blob: 2d16fa6b8c2d8fb6a40ea5b3f9ea57c77a34d1af [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
4 */
5#include <linux/gfp.h>
6#include <linux/list.h>
7#include <linux/init.h>
8#include <linux/module.h>
9#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +040010#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070011#include <linux/sysctl.h>
12#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070013#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070015#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080016#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080017#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080018#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070019#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070020#include <linux/sysfs.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070021
David Gibson63551ae2005-06-21 17:14:44 -070022#include <asm/page.h>
23#include <asm/pgtable.h>
Adrian Bunk78a34ae2008-07-28 15:46:30 -070024#include <asm/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070025
26#include <linux/hugetlb.h>
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;
Mel Gorman396faf02007-07-17 04:03:13 -0700468 struct zonelist *zonelist = huge_zonelist(vma, address,
Mel Gorman19770b32008-04-28 02:12:18 -0700469 htlb_alloc_mask, &mpol, &nodemask);
Mel Gormandd1a2392008-04-28 02:12:17 -0700470 struct zone *zone;
471 struct zoneref *z;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472
Mel Gormana1e78772008-07-23 21:27:23 -0700473 /*
474 * A child process with MAP_PRIVATE mappings created by their parent
475 * have no page reserves. This check ensures that reservations are
476 * not "stolen". The child may still get SIGKILLed
477 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700478 if (!vma_has_reserves(vma) &&
Andi Kleena5516432008-07-23 21:27:41 -0700479 h->free_huge_pages - h->resv_huge_pages == 0)
Mel Gormana1e78772008-07-23 21:27:23 -0700480 return NULL;
481
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700482 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700483 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700484 return NULL;
485
Mel Gorman19770b32008-04-28 02:12:18 -0700486 for_each_zone_zonelist_nodemask(zone, z, zonelist,
487 MAX_NR_ZONES - 1, nodemask) {
Mel Gorman54a6eb52008-04-28 02:12:16 -0700488 nid = zone_to_nid(zone);
489 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
Andi Kleena5516432008-07-23 21:27:41 -0700490 !list_empty(&h->hugepage_freelists[nid])) {
491 page = list_entry(h->hugepage_freelists[nid].next,
Andrew Morton3abf7af2007-07-19 01:49:08 -0700492 struct page, lru);
493 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700494 h->free_huge_pages--;
495 h->free_huge_pages_node[nid]--;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700496
497 if (!avoid_reserve)
Andi Kleena5516432008-07-23 21:27:41 -0700498 decrement_hugepage_resv_vma(h, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700499
Ken Chen5ab3ee72007-07-23 18:44:00 -0700500 break;
Andrew Morton3abf7af2007-07-19 01:49:08 -0700501 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 }
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700503 mpol_cond_put(mpol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 return page;
505}
506
Andi Kleena5516432008-07-23 21:27:41 -0700507static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700508{
509 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700510
Andy Whitcroft18229df2008-11-06 12:53:27 -0800511 VM_BUG_ON(h->order >= MAX_ORDER);
512
Andi Kleena5516432008-07-23 21:27:41 -0700513 h->nr_huge_pages--;
514 h->nr_huge_pages_node[page_to_nid(page)]--;
515 for (i = 0; i < pages_per_huge_page(h); i++) {
Adam Litke6af2acb2007-10-16 01:26:16 -0700516 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
517 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
518 1 << PG_private | 1<< PG_writeback);
519 }
520 set_compound_page_dtor(page, NULL);
521 set_page_refcounted(page);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700522 arch_release_hugepage(page);
Andi Kleena5516432008-07-23 21:27:41 -0700523 __free_pages(page, huge_page_order(h));
Adam Litke6af2acb2007-10-16 01:26:16 -0700524}
525
Andi Kleene5ff2152008-07-23 21:27:42 -0700526struct hstate *size_to_hstate(unsigned long size)
527{
528 struct hstate *h;
529
530 for_each_hstate(h) {
531 if (huge_page_size(h) == size)
532 return h;
533 }
534 return NULL;
535}
536
David Gibson27a85ef2006-03-22 00:08:56 -0800537static void free_huge_page(struct page *page)
538{
Andi Kleena5516432008-07-23 21:27:41 -0700539 /*
540 * Can't pass hstate in here because it is called from the
541 * compound page destructor.
542 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700543 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700544 int nid = page_to_nid(page);
Adam Litkec79fb752007-11-14 16:59:38 -0800545 struct address_space *mapping;
David Gibson27a85ef2006-03-22 00:08:56 -0800546
Adam Litkec79fb752007-11-14 16:59:38 -0800547 mapping = (struct address_space *) page_private(page);
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800548 set_page_private(page, 0);
Adam Litke7893d1d2007-10-16 01:26:18 -0700549 BUG_ON(page_count(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800550 INIT_LIST_HEAD(&page->lru);
551
552 spin_lock(&hugetlb_lock);
Andi Kleenaa888a72008-07-23 21:27:47 -0700553 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
Andi Kleena5516432008-07-23 21:27:41 -0700554 update_and_free_page(h, page);
555 h->surplus_huge_pages--;
556 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700557 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700558 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700559 }
David Gibson27a85ef2006-03-22 00:08:56 -0800560 spin_unlock(&hugetlb_lock);
Adam Litkec79fb752007-11-14 16:59:38 -0800561 if (mapping)
Adam Litke9a119c02007-11-14 16:59:41 -0800562 hugetlb_put_quota(mapping, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800563}
564
Andi Kleena5516432008-07-23 21:27:41 -0700565static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700566{
567 set_compound_page_dtor(page, free_huge_page);
568 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700569 h->nr_huge_pages++;
570 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700571 spin_unlock(&hugetlb_lock);
572 put_page(page); /* free it into the hugepage allocator */
573}
574
Wu Fengguang20a03072009-06-16 15:32:22 -0700575static void prep_compound_gigantic_page(struct page *page, unsigned long order)
576{
577 int i;
578 int nr_pages = 1 << order;
579 struct page *p = page + 1;
580
581 /* we rely on prep_new_huge_page to set the destructor */
582 set_compound_order(page, order);
583 __SetPageHead(page);
584 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
585 __SetPageTail(p);
586 p->first_page = page;
587 }
588}
589
590int PageHuge(struct page *page)
591{
592 compound_page_dtor *dtor;
593
594 if (!PageCompound(page))
595 return 0;
596
597 page = compound_head(page);
598 dtor = get_compound_page_dtor(page);
599
600 return dtor == free_huge_page;
601}
602
Andi Kleena5516432008-07-23 21:27:41 -0700603static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700606
Andi Kleenaa888a72008-07-23 21:27:47 -0700607 if (h->order >= MAX_ORDER)
608 return NULL;
609
Mel Gorman6484eb32009-06-16 15:31:54 -0700610 page = alloc_pages_exact_node(nid,
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700611 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
612 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700613 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700615 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700616 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700617 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700618 }
Andi Kleena5516432008-07-23 21:27:41 -0700619 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700621
622 return page;
623}
624
Andi Kleen5ced66c2008-07-23 21:27:45 -0700625/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800626 * common helper functions for hstate_next_node_to_{alloc|free}.
627 * We may have allocated or freed a huge page based on a different
628 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
629 * be outside of *nodes_allowed. Ensure that we use an allowed
630 * node for alloc or free.
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800631 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800632static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800633{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800634 nid = next_node(nid, *nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800635 if (nid == MAX_NUMNODES)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800636 nid = first_node(*nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800637 VM_BUG_ON(nid >= MAX_NUMNODES);
638
639 return nid;
640}
641
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800642static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
Andi Kleen5ced66c2008-07-23 21:27:45 -0700643{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800644 if (!node_isset(nid, *nodes_allowed))
645 nid = next_node_allowed(nid, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800646 return nid;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700647}
648
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800649/*
650 * returns the previously saved node ["this node"] from which to
651 * allocate a persistent huge page for the pool and advance the
652 * next node from which to allocate, handling wrap at end of node
653 * mask.
654 */
655static int hstate_next_node_to_alloc(struct hstate *h,
656 nodemask_t *nodes_allowed)
657{
658 int nid;
659
660 VM_BUG_ON(!nodes_allowed);
661
662 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
663 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
664
665 return nid;
666}
667
668static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700669{
670 struct page *page;
671 int start_nid;
672 int next_nid;
673 int ret = 0;
674
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800675 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700676 next_nid = start_nid;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700677
678 do {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700679 page = alloc_fresh_huge_page_node(h, next_nid);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800680 if (page) {
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700681 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800682 break;
683 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800684 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800685 } while (next_nid != start_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700686
Adam Litke3b116302008-04-28 02:13:06 -0700687 if (ret)
688 count_vm_event(HTLB_BUDDY_PGALLOC);
689 else
690 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
691
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700692 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693}
694
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700695/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800696 * helper for free_pool_huge_page() - return the previously saved
697 * node ["this node"] from which to free a huge page. Advance the
698 * next node id whether or not we find a free huge page to free so
699 * that the next attempt to free addresses the next node.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700700 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800701static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700702{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800703 int nid;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800704
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800705 VM_BUG_ON(!nodes_allowed);
706
707 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
708 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
709
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800710 return nid;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700711}
712
713/*
714 * Free huge page from pool from next node to free.
