int nid;
1595
1596 for (nid = 0; nid < nr_node_ids; nid++) {
1597 struct node_hstate *nhs = &node_hstates[nid];
1598 int i;
1599 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1600 if (nhs->hstate_kobjs[i] == kobj) {
1601 if (nidp)
1602 *nidp = nid;
1603 return &hstates[i];
1604 }
1605 }
1606
1607 BUG();
1608 return NULL;
1609}
1610
1611/*
1612 * Unregister hstate attributes from a single node sysdev.
1613 * No-op if no hstate attributes attached.
1614 */
1615void hugetlb_unregister_node(struct node *node)
1616{
1617 struct hstate *h;
1618 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1619
1620 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001621 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001622
1623 for_each_hstate(h)
1624 if (nhs->hstate_kobjs[h - hstates]) {
1625 kobject_put(nhs->hstate_kobjs[h - hstates]);
1626 nhs->hstate_kobjs[h - hstates] = NULL;
1627 }
1628
1629 kobject_put(nhs->hugepages_kobj);
1630 nhs->hugepages_kobj = NULL;
1631}
1632
1633/*
1634 * hugetlb module exit: unregister hstate attributes from node sysdevs
1635 * that have them.
1636 */
1637static void hugetlb_unregister_all_nodes(void)
1638{
1639 int nid;
1640
1641 /*
1642 * disable node sysdev registrations.
1643 */
1644 register_hugetlbfs_with_node(NULL, NULL);
1645
1646 /*
1647 * remove hstate attributes from any nodes that have them.
1648 */
1649 for (nid = 0; nid < nr_node_ids; nid++)
1650 hugetlb_unregister_node(&node_devices[nid]);
1651}
1652
1653/*
1654 * Register hstate attributes for a single node sysdev.
1655 * No-op if attributes already registered.
1656 */
1657void hugetlb_register_node(struct node *node)
1658{
1659 struct hstate *h;
1660 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1661 int err;
1662
1663 if (nhs->hugepages_kobj)
1664 return; /* already allocated */
1665
1666 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1667 &node->sysdev.kobj);
1668 if (!nhs->hugepages_kobj)
1669 return;
1670
1671 for_each_hstate(h) {
1672 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1673 nhs->hstate_kobjs,
1674 &per_node_hstate_attr_group);
1675 if (err) {
1676 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1677 " for node %d\n",
1678 h->name, node->sysdev.id);
1679 hugetlb_unregister_node(node);
1680 break;
1681 }
1682 }
1683}
1684
1685/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001686 * hugetlb init time: register hstate attributes for all registered node
1687 * sysdevs of nodes that have memory. All on-line nodes should have
1688 * registered their associated sysdev by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001689 */
1690static void hugetlb_register_all_nodes(void)
1691{
1692 int nid;
1693
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001694 for_each_node_state(nid, N_HIGH_MEMORY) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001695 struct node *node = &node_devices[nid];
1696 if (node->sysdev.id == nid)
1697 hugetlb_register_node(node);
1698 }
1699
1700 /*
1701 * Let the node sysdev driver know we're here so it can
1702 * [un]register hstate attributes on node hotplug.
1703 */
1704 register_hugetlbfs_with_node(hugetlb_register_node,
1705 hugetlb_unregister_node);
1706}
1707#else /* !CONFIG_NUMA */
1708
1709static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1710{
1711 BUG();
1712 if (nidp)
1713 *nidp = -1;
1714 return NULL;
1715}
1716
1717static void hugetlb_unregister_all_nodes(void) { }
1718
1719static void hugetlb_register_all_nodes(void) { }
1720
1721#endif
1722
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001723static void __exit hugetlb_exit(void)
1724{
1725 struct hstate *h;
1726
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001727 hugetlb_unregister_all_nodes();
1728
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001729 for_each_hstate(h) {
1730 kobject_put(hstate_kobjs[h - hstates]);
1731 }
1732
1733 kobject_put(hugepages_kobj);
1734}
1735module_exit(hugetlb_exit);
1736
1737static int __init hugetlb_init(void)
1738{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001739 /* Some platform decide whether they support huge pages at boot
1740 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1741 * there is no such support
1742 */
1743 if (HPAGE_SHIFT == 0)
1744 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001745
Nick Piggine11bfbf2008-07-23 21:27:52 -07001746 if (!size_to_hstate(default_hstate_size)) {
1747 default_hstate_size = HPAGE_SIZE;
1748 if (!size_to_hstate(default_hstate_size))
1749 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001750 }
Nick Piggine11bfbf2008-07-23 21:27:52 -07001751 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1752 if (default_hstate_max_huge_pages)
1753 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001754
1755 hugetlb_init_hstates();
1756
Andi Kleenaa888a72008-07-23 21:27:47 -07001757 gather_bootmem_prealloc();
1758
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001759 report_hugepages();
1760
1761 hugetlb_sysfs_init();
1762
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001763 hugetlb_register_all_nodes();
1764
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001765 return 0;
1766}
1767module_init(hugetlb_init);
1768
1769/* Should be called on processing a hugepagesz=... option */
1770void __init hugetlb_add_hstate(unsigned order)
1771{
1772 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001773 unsigned long i;
1774
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001775 if (size_to_hstate(PAGE_SIZE << order)) {
1776 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1777 return;
1778 }
1779 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1780 BUG_ON(order == 0);
1781 h = &hstates[max_hstate++];
1782 h->order = order;
1783 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001784 h->nr_huge_pages = 0;
1785 h->free_huge_pages = 0;
1786 for (i = 0; i < MAX_NUMNODES; ++i)
1787 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001788 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1789 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001790 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1791 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001792
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001793 parsed_hstate = h;
1794}
1795
Nick Piggine11bfbf2008-07-23 21:27:52 -07001796static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001797{
1798 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001799 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001800
1801 /*
1802 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1803 * so this hugepages= parameter goes to the "default hstate".
