Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Generic hugetlb support. | |
3 | * (C) William Irwin, April 2004 | |
4 | */ | |
1da177e4 LT |
5 | #include <linux/list.h> |
6 | #include <linux/init.h> | |
7 | #include <linux/module.h> | |
8 | #include <linux/mm.h> | |
e1759c21 | 9 | #include <linux/seq_file.h> |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
cddb8a5c | 12 | #include <linux/mmu_notifier.h> |
1da177e4 | 13 | #include <linux/nodemask.h> |
63551ae0 | 14 | #include <linux/pagemap.h> |
5da7ca86 | 15 | #include <linux/mempolicy.h> |
aea47ff3 | 16 | #include <linux/cpuset.h> |
3935baa9 | 17 | #include <linux/mutex.h> |
aa888a74 | 18 | #include <linux/bootmem.h> |
a3437870 | 19 | #include <linux/sysfs.h> |
5a0e3ad6 | 20 | #include <linux/slab.h> |
0fe6e20b | 21 | #include <linux/rmap.h> |
fd6a03ed NH |
22 | #include <linux/swap.h> |
23 | #include <linux/swapops.h> | |
d6606683 | 24 | |
63551ae0 DG |
25 | #include <asm/page.h> |
26 | #include <asm/pgtable.h> | |
78a34ae2 | 27 | #include <asm/io.h> |
63551ae0 DG |
28 | |
29 | #include <linux/hugetlb.h> | |
9a305230 | 30 | #include <linux/node.h> |
7835e98b | 31 | #include "internal.h" |
1da177e4 LT |
32 | |
33 | const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; | |
396faf03 MG |
34 | static gfp_t htlb_alloc_mask = GFP_HIGHUSER; |
35 | unsigned long hugepages_treat_as_movable; | |
a5516438 | 36 | |
e5ff2159 AK |
37 | static int max_hstate; |
38 | unsigned int default_hstate_idx; | |
39 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
40 | ||
53ba51d2 JT |
41 | __initdata LIST_HEAD(huge_boot_pages); |
42 | ||
e5ff2159 AK |
43 | /* for command line parsing */ |
44 | static struct hstate * __initdata parsed_hstate; | |
45 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 46 | static unsigned long __initdata default_hstate_size; |
e5ff2159 AK |
47 | |
48 | #define for_each_hstate(h) \ | |
49 | for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++) | |
396faf03 | 50 | |
3935baa9 DG |
51 | /* |
52 | * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages | |
53 | */ | |
54 | static DEFINE_SPINLOCK(hugetlb_lock); | |
0bd0f9fb | 55 | |
96822904 AW |
56 | /* |
57 | * Region tracking -- allows tracking of reservations and instantiated pages | |
58 | * across the pages in a mapping. | |
84afd99b AW |
59 | * |
60 | * The region data structures are protected by a combination of the mmap_sem | |
61 | * and the hugetlb_instantion_mutex. To access or modify a region the caller | |
62 | * must either hold the mmap_sem for write, or the mmap_sem for read and | |
63 | * the hugetlb_instantiation mutex: | |
64 | * | |
65 | * down_write(&mm->mmap_sem); | |
66 | * or | |
67 | * down_read(&mm->mmap_sem); | |
68 | * mutex_lock(&hugetlb_instantiation_mutex); | |
96822904 AW |
69 | */ |
70 | struct file_region { | |
71 | struct list_head link; | |
72 | long from; | |
73 | long to; | |
74 | }; | |
75 | ||
76 | static long region_add(struct list_head *head, long f, long t) | |
77 | { | |
78 | struct file_region *rg, *nrg, *trg; | |
79 | ||
80 | /* Locate the region we are either in or before. */ | |
81 | list_for_each_entry(rg, head, link) | |
82 | if (f <= rg->to) | |
83 | break; | |
84 | ||
85 | /* Round our left edge to the current segment if it encloses us. */ | |
86 | if (f > rg->from) | |
87 | f = rg->from; | |
88 | ||
89 | /* Check for and consume any regions we now overlap with. */ | |
90 | nrg = rg; | |
91 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
92 | if (&rg->link == head) | |
93 | break; | |
94 | if (rg->from > t) | |
95 | break; | |
96 | ||
97 | /* If this area reaches higher then extend our area to | |
98 | * include it completely. If this is not the first area | |
99 | * which we intend to reuse, free it. */ | |
100 | if (rg->to > t) | |
101 | t = rg->to; | |
102 | if (rg != nrg) { | |
103 | list_del(&rg->link); | |
104 | kfree(rg); | |
105 | } | |
106 | } | |
107 | nrg->from = f; | |
108 | nrg->to = t; | |
109 | return 0; | |
110 | } | |
111 | ||
112 | static long region_chg(struct list_head *head, long f, long t) | |
113 | { | |
114 | struct file_region *rg, *nrg; | |
115 | long chg = 0; | |
116 | ||
117 | /* Locate the region we are before or in. */ | |
118 | list_for_each_entry(rg, head, link) | |
119 | if (f <= rg->to) | |
120 | break; | |
121 | ||
122 | /* If we are below the current region then a new region is required. | |
123 | * Subtle, allocate a new region at the position but make it zero | |
124 | * size such that we can guarantee to record the reservation. */ | |
125 | if (&rg->link == head || t < rg->from) { | |
126 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
127 | if (!nrg) | |
128 | return -ENOMEM; | |
129 | nrg->from = f; | |
130 | nrg->to = f; | |
131 | INIT_LIST_HEAD(&nrg->link); | |
132 | list_add(&nrg->link, rg->link.prev); | |
133 | ||
134 | return t - f; | |
135 | } | |
136 | ||
137 | /* Round our left edge to the current segment if it encloses us. */ | |
138 | if (f > rg->from) | |
139 | f = rg->from; | |
140 | chg = t - f; | |
141 | ||
142 | /* Check for and consume any regions we now overlap with. */ | |
143 | list_for_each_entry(rg, rg->link.prev, link) { | |
144 | if (&rg->link == head) | |
145 | break; | |
146 | if (rg->from > t) | |
147 | return chg; | |
148 | ||
149 | /* We overlap with this area, if it extends futher than | |
150 | * us then we must extend ourselves. Account for its | |
151 | * existing reservation. */ | |
152 | if (rg->to > t) { | |
153 | chg += rg->to - t; | |
154 | t = rg->to; | |
155 | } | |
156 | chg -= rg->to - rg->from; | |
157 | } | |
158 | return chg; | |
159 | } | |
160 | ||
161 | static long region_truncate(struct list_head *head, long end) | |
162 | { | |
163 | struct file_region *rg, *trg; | |
164 | long chg = 0; | |
165 | ||
166 | /* Locate the region we are either in or before. */ | |
167 | list_for_each_entry(rg, head, link) | |
168 | if (end <= rg->to) | |
169 | break; | |
170 | if (&rg->link == head) | |
171 | return 0; | |
172 | ||
173 | /* If we are in the middle of a region then adjust it. */ | |
174 | if (end > rg->from) { | |
175 | chg = rg->to - end; | |
176 | rg->to = end; | |
177 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
178 | } | |
179 | ||
180 | /* Drop any remaining regions. */ | |
181 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
182 | if (&rg->link == head) | |
183 | break; | |
184 | chg += rg->to - rg->from; | |
185 | list_del(&rg->link); | |
186 | kfree(rg); | |
187 | } | |
188 | return chg; | |
189 | } | |
190 | ||
84afd99b AW |
191 | static long region_count(struct list_head *head, long f, long t) |
192 | { | |
193 | struct file_region *rg; | |
194 | long chg = 0; | |
195 | ||
196 | /* Locate each segment we overlap with, and count that overlap. */ | |
197 | list_for_each_entry(rg, head, link) { | |
198 | int seg_from; | |
199 | int seg_to; | |
200 | ||
201 | if (rg->to <= f) | |
202 | continue; | |
203 | if (rg->from >= t) | |
204 | break; | |
205 | ||
206 | seg_from = max(rg->from, f); | |
207 | seg_to = min(rg->to, t); | |
208 | ||
209 | chg += seg_to - seg_from; | |
210 | } | |
211 | ||
212 | return chg; | |
213 | } | |
214 | ||
e7c4b0bf AW |
215 | /* |
216 | * Convert the address within this vma to the page offset within | |
217 | * the mapping, in pagecache page units; huge pages here. | |
218 | */ | |
a5516438 AK |
219 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
220 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 221 | { |
a5516438 AK |
222 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
223 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
224 | } |
225 | ||
0fe6e20b NH |
226 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
227 | unsigned long address) | |
228 | { | |
229 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
230 | } | |
231 | ||
08fba699 MG |
232 | /* |
233 | * Return the size of the pages allocated when backing a VMA. In the majority | |
234 | * cases this will be same size as used by the page table entries. | |
235 | */ | |
236 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
237 | { | |
238 | struct hstate *hstate; | |
239 | ||
240 | if (!is_vm_hugetlb_page(vma)) | |
241 | return PAGE_SIZE; | |
242 | ||
243 | hstate = hstate_vma(vma); | |
244 | ||
245 | return 1UL << (hstate->order + PAGE_SHIFT); | |
246 | } | |
f340ca0f | 247 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 248 | |
3340289d MG |
249 | /* |
250 | * Return the page size being used by the MMU to back a VMA. In the majority | |
251 | * of cases, the page size used by the kernel matches the MMU size. On | |
252 | * architectures where it differs, an architecture-specific version of this | |
253 | * function is required. | |
254 | */ | |
255 | #ifndef vma_mmu_pagesize | |
256 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
257 | { | |
258 | return vma_kernel_pagesize(vma); | |
259 | } | |
260 | #endif | |
261 | ||
84afd99b AW |
262 | /* |
263 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
264 | * bits of the reservation map pointer, which are always clear due to | |
265 | * alignment. | |
266 | */ | |
267 | #define HPAGE_RESV_OWNER (1UL << 0) | |
268 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 269 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 270 | |
a1e78772 MG |
271 | /* |
272 | * These helpers are used to track how many pages are reserved for | |
273 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
274 | * is guaranteed to have their future faults succeed. | |
275 | * | |
276 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
277 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
278 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
279 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
280 | * |
281 | * The private mapping reservation is represented in a subtly different | |
282 | * manner to a shared mapping. A shared mapping has a region map associated | |
283 | * with the underlying file, this region map represents the backing file | |
284 | * pages which have ever had a reservation assigned which this persists even | |
285 | * after the page is instantiated. A private mapping has a region map | |
286 | * associated with the original mmap which is attached to all VMAs which | |
287 | * reference it, this region map represents those offsets which have consumed | |
288 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 289 | */ |
e7c4b0bf AW |
290 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
291 | { | |
292 | return (unsigned long)vma->vm_private_data; | |
293 | } | |
294 | ||
295 | static void set_vma_private_data(struct vm_area_struct *vma, | |
296 | unsigned long value) | |
297 | { | |
298 | vma->vm_private_data = (void *)value; | |
299 | } | |
300 | ||
84afd99b AW |
301 | struct resv_map { |
302 | struct kref refs; | |
303 | struct list_head regions; | |
304 | }; | |
305 | ||
2a4b3ded | 306 | static struct resv_map *resv_map_alloc(void) |
84afd99b AW |
307 | { |
308 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
309 | if (!resv_map) | |
310 | return NULL; | |
311 | ||
312 | kref_init(&resv_map->refs); | |
313 | INIT_LIST_HEAD(&resv_map->regions); | |
314 | ||
315 | return resv_map; | |
316 | } | |
317 | ||
2a4b3ded | 318 | static void resv_map_release(struct kref *ref) |
84afd99b AW |
319 | { |
320 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
321 | ||
322 | /* Clear out any active regions before we release the map. */ | |
323 | region_truncate(&resv_map->regions, 0); | |
324 | kfree(resv_map); | |
325 | } | |
326 | ||
327 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) | |
a1e78772 MG |
328 | { |
329 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 330 | if (!(vma->vm_flags & VM_MAYSHARE)) |
84afd99b AW |
331 | return (struct resv_map *)(get_vma_private_data(vma) & |
332 | ~HPAGE_RESV_MASK); | |
2a4b3ded | 333 | return NULL; |
a1e78772 MG |
334 | } |
335 | ||
84afd99b | 336 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 MG |
337 | { |
338 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 339 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
a1e78772 | 340 | |
84afd99b AW |
341 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
342 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
343 | } |
344 | ||
345 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
346 | { | |
04f2cbe3 | 347 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
f83a275d | 348 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
e7c4b0bf AW |
349 | |
350 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
351 | } |
352 | ||
353 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
354 | { | |
355 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
e7c4b0bf AW |
356 | |
357 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
358 | } |
359 | ||
360 | /* Decrement the reserved pages in the hugepage pool by one */ | |
a5516438 AK |
361 | static void decrement_hugepage_resv_vma(struct hstate *h, |
362 | struct vm_area_struct *vma) | |
a1e78772 | 363 | { |
c37f9fb1 AW |
364 | if (vma->vm_flags & VM_NORESERVE) |
365 | return; | |
366 | ||
f83a275d | 367 | if (vma->vm_flags & VM_MAYSHARE) { |
a1e78772 | 368 | /* Shared mappings always use reserves */ |
a5516438 | 369 | h->resv_huge_pages--; |
84afd99b | 370 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
a1e78772 MG |
371 | /* |
372 | * Only the process that called mmap() has reserves for | |
373 | * private mappings. | |
374 | */ | |
a5516438 | 375 | h->resv_huge_pages--; |
a1e78772 MG |
376 | } |
377 | } | |
378 | ||
04f2cbe3 | 379 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