715 * Attempt to keep persistent huge pages more or less
716 * balanced over allowed nodes.
717 * Called with hugetlb_lock locked.
718 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800719static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
720 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700721{
722 int start_nid;
723 int next_nid;
724 int ret = 0;
725
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800726 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700727 next_nid = start_nid;
728
729 do {
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700730 /*
731 * If we're returning unused surplus pages, only examine
732 * nodes with surplus pages.
733 */
734 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
735 !list_empty(&h->hugepage_freelists[next_nid])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700736 struct page *page =
737 list_entry(h->hugepage_freelists[next_nid].next,
738 struct page, lru);
739 list_del(&page->lru);
740 h->free_huge_pages--;
741 h->free_huge_pages_node[next_nid]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700742 if (acct_surplus) {
743 h->surplus_huge_pages--;
744 h->surplus_huge_pages_node[next_nid]--;
745 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700746 update_and_free_page(h, page);
747 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800748 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700749 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800750 next_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800751 } while (next_nid != start_nid);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700752
753 return ret;
754}
755
Andi Kleena5516432008-07-23 21:27:41 -0700756static struct page *alloc_buddy_huge_page(struct hstate *h,
757 struct vm_area_struct *vma, unsigned long address)
Adam Litke7893d1d2007-10-16 01:26:18 -0700758{
759 struct page *page;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800760 unsigned int nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700761
Andi Kleenaa888a72008-07-23 21:27:47 -0700762 if (h->order >= MAX_ORDER)
763 return NULL;
764
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800765 /*
766 * Assume we will successfully allocate the surplus page to
767 * prevent racing processes from causing the surplus to exceed
768 * overcommit
769 *
770 * This however introduces a different race, where a process B
771 * tries to grow the static hugepage pool while alloc_pages() is
772 * called by process A. B will only examine the per-node
773 * counters in determining if surplus huge pages can be
774 * converted to normal huge pages in adjust_pool_surplus(). A
775 * won't be able to increment the per-node counter, until the
776 * lock is dropped by B, but B doesn't drop hugetlb_lock until
777 * no more huge pages can be converted from surplus to normal
778 * state (and doesn't try to convert again). Thus, we have a
779 * case where a surplus huge page exists, the pool is grown, and
780 * the surplus huge page still exists after, even though it
781 * should just have been converted to a normal huge page. This
782 * does not leak memory, though, as the hugepage will be freed
783 * once it is out of use. It also does not allow the counters to
784 * go out of whack in adjust_pool_surplus() as we don't modify
785 * the node values until we've gotten the hugepage and only the
786 * per-node value is checked there.
787 */
788 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700789 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800790 spin_unlock(&hugetlb_lock);
791 return NULL;
792 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700793 h->nr_huge_pages++;
794 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800795 }
796 spin_unlock(&hugetlb_lock);
797
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700798 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
799 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700800 huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800801
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700802 if (page && arch_prepare_hugepage(page)) {
803 __free_pages(page, huge_page_order(h));
804 return NULL;
805 }
806
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800807 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700808 if (page) {
Adam Litke2668db92008-03-10 11:43:50 -0700809 /*
810 * This page is now managed by the hugetlb allocator and has
811 * no users -- drop the buddy allocator's reference.
812 */
813 put_page_testzero(page);
814 VM_BUG_ON(page_count(page));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800815 nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700816 set_compound_page_dtor(page, free_huge_page);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800817 /*
818 * We incremented the global counters already
819 */
Andi Kleena5516432008-07-23 21:27:41 -0700820 h->nr_huge_pages_node[nid]++;
821 h->surplus_huge_pages_node[nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700822 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800823 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700824 h->nr_huge_pages--;
825 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700826 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700827 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800828 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700829
830 return page;
831}
832
Adam Litkee4e574b2007-10-16 01:26:19 -0700833/*
834 * Increase the hugetlb pool such that it can accomodate a reservation
835 * of size 'delta'.
836 */
Andi Kleena5516432008-07-23 21:27:41 -0700837static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -0700838{
839 struct list_head surplus_list;
840 struct page *page, *tmp;
841 int ret, i;
842 int needed, allocated;
843
Andi Kleena5516432008-07-23 21:27:41 -0700844 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800845 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -0700846 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700847 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800848 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700849
850 allocated = 0;
851 INIT_LIST_HEAD(&surplus_list);
852
853 ret = -ENOMEM;
854retry:
855 spin_unlock(&hugetlb_lock);
856 for (i = 0; i < needed; i++) {
Andi Kleena5516432008-07-23 21:27:41 -0700857 page = alloc_buddy_huge_page(h, NULL, 0);
Adam Litkee4e574b2007-10-16 01:26:19 -0700858 if (!page) {
859 /*
860 * We were not able to allocate enough pages to
861 * satisfy the entire reservation so we free what
862 * we've allocated so far.
863 */
864 spin_lock(&hugetlb_lock);
865 needed = 0;
866 goto free;
867 }
868
869 list_add(&page->lru, &surplus_list);
870 }
871 allocated += needed;
872
873 /*
874 * After retaking hugetlb_lock, we need to recalculate 'needed'
875 * because either resv_huge_pages or free_huge_pages may have changed.
876 */
877 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700878 needed = (h->resv_huge_pages + delta) -
879 (h->free_huge_pages + allocated);
Adam Litkee4e574b2007-10-16 01:26:19 -0700880 if (needed > 0)
881 goto retry;
882
883 /*
884 * The surplus_list now contains _at_least_ the number of extra pages
885 * needed to accomodate the reservation. Add the appropriate number
886 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -0800887 * allocator. Commit the entire reservation here to prevent another
888 * process from stealing the pages as they are added to the pool but
889 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -0700890 */
891 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -0700892 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700893 ret = 0;
894free:
Adam Litke19fc3f02008-04-28 02:12:20 -0700895 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -0700896 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -0700897 if ((--needed) < 0)
898 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700899 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700900 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -0700901 }
902
903 /* Free unnecessary surplus pages to the buddy allocator */
904 if (!list_empty(&surplus_list)) {
905 spin_unlock(&hugetlb_lock);
906 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
907 list_del(&page->lru);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700908 /*
Adam Litke2668db92008-03-10 11:43:50 -0700909 * The page has a reference count of zero already, so
910 * call free_huge_page directly instead of using
911 * put_page. This must be done with hugetlb_lock
Adam Litkeaf767cb2007-10-16 01:26:25 -0700912 * unlocked which is safe because free_huge_page takes
913 * hugetlb_lock before deciding how to free the page.
914 */
Adam Litke2668db92008-03-10 11:43:50 -0700915 free_huge_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700916 }
Adam Litke19fc3f02008-04-28 02:12:20 -0700917 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -0700918 }
919
920 return ret;
921}
922
923/*
924 * When releasing a hugetlb pool reservation, any surplus pages that were
925 * allocated to satisfy the reservation must be explicitly freed if they were
926 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700927 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -0700928 */
Andi Kleena5516432008-07-23 21:27:41 -0700929static void return_unused_surplus_pages(struct hstate *h,
930 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -0700931{
Adam Litkee4e574b2007-10-16 01:26:19 -0700932 unsigned long nr_pages;
933
Adam Litkeac09b3a2008-03-04 14:29:38 -0800934 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -0700935 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800936
Andi Kleenaa888a72008-07-23 21:27:47 -0700937 /* Cannot return gigantic pages currently */
938 if (h->order >= MAX_ORDER)
939 return;
940
Andi Kleena5516432008-07-23 21:27:41 -0700941 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -0700942
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700943 /*
944 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -0800945 * evenly across all nodes with memory. Iterate across these nodes
946 * until we can no longer free unreserved surplus pages. This occurs
947 * when the nodes with surplus pages have no free pages.
948 * free_pool_huge_page() will balance the the freed pages across the
949 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700950 */
951 while (nr_pages--) {
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -0800952 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700953 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700954 }
955}
956
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700957/*
958 * Determine if the huge page at addr within the vma has an associated
959 * reservation. Where it does not we will need to logically increase
960 * reservation and actually increase quota before an allocation can occur.
961 * Where any new reservation would be required the reservation change is
962 * prepared, but not committed. Once the page has been quota'd allocated
963 * an instantiated the change should be committed via vma_commit_reservation.
964 * No action is required on failure.