1804 */
1805 if (!max_hstate)
1806 mhp = &default_hstate_max_huge_pages;
1807 else
1808 mhp = &parsed_hstate->max_huge_pages;
1809
Andi Kleen8faa8b02008-07-23 21:27:48 -07001810 if (mhp == last_mhp) {
1811 printk(KERN_WARNING "hugepages= specified twice without "
1812 "interleaving hugepagesz=, ignoring\n");
1813 return 1;
1814 }
1815
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001816 if (sscanf(s, "%lu", mhp) <= 0)
1817 *mhp = 0;
1818
Andi Kleen8faa8b02008-07-23 21:27:48 -07001819 /*
1820 * Global state is always initialized later in hugetlb_init.
1821 * But we need to allocate >= MAX_ORDER hstates here early to still
1822 * use the bootmem allocator.
1823 */
1824 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1825 hugetlb_hstate_alloc_pages(parsed_hstate);
1826
1827 last_mhp = mhp;
1828
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001829 return 1;
1830}
Nick Piggine11bfbf2008-07-23 21:27:52 -07001831__setup("hugepages=", hugetlb_nrpages_setup);
1832
1833static int __init hugetlb_default_setup(char *s)
1834{
1835 default_hstate_size = memparse(s, &s);
1836 return 1;
1837}
1838__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001839
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07001840static unsigned int cpuset_mems_nr(unsigned int *array)
1841{
1842 int node;
1843 unsigned int nr = 0;
1844
1845 for_each_node_mask(node, cpuset_current_mems_allowed)
1846 nr += array[node];
1847
1848 return nr;
1849}
1850
1851#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001852static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1853 struct ctl_table *table, int write,
1854 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855{
Andi Kleene5ff2152008-07-23 21:27:42 -07001856 struct hstate *h = &default_hstate;
1857 unsigned long tmp;
1858
1859 if (!write)
1860 tmp = h->max_huge_pages;
1861
1862 table->data = &tmp;
1863 table->maxlen = sizeof(unsigned long);
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001864 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001865
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001866 if (write) {
David Rientjesbad44b52009-12-14 17:58:38 -08001867 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1868 GFP_KERNEL | __GFP_NORETRY);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001869 if (!(obey_mempolicy &&
1870 init_nodemask_of_mempolicy(nodes_allowed))) {
1871 NODEMASK_FREE(nodes_allowed);
1872 nodes_allowed = &node_states[N_HIGH_MEMORY];
1873 }
1874 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1875
1876 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1877 NODEMASK_FREE(nodes_allowed);
1878 }
Andi Kleene5ff2152008-07-23 21:27:42 -07001879
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880 return 0;
1881}
Mel Gorman396faf02007-07-17 04:03:13 -07001882
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001883int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1884 void __user *buffer, size_t *length, loff_t *ppos)
1885{
1886
1887 return hugetlb_sysctl_handler_common(false, table, write,
1888 buffer, length, ppos);
1889}
1890
1891#ifdef CONFIG_NUMA
1892int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1893 void __user *buffer, size_t *length, loff_t *ppos)
1894{
1895 return hugetlb_sysctl_handler_common(true, table, write,
1896 buffer, length, ppos);
1897}
1898#endif /* CONFIG_NUMA */
1899
Mel Gorman396faf02007-07-17 04:03:13 -07001900int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001901 void __user *buffer,
Mel Gorman396faf02007-07-17 04:03:13 -07001902 size_t *length, loff_t *ppos)
1903{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001904 proc_dointvec(table, write, buffer, length, ppos);
Mel Gorman396faf02007-07-17 04:03:13 -07001905 if (hugepages_treat_as_movable)
1906 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1907 else
1908 htlb_alloc_mask = GFP_HIGHUSER;
1909 return 0;
1910}
1911
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001912int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001913 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001914 size_t *length, loff_t *ppos)
1915{
Andi Kleena5516432008-07-23 21:27:41 -07001916 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07001917 unsigned long tmp;
1918
1919 if (!write)
1920 tmp = h->nr_overcommit_huge_pages;
1921
1922 table->data = &tmp;
1923 table->maxlen = sizeof(unsigned long);
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001924 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001925
1926 if (write) {
1927 spin_lock(&hugetlb_lock);
1928 h->nr_overcommit_huge_pages = tmp;
1929 spin_unlock(&hugetlb_lock);
1930 }
1931
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001932 return 0;
1933}
1934
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935#endif /* CONFIG_SYSCTL */
1936
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001937void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938{
