blob: e83ad2c9228c1242a582920c3a03b691f5dd56f9 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
4 */
5#include <linux/gfp.h>
6#include <linux/list.h>
7#include <linux/init.h>
8#include <linux/module.h>
9#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +040010#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070011#include <linux/sysctl.h>
12#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070013#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070015#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080016#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080017#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080018#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070019#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070020#include <linux/sysfs.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070021
David Gibson63551ae2005-06-21 17:14:44 -070022#include <asm/page.h>
23#include <asm/pgtable.h>
Adrian Bunk78a34ae2008-07-28 15:46:30 -070024#include <asm/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070025
26#include <linux/hugetlb.h>
Nick Piggin7835e982006-03-22 00:08:40 -080027#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070028
29const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
Mel Gorman396faf02007-07-17 04:03:13 -070030static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
31unsigned long hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070032
Andi Kleene5ff2152008-07-23 21:27:42 -070033static int max_hstate;
34unsigned int default_hstate_idx;
35struct hstate hstates[HUGE_MAX_HSTATE];
36
Jon Tollefson53ba51d2008-07-23 21:27:52 -070037__initdata LIST_HEAD(huge_boot_pages);
38
Andi Kleene5ff2152008-07-23 21:27:42 -070039/* for command line parsing */
40static struct hstate * __initdata parsed_hstate;
41static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070042static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070043
44#define for_each_hstate(h) \
45 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
Mel Gorman396faf02007-07-17 04:03:13 -070046
David Gibson3935baa2006-03-22 00:08:53 -080047/*
48 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
49 */
50static DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080051
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -070052/*
Andy Whitcroft96822902008-07-23 21:27:29 -070053 * Region tracking -- allows tracking of reservations and instantiated pages
54 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -070055 *
56 * The region data structures are protected by a combination of the mmap_sem
57 * and the hugetlb_instantion_mutex. To access or modify a region the caller
58 * must either hold the mmap_sem for write, or the mmap_sem for read and
59 * the hugetlb_instantiation mutex:
60 *
61 * down_write(&mm->mmap_sem);
62 * or
63 * down_read(&mm->mmap_sem);
64 * mutex_lock(&hugetlb_instantiation_mutex);
Andy Whitcroft96822902008-07-23 21:27:29 -070065 */
66struct file_region {
67 struct list_head link;
68 long from;
69 long to;
70};
71
72static long region_add(struct list_head *head, long f, long t)
73{
74 struct file_region *rg, *nrg, *trg;
75
76 /* Locate the region we are either in or before. */
77 list_for_each_entry(rg, head, link)
78 if (f <= rg->to)
79 break;
80
81 /* Round our left edge to the current segment if it encloses us. */
82 if (f > rg->from)
83 f = rg->from;
84
85 /* Check for and consume any regions we now overlap with. */
86 nrg = rg;
87 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
88 if (&rg->link == head)
89 break;
90 if (rg->from > t)
91 break;
92
93 /* If this area reaches higher then extend our area to
94 * include it completely. If this is not the first area
95 * which we intend to reuse, free it. */
96 if (rg->to > t)
97 t = rg->to;
98 if (rg != nrg) {
99 list_del(&rg->link);
100 kfree(rg);
101 }
102 }
103 nrg->from = f;
104 nrg->to = t;
105 return 0;
106}
107
108static long region_chg(struct list_head *head, long f, long t)
109{
110 struct file_region *rg, *nrg;
111 long chg = 0;
112
113 /* Locate the region we are before or in. */
114 list_for_each_entry(rg, head, link)
115 if (f <= rg->to)
116 break;
117
118 /* If we are below the current region then a new region is required.
119 * Subtle, allocate a new region at the position but make it zero
120 * size such that we can guarantee to record the reservation. */
121 if (&rg->link == head || t < rg->from) {
122 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
123 if (!nrg)
124 return -ENOMEM;
125 nrg->from = f;
126 nrg->to = f;
127 INIT_LIST_HEAD(&nrg->link);
128 list_add(&nrg->link, rg->link.prev);
129
130 return t - f;
131 }
132
133 /* Round our left edge to the current segment if it encloses us. */
134 if (f > rg->from)
135 f = rg->from;
136 chg = t - f;
137
138 /* Check for and consume any regions we now overlap with. */
139 list_for_each_entry(rg, rg->link.prev, link) {
140 if (&rg->link == head)
141 break;
142 if (rg->from > t)
143 return chg;
144
145 /* We overlap with this area, if it extends futher than
146 * us then we must extend ourselves. Account for its
147 * existing reservation. */
148 if (rg->to > t) {
149 chg += rg->to - t;
150 t = rg->to;
151 }
152 chg -= rg->to - rg->from;
153 }
154 return chg;
155}
156
157static long region_truncate(struct list_head *head, long end)
158{
159 struct file_region *rg, *trg;
160 long chg = 0;
161
162 /* Locate the region we are either in or before. */
163 list_for_each_entry(rg, head, link)
164 if (end <= rg->to)
165 break;
166 if (&rg->link == head)
167 return 0;
168
169 /* If we are in the middle of a region then adjust it. */
170 if (end > rg->from) {
171 chg = rg->to - end;
172 rg->to = end;
173 rg = list_entry(rg->link.next, typeof(*rg), link);
174 }
175
176 /* Drop any remaining regions. */
177 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
178 if (&rg->link == head)
179 break;
180 chg += rg->to - rg->from;
181 list_del(&rg->link);
182 kfree(rg);
183 }
184 return chg;
185}
186
Andy Whitcroft84afd992008-07-23 21:27:32 -0700187static long region_count(struct list_head *head, long f, long t)
188{
189 struct file_region *rg;
190 long chg = 0;
191
192 /* Locate each segment we overlap with, and count that overlap. */
193 list_for_each_entry(rg, head, link) {
194 int seg_from;
195 int seg_to;
196
197 if (rg->to <= f)
198 continue;
199 if (rg->from >= t)
200 break;
201
202 seg_from = max(rg->from, f);
203 seg_to = min(rg->to, t);
204
205 chg += seg_to - seg_from;
206 }
207
208 return chg;
209}
210
Andy Whitcroft96822902008-07-23 21:27:29 -0700211/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700212 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700213 * the mapping, in pagecache page units; huge pages here.
214 */
Andi Kleena5516432008-07-23 21:27:41 -0700215static pgoff_t vma_hugecache_offset(struct hstate *h,
216 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700217{
Andi Kleena5516432008-07-23 21:27:41 -0700218 return ((address - vma->vm_start) >> huge_page_shift(h)) +
219 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700220}
221
Andy Whitcroft84afd992008-07-23 21:27:32 -0700222/*
Mel Gorman08fba692009-01-06 14:38:53 -0800223 * Return the size of the pages allocated when backing a VMA. In the majority
224 * cases this will be same size as used by the page table entries.
225 */
226unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
227{
228 struct hstate *hstate;
229
230 if (!is_vm_hugetlb_page(vma))
231 return PAGE_SIZE;
232
233 hstate = hstate_vma(vma);
234
235 return 1UL << (hstate->order + PAGE_SHIFT);
236}
237
238/*
Mel Gorman33402892009-01-06 14:38:54 -0800239 * Return the page size being used by the MMU to back a VMA. In the majority
240 * of cases, the page size used by the kernel matches the MMU size. On
241 * architectures where it differs, an architecture-specific version of this
242 * function is required.
243 */
244#ifndef vma_mmu_pagesize
245unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
246{
247 return vma_kernel_pagesize(vma);
248}
249#endif
250
251/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700252 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
253 * bits of the reservation map pointer, which are always clear due to
254 * alignment.
255 */
256#define HPAGE_RESV_OWNER (1UL << 0)
257#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700258#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700259
Mel Gormana1e78772008-07-23 21:27:23 -0700260/*
261 * These helpers are used to track how many pages are reserved for
262 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
263 * is guaranteed to have their future faults succeed.
264 *
265 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
266 * the reserve counters are updated with the hugetlb_lock held. It is safe
267 * to reset the VMA at fork() time as it is not in use yet and there is no
268 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700269 *
270 * The private mapping reservation is represented in a subtly different
271 * manner to a shared mapping. A shared mapping has a region map associated
272 * with the underlying file, this region map represents the backing file
273 * pages which have ever had a reservation assigned which this persists even
274 * after the page is instantiated. A private mapping has a region map
275 * associated with the original mmap which is attached to all VMAs which
276 * reference it, this region map represents those offsets which have consumed
277 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700278 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700279static unsigned long get_vma_private_data(struct vm_area_struct *vma)
280{
281 return (unsigned long)vma->vm_private_data;
282}
283
284static void set_vma_private_data(struct vm_area_struct *vma,
285 unsigned long value)
286{
287 vma->vm_private_data = (void *)value;
288}
289
Andy Whitcroft84afd992008-07-23 21:27:32 -0700290struct resv_map {
291 struct kref refs;
292 struct list_head regions;
293};
294
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700295static struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700296{
297 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
298 if (!resv_map)
299 return NULL;
300
301 kref_init(&resv_map->refs);
302 INIT_LIST_HEAD(&resv_map->regions);
303
304 return resv_map;
305}
306
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700307static void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700308{
309 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
310
311 /* Clear out any active regions before we release the map. */
312 region_truncate(&resv_map->regions, 0);
313 kfree(resv_map);
314}
315
316static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700317{
318 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700319 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft84afd992008-07-23 21:27:32 -0700320 return (struct resv_map *)(get_vma_private_data(vma) &
321 ~HPAGE_RESV_MASK);
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700322 return NULL;
Mel Gormana1e78772008-07-23 21:27:23 -0700323}
324
Andy Whitcroft84afd992008-07-23 21:27:32 -0700325static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700326{
327 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700328 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Mel Gormana1e78772008-07-23 21:27:23 -0700329
Andy Whitcroft84afd992008-07-23 21:27:32 -0700330 set_vma_private_data(vma, (get_vma_private_data(vma) &
331 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700332}
333
334static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
335{
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700336 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700337 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700338
339 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700340}
341
342static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
343{
344 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700345
346 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700347}
348
349/* Decrement the reserved pages in the hugepage pool by one */
Andi Kleena5516432008-07-23 21:27:41 -0700350static void decrement_hugepage_resv_vma(struct hstate *h,
351 struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700352{
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700353 if (vma->vm_flags & VM_NORESERVE)
354 return;
355
Mel Gormanf83a2752009-05-28 14:34:40 -0700356 if (vma->vm_flags & VM_MAYSHARE) {
Mel Gormana1e78772008-07-23 21:27:23 -0700357 /* Shared mappings always use reserves */
Andi Kleena5516432008-07-23 21:27:41 -0700358 h->resv_huge_pages--;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700359 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Mel Gormana1e78772008-07-23 21:27:23 -0700360 /*
361 * Only the process that called mmap() has reserves for
362 * private mappings.