965 */
Roel Kluine2f17d92009-03-31 15:23:15 -0700966static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -0700967 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700968{
969 struct address_space *mapping = vma->vm_file->f_mapping;
970 struct inode *inode = mapping->host;
971
Mel Gormanf83a2752009-05-28 14:34:40 -0700972 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -0700973 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700974 return region_chg(&inode->i_mapping->private_list,
975 idx, idx + 1);
976
Andy Whitcroft84afd992008-07-23 21:27:32 -0700977 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
978 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700979
Andy Whitcroft84afd992008-07-23 21:27:32 -0700980 } else {
Roel Kluine2f17d92009-03-31 15:23:15 -0700981 long err;
Andi Kleena5516432008-07-23 21:27:41 -0700982 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700983 struct resv_map *reservations = vma_resv_map(vma);
984
985 err = region_chg(&reservations->regions, idx, idx + 1);
986 if (err < 0)
987 return err;
988 return 0;
989 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700990}
Andi Kleena5516432008-07-23 21:27:41 -0700991static void vma_commit_reservation(struct hstate *h,
992 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700993{
994 struct address_space *mapping = vma->vm_file->f_mapping;
995 struct inode *inode = mapping->host;
996
Mel Gormanf83a2752009-05-28 14:34:40 -0700997 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -0700998 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700999 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001000
1001 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -07001002 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001003 struct resv_map *reservations = vma_resv_map(vma);
1004
1005 /* Mark this page used in the map. */
1006 region_add(&reservations->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001007 }
1008}
1009
David Gibson27a85ef2006-03-22 00:08:56 -08001010static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001011 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012{
Andi Kleena5516432008-07-23 21:27:41 -07001013 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001014 struct page *page;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001015 struct address_space *mapping = vma->vm_file->f_mapping;
Mel Gormana1e78772008-07-23 21:27:23 -07001016 struct inode *inode = mapping->host;
Roel Kluine2f17d92009-03-31 15:23:15 -07001017 long chg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001018
Mel Gormana1e78772008-07-23 21:27:23 -07001019 /*
1020 * Processes that did not create the mapping will have no reserves and
1021 * will not have accounted against quota. Check that the quota can be
1022 * made before satisfying the allocation
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001023 * MAP_NORESERVE mappings may also need pages and quota allocated
1024 * if no reserve mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001025 */
Andi Kleena5516432008-07-23 21:27:41 -07001026 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001027 if (chg < 0)
1028 return ERR_PTR(chg);
1029 if (chg)
Mel Gormana1e78772008-07-23 21:27:23 -07001030 if (hugetlb_get_quota(inode->i_mapping, chg))
1031 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001032
1033 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001034 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
Mel Gormana1e78772008-07-23 21:27:23 -07001035 spin_unlock(&hugetlb_lock);
1036
1037 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07001038 page = alloc_buddy_huge_page(h, vma, addr);
Mel Gormana1e78772008-07-23 21:27:23 -07001039 if (!page) {
1040 hugetlb_put_quota(inode->i_mapping, chg);
1041 return ERR_PTR(-VM_FAULT_OOM);
1042 }
1043 }
1044
1045 set_page_refcounted(page);
1046 set_page_private(page, (unsigned long) mapping);
1047
Andi Kleena5516432008-07-23 21:27:41 -07001048 vma_commit_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001049
Adam Litke90d8b7e2007-11-14 16:59:42 -08001050 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001051}
1052
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001053int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001054{
1055 struct huge_bootmem_page *m;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001056 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
Andi Kleenaa888a72008-07-23 21:27:47 -07001057
1058 while (nr_nodes) {
1059 void *addr;
1060
1061 addr = __alloc_bootmem_node_nopanic(
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001062 NODE_DATA(hstate_next_node_to_alloc(h,
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001063 &node_states[N_HIGH_MEMORY])),
Andi Kleenaa888a72008-07-23 21:27:47 -07001064 huge_page_size(h), huge_page_size(h), 0);
1065
1066 if (addr) {
1067 /*
1068 * Use the beginning of the huge page to store the
1069 * huge_bootmem_page struct (until gather_bootmem
1070 * puts them into the mem_map).
1071 */
1072 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001073 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001074 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001075 nr_nodes--;
1076 }
1077 return 0;
1078
1079found:
1080 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1081 /* Put them into a private list first because mem_map is not up yet */
1082 list_add(&m->list, &huge_boot_pages);
1083 m->hstate = h;
1084 return 1;
1085}
1086
Andy Whitcroft18229df2008-11-06 12:53:27 -08001087static void prep_compound_huge_page(struct page *page, int order)
1088{
1089 if (unlikely(order > (MAX_ORDER - 1)))
1090 prep_compound_gigantic_page(page, order);
1091 else
1092 prep_compound_page(page, order);
1093}
1094
Andi Kleenaa888a72008-07-23 21:27:47 -07001095/* Put bootmem huge pages into the standard lists after mem_map is up */
1096static void __init gather_bootmem_prealloc(void)
1097{
1098 struct huge_bootmem_page *m;
1099
1100 list_for_each_entry(m, &huge_boot_pages, list) {
1101 struct page *page = virt_to_page(m);
1102 struct hstate *h = m->hstate;
1103 __ClearPageReserved(page);
1104 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001105 prep_compound_huge_page(page, h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001106 prep_new_huge_page(h, page, page_to_nid(page));
1107 }
1108}
1109
Andi Kleen8faa8b02008-07-23 21:27:48 -07001110static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111{
1112 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113
Andi Kleene5ff2152008-07-23 21:27:42 -07001114 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001115 if (h->order >= MAX_ORDER) {
1116 if (!alloc_bootmem_huge_page(h))
1117 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001118 } else if (!alloc_fresh_huge_page(h,
1119 &node_states[N_HIGH_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001122 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001123}
1124
1125static void __init hugetlb_init_hstates(void)
1126{
1127 struct hstate *h;
1128
1129 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001130 /* oversize hugepages were init'ed in early boot */
1131 if (h->order < MAX_ORDER)
1132 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001133 }
1134}
1135
Andi Kleen4abd32d2008-07-23 21:27:49 -07001136static char * __init memfmt(char *buf, unsigned long n)
1137{
1138 if (n >= (1UL << 30))
1139 sprintf(buf, "%lu GB", n >> 30);
1140 else if (n >= (1UL << 20))
1141 sprintf(buf, "%lu MB", n >> 20);
1142 else
1143 sprintf(buf, "%lu KB", n >> 10);
1144 return buf;
1145}
1146
Andi Kleene5ff2152008-07-23 21:27:42 -07001147static void __init report_hugepages(void)
1148{
1149 struct hstate *h;
1150
1151 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001152 char buf[32];
1153 printk(KERN_INFO "HugeTLB registered %s page size, "
1154 "pre-allocated %ld pages\n",
1155 memfmt(buf, huge_page_size(h)),
1156 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001157 }
1158}
1159
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001161static void try_to_free_low(struct hstate *h, unsigned long count,
1162 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001164 int i;
1165
Andi Kleenaa888a72008-07-23 21:27:47 -07001166 if (h->order >= MAX_ORDER)
1167 return;
1168
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001169 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001171 struct list_head *freel = &h->hugepage_freelists[i];
1172 list_for_each_entry_safe(page, next, freel, lru) {
1173 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001174 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 if (PageHighMem(page))
1176 continue;
1177 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001178 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001179 h->free_huge_pages--;
1180 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 }
1182 }
1183}
1184#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001185static inline void try_to_free_low(struct hstate *h, unsigned long count,
1186 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187{
1188}
1189#endif
1190
Wu Fengguang20a03072009-06-16 15:32:22 -07001191/*
1192 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1193 * balanced by operating on them in a round-robin fashion.
1194 * Returns 1 if an adjustment was made.
1195 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001196static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1197 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001198{
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001199 int start_nid, next_nid;
Wu Fengguang20a03072009-06-16 15:32:22 -07001200 int ret = 0;
1201
1202 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001203
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001204 if (delta < 0)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001205 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001206 else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001207 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001208 next_nid = start_nid;
1209
1210 do {
1211 int nid = next_nid;
1212 if (delta < 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001213 /*
1214 * To shrink on this node, there must be a surplus page
1215 */
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001216 if (!h->surplus_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001217 next_nid = hstate_next_node_to_alloc(h,
1218 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001219 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001220 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001221 }
1222 if (delta > 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001223 /*
1224 * Surplus cannot exceed the total number of pages
1225 */
1226 if (h->surplus_huge_pages_node[nid] >=
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001227 h->nr_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001228 next_nid = hstate_next_node_to_free(h,
1229 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001230 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001231 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001232 }
Wu Fengguang20a03072009-06-16 15:32:22 -07001233
1234 h->surplus_huge_pages += delta;
1235 h->surplus_huge_pages_node[nid] += delta;
1236 ret = 1;
1237 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001238 } while (next_nid != start_nid);
Wu Fengguang20a03072009-06-16 15:32:22 -07001239
Wu Fengguang20a03072009-06-16 15:32:22 -07001240 return ret;
1241}
1242
Andi Kleena5516432008-07-23 21:27:41 -07001243#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001244static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1245 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246{
Adam Litke7893d1d2007-10-16 01:26:18 -07001247 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248
Andi Kleenaa888a72008-07-23 21:27:47 -07001249 if (h->order >= MAX_ORDER)
1250 return h->max_huge_pages;
1251
Adam Litke7893d1d2007-10-16 01:26:18 -07001252 /*
1253 * Increase the pool size
1254 * First take pages out of surplus state. Then make up the
1255 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001256 *
1257 * We might race with alloc_buddy_huge_page() here and be unable
1258 * to convert a surplus huge page to a normal huge page. That is
1259 * not critical, though, it just means the overall size of the
1260 * pool might be one hugepage larger than it needs to be, but
1261 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001262 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001264 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001265 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001266 break;
1267 }
1268
Andi Kleena5516432008-07-23 21:27:41 -07001269 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001270 /*
1271 * If this allocation races such that we no longer need the
1272 * page, free_huge_page will handle it by freeing the page
1273 * and reducing the surplus.