380 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
381 | { | |
382 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 383 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
384 | vma->vm_private_data = (void *)0; |
385 | } | |
386 | ||
387 | /* Returns true if the VMA has associated reserve pages */ | |
7f09ca51 | 388 | static int vma_has_reserves(struct vm_area_struct *vma) |
a1e78772 | 389 | { |
f83a275d | 390 | if (vma->vm_flags & VM_MAYSHARE) |
7f09ca51 MG |
391 | return 1; |
392 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) | |
393 | return 1; | |
394 | return 0; | |
a1e78772 MG |
395 | } |
396 | ||
69d177c2 AW |
397 | static void clear_gigantic_page(struct page *page, |
398 | unsigned long addr, unsigned long sz) | |
399 | { | |
400 | int i; | |
401 | struct page *p = page; | |
402 | ||
403 | might_sleep(); | |
404 | for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) { | |
405 | cond_resched(); | |
406 | clear_user_highpage(p, addr + i * PAGE_SIZE); | |
407 | } | |
408 | } | |
a5516438 AK |
409 | static void clear_huge_page(struct page *page, |
410 | unsigned long addr, unsigned long sz) | |
79ac6ba4 DG |
411 | { |
412 | int i; | |
413 | ||
74dbdd23 | 414 | if (unlikely(sz/PAGE_SIZE > MAX_ORDER_NR_PAGES)) { |
ebdd4aea HE |
415 | clear_gigantic_page(page, addr, sz); |
416 | return; | |
417 | } | |
69d177c2 | 418 | |
79ac6ba4 | 419 | might_sleep(); |
a5516438 | 420 | for (i = 0; i < sz/PAGE_SIZE; i++) { |
79ac6ba4 | 421 | cond_resched(); |
281e0e3b | 422 | clear_user_highpage(page + i, addr + i * PAGE_SIZE); |
79ac6ba4 DG |
423 | } |
424 | } | |
425 | ||
69d177c2 AW |
426 | static void copy_gigantic_page(struct page *dst, struct page *src, |
427 | unsigned long addr, struct vm_area_struct *vma) | |
428 | { | |
429 | int i; | |
430 | struct hstate *h = hstate_vma(vma); | |
431 | struct page *dst_base = dst; | |
432 | struct page *src_base = src; | |
433 | might_sleep(); | |
434 | for (i = 0; i < pages_per_huge_page(h); ) { | |
435 | cond_resched(); | |
436 | copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); | |
437 | ||
438 | i++; | |
439 | dst = mem_map_next(dst, dst_base, i); | |
440 | src = mem_map_next(src, src_base, i); | |
441 | } | |
442 | } | |
79ac6ba4 | 443 | static void copy_huge_page(struct page *dst, struct page *src, |
9de455b2 | 444 | unsigned long addr, struct vm_area_struct *vma) |
79ac6ba4 DG |
445 | { |
446 | int i; | |
a5516438 | 447 | struct hstate *h = hstate_vma(vma); |
79ac6ba4 | 448 | |
ebdd4aea HE |
449 | if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { |
450 | copy_gigantic_page(dst, src, addr, vma); | |
451 | return; | |
452 | } | |
69d177c2 | 453 | |
79ac6ba4 | 454 | might_sleep(); |
a5516438 | 455 | for (i = 0; i < pages_per_huge_page(h); i++) { |
79ac6ba4 | 456 | cond_resched(); |
9de455b2 | 457 | copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); |
79ac6ba4 DG |
458 | } |
459 | } | |
460 | ||
a5516438 | 461 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
462 | { |
463 | int nid = page_to_nid(page); | |
a5516438 AK |
464 | list_add(&page->lru, &h->hugepage_freelists[nid]); |
465 | h->free_huge_pages++; | |
466 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
467 | } |
468 | ||
a5516438 AK |
469 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
470 | struct vm_area_struct *vma, | |
04f2cbe3 | 471 | unsigned long address, int avoid_reserve) |
1da177e4 | 472 | { |
31a5c6e4 | 473 | int nid; |
1da177e4 | 474 | struct page *page = NULL; |
480eccf9 | 475 | struct mempolicy *mpol; |
19770b32 | 476 | nodemask_t *nodemask; |
c0ff7453 | 477 | struct zonelist *zonelist; |
dd1a239f MG |
478 | struct zone *zone; |
479 | struct zoneref *z; | |
1da177e4 | 480 | |
c0ff7453 MX |
481 | get_mems_allowed(); |
482 | zonelist = huge_zonelist(vma, address, | |
483 | htlb_alloc_mask, &mpol, &nodemask); | |
a1e78772 MG |
484 | /* |
485 | * A child process with MAP_PRIVATE mappings created by their parent | |
486 | * have no page reserves. This check ensures that reservations are | |
487 | * not "stolen". The child may still get SIGKILLed | |
488 | */ | |
7f09ca51 | 489 | if (!vma_has_reserves(vma) && |
a5516438 | 490 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 491 | goto err; |
a1e78772 | 492 | |
04f2cbe3 | 493 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 494 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 495 | goto err;; |
04f2cbe3 | 496 | |
19770b32 MG |
497 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
498 | MAX_NR_ZONES - 1, nodemask) { | |
54a6eb5c MG |
499 | nid = zone_to_nid(zone); |
500 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) && | |
a5516438 AK |
501 | !list_empty(&h->hugepage_freelists[nid])) { |
502 | page = list_entry(h->hugepage_freelists[nid].next, | |
3abf7afd AM |
503 | struct page, lru); |
504 | list_del(&page->lru); | |
a5516438 AK |
505 | h->free_huge_pages--; |
506 | h->free_huge_pages_node[nid]--; | |
04f2cbe3 MG |
507 | |
508 | if (!avoid_reserve) | |
a5516438 | 509 | decrement_hugepage_resv_vma(h, vma); |
a1e78772 | 510 | |
5ab3ee7b | 511 | break; |
3abf7afd | 512 | } |
1da177e4 | 513 | } |
c0ff7453 | 514 | err: |
52cd3b07 | 515 | mpol_cond_put(mpol); |
c0ff7453 | 516 | put_mems_allowed(); |
1da177e4 LT |
517 | return page; |
518 | } | |
519 | ||
a5516438 | 520 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
521 | { |
522 | int i; | |
a5516438 | 523 | |
18229df5 AW |
524 | VM_BUG_ON(h->order >= MAX_ORDER); |
525 | ||
a5516438 AK |
526 | h->nr_huge_pages--; |
527 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
528 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
6af2acb6 AL |
529 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | |
530 | 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | | |
531 | 1 << PG_private | 1<< PG_writeback); | |
532 | } | |
533 | set_compound_page_dtor(page, NULL); | |
534 | set_page_refcounted(page); | |
7f2e9525 | 535 | arch_release_hugepage(page); |
a5516438 | 536 | __free_pages(page, huge_page_order(h)); |
6af2acb6 AL |
537 | } |
538 | ||
e5ff2159 AK |
539 | struct hstate *size_to_hstate(unsigned long size) |
540 | { | |
541 | struct hstate *h; | |
542 | ||
543 | for_each_hstate(h) { | |
544 | if (huge_page_size(h) == size) | |
545 | return h; | |
546 | } | |
547 | return NULL; | |
548 | } | |
549 | ||
27a85ef1 DG |
550 | static void free_huge_page(struct page *page) |
551 | { | |
a5516438 AK |
552 | /* |
553 | * Can't pass hstate in here because it is called from the | |
554 | * compound page destructor. | |
555 | */ | |
e5ff2159 | 556 | struct hstate *h = page_hstate(page); |
7893d1d5 | 557 | int nid = page_to_nid(page); |
c79fb75e | 558 | struct address_space *mapping; |
27a85ef1 | 559 | |
c79fb75e | 560 | mapping = (struct address_space *) page_private(page); |
e5df70ab | 561 | set_page_private(page, 0); |
23be7468 | 562 | page->mapping = NULL; |
7893d1d5 | 563 | BUG_ON(page_count(page)); |
0fe6e20b | 564 | BUG_ON(page_mapcount(page)); |
27a85ef1 DG |
565 | INIT_LIST_HEAD(&page->lru); |
566 | ||
567 | spin_lock(&hugetlb_lock); | |
aa888a74 | 568 | if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { |
a5516438 AK |
569 | update_and_free_page(h, page); |
570 | h->surplus_huge_pages--; | |
571 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 572 | } else { |
a5516438 | 573 | enqueue_huge_page(h, page); |
7893d1d5 | 574 | } |
27a85ef1 | 575 | spin_unlock(&hugetlb_lock); |
c79fb75e | 576 | if (mapping) |
9a119c05 | 577 | hugetlb_put_quota(mapping, 1); |
27a85ef1 DG |
578 | } |
579 | ||
a5516438 | 580 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 AK |
581 | { |
582 | set_compound_page_dtor(page, free_huge_page); | |
583 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
584 | h->nr_huge_pages++; |
585 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
586 | spin_unlock(&hugetlb_lock); |
587 | put_page(page); /* free it into the hugepage allocator */ | |
588 | } | |
589 | ||
20a0307c WF |
590 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
591 | { | |
592 | int i; | |
593 | int nr_pages = 1 << order; | |
594 | struct page *p = page + 1; | |
595 | ||
596 | /* we rely on prep_new_huge_page to set the destructor */ | |
597 | set_compound_order(page, order); | |
598 | __SetPageHead(page); | |
599 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
600 | __SetPageTail(p); | |
601 | p->first_page = page; | |
602 | } | |
603 | } | |
604 | ||
605 | int PageHuge(struct page *page) | |
606 | { | |
607 | compound_page_dtor *dtor; | |
608 | ||
609 | if (!PageCompound(page)) | |
610 | return 0; | |
611 | ||
612 | page = compound_head(page); | |
613 | dtor = get_compound_page_dtor(page); | |
614 | ||
615 | return dtor == free_huge_page; | |
616 | } | |
617 | ||
a5516438 | 618 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 619 | { |
1da177e4 | 620 | struct page *page; |
f96efd58 | 621 | |
aa888a74 AK |
622 | if (h->order >= MAX_ORDER) |
623 | return NULL; | |
624 | ||
6484eb3e | 625 | page = alloc_pages_exact_node(nid, |
551883ae NA |
626 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| |
627 | __GFP_REPEAT|__GFP_NOWARN, | |
a5516438 | 628 | huge_page_order(h)); |
1da177e4 | 629 | if (page) { |
7f2e9525 | 630 | if (arch_prepare_hugepage(page)) { |
caff3a2c | 631 | __free_pages(page, huge_page_order(h)); |
7b8ee84d | 632 | return NULL; |
7f2e9525 | 633 | } |
a5516438 | 634 | prep_new_huge_page(h, page, nid); |
1da177e4 | 635 | } |
63b4613c NA |
636 | |
637 | return page; | |
638 | } | |
639 | ||
9a76db09 | 640 | /* |
6ae11b27 LS |
641 | * common helper functions for hstate_next_node_to_{alloc|free}. |
642 | * We may have allocated or freed a huge page based on a different | |
643 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
644 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
645 | * node for alloc or free. | |
9a76db09 | 646 | */ |
6ae11b27 | 647 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) |
9a76db09 | 648 | { |
6ae11b27 | 649 | nid = next_node(nid, *nodes_allowed); |
9a76db09 | 650 | if (nid == MAX_NUMNODES) |
6ae11b27 | 651 | nid = first_node(*nodes_allowed); |
9a76db09 LS |
652 | VM_BUG_ON(nid >= MAX_NUMNODES); |
653 | ||
654 | return nid; | |
655 | } | |
656 | ||
6ae11b27 LS |
657 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) |
658 | { | |
659 | if (!node_isset(nid, *nodes_allowed)) | |
660 | nid = next_node_allowed(nid, nodes_allowed); | |
661 | return nid; | |
662 | } | |
663 | ||
5ced66c9 | 664 | /* |
6ae11b27 LS |
665 | * returns the previously saved node ["this node"] from which to |
666 | * allocate a persistent huge page for the pool and advance the | |
667 | * next node from which to allocate, handling wrap at end of node | |
668 | * mask. | |
5ced66c9 | 669 | */ |
6ae11b27 LS |
670 | static int hstate_next_node_to_alloc(struct hstate *h, |
671 | nodemask_t *nodes_allowed) | |
5ced66c9 | 672 | { |
6ae11b27 LS |
673 | int nid; |
674 | ||
675 | VM_BUG_ON(!nodes_allowed); | |
676 | ||
677 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
678 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 679 | |
9a76db09 | 680 | return nid; |
5ced66c9 AK |
681 | } |
682 | ||
6ae11b27 | 683 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
63b4613c NA |
684 | { |
685 | struct page *page; | |
686 | int start_nid; | |
687 | int next_nid; | |
688 | int ret = 0; | |
689 | ||
6ae11b27 | 690 | start_nid = hstate_next_node_to_alloc(h, nodes_allowed); |
e8c5c824 | 691 | next_nid = start_nid; |
63b4613c NA |
692 | |
693 | do { | |
e8c5c824 | 694 | page = alloc_fresh_huge_page_node(h, next_nid); |
9a76db09 | 695 | if (page) { |
63b4613c | 696 | ret = 1; |
9a76db09 LS |
697 | break; |
698 | } | |
6ae11b27 | 699 | next_nid = hstate_next_node_to_alloc(h, nodes_allowed); |
9a76db09 | 700 | } while (next_nid != start_nid); |
63b4613c | 701 | |
3b116300 AL |
702 | if (ret) |
703 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
704 | else | |
705 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
706 | ||
63b4613c | 707 | return ret; |
1da177e4 LT |
708 | } |
709 | ||
e8c5c824 | 710 | /* |
6ae11b27 LS |
711 | * helper for free_pool_huge_page() - return the previously saved |
712 | * node ["this node"] from which to free a huge page. Advance the | |
713 | * next node id whether or not we find a free huge page to free so | |
714 | * that the next attempt to free addresses the next node. | |
e8c5c824 | 715 | */ |
6ae11b27 | 716 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) |
e8c5c824 | 717 | { |
6ae11b27 LS |
718 | int nid; |
719 | ||
720 | VM_BUG_ON(!nodes_allowed); | |
721 | ||
722 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
723 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 724 | |
9a76db09 | 725 | return nid; |
e8c5c824 LS |
726 | } |
727 | ||
728 | /* | |
729 | * Free huge page from pool from next node to free. | |
730 | * Attempt to keep persistent huge pages more or less | |
731 | * balanced over allowed nodes. | |
732 | * Called with hugetlb_lock locked. | |
733 | */ | |
6ae11b27 LS |
734 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
735 | bool acct_surplus) | |
e8c5c824 LS |
736 | { |
737 | int start_nid; | |
738 | int next_nid; | |
739 | int ret = 0; | |
740 | ||
6ae11b27 | 741 | start_nid = hstate_next_node_to_free(h, nodes_allowed); |
e8c5c824 LS |
742 | next_nid = start_nid; |
743 | ||
744 | do { | |
685f3457 LS |
745 | /* |
746 | * If we're returning unused surplus pages, only examine | |
747 | * nodes with surplus pages. | |
748 | */ | |
749 | if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && | |
750 | !list_empty(&h->hugepage_freelists[next_nid])) { | |
e8c5c824 LS |
751 | struct page *page = |