Andi Kleena5516432008-07-23 21:27:41 -07001939 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001940 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001941 "HugePages_Total: %5lu\n"
1942 "HugePages_Free: %5lu\n"
1943 "HugePages_Rsvd: %5lu\n"
1944 "HugePages_Surp: %5lu\n"
1945 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07001946 h->nr_huge_pages,
1947 h->free_huge_pages,
1948 h->resv_huge_pages,
1949 h->surplus_huge_pages,
1950 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951}
1952
1953int hugetlb_report_node_meminfo(int nid, char *buf)
1954{
Andi Kleena5516432008-07-23 21:27:41 -07001955 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956 return sprintf(buf,
1957 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07001958 "Node %d HugePages_Free: %5u\n"
1959 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07001960 nid, h->nr_huge_pages_node[nid],
1961 nid, h->free_huge_pages_node[nid],
1962 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963}
1964
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1966unsigned long hugetlb_total_pages(void)
1967{
Andi Kleena5516432008-07-23 21:27:41 -07001968 struct hstate *h = &default_hstate;
1969 return h->nr_huge_pages * pages_per_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971
Andi Kleena5516432008-07-23 21:27:41 -07001972static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001973{
1974 int ret = -ENOMEM;
1975
1976 spin_lock(&hugetlb_lock);
1977 /*
1978 * When cpuset is configured, it breaks the strict hugetlb page
1979 * reservation as the accounting is done on a global variable. Such
1980 * reservation is completely rubbish in the presence of cpuset because
1981 * the reservation is not checked against page availability for the
1982 * current cpuset. Application can still potentially OOM'ed by kernel
1983 * with lack of free htlb page in cpuset that the task is in.
1984 * Attempt to enforce strict accounting with cpuset is almost
1985 * impossible (or too ugly) because cpuset is too fluid that
1986 * task or memory node can be dynamically moved between cpusets.
1987 *
1988 * The change of semantics for shared hugetlb mapping with cpuset is
1989 * undesirable. However, in order to preserve some of the semantics,
1990 * we fall back to check against current free page availability as
1991 * a best attempt and hopefully to minimize the impact of changing
1992 * semantics that cpuset has.
1993 */
1994 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001995 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001996 goto out;
1997
Andi Kleena5516432008-07-23 21:27:41 -07001998 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1999 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002000 goto out;
2001 }
2002 }
2003
2004 ret = 0;
2005 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002006 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002007
2008out:
2009 spin_unlock(&hugetlb_lock);
2010 return ret;
2011}
2012
Andy Whitcroft84afd992008-07-23 21:27:32 -07002013static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2014{
2015 struct resv_map *reservations = vma_resv_map(vma);
2016
2017 /*
2018 * This new VMA should share its siblings reservation map if present.
2019 * The VMA will only ever have a valid reservation map pointer where
2020 * it is being copied for another still existing VMA. As that VMA
2021 * has a reference to the reservation map it cannot dissappear until
2022 * after this open call completes. It is therefore safe to take a
2023 * new reference here without additional locking.
2024 */
2025 if (reservations)
2026 kref_get(&reservations->refs);
2027}
2028
Mel Gormana1e78772008-07-23 21:27:23 -07002029static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2030{
Andi Kleena5516432008-07-23 21:27:41 -07002031 struct hstate *h = hstate_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002032 struct resv_map *reservations = vma_resv_map(vma);
2033 unsigned long reserve;
2034 unsigned long start;
2035 unsigned long end;
2036
2037 if (reservations) {
Andi Kleena5516432008-07-23 21:27:41 -07002038 start = vma_hugecache_offset(h, vma, vma->vm_start);
2039 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002040
2041 reserve = (end - start) -
2042 region_count(&reservations->regions, start, end);
2043
2044 kref_put(&reservations->refs, resv_map_release);
2045
Adam Litke7251ff782008-07-23 21:27:59 -07002046 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07002047 hugetlb_acct_memory(h, -reserve);
Adam Litke7251ff782008-07-23 21:27:59 -07002048 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2049 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002050 }
Mel Gormana1e78772008-07-23 21:27:23 -07002051}
2052
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053/*
2054 * We cannot handle pagefaults against hugetlb pages at all. They cause
2055 * handle_mm_fault() to try to instantiate regular-sized pages in the
2056 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2057 * this far.