1274 */
1275 spin_unlock(&hugetlb_lock);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001276 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001277 spin_lock(&hugetlb_lock);
1278 if (!ret)
1279 goto out;
1280
Mel Gorman536240f2009-12-14 17:59:56 -08001281 /* Bail for signals. Probably ctrl-c from user */
1282 if (signal_pending(current))
1283 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001284 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001285
1286 /*
1287 * Decrease the pool size
1288 * First return free pages to the buddy allocator (being careful
1289 * to keep enough around to satisfy reservations). Then place
1290 * pages into surplus state as needed so the pool will shrink
1291 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001292 *
1293 * By placing pages into the surplus state independent of the
1294 * overcommit value, we are allowing the surplus pool size to
1295 * exceed overcommit. There are few sane options here. Since
1296 * alloc_buddy_huge_page() is checking the global counter,
1297 * though, we'll note that we're not allowed to exceed surplus
1298 * and won't grow the pool anywhere else. Not until one of the
1299 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001300 */
Andi Kleena5516432008-07-23 21:27:41 -07001301 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001302 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001303 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001304 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001305 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307 }
Andi Kleena5516432008-07-23 21:27:41 -07001308 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001309 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001310 break;
1311 }
1312out:
Andi Kleena5516432008-07-23 21:27:41 -07001313 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001315 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316}
1317
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001318#define HSTATE_ATTR_RO(_name) \
1319 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1320
1321#define HSTATE_ATTR(_name) \
1322 static struct kobj_attribute _name##_attr = \
1323 __ATTR(_name, 0644, _name##_show, _name##_store)
1324
1325static struct kobject *hugepages_kobj;
1326static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1327
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001328static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1329
1330static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001331{
1332 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001333
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001334 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001335 if (hstate_kobjs[i] == kobj) {
1336 if (nidp)
1337 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001338 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001339 }
1340
1341 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001342}
1343
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001344static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001345 struct kobj_attribute *attr, char *buf)
1346{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001347 struct hstate *h;
1348 unsigned long nr_huge_pages;
1349 int nid;
1350
1351 h = kobj_to_hstate(kobj, &nid);
1352 if (nid == NUMA_NO_NODE)
1353 nr_huge_pages = h->nr_huge_pages;
1354 else
1355 nr_huge_pages = h->nr_huge_pages_node[nid];
1356
1357 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001358}
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001359static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1360 struct kobject *kobj, struct kobj_attribute *attr,
1361 const char *buf, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001362{
1363 int err;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001364 int nid;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001365 unsigned long count;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001366 struct hstate *h;
David Rientjesbad44b52009-12-14 17:58:38 -08001367 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001368
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001369 err = strict_strtoul(buf, 10, &count);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001370 if (err)
1371 return 0;
1372
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001373 h = kobj_to_hstate(kobj, &nid);
1374 if (nid == NUMA_NO_NODE) {
1375 /*
1376 * global hstate attribute
1377 */
1378 if (!(obey_mempolicy &&
1379 init_nodemask_of_mempolicy(nodes_allowed))) {
1380 NODEMASK_FREE(nodes_allowed);
1381 nodes_allowed = &node_states[N_HIGH_MEMORY];
1382 }
1383 } else if (nodes_allowed) {
1384 /*
1385 * per node hstate attribute: adjust count to global,
1386 * but restrict alloc/free to the specified node.
1387 */
1388 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1389 init_nodemask_of_node(nodes_allowed, nid);
1390 } else
1391 nodes_allowed = &node_states[N_HIGH_MEMORY];
1392
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001393 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001394
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001395 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001396 NODEMASK_FREE(nodes_allowed);
1397
1398 return len;
1399}
1400
1401static ssize_t nr_hugepages_show(struct kobject *kobj,
1402 struct kobj_attribute *attr, char *buf)
1403{
1404 return nr_hugepages_show_common(kobj, attr, buf);
1405}
1406
1407static ssize_t nr_hugepages_store(struct kobject *kobj,
1408 struct kobj_attribute *attr, const char *buf, size_t len)
1409{
1410 return nr_hugepages_store_common(false, kobj, attr, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001411}
1412HSTATE_ATTR(nr_hugepages);
1413
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001414#ifdef CONFIG_NUMA
1415
1416/*
1417 * hstate attribute for optionally mempolicy-based constraint on persistent
1418 * huge page alloc/free.
1419 */
1420static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1421 struct kobj_attribute *attr, char *buf)
1422{
1423 return nr_hugepages_show_common(kobj, attr, buf);
1424}
1425
1426static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1427 struct kobj_attribute *attr, const char *buf, size_t len)
1428{
1429 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1430}
1431HSTATE_ATTR(nr_hugepages_mempolicy);
1432#endif
1433
1434
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001435static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1436 struct kobj_attribute *attr, char *buf)
1437{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001438 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001439 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1440}
1441static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1442 struct kobj_attribute *attr, const char *buf, size_t count)
1443{
1444 int err;
1445 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001446 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001447
1448 err = strict_strtoul(buf, 10, &input);
1449 if (err)
1450 return 0;
1451
1452 spin_lock(&hugetlb_lock);
1453 h->nr_overcommit_huge_pages = input;
1454 spin_unlock(&hugetlb_lock);
1455
1456 return count;
1457}
1458HSTATE_ATTR(nr_overcommit_hugepages);
1459
1460static ssize_t free_hugepages_show(struct kobject *kobj,
1461 struct kobj_attribute *attr, char *buf)
1462{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001463 struct hstate *h;
1464 unsigned long free_huge_pages;
1465 int nid;
1466
1467 h = kobj_to_hstate(kobj, &nid);
1468 if (nid == NUMA_NO_NODE)
1469 free_huge_pages = h->free_huge_pages;
1470 else
1471 free_huge_pages = h->free_huge_pages_node[nid];
1472
1473 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001474}
1475HSTATE_ATTR_RO(free_hugepages);
1476
1477static ssize_t resv_hugepages_show(struct kobject *kobj,
1478 struct kobj_attribute *attr, char *buf)
1479{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001480 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001481 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1482}
1483HSTATE_ATTR_RO(resv_hugepages);
1484
1485static ssize_t surplus_hugepages_show(struct kobject *kobj,
1486 struct kobj_attribute *attr, char *buf)
1487{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001488 struct hstate *h;
1489 unsigned long surplus_huge_pages;
1490 int nid;
1491
1492 h = kobj_to_hstate(kobj, &nid);
1493 if (nid == NUMA_NO_NODE)
1494 surplus_huge_pages = h->surplus_huge_pages;
1495 else
1496 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1497
1498 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001499}
1500HSTATE_ATTR_RO(surplus_hugepages);
1501
1502static struct attribute *hstate_attrs[] = {
1503 &nr_hugepages_attr.attr,
1504 &nr_overcommit_hugepages_attr.attr,
1505 &free_hugepages_attr.attr,
1506 &resv_hugepages_attr.attr,
1507 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001508#ifdef CONFIG_NUMA
1509 &nr_hugepages_mempolicy_attr.attr,
1510#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001511 NULL,
1512};
1513
1514static struct attribute_group hstate_attr_group = {
1515 .attrs = hstate_attrs,
1516};
1517
Jeff Mahoney094e9532010-02-02 13:44:14 -08001518static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1519 struct kobject **hstate_kobjs,
1520 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001521{
1522 int retval;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001523 int hi = h - hstates;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001524
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001525 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1526 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001527 return -ENOMEM;
1528
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001529 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001530 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001531 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001532
1533 return retval;
1534}
1535
1536static void __init hugetlb_sysfs_init(void)
1537{
1538 struct hstate *h;
1539 int err;
1540
1541 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1542 if (!hugepages_kobj)
1543 return;
1544
1545 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001546 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1547 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001548 if (err)
1549 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1550 h->name);
1551 }
1552}
1553
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001554#ifdef CONFIG_NUMA
1555
1556/*
1557 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1558 * with node sysdevs in node_devices[] using a parallel array. The array
1559 * index of a node sysdev or _hstate == node id.
1560 * This is here to avoid any static dependency of the node sysdev driver, in
1561 * the base kernel, on the hugetlb module.
1562 */
1563struct node_hstate {
1564 struct kobject *hugepages_kobj;
1565 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1566};
1567struct node_hstate node_hstates[MAX_NUMNODES];
1568
1569/*
1570 * A subset of global hstate attributes for node sysdevs
1571 */
1572static struct attribute *per_node_hstate_attrs[] = {
1573 &nr_hugepages_attr.attr,
1574 &free_hugepages_attr.attr,
1575 &surplus_hugepages_attr.attr,
1576 NULL,
1577};
1578
1579static struct attribute_group per_node_hstate_attr_group = {
1580 .attrs = per_node_hstate_attrs,
1581};
1582
1583/*
1584 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1585 * Returns node id via non-NULL nidp.
1586 */
1587static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1588{
1589 int nid;
1590
1591 for (nid = 0; nid < nr_node_ids; nid++) {
1592 struct node_hstate *nhs = &node_hstates[nid];
1593 int i;
1594 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1595 if (nhs->hstate_kobjs[i] == kobj) {
1596 if (nidp)
1597 *nidp = nid;
1598 return &hstates[i];
1599 }
1600 }
1601
1602 BUG();
1603 return NULL;
1604}
1605
1606/*
1607 * Unregister hstate attributes from a single node sysdev.