363 */
Andi Kleena5516432008-07-23 21:27:41 -0700364 h->resv_huge_pages--;
Mel Gormana1e78772008-07-23 21:27:23 -0700365 }
366}
367
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700368/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700369void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
370{
371 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700372 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700373 vma->vm_private_data = (void *)0;
374}
375
376/* Returns true if the VMA has associated reserve pages */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700377static int vma_has_reserves(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700378{
Mel Gormanf83a2752009-05-28 14:34:40 -0700379 if (vma->vm_flags & VM_MAYSHARE)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700380 return 1;
381 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
382 return 1;
383 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700384}
385
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800386static void clear_gigantic_page(struct page *page,
387 unsigned long addr, unsigned long sz)
388{
389 int i;
390 struct page *p = page;
391
392 might_sleep();
393 for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
394 cond_resched();
395 clear_user_highpage(p, addr + i * PAGE_SIZE);
396 }
397}
Andi Kleena5516432008-07-23 21:27:41 -0700398static void clear_huge_page(struct page *page,
399 unsigned long addr, unsigned long sz)
David Gibson79ac6ba2006-03-22 00:08:51 -0800400{
401 int i;
402
Hannes Ederebdd4ae2009-01-06 14:39:58 -0800403 if (unlikely(sz > MAX_ORDER_NR_PAGES)) {
404 clear_gigantic_page(page, addr, sz);
405 return;
406 }
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800407
David Gibson79ac6ba2006-03-22 00:08:51 -0800408 might_sleep();
Andi Kleena5516432008-07-23 21:27:41 -0700409 for (i = 0; i < sz/PAGE_SIZE; i++) {
David Gibson79ac6ba2006-03-22 00:08:51 -0800410 cond_resched();
Ralf Baechle281e0e32007-10-01 01:20:10 -0700411 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
David Gibson79ac6ba2006-03-22 00:08:51 -0800412 }
413}
414
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800415static void copy_gigantic_page(struct page *dst, struct page *src,
416 unsigned long addr, struct vm_area_struct *vma)
417{
418 int i;
419 struct hstate *h = hstate_vma(vma);
420 struct page *dst_base = dst;
421 struct page *src_base = src;
422 might_sleep();
423 for (i = 0; i < pages_per_huge_page(h); ) {
424 cond_resched();
425 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
426
427 i++;
428 dst = mem_map_next(dst, dst_base, i);
429 src = mem_map_next(src, src_base, i);
430 }
431}
David Gibson79ac6ba2006-03-22 00:08:51 -0800432static void copy_huge_page(struct page *dst, struct page *src,
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000433 unsigned long addr, struct vm_area_struct *vma)
David Gibson79ac6ba2006-03-22 00:08:51 -0800434{
435 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700436 struct hstate *h = hstate_vma(vma);
David Gibson79ac6ba2006-03-22 00:08:51 -0800437
Hannes Ederebdd4ae2009-01-06 14:39:58 -0800438 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
439 copy_gigantic_page(dst, src, addr, vma);
440 return;
441 }
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800442
David Gibson79ac6ba2006-03-22 00:08:51 -0800443 might_sleep();
Andi Kleena5516432008-07-23 21:27:41 -0700444 for (i = 0; i < pages_per_huge_page(h); i++) {
David Gibson79ac6ba2006-03-22 00:08:51 -0800445 cond_resched();
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000446 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
David Gibson79ac6ba2006-03-22 00:08:51 -0800447 }
448}
449
Andi Kleena5516432008-07-23 21:27:41 -0700450static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451{
452 int nid = page_to_nid(page);
Andi Kleena5516432008-07-23 21:27:41 -0700453 list_add(&page->lru, &h->hugepage_freelists[nid]);
454 h->free_huge_pages++;
455 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456}
457
Andi Kleena5516432008-07-23 21:27:41 -0700458static struct page *dequeue_huge_page(struct hstate *h)
Nishanth Aravamudan348e1e02008-03-04 14:29:42 -0800459{
460 int nid;
461 struct page *page = NULL;
462
463 for (nid = 0; nid < MAX_NUMNODES; ++nid) {
Andi Kleena5516432008-07-23 21:27:41 -0700464 if (!list_empty(&h->hugepage_freelists[nid])) {
465 page = list_entry(h->hugepage_freelists[nid].next,
Nishanth Aravamudan348e1e02008-03-04 14:29:42 -0800466 struct page, lru);
467 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700468 h->free_huge_pages--;
469 h->free_huge_pages_node[nid]--;
Nishanth Aravamudan348e1e02008-03-04 14:29:42 -0800470 break;
471 }
472 }
473 return page;
474}
475
Andi Kleena5516432008-07-23 21:27:41 -0700476static struct page *dequeue_huge_page_vma(struct hstate *h,
477 struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700478 unsigned long address, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479{
Nishanth Aravamudan31a5c6e2007-07-15 23:38:02 -0700480 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700482 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700483 nodemask_t *nodemask;
Mel Gorman396faf02007-07-17 04:03:13 -0700484 struct zonelist *zonelist = huge_zonelist(vma, address,
Mel Gorman19770b32008-04-28 02:12:18 -0700485 htlb_alloc_mask, &mpol, &nodemask);
Mel Gormandd1a2392008-04-28 02:12:17 -0700486 struct zone *zone;
487 struct zoneref *z;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488
Mel Gormana1e78772008-07-23 21:27:23 -0700489 /*
490 * A child process with MAP_PRIVATE mappings created by their parent
491 * have no page reserves. This check ensures that reservations are
492 * not "stolen". The child may still get SIGKILLed
493 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700494 if (!vma_has_reserves(vma) &&
Andi Kleena5516432008-07-23 21:27:41 -0700495 h->free_huge_pages - h->resv_huge_pages == 0)
Mel Gormana1e78772008-07-23 21:27:23 -0700496 return NULL;
497
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700498 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700499 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700500 return NULL;
501
Mel Gorman19770b32008-04-28 02:12:18 -0700502 for_each_zone_zonelist_nodemask(zone, z, zonelist,
503 MAX_NR_ZONES - 1, nodemask) {
Mel Gorman54a6eb52008-04-28 02:12:16 -0700504 nid = zone_to_nid(zone);
505 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
Andi Kleena5516432008-07-23 21:27:41 -0700506 !list_empty(&h->hugepage_freelists[nid])) {
507 page = list_entry(h->hugepage_freelists[nid].next,
Andrew Morton3abf7af2007-07-19 01:49:08 -0700508 struct page, lru);
509 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700510 h->free_huge_pages--;
511 h->free_huge_pages_node[nid]--;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700512
513 if (!avoid_reserve)
Andi Kleena5516432008-07-23 21:27:41 -0700514 decrement_hugepage_resv_vma(h, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700515
Ken Chen5ab3ee72007-07-23 18:44:00 -0700516 break;
Andrew Morton3abf7af2007-07-19 01:49:08 -0700517 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 }
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700519 mpol_cond_put(mpol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 return page;
521}
522
Andi Kleena5516432008-07-23 21:27:41 -0700523static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700524{
525 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700526
Andy Whitcroft18229df2008-11-06 12:53:27 -0800527 VM_BUG_ON(h->order >= MAX_ORDER);
528
Andi Kleena5516432008-07-23 21:27:41 -0700529 h->nr_huge_pages--;
530 h->nr_huge_pages_node[page_to_nid(page)]--;
531 for (i = 0; i < pages_per_huge_page(h); i++) {
Adam Litke6af2acb2007-10-16 01:26:16 -0700532 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
533 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
534 1 << PG_private | 1<< PG_writeback);
535 }
536 set_compound_page_dtor(page, NULL);
537 set_page_refcounted(page);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700538 arch_release_hugepage(page);
Andi Kleena5516432008-07-23 21:27:41 -0700539 __free_pages(page, huge_page_order(h));
Adam Litke6af2acb2007-10-16 01:26:16 -0700540}
541
Andi Kleene5ff2152008-07-23 21:27:42 -0700542struct hstate *size_to_hstate(unsigned long size)
543{
544 struct hstate *h;
545
546 for_each_hstate(h) {
547 if (huge_page_size(h) == size)
548 return h;
549 }
550 return NULL;
551}
552
David Gibson27a85ef2006-03-22 00:08:56 -0800553static void free_huge_page(struct page *page)
554{
Andi Kleena5516432008-07-23 21:27:41 -0700555 /*
556 * Can't pass hstate in here because it is called from the
557 * compound page destructor.
558 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700559 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700560 int nid = page_to_nid(page);
Adam Litkec79fb752007-11-14 16:59:38 -0800561 struct address_space *mapping;
David Gibson27a85ef2006-03-22 00:08:56 -0800562
Adam Litkec79fb752007-11-14 16:59:38 -0800563 mapping = (struct address_space *) page_private(page);
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800564 set_page_private(page, 0);
Adam Litke7893d1d2007-10-16 01:26:18 -0700565 BUG_ON(page_count(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800566 INIT_LIST_HEAD(&page->lru);
567
568 spin_lock(&hugetlb_lock);
Andi Kleenaa888a72008-07-23 21:27:47 -0700569 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
Andi Kleena5516432008-07-23 21:27:41 -0700570 update_and_free_page(h, page);
571 h->surplus_huge_pages--;
572 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700573 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700574 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700575 }
David Gibson27a85ef2006-03-22 00:08:56 -0800576 spin_unlock(&hugetlb_lock);
Adam Litkec79fb752007-11-14 16:59:38 -0800577 if (mapping)
Adam Litke9a119c02007-11-14 16:59:41 -0800578 hugetlb_put_quota(mapping, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800579}
580
Adam Litke7893d1d2007-10-16 01:26:18 -0700581/*
582 * Increment or decrement surplus_huge_pages. Keep node-specific counters
583 * balanced by operating on them in a round-robin fashion.
