]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Generic hugetlb support. | |
6d49e352 | 3 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 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> |
3b32123d | 16 | #include <linux/compiler.h> |
aea47ff3 | 17 | #include <linux/cpuset.h> |
3935baa9 | 18 | #include <linux/mutex.h> |
aa888a74 | 19 | #include <linux/bootmem.h> |
a3437870 | 20 | #include <linux/sysfs.h> |
5a0e3ad6 | 21 | #include <linux/slab.h> |
0fe6e20b | 22 | #include <linux/rmap.h> |
fd6a03ed NH |
23 | #include <linux/swap.h> |
24 | #include <linux/swapops.h> | |
c8721bbb | 25 | #include <linux/page-isolation.h> |
8382d914 | 26 | #include <linux/jhash.h> |
d6606683 | 27 | |
63551ae0 DG |
28 | #include <asm/page.h> |
29 | #include <asm/pgtable.h> | |
24669e58 | 30 | #include <asm/tlb.h> |
63551ae0 | 31 | |
24669e58 | 32 | #include <linux/io.h> |
63551ae0 | 33 | #include <linux/hugetlb.h> |
9dd540e2 | 34 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 35 | #include <linux/node.h> |
7835e98b | 36 | #include "internal.h" |
1da177e4 | 37 | |
753162cd | 38 | int hugepages_treat_as_movable; |
a5516438 | 39 | |
c3f38a38 | 40 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
41 | unsigned int default_hstate_idx; |
42 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
641844f5 NH |
43 | /* |
44 | * Minimum page order among possible hugepage sizes, set to a proper value | |
45 | * at boot time. | |
46 | */ | |
47 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 48 | |
53ba51d2 JT |
49 | __initdata LIST_HEAD(huge_boot_pages); |
50 | ||
e5ff2159 AK |
51 | /* for command line parsing */ |
52 | static struct hstate * __initdata parsed_hstate; | |
53 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 54 | static unsigned long __initdata default_hstate_size; |
e5ff2159 | 55 | |
3935baa9 | 56 | /* |
31caf665 NH |
57 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
58 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 59 | */ |
c3f38a38 | 60 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 61 | |
8382d914 DB |
62 | /* |
63 | * Serializes faults on the same logical page. This is used to | |
64 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
65 | */ | |
66 | static int num_fault_mutexes; | |
67 | static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp; | |
68 | ||
7ca02d0a MK |
69 | /* Forward declaration */ |
70 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
71 | ||
90481622 DG |
72 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
73 | { | |
74 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
75 | ||
76 | spin_unlock(&spool->lock); | |
77 | ||
78 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
79 | * remain, give up any reservations mased on minimum size and |
80 | * free the subpool */ | |
81 | if (free) { | |
82 | if (spool->min_hpages != -1) | |
83 | hugetlb_acct_memory(spool->hstate, | |
84 | -spool->min_hpages); | |
90481622 | 85 | kfree(spool); |
7ca02d0a | 86 | } |
90481622 DG |
87 | } |
88 | ||
7ca02d0a MK |
89 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
90 | long min_hpages) | |
90481622 DG |
91 | { |
92 | struct hugepage_subpool *spool; | |
93 | ||
c6a91820 | 94 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
95 | if (!spool) |
96 | return NULL; | |
97 | ||
98 | spin_lock_init(&spool->lock); | |
99 | spool->count = 1; | |
7ca02d0a MK |
100 | spool->max_hpages = max_hpages; |
101 | spool->hstate = h; | |
102 | spool->min_hpages = min_hpages; | |
103 | ||
104 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
105 | kfree(spool); | |
106 | return NULL; | |
107 | } | |
108 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
109 | |
110 | return spool; | |
111 | } | |
112 | ||
113 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
114 | { | |
115 | spin_lock(&spool->lock); | |
116 | BUG_ON(!spool->count); | |
117 | spool->count--; | |