752 | list_entry(h->hugepage_freelists[next_nid].next, | |
753 | struct page, lru); | |
754 | list_del(&page->lru); | |
755 | h->free_huge_pages--; | |
756 | h->free_huge_pages_node[next_nid]--; | |
685f3457 LS |
757 | if (acct_surplus) { |
758 | h->surplus_huge_pages--; | |
759 | h->surplus_huge_pages_node[next_nid]--; | |
760 | } | |
e8c5c824 LS |
761 | update_and_free_page(h, page); |
762 | ret = 1; | |
9a76db09 | 763 | break; |
e8c5c824 | 764 | } |
6ae11b27 | 765 | next_nid = hstate_next_node_to_free(h, nodes_allowed); |
9a76db09 | 766 | } while (next_nid != start_nid); |
e8c5c824 LS |
767 | |
768 | return ret; | |
769 | } | |
770 | ||
a5516438 AK |
771 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
772 | struct vm_area_struct *vma, unsigned long address) | |
7893d1d5 AL |
773 | { |
774 | struct page *page; | |
d1c3fb1f | 775 | unsigned int nid; |
7893d1d5 | 776 | |
aa888a74 AK |
777 | if (h->order >= MAX_ORDER) |
778 | return NULL; | |
779 | ||
d1c3fb1f NA |
780 | /* |
781 | * Assume we will successfully allocate the surplus page to | |
782 | * prevent racing processes from causing the surplus to exceed | |
783 | * overcommit | |
784 | * | |
785 | * This however introduces a different race, where a process B | |
786 | * tries to grow the static hugepage pool while alloc_pages() is | |
787 | * called by process A. B will only examine the per-node | |
788 | * counters in determining if surplus huge pages can be | |
789 | * converted to normal huge pages in adjust_pool_surplus(). A | |
790 | * won't be able to increment the per-node counter, until the | |
791 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
792 | * no more huge pages can be converted from surplus to normal | |
793 | * state (and doesn't try to convert again). Thus, we have a | |
794 | * case where a surplus huge page exists, the pool is grown, and | |
795 | * the surplus huge page still exists after, even though it | |
796 | * should just have been converted to a normal huge page. This | |
797 | * does not leak memory, though, as the hugepage will be freed | |
798 | * once it is out of use. It also does not allow the counters to | |
799 | * go out of whack in adjust_pool_surplus() as we don't modify | |
800 | * the node values until we've gotten the hugepage and only the | |
801 | * per-node value is checked there. | |
802 | */ | |
803 | spin_lock(&hugetlb_lock); | |
a5516438 | 804 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
805 | spin_unlock(&hugetlb_lock); |
806 | return NULL; | |
807 | } else { | |
a5516438 AK |
808 | h->nr_huge_pages++; |
809 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
810 | } |
811 | spin_unlock(&hugetlb_lock); | |
812 | ||
551883ae NA |
813 | page = alloc_pages(htlb_alloc_mask|__GFP_COMP| |
814 | __GFP_REPEAT|__GFP_NOWARN, | |
a5516438 | 815 | huge_page_order(h)); |
d1c3fb1f | 816 | |
caff3a2c GS |
817 | if (page && arch_prepare_hugepage(page)) { |
818 | __free_pages(page, huge_page_order(h)); | |
819 | return NULL; | |
820 | } | |
821 | ||
d1c3fb1f | 822 | spin_lock(&hugetlb_lock); |
7893d1d5 | 823 | if (page) { |
2668db91 AL |
824 | /* |
825 | * This page is now managed by the hugetlb allocator and has | |
826 | * no users -- drop the buddy allocator's reference. | |
827 | */ | |
828 | put_page_testzero(page); | |
829 | VM_BUG_ON(page_count(page)); | |
d1c3fb1f | 830 | nid = page_to_nid(page); |
7893d1d5 | 831 | set_compound_page_dtor(page, free_huge_page); |
d1c3fb1f NA |
832 | /* |
833 | * We incremented the global counters already | |
834 | */ | |
a5516438 AK |
835 | h->nr_huge_pages_node[nid]++; |
836 | h->surplus_huge_pages_node[nid]++; | |
3b116300 | 837 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 838 | } else { |
a5516438 AK |
839 | h->nr_huge_pages--; |
840 | h->surplus_huge_pages--; | |
3b116300 | 841 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 842 | } |
d1c3fb1f | 843 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
844 | |
845 | return page; | |
846 | } | |
847 | ||
e4e574b7 AL |
848 | /* |
849 | * Increase the hugetlb pool such that it can accomodate a reservation | |
850 | * of size 'delta'. | |
851 | */ | |
a5516438 | 852 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
853 | { |
854 | struct list_head surplus_list; | |
855 | struct page *page, *tmp; | |
856 | int ret, i; | |
857 | int needed, allocated; | |
858 | ||
a5516438 | 859 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 860 | if (needed <= 0) { |
a5516438 | 861 | h->resv_huge_pages += delta; |
e4e574b7 | 862 | return 0; |
ac09b3a1 | 863 | } |
e4e574b7 AL |
864 | |
865 | allocated = 0; | |
866 | INIT_LIST_HEAD(&surplus_list); | |
867 | ||
868 | ret = -ENOMEM; | |
869 | retry: | |
870 | spin_unlock(&hugetlb_lock); | |
871 | for (i = 0; i < needed; i++) { | |
a5516438 | 872 | page = alloc_buddy_huge_page(h, NULL, 0); |
e4e574b7 AL |
873 | if (!page) { |
874 | /* | |
875 | * We were not able to allocate enough pages to | |
876 | * satisfy the entire reservation so we free what | |
877 | * we've allocated so far. | |
878 | */ | |
879 | spin_lock(&hugetlb_lock); | |
880 | needed = 0; | |
881 | goto free; | |
882 | } | |
883 | ||
884 | list_add(&page->lru, &surplus_list); | |
885 | } | |
886 | allocated += needed; | |
887 | ||
888 | /* | |
889 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
890 | * because either resv_huge_pages or free_huge_pages may have changed. | |
891 | */ | |
892 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
893 | needed = (h->resv_huge_pages + delta) - |
894 | (h->free_huge_pages + allocated); | |
e4e574b7 AL |
895 | if (needed > 0) |
896 | goto retry; | |
897 | ||
898 | /* | |
899 | * The surplus_list now contains _at_least_ the number of extra pages | |
900 | * needed to accomodate the reservation. Add the appropriate number | |
901 | * of pages to the hugetlb pool and free the extras back to the buddy | |
ac09b3a1 AL |
902 | * allocator. Commit the entire reservation here to prevent another |
903 | * process from stealing the pages as they are added to the pool but | |
904 | * before they are reserved. | |
e4e574b7 AL |
905 | */ |
906 | needed += allocated; | |
a5516438 | 907 | h->resv_huge_pages += delta; |
e4e574b7 AL |
908 | ret = 0; |
909 | free: | |
19fc3f0a | 910 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 911 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
912 | if ((--needed) < 0) |
913 | break; | |
e4e574b7 | 914 | list_del(&page->lru); |
a5516438 | 915 | enqueue_huge_page(h, page); |
19fc3f0a AL |
916 | } |
917 | ||
918 | /* Free unnecessary surplus pages to the buddy allocator */ | |
919 | if (!list_empty(&surplus_list)) { | |
920 | spin_unlock(&hugetlb_lock); | |
921 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { | |
922 | list_del(&page->lru); | |
af767cbd | 923 | /* |
2668db91 AL |
924 | * The page has a reference count of zero already, so |
925 | * call free_huge_page directly instead of using | |
926 | * put_page. This must be done with hugetlb_lock | |
af767cbd AL |
927 | * unlocked which is safe because free_huge_page takes |
928 | * hugetlb_lock before deciding how to free the page. | |
929 | */ | |
2668db91 | 930 | free_huge_page(page); |
af767cbd | 931 | } |
19fc3f0a | 932 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
933 | } |
934 | ||
935 | return ret; | |
936 | } | |
937 | ||
938 | /* | |
939 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
940 | * allocated to satisfy the reservation must be explicitly freed if they were | |
941 | * never used. | |
685f3457 | 942 | * Called with hugetlb_lock held. |
e4e574b7 | 943 | */ |
a5516438 AK |
944 | static void return_unused_surplus_pages(struct hstate *h, |
945 | unsigned long unused_resv_pages) | |
e4e574b7 | 946 | { |
e4e574b7 AL |
947 | unsigned long nr_pages; |
948 | ||
ac09b3a1 | 949 | /* Uncommit the reservation */ |
a5516438 | 950 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 951 | |
aa888a74 AK |
952 | /* Cannot return gigantic pages currently */ |
953 | if (h->order >= MAX_ORDER) | |
954 | return; | |
955 | ||
a5516438 | 956 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 957 | |
685f3457 LS |
958 | /* |
959 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
960 | * evenly across all nodes with memory. Iterate across these nodes |
961 | * until we can no longer free unreserved surplus pages. This occurs | |
962 | * when the nodes with surplus pages have no free pages. | |
963 | * free_pool_huge_page() will balance the the freed pages across the | |
964 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
965 | */ |
966 | while (nr_pages--) { | |
9b5e5d0f | 967 | if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1)) |
685f3457 | 968 | break; |
e4e574b7 AL |
969 | } |
970 | } | |
971 | ||
c37f9fb1 AW |
972 | /* |
973 | * Determine if the huge page at addr within the vma has an associated | |
974 | * reservation. Where it does not we will need to logically increase | |
975 | * reservation and actually increase quota before an allocation can occur. | |
976 | * Where any new reservation would be required the reservation change is | |
977 | * prepared, but not committed. Once the page has been quota'd allocated | |
978 | * an instantiated the change should be committed via vma_commit_reservation. | |
979 | * No action is required on failure. | |
980 | */ | |
e2f17d94 | 981 | static long vma_needs_reservation(struct hstate *h, |
a5516438 | 982 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 AW |
983 | { |
984 | struct address_space *mapping = vma->vm_file->f_mapping; | |
985 | struct inode *inode = mapping->host; | |
986 | ||
f83a275d | 987 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 988 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 AW |
989 | return region_chg(&inode->i_mapping->private_list, |
990 | idx, idx + 1); | |
991 | ||
84afd99b AW |
992 | } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
993 | return 1; | |
c37f9fb1 | 994 | |
84afd99b | 995 | } else { |
e2f17d94 | 996 | long err; |
a5516438 | 997 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
84afd99b AW |
998 | struct resv_map *reservations = vma_resv_map(vma); |
999 | ||
1000 | err = region_chg(&reservations->regions, idx, idx + 1); | |
1001 | if (err < 0) | |
1002 | return err; | |
1003 | return 0; | |
1004 | } | |
c37f9fb1 | 1005 | } |
a5516438 AK |
1006 | static void vma_commit_reservation(struct hstate *h, |
1007 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 AW |
1008 | { |
1009 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1010 | struct inode *inode = mapping->host; | |
1011 | ||
f83a275d | 1012 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1013 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 | 1014 | region_add(&inode->i_mapping->private_list, idx, idx + 1); |
84afd99b AW |
1015 | |
1016 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { | |
a5516438 | 1017 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
84afd99b AW |
1018 | struct resv_map *reservations = vma_resv_map(vma); |
1019 | ||
1020 | /* Mark this page used in the map. */ | |
1021 | region_add(&reservations->regions, idx, idx + 1); | |
c37f9fb1 AW |
1022 | } |
1023 | } | |
1024 | ||
a1e78772 | 1025 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1026 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1027 | { |
a5516438 | 1028 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1029 | struct page *page; |
a1e78772 MG |
1030 | struct address_space *mapping = vma->vm_file->f_mapping; |
1031 | struct inode *inode = mapping->host; | |
e2f17d94 | 1032 | long chg; |
a1e78772 MG |
1033 | |
1034 | /* | |
1035 | * Processes that did not create the mapping will have no reserves and | |
1036 | * will not have accounted against quota. Check that the quota can be | |
1037 | * made before satisfying the allocation | |
c37f9fb1 AW |
1038 | * MAP_NORESERVE mappings may also need pages and quota allocated |
1039 | * if no reserve mapping overlaps. | |
a1e78772 | 1040 | */ |
a5516438 | 1041 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 AW |
1042 | if (chg < 0) |
1043 | return ERR_PTR(chg); | |
1044 | if (chg) | |
a1e78772 MG |
1045 | if (hugetlb_get_quota(inode->i_mapping, chg)) |
1046 | return ERR_PTR(-ENOSPC); | |
1da177e4 LT |
1047 | |
1048 | spin_lock(&hugetlb_lock); | |
a5516438 | 1049 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); |
1da177e4 | 1050 | spin_unlock(&hugetlb_lock); |
b45b5bd6 | 1051 | |
68842c9b | 1052 | if (!page) { |
a5516438 | 1053 | page = alloc_buddy_huge_page(h, vma, addr); |
68842c9b | 1054 | if (!page) { |
a1e78772 | 1055 | hugetlb_put_quota(inode->i_mapping, chg); |
4a6018f7 | 1056 | return ERR_PTR(-VM_FAULT_SIGBUS); |
68842c9b KC |
1057 | } |
1058 | } | |
348ea204 | 1059 | |
a1e78772 MG |
1060 | set_page_refcounted(page); |
1061 | set_page_private(page, (unsigned long) mapping); | |
90d8b7e6 | 1062 | |
a5516438 | 1063 | vma_commit_reservation(h, vma, addr); |
c37f9fb1 | 1064 | |
90d8b7e6 | 1065 | return page; |
b45b5bd6 DG |
1066 | } |
1067 | ||
91f47662 | 1068 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1069 | { |
1070 | struct huge_bootmem_page *m; | |
9b5e5d0f | 1071 | int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); |
aa888a74 AK |
1072 | |
1073 | while (nr_nodes) { | |
1074 | void *addr; | |
1075 | ||
1076 | addr = __alloc_bootmem_node_nopanic( | |
6ae11b27 | 1077 | NODE_DATA(hstate_next_node_to_alloc(h, |
9b5e5d0f | 1078 | &node_states[N_HIGH_MEMORY])), |
aa888a74 AK |
1079 | huge_page_size(h), huge_page_size(h), 0); |
1080 | ||
1081 | if (addr) { | |
1082 | /* | |
1083 | * Use the beginning of the huge page to store the | |