584 * Returns 1 if an adjustment was made.
585 */
Andi Kleena5516432008-07-23 21:27:41 -0700586static int adjust_pool_surplus(struct hstate *h, int delta)
Adam Litke7893d1d2007-10-16 01:26:18 -0700587{
588 static int prev_nid;
589 int nid = prev_nid;
590 int ret = 0;
591
592 VM_BUG_ON(delta != -1 && delta != 1);
593 do {
594 nid = next_node(nid, node_online_map);
595 if (nid == MAX_NUMNODES)
596 nid = first_node(node_online_map);
597
598 /* To shrink on this node, there must be a surplus page */
Andi Kleena5516432008-07-23 21:27:41 -0700599 if (delta < 0 && !h->surplus_huge_pages_node[nid])
Adam Litke7893d1d2007-10-16 01:26:18 -0700600 continue;
601 /* Surplus cannot exceed the total number of pages */
Andi Kleena5516432008-07-23 21:27:41 -0700602 if (delta > 0 && h->surplus_huge_pages_node[nid] >=
603 h->nr_huge_pages_node[nid])
Adam Litke7893d1d2007-10-16 01:26:18 -0700604 continue;
605
Andi Kleena5516432008-07-23 21:27:41 -0700606 h->surplus_huge_pages += delta;
607 h->surplus_huge_pages_node[nid] += delta;
Adam Litke7893d1d2007-10-16 01:26:18 -0700608 ret = 1;
609 break;
610 } while (nid != prev_nid);
611
612 prev_nid = nid;
613 return ret;
614}
615
Andi Kleena5516432008-07-23 21:27:41 -0700616static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700617{
618 set_compound_page_dtor(page, free_huge_page);
619 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700620 h->nr_huge_pages++;
621 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700622 spin_unlock(&hugetlb_lock);
623 put_page(page); /* free it into the hugepage allocator */
624}
625
Andi Kleena5516432008-07-23 21:27:41 -0700626static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700629
Andi Kleenaa888a72008-07-23 21:27:47 -0700630 if (h->order >= MAX_ORDER)
631 return NULL;
632
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700633 page = alloc_pages_node(nid,
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700634 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
635 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700636 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700638 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700639 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700640 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700641 }
Andi Kleena5516432008-07-23 21:27:41 -0700642 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700644
645 return page;
646}
647
Andi Kleen5ced66c2008-07-23 21:27:45 -0700648/*
649 * Use a helper variable to find the next node and then
650 * copy it back to hugetlb_next_nid afterwards:
651 * otherwise there's a window in which a racer might
652 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
653 * But we don't need to use a spin_lock here: it really
654 * doesn't matter if occasionally a racer chooses the
655 * same nid as we do. Move nid forward in the mask even
656 * if we just successfully allocated a hugepage so that
657 * the next caller gets hugepages on the next node.
658 */
659static int hstate_next_node(struct hstate *h)
660{
661 int next_nid;
662 next_nid = next_node(h->hugetlb_next_nid, node_online_map);
663 if (next_nid == MAX_NUMNODES)
664 next_nid = first_node(node_online_map);
665 h->hugetlb_next_nid = next_nid;
666 return next_nid;
667}
668
Andi Kleena5516432008-07-23 21:27:41 -0700669static int alloc_fresh_huge_page(struct hstate *h)
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700670{
671 struct page *page;
672 int start_nid;
673 int next_nid;
674 int ret = 0;
675
Andi Kleena5516432008-07-23 21:27:41 -0700676 start_nid = h->hugetlb_next_nid;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700677
678 do {
Andi Kleena5516432008-07-23 21:27:41 -0700679 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700680 if (page)
681 ret = 1;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700682 next_nid = hstate_next_node(h);
Andi Kleena5516432008-07-23 21:27:41 -0700683 } while (!page && h->hugetlb_next_nid != start_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700684
Adam Litke3b116302008-04-28 02:13:06 -0700685 if (ret)
686 count_vm_event(HTLB_BUDDY_PGALLOC);
687 else
688 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
689
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700690 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691}
692
Andi Kleena5516432008-07-23 21:27:41 -0700693static struct page *alloc_buddy_huge_page(struct hstate *h,
694 struct vm_area_struct *vma, unsigned long address)
Adam Litke7893d1d2007-10-16 01:26:18 -0700695{
696 struct page *page;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800697 unsigned int nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700698
Andi Kleenaa888a72008-07-23 21:27:47 -0700699 if (h->order >= MAX_ORDER)
700 return NULL;
701
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800702 /*
703 * Assume we will successfully allocate the surplus page to
704 * prevent racing processes from causing the surplus to exceed
705 * overcommit
706 *
707 * This however introduces a different race, where a process B
708 * tries to grow the static hugepage pool while alloc_pages() is
709 * called by process A. B will only examine the per-node
710 * counters in determining if surplus huge pages can be
711 * converted to normal huge pages in adjust_pool_surplus(). A
712 * won't be able to increment the per-node counter, until the
713 * lock is dropped by B, but B doesn't drop hugetlb_lock until
714 * no more huge pages can be converted from surplus to normal
715 * state (and doesn't try to convert again). Thus, we have a
716 * case where a surplus huge page exists, the pool is grown, and
717 * the surplus huge page still exists after, even though it
718 * should just have been converted to a normal huge page. This
719 * does not leak memory, though, as the hugepage will be freed
720 * once it is out of use. It also does not allow the counters to
721 * go out of whack in adjust_pool_surplus() as we don't modify
722 * the node values until we've gotten the hugepage and only the
723 * per-node value is checked there.
724 */
725 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700726 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800727 spin_unlock(&hugetlb_lock);
728 return NULL;
729 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700730 h->nr_huge_pages++;
731 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800732 }
733 spin_unlock(&hugetlb_lock);
734
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700735 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
736 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700737 huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800738
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700739 if (page && arch_prepare_hugepage(page)) {
740 __free_pages(page, huge_page_order(h));
741 return NULL;
742 }
743
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800744 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700745 if (page) {
Adam Litke2668db92008-03-10 11:43:50 -0700746 /*
747 * This page is now managed by the hugetlb allocator and has
748 * no users -- drop the buddy allocator's reference.
749 */
750 put_page_testzero(page);
751 VM_BUG_ON(page_count(page));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800752 nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700753 set_compound_page_dtor(page, free_huge_page);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800754 /*
755 * We incremented the global counters already
756 */
Andi Kleena5516432008-07-23 21:27:41 -0700757 h->nr_huge_pages_node[nid]++;
758 h->surplus_huge_pages_node[nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700759 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800760 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700761 h->nr_huge_pages--;
762 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700763 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700764 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800765 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700766
767 return page;
768}
769
Adam Litkee4e574b2007-10-16 01:26:19 -0700770/*
771 * Increase the hugetlb pool such that it can accomodate a reservation
772 * of size 'delta'.
773 */
Andi Kleena5516432008-07-23 21:27:41 -0700774static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -0700775{
776 struct list_head surplus_list;
777 struct page *page, *tmp;
778 int ret, i;
779 int needed, allocated;
780
Andi Kleena5516432008-07-23 21:27:41 -0700781 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800782 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -0700783 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700784 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800785 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700786
787 allocated = 0;
788 INIT_LIST_HEAD(&surplus_list);
789
790 ret = -ENOMEM;
791retry:
792 spin_unlock(&hugetlb_lock);
793 for (i = 0; i < needed; i++) {
Andi Kleena5516432008-07-23 21:27:41 -0700794 page = alloc_buddy_huge_page(h, NULL, 0);
Adam Litkee4e574b2007-10-16 01:26:19 -0700795 if (!page) {
796 /*
797 * We were not able to allocate enough pages to
798 * satisfy the entire reservation so we free what
799 * we've allocated so far.
800 */
801 spin_lock(&hugetlb_lock);
802 needed = 0;
803 goto free;
804 }
805
806 list_add(&page->lru, &surplus_list);
807 }
808 allocated += needed;
809
810 /*
811 * After retaking hugetlb_lock, we need to recalculate 'needed'
812 * because either resv_huge_pages or free_huge_pages may have changed.
813 */
814 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700815 needed = (h->resv_huge_pages + delta) -
816 (h->free_huge_pages + allocated);
Adam Litkee4e574b2007-10-16 01:26:19 -0700817 if (needed > 0)
818 goto retry;
819
820 /*
821 * The surplus_list now contains _at_least_ the number of extra pages
822 * needed to accomodate the reservation. Add the appropriate number
823 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -0800824 * allocator. Commit the entire reservation here to prevent another
825 * process from stealing the pages as they are added to the pool but
826 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -0700827 */
828 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -0700829 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700830 ret = 0;
831free:
Adam Litke19fc3f02008-04-28 02:12:20 -0700832 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -0700833 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -0700834 if ((--needed) < 0)
835 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700836 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700837 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -0700838 }
839
840 /* Free unnecessary surplus pages to the buddy allocator */
841 if (!list_empty(&surplus_list)) {
842 spin_unlock(&hugetlb_lock);
843 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
844 list_del(&page->lru);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700845 /*
Adam Litke2668db92008-03-10 11:43:50 -0700846 * The page has a reference count of zero already, so
847 * call free_huge_page directly instead of using
848 * put_page. This must be done with hugetlb_lock
Adam Litkeaf767cb2007-10-16 01:26:25 -0700849 * unlocked which is safe because free_huge_page takes
850 * hugetlb_lock before deciding how to free the page.
851 */
Adam Litke2668db92008-03-10 11:43:50 -0700852 free_huge_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700853 }
Adam Litke19fc3f02008-04-28 02:12:20 -0700854 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -0700855 }
856
857 return ret;
858}
859
860/*
861 * When releasing a hugetlb pool reservation, any surplus pages that were
862 * allocated to satisfy the reservation must be explicitly freed if they were
863 * never used.