118 | unlock_or_release_subpool(spool); | |
119 | } | |
120 | ||
1c5ecae3 MK |
121 | /* |
122 | * Subpool accounting for allocating and reserving pages. | |
123 | * Return -ENOMEM if there are not enough resources to satisfy the | |
124 | * the request. Otherwise, return the number of pages by which the | |
125 | * global pools must be adjusted (upward). The returned value may | |
126 | * only be different than the passed value (delta) in the case where | |
127 | * a subpool minimum size must be manitained. | |
128 | */ | |
129 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
130 | long delta) |
131 | { | |
1c5ecae3 | 132 | long ret = delta; |
90481622 DG |
133 | |
134 | if (!spool) | |
1c5ecae3 | 135 | return ret; |
90481622 DG |
136 | |
137 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
138 | |
139 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
140 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
141 | spool->used_hpages += delta; | |
142 | else { | |
143 | ret = -ENOMEM; | |
144 | goto unlock_ret; | |
145 | } | |
90481622 | 146 | } |
90481622 | 147 | |
1c5ecae3 MK |
148 | if (spool->min_hpages != -1) { /* minimum size accounting */ |
149 | if (delta > spool->rsv_hpages) { | |
150 | /* | |
151 | * Asking for more reserves than those already taken on | |
152 | * behalf of subpool. Return difference. | |
153 | */ | |
154 | ret = delta - spool->rsv_hpages; | |
155 | spool->rsv_hpages = 0; | |
156 | } else { | |
157 | ret = 0; /* reserves already accounted for */ | |
158 | spool->rsv_hpages -= delta; | |
159 | } | |
160 | } | |
161 | ||
162 | unlock_ret: | |
163 | spin_unlock(&spool->lock); | |
90481622 DG |
164 | return ret; |
165 | } | |
166 | ||
1c5ecae3 MK |
167 | /* |
168 | * Subpool accounting for freeing and unreserving pages. | |
169 | * Return the number of global page reservations that must be dropped. | |
170 | * The return value may only be different than the passed value (delta) | |
171 | * in the case where a subpool minimum size must be maintained. | |
172 | */ | |
173 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
174 | long delta) |
175 | { | |
1c5ecae3 MK |
176 | long ret = delta; |
177 | ||
90481622 | 178 | if (!spool) |
1c5ecae3 | 179 | return delta; |
90481622 DG |
180 | |
181 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
182 | |
183 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
184 | spool->used_hpages -= delta; | |
185 | ||
186 | if (spool->min_hpages != -1) { /* minimum size accounting */ | |
187 | if (spool->rsv_hpages + delta <= spool->min_hpages) | |
188 | ret = 0; | |
189 | else | |
190 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
191 | ||
192 | spool->rsv_hpages += delta; | |
193 | if (spool->rsv_hpages > spool->min_hpages) | |
194 | spool->rsv_hpages = spool->min_hpages; | |
195 | } | |
196 | ||
197 | /* | |
198 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
199 | * quota reference, free it now. | |
200 | */ | |
90481622 | 201 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
202 | |
203 | return ret; | |
90481622 DG |
204 | } |
205 | ||
206 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
207 | { | |
208 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
209 | } | |
210 | ||
211 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
212 | { | |
496ad9aa | 213 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
214 | } |
215 | ||
96822904 AW |
216 | /* |
217 | * Region tracking -- allows tracking of reservations and instantiated pages | |
218 | * across the pages in a mapping. | |
84afd99b | 219 | * |
1dd308a7 MK |
220 | * The region data structures are embedded into a resv_map and protected |
221 | * by a resv_map's lock. The set of regions within the resv_map represent | |
222 | * reservations for huge pages, or huge pages that have already been | |
223 | * instantiated within the map. The from and to elements are huge page | |