2058 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002059static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060{
2061 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002062 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002063}
2064
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002065const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002066 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002067 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002068 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069};
2070
David Gibson1e8f8892006-01-06 00:10:44 -08002071static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2072 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002073{
2074 pte_t entry;
2075
David Gibson1e8f8892006-01-06 00:10:44 -08002076 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -07002077 entry =
2078 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2079 } else {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002080 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002081 }
2082 entry = pte_mkyoung(entry);
2083 entry = pte_mkhuge(entry);
2084
2085 return entry;
2086}
2087
David Gibson1e8f8892006-01-06 00:10:44 -08002088static void set_huge_ptep_writable(struct vm_area_struct *vma,
2089 unsigned long address, pte_t *ptep)
2090{
2091 pte_t entry;
2092
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002093 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2094 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
Russell King4b3073e2009-12-18 16:40:18 +00002095 update_mmu_cache(vma, address, ptep);
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07002096 }
David Gibson1e8f8892006-01-06 00:10:44 -08002097}
2098
2099
David Gibson63551ae2005-06-21 17:14:44 -07002100int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2101 struct vm_area_struct *vma)
2102{
2103 pte_t *src_pte, *dst_pte, entry;
2104 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002105 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002106 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002107 struct hstate *h = hstate_vma(vma);
2108 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08002109
2110 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002111
Andi Kleena5516432008-07-23 21:27:41 -07002112 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Hugh Dickinsc74df322005-10-29 18:16:23 -07002113 src_pte = huge_pte_offset(src, addr);
2114 if (!src_pte)
2115 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002116 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07002117 if (!dst_pte)
2118 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08002119
2120 /* If the pagetables are shared don't copy or take references */
2121 if (dst_pte == src_pte)
2122 continue;
2123
Hugh Dickinsc74df322005-10-29 18:16:23 -07002124 spin_lock(&dst->page_table_lock);
Nick Piggin46478752008-06-05 22:45:57 -07002125 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002126 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002127 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002128 huge_ptep_set_wrprotect(src, addr, src_pte);
2129 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002130 ptepage = pte_page(entry);
2131 get_page(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002132 set_huge_pte_at(dst, addr, dst_pte, entry);
2133 }
2134 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -07002135 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002136 }
2137 return 0;
2138
2139nomem:
2140 return -ENOMEM;
2141}
2142
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002143void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002144 unsigned long end, struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002145{
2146 struct mm_struct *mm = vma->vm_mm;
2147 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002148 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002149 pte_t pte;
2150 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002151 struct page *tmp;
Andi Kleena5516432008-07-23 21:27:41 -07002152 struct hstate *h = hstate_vma(vma);
2153 unsigned long sz = huge_page_size(h);
2154
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -08002155 /*
2156 * A page gathering list, protected by per file i_mmap_lock. The
2157 * lock is used to avoid list corruption from multiple unmapping
2158 * of the same page since we are using page->lru.
2159 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002160 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -07002161
2162 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002163 BUG_ON(start & ~huge_page_mask(h));
2164 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002165
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07002166 mmu_notifier_invalidate_range_start(mm, start, end);
Hugh Dickins508034a2005-10-29 18:16:30 -07002167 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002168 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002169 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002170 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002171 continue;
2172
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002173 if (huge_pmd_unshare(mm, &address, ptep))
2174 continue;
2175
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002176 /*
2177 * If a reference page is supplied, it is because a specific
2178 * page is being unmapped, not a range. Ensure the page we
2179 * are about to unmap is the actual page of interest.
2180 */
2181 if (ref_page) {
2182 pte = huge_ptep_get(ptep);
2183 if (huge_pte_none(pte))
2184 continue;
2185 page = pte_page(pte);
2186 if (page != ref_page)
2187 continue;
2188
2189 /*
2190 * Mark the VMA as having unmapped its page so that
2191 * future faults in this VMA will fail rather than
2192 * looking like data was lost
2193 */
2194 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2195 }
2196
David Gibsonc7546f82005-08-05 11:59:35 -07002197 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002198 if (huge_pte_none(pte))
David Gibson63551ae2005-06-21 17:14:44 -07002199 continue;
David Gibsonc7546f82005-08-05 11:59:35 -07002200
David Gibson63551ae2005-06-21 17:14:44 -07002201 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -08002202 if (pte_dirty(pte))
2203 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002204 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -07002205 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -07002207 flush_tlb_range(vma, start, end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07002208 mmu_notifier_invalidate_range_end(mm, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002209 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2210 list_del(&page->lru);
2211 put_page(page);
2212 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213}
David Gibson63551ae2005-06-21 17:14:44 -07002214
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002215void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002216 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002217{
Andi Kleena137e1c2008-07-23 21:27:43 -07002218 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2219 __unmap_hugepage_range(vma, start, end, ref_page);
2220 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002221}
2222
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002223/*
2224 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2225 * mappping it owns the reserve page for. The intention is to unmap the page
2226 * from other VMAs and let the children be SIGKILLed if they are faulting the
2227 * same region.