1608 * No-op if no hstate attributes attached.
1609 */
1610void hugetlb_unregister_node(struct node *node)
1611{
1612 struct hstate *h;
1613 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1614
1615 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001616 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001617
1618 for_each_hstate(h)
1619 if (nhs->hstate_kobjs[h - hstates]) {
1620 kobject_put(nhs->hstate_kobjs[h - hstates]);
1621 nhs->hstate_kobjs[h - hstates] = NULL;
1622 }
1623
1624 kobject_put(nhs->hugepages_kobj);
1625 nhs->hugepages_kobj = NULL;
1626}
1627
1628/*
1629 * hugetlb module exit: unregister hstate attributes from node sysdevs
1630 * that have them.
1631 */
1632static void hugetlb_unregister_all_nodes(void)
1633{
1634 int nid;
1635
1636 /*
1637 * disable node sysdev registrations.
1638 */
1639 register_hugetlbfs_with_node(NULL, NULL);
1640
1641 /*
1642 * remove hstate attributes from any nodes that have them.
1643 */
1644 for (nid = 0; nid < nr_node_ids; nid++)
1645 hugetlb_unregister_node(&node_devices[nid]);
1646}
1647
1648/*
1649 * Register hstate attributes for a single node sysdev.
1650 * No-op if attributes already registered.
1651 */
1652void hugetlb_register_node(struct node *node)
1653{
1654 struct hstate *h;
1655 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1656 int err;
1657
1658 if (nhs->hugepages_kobj)
1659 return; /* already allocated */
1660
1661 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1662 &node->sysdev.kobj);
1663 if (!nhs->hugepages_kobj)
1664 return;
1665
1666 for_each_hstate(h) {
1667 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1668 nhs->hstate_kobjs,
1669 &per_node_hstate_attr_group);
1670 if (err) {
1671 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1672 " for node %d\n",
1673 h->name, node->sysdev.id);
1674 hugetlb_unregister_node(node);
1675 break;
1676 }
1677 }
1678}
1679
1680/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001681 * hugetlb init time: register hstate attributes for all registered node
1682 * sysdevs of nodes that have memory. All on-line nodes should have
1683 * registered their associated sysdev by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001684 */
1685static void hugetlb_register_all_nodes(void)
1686{
1687 int nid;
1688
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001689 for_each_node_state(nid, N_HIGH_MEMORY) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001690 struct node *node = &node_devices[nid];
1691 if (node->sysdev.id == nid)
1692 hugetlb_register_node(node);
1693 }
1694
1695 /*
1696 * Let the node sysdev driver know we're here so it can
1697 * [un]register hstate attributes on node hotplug.
1698 */
1699 register_hugetlbfs_with_node(hugetlb_register_node,
1700 hugetlb_unregister_node);
1701}
1702#else /* !CONFIG_NUMA */
1703
1704static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1705{
1706 BUG();
1707 if (nidp)
1708 *nidp = -1;
1709 return NULL;
1710}
1711
1712static void hugetlb_unregister_all_nodes(void) { }
1713
1714static void hugetlb_register_all_nodes(void) { }
1715
1716#endif
1717
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001718static void __exit hugetlb_exit(void)
1719{
1720 struct hstate *h;
1721
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001722 hugetlb_unregister_all_nodes();
1723
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001724 for_each_hstate(h) {
1725 kobject_put(hstate_kobjs[h - hstates]);
1726 }
1727
1728 kobject_put(hugepages_kobj);
1729}
1730module_exit(hugetlb_exit);
1731
1732static int __init hugetlb_init(void)
1733{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001734 /* Some platform decide whether they support huge pages at boot
1735 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1736 * there is no such support
1737 */
1738 if (HPAGE_SHIFT == 0)
1739 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001740
Nick Piggine11bfbf2008-07-23 21:27:52 -07001741 if (!size_to_hstate(default_hstate_size)) {
1742 default_hstate_size = HPAGE_SIZE;
1743 if (!size_to_hstate(default_hstate_size))
1744 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001745 }
Nick Piggine11bfbf2008-07-23 21:27:52 -07001746 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1747 if (default_hstate_max_huge_pages)
1748 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001749
1750 hugetlb_init_hstates();
1751
Andi Kleenaa888a72008-07-23 21:27:47 -07001752 gather_bootmem_prealloc();
1753
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001754 report_hugepages();
1755
1756 hugetlb_sysfs_init();
1757
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001758 hugetlb_register_all_nodes();
1759
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001760 return 0;
1761}
1762module_init(hugetlb_init);
1763
1764/* Should be called on processing a hugepagesz=... option */
1765void __init hugetlb_add_hstate(unsigned order)
1766{
1767 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001768 unsigned long i;
1769
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001770 if (size_to_hstate(PAGE_SIZE << order)) {
1771 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1772 return;
1773 }
1774 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1775 BUG_ON(order == 0);
1776 h = &hstates[max_hstate++];
1777 h->order = order;
1778 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001779 h->nr_huge_pages = 0;
1780 h->free_huge_pages = 0;
1781 for (i = 0; i < MAX_NUMNODES; ++i)
1782 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001783 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1784 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001785 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1786 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001787
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001788 parsed_hstate = h;
1789}
1790
Nick Piggine11bfbf2008-07-23 21:27:52 -07001791static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001792{
1793 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001794 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001795
1796 /*
1797 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1798 * so this hugepages= parameter goes to the "default hstate".
1799 */
1800 if (!max_hstate)
1801 mhp = &default_hstate_max_huge_pages;
1802 else
1803 mhp = &parsed_hstate->max_huge_pages;
1804
Andi Kleen8faa8b02008-07-23 21:27:48 -07001805 if (mhp == last_mhp) {
1806 printk(KERN_WARNING "hugepages= specified twice without "
1807 "interleaving hugepagesz=, ignoring\n");
1808 return 1;
1809 }
1810
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001811 if (sscanf(s, "%lu", mhp) <= 0)
1812 *mhp = 0;
1813
Andi Kleen8faa8b02008-07-23 21:27:48 -07001814 /*
1815 * Global state is always initialized later in hugetlb_init.
1816 * But we need to allocate >= MAX_ORDER hstates here early to still
1817 * use the bootmem allocator.
1818 */
1819 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1820 hugetlb_hstate_alloc_pages(parsed_hstate);
1821
1822 last_mhp = mhp;
1823
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001824 return 1;
1825}
Nick Piggine11bfbf2008-07-23 21:27:52 -07001826__setup("hugepages=", hugetlb_nrpages_setup);
1827
1828static int __init hugetlb_default_setup(char *s)
1829{
1830 default_hstate_size = memparse(s, &s);
1831 return 1;
1832}
1833__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001834
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07001835static unsigned int cpuset_mems_nr(unsigned int *array)
1836{
1837 int node;
1838 unsigned int nr = 0;
1839
1840 for_each_node_mask(node, cpuset_current_mems_allowed)
1841 nr += array[node];
1842
1843 return nr;
1844}
1845
1846#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001847static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1848 struct ctl_table *table, int write,
1849 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850{
Andi Kleene5ff2152008-07-23 21:27:42 -07001851 struct hstate *h = &default_hstate;
1852 unsigned long tmp;
1853
1854 if (!write)
1855 tmp = h->max_huge_pages;
1856
1857 table->data = &tmp;
1858 table->maxlen = sizeof(unsigned long);
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001859 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001860
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001861 if (write) {
David Rientjesbad44b52009-12-14 17:58:38 -08001862 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1863 GFP_KERNEL | __GFP_NORETRY);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001864 if (!(obey_mempolicy &&
1865 init_nodemask_of_mempolicy(nodes_allowed))) {
1866 NODEMASK_FREE(nodes_allowed);
1867 nodes_allowed = &node_states[N_HIGH_MEMORY];
1868 }
1869 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1870
1871 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1872 NODEMASK_FREE(nodes_allowed);
1873 }
Andi Kleene5ff2152008-07-23 21:27:42 -07001874
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 return 0;
1876}
Mel Gorman396faf02007-07-17 04:03:13 -07001877
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001878int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1879 void __user *buffer, size_t *length, loff_t *ppos)
1880{
1881
1882 return hugetlb_sysctl_handler_common(false, table, write,
1883 buffer, length, ppos);
1884}
1885
1886#ifdef CONFIG_NUMA
1887int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1888 void __user *buffer, size_t *length, loff_t *ppos)
1889{
1890 return hugetlb_sysctl_handler_common(true, table, write,
1891 buffer, length, ppos);
1892}
1893#endif /* CONFIG_NUMA */
1894
Mel Gorman396faf02007-07-17 04:03:13 -07001895int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001896 void __user *buffer,
Mel Gorman396faf02007-07-17 04:03:13 -07001897 size_t *length, loff_t *ppos)
1898{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001899 proc_dointvec(table, write, buffer, length, ppos);
Mel Gorman396faf02007-07-17 04:03:13 -07001900 if (hugepages_treat_as_movable)
1901 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1902 else
1903 htlb_alloc_mask = GFP_HIGHUSER;
1904 return 0;
1905}
1906
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001907int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001908 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001909 size_t *length, loff_t *ppos)
1910{
Andi Kleena5516432008-07-23 21:27:41 -07001911 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07001912 unsigned long tmp;
1913
1914 if (!write)
1915 tmp = h->nr_overcommit_huge_pages;
1916
1917 table->data = &tmp;
1918 table->maxlen = sizeof(unsigned long);
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001919 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001920
1921 if (write) {
1922 spin_lock(&hugetlb_lock);
1923 h->nr_overcommit_huge_pages = tmp;
1924 spin_unlock(&hugetlb_lock);
1925 }
1926
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001927 return 0;
1928}
1929
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930#endif /* CONFIG_SYSCTL */
1931
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001932void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933{
Andi Kleena5516432008-07-23 21:27:41 -07001934 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001935 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001936 "HugePages_Total: %5lu\n"
1937 "HugePages_Free: %5lu\n"
1938 "HugePages_Rsvd: %5lu\n"
1939 "HugePages_Surp: %5lu\n"
1940 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07001941 h->nr_huge_pages,
1942 h->free_huge_pages,
1943 h->resv_huge_pages,
1944 h->surplus_huge_pages,
1945 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946}
1947
1948int hugetlb_report_node_meminfo(int nid, char *buf)
1949{
Andi Kleena5516432008-07-23 21:27:41 -07001950 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951 return sprintf(buf,
1952 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07001953 "Node %d HugePages_Free: %5u\n"
1954 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07001955 nid, h->nr_huge_pages_node[nid],
1956 nid, h->free_huge_pages_node[nid],
1957 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958}
1959
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1961unsigned long hugetlb_total_pages(void)
1962{
Andi Kleena5516432008-07-23 21:27:41 -07001963 struct hstate *h = &default_hstate;
1964 return h->nr_huge_pages * pages_per_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001966
Andi Kleena5516432008-07-23 21:27:41 -07001967static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001968{
1969 int ret = -ENOMEM;
1970
1971 spin_lock(&hugetlb_lock);
1972 /*
1973 * When cpuset is configured, it breaks the strict hugetlb page
1974 * reservation as the accounting is done on a global variable. Such
1975 * reservation is completely rubbish in the presence of cpuset because
1976 * the reservation is not checked against page availability for the
1977 * current cpuset. Application can still potentially OOM'ed by kernel
1978 * with lack of free htlb page in cpuset that the task is in.