224 | * indicies into the associated mapping. from indicates the starting index | |
225 | * of the region. to represents the first index past the end of the region. | |
226 | * | |
227 | * For example, a file region structure with from == 0 and to == 4 represents | |
228 | * four huge pages in a mapping. It is important to note that the to element | |
229 | * represents the first element past the end of the region. This is used in | |
230 | * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. | |
231 | * | |
232 | * Interval notation of the form [from, to) will be used to indicate that | |
233 | * the endpoint from is inclusive and to is exclusive. | |
96822904 AW |
234 | */ |
235 | struct file_region { | |
236 | struct list_head link; | |
237 | long from; | |
238 | long to; | |
239 | }; | |
240 | ||
1dd308a7 MK |
241 | /* |
242 | * Add the huge page range represented by [f, t) to the reserve | |
243 | * map. Existing regions will be expanded to accommodate the | |
244 | * specified range. We know only existing regions need to be | |
245 | * expanded, because region_add is only called after region_chg | |
246 | * with the same range. If a new file_region structure must | |
247 | * be allocated, it is done in region_chg. | |
248 | */ | |
1406ec9b | 249 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 250 | { |
1406ec9b | 251 | struct list_head *head = &resv->regions; |
96822904 AW |
252 | struct file_region *rg, *nrg, *trg; |
253 | ||
7b24d861 | 254 | spin_lock(&resv->lock); |
96822904 AW |
255 | /* Locate the region we are either in or before. */ |
256 | list_for_each_entry(rg, head, link) | |
257 | if (f <= rg->to) | |
258 | break; | |
259 | ||
260 | /* Round our left edge to the current segment if it encloses us. */ | |
261 | if (f > rg->from) | |
262 | f = rg->from; | |
263 | ||
264 | /* Check for and consume any regions we now overlap with. */ | |
265 | nrg = rg; | |
266 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
267 | if (&rg->link == head) | |
268 | break; | |
269 | if (rg->from > t) | |
270 | break; | |
271 | ||
272 | /* If this area reaches higher then extend our area to | |
273 | * include it completely. If this is not the first area | |
274 | * which we intend to reuse, free it. */ | |
275 | if (rg->to > t) | |
276 | t = rg->to; | |
277 | if (rg != nrg) { | |
278 | list_del(&rg->link); | |
279 | kfree(rg); | |
280 | } | |
281 | } | |
282 | nrg->from = f; | |
283 | nrg->to = t; | |
7b24d861 | 284 | spin_unlock(&resv->lock); |
96822904 AW |
285 | return 0; |
286 | } | |
287 | ||
1dd308a7 MK |
288 | /* |
289 | * Examine the existing reserve map and determine how many | |
290 | * huge pages in the specified range [f, t) are NOT currently | |
291 | * represented. This routine is called before a subsequent | |
292 | * call to region_add that will actually modify the reserve | |
293 | * map to add the specified range [f, t). region_chg does | |
294 | * not change the number of huge pages represented by the | |
295 | * map. However, if the existing regions in the map can not | |
296 | * be expanded to represent the new range, a new file_region | |
297 | * structure is added to the map as a placeholder. This is | |
298 | * so that the subsequent region_add call will have all the | |
299 | * regions it needs and will not fail. | |
300 | * | |
301 | * Returns the number of huge pages that need to be added | |
302 | * to the existing reservation map for the range [f, t). | |
303 | * This number is greater or equal to zero. -ENOMEM is | |
304 | * returned if a new file_region structure is needed and can | |
305 | * not be allocated. | |
306 | */ | |
1406ec9b | 307 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 308 | { |
1406ec9b | 309 | struct list_head *head = &resv->regions; |
7b24d861 | 310 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
311 | long chg = 0; |
312 | ||
7b24d861 DB |