2228 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07002229static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2230 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002231{
Adam Litke75266742008-11-12 13:24:56 -08002232 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002233 struct vm_area_struct *iter_vma;
2234 struct address_space *mapping;
2235 struct prio_tree_iter iter;
2236 pgoff_t pgoff;
2237
2238 /*
2239 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2240 * from page cache lookup which is in HPAGE_SIZE units.
2241 */
Adam Litke75266742008-11-12 13:24:56 -08002242 address = address & huge_page_mask(h);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002243 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2244 + (vma->vm_pgoff >> PAGE_SHIFT);
2245 mapping = (struct address_space *)page_private(page);
2246
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002247 /*
2248 * Take the mapping lock for the duration of the table walk. As
2249 * this mapping should be shared between all the VMAs,
2250 * __unmap_hugepage_range() is called as the lock is already held
2251 */
2252 spin_lock(&mapping->i_mmap_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002253 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2254 /* Do not unmap the current VMA */
2255 if (iter_vma == vma)
2256 continue;
2257
2258 /*
2259 * Unmap the page from other VMAs without their own reserves.
2260 * They get marked to be SIGKILLed if they fault in these
2261 * areas. This is because a future no-page fault on this VMA
2262 * could insert a zeroed page instead of the data existing
2263 * from the time of fork. This would look like data corruption
2264 */
2265 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002266 __unmap_hugepage_range(iter_vma,
Adam Litke75266742008-11-12 13:24:56 -08002267 address, address + huge_page_size(h),
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002268 page);
2269 }
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002270 spin_unlock(&mapping->i_mmap_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002271
2272 return 1;
2273}
2274
David Gibson1e8f8892006-01-06 00:10:44 -08002275static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002276 unsigned long address, pte_t *ptep, pte_t pte,
2277 struct page *pagecache_page)
David Gibson1e8f8892006-01-06 00:10:44 -08002278{
Andi Kleena5516432008-07-23 21:27:41 -07002279 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002280 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -08002281 int avoidcopy;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002282 int outside_reserve = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002283
2284 old_page = pte_page(pte);
2285
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002286retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002287 /* If no-one else is actually using this page, avoid the copy
2288 * and just make the page writable */
2289 avoidcopy = (page_count(old_page) == 1);
2290 if (avoidcopy) {
2291 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002292 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002293 }
2294
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002295 /*
2296 * If the process that created a MAP_PRIVATE mapping is about to
2297 * perform a COW due to a shared page count, attempt to satisfy
2298 * the allocation without using the existing reserves. The pagecache
2299 * page is used to determine if the reserve at this address was
2300 * consumed or not. If reserves were used, a partial faulted mapping
2301 * at the time of fork() could consume its reserves on COW instead
2302 * of the full address range.
2303 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002304 if (!(vma->vm_flags & VM_MAYSHARE) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002305 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2306 old_page != pagecache_page)
2307 outside_reserve = 1;
2308
David Gibson1e8f8892006-01-06 00:10:44 -08002309 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002310
2311 /* Drop page_table_lock as buddy allocator may be called */
2312 spin_unlock(&mm->page_table_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002313 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002314
Adam Litke2fc39ce2007-11-14 16:59:39 -08002315 if (IS_ERR(new_page)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002316 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002317
2318 /*
2319 * If a process owning a MAP_PRIVATE mapping fails to COW,
2320 * it is due to references held by a child and an insufficient
2321 * huge page pool. To guarantee the original mappers
2322 * reliability, unmap the page from child processes. The child
2323 * may get SIGKILLed if it later faults.