1979 * Attempt to enforce strict accounting with cpuset is almost
1980 * impossible (or too ugly) because cpuset is too fluid that
1981 * task or memory node can be dynamically moved between cpusets.
1982 *
1983 * The change of semantics for shared hugetlb mapping with cpuset is
1984 * undesirable. However, in order to preserve some of the semantics,
1985 * we fall back to check against current free page availability as
1986 * a best attempt and hopefully to minimize the impact of changing
1987 * semantics that cpuset has.
1988 */
1989 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001990 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001991 goto out;
1992
Andi Kleena5516432008-07-23 21:27:41 -07001993 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1994 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001995 goto out;
1996 }
1997 }
1998
1999 ret = 0;
2000 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002001 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002002
2003out:
2004 spin_unlock(&hugetlb_lock);
2005 return ret;
2006}
2007
Andy Whitcroft84afd992008-07-23 21:27:32 -07002008static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2009{
2010 struct resv_map *reservations = vma_resv_map(vma);
2011
2012 /*
2013 * This new VMA should share its siblings reservation map if present.
2014 * The VMA will only ever have a valid reservation map pointer where
2015 * it is being copied for another still existing VMA. As that VMA
2016 * has a reference to the reservation map it cannot dissappear until
2017 * after this open call completes. It is therefore safe to take a
2018 * new reference here without additional locking.
2019 */
2020 if (reservations)
2021 kref_get(&reservations->refs);
2022}
2023
Mel Gormana1e78772008-07-23 21:27:23 -07002024static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2025{
Andi Kleena5516432008-07-23 21:27:41 -07002026 struct hstate *h = hstate_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002027 struct resv_map *reservations = vma_resv_map(vma);
2028 unsigned long reserve;
2029 unsigned long start;
2030 unsigned long end;
2031
2032 if (reservations) {
Andi Kleena5516432008-07-23 21:27:41 -07002033 start = vma_hugecache_offset(h, vma, vma->vm_start);
2034 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002035
2036 reserve = (end - start) -
2037 region_count(&reservations->regions, start, end);
2038
2039 kref_put(&reservations->refs, resv_map_release);
2040
Adam Litke7251ff782008-07-23 21:27:59 -07002041 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07002042 hugetlb_acct_memory(h, -reserve);
Adam Litke7251ff782008-07-23 21:27:59 -07002043 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2044 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002045 }
Mel Gormana1e78772008-07-23 21:27:23 -07002046}
2047
Linus Torvalds1da177e2005-04-16 15:20:36 -07002048/*
2049 * We cannot handle pagefaults against hugetlb pages at all. They cause
2050 * handle_mm_fault() to try to instantiate regular-sized pages in the
2051 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2052 * this far.
2053 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002054static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055{
2056 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002057 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058}
2059
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002060const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002061 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002062 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002063 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064};
2065
David Gibson1e8f8892006-01-06 00:10:44 -08002066static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2067 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002068{
2069 pte_t entry;
2070
David Gibson1e8f8892006-01-06 00:10:44 -08002071 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -07002072 entry =
2073 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2074 } else {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002075 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002076 }
2077 entry = pte_mkyoung(entry);
2078 entry = pte_mkhuge(entry);
2079
2080 return entry;
2081}
2082
David Gibson1e8f8892006-01-06 00:10:44 -08002083static void set_huge_ptep_writable(struct vm_area_struct *vma,
2084 unsigned long address, pte_t *ptep)
2085{
2086 pte_t entry;
2087
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002088 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2089 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07002090 update_mmu_cache(vma, address, entry);
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07002091 }
David Gibson1e8f8892006-01-06 00:10:44 -08002092}
2093
2094
David Gibson63551ae2005-06-21 17:14:44 -07002095int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2096 struct vm_area_struct *vma)
2097{
2098 pte_t *src_pte, *dst_pte, entry;
2099 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002100 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002101 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002102 struct hstate *h = hstate_vma(vma);
2103 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08002104
2105 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002106
Andi Kleena5516432008-07-23 21:27:41 -07002107 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Hugh Dickinsc74df322005-10-29 18:16:23 -07002108 src_pte = huge_pte_offset(src, addr);
2109 if (!src_pte)
2110 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002111 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07002112 if (!dst_pte)
2113 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08002114
2115 /* If the pagetables are shared don't copy or take references */
2116 if (dst_pte == src_pte)
2117 continue;
2118
Hugh Dickinsc74df322005-10-29 18:16:23 -07002119 spin_lock(&dst->page_table_lock);
Nick Piggin46478752008-06-05 22:45:57 -07002120 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002121 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002122 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002123 huge_ptep_set_wrprotect(src, addr, src_pte);
2124 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002125 ptepage = pte_page(entry);
2126 get_page(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002127 set_huge_pte_at(dst, addr, dst_pte, entry);
2128 }
2129 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -07002130 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002131 }
2132 return 0;
2133
2134nomem:
2135 return -ENOMEM;
2136}
2137
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002138void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002139 unsigned long end, struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002140{
2141 struct mm_struct *mm = vma->vm_mm;
2142 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002143 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002144 pte_t pte;
2145 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002146 struct page *tmp;
Andi Kleena5516432008-07-23 21:27:41 -07002147 struct hstate *h = hstate_vma(vma);
2148 unsigned long sz = huge_page_size(h);
2149
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -08002150 /*
2151 * A page gathering list, protected by per file i_mmap_lock. The
2152 * lock is used to avoid list corruption from multiple unmapping
2153 * of the same page since we are using page->lru.
2154 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002155 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -07002156
2157 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002158 BUG_ON(start & ~huge_page_mask(h));
2159 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002160
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07002161 mmu_notifier_invalidate_range_start(mm, start, end);
Hugh Dickins508034a2005-10-29 18:16:30 -07002162 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002163 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002164 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002165 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002166 continue;
2167
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002168 if (huge_pmd_unshare(mm, &address, ptep))
2169 continue;
2170
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002171 /*
2172 * If a reference page is supplied, it is because a specific
2173 * page is being unmapped, not a range. Ensure the page we
2174 * are about to unmap is the actual page of interest.
2175 */
2176 if (ref_page) {
2177 pte = huge_ptep_get(ptep);
2178 if (huge_pte_none(pte))
2179 continue;
2180 page = pte_page(pte);
2181 if (page != ref_page)
2182 continue;
2183
2184 /*
2185 * Mark the VMA as having unmapped its page so that
2186 * future faults in this VMA will fail rather than
2187 * looking like data was lost
2188 */
2189 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2190 }
2191
David Gibsonc7546f82005-08-05 11:59:35 -07002192 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002193 if (huge_pte_none(pte))
David Gibson63551ae2005-06-21 17:14:44 -07002194 continue;
David Gibsonc7546f82005-08-05 11:59:35 -07002195
David Gibson63551ae2005-06-21 17:14:44 -07002196 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -08002197 if (pte_dirty(pte))
2198 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002199 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -07002200 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -07002202 flush_tlb_range(vma, start, end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07002203 mmu_notifier_invalidate_range_end(mm, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002204 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2205 list_del(&page->lru);
2206 put_page(page);
2207 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208}
David Gibson63551ae2005-06-21 17:14:44 -07002209
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002210void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002211 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002212{
Andi Kleena137e1c2008-07-23 21:27:43 -07002213 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2214 __unmap_hugepage_range(vma, start, end, ref_page);
2215 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002216}
2217
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002218/*
2219 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2220 * mappping it owns the reserve page for. The intention is to unmap the page
2221 * from other VMAs and let the children be SIGKILLed if they are faulting the
2222 * same region.