864 */
Andi Kleena5516432008-07-23 21:27:41 -0700865static void return_unused_surplus_pages(struct hstate *h,
866 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -0700867{
868 static int nid = -1;
869 struct page *page;
870 unsigned long nr_pages;
871
Nishanth Aravamudan11320d12008-03-26 14:40:20 -0700872 /*
873 * We want to release as many surplus pages as possible, spread
874 * evenly across all nodes. Iterate across all nodes until we
875 * can no longer free unreserved surplus pages. This occurs when
876 * the nodes with surplus pages have no free pages.
877 */
878 unsigned long remaining_iterations = num_online_nodes();
879
Adam Litkeac09b3a2008-03-04 14:29:38 -0800880 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -0700881 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800882
Andi Kleenaa888a72008-07-23 21:27:47 -0700883 /* Cannot return gigantic pages currently */
884 if (h->order >= MAX_ORDER)
885 return;
886
Andi Kleena5516432008-07-23 21:27:41 -0700887 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -0700888
Nishanth Aravamudan11320d12008-03-26 14:40:20 -0700889 while (remaining_iterations-- && nr_pages) {
Adam Litkee4e574b2007-10-16 01:26:19 -0700890 nid = next_node(nid, node_online_map);
891 if (nid == MAX_NUMNODES)
892 nid = first_node(node_online_map);
893
Andi Kleena5516432008-07-23 21:27:41 -0700894 if (!h->surplus_huge_pages_node[nid])
Adam Litkee4e574b2007-10-16 01:26:19 -0700895 continue;
896
Andi Kleena5516432008-07-23 21:27:41 -0700897 if (!list_empty(&h->hugepage_freelists[nid])) {
898 page = list_entry(h->hugepage_freelists[nid].next,
Adam Litkee4e574b2007-10-16 01:26:19 -0700899 struct page, lru);
900 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700901 update_and_free_page(h, page);
902 h->free_huge_pages--;
903 h->free_huge_pages_node[nid]--;
904 h->surplus_huge_pages--;
905 h->surplus_huge_pages_node[nid]--;
Adam Litkee4e574b2007-10-16 01:26:19 -0700906 nr_pages--;
Nishanth Aravamudan11320d12008-03-26 14:40:20 -0700907 remaining_iterations = num_online_nodes();
Adam Litkee4e574b2007-10-16 01:26:19 -0700908 }
909 }
910}
911
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700912/*
913 * Determine if the huge page at addr within the vma has an associated
914 * reservation. Where it does not we will need to logically increase
915 * reservation and actually increase quota before an allocation can occur.
916 * Where any new reservation would be required the reservation change is
917 * prepared, but not committed. Once the page has been quota'd allocated
918 * an instantiated the change should be committed via vma_commit_reservation.
919 * No action is required on failure.
920 */
Roel Kluine2f17d92009-03-31 15:23:15 -0700921static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -0700922 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700923{
924 struct address_space *mapping = vma->vm_file->f_mapping;
925 struct inode *inode = mapping->host;
926
Mel Gormanf83a2752009-05-28 14:34:40 -0700927 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -0700928 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700929 return region_chg(&inode->i_mapping->private_list,
930 idx, idx + 1);
931
Andy Whitcroft84afd992008-07-23 21:27:32 -0700932 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
933 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700934
Andy Whitcroft84afd992008-07-23 21:27:32 -0700935 } else {
Roel Kluine2f17d92009-03-31 15:23:15 -0700936 long err;
Andi Kleena5516432008-07-23 21:27:41 -0700937 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700938 struct resv_map *reservations = vma_resv_map(vma);
939
940 err = region_chg(&reservations->regions, idx, idx + 1);
941 if (err < 0)
942 return err;
943 return 0;
944 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700945}
Andi Kleena5516432008-07-23 21:27:41 -0700946static void vma_commit_reservation(struct hstate *h,
947 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700948{
949 struct address_space *mapping = vma->vm_file->f_mapping;
950 struct inode *inode = mapping->host;
951
Mel Gormanf83a2752009-05-28 14:34:40 -0700952 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -0700953 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700954 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700955
956 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -0700957 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700958 struct resv_map *reservations = vma_resv_map(vma);
959
960 /* Mark this page used in the map. */
961 region_add(&reservations->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700962 }
963}
964
David Gibson27a85ef2006-03-22 00:08:56 -0800965static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700966 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967{
Andi Kleena5516432008-07-23 21:27:41 -0700968 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -0800969 struct page *page;
Adam Litke2fc39ce2007-11-14 16:59:39 -0800970 struct address_space *mapping = vma->vm_file->f_mapping;
Mel Gormana1e78772008-07-23 21:27:23 -0700971 struct inode *inode = mapping->host;
Roel Kluine2f17d92009-03-31 15:23:15 -0700972 long chg;
Adam Litke2fc39ce2007-11-14 16:59:39 -0800973
Mel Gormana1e78772008-07-23 21:27:23 -0700974 /*
975 * Processes that did not create the mapping will have no reserves and
976 * will not have accounted against quota. Check that the quota can be
977 * made before satisfying the allocation
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700978 * MAP_NORESERVE mappings may also need pages and quota allocated
979 * if no reserve mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -0700980 */
Andi Kleena5516432008-07-23 21:27:41 -0700981 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700982 if (chg < 0)
983 return ERR_PTR(chg);
984 if (chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700985 if (hugetlb_get_quota(inode->i_mapping, chg))
986 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -0700987
988 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700989 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
Mel Gormana1e78772008-07-23 21:27:23 -0700990 spin_unlock(&hugetlb_lock);
991
992 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -0700993 page = alloc_buddy_huge_page(h, vma, addr);
Mel Gormana1e78772008-07-23 21:27:23 -0700994 if (!page) {
995 hugetlb_put_quota(inode->i_mapping, chg);
996 return ERR_PTR(-VM_FAULT_OOM);
997 }
998 }
999
1000 set_page_refcounted(page);
1001 set_page_private(page, (unsigned long) mapping);
1002
Andi Kleena5516432008-07-23 21:27:41 -07001003 vma_commit_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001004
Adam Litke90d8b7e2007-11-14 16:59:42 -08001005 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001006}
1007
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001008int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001009{
1010 struct huge_bootmem_page *m;
1011 int nr_nodes = nodes_weight(node_online_map);
1012
1013 while (nr_nodes) {
1014 void *addr;
1015
1016 addr = __alloc_bootmem_node_nopanic(
1017 NODE_DATA(h->hugetlb_next_nid),
1018 huge_page_size(h), huge_page_size(h), 0);
1019
1020 if (addr) {
1021 /*
1022 * Use the beginning of the huge page to store the
1023 * huge_bootmem_page struct (until gather_bootmem
1024 * puts them into the mem_map).
1025 */
1026 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001027 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001028 }
1029 hstate_next_node(h);
1030 nr_nodes--;
1031 }
1032 return 0;
1033
1034found:
1035 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1036 /* Put them into a private list first because mem_map is not up yet */
1037 list_add(&m->list, &huge_boot_pages);
1038 m->hstate = h;
1039 return 1;
1040}
1041
Andy Whitcroft18229df2008-11-06 12:53:27 -08001042static void prep_compound_huge_page(struct page *page, int order)
1043{
1044 if (unlikely(order > (MAX_ORDER - 1)))
1045 prep_compound_gigantic_page(page, order);
1046 else
1047 prep_compound_page(page, order);
1048}
1049
Andi Kleenaa888a72008-07-23 21:27:47 -07001050/* Put bootmem huge pages into the standard lists after mem_map is up */
1051static void __init gather_bootmem_prealloc(void)
1052{
1053 struct huge_bootmem_page *m;
1054
1055 list_for_each_entry(m, &huge_boot_pages, list) {
1056 struct page *page = virt_to_page(m);
1057 struct hstate *h = m->hstate;
1058 __ClearPageReserved(page);
1059 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001060 prep_compound_huge_page(page, h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001061 prep_new_huge_page(h, page, page_to_nid(page));
1062 }
1063}
1064
Andi Kleen8faa8b02008-07-23 21:27:48 -07001065static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066{
1067 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068
Andi Kleene5ff2152008-07-23 21:27:42 -07001069 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001070 if (h->order >= MAX_ORDER) {
1071 if (!alloc_bootmem_huge_page(h))
1072 break;
1073 } else if (!alloc_fresh_huge_page(h))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001076 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001077}
1078
1079static void __init hugetlb_init_hstates(void)
1080{
1081 struct hstate *h;
1082
1083 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001084 /* oversize hugepages were init'ed in early boot */
1085 if (h->order < MAX_ORDER)
1086 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001087 }
1088}
1089
Andi Kleen4abd32d2008-07-23 21:27:49 -07001090static char * __init memfmt(char *buf, unsigned long n)
1091{
1092 if (n >= (1UL << 30))
1093 sprintf(buf, "%lu GB", n >> 30);
1094 else if (n >= (1UL << 20))
1095 sprintf(buf, "%lu MB", n >> 20);
1096 else
1097 sprintf(buf, "%lu KB", n >> 10);
1098 return buf;
1099}
1100
Andi Kleene5ff2152008-07-23 21:27:42 -07001101static void __init report_hugepages(void)
1102{
1103 struct hstate *h;
1104
1105 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001106 char buf[32];
1107 printk(KERN_INFO "HugeTLB registered %s page size, "
1108 "pre-allocated %ld pages\n",
1109 memfmt(buf, huge_page_size(h)),
1110 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001111 }
1112}
1113
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114#ifdef CONFIG_HIGHMEM
Andi Kleena5516432008-07-23 21:27:41 -07001115static void try_to_free_low(struct hstate *h, unsigned long count)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001117 int i;
1118
Andi Kleenaa888a72008-07-23 21:27:47 -07001119 if (h->order >= MAX_ORDER)
1120 return;
1121
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 for (i = 0; i < MAX_NUMNODES; ++i) {
1123 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001124 struct list_head *freel = &h->hugepage_freelists[i];
1125 list_for_each_entry_safe(page, next, freel, lru) {
1126 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001127 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 if (PageHighMem(page))
1129 continue;
1130 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001131 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001132 h->free_huge_pages--;
1133 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 }
1135 }
1136}
1137#else
Andi Kleena5516432008-07-23 21:27:41 -07001138static inline void try_to_free_low(struct hstate *h, unsigned long count)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139{
1140}
1141#endif
1142
Andi Kleena5516432008-07-23 21:27:41 -07001143#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Andi Kleene5ff2152008-07-23 21:27:42 -07001144static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145{
Adam Litke7893d1d2007-10-16 01:26:18 -07001146 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147
Andi Kleenaa888a72008-07-23 21:27:47 -07001148 if (h->order >= MAX_ORDER)
1149 return h->max_huge_pages;
1150
Adam Litke7893d1d2007-10-16 01:26:18 -07001151 /*
1152 * Increase the pool size
1153 * First take pages out of surplus state. Then make up the
1154 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001155 *
1156 * We might race with alloc_buddy_huge_page() here and be unable
1157 * to convert a surplus huge page to a normal huge page. That is
1158 * not critical, though, it just means the overall size of the
1159 * pool might be one hugepage larger than it needs to be, but
1160 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001161 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001163 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1164 if (!adjust_pool_surplus(h, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001165 break;
1166 }
1167
Andi Kleena5516432008-07-23 21:27:41 -07001168 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001169 /*
1170 * If this allocation races such that we no longer need the
1171 * page, free_huge_page will handle it by freeing the page
1172 * and reducing the surplus.