1084 | * huge_bootmem_page struct (until gather_bootmem | |
1085 | * puts them into the mem_map). | |
1086 | */ | |
1087 | m = addr; | |
91f47662 | 1088 | goto found; |
aa888a74 | 1089 | } |
aa888a74 AK |
1090 | nr_nodes--; |
1091 | } | |
1092 | return 0; | |
1093 | ||
1094 | found: | |
1095 | BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); | |
1096 | /* Put them into a private list first because mem_map is not up yet */ | |
1097 | list_add(&m->list, &huge_boot_pages); | |
1098 | m->hstate = h; | |
1099 | return 1; | |
1100 | } | |
1101 | ||
18229df5 AW |
1102 | static void prep_compound_huge_page(struct page *page, int order) |
1103 | { | |
1104 | if (unlikely(order > (MAX_ORDER - 1))) | |
1105 | prep_compound_gigantic_page(page, order); | |
1106 | else | |
1107 | prep_compound_page(page, order); | |
1108 | } | |
1109 | ||
aa888a74 AK |
1110 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1111 | static void __init gather_bootmem_prealloc(void) | |
1112 | { | |
1113 | struct huge_bootmem_page *m; | |
1114 | ||
1115 | list_for_each_entry(m, &huge_boot_pages, list) { | |
1116 | struct page *page = virt_to_page(m); | |
1117 | struct hstate *h = m->hstate; | |
1118 | __ClearPageReserved(page); | |
1119 | WARN_ON(page_count(page) != 1); | |
18229df5 | 1120 | prep_compound_huge_page(page, h->order); |
aa888a74 AK |
1121 | prep_new_huge_page(h, page, page_to_nid(page)); |
1122 | } | |
1123 | } | |
1124 | ||
8faa8b07 | 1125 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1126 | { |
1127 | unsigned long i; | |
a5516438 | 1128 | |
e5ff2159 | 1129 | for (i = 0; i < h->max_huge_pages; ++i) { |
aa888a74 AK |
1130 | if (h->order >= MAX_ORDER) { |
1131 | if (!alloc_bootmem_huge_page(h)) | |
1132 | break; | |
9b5e5d0f LS |
1133 | } else if (!alloc_fresh_huge_page(h, |
1134 | &node_states[N_HIGH_MEMORY])) | |
1da177e4 | 1135 | break; |
1da177e4 | 1136 | } |
8faa8b07 | 1137 | h->max_huge_pages = i; |
e5ff2159 AK |
1138 | } |
1139 | ||
1140 | static void __init hugetlb_init_hstates(void) | |
1141 | { | |
1142 | struct hstate *h; | |
1143 | ||
1144 | for_each_hstate(h) { | |
8faa8b07 AK |
1145 | /* oversize hugepages were init'ed in early boot */ |
1146 | if (h->order < MAX_ORDER) | |
1147 | hugetlb_hstate_alloc_pages(h); | |
e5ff2159 AK |
1148 | } |
1149 | } | |
1150 | ||
4abd32db AK |
1151 | static char * __init memfmt(char *buf, unsigned long n) |
1152 | { | |
1153 | if (n >= (1UL << 30)) | |
1154 | sprintf(buf, "%lu GB", n >> 30); | |
1155 | else if (n >= (1UL << 20)) | |
1156 | sprintf(buf, "%lu MB", n >> 20); | |
1157 | else | |
1158 | sprintf(buf, "%lu KB", n >> 10); | |
1159 | return buf; | |
1160 | } | |
1161 | ||
e5ff2159 AK |
1162 | static void __init report_hugepages(void) |
1163 | { | |
1164 | struct hstate *h; | |
1165 | ||
1166 | for_each_hstate(h) { | |
4abd32db AK |
1167 | char buf[32]; |
1168 | printk(KERN_INFO "HugeTLB registered %s page size, " | |
1169 | "pre-allocated %ld pages\n", | |
1170 | memfmt(buf, huge_page_size(h)), | |
1171 | h->free_huge_pages); | |
e5ff2159 AK |
1172 | } |
1173 | } | |
1174 | ||
1da177e4 | 1175 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1176 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1177 | nodemask_t *nodes_allowed) | |
1da177e4 | 1178 | { |
4415cc8d CL |
1179 | int i; |
1180 | ||
aa888a74 AK |
1181 | if (h->order >= MAX_ORDER) |
1182 | return; | |
1183 | ||
6ae11b27 | 1184 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1185 | struct page *page, *next; |
a5516438 AK |
1186 | struct list_head *freel = &h->hugepage_freelists[i]; |
1187 | list_for_each_entry_safe(page, next, freel, lru) { | |
1188 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1189 | return; |
1da177e4 LT |
1190 | if (PageHighMem(page)) |
1191 | continue; | |
1192 | list_del(&page->lru); | |
e5ff2159 | 1193 | update_and_free_page(h, page); |
a5516438 AK |
1194 | h->free_huge_pages--; |
1195 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1196 | } |
1197 | } | |
1198 | } | |
1199 | #else | |
6ae11b27 LS |
1200 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1201 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1202 | { |
1203 | } | |
1204 | #endif | |
1205 | ||
20a0307c WF |
1206 | /* |
1207 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1208 | * balanced by operating on them in a round-robin fashion. | |
1209 | * Returns 1 if an adjustment was made. | |
1210 | */ | |
6ae11b27 LS |
1211 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1212 | int delta) | |
20a0307c | 1213 | { |
e8c5c824 | 1214 | int start_nid, next_nid; |
20a0307c WF |
1215 | int ret = 0; |
1216 | ||
1217 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1218 | |
e8c5c824 | 1219 | if (delta < 0) |
6ae11b27 | 1220 | start_nid = hstate_next_node_to_alloc(h, nodes_allowed); |
e8c5c824 | 1221 | else |
6ae11b27 | 1222 | start_nid = hstate_next_node_to_free(h, nodes_allowed); |
e8c5c824 LS |
1223 | next_nid = start_nid; |
1224 | ||
1225 | do { | |
1226 | int nid = next_nid; | |
1227 | if (delta < 0) { | |
e8c5c824 LS |
1228 | /* |
1229 | * To shrink on this node, there must be a surplus page | |
1230 | */ | |
9a76db09 | 1231 | if (!h->surplus_huge_pages_node[nid]) { |
6ae11b27 LS |
1232 | next_nid = hstate_next_node_to_alloc(h, |
1233 | nodes_allowed); | |
e8c5c824 | 1234 | continue; |
9a76db09 | 1235 | } |
e8c5c824 LS |
1236 | } |
1237 | if (delta > 0) { | |
e8c5c824 LS |
1238 | /* |
1239 | * Surplus cannot exceed the total number of pages | |
1240 | */ | |
1241 | if (h->surplus_huge_pages_node[nid] >= | |
9a76db09 | 1242 | h->nr_huge_pages_node[nid]) { |
6ae11b27 LS |
1243 | next_nid = hstate_next_node_to_free(h, |
1244 | nodes_allowed); | |
e8c5c824 | 1245 | continue; |
9a76db09 | 1246 | } |
e8c5c824 | 1247 | } |
20a0307c WF |
1248 | |
1249 | h->surplus_huge_pages += delta; | |
1250 | h->surplus_huge_pages_node[nid] += delta; | |
1251 | ret = 1; | |
1252 | break; | |
e8c5c824 | 1253 | } while (next_nid != start_nid); |
20a0307c | 1254 | |
20a0307c WF |
1255 | return ret; |
1256 | } | |
1257 | ||
a5516438 | 1258 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1259 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1260 | nodemask_t *nodes_allowed) | |
1da177e4 | 1261 | { |
7893d1d5 | 1262 | unsigned long min_count, ret; |
1da177e4 | 1263 | |
aa888a74 AK |
1264 | if (h->order >= MAX_ORDER) |
1265 | return h->max_huge_pages; | |
1266 | ||
7893d1d5 AL |
1267 | /* |
1268 | * Increase the pool size | |
1269 | * First take pages out of surplus state. Then make up the | |
1270 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1271 | * |
1272 | * We might race with alloc_buddy_huge_page() here and be unable | |
1273 | * to convert a surplus huge page to a normal huge page. That is | |
1274 | * not critical, though, it just means the overall size of the | |
1275 | * pool might be one hugepage larger than it needs to be, but | |
1276 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1277 | */ |
1da177e4 | 1278 | spin_lock(&hugetlb_lock); |
a5516438 | 1279 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 1280 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
1281 | break; |
1282 | } | |
1283 | ||
a5516438 | 1284 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1285 | /* |
1286 | * If this allocation races such that we no longer need the | |
1287 | * page, free_huge_page will handle it by freeing the page | |
1288 | * and reducing the surplus. | |
1289 | */ | |
1290 | spin_unlock(&hugetlb_lock); | |
6ae11b27 | 1291 | ret = alloc_fresh_huge_page(h, nodes_allowed); |
7893d1d5 AL |
1292 | spin_lock(&hugetlb_lock); |
1293 | if (!ret) | |
1294 | goto out; | |
1295 | ||
536240f2 MG |
1296 | /* Bail for signals. Probably ctrl-c from user */ |
1297 | if (signal_pending(current)) | |
1298 | goto out; | |
7893d1d5 | 1299 | } |
7893d1d5 AL |
1300 | |
1301 | /* | |
1302 | * Decrease the pool size | |
1303 | * First return free pages to the buddy allocator (being careful | |
1304 | * to keep enough around to satisfy reservations). Then place | |
1305 | * pages into surplus state as needed so the pool will shrink | |
1306 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1307 | * |
1308 | * By placing pages into the surplus state independent of the | |
1309 | * overcommit value, we are allowing the surplus pool size to | |
1310 | * exceed overcommit. There are few sane options here. Since | |
1311 | * alloc_buddy_huge_page() is checking the global counter, | |
1312 | * though, we'll note that we're not allowed to exceed surplus | |
1313 | * and won't grow the pool anywhere else. Not until one of the | |
1314 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1315 | */ |
a5516438 | 1316 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1317 | min_count = max(count, min_count); |
6ae11b27 | 1318 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 1319 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 1320 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 1321 | break; |
1da177e4 | 1322 | } |
a5516438 | 1323 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 1324 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
1325 | break; |
1326 | } | |
1327 | out: | |
a5516438 | 1328 | ret = persistent_huge_pages(h); |
1da177e4 | 1329 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1330 | return ret; |
1da177e4 LT |
1331 | } |
1332 | ||
a3437870 NA |
1333 | #define HSTATE_ATTR_RO(_name) \ |
1334 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1335 | ||
1336 | #define HSTATE_ATTR(_name) \ | |
1337 | static struct kobj_attribute _name##_attr = \ | |
1338 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1339 | ||
1340 | static struct kobject *hugepages_kobj; | |
1341 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1342 | ||
9a305230 LS |
1343 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
1344 | ||
1345 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
1346 | { |
1347 | int i; | |
9a305230 | 1348 | |
a3437870 | 1349 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
1350 | if (hstate_kobjs[i] == kobj) { |
1351 | if (nidp) | |
1352 | *nidp = NUMA_NO_NODE; | |
a3437870 | 1353 | return &hstates[i]; |
9a305230 LS |
1354 | } |
1355 | ||
1356 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
1357 | } |
1358 | ||
06808b08 | 1359 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
1360 | struct kobj_attribute *attr, char *buf) |
1361 | { | |
9a305230 LS |
1362 | struct hstate *h; |
1363 | unsigned long nr_huge_pages; | |
1364 | int nid; | |
1365 | ||
1366 | h = kobj_to_hstate(kobj, &nid); | |
1367 | if (nid == NUMA_NO_NODE) | |
1368 | nr_huge_pages = h->nr_huge_pages; | |
1369 | else | |
1370 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
1371 | ||
1372 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 1373 | } |
06808b08 LS |
1374 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
1375 | struct kobject *kobj, struct kobj_attribute *attr, | |
1376 | const char *buf, size_t len) | |
a3437870 NA |
1377 | { |
1378 | int err; | |
9a305230 | 1379 | int nid; |
06808b08 | 1380 | unsigned long count; |
9a305230 | 1381 | struct hstate *h; |
bad44b5b | 1382 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 1383 | |
06808b08 | 1384 | err = strict_strtoul(buf, 10, &count); |
a3437870 NA |
1385 | if (err) |
1386 | return 0; | |
1387 | ||
9a305230 LS |
1388 | h = kobj_to_hstate(kobj, &nid); |
1389 | if (nid == NUMA_NO_NODE) { | |
1390 | /* | |
1391 | * global hstate attribute | |
1392 | */ | |
1393 | if (!(obey_mempolicy && | |
1394 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1395 | NODEMASK_FREE(nodes_allowed); | |
1396 | nodes_allowed = &node_states[N_HIGH_MEMORY]; | |
1397 | } | |
1398 | } else if (nodes_allowed) { | |
1399 | /* | |
1400 | * per node hstate attribute: adjust count to global, | |
1401 | * but restrict alloc/free to the specified node. | |
1402 | */ | |
1403 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
1404 | init_nodemask_of_node(nodes_allowed, nid); | |
1405 | } else | |
1406 | nodes_allowed = &node_states[N_HIGH_MEMORY]; | |
1407 | ||
06808b08 | 1408 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 1409 | |
9b5e5d0f | 1410 | if (nodes_allowed != &node_states[N_HIGH_MEMORY]) |
06808b08 LS |
1411 | NODEMASK_FREE(nodes_allowed); |
1412 | ||
1413 | return len; | |
1414 | } | |
1415 | ||
1416 | static ssize_t nr_hugepages_show(struct kobject *kobj, | |
1417 | struct kobj_attribute *attr, char *buf) | |
1418 | { | |
1419 | return nr_hugepages_show_common(kobj, attr, buf); | |
1420 | } | |
1421 | ||
1422 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1423 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1424 | { | |
1425 | return nr_hugepages_store_common(false, kobj, attr, buf, len); | |
a3437870 NA |
1426 | } |
1427 | HSTATE_ATTR(nr_hugepages); | |
1428 | ||
06808b08 LS |
1429 | #ifdef CONFIG_NUMA |
1430 | ||
1431 | /* | |
1432 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
1433 | * huge page alloc/free. | |
1434 | */ | |
1435 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
1436 | struct kobj_attribute *attr, char *buf) | |
1437 | { | |
1438 | return nr_hugepages_show_common(kobj, attr, buf); | |
1439 | } | |
1440 | ||
1441 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
1442 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1443 | { | |
1444 | return nr_hugepages_store_common(true, kobj, attr, buf, len); | |
1445 | } | |
1446 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
1447 | #endif | |
1448 | ||
1449 | ||
a3437870 NA |
1450 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
1451 | struct kobj_attribute *attr, char *buf) | |
1452 | { | |