2223 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07002224static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2225 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002226{
Adam Litke75266742008-11-12 13:24:56 -08002227 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002228 struct vm_area_struct *iter_vma;
2229 struct address_space *mapping;
2230 struct prio_tree_iter iter;
2231 pgoff_t pgoff;
2232
2233 /*
2234 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2235 * from page cache lookup which is in HPAGE_SIZE units.
2236 */
Adam Litke75266742008-11-12 13:24:56 -08002237 address = address & huge_page_mask(h);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002238 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2239 + (vma->vm_pgoff >> PAGE_SHIFT);
2240 mapping = (struct address_space *)page_private(page);
2241
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002242 /*
2243 * Take the mapping lock for the duration of the table walk. As
2244 * this mapping should be shared between all the VMAs,
2245 * __unmap_hugepage_range() is called as the lock is already held
2246 */
2247 spin_lock(&mapping->i_mmap_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002248 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2249 /* Do not unmap the current VMA */
2250 if (iter_vma == vma)
2251 continue;
2252
2253 /*
2254 * Unmap the page from other VMAs without their own reserves.
2255 * They get marked to be SIGKILLed if they fault in these
2256 * areas. This is because a future no-page fault on this VMA
2257 * could insert a zeroed page instead of the data existing
2258 * from the time of fork. This would look like data corruption
2259 */
2260 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002261 __unmap_hugepage_range(iter_vma,
Adam Litke75266742008-11-12 13:24:56 -08002262 address, address + huge_page_size(h),
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002263 page);
2264 }
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002265 spin_unlock(&mapping->i_mmap_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002266
2267 return 1;
2268}
2269
David Gibson1e8f8892006-01-06 00:10:44 -08002270static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002271 unsigned long address, pte_t *ptep, pte_t pte,
2272 struct page *pagecache_page)
David Gibson1e8f8892006-01-06 00:10:44 -08002273{
Andi Kleena5516432008-07-23 21:27:41 -07002274 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002275 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -08002276 int avoidcopy;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002277 int outside_reserve = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002278
2279 old_page = pte_page(pte);
2280
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002281retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002282 /* If no-one else is actually using this page, avoid the copy
2283 * and just make the page writable */
2284 avoidcopy = (page_count(old_page) == 1);
2285 if (avoidcopy) {
2286 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002287 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002288 }
2289
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002290 /*
2291 * If the process that created a MAP_PRIVATE mapping is about to
2292 * perform a COW due to a shared page count, attempt to satisfy
2293 * the allocation without using the existing reserves. The pagecache
2294 * page is used to determine if the reserve at this address was
2295 * consumed or not. If reserves were used, a partial faulted mapping
2296 * at the time of fork() could consume its reserves on COW instead
2297 * of the full address range.
2298 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002299 if (!(vma->vm_flags & VM_MAYSHARE) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002300 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2301 old_page != pagecache_page)
2302 outside_reserve = 1;
2303
David Gibson1e8f8892006-01-06 00:10:44 -08002304 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002305
2306 /* Drop page_table_lock as buddy allocator may be called */
2307 spin_unlock(&mm->page_table_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002308 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002309
Adam Litke2fc39ce2007-11-14 16:59:39 -08002310 if (IS_ERR(new_page)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002311 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002312
2313 /*
2314 * If a process owning a MAP_PRIVATE mapping fails to COW,
2315 * it is due to references held by a child and an insufficient
2316 * huge page pool. To guarantee the original mappers
2317 * reliability, unmap the page from child processes. The child
2318 * may get SIGKILLed if it later faults.
2319 */
2320 if (outside_reserve) {
2321 BUG_ON(huge_pte_none(pte));
2322 if (unmap_ref_private(mm, vma, old_page, address)) {
2323 BUG_ON(page_count(old_page) != 1);
2324 BUG_ON(huge_pte_none(pte));
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002325 spin_lock(&mm->page_table_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002326 goto retry_avoidcopy;
2327 }
2328 WARN_ON_ONCE(1);
2329 }
2330
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002331 /* Caller expects lock to be held */
2332 spin_lock(&mm->page_table_lock);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002333 return -PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002334 }
2335
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00002336 copy_huge_page(new_page, old_page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002337 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002338
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002339 /*
2340 * Retake the page_table_lock to check for racing updates
2341 * before the page tables are altered
2342 */
2343 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002344 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002345 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002346 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002347 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002348 set_huge_pte_at(mm, address, ptep,
2349 make_huge_pte(vma, new_page, 1));
2350 /* Make the old page be freed below */
2351 new_page = old_page;
2352 }
2353 page_cache_release(new_page);
2354 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -07002355 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002356}
2357
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002358/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002359static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2360 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002361{
2362 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002363 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002364
2365 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002366 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002367
2368 return find_lock_page(mapping, idx);
2369}
2370
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002371/*
2372 * Return whether there is a pagecache page to back given address within VMA.
2373 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2374 */
2375static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002376 struct vm_area_struct *vma, unsigned long address)
2377{
2378 struct address_space *mapping;
2379 pgoff_t idx;
2380 struct page *page;
2381
2382 mapping = vma->vm_file->f_mapping;
2383 idx = vma_hugecache_offset(h, vma, address);
2384
2385 page = find_get_page(mapping, idx);
2386 if (page)
2387 put_page(page);
2388 return page != NULL;
2389}
2390
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002391static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002392 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002393{
Andi Kleena5516432008-07-23 21:27:41 -07002394 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002395 int ret = VM_FAULT_SIGBUS;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002396 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07002397 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002398 struct page *page;
2399 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08002400 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -07002401
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002402 /*
2403 * Currently, we are forced to kill the process in the event the
2404 * original mapper has unmapped pages from the child due to a failed
2405 * COW. Warn that such a situation has occured as it may not be obvious
2406 */
2407 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2408 printk(KERN_WARNING
2409 "PID %d killed due to inadequate hugepage pool\n",
2410 current->pid);
2411 return ret;
2412 }
2413
Adam Litke4c887262005-10-29 18:16:46 -07002414 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002415 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07002416
2417 /*
2418 * Use page lock to guard against racing truncation
2419 * before we get page_table_lock.
2420 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002421retry:
2422 page = find_lock_page(mapping, idx);
2423 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002424 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002425 if (idx >= size)
2426 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002427 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002428 if (IS_ERR(page)) {
2429 ret = -PTR_ERR(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002430 goto out;
2431 }
Andi Kleena5516432008-07-23 21:27:41 -07002432 clear_huge_page(page, address, huge_page_size(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002433 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002434
Mel Gormanf83a2752009-05-28 14:34:40 -07002435 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002436 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002437 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002438
2439 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2440 if (err) {
2441 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002442 if (err == -EEXIST)
2443 goto retry;
2444 goto out;
2445 }
Ken Chen45c682a2007-11-14 16:59:44 -08002446
2447 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002448 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002449 spin_unlock(&inode->i_lock);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002450 } else
2451 lock_page(page);
2452 }
David Gibson1e8f8892006-01-06 00:10:44 -08002453
Andy Whitcroft57303d82008-08-12 15:08:47 -07002454 /*
2455 * If we are going to COW a private mapping later, we examine the
2456 * pending reservations for this page now. This will ensure that
2457 * any allocations necessary to record that reservation occur outside
2458 * the spinlock.
2459 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002460 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002461 if (vma_needs_reservation(h, vma, address) < 0) {
2462 ret = VM_FAULT_OOM;
2463 goto backout_unlocked;
2464 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002465
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002466 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002467 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002468 if (idx >= size)
2469 goto backout;
2470
Nick Piggin83c54072007-07-19 01:47:05 -07002471 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002472 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002473 goto backout;
2474
David Gibson1e8f8892006-01-06 00:10:44 -08002475 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2476 && (vma->vm_flags & VM_SHARED)));
2477 set_huge_pte_at(mm, address, ptep, new_pte);
2478
Hugh Dickins788c7df2009-06-23 13:49:05 +01002479 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002480 /* Optimization, do the COW without a second fault */
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002481 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
David Gibson1e8f8892006-01-06 00:10:44 -08002482 }
2483
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002484 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07002485 unlock_page(page);
2486out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002487 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002488
2489backout:
2490 spin_unlock(&mm->page_table_lock);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002491backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002492 unlock_page(page);
2493 put_page(page);
2494 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002495}
2496
Adam Litke86e52162006-01-06 00:10:43 -08002497int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002498 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08002499{
2500 pte_t *ptep;
2501 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -08002502 int ret;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002503 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002504 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002505 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002506
Andi Kleena5516432008-07-23 21:27:41 -07002507 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002508 if (!ptep)
2509 return VM_FAULT_OOM;
2510
David Gibson3935baa2006-03-22 00:08:53 -08002511 /*
2512 * Serialize hugepage allocation and instantiation, so that we don't
2513 * get spurious allocation failures if two CPUs race to instantiate
2514 * the same page in the page cache.