1173 */
1174 spin_unlock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001175 ret = alloc_fresh_huge_page(h);
Adam Litke7893d1d2007-10-16 01:26:18 -07001176 spin_lock(&hugetlb_lock);
1177 if (!ret)
1178 goto out;
1179
1180 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001181
1182 /*
1183 * Decrease the pool size
1184 * First return free pages to the buddy allocator (being careful
1185 * to keep enough around to satisfy reservations). Then place
1186 * pages into surplus state as needed so the pool will shrink
1187 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001188 *
1189 * By placing pages into the surplus state independent of the
1190 * overcommit value, we are allowing the surplus pool size to
1191 * exceed overcommit. There are few sane options here. Since
1192 * alloc_buddy_huge_page() is checking the global counter,
1193 * though, we'll note that we're not allowed to exceed surplus
1194 * and won't grow the pool anywhere else. Not until one of the
1195 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001196 */
Andi Kleena5516432008-07-23 21:27:41 -07001197 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001198 min_count = max(count, min_count);
Andi Kleena5516432008-07-23 21:27:41 -07001199 try_to_free_low(h, min_count);
1200 while (min_count < persistent_huge_pages(h)) {
1201 struct page *page = dequeue_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202 if (!page)
1203 break;
Andi Kleena5516432008-07-23 21:27:41 -07001204 update_and_free_page(h, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 }
Andi Kleena5516432008-07-23 21:27:41 -07001206 while (count < persistent_huge_pages(h)) {
1207 if (!adjust_pool_surplus(h, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001208 break;
1209 }
1210out:
Andi Kleena5516432008-07-23 21:27:41 -07001211 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001213 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214}
1215
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001216#define HSTATE_ATTR_RO(_name) \
1217 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1218
1219#define HSTATE_ATTR(_name) \
1220 static struct kobj_attribute _name##_attr = \
1221 __ATTR(_name, 0644, _name##_show, _name##_store)
1222
1223static struct kobject *hugepages_kobj;
1224static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1225
1226static struct hstate *kobj_to_hstate(struct kobject *kobj)
1227{
1228 int i;
1229 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1230 if (hstate_kobjs[i] == kobj)
1231 return &hstates[i];
1232 BUG();
1233 return NULL;
1234}
1235
1236static ssize_t nr_hugepages_show(struct kobject *kobj,
1237 struct kobj_attribute *attr, char *buf)
1238{
1239 struct hstate *h = kobj_to_hstate(kobj);
1240 return sprintf(buf, "%lu\n", h->nr_huge_pages);
1241}
1242static ssize_t nr_hugepages_store(struct kobject *kobj,
1243 struct kobj_attribute *attr, const char *buf, size_t count)
1244{
1245 int err;
1246 unsigned long input;
1247 struct hstate *h = kobj_to_hstate(kobj);
1248
1249 err = strict_strtoul(buf, 10, &input);
1250 if (err)
1251 return 0;
1252
1253 h->max_huge_pages = set_max_huge_pages(h, input);
1254
1255 return count;
1256}
1257HSTATE_ATTR(nr_hugepages);
1258
1259static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1260 struct kobj_attribute *attr, char *buf)
1261{
1262 struct hstate *h = kobj_to_hstate(kobj);
1263 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1264}
1265static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1266 struct kobj_attribute *attr, const char *buf, size_t count)
1267{
1268 int err;
1269 unsigned long input;
1270 struct hstate *h = kobj_to_hstate(kobj);
1271
1272 err = strict_strtoul(buf, 10, &input);
1273 if (err)
1274 return 0;
1275
1276 spin_lock(&hugetlb_lock);
1277 h->nr_overcommit_huge_pages = input;
1278 spin_unlock(&hugetlb_lock);
1279
1280 return count;
1281}
1282HSTATE_ATTR(nr_overcommit_hugepages);
1283
1284static ssize_t free_hugepages_show(struct kobject *kobj,
1285 struct kobj_attribute *attr, char *buf)
1286{
1287 struct hstate *h = kobj_to_hstate(kobj);
1288 return sprintf(buf, "%lu\n", h->free_huge_pages);
1289}
1290HSTATE_ATTR_RO(free_hugepages);
1291
1292static ssize_t resv_hugepages_show(struct kobject *kobj,
1293 struct kobj_attribute *attr, char *buf)
1294{
1295 struct hstate *h = kobj_to_hstate(kobj);
1296 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1297}
1298HSTATE_ATTR_RO(resv_hugepages);
1299
1300static ssize_t surplus_hugepages_show(struct kobject *kobj,
1301 struct kobj_attribute *attr, char *buf)
1302{
1303 struct hstate *h = kobj_to_hstate(kobj);
1304 return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1305}
1306HSTATE_ATTR_RO(surplus_hugepages);
1307
1308static struct attribute *hstate_attrs[] = {
1309 &nr_hugepages_attr.attr,
1310 &nr_overcommit_hugepages_attr.attr,
1311 &free_hugepages_attr.attr,
1312 &resv_hugepages_attr.attr,
1313 &surplus_hugepages_attr.attr,
1314 NULL,
1315};
1316
1317static struct attribute_group hstate_attr_group = {
1318 .attrs = hstate_attrs,
1319};
1320
1321static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1322{
1323 int retval;
1324
1325 hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1326 hugepages_kobj);
1327 if (!hstate_kobjs[h - hstates])
1328 return -ENOMEM;
1329
1330 retval = sysfs_create_group(hstate_kobjs[h - hstates],
1331 &hstate_attr_group);
1332 if (retval)
1333 kobject_put(hstate_kobjs[h - hstates]);
1334
1335 return retval;
1336}
1337
1338static void __init hugetlb_sysfs_init(void)
1339{
1340 struct hstate *h;
1341 int err;
1342
1343 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1344 if (!hugepages_kobj)
1345 return;
1346
1347 for_each_hstate(h) {
1348 err = hugetlb_sysfs_add_hstate(h);
1349 if (err)
1350 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1351 h->name);
1352 }
1353}
1354
1355static void __exit hugetlb_exit(void)
1356{
1357 struct hstate *h;
1358
1359 for_each_hstate(h) {
1360 kobject_put(hstate_kobjs[h - hstates]);
1361 }
1362
1363 kobject_put(hugepages_kobj);
1364}
1365module_exit(hugetlb_exit);
1366
1367static int __init hugetlb_init(void)
1368{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001369 /* Some platform decide whether they support huge pages at boot
1370 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1371 * there is no such support
1372 */
1373 if (HPAGE_SHIFT == 0)
1374 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001375
Nick Piggine11bfbf2008-07-23 21:27:52 -07001376 if (!size_to_hstate(default_hstate_size)) {
1377 default_hstate_size = HPAGE_SIZE;
1378 if (!size_to_hstate(default_hstate_size))
1379 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001380 }
Nick Piggine11bfbf2008-07-23 21:27:52 -07001381 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1382 if (default_hstate_max_huge_pages)
1383 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001384
1385 hugetlb_init_hstates();
1386
Andi Kleenaa888a72008-07-23 21:27:47 -07001387 gather_bootmem_prealloc();
1388
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001389 report_hugepages();
1390
1391 hugetlb_sysfs_init();
1392
1393 return 0;
1394}
1395module_init(hugetlb_init);
1396
1397/* Should be called on processing a hugepagesz=... option */
1398void __init hugetlb_add_hstate(unsigned order)
1399{
1400 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001401 unsigned long i;
1402
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001403 if (size_to_hstate(PAGE_SIZE << order)) {
1404 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1405 return;
1406 }
1407 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1408 BUG_ON(order == 0);
1409 h = &hstates[max_hstate++];
1410 h->order = order;
1411 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001412 h->nr_huge_pages = 0;
1413 h->free_huge_pages = 0;
1414 for (i = 0; i < MAX_NUMNODES; ++i)
1415 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1416 h->hugetlb_next_nid = first_node(node_online_map);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001417 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1418 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001419
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001420 parsed_hstate = h;
1421}
1422
Nick Piggine11bfbf2008-07-23 21:27:52 -07001423static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001424{
1425 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001426 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001427
1428 /*
1429 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1430 * so this hugepages= parameter goes to the "default hstate".
1431 */
1432 if (!max_hstate)
1433 mhp = &default_hstate_max_huge_pages;
1434 else
1435 mhp = &parsed_hstate->max_huge_pages;
1436
Andi Kleen8faa8b02008-07-23 21:27:48 -07001437 if (mhp == last_mhp) {
1438 printk(KERN_WARNING "hugepages= specified twice without "
1439 "interleaving hugepagesz=, ignoring\n");
1440 return 1;
1441 }
1442
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001443 if (sscanf(s, "%lu", mhp) <= 0)
1444 *mhp = 0;
1445
Andi Kleen8faa8b02008-07-23 21:27:48 -07001446 /*
1447 * Global state is always initialized later in hugetlb_init.