2324 */
2325 if (outside_reserve) {
2326 BUG_ON(huge_pte_none(pte));
2327 if (unmap_ref_private(mm, vma, old_page, address)) {
2328 BUG_ON(page_count(old_page) != 1);
2329 BUG_ON(huge_pte_none(pte));
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002330 spin_lock(&mm->page_table_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002331 goto retry_avoidcopy;
2332 }
2333 WARN_ON_ONCE(1);
2334 }
2335
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002336 /* Caller expects lock to be held */
2337 spin_lock(&mm->page_table_lock);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002338 return -PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002339 }
2340
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002341 copy_huge_page(new_page, old_page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002342 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002343
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002344 /*
2345 * Retake the page_table_lock to check for racing updates
2346 * before the page tables are altered
2347 */
2348 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002349 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002350 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002351 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002352 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002353 set_huge_pte_at(mm, address, ptep,
2354 make_huge_pte(vma, new_page, 1));
2355 /* Make the old page be freed below */
2356 new_page = old_page;
2357 }
2358 page_cache_release(new_page);
2359 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -07002360 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002361}
2362
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002363/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002364static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2365 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002366{
2367 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002368 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002369
2370 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002371 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002372
2373 return find_lock_page(mapping, idx);
2374}
2375
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002376/*
2377 * Return whether there is a pagecache page to back given address within VMA.
2378 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2379 */
2380static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002381 struct vm_area_struct *vma, unsigned long address)
2382{
2383 struct address_space *mapping;
2384 pgoff_t idx;
2385 struct page *page;
2386
2387 mapping = vma->vm_file->f_mapping;
2388 idx = vma_hugecache_offset(h, vma, address);
2389
2390 page = find_get_page(mapping, idx);
2391 if (page)
2392 put_page(page);
2393 return page != NULL;
2394}
2395
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002396static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002397 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002398{
Andi Kleena5516432008-07-23 21:27:41 -07002399 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002400 int ret = VM_FAULT_SIGBUS;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002401 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07002402 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002403 struct page *page;
2404 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08002405 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -07002406
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002407 /*
2408 * Currently, we are forced to kill the process in the event the
2409 * original mapper has unmapped pages from the child due to a failed
2410 * COW. Warn that such a situation has occured as it may not be obvious
2411 */
2412 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2413 printk(KERN_WARNING
2414 "PID %d killed due to inadequate hugepage pool\n",
2415 current->pid);
2416 return ret;
2417 }
2418
Adam Litke4c887262005-10-29 18:16:46 -07002419 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002420 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07002421
2422 /*
2423 * Use page lock to guard against racing truncation
2424 * before we get page_table_lock.
2425 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002426retry:
2427 page = find_lock_page(mapping, idx);
2428 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002429 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002430 if (idx >= size)
2431 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002432 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002433 if (IS_ERR(page)) {
2434 ret = -PTR_ERR(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002435 goto out;
2436 }
Andi Kleena5516432008-07-23 21:27:41 -07002437 clear_huge_page(page, address, huge_page_size(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002438 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002439
Mel Gormanf83a2752009-05-28 14:34:40 -07002440 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002441 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002442 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002443
2444 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2445 if (err) {
2446 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002447 if (err == -EEXIST)
2448 goto retry;
2449 goto out;
2450 }
Ken Chen45c682a2007-11-14 16:59:44 -08002451
2452 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002453 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002454 spin_unlock(&inode->i_lock);
Mel Gorman23be7462010-04-23 13:17:56 -04002455 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002456 lock_page(page);
Mel Gorman23be7462010-04-23 13:17:56 -04002457 page->mapping = HUGETLB_POISON;
2458 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08002459 }
David Gibson1e8f8892006-01-06 00:10:44 -08002460
Andy Whitcroft57303d82008-08-12 15:08:47 -07002461 /*
2462 * If we are going to COW a private mapping later, we examine the
2463 * pending reservations for this page now. This will ensure that
2464 * any allocations necessary to record that reservation occur outside
2465 * the spinlock.
2466 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002467 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002468 if (vma_needs_reservation(h, vma, address) < 0) {
2469 ret = VM_FAULT_OOM;
2470 goto backout_unlocked;
2471 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002472
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002473 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002474 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002475 if (idx >= size)
2476 goto backout;
2477
Nick Piggin83c54072007-07-19 01:47:05 -07002478 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002479 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002480 goto backout;
2481
David Gibson1e8f8892006-01-06 00:10:44 -08002482 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2483 && (vma->vm_flags & VM_SHARED)));
2484 set_huge_pte_at(mm, address, ptep, new_pte);
2485
Hugh Dickins788c7df2009-06-23 13:49:05 +01002486 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002487 /* Optimization, do the COW without a second fault */
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002488 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
David Gibson1e8f8892006-01-06 00:10:44 -08002489 }
2490
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002491 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07002492 unlock_page(page);
2493out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002494 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002495
2496backout:
2497 spin_unlock(&mm->page_table_lock);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002498backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002499 unlock_page(page);
2500 put_page(page);
2501 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002502}
2503
Adam Litke86e52162006-01-06 00:10:43 -08002504int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002505 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08002506{
2507 pte_t *ptep;
2508 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -08002509 int ret;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002510 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002511 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002512 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002513
Andi Kleena5516432008-07-23 21:27:41 -07002514 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002515 if (!ptep)
2516 return VM_FAULT_OOM;
2517
David Gibson3935baa2006-03-22 00:08:53 -08002518 /*
2519 * Serialize hugepage allocation and instantiation, so that we don't
2520 * get spurious allocation failures if two CPUs race to instantiate
2521 * the same page in the page cache.