9a305230 | 1453 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1454 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
1455 | } | |
1456 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, | |
1457 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1458 | { | |
1459 | int err; | |
1460 | unsigned long input; | |
9a305230 | 1461 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1462 | |
1463 | err = strict_strtoul(buf, 10, &input); | |
1464 | if (err) | |
1465 | return 0; | |
1466 | ||
1467 | spin_lock(&hugetlb_lock); | |
1468 | h->nr_overcommit_huge_pages = input; | |
1469 | spin_unlock(&hugetlb_lock); | |
1470 | ||
1471 | return count; | |
1472 | } | |
1473 | HSTATE_ATTR(nr_overcommit_hugepages); | |
1474 | ||
1475 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
1476 | struct kobj_attribute *attr, char *buf) | |
1477 | { | |
9a305230 LS |
1478 | struct hstate *h; |
1479 | unsigned long free_huge_pages; | |
1480 | int nid; | |
1481 | ||
1482 | h = kobj_to_hstate(kobj, &nid); | |
1483 | if (nid == NUMA_NO_NODE) | |
1484 | free_huge_pages = h->free_huge_pages; | |
1485 | else | |
1486 | free_huge_pages = h->free_huge_pages_node[nid]; | |
1487 | ||
1488 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
1489 | } |
1490 | HSTATE_ATTR_RO(free_hugepages); | |
1491 | ||
1492 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
1493 | struct kobj_attribute *attr, char *buf) | |
1494 | { | |
9a305230 | 1495 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1496 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
1497 | } | |
1498 | HSTATE_ATTR_RO(resv_hugepages); | |
1499 | ||
1500 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
1501 | struct kobj_attribute *attr, char *buf) | |
1502 | { | |
9a305230 LS |
1503 | struct hstate *h; |
1504 | unsigned long surplus_huge_pages; | |
1505 | int nid; | |
1506 | ||
1507 | h = kobj_to_hstate(kobj, &nid); | |
1508 | if (nid == NUMA_NO_NODE) | |
1509 | surplus_huge_pages = h->surplus_huge_pages; | |
1510 | else | |
1511 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
1512 | ||
1513 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
1514 | } |
1515 | HSTATE_ATTR_RO(surplus_hugepages); | |
1516 | ||
1517 | static struct attribute *hstate_attrs[] = { | |
1518 | &nr_hugepages_attr.attr, | |
1519 | &nr_overcommit_hugepages_attr.attr, | |
1520 | &free_hugepages_attr.attr, | |
1521 | &resv_hugepages_attr.attr, | |
1522 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
1523 | #ifdef CONFIG_NUMA |
1524 | &nr_hugepages_mempolicy_attr.attr, | |
1525 | #endif | |
a3437870 NA |
1526 | NULL, |
1527 | }; | |
1528 | ||
1529 | static struct attribute_group hstate_attr_group = { | |
1530 | .attrs = hstate_attrs, | |
1531 | }; | |
1532 | ||
094e9539 JM |
1533 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
1534 | struct kobject **hstate_kobjs, | |
1535 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
1536 | { |
1537 | int retval; | |
9a305230 | 1538 | int hi = h - hstates; |
a3437870 | 1539 | |
9a305230 LS |
1540 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
1541 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
1542 | return -ENOMEM; |
1543 | ||
9a305230 | 1544 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 1545 | if (retval) |
9a305230 | 1546 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
1547 | |
1548 | return retval; | |
1549 | } | |
1550 | ||
1551 | static void __init hugetlb_sysfs_init(void) | |
1552 | { | |
1553 | struct hstate *h; | |
1554 | int err; | |
1555 | ||
1556 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
1557 | if (!hugepages_kobj) | |
1558 | return; | |
1559 | ||
1560 | for_each_hstate(h) { | |
9a305230 LS |
1561 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
1562 | hstate_kobjs, &hstate_attr_group); | |
a3437870 NA |
1563 | if (err) |
1564 | printk(KERN_ERR "Hugetlb: Unable to add hstate %s", | |
1565 | h->name); | |
1566 | } | |
1567 | } | |
1568 | ||
9a305230 LS |
1569 | #ifdef CONFIG_NUMA |
1570 | ||
1571 | /* | |
1572 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
1573 | * with node sysdevs in node_devices[] using a parallel array. The array | |
1574 | * index of a node sysdev or _hstate == node id. | |
1575 | * This is here to avoid any static dependency of the node sysdev driver, in | |
1576 | * the base kernel, on the hugetlb module. | |
1577 | */ | |
1578 | struct node_hstate { | |
1579 | struct kobject *hugepages_kobj; | |
1580 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1581 | }; | |
1582 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
1583 | ||
1584 | /* | |
1585 | * A subset of global hstate attributes for node sysdevs | |
1586 | */ | |
1587 | static struct attribute *per_node_hstate_attrs[] = { | |
1588 | &nr_hugepages_attr.attr, | |
1589 | &free_hugepages_attr.attr, | |
1590 | &surplus_hugepages_attr.attr, | |
1591 | NULL, | |
1592 | }; | |
1593 | ||
1594 | static struct attribute_group per_node_hstate_attr_group = { | |
1595 | .attrs = per_node_hstate_attrs, | |
1596 | }; | |
1597 | ||
1598 | /* | |
1599 | * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj. | |
1600 | * Returns node id via non-NULL nidp. | |
1601 | */ | |
1602 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1603 | { | |
1604 | int nid; | |
1605 | ||
1606 | for (nid = 0; nid < nr_node_ids; nid++) { | |
1607 | struct node_hstate *nhs = &node_hstates[nid]; | |
1608 | int i; | |
1609 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
1610 | if (nhs->hstate_kobjs[i] == kobj) { | |
1611 | if (nidp) | |
1612 | *nidp = nid; | |
1613 | return &hstates[i]; | |
1614 | } | |
1615 | } | |
1616 | ||
1617 | BUG(); | |
1618 | return NULL; | |
1619 | } | |
1620 | ||
1621 | /* | |
1622 | * Unregister hstate attributes from a single node sysdev. | |
1623 | * No-op if no hstate attributes attached. | |
1624 | */ | |
1625 | void hugetlb_unregister_node(struct node *node) | |
1626 | { | |
1627 | struct hstate *h; | |
1628 | struct node_hstate *nhs = &node_hstates[node->sysdev.id]; | |
1629 | ||
1630 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 1631 | return; /* no hstate attributes */ |
9a305230 LS |
1632 | |
1633 | for_each_hstate(h) | |
1634 | if (nhs->hstate_kobjs[h - hstates]) { | |
1635 | kobject_put(nhs->hstate_kobjs[h - hstates]); | |
1636 | nhs->hstate_kobjs[h - hstates] = NULL; | |
1637 | } | |
1638 | ||
1639 | kobject_put(nhs->hugepages_kobj); | |
1640 | nhs->hugepages_kobj = NULL; | |
1641 | } | |
1642 | ||
1643 | /* | |
1644 | * hugetlb module exit: unregister hstate attributes from node sysdevs | |
1645 | * that have them. | |
1646 | */ | |
1647 | static void hugetlb_unregister_all_nodes(void) | |
1648 | { | |
1649 | int nid; | |
1650 | ||
1651 | /* | |
1652 | * disable node sysdev registrations. | |
1653 | */ | |
1654 | register_hugetlbfs_with_node(NULL, NULL); | |
1655 | ||
1656 | /* | |
1657 | * remove hstate attributes from any nodes that have them. | |
1658 | */ | |
1659 | for (nid = 0; nid < nr_node_ids; nid++) | |
1660 | hugetlb_unregister_node(&node_devices[nid]); | |
1661 | } | |
1662 | ||
1663 | /* | |
1664 | * Register hstate attributes for a single node sysdev. | |
1665 | * No-op if attributes already registered. | |
1666 | */ | |
1667 | void hugetlb_register_node(struct node *node) | |
1668 | { | |
1669 | struct hstate *h; | |
1670 | struct node_hstate *nhs = &node_hstates[node->sysdev.id]; | |
1671 | int err; | |
1672 | ||
1673 | if (nhs->hugepages_kobj) | |
1674 | return; /* already allocated */ | |
1675 | ||
1676 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
1677 | &node->sysdev.kobj); | |
1678 | if (!nhs->hugepages_kobj) | |
1679 | return; | |
1680 | ||
1681 | for_each_hstate(h) { | |
1682 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
1683 | nhs->hstate_kobjs, | |
1684 | &per_node_hstate_attr_group); | |
1685 | if (err) { | |
1686 | printk(KERN_ERR "Hugetlb: Unable to add hstate %s" | |
1687 | " for node %d\n", | |
1688 | h->name, node->sysdev.id); | |
1689 | hugetlb_unregister_node(node); | |
1690 | break; | |
1691 | } | |
1692 | } | |
1693 | } | |
1694 | ||
1695 | /* | |
9b5e5d0f LS |
1696 | * hugetlb init time: register hstate attributes for all registered node |
1697 | * sysdevs of nodes that have memory. All on-line nodes should have | |
1698 | * registered their associated sysdev by this time. | |
9a305230 LS |
1699 | */ |
1700 | static void hugetlb_register_all_nodes(void) | |
1701 | { | |
1702 | int nid; | |
1703 | ||
9b5e5d0f | 1704 | for_each_node_state(nid, N_HIGH_MEMORY) { |
9a305230 LS |
1705 | struct node *node = &node_devices[nid]; |
1706 | if (node->sysdev.id == nid) | |
1707 | hugetlb_register_node(node); | |
1708 | } | |
1709 | ||
1710 | /* | |
1711 | * Let the node sysdev driver know we're here so it can | |
1712 | * [un]register hstate attributes on node hotplug. | |
1713 | */ | |
1714 | register_hugetlbfs_with_node(hugetlb_register_node, | |
1715 | hugetlb_unregister_node); | |
1716 | } | |
1717 | #else /* !CONFIG_NUMA */ | |
1718 | ||
1719 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1720 | { | |
1721 | BUG(); | |
1722 | if (nidp) | |
1723 | *nidp = -1; | |
1724 | return NULL; | |
1725 | } | |
1726 | ||
1727 | static void hugetlb_unregister_all_nodes(void) { } | |
1728 | ||
1729 | static void hugetlb_register_all_nodes(void) { } | |
1730 | ||
1731 | #endif | |
1732 | ||
a3437870 NA |
1733 | static void __exit hugetlb_exit(void) |
1734 | { | |
1735 | struct hstate *h; | |
1736 | ||
9a305230 LS |
1737 | hugetlb_unregister_all_nodes(); |
1738 | ||
a3437870 NA |
1739 | for_each_hstate(h) { |
1740 | kobject_put(hstate_kobjs[h - hstates]); | |
1741 | } | |
1742 | ||
1743 | kobject_put(hugepages_kobj); | |
1744 | } | |
1745 | module_exit(hugetlb_exit); | |
1746 | ||
1747 | static int __init hugetlb_init(void) | |
1748 | { | |
0ef89d25 BH |
1749 | /* Some platform decide whether they support huge pages at boot |
1750 | * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when | |
1751 | * there is no such support | |
1752 | */ | |
1753 | if (HPAGE_SHIFT == 0) | |
1754 | return 0; | |
a3437870 | 1755 | |
e11bfbfc NP |
1756 | if (!size_to_hstate(default_hstate_size)) { |
1757 | default_hstate_size = HPAGE_SIZE; | |
1758 | if (!size_to_hstate(default_hstate_size)) | |
1759 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 1760 | } |
e11bfbfc NP |
1761 | default_hstate_idx = size_to_hstate(default_hstate_size) - hstates; |
1762 | if (default_hstate_max_huge_pages) | |
1763 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
1764 | |
1765 | hugetlb_init_hstates(); | |
1766 | ||
aa888a74 AK |
1767 | gather_bootmem_prealloc(); |
1768 | ||
a3437870 NA |
1769 | report_hugepages(); |
1770 | ||
1771 | hugetlb_sysfs_init(); | |
1772 | ||
9a305230 LS |
1773 | hugetlb_register_all_nodes(); |
1774 | ||
a3437870 NA |
1775 | return 0; |
1776 | } | |
1777 | module_init(hugetlb_init); | |
1778 | ||
1779 | /* Should be called on processing a hugepagesz=... option */ | |
1780 | void __init hugetlb_add_hstate(unsigned order) | |
1781 | { | |
1782 | struct hstate *h; | |
8faa8b07 AK |
1783 | unsigned long i; |
1784 | ||
a3437870 NA |
1785 | if (size_to_hstate(PAGE_SIZE << order)) { |
1786 | printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); | |
1787 | return; | |
1788 | } | |
1789 | BUG_ON(max_hstate >= HUGE_MAX_HSTATE); | |
1790 | BUG_ON(order == 0); | |
1791 | h = &hstates[max_hstate++]; | |
1792 | h->order = order; | |
1793 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
1794 | h->nr_huge_pages = 0; |
1795 | h->free_huge_pages = 0; | |
1796 | for (i = 0; i < MAX_NUMNODES; ++i) | |
1797 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
9b5e5d0f LS |
1798 | h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]); |
1799 | h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]); | |
a3437870 NA |
1800 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
1801 | huge_page_size(h)/1024); | |
8faa8b07 | 1802 | |
a3437870 NA |
1803 | parsed_hstate = h; |
1804 | } | |
1805 | ||
e11bfbfc | 1806 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
1807 | { |
1808 | unsigned long *mhp; | |
8faa8b07 | 1809 | static unsigned long *last_mhp; |
a3437870 NA |
1810 | |
1811 | /* | |
1812 | * !max_hstate means we haven't parsed a hugepagesz= parameter yet, | |
1813 | * so this hugepages= parameter goes to the "default hstate". | |
1814 | */ | |
1815 | if (!max_hstate) | |
1816 | mhp = &default_hstate_max_huge_pages; | |
1817 | else | |
1818 | mhp = &parsed_hstate->max_huge_pages; | |
1819 | ||
8faa8b07 AK |
1820 | if (mhp == last_mhp) { |
1821 | printk(KERN_WARNING "hugepages= specified twice without " | |
1822 | "interleaving hugepagesz=, ignoring\n"); | |
1823 | return 1; | |
1824 | } | |
1825 | ||
a3437870 NA |
1826 | if (sscanf(s, "%lu", mhp) <= 0) |
1827 | *mhp = 0; | |
1828 | ||
8faa8b07 AK |
1829 | /* |
1830 | * Global state is always initialized later in hugetlb_init. | |
1831 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
1832 | * use the bootmem allocator. | |
1833 | */ | |
1834 | if (max_hstate && parsed_hstate->order >= MAX_ORDER) | |
1835 | hugetlb_hstate_alloc_pages(parsed_hstate); | |
1836 | ||
1837 | last_mhp = mhp; | |
1838 | ||
a3437870 NA |
1839 | return 1; |
1840 | } | |
e11bfbfc NP |
1841 | __setup("hugepages=", hugetlb_nrpages_setup); |
1842 | ||
1843 | static int __init hugetlb_default_setup(char *s) | |
1844 | { | |
1845 | default_hstate_size = memparse(s, &s); | |
1846 | return 1; | |
1847 | } | |
1848 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 1849 | |
8a213460 NA |
1850 | static unsigned int cpuset_mems_nr(unsigned int *array) |
1851 | { | |
1852 | int node; | |