1448 * But we need to allocate >= MAX_ORDER hstates here early to still
1449 * use the bootmem allocator.
1450 */
1451 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1452 hugetlb_hstate_alloc_pages(parsed_hstate);
1453
1454 last_mhp = mhp;
1455
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001456 return 1;
1457}
Nick Piggine11bfbf2008-07-23 21:27:52 -07001458__setup("hugepages=", hugetlb_nrpages_setup);
1459
1460static int __init hugetlb_default_setup(char *s)
1461{
1462 default_hstate_size = memparse(s, &s);
1463 return 1;
1464}
1465__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001466
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07001467static unsigned int cpuset_mems_nr(unsigned int *array)
1468{
1469 int node;
1470 unsigned int nr = 0;
1471
1472 for_each_node_mask(node, cpuset_current_mems_allowed)
1473 nr += array[node];
1474
1475 return nr;
1476}
1477
1478#ifdef CONFIG_SYSCTL
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1480 struct file *file, void __user *buffer,
1481 size_t *length, loff_t *ppos)
1482{
Andi Kleene5ff2152008-07-23 21:27:42 -07001483 struct hstate *h = &default_hstate;
1484 unsigned long tmp;
1485
1486 if (!write)
1487 tmp = h->max_huge_pages;
1488
1489 table->data = &tmp;
1490 table->maxlen = sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001492
1493 if (write)
1494 h->max_huge_pages = set_max_huge_pages(h, tmp);
1495
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496 return 0;
1497}
Mel Gorman396faf02007-07-17 04:03:13 -07001498
1499int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1500 struct file *file, void __user *buffer,
1501 size_t *length, loff_t *ppos)
1502{
1503 proc_dointvec(table, write, file, buffer, length, ppos);
1504 if (hugepages_treat_as_movable)
1505 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1506 else
1507 htlb_alloc_mask = GFP_HIGHUSER;
1508 return 0;
1509}
1510
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001511int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1512 struct file *file, void __user *buffer,
1513 size_t *length, loff_t *ppos)
1514{
Andi Kleena5516432008-07-23 21:27:41 -07001515 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07001516 unsigned long tmp;
1517
1518 if (!write)
1519 tmp = h->nr_overcommit_huge_pages;
1520
1521 table->data = &tmp;
1522 table->maxlen = sizeof(unsigned long);
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001523 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001524
1525 if (write) {
1526 spin_lock(&hugetlb_lock);
1527 h->nr_overcommit_huge_pages = tmp;
1528 spin_unlock(&hugetlb_lock);
1529 }
1530
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001531 return 0;
1532}
1533
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534#endif /* CONFIG_SYSCTL */
1535
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001536void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537{
Andi Kleena5516432008-07-23 21:27:41 -07001538 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001539 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001540 "HugePages_Total: %5lu\n"
1541 "HugePages_Free: %5lu\n"
1542 "HugePages_Rsvd: %5lu\n"
1543 "HugePages_Surp: %5lu\n"
1544 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07001545 h->nr_huge_pages,
1546 h->free_huge_pages,
1547 h->resv_huge_pages,
1548 h->surplus_huge_pages,
1549 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550}
1551
1552int hugetlb_report_node_meminfo(int nid, char *buf)
1553{
Andi Kleena5516432008-07-23 21:27:41 -07001554 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555 return sprintf(buf,
1556 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07001557 "Node %d HugePages_Free: %5u\n"
1558 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07001559 nid, h->nr_huge_pages_node[nid],
1560 nid, h->free_huge_pages_node[nid],
1561 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562}
1563
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1565unsigned long hugetlb_total_pages(void)
1566{
Andi Kleena5516432008-07-23 21:27:41 -07001567 struct hstate *h = &default_hstate;
1568 return h->nr_huge_pages * pages_per_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570
Andi Kleena5516432008-07-23 21:27:41 -07001571static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001572{
1573 int ret = -ENOMEM;
1574
1575 spin_lock(&hugetlb_lock);
1576 /*
1577 * When cpuset is configured, it breaks the strict hugetlb page
1578 * reservation as the accounting is done on a global variable. Such
1579 * reservation is completely rubbish in the presence of cpuset because
1580 * the reservation is not checked against page availability for the
1581 * current cpuset. Application can still potentially OOM'ed by kernel
1582 * with lack of free htlb page in cpuset that the task is in.
1583 * Attempt to enforce strict accounting with cpuset is almost
1584 * impossible (or too ugly) because cpuset is too fluid that
1585 * task or memory node can be dynamically moved between cpusets.
1586 *
1587 * The change of semantics for shared hugetlb mapping with cpuset is
1588 * undesirable. However, in order to preserve some of the semantics,
1589 * we fall back to check against current free page availability as
1590 * a best attempt and hopefully to minimize the impact of changing
1591 * semantics that cpuset has.
1592 */
1593 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001594 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001595 goto out;
1596
Andi Kleena5516432008-07-23 21:27:41 -07001597 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1598 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001599 goto out;
1600 }
1601 }
1602
1603 ret = 0;
1604 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07001605 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001606
1607out:
1608 spin_unlock(&hugetlb_lock);
1609 return ret;
1610}
1611
Andy Whitcroft84afd992008-07-23 21:27:32 -07001612static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1613{
1614 struct resv_map *reservations = vma_resv_map(vma);
1615
1616 /*
1617 * This new VMA should share its siblings reservation map if present.
1618 * The VMA will only ever have a valid reservation map pointer where
1619 * it is being copied for another still existing VMA. As that VMA
1620 * has a reference to the reservation map it cannot dissappear until
1621 * after this open call completes. It is therefore safe to take a
1622 * new reference here without additional locking.
1623 */
1624 if (reservations)
1625 kref_get(&reservations->refs);
1626}
1627
Mel Gormana1e78772008-07-23 21:27:23 -07001628static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1629{
Andi Kleena5516432008-07-23 21:27:41 -07001630 struct hstate *h = hstate_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001631 struct resv_map *reservations = vma_resv_map(vma);
1632 unsigned long reserve;
1633 unsigned long start;
1634 unsigned long end;
1635
1636 if (reservations) {
Andi Kleena5516432008-07-23 21:27:41 -07001637 start = vma_hugecache_offset(h, vma, vma->vm_start);
1638 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001639
1640 reserve = (end - start) -
1641 region_count(&reservations->regions, start, end);
1642
1643 kref_put(&reservations->refs, resv_map_release);
1644
Adam Litke7251ff782008-07-23 21:27:59 -07001645 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07001646 hugetlb_acct_memory(h, -reserve);
Adam Litke7251ff782008-07-23 21:27:59 -07001647 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1648 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07001649 }
Mel Gormana1e78772008-07-23 21:27:23 -07001650}
1651
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652/*
1653 * We cannot handle pagefaults against hugetlb pages at all. They cause
1654 * handle_mm_fault() to try to instantiate regular-sized pages in the
1655 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1656 * this far.
1657 */
Nick Piggind0217ac2007-07-19 01:47:03 -07001658static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659{
1660 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07001661 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662}
1663
1664struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07001665 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07001666 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07001667 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668};
1669
David Gibson1e8f8892006-01-06 00:10:44 -08001670static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1671 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07001672{
1673 pte_t entry;
1674
David Gibson1e8f8892006-01-06 00:10:44 -08001675 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -07001676 entry =
1677 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1678 } else {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001679 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07001680 }
1681 entry = pte_mkyoung(entry);
1682 entry = pte_mkhuge(entry);
1683
1684 return entry;
1685}
1686
David Gibson1e8f8892006-01-06 00:10:44 -08001687static void set_huge_ptep_writable(struct vm_area_struct *vma,
1688 unsigned long address, pte_t *ptep)
1689{
1690 pte_t entry;
1691
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001692 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1693 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07001694 update_mmu_cache(vma, address, entry);
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07001695 }
David Gibson1e8f8892006-01-06 00:10:44 -08001696}
1697
1698
David Gibson63551ae2005-06-21 17:14:44 -07001699int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1700 struct vm_area_struct *vma)
1701{
1702 pte_t *src_pte, *dst_pte, entry;
1703 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07001704 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08001705 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07001706 struct hstate *h = hstate_vma(vma);
1707 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08001708
1709 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07001710
Andi Kleena5516432008-07-23 21:27:41 -07001711 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Hugh Dickinsc74df322005-10-29 18:16:23 -07001712 src_pte = huge_pte_offset(src, addr);
1713 if (!src_pte)
1714 continue;
Andi Kleena5516432008-07-23 21:27:41 -07001715 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07001716 if (!dst_pte)
1717 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08001718
1719 /* If the pagetables are shared don't copy or take references */
1720 if (dst_pte == src_pte)
1721 continue;
1722
Hugh Dickinsc74df322005-10-29 18:16:23 -07001723 spin_lock(&dst->page_table_lock);
Nick Piggin46478752008-06-05 22:45:57 -07001724 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001725 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08001726 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001727 huge_ptep_set_wrprotect(src, addr, src_pte);
1728 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07001729 ptepage = pte_page(entry);
1730 get_page(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07001731 set_huge_pte_at(dst, addr, dst_pte, entry);
1732 }
1733 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -07001734 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07001735 }
1736 return 0;
1737
1738nomem:
1739 return -ENOMEM;
1740}
1741
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001742void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001743 unsigned long end, struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07001744{
1745 struct mm_struct *mm = vma->vm_mm;
1746 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07001747 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07001748 pte_t pte;
1749 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001750 struct page *tmp;
Andi Kleena5516432008-07-23 21:27:41 -07001751 struct hstate *h = hstate_vma(vma);
1752 unsigned long sz = huge_page_size(h);
1753
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -08001754 /*
1755 * A page gathering list, protected by per file i_mmap_lock. The
1756 * lock is used to avoid list corruption from multiple unmapping
1757 * of the same page since we are using page->lru.