2515 */
2516 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002517 entry = huge_ptep_get(ptep);
2518 if (huge_pte_none(entry)) {
Hugh Dickins788c7df2009-06-23 13:49:05 +01002519 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07002520 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002521 }
Adam Litke86e52162006-01-06 00:10:43 -08002522
Nick Piggin83c54072007-07-19 01:47:05 -07002523 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002524
Andy Whitcroft57303d82008-08-12 15:08:47 -07002525 /*
2526 * If we are going to COW the mapping later, we examine the pending
2527 * reservations for this page now. This will ensure that any
2528 * allocations necessary to record that reservation occur outside the
2529 * spinlock. For private mappings, we also lookup the pagecache
2530 * page now as it is used to determine if a reservation has been
2531 * consumed.
2532 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002533 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002534 if (vma_needs_reservation(h, vma, address) < 0) {
2535 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002536 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002537 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002538
Mel Gormanf83a2752009-05-28 14:34:40 -07002539 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07002540 pagecache_page = hugetlbfs_pagecache_page(h,
2541 vma, address);
2542 }
2543
David Gibson1e8f8892006-01-06 00:10:44 -08002544 spin_lock(&mm->page_table_lock);
2545 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002546 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2547 goto out_page_table_lock;
2548
2549
Hugh Dickins788c7df2009-06-23 13:49:05 +01002550 if (flags & FAULT_FLAG_WRITE) {
David Gibsonb4d1d992008-10-15 22:01:11 -07002551 if (!pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002552 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2553 pagecache_page);
David Gibsonb4d1d992008-10-15 22:01:11 -07002554 goto out_page_table_lock;
2555 }
2556 entry = pte_mkdirty(entry);
2557 }
2558 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01002559 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2560 flags & FAULT_FLAG_WRITE))
David Gibsonb4d1d992008-10-15 22:01:11 -07002561 update_mmu_cache(vma, address, entry);
2562
2563out_page_table_lock:
David Gibson1e8f8892006-01-06 00:10:44 -08002564 spin_unlock(&mm->page_table_lock);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002565
2566 if (pagecache_page) {
2567 unlock_page(pagecache_page);
2568 put_page(pagecache_page);
2569 }
2570
David Gibsonb4d1d992008-10-15 22:01:11 -07002571out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08002572 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08002573
2574 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08002575}
2576
Andi Kleenceb86872008-07-23 21:27:50 -07002577/* Can be overriden by architectures */
2578__attribute__((weak)) struct page *
2579follow_huge_pud(struct mm_struct *mm, unsigned long address,
2580 pud_t *pud, int write)
2581{
2582 BUG();
2583 return NULL;
2584}
2585
David Gibson63551ae2005-06-21 17:14:44 -07002586int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2587 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -08002588 unsigned long *position, int *length, int i,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002589 unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07002590{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002591 unsigned long pfn_offset;
2592 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -07002593 int remainder = *length;
Andi Kleena5516432008-07-23 21:27:41 -07002594 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07002595
Hugh Dickins1c598272005-10-19 21:23:43 -07002596 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002597 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07002598 pte_t *pte;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002599 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07002600 struct page *page;
2601
2602 /*
2603 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002604 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07002605 * first, for the page indexing below to work.
2606 */
Andi Kleena5516432008-07-23 21:27:41 -07002607 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002608 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07002609
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002610 /*
2611 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002612 * an error where there's an empty slot with no huge pagecache
2613 * to back it. This way, we avoid allocating a hugepage, and
2614 * the sparse dumpfile avoids allocating disk blocks, but its
2615 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002616 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002617 if (absent && (flags & FOLL_DUMP) &&
2618 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002619 remainder = 0;
2620 break;
2621 }
2622
2623 if (absent ||
2624 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07002625 int ret;
2626
2627 spin_unlock(&mm->page_table_lock);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002628 ret = hugetlb_fault(mm, vma, vaddr,
2629 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litke4c887262005-10-29 18:16:46 -07002630 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07002631 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07002632 continue;
2633
2634 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002635 break;
2636 }
David Gibson63551ae2005-06-21 17:14:44 -07002637
Andi Kleena5516432008-07-23 21:27:41 -07002638 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002639 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002640same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002641 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002642 pages[i] = mem_map_offset(page, pfn_offset);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002643 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002644 }
David Gibson63551ae2005-06-21 17:14:44 -07002645
2646 if (vmas)
2647 vmas[i] = vma;
2648
2649 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002650 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07002651 --remainder;
2652 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002653 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07002654 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002655 /*
2656 * We use pfn_offset to avoid touching the pageframes
2657 * of this compound page.
2658 */
2659 goto same_page;
2660 }
David Gibson63551ae2005-06-21 17:14:44 -07002661 }
Hugh Dickins1c598272005-10-19 21:23:43 -07002662 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002663 *length = remainder;
2664 *position = vaddr;
2665
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002666 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07002667}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002668
2669void hugetlb_change_protection(struct vm_area_struct *vma,
2670 unsigned long address, unsigned long end, pgprot_t newprot)
2671{
2672 struct mm_struct *mm = vma->vm_mm;
2673 unsigned long start = address;
2674 pte_t *ptep;
2675 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07002676 struct hstate *h = hstate_vma(vma);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002677
2678 BUG_ON(address >= end);
2679 flush_cache_range(vma, address, end);
2680
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002681 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002682 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002683 for (; address < end; address += huge_page_size(h)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002684 ptep = huge_pte_offset(mm, address);
2685 if (!ptep)
2686 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002687 if (huge_pmd_unshare(mm, &address, ptep))
2688 continue;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002689 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002690 pte = huge_ptep_get_and_clear(mm, address, ptep);
2691 pte = pte_mkhuge(pte_modify(pte, newprot));
2692 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002693 }
2694 }
2695 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002696 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002697
2698 flush_tlb_range(vma, start, end);
2699}
2700
Mel Gormana1e78772008-07-23 21:27:23 -07002701int hugetlb_reserve_pages(struct inode *inode,
2702 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00002703 struct vm_area_struct *vma,
2704 int acctflag)
Adam Litkee4e574b2007-10-16 01:26:19 -07002705{
Mel Gorman17c9d122009-02-11 16:34:16 +00002706 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07002707 struct hstate *h = hstate_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07002708
Mel Gormana1e78772008-07-23 21:27:23 -07002709 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00002710 * Only apply hugepage reservation if asked. At fault time, an
2711 * attempt will be made for VM_NORESERVE to allocate a page
2712 * and filesystem quota without using reserves
2713 */
2714 if (acctflag & VM_NORESERVE)
2715 return 0;
2716
2717 /*
Mel Gormana1e78772008-07-23 21:27:23 -07002718 * Shared mappings base their reservation on the number of pages that
2719 * are already allocated on behalf of the file. Private mappings need
2720 * to reserve the full area even if read-only as mprotect() may be
2721 * called to make the mapping read-write. Assume !vma is a shm mapping
2722 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002723 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gormana1e78772008-07-23 21:27:23 -07002724 chg = region_chg(&inode->i_mapping->private_list, from, to);
Mel Gorman5a6fe122009-02-10 14:02:27 +00002725 else {
2726 struct resv_map *resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00002727 if (!resv_map)
2728 return -ENOMEM;
2729
Mel Gorman17c9d122009-02-11 16:34:16 +00002730 chg = to - from;
2731
Mel Gorman5a6fe122009-02-10 14:02:27 +00002732 set_vma_resv_map(vma, resv_map);
2733 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2734 }
2735
Mel Gorman17c9d122009-02-11 16:34:16 +00002736 if (chg < 0)
2737 return chg;
2738
2739 /* There must be enough filesystem quota for the mapping */
2740 if (hugetlb_get_quota(inode->i_mapping, chg))
2741 return -ENOSPC;
2742
2743 /*
2744 * Check enough hugepages are available for the reservation.
2745 * Hand back the quota if there are not
2746 */
2747 ret = hugetlb_acct_memory(h, chg);
2748 if (ret < 0) {
2749 hugetlb_put_quota(inode->i_mapping, chg);
2750 return ret;
2751 }
2752
2753 /*
2754 * Account for the reservations made. Shared mappings record regions
2755 * that have reservations as they are shared by multiple VMAs.
2756 * When the last VMA disappears, the region map says how much
2757 * the reservation was and the page cache tells how much of
2758 * the reservation was consumed. Private mappings are per-VMA and
2759 * only the consumed reservations are tracked. When the VMA
2760 * disappears, the original reservation is the VMA size and the
2761 * consumed reservations are stored in the map. Hence, nothing
2762 * else has to be done for private mappings here
2763 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002764 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gorman17c9d122009-02-11 16:34:16 +00002765 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002766 return 0;
2767}
2768
2769void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2770{
Andi Kleena5516432008-07-23 21:27:41 -07002771 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002772 long chg = region_truncate(&inode->i_mapping->private_list, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08002773
2774 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07002775 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08002776 spin_unlock(&inode->i_lock);
2777
Adam Litke90d8b7e2007-11-14 16:59:42 -08002778 hugetlb_put_quota(inode->i_mapping, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07002779 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002780}