2522 */
2523 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002524 entry = huge_ptep_get(ptep);
2525 if (huge_pte_none(entry)) {
Hugh Dickins788c7df2009-06-23 13:49:05 +01002526 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07002527 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002528 }
Adam Litke86e52162006-01-06 00:10:43 -08002529
Nick Piggin83c54072007-07-19 01:47:05 -07002530 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002531
Andy Whitcroft57303d82008-08-12 15:08:47 -07002532 /*
2533 * If we are going to COW the mapping later, we examine the pending
2534 * reservations for this page now. This will ensure that any
2535 * allocations necessary to record that reservation occur outside the
2536 * spinlock. For private mappings, we also lookup the pagecache
2537 * page now as it is used to determine if a reservation has been
2538 * consumed.
2539 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002540 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002541 if (vma_needs_reservation(h, vma, address) < 0) {
2542 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002543 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002544 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002545
Mel Gormanf83a2752009-05-28 14:34:40 -07002546 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07002547 pagecache_page = hugetlbfs_pagecache_page(h,
2548 vma, address);
2549 }
2550
David Gibson1e8f8892006-01-06 00:10:44 -08002551 spin_lock(&mm->page_table_lock);
2552 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002553 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2554 goto out_page_table_lock;
2555
2556
Hugh Dickins788c7df2009-06-23 13:49:05 +01002557 if (flags & FAULT_FLAG_WRITE) {
David Gibsonb4d1d992008-10-15 22:01:11 -07002558 if (!pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002559 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2560 pagecache_page);
David Gibsonb4d1d992008-10-15 22:01:11 -07002561 goto out_page_table_lock;
2562 }
2563 entry = pte_mkdirty(entry);
2564 }
2565 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01002566 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2567 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00002568 update_mmu_cache(vma, address, ptep);
David Gibsonb4d1d992008-10-15 22:01:11 -07002569
2570out_page_table_lock:
David Gibson1e8f8892006-01-06 00:10:44 -08002571 spin_unlock(&mm->page_table_lock);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002572
2573 if (pagecache_page) {
2574 unlock_page(pagecache_page);
2575 put_page(pagecache_page);
2576 }
2577
David Gibsonb4d1d992008-10-15 22:01:11 -07002578out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08002579 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08002580
2581 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08002582}
2583
Andi Kleenceb86872008-07-23 21:27:50 -07002584/* Can be overriden by architectures */
2585__attribute__((weak)) struct page *
2586follow_huge_pud(struct mm_struct *mm, unsigned long address,
2587 pud_t *pud, int write)
2588{
2589 BUG();
2590 return NULL;
2591}
2592
David Gibson63551ae2005-06-21 17:14:44 -07002593int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2594 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -08002595 unsigned long *position, int *length, int i,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002596 unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07002597{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002598 unsigned long pfn_offset;
2599 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -07002600 int remainder = *length;
Andi Kleena5516432008-07-23 21:27:41 -07002601 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07002602
Hugh Dickins1c598272005-10-19 21:23:43 -07002603 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002604 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07002605 pte_t *pte;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002606 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07002607 struct page *page;
2608
2609 /*
2610 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002611 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07002612 * first, for the page indexing below to work.
2613 */
Andi Kleena5516432008-07-23 21:27:41 -07002614 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002615 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07002616
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002617 /*
2618 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002619 * an error where there's an empty slot with no huge pagecache
2620 * to back it. This way, we avoid allocating a hugepage, and
2621 * the sparse dumpfile avoids allocating disk blocks, but its
2622 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002623 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002624 if (absent && (flags & FOLL_DUMP) &&
2625 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002626 remainder = 0;
2627 break;
2628 }
2629
2630 if (absent ||
2631 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07002632 int ret;
2633
2634 spin_unlock(&mm->page_table_lock);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002635 ret = hugetlb_fault(mm, vma, vaddr,
2636 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litke4c887262005-10-29 18:16:46 -07002637 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07002638 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07002639 continue;
2640
2641 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002642 break;
2643 }
David Gibson63551ae2005-06-21 17:14:44 -07002644
Andi Kleena5516432008-07-23 21:27:41 -07002645 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002646 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002647same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002648 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002649 pages[i] = mem_map_offset(page, pfn_offset);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002650 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002651 }
David Gibson63551ae2005-06-21 17:14:44 -07002652
2653 if (vmas)
2654 vmas[i] = vma;
2655
2656 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002657 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07002658 --remainder;
2659 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002660 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07002661 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002662 /*
2663 * We use pfn_offset to avoid touching the pageframes
2664 * of this compound page.