1853 | unsigned int nr = 0; | |
1854 | ||
1855 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
1856 | nr += array[node]; | |
1857 | ||
1858 | return nr; | |
1859 | } | |
1860 | ||
1861 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
1862 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
1863 | struct ctl_table *table, int write, | |
1864 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 1865 | { |
e5ff2159 AK |
1866 | struct hstate *h = &default_hstate; |
1867 | unsigned long tmp; | |
1868 | ||
1869 | if (!write) | |
1870 | tmp = h->max_huge_pages; | |
1871 | ||
1872 | table->data = &tmp; | |
1873 | table->maxlen = sizeof(unsigned long); | |
8d65af78 | 1874 | proc_doulongvec_minmax(table, write, buffer, length, ppos); |
e5ff2159 | 1875 | |
06808b08 | 1876 | if (write) { |
bad44b5b DR |
1877 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, |
1878 | GFP_KERNEL | __GFP_NORETRY); | |
06808b08 LS |
1879 | if (!(obey_mempolicy && |
1880 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1881 | NODEMASK_FREE(nodes_allowed); | |
1882 | nodes_allowed = &node_states[N_HIGH_MEMORY]; | |
1883 | } | |
1884 | h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); | |
1885 | ||
1886 | if (nodes_allowed != &node_states[N_HIGH_MEMORY]) | |
1887 | NODEMASK_FREE(nodes_allowed); | |
1888 | } | |
e5ff2159 | 1889 | |
1da177e4 LT |
1890 | return 0; |
1891 | } | |
396faf03 | 1892 | |
06808b08 LS |
1893 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
1894 | void __user *buffer, size_t *length, loff_t *ppos) | |
1895 | { | |
1896 | ||
1897 | return hugetlb_sysctl_handler_common(false, table, write, | |
1898 | buffer, length, ppos); | |
1899 | } | |
1900 | ||
1901 | #ifdef CONFIG_NUMA | |
1902 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
1903 | void __user *buffer, size_t *length, loff_t *ppos) | |
1904 | { | |
1905 | return hugetlb_sysctl_handler_common(true, table, write, | |
1906 | buffer, length, ppos); | |
1907 | } | |
1908 | #endif /* CONFIG_NUMA */ | |
1909 | ||
396faf03 | 1910 | int hugetlb_treat_movable_handler(struct ctl_table *table, int write, |
8d65af78 | 1911 | void __user *buffer, |
396faf03 MG |
1912 | size_t *length, loff_t *ppos) |
1913 | { | |
8d65af78 | 1914 | proc_dointvec(table, write, buffer, length, ppos); |
396faf03 MG |
1915 | if (hugepages_treat_as_movable) |
1916 | htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; | |
1917 | else | |
1918 | htlb_alloc_mask = GFP_HIGHUSER; | |
1919 | return 0; | |
1920 | } | |
1921 | ||
a3d0c6aa | 1922 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 1923 | void __user *buffer, |
a3d0c6aa NA |
1924 | size_t *length, loff_t *ppos) |
1925 | { | |
a5516438 | 1926 | struct hstate *h = &default_hstate; |
e5ff2159 AK |
1927 | unsigned long tmp; |
1928 | ||
1929 | if (!write) | |
1930 | tmp = h->nr_overcommit_huge_pages; | |
1931 | ||
1932 | table->data = &tmp; | |
1933 | table->maxlen = sizeof(unsigned long); | |
8d65af78 | 1934 | proc_doulongvec_minmax(table, write, buffer, length, ppos); |
e5ff2159 AK |
1935 | |
1936 | if (write) { | |
1937 | spin_lock(&hugetlb_lock); | |
1938 | h->nr_overcommit_huge_pages = tmp; | |
1939 | spin_unlock(&hugetlb_lock); | |
1940 | } | |
1941 | ||
a3d0c6aa NA |
1942 | return 0; |
1943 | } | |
1944 | ||
1da177e4 LT |
1945 | #endif /* CONFIG_SYSCTL */ |
1946 | ||
e1759c21 | 1947 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 1948 | { |
a5516438 | 1949 | struct hstate *h = &default_hstate; |
e1759c21 | 1950 | seq_printf(m, |
4f98a2fe RR |
1951 | "HugePages_Total: %5lu\n" |
1952 | "HugePages_Free: %5lu\n" | |
1953 | "HugePages_Rsvd: %5lu\n" | |
1954 | "HugePages_Surp: %5lu\n" | |
1955 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
1956 | h->nr_huge_pages, |
1957 | h->free_huge_pages, | |
1958 | h->resv_huge_pages, | |
1959 | h->surplus_huge_pages, | |
1960 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
1961 | } |
1962 | ||
1963 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
1964 | { | |
a5516438 | 1965 | struct hstate *h = &default_hstate; |
1da177e4 LT |
1966 | return sprintf(buf, |
1967 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
1968 | "Node %d HugePages_Free: %5u\n" |
1969 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
1970 | nid, h->nr_huge_pages_node[nid], |
1971 | nid, h->free_huge_pages_node[nid], | |
1972 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
1973 | } |
1974 | ||
1da177e4 LT |
1975 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
1976 | unsigned long hugetlb_total_pages(void) | |
1977 | { | |
a5516438 AK |
1978 | struct hstate *h = &default_hstate; |
1979 | return h->nr_huge_pages * pages_per_huge_page(h); | |
1da177e4 | 1980 | } |
1da177e4 | 1981 | |
a5516438 | 1982 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
1983 | { |
1984 | int ret = -ENOMEM; | |
1985 | ||
1986 | spin_lock(&hugetlb_lock); | |
1987 | /* | |
1988 | * When cpuset is configured, it breaks the strict hugetlb page | |
1989 | * reservation as the accounting is done on a global variable. Such | |
1990 | * reservation is completely rubbish in the presence of cpuset because | |
1991 | * the reservation is not checked against page availability for the | |
1992 | * current cpuset. Application can still potentially OOM'ed by kernel | |
1993 | * with lack of free htlb page in cpuset that the task is in. | |
1994 | * Attempt to enforce strict accounting with cpuset is almost | |
1995 | * impossible (or too ugly) because cpuset is too fluid that | |
1996 | * task or memory node can be dynamically moved between cpusets. | |
1997 | * | |
1998 | * The change of semantics for shared hugetlb mapping with cpuset is | |
1999 | * undesirable. However, in order to preserve some of the semantics, | |
2000 | * we fall back to check against current free page availability as | |
2001 | * a best attempt and hopefully to minimize the impact of changing | |
2002 | * semantics that cpuset has. | |
2003 | */ | |
2004 | if (delta > 0) { | |
a5516438 | 2005 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2006 | goto out; |
2007 | ||
a5516438 AK |
2008 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2009 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2010 | goto out; |
2011 | } | |
2012 | } | |
2013 | ||
2014 | ret = 0; | |
2015 | if (delta < 0) | |
a5516438 | 2016 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2017 | |
2018 | out: | |
2019 | spin_unlock(&hugetlb_lock); | |
2020 | return ret; | |
2021 | } | |
2022 | ||
84afd99b AW |
2023 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2024 | { | |
2025 | struct resv_map *reservations = vma_resv_map(vma); | |
2026 | ||
2027 | /* | |
2028 | * This new VMA should share its siblings reservation map if present. | |
2029 | * The VMA will only ever have a valid reservation map pointer where | |
2030 | * it is being copied for another still existing VMA. As that VMA | |
2031 | * has a reference to the reservation map it cannot dissappear until | |
2032 | * after this open call completes. It is therefore safe to take a | |
2033 | * new reference here without additional locking. | |
2034 | */ | |
2035 | if (reservations) | |
2036 | kref_get(&reservations->refs); | |
2037 | } | |
2038 | ||
a1e78772 MG |
2039 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2040 | { | |
a5516438 | 2041 | struct hstate *h = hstate_vma(vma); |
84afd99b AW |
2042 | struct resv_map *reservations = vma_resv_map(vma); |
2043 | unsigned long reserve; | |
2044 | unsigned long start; | |
2045 | unsigned long end; | |
2046 | ||
2047 | if (reservations) { | |
a5516438 AK |
2048 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2049 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b AW |
2050 | |
2051 | reserve = (end - start) - | |
2052 | region_count(&reservations->regions, start, end); | |
2053 | ||
2054 | kref_put(&reservations->refs, resv_map_release); | |
2055 | ||
7251ff78 | 2056 | if (reserve) { |
a5516438 | 2057 | hugetlb_acct_memory(h, -reserve); |
7251ff78 AL |
2058 | hugetlb_put_quota(vma->vm_file->f_mapping, reserve); |
2059 | } | |
84afd99b | 2060 | } |
a1e78772 MG |
2061 | } |
2062 | ||
1da177e4 LT |
2063 | /* |
2064 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2065 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2066 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2067 | * this far. | |
2068 | */ | |
d0217ac0 | 2069 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2070 | { |
2071 | BUG(); | |
d0217ac0 | 2072 | return 0; |
1da177e4 LT |
2073 | } |
2074 | ||
f0f37e2f | 2075 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2076 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2077 | .open = hugetlb_vm_op_open, |
a1e78772 | 2078 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2079 | }; |
2080 | ||
1e8f889b DG |
2081 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2082 | int writable) | |
63551ae0 DG |
2083 | { |
2084 | pte_t entry; | |
2085 | ||
1e8f889b | 2086 | if (writable) { |
63551ae0 DG |
2087 | entry = |
2088 | pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); | |
2089 | } else { | |
7f2e9525 | 2090 | entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); |
63551ae0 DG |
2091 | } |
2092 | entry = pte_mkyoung(entry); | |
2093 | entry = pte_mkhuge(entry); | |
2094 | ||
2095 | return entry; | |
2096 | } | |
2097 | ||
1e8f889b DG |
2098 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2099 | unsigned long address, pte_t *ptep) | |
2100 | { | |
2101 | pte_t entry; | |
2102 | ||
7f2e9525 GS |
2103 | entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); |
2104 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { | |
4b3073e1 | 2105 | update_mmu_cache(vma, address, ptep); |
8dab5241 | 2106 | } |
1e8f889b DG |
2107 | } |
2108 | ||
2109 | ||
63551ae0 DG |
2110 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2111 | struct vm_area_struct *vma) | |
2112 | { | |
2113 | pte_t *src_pte, *dst_pte, entry; | |
2114 | struct page *ptepage; | |
1c59827d | 2115 | unsigned long addr; |
1e8f889b | 2116 | int cow; |
a5516438 AK |
2117 | struct hstate *h = hstate_vma(vma); |
2118 | unsigned long sz = huge_page_size(h); | |
1e8f889b DG |
2119 | |
2120 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2121 | |
a5516438 | 2122 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
c74df32c HD |
2123 | src_pte = huge_pte_offset(src, addr); |
2124 | if (!src_pte) | |
2125 | continue; | |
a5516438 | 2126 | dst_pte = huge_pte_alloc(dst, addr, sz); |
63551ae0 DG |
2127 | if (!dst_pte) |
2128 | goto nomem; | |
c5c99429 LW |
2129 | |
2130 | /* If the pagetables are shared don't copy or take references */ | |
2131 | if (dst_pte == src_pte) | |
2132 | continue; | |
2133 | ||
c74df32c | 2134 | spin_lock(&dst->page_table_lock); |
46478758 | 2135 | spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); |
7f2e9525 | 2136 | if (!huge_pte_none(huge_ptep_get(src_pte))) { |
1e8f889b | 2137 | if (cow) |
7f2e9525 GS |
2138 | huge_ptep_set_wrprotect(src, addr, src_pte); |
2139 | entry = huge_ptep_get(src_pte); | |
1c59827d HD |
2140 | ptepage = pte_page(entry); |
2141 | get_page(ptepage); | |
0fe6e20b | 2142 | page_dup_rmap(ptepage); |
1c59827d HD |
2143 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2144 | } | |
2145 | spin_unlock(&src->page_table_lock); | |
c74df32c | 2146 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
2147 | } |
2148 | return 0; | |
2149 | ||
2150 | nomem: | |
2151 | return -ENOMEM; | |
2152 | } | |
2153 | ||
fd6a03ed NH |
2154 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) |
2155 | { | |
2156 | swp_entry_t swp; | |
2157 | ||
2158 | if (huge_pte_none(pte) || pte_present(pte)) | |
2159 | return 0; | |
2160 | swp = pte_to_swp_entry(pte); | |
2161 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) { | |
2162 | return 1; | |
2163 | } else | |
2164 | return 0; | |
2165 | } | |
2166 | ||
502717f4 | 2167 | void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2168 | unsigned long end, struct page *ref_page) |
63551ae0 DG |
2169 | { |
2170 | struct mm_struct *mm = vma->vm_mm; | |
2171 | unsigned long address; | |
c7546f8f | 2172 | pte_t *ptep; |
63551ae0 DG |
2173 | pte_t pte; |
2174 | struct page *page; | |
fe1668ae | 2175 | struct page *tmp; |
a5516438 AK |
2176 | struct hstate *h = hstate_vma(vma); |
2177 | unsigned long sz = huge_page_size(h); | |
2178 | ||
c0a499c2 CK |
2179 | /* |
2180 | * A page gathering list, protected by per file i_mmap_lock. The | |
2181 | * lock is used to avoid list corruption from multiple unmapping | |
2182 | * of the same page since we are using page->lru. | |
2183 | */ | |
fe1668ae | 2184 | LIST_HEAD(page_list); |
63551ae0 DG |
2185 | |
2186 | WARN_ON(!is_vm_hugetlb_page(vma)); | |
a5516438 AK |
2187 | BUG_ON(start & ~huge_page_mask(h)); |
2188 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 2189 | |
cddb8a5c | 2190 | mmu_notifier_invalidate_range_start(mm, start, end); |
508034a3 | 2191 | spin_lock(&mm->page_table_lock); |
a5516438 | 2192 | for (address = start; address < end; address += sz) { |
c7546f8f | 2193 | ptep = huge_pte_offset(mm, address); |
4c887265 | 2194 | if (!ptep) |
c7546f8f DG |
2195 | continue; |
2196 | ||
39dde65c CK |
2197 | if (huge_pmd_unshare(mm, &address, ptep)) |
2198 | continue; | |
2199 | ||
04f2cbe3 MG |
2200 | /* |
2201 | * If a reference page is supplied, it is because a specific | |
2202 | * page is being unmapped, not a range. Ensure the page we | |
2203 | * are about to unmap is the actual page of interest. | |
2204 | */ | |
2205 | if (ref_page) { | |
2206 | pte = huge_ptep_get(ptep); | |
2207 | if (huge_pte_none(pte)) | |
2208 | continue; | |
2209 | page = pte_page(pte); | |