1758 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001759 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -07001760
1761 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07001762 BUG_ON(start & ~huge_page_mask(h));
1763 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07001764
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001765 mmu_notifier_invalidate_range_start(mm, start, end);
Hugh Dickins508034a2005-10-29 18:16:30 -07001766 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001767 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07001768 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07001769 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07001770 continue;
1771
Chen, Kenneth W39dde652006-12-06 20:32:03 -08001772 if (huge_pmd_unshare(mm, &address, ptep))
1773 continue;
1774
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001775 /*
1776 * If a reference page is supplied, it is because a specific
1777 * page is being unmapped, not a range. Ensure the page we
1778 * are about to unmap is the actual page of interest.
1779 */
1780 if (ref_page) {
1781 pte = huge_ptep_get(ptep);
1782 if (huge_pte_none(pte))
1783 continue;
1784 page = pte_page(pte);
1785 if (page != ref_page)
1786 continue;
1787
1788 /*
1789 * Mark the VMA as having unmapped its page so that
1790 * future faults in this VMA will fail rather than
1791 * looking like data was lost
1792 */
1793 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1794 }
1795
David Gibsonc7546f82005-08-05 11:59:35 -07001796 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001797 if (huge_pte_none(pte))
David Gibson63551ae2005-06-21 17:14:44 -07001798 continue;
David Gibsonc7546f82005-08-05 11:59:35 -07001799
David Gibson63551ae2005-06-21 17:14:44 -07001800 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -08001801 if (pte_dirty(pte))
1802 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001803 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -07001804 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -07001806 flush_tlb_range(vma, start, end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001807 mmu_notifier_invalidate_range_end(mm, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001808 list_for_each_entry_safe(page, tmp, &page_list, lru) {
1809 list_del(&page->lru);
1810 put_page(page);
1811 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812}
David Gibson63551ae2005-06-21 17:14:44 -07001813
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001814void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001815 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001816{
Andi Kleena137e1c2008-07-23 21:27:43 -07001817 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1818 __unmap_hugepage_range(vma, start, end, ref_page);
1819 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001820}
1821
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001822/*
1823 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1824 * mappping it owns the reserve page for. The intention is to unmap the page
1825 * from other VMAs and let the children be SIGKILLed if they are faulting the
1826 * same region.
1827 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07001828static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
1829 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001830{
Adam Litke75266742008-11-12 13:24:56 -08001831 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001832 struct vm_area_struct *iter_vma;
1833 struct address_space *mapping;
1834 struct prio_tree_iter iter;
1835 pgoff_t pgoff;
1836
1837 /*
1838 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1839 * from page cache lookup which is in HPAGE_SIZE units.
1840 */
Adam Litke75266742008-11-12 13:24:56 -08001841 address = address & huge_page_mask(h);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001842 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1843 + (vma->vm_pgoff >> PAGE_SHIFT);
1844 mapping = (struct address_space *)page_private(page);
1845
1846 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1847 /* Do not unmap the current VMA */
1848 if (iter_vma == vma)
1849 continue;
1850
1851 /*
1852 * Unmap the page from other VMAs without their own reserves.
1853 * They get marked to be SIGKILLed if they fault in these
1854 * areas. This is because a future no-page fault on this VMA
1855 * could insert a zeroed page instead of the data existing
1856 * from the time of fork. This would look like data corruption
1857 */
1858 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1859 unmap_hugepage_range(iter_vma,
Adam Litke75266742008-11-12 13:24:56 -08001860 address, address + huge_page_size(h),
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001861 page);
1862 }
1863
1864 return 1;
1865}
1866
David Gibson1e8f8892006-01-06 00:10:44 -08001867static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001868 unsigned long address, pte_t *ptep, pte_t pte,
1869 struct page *pagecache_page)
David Gibson1e8f8892006-01-06 00:10:44 -08001870{
Andi Kleena5516432008-07-23 21:27:41 -07001871 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08001872 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -08001873 int avoidcopy;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001874 int outside_reserve = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08001875
1876 old_page = pte_page(pte);
1877
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001878retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08001879 /* If no-one else is actually using this page, avoid the copy
1880 * and just make the page writable */
1881 avoidcopy = (page_count(old_page) == 1);
1882 if (avoidcopy) {
1883 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07001884 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08001885 }
1886
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001887 /*
1888 * If the process that created a MAP_PRIVATE mapping is about to
1889 * perform a COW due to a shared page count, attempt to satisfy
1890 * the allocation without using the existing reserves. The pagecache
1891 * page is used to determine if the reserve at this address was
1892 * consumed or not. If reserves were used, a partial faulted mapping
1893 * at the time of fork() could consume its reserves on COW instead
1894 * of the full address range.
1895 */
Mel Gormanf83a2752009-05-28 14:34:40 -07001896 if (!(vma->vm_flags & VM_MAYSHARE) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001897 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1898 old_page != pagecache_page)
1899 outside_reserve = 1;
1900
David Gibson1e8f8892006-01-06 00:10:44 -08001901 page_cache_get(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001902 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08001903
Adam Litke2fc39ce2007-11-14 16:59:39 -08001904 if (IS_ERR(new_page)) {
David Gibson1e8f8892006-01-06 00:10:44 -08001905 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001906
1907 /*
1908 * If a process owning a MAP_PRIVATE mapping fails to COW,
1909 * it is due to references held by a child and an insufficient
1910 * huge page pool. To guarantee the original mappers
1911 * reliability, unmap the page from child processes. The child
1912 * may get SIGKILLed if it later faults.
1913 */
1914 if (outside_reserve) {
1915 BUG_ON(huge_pte_none(pte));
1916 if (unmap_ref_private(mm, vma, old_page, address)) {
1917 BUG_ON(page_count(old_page) != 1);
1918 BUG_ON(huge_pte_none(pte));
1919 goto retry_avoidcopy;
1920 }
1921 WARN_ON_ONCE(1);
1922 }
1923
Adam Litke2fc39ce2007-11-14 16:59:39 -08001924 return -PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08001925 }
1926
1927 spin_unlock(&mm->page_table_lock);
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00001928 copy_huge_page(new_page, old_page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08001929 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08001930 spin_lock(&mm->page_table_lock);
1931
Andi Kleena5516432008-07-23 21:27:41 -07001932 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001933 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08001934 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07001935 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08001936 set_huge_pte_at(mm, address, ptep,
1937 make_huge_pte(vma, new_page, 1));
1938 /* Make the old page be freed below */
1939 new_page = old_page;
1940 }
1941 page_cache_release(new_page);
1942 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -07001943 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08001944}
1945
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001946/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07001947static struct page *hugetlbfs_pagecache_page(struct hstate *h,
1948 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001949{
1950 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07001951 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001952
1953 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07001954 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001955
1956 return find_lock_page(mapping, idx);
1957}
1958
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07001959static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
David Gibson1e8f8892006-01-06 00:10:44 -08001960 unsigned long address, pte_t *ptep, int write_access)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01001961{
Andi Kleena5516432008-07-23 21:27:41 -07001962 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01001963 int ret = VM_FAULT_SIGBUS;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07001964 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07001965 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07001966 struct page *page;
1967 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08001968 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -07001969
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001970 /*
1971 * Currently, we are forced to kill the process in the event the
1972 * original mapper has unmapped pages from the child due to a failed
1973 * COW. Warn that such a situation has occured as it may not be obvious
1974 */
1975 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1976 printk(KERN_WARNING
1977 "PID %d killed due to inadequate hugepage pool\n",
1978 current->pid);
1979 return ret;
1980 }
1981
Adam Litke4c887262005-10-29 18:16:46 -07001982 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07001983 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07001984
1985 /*
1986 * Use page lock to guard against racing truncation
1987 * before we get page_table_lock.
1988 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08001989retry:
1990 page = find_lock_page(mapping, idx);
1991 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07001992 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07001993 if (idx >= size)
1994 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001995 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08001996 if (IS_ERR(page)) {
1997 ret = -PTR_ERR(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08001998 goto out;
1999 }
Andi Kleena5516432008-07-23 21:27:41 -07002000 clear_huge_page(page, address, huge_page_size(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002001 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002002
Mel Gormanf83a2752009-05-28 14:34:40 -07002003 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002004 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002005 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002006
2007 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2008 if (err) {
2009 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002010 if (err == -EEXIST)
2011 goto retry;
2012 goto out;
2013 }
Ken Chen45c682a2007-11-14 16:59:44 -08002014
2015 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002016 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002017 spin_unlock(&inode->i_lock);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002018 } else
2019 lock_page(page);
2020 }
David Gibson1e8f8892006-01-06 00:10:44 -08002021
Andy Whitcroft57303d82008-08-12 15:08:47 -07002022 /*
2023 * If we are going to COW a private mapping later, we examine the
2024 * pending reservations for this page now. This will ensure that
2025 * any allocations necessary to record that reservation occur outside
2026 * the spinlock.
2027 */
2028 if (write_access && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002029 if (vma_needs_reservation(h, vma, address) < 0) {
2030 ret = VM_FAULT_OOM;
2031 goto backout_unlocked;
2032 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002033
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002034 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002035 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002036 if (idx >= size)
2037 goto backout;
2038
Nick Piggin83c54072007-07-19 01:47:05 -07002039 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002040 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002041 goto backout;
2042
David Gibson1e8f8892006-01-06 00:10:44 -08002043 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2044 && (vma->vm_flags & VM_SHARED)));
2045 set_huge_pte_at(mm, address, ptep, new_pte);
2046
2047 if (write_access && !(vma->vm_flags & VM_SHARED)) {
2048 /* Optimization, do the COW without a second fault */
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002049 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
David Gibson1e8f8892006-01-06 00:10:44 -08002050 }
2051
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002052 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07002053 unlock_page(page);
2054out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002055 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002056
2057backout:
2058 spin_unlock(&mm->page_table_lock);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002059backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002060 unlock_page(page);
2061 put_page(page);
2062 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002063}
2064
Adam Litke86e52162006-01-06 00:10:43 -08002065int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2066 unsigned long address, int write_access)
2067{
2068 pte_t *ptep;
2069 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -08002070 int ret;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002071 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002072 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002073 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002074
Andi Kleena5516432008-07-23 21:27:41 -07002075 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002076 if (!ptep)
2077 return VM_FAULT_OOM;
2078
David Gibson3935baa2006-03-22 00:08:53 -08002079 /*
2080 * Serialize hugepage allocation and instantiation, so that we don't
2081 * get spurious allocation failures if two CPUs race to instantiate
2082 * the same page in the page cache.