2665 */
2666 goto same_page;
2667 }
David Gibson63551ae2005-06-21 17:14:44 -07002668 }
Hugh Dickins1c598272005-10-19 21:23:43 -07002669 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002670 *length = remainder;
2671 *position = vaddr;
2672
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002673 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07002674}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002675
2676void hugetlb_change_protection(struct vm_area_struct *vma,
2677 unsigned long address, unsigned long end, pgprot_t newprot)
2678{
2679 struct mm_struct *mm = vma->vm_mm;
2680 unsigned long start = address;
2681 pte_t *ptep;
2682 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07002683 struct hstate *h = hstate_vma(vma);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002684
2685 BUG_ON(address >= end);
2686 flush_cache_range(vma, address, end);
2687
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002688 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002689 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002690 for (; address < end; address += huge_page_size(h)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002691 ptep = huge_pte_offset(mm, address);
2692 if (!ptep)
2693 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002694 if (huge_pmd_unshare(mm, &address, ptep))
2695 continue;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002696 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002697 pte = huge_ptep_get_and_clear(mm, address, ptep);
2698 pte = pte_mkhuge(pte_modify(pte, newprot));
2699 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002700 }
2701 }
2702 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002703 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002704
2705 flush_tlb_range(vma, start, end);
2706}
2707
Mel Gormana1e78772008-07-23 21:27:23 -07002708int hugetlb_reserve_pages(struct inode *inode,
2709 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00002710 struct vm_area_struct *vma,
2711 int acctflag)
Adam Litkee4e574b2007-10-16 01:26:19 -07002712{
Mel Gorman17c9d122009-02-11 16:34:16 +00002713 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07002714 struct hstate *h = hstate_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07002715
Mel Gormana1e78772008-07-23 21:27:23 -07002716 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00002717 * Only apply hugepage reservation if asked. At fault time, an
2718 * attempt will be made for VM_NORESERVE to allocate a page
2719 * and filesystem quota without using reserves
2720 */
2721 if (acctflag & VM_NORESERVE)
2722 return 0;
2723
2724 /*
Mel Gormana1e78772008-07-23 21:27:23 -07002725 * Shared mappings base their reservation on the number of pages that
2726 * are already allocated on behalf of the file. Private mappings need
2727 * to reserve the full area even if read-only as mprotect() may be
2728 * called to make the mapping read-write. Assume !vma is a shm mapping
2729 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002730 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gormana1e78772008-07-23 21:27:23 -07002731 chg = region_chg(&inode->i_mapping->private_list, from, to);
Mel Gorman5a6fe122009-02-10 14:02:27 +00002732 else {
2733 struct resv_map *resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00002734 if (!resv_map)
2735 return -ENOMEM;
2736
Mel Gorman17c9d122009-02-11 16:34:16 +00002737 chg = to - from;
2738
Mel Gorman5a6fe122009-02-10 14:02:27 +00002739 set_vma_resv_map(vma, resv_map);
2740 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2741 }
2742
Mel Gorman17c9d122009-02-11 16:34:16 +00002743 if (chg < 0)
2744 return chg;
2745
2746 /* There must be enough filesystem quota for the mapping */
2747 if (hugetlb_get_quota(inode->i_mapping, chg))
2748 return -ENOSPC;
2749
2750 /*
2751 * Check enough hugepages are available for the reservation.
2752 * Hand back the quota if there are not
2753 */
2754 ret = hugetlb_acct_memory(h, chg);
2755 if (ret < 0) {
2756 hugetlb_put_quota(inode->i_mapping, chg);
2757 return ret;
2758 }
2759
2760 /*
2761 * Account for the reservations made. Shared mappings record regions
2762 * that have reservations as they are shared by multiple VMAs.
2763 * When the last VMA disappears, the region map says how much
2764 * the reservation was and the page cache tells how much of
2765 * the reservation was consumed. Private mappings are per-VMA and
2766 * only the consumed reservations are tracked. When the VMA
2767 * disappears, the original reservation is the VMA size and the
2768 * consumed reservations are stored in the map. Hence, nothing
2769 * else has to be done for private mappings here
2770 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002771 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gorman17c9d122009-02-11 16:34:16 +00002772 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002773 return 0;
2774}
2775
2776void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2777{
Andi Kleena5516432008-07-23 21:27:41 -07002778 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002779 long chg = region_truncate(&inode->i_mapping->private_list, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08002780
2781 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07002782 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08002783 spin_unlock(&inode->i_lock);
2784
Adam Litke90d8b7e2007-11-14 16:59:42 -08002785 hugetlb_put_quota(inode->i_mapping, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07002786 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002787}