2210 | if (page != ref_page) | |
2211 | continue; | |
2212 | ||
2213 | /* | |
2214 | * Mark the VMA as having unmapped its page so that | |
2215 | * future faults in this VMA will fail rather than | |
2216 | * looking like data was lost | |
2217 | */ | |
2218 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
2219 | } | |
2220 | ||
c7546f8f | 2221 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
7f2e9525 | 2222 | if (huge_pte_none(pte)) |
63551ae0 | 2223 | continue; |
c7546f8f | 2224 | |
fd6a03ed NH |
2225 | /* |
2226 | * HWPoisoned hugepage is already unmapped and dropped reference | |
2227 | */ | |
2228 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) | |
2229 | continue; | |
2230 | ||
63551ae0 | 2231 | page = pte_page(pte); |
6649a386 KC |
2232 | if (pte_dirty(pte)) |
2233 | set_page_dirty(page); | |
fe1668ae | 2234 | list_add(&page->lru, &page_list); |
63551ae0 | 2235 | } |
1da177e4 | 2236 | spin_unlock(&mm->page_table_lock); |
508034a3 | 2237 | flush_tlb_range(vma, start, end); |
cddb8a5c | 2238 | mmu_notifier_invalidate_range_end(mm, start, end); |
fe1668ae | 2239 | list_for_each_entry_safe(page, tmp, &page_list, lru) { |
0fe6e20b | 2240 | page_remove_rmap(page); |
fe1668ae CK |
2241 | list_del(&page->lru); |
2242 | put_page(page); | |
2243 | } | |
1da177e4 | 2244 | } |
63551ae0 | 2245 | |
502717f4 | 2246 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2247 | unsigned long end, struct page *ref_page) |
502717f4 | 2248 | { |
a137e1cc AK |
2249 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); |
2250 | __unmap_hugepage_range(vma, start, end, ref_page); | |
2251 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); | |
502717f4 CK |
2252 | } |
2253 | ||
04f2cbe3 MG |
2254 | /* |
2255 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
2256 | * mappping it owns the reserve page for. The intention is to unmap the page | |
2257 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
2258 | * same region. | |
2259 | */ | |
2a4b3ded HH |
2260 | static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
2261 | struct page *page, unsigned long address) | |
04f2cbe3 | 2262 | { |
7526674d | 2263 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
2264 | struct vm_area_struct *iter_vma; |
2265 | struct address_space *mapping; | |
2266 | struct prio_tree_iter iter; | |
2267 | pgoff_t pgoff; | |
2268 | ||
2269 | /* | |
2270 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
2271 | * from page cache lookup which is in HPAGE_SIZE units. | |
2272 | */ | |
7526674d | 2273 | address = address & huge_page_mask(h); |
04f2cbe3 MG |
2274 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) |
2275 | + (vma->vm_pgoff >> PAGE_SHIFT); | |
2276 | mapping = (struct address_space *)page_private(page); | |
2277 | ||
4eb2b1dc MG |
2278 | /* |
2279 | * Take the mapping lock for the duration of the table walk. As | |
2280 | * this mapping should be shared between all the VMAs, | |
2281 | * __unmap_hugepage_range() is called as the lock is already held | |
2282 | */ | |
2283 | spin_lock(&mapping->i_mmap_lock); | |
04f2cbe3 MG |
2284 | vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { |
2285 | /* Do not unmap the current VMA */ | |
2286 | if (iter_vma == vma) | |
2287 | continue; | |
2288 | ||
2289 | /* | |
2290 | * Unmap the page from other VMAs without their own reserves. | |
2291 | * They get marked to be SIGKILLed if they fault in these | |
2292 | * areas. This is because a future no-page fault on this VMA | |
2293 | * could insert a zeroed page instead of the data existing | |
2294 | * from the time of fork. This would look like data corruption | |
2295 | */ | |
2296 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
4eb2b1dc | 2297 | __unmap_hugepage_range(iter_vma, |
7526674d | 2298 | address, address + huge_page_size(h), |
04f2cbe3 MG |
2299 | page); |
2300 | } | |
4eb2b1dc | 2301 | spin_unlock(&mapping->i_mmap_lock); |
04f2cbe3 MG |
2302 | |
2303 | return 1; | |
2304 | } | |
2305 | ||
0fe6e20b NH |
2306 | /* |
2307 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
2308 | */ | |
1e8f889b | 2309 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 MG |
2310 | unsigned long address, pte_t *ptep, pte_t pte, |
2311 | struct page *pagecache_page) | |
1e8f889b | 2312 | { |
a5516438 | 2313 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 2314 | struct page *old_page, *new_page; |
79ac6ba4 | 2315 | int avoidcopy; |
04f2cbe3 | 2316 | int outside_reserve = 0; |
1e8f889b DG |
2317 | |
2318 | old_page = pte_page(pte); | |
2319 | ||
04f2cbe3 | 2320 | retry_avoidcopy: |
1e8f889b DG |
2321 | /* If no-one else is actually using this page, avoid the copy |
2322 | * and just make the page writable */ | |
0fe6e20b | 2323 | avoidcopy = (page_mapcount(old_page) == 1); |
1e8f889b | 2324 | if (avoidcopy) { |
0fe6e20b NH |
2325 | if (!trylock_page(old_page)) |
2326 | if (PageAnon(old_page)) | |
2327 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 2328 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 2329 | return 0; |
1e8f889b DG |
2330 | } |
2331 | ||
04f2cbe3 MG |
2332 | /* |
2333 | * If the process that created a MAP_PRIVATE mapping is about to | |
2334 | * perform a COW due to a shared page count, attempt to satisfy | |
2335 | * the allocation without using the existing reserves. The pagecache | |
2336 | * page is used to determine if the reserve at this address was | |
2337 | * consumed or not. If reserves were used, a partial faulted mapping | |
2338 | * at the time of fork() could consume its reserves on COW instead | |
2339 | * of the full address range. | |
2340 | */ | |
f83a275d | 2341 | if (!(vma->vm_flags & VM_MAYSHARE) && |
04f2cbe3 MG |
2342 | is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
2343 | old_page != pagecache_page) | |
2344 | outside_reserve = 1; | |
2345 | ||
1e8f889b | 2346 | page_cache_get(old_page); |
b76c8cfb LW |
2347 | |
2348 | /* Drop page_table_lock as buddy allocator may be called */ | |
2349 | spin_unlock(&mm->page_table_lock); | |
04f2cbe3 | 2350 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 2351 | |
2fc39cec | 2352 | if (IS_ERR(new_page)) { |
1e8f889b | 2353 | page_cache_release(old_page); |
04f2cbe3 MG |
2354 | |
2355 | /* | |
2356 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
2357 | * it is due to references held by a child and an insufficient | |
2358 | * huge page pool. To guarantee the original mappers | |
2359 | * reliability, unmap the page from child processes. The child | |
2360 | * may get SIGKILLed if it later faults. | |
2361 | */ | |
2362 | if (outside_reserve) { | |
2363 | BUG_ON(huge_pte_none(pte)); | |
2364 | if (unmap_ref_private(mm, vma, old_page, address)) { | |
2365 | BUG_ON(page_count(old_page) != 1); | |
2366 | BUG_ON(huge_pte_none(pte)); | |
b76c8cfb | 2367 | spin_lock(&mm->page_table_lock); |
04f2cbe3 MG |
2368 | goto retry_avoidcopy; |
2369 | } | |
2370 | WARN_ON_ONCE(1); | |
2371 | } | |
2372 | ||
b76c8cfb LW |
2373 | /* Caller expects lock to be held */ |
2374 | spin_lock(&mm->page_table_lock); | |
2fc39cec | 2375 | return -PTR_ERR(new_page); |
1e8f889b DG |
2376 | } |
2377 | ||
0fe6e20b NH |
2378 | /* |
2379 | * When the original hugepage is shared one, it does not have | |
2380 | * anon_vma prepared. | |
2381 | */ | |
2382 | if (unlikely(anon_vma_prepare(vma))) | |
2383 | return VM_FAULT_OOM; | |
2384 | ||
9de455b2 | 2385 | copy_huge_page(new_page, old_page, address, vma); |
0ed361de | 2386 | __SetPageUptodate(new_page); |
1e8f889b | 2387 | |
b76c8cfb LW |
2388 | /* |
2389 | * Retake the page_table_lock to check for racing updates | |
2390 | * before the page tables are altered | |
2391 | */ | |
2392 | spin_lock(&mm->page_table_lock); | |
a5516438 | 2393 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
7f2e9525 | 2394 | if (likely(pte_same(huge_ptep_get(ptep), pte))) { |
1e8f889b | 2395 | /* Break COW */ |
8fe627ec | 2396 | huge_ptep_clear_flush(vma, address, ptep); |
1e8f889b DG |
2397 | set_huge_pte_at(mm, address, ptep, |
2398 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b NH |
2399 | page_remove_rmap(old_page); |
2400 | hugepage_add_anon_rmap(new_page, vma, address); | |
1e8f889b DG |
2401 | /* Make the old page be freed below */ |
2402 | new_page = old_page; | |
2403 | } | |
2404 | page_cache_release(new_page); | |
2405 | page_cache_release(old_page); | |
83c54070 | 2406 | return 0; |
1e8f889b DG |
2407 | } |
2408 | ||
04f2cbe3 | 2409 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
2410 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
2411 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
2412 | { |
2413 | struct address_space *mapping; | |
e7c4b0bf | 2414 | pgoff_t idx; |
04f2cbe3 MG |
2415 | |
2416 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 2417 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
2418 | |
2419 | return find_lock_page(mapping, idx); | |
2420 | } | |
2421 | ||
3ae77f43 HD |
2422 | /* |
2423 | * Return whether there is a pagecache page to back given address within VMA. | |
2424 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
2425 | */ | |
2426 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
2427 | struct vm_area_struct *vma, unsigned long address) |
2428 | { | |
2429 | struct address_space *mapping; | |
2430 | pgoff_t idx; | |
2431 | struct page *page; | |
2432 | ||
2433 | mapping = vma->vm_file->f_mapping; | |
2434 | idx = vma_hugecache_offset(h, vma, address); | |
2435 | ||
2436 | page = find_get_page(mapping, idx); | |
2437 | if (page) | |
2438 | put_page(page); | |
2439 | return page != NULL; | |
2440 | } | |
2441 | ||
a1ed3dda | 2442 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2443 | unsigned long address, pte_t *ptep, unsigned int flags) |
ac9b9c66 | 2444 | { |
a5516438 | 2445 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 2446 | int ret = VM_FAULT_SIGBUS; |
e7c4b0bf | 2447 | pgoff_t idx; |
4c887265 | 2448 | unsigned long size; |
4c887265 AL |
2449 | struct page *page; |
2450 | struct address_space *mapping; | |
1e8f889b | 2451 | pte_t new_pte; |
4c887265 | 2452 | |
04f2cbe3 MG |
2453 | /* |
2454 | * Currently, we are forced to kill the process in the event the | |
2455 | * original mapper has unmapped pages from the child due to a failed | |
2456 | * COW. Warn that such a situation has occured as it may not be obvious | |
2457 | */ | |
2458 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
2459 | printk(KERN_WARNING | |
2460 | "PID %d killed due to inadequate hugepage pool\n", | |
2461 | current->pid); | |
2462 | return ret; | |
2463 | } | |
2464 | ||
4c887265 | 2465 | mapping = vma->vm_file->f_mapping; |
a5516438 | 2466 | idx = vma_hugecache_offset(h, vma, address); |
4c887265 AL |
2467 | |
2468 | /* | |
2469 | * Use page lock to guard against racing truncation | |
2470 | * before we get page_table_lock. | |
2471 | */ | |
6bda666a CL |
2472 | retry: |
2473 | page = find_lock_page(mapping, idx); | |
2474 | if (!page) { | |
a5516438 | 2475 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
2476 | if (idx >= size) |
2477 | goto out; | |
04f2cbe3 | 2478 | page = alloc_huge_page(vma, address, 0); |
2fc39cec AL |
2479 | if (IS_ERR(page)) { |
2480 | ret = -PTR_ERR(page); | |
6bda666a CL |
2481 | goto out; |
2482 | } | |
a5516438 | 2483 | clear_huge_page(page, address, huge_page_size(h)); |
0ed361de | 2484 | __SetPageUptodate(page); |
ac9b9c66 | 2485 | |
f83a275d | 2486 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 2487 | int err; |
45c682a6 | 2488 | struct inode *inode = mapping->host; |
6bda666a CL |
2489 | |
2490 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
2491 | if (err) { | |
2492 | put_page(page); | |
6bda666a CL |
2493 | if (err == -EEXIST) |
2494 | goto retry; | |
2495 | goto out; | |
2496 | } | |
45c682a6 KC |
2497 | |
2498 | spin_lock(&inode->i_lock); | |
a5516438 | 2499 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 2500 | spin_unlock(&inode->i_lock); |
0fe6e20b | 2501 | page_dup_rmap(page); |
23be7468 | 2502 | } else { |
6bda666a | 2503 | lock_page(page); |
0fe6e20b NH |
2504 | if (unlikely(anon_vma_prepare(vma))) { |
2505 | ret = VM_FAULT_OOM; | |
2506 | goto backout_unlocked; | |
2507 | } | |
2508 | hugepage_add_new_anon_rmap(page, vma, address); | |
23be7468 | 2509 | } |
0fe6e20b NH |
2510 | } else { |
2511 | page_dup_rmap(page); | |
6bda666a | 2512 | } |
1e8f889b | 2513 | |
fd6a03ed NH |
2514 | /* |
2515 | * Since memory error handler replaces pte into hwpoison swap entry | |
2516 | * at the time of error handling, a process which reserved but not have | |
2517 | * the mapping to the error hugepage does not have hwpoison swap entry. | |
2518 | * So we need to block accesses from such a process by checking | |
2519 | * PG_hwpoison bit here. | |
2520 | */ | |
2521 | if (unlikely(PageHWPoison(page))) { | |
2522 | ret = VM_FAULT_HWPOISON; | |
2523 | goto backout_unlocked; | |
2524 | } | |
2525 | ||
57303d80 AW |
2526 | /* |
2527 | * If we are going to COW a private mapping later, we examine the | |
2528 | * pending reservations for this page now. This will ensure that | |
2529 | * any allocations necessary to record that reservation occur outside | |
2530 | * the spinlock. | |
2531 | */ | |
788c7df4 | 2532 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
2b26736c AW |
2533 | if (vma_needs_reservation(h, vma, address) < 0) { |
2534 | ret = VM_FAULT_OOM; | |
2535 | goto backout_unlocked; | |
2536 | } | |
57303d80 | 2537 | |
ac9b9c66 | 2538 | spin_lock(&mm->page_table_lock); |
a5516438 | 2539 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
2540 | if (idx >= size) |
2541 | goto backout; | |
2542 | ||
83c54070 | 2543 | ret = 0; |