2083 */
2084 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002085 entry = huge_ptep_get(ptep);
2086 if (huge_pte_none(entry)) {
David Gibson3935baa2006-03-22 00:08:53 -08002087 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
David Gibsonb4d1d992008-10-15 22:01:11 -07002088 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002089 }
Adam Litke86e52162006-01-06 00:10:43 -08002090
Nick Piggin83c54072007-07-19 01:47:05 -07002091 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002092
Andy Whitcroft57303d82008-08-12 15:08:47 -07002093 /*
2094 * If we are going to COW the mapping later, we examine the pending
2095 * reservations for this page now. This will ensure that any
2096 * allocations necessary to record that reservation occur outside the
2097 * spinlock. For private mappings, we also lookup the pagecache
2098 * page now as it is used to determine if a reservation has been
2099 * consumed.
2100 */
2101 if (write_access && !pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002102 if (vma_needs_reservation(h, vma, address) < 0) {
2103 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002104 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002105 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002106
Mel Gormanf83a2752009-05-28 14:34:40 -07002107 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07002108 pagecache_page = hugetlbfs_pagecache_page(h,
2109 vma, address);
2110 }
2111
David Gibson1e8f8892006-01-06 00:10:44 -08002112 spin_lock(&mm->page_table_lock);
2113 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002114 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2115 goto out_page_table_lock;
2116
2117
2118 if (write_access) {
2119 if (!pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002120 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2121 pagecache_page);
David Gibsonb4d1d992008-10-15 22:01:11 -07002122 goto out_page_table_lock;
2123 }
2124 entry = pte_mkdirty(entry);
2125 }
2126 entry = pte_mkyoung(entry);
2127 if (huge_ptep_set_access_flags(vma, address, ptep, entry, write_access))
2128 update_mmu_cache(vma, address, entry);
2129
2130out_page_table_lock:
David Gibson1e8f8892006-01-06 00:10:44 -08002131 spin_unlock(&mm->page_table_lock);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002132
2133 if (pagecache_page) {
2134 unlock_page(pagecache_page);
2135 put_page(pagecache_page);
2136 }
2137
David Gibsonb4d1d992008-10-15 22:01:11 -07002138out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08002139 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08002140
2141 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08002142}
2143
Andi Kleenceb86872008-07-23 21:27:50 -07002144/* Can be overriden by architectures */
2145__attribute__((weak)) struct page *
2146follow_huge_pud(struct mm_struct *mm, unsigned long address,
2147 pud_t *pud, int write)
2148{
2149 BUG();
2150 return NULL;
2151}
2152
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002153static int huge_zeropage_ok(pte_t *ptep, int write, int shared)
2154{
2155 if (!ptep || write || shared)
2156 return 0;
2157 else
2158 return huge_pte_none(huge_ptep_get(ptep));
2159}
2160
David Gibson63551ae2005-06-21 17:14:44 -07002161int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2162 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -08002163 unsigned long *position, int *length, int i,
2164 int write)
David Gibson63551ae2005-06-21 17:14:44 -07002165{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002166 unsigned long pfn_offset;
2167 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -07002168 int remainder = *length;
Andi Kleena5516432008-07-23 21:27:41 -07002169 struct hstate *h = hstate_vma(vma);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002170 int zeropage_ok = 0;
2171 int shared = vma->vm_flags & VM_SHARED;
David Gibson63551ae2005-06-21 17:14:44 -07002172
Hugh Dickins1c598272005-10-19 21:23:43 -07002173 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002174 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07002175 pte_t *pte;
2176 struct page *page;
2177
2178 /*
2179 * Some archs (sparc64, sh*) have multiple pte_ts to
2180 * each hugepage. We have to make * sure we get the
2181 * first, for the page indexing below to work.
2182 */
Andi Kleena5516432008-07-23 21:27:41 -07002183 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002184 if (huge_zeropage_ok(pte, write, shared))
2185 zeropage_ok = 1;
Adam Litke4c887262005-10-29 18:16:46 -07002186
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002187 if (!pte ||
2188 (huge_pte_none(huge_ptep_get(pte)) && !zeropage_ok) ||
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002189 (write && !pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07002190 int ret;
2191
2192 spin_unlock(&mm->page_table_lock);
Adam Litke5b23dbe2007-11-14 16:59:33 -08002193 ret = hugetlb_fault(mm, vma, vaddr, write);
Adam Litke4c887262005-10-29 18:16:46 -07002194 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07002195 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07002196 continue;
2197
2198 remainder = 0;
2199 if (!i)
2200 i = -EFAULT;
2201 break;
2202 }
David Gibson63551ae2005-06-21 17:14:44 -07002203
Andi Kleena5516432008-07-23 21:27:41 -07002204 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002205 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002206same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002207 if (pages) {
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002208 if (zeropage_ok)
2209 pages[i] = ZERO_PAGE(0);
2210 else
Andy Whitcroft69d177c2008-11-06 12:53:26 -08002211 pages[i] = mem_map_offset(page, pfn_offset);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002212 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002213 }
David Gibson63551ae2005-06-21 17:14:44 -07002214
2215 if (vmas)
2216 vmas[i] = vma;
2217
2218 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002219 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07002220 --remainder;
2221 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002222 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07002223 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002224 /*
2225 * We use pfn_offset to avoid touching the pageframes
2226 * of this compound page.
2227 */
2228 goto same_page;
2229 }
David Gibson63551ae2005-06-21 17:14:44 -07002230 }
Hugh Dickins1c598272005-10-19 21:23:43 -07002231 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002232 *length = remainder;
2233 *position = vaddr;
2234
2235 return i;
2236}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002237
2238void hugetlb_change_protection(struct vm_area_struct *vma,
2239 unsigned long address, unsigned long end, pgprot_t newprot)
2240{
2241 struct mm_struct *mm = vma->vm_mm;
2242 unsigned long start = address;
2243 pte_t *ptep;
2244 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07002245 struct hstate *h = hstate_vma(vma);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002246
2247 BUG_ON(address >= end);
2248 flush_cache_range(vma, address, end);
2249
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002250 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002251 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002252 for (; address < end; address += huge_page_size(h)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002253 ptep = huge_pte_offset(mm, address);
2254 if (!ptep)
2255 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002256 if (huge_pmd_unshare(mm, &address, ptep))
2257 continue;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002258 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002259 pte = huge_ptep_get_and_clear(mm, address, ptep);
2260 pte = pte_mkhuge(pte_modify(pte, newprot));
2261 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002262 }
2263 }
2264 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002265 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002266
2267 flush_tlb_range(vma, start, end);
2268}
2269
Mel Gormana1e78772008-07-23 21:27:23 -07002270int hugetlb_reserve_pages(struct inode *inode,
2271 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00002272 struct vm_area_struct *vma,
2273 int acctflag)
Adam Litkee4e574b2007-10-16 01:26:19 -07002274{
Mel Gorman17c9d122009-02-11 16:34:16 +00002275 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07002276 struct hstate *h = hstate_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07002277
Mel Gormana1e78772008-07-23 21:27:23 -07002278 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00002279 * Only apply hugepage reservation if asked. At fault time, an
2280 * attempt will be made for VM_NORESERVE to allocate a page
2281 * and filesystem quota without using reserves
2282 */
2283 if (acctflag & VM_NORESERVE)
2284 return 0;
2285
2286 /*
Mel Gormana1e78772008-07-23 21:27:23 -07002287 * Shared mappings base their reservation on the number of pages that
2288 * are already allocated on behalf of the file. Private mappings need
2289 * to reserve the full area even if read-only as mprotect() may be
2290 * called to make the mapping read-write. Assume !vma is a shm mapping
2291 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002292 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gormana1e78772008-07-23 21:27:23 -07002293 chg = region_chg(&inode->i_mapping->private_list, from, to);
Mel Gorman5a6fe122009-02-10 14:02:27 +00002294 else {
2295 struct resv_map *resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00002296 if (!resv_map)
2297 return -ENOMEM;
2298
Mel Gorman17c9d122009-02-11 16:34:16 +00002299 chg = to - from;
2300
Mel Gorman5a6fe122009-02-10 14:02:27 +00002301 set_vma_resv_map(vma, resv_map);
2302 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2303 }
2304
Mel Gorman17c9d122009-02-11 16:34:16 +00002305 if (chg < 0)
2306 return chg;
2307
2308 /* There must be enough filesystem quota for the mapping */
2309 if (hugetlb_get_quota(inode->i_mapping, chg))
2310 return -ENOSPC;
2311
2312 /*
2313 * Check enough hugepages are available for the reservation.
2314 * Hand back the quota if there are not
2315 */
2316 ret = hugetlb_acct_memory(h, chg);
2317 if (ret < 0) {
2318 hugetlb_put_quota(inode->i_mapping, chg);
2319 return ret;
2320 }
2321
2322 /*
2323 * Account for the reservations made. Shared mappings record regions
2324 * that have reservations as they are shared by multiple VMAs.
2325 * When the last VMA disappears, the region map says how much
2326 * the reservation was and the page cache tells how much of
2327 * the reservation was consumed. Private mappings are per-VMA and
2328 * only the consumed reservations are tracked. When the VMA
2329 * disappears, the original reservation is the VMA size and the
2330 * consumed reservations are stored in the map. Hence, nothing
2331 * else has to be done for private mappings here
2332 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002333 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gorman17c9d122009-02-11 16:34:16 +00002334 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002335 return 0;
2336}
2337
2338void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2339{
Andi Kleena5516432008-07-23 21:27:41 -07002340 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002341 long chg = region_truncate(&inode->i_mapping->private_list, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08002342
2343 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002344 inode->i_blocks -= blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002345 spin_unlock(&inode->i_lock);
2346
Adam Litke90d8b7e2007-11-14 16:59:42 -08002347 hugetlb_put_quota(inode->i_mapping, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07002348 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002349}