7f2e9525 | 2544 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
2545 | goto backout; |
2546 | ||
1e8f889b DG |
2547 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
2548 | && (vma->vm_flags & VM_SHARED))); | |
2549 | set_huge_pte_at(mm, address, ptep, new_pte); | |
2550 | ||
788c7df4 | 2551 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 2552 | /* Optimization, do the COW without a second fault */ |
04f2cbe3 | 2553 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); |
1e8f889b DG |
2554 | } |
2555 | ||
ac9b9c66 | 2556 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
2557 | unlock_page(page); |
2558 | out: | |
ac9b9c66 | 2559 | return ret; |
4c887265 AL |
2560 | |
2561 | backout: | |
2562 | spin_unlock(&mm->page_table_lock); | |
2b26736c | 2563 | backout_unlocked: |
4c887265 AL |
2564 | unlock_page(page); |
2565 | put_page(page); | |
2566 | goto out; | |
ac9b9c66 HD |
2567 | } |
2568 | ||
86e5216f | 2569 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2570 | unsigned long address, unsigned int flags) |
86e5216f AL |
2571 | { |
2572 | pte_t *ptep; | |
2573 | pte_t entry; | |
1e8f889b | 2574 | int ret; |
0fe6e20b | 2575 | struct page *page = NULL; |
57303d80 | 2576 | struct page *pagecache_page = NULL; |
3935baa9 | 2577 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
a5516438 | 2578 | struct hstate *h = hstate_vma(vma); |
86e5216f | 2579 | |
fd6a03ed NH |
2580 | ptep = huge_pte_offset(mm, address); |
2581 | if (ptep) { | |
2582 | entry = huge_ptep_get(ptep); | |
2583 | if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
2584 | return VM_FAULT_HWPOISON; | |
2585 | } | |
2586 | ||
a5516438 | 2587 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
2588 | if (!ptep) |
2589 | return VM_FAULT_OOM; | |
2590 | ||
3935baa9 DG |
2591 | /* |
2592 | * Serialize hugepage allocation and instantiation, so that we don't | |
2593 | * get spurious allocation failures if two CPUs race to instantiate | |
2594 | * the same page in the page cache. | |
2595 | */ | |
2596 | mutex_lock(&hugetlb_instantiation_mutex); | |
7f2e9525 GS |
2597 | entry = huge_ptep_get(ptep); |
2598 | if (huge_pte_none(entry)) { | |
788c7df4 | 2599 | ret = hugetlb_no_page(mm, vma, address, ptep, flags); |
b4d1d99f | 2600 | goto out_mutex; |
3935baa9 | 2601 | } |
86e5216f | 2602 | |
83c54070 | 2603 | ret = 0; |
1e8f889b | 2604 | |
57303d80 AW |
2605 | /* |
2606 | * If we are going to COW the mapping later, we examine the pending | |
2607 | * reservations for this page now. This will ensure that any | |
2608 | * allocations necessary to record that reservation occur outside the | |
2609 | * spinlock. For private mappings, we also lookup the pagecache | |
2610 | * page now as it is used to determine if a reservation has been | |
2611 | * consumed. | |
2612 | */ | |
788c7df4 | 2613 | if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { |
2b26736c AW |
2614 | if (vma_needs_reservation(h, vma, address) < 0) { |
2615 | ret = VM_FAULT_OOM; | |
b4d1d99f | 2616 | goto out_mutex; |
2b26736c | 2617 | } |
57303d80 | 2618 | |
f83a275d | 2619 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
2620 | pagecache_page = hugetlbfs_pagecache_page(h, |
2621 | vma, address); | |
2622 | } | |
2623 | ||
0fe6e20b NH |
2624 | if (!pagecache_page) { |
2625 | page = pte_page(entry); | |
2626 | lock_page(page); | |
2627 | } | |
2628 | ||
1e8f889b DG |
2629 | spin_lock(&mm->page_table_lock); |
2630 | /* Check for a racing update before calling hugetlb_cow */ | |
b4d1d99f DG |
2631 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
2632 | goto out_page_table_lock; | |
2633 | ||
2634 | ||
788c7df4 | 2635 | if (flags & FAULT_FLAG_WRITE) { |
b4d1d99f | 2636 | if (!pte_write(entry)) { |
57303d80 AW |
2637 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
2638 | pagecache_page); | |
b4d1d99f DG |
2639 | goto out_page_table_lock; |
2640 | } | |
2641 | entry = pte_mkdirty(entry); | |
2642 | } | |
2643 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
2644 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
2645 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 2646 | update_mmu_cache(vma, address, ptep); |
b4d1d99f DG |
2647 | |
2648 | out_page_table_lock: | |
1e8f889b | 2649 | spin_unlock(&mm->page_table_lock); |
57303d80 AW |
2650 | |
2651 | if (pagecache_page) { | |
2652 | unlock_page(pagecache_page); | |
2653 | put_page(pagecache_page); | |
0fe6e20b NH |
2654 | } else { |
2655 | unlock_page(page); | |
57303d80 AW |
2656 | } |
2657 | ||
b4d1d99f | 2658 | out_mutex: |
3935baa9 | 2659 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
2660 | |
2661 | return ret; | |
86e5216f AL |
2662 | } |
2663 | ||
ceb86879 AK |
2664 | /* Can be overriden by architectures */ |
2665 | __attribute__((weak)) struct page * | |
2666 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
2667 | pud_t *pud, int write) | |
2668 | { | |
2669 | BUG(); | |
2670 | return NULL; | |
2671 | } | |
2672 | ||
63551ae0 DG |
2673 | int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
2674 | struct page **pages, struct vm_area_struct **vmas, | |
5b23dbe8 | 2675 | unsigned long *position, int *length, int i, |
2a15efc9 | 2676 | unsigned int flags) |
63551ae0 | 2677 | { |
d5d4b0aa CK |
2678 | unsigned long pfn_offset; |
2679 | unsigned long vaddr = *position; | |
63551ae0 | 2680 | int remainder = *length; |
a5516438 | 2681 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 2682 | |
1c59827d | 2683 | spin_lock(&mm->page_table_lock); |
63551ae0 | 2684 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 2685 | pte_t *pte; |
2a15efc9 | 2686 | int absent; |
4c887265 | 2687 | struct page *page; |
63551ae0 | 2688 | |
4c887265 AL |
2689 | /* |
2690 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 2691 | * each hugepage. We have to make sure we get the |
4c887265 AL |
2692 | * first, for the page indexing below to work. |
2693 | */ | |
a5516438 | 2694 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
2a15efc9 HD |
2695 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
2696 | ||
2697 | /* | |
2698 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
2699 | * an error where there's an empty slot with no huge pagecache |
2700 | * to back it. This way, we avoid allocating a hugepage, and | |
2701 | * the sparse dumpfile avoids allocating disk blocks, but its | |
2702 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 2703 | */ |
3ae77f43 HD |
2704 | if (absent && (flags & FOLL_DUMP) && |
2705 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
2a15efc9 HD |
2706 | remainder = 0; |
2707 | break; | |
2708 | } | |
63551ae0 | 2709 | |
2a15efc9 HD |
2710 | if (absent || |
2711 | ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 2712 | int ret; |
63551ae0 | 2713 | |
4c887265 | 2714 | spin_unlock(&mm->page_table_lock); |
2a15efc9 HD |
2715 | ret = hugetlb_fault(mm, vma, vaddr, |
2716 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
4c887265 | 2717 | spin_lock(&mm->page_table_lock); |
a89182c7 | 2718 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 2719 | continue; |
63551ae0 | 2720 | |
4c887265 | 2721 | remainder = 0; |
4c887265 AL |
2722 | break; |
2723 | } | |
2724 | ||
a5516438 | 2725 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 2726 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 2727 | same_page: |
d6692183 | 2728 | if (pages) { |
2a15efc9 | 2729 | pages[i] = mem_map_offset(page, pfn_offset); |
4b2e38ad | 2730 | get_page(pages[i]); |
d6692183 | 2731 | } |
63551ae0 DG |
2732 | |
2733 | if (vmas) | |
2734 | vmas[i] = vma; | |
2735 | ||
2736 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 2737 | ++pfn_offset; |
63551ae0 DG |
2738 | --remainder; |
2739 | ++i; | |
d5d4b0aa | 2740 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 2741 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
2742 | /* |
2743 | * We use pfn_offset to avoid touching the pageframes | |
2744 | * of this compound page. | |
2745 | */ | |
2746 | goto same_page; | |
2747 | } | |
63551ae0 | 2748 | } |
1c59827d | 2749 | spin_unlock(&mm->page_table_lock); |
63551ae0 DG |
2750 | *length = remainder; |
2751 | *position = vaddr; | |
2752 | ||
2a15efc9 | 2753 | return i ? i : -EFAULT; |
63551ae0 | 2754 | } |
8f860591 ZY |
2755 | |
2756 | void hugetlb_change_protection(struct vm_area_struct *vma, | |
2757 | unsigned long address, unsigned long end, pgprot_t newprot) | |
2758 | { | |
2759 | struct mm_struct *mm = vma->vm_mm; | |
2760 | unsigned long start = address; | |
2761 | pte_t *ptep; | |
2762 | pte_t pte; | |
a5516438 | 2763 | struct hstate *h = hstate_vma(vma); |
8f860591 ZY |
2764 | |
2765 | BUG_ON(address >= end); | |
2766 | flush_cache_range(vma, address, end); | |
2767 | ||
39dde65c | 2768 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 | 2769 | spin_lock(&mm->page_table_lock); |
a5516438 | 2770 | for (; address < end; address += huge_page_size(h)) { |
8f860591 ZY |
2771 | ptep = huge_pte_offset(mm, address); |
2772 | if (!ptep) | |
2773 | continue; | |
39dde65c CK |
2774 | if (huge_pmd_unshare(mm, &address, ptep)) |
2775 | continue; | |
7f2e9525 | 2776 | if (!huge_pte_none(huge_ptep_get(ptep))) { |
8f860591 ZY |
2777 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
2778 | pte = pte_mkhuge(pte_modify(pte, newprot)); | |
2779 | set_huge_pte_at(mm, address, ptep, pte); | |
8f860591 ZY |
2780 | } |
2781 | } | |
2782 | spin_unlock(&mm->page_table_lock); | |
39dde65c | 2783 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
2784 | |
2785 | flush_tlb_range(vma, start, end); | |
2786 | } | |
2787 | ||
a1e78772 MG |
2788 | int hugetlb_reserve_pages(struct inode *inode, |
2789 | long from, long to, | |
5a6fe125 MG |
2790 | struct vm_area_struct *vma, |
2791 | int acctflag) | |
e4e574b7 | 2792 | { |
17c9d12e | 2793 | long ret, chg; |
a5516438 | 2794 | struct hstate *h = hstate_inode(inode); |
e4e574b7 | 2795 | |
17c9d12e MG |
2796 | /* |
2797 | * Only apply hugepage reservation if asked. At fault time, an | |
2798 | * attempt will be made for VM_NORESERVE to allocate a page | |
2799 | * and filesystem quota without using reserves | |
2800 | */ | |
2801 | if (acctflag & VM_NORESERVE) | |
2802 | return 0; | |
2803 | ||
a1e78772 MG |
2804 | /* |
2805 | * Shared mappings base their reservation on the number of pages that | |
2806 | * are already allocated on behalf of the file. Private mappings need | |
2807 | * to reserve the full area even if read-only as mprotect() may be | |
2808 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
2809 | */ | |
f83a275d | 2810 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
a1e78772 | 2811 | chg = region_chg(&inode->i_mapping->private_list, from, to); |
17c9d12e MG |
2812 | else { |
2813 | struct resv_map *resv_map = resv_map_alloc(); | |
2814 | if (!resv_map) | |
2815 | return -ENOMEM; | |
2816 | ||
a1e78772 | 2817 | chg = to - from; |
84afd99b | 2818 | |
17c9d12e MG |
2819 | set_vma_resv_map(vma, resv_map); |
2820 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
2821 | } | |
2822 | ||
e4e574b7 AL |
2823 | if (chg < 0) |
2824 | return chg; | |
8a630112 | 2825 | |
17c9d12e | 2826 | /* There must be enough filesystem quota for the mapping */ |
90d8b7e6 AL |
2827 | if (hugetlb_get_quota(inode->i_mapping, chg)) |
2828 | return -ENOSPC; | |
5a6fe125 MG |
2829 | |
2830 | /* | |
17c9d12e MG |
2831 | * Check enough hugepages are available for the reservation. |
2832 | * Hand back the quota if there are not | |
5a6fe125 | 2833 | */ |
a5516438 | 2834 | ret = hugetlb_acct_memory(h, chg); |
68842c9b KC |
2835 | if (ret < 0) { |
2836 | hugetlb_put_quota(inode->i_mapping, chg); | |
a43a8c39 | 2837 | return ret; |
68842c9b | 2838 | } |
17c9d12e MG |
2839 | |
2840 | /* | |
2841 | * Account for the reservations made. Shared mappings record regions | |
2842 | * that have reservations as they are shared by multiple VMAs. | |
2843 | * When the last VMA disappears, the region map says how much | |
2844 | * the reservation was and the page cache tells how much of | |
2845 | * the reservation was consumed. Private mappings are per-VMA and | |
2846 | * only the consumed reservations are tracked. When the VMA | |
2847 | * disappears, the original reservation is the VMA size and the | |
2848 | * consumed reservations are stored in the map. Hence, nothing | |
2849 | * else has to be done for private mappings here | |
2850 | */ | |
f83a275d | 2851 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
a1e78772 | 2852 | region_add(&inode->i_mapping->private_list, from, to); |
a43a8c39 CK |
2853 | return 0; |
2854 | } | |
2855 | ||
2856 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
2857 | { | |
a5516438 | 2858 | struct hstate *h = hstate_inode(inode); |
a43a8c39 | 2859 | long chg = region_truncate(&inode->i_mapping->private_list, offset); |
45c682a6 KC |
2860 | |
2861 | spin_lock(&inode->i_lock); | |
e4c6f8be | 2862 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
2863 | spin_unlock(&inode->i_lock); |
2864 | ||
90d8b7e6 | 2865 | hugetlb_put_quota(inode->i_mapping, (chg - freed)); |
a5516438 | 2866 | hugetlb_acct_memory(h, -(chg - freed)); |
a43a8c39 | 2867 | } |
93f70f90 NH |
2868 | |
2869 | /* | |
2870 | * This function is called from memory failure code. | |
2871 | * Assume the caller holds page lock of the head page. | |
2872 | */ | |
2873 | void __isolate_hwpoisoned_huge_page(struct page *hpage) | |
2874 | { | |
2875 | struct hstate *h = page_hstate(hpage); | |
2876 | int nid = page_to_nid(hpage); | |
2877 | ||
2878 | spin_lock(&hugetlb_lock); | |
2879 | list_del(&hpage->lru); | |
2880 | h->free_huge_pages--; | |
2881 | h->free_huge_pages_node[nid]--; | |
2882 | spin_unlock(&hugetlb_lock); | |
2883 | } |