313 | retry: |
314 | spin_lock(&resv->lock); | |
96822904 AW |
315 | /* Locate the region we are before or in. */ |
316 | list_for_each_entry(rg, head, link) | |
317 | if (f <= rg->to) | |
318 | break; | |
319 | ||
320 | /* If we are below the current region then a new region is required. | |
321 | * Subtle, allocate a new region at the position but make it zero | |
322 | * size such that we can guarantee to record the reservation. */ | |
323 | if (&rg->link == head || t < rg->from) { | |
7b24d861 DB |
324 | if (!nrg) { |
325 | spin_unlock(&resv->lock); | |
326 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
327 | if (!nrg) | |
328 | return -ENOMEM; | |
329 | ||
330 | nrg->from = f; | |
331 | nrg->to = f; | |
332 | INIT_LIST_HEAD(&nrg->link); | |
333 | goto retry; | |
334 | } | |
96822904 | 335 | |
7b24d861 DB |
336 | list_add(&nrg->link, rg->link.prev); |
337 | chg = t - f; | |
338 | goto out_nrg; | |
96822904 AW |
339 | } |
340 | ||
341 | /* Round our left edge to the current segment if it encloses us. */ | |
342 | if (f > rg->from) | |
343 | f = rg->from; | |
344 | chg = t - f; | |
345 | ||
346 | /* Check for and consume any regions we now overlap with. */ | |
347 | list_for_each_entry(rg, rg->link.prev, link) { | |
348 | if (&rg->link == head) | |
349 | break; | |
350 | if (rg->from > t) | |
7b24d861 | 351 | goto out; |
96822904 | 352 | |
25985edc | 353 | /* We overlap with this area, if it extends further than |
96822904 AW |
354 | * us then we must extend ourselves. Account for its |
355 | * existing reservation. */ | |
356 | if (rg->to > t) { | |
357 | chg += rg->to - t; | |
358 | t = rg->to; | |
359 | } | |
360 | chg -= rg->to - rg->from; | |
361 | } | |
7b24d861 DB |
362 | |
363 | out: | |
364 | spin_unlock(&resv->lock); | |
365 | /* We already know we raced and no longer need the new region */ | |
366 | kfree(nrg); | |
367 | return chg; | |
368 | out_nrg: | |
369 | spin_unlock(&resv->lock); | |
96822904 AW |
370 | return chg; |
371 | } | |
372 | ||
1dd308a7 MK |
373 | /* |
374 | * Truncate the reserve map at index 'end'. Modify/truncate any | |
375 | * region which contains end. Delete any regions past end. | |
376 | * Return the number of huge pages removed from the map. | |
377 | */ | |
1406ec9b | 378 | static long region_truncate(struct resv_map *resv, long end) |
96822904 | 379 | { |
1406ec9b | 380 | struct list_head *head = &resv->regions; |
96822904 AW |
381 | struct file_region *rg, *trg; |
382 | long chg = 0; | |
383 | ||
7b24d861 | 384 | spin_lock(&resv->lock); |
96822904 AW |
385 | /* Locate the region we are either in or before. */ |
386 | list_for_each_entry(rg, head, link) | |
387 | if (end <= rg->to) | |
388 | break; | |
389 | if (&rg->link == head) | |
7b24d861 | 390 | goto out; |
96822904 AW |
391 | |
392 | /* If we are in the middle of a region then adjust it. */ | |
393 | if (end > rg->from) { | |
394 | chg = rg->to - end; | |
395 | rg->to = end; | |
396 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
397 | } | |
398 | ||
399 | /* Drop any remaining regions. */ | |
400 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
401 | if (&rg->link == head) | |
402 | break; | |
403 | chg += rg->to - rg->from; | |
404 | list_del(&rg->link); | |
405 | kfree(rg); | |
406 | } | |
7b24d861 DB |
407 | |
408 | out: | |
409 | spin_unlock(&resv->lock); | |
96822904 AW |
410 | return chg; |
411 | } | |
412 | ||
1dd308a7 MK |
413 | /* |
414 | * Count and return the number of huge pages in the reserve map | |
415 | * that intersect with the range [f, t). | |
416 | */ | |
1406ec9b | 417 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 418 | { |
1406ec9b | 419 | struct list_head *head = &resv->regions; |
84afd99b AW |
420 | struct file_region *rg; |
421 | long chg = 0; | |
422 | ||
7b24d861 | 423 | spin_lock(&resv->lock); |
84afd99b AW |
424 | /* Locate each segment we overlap with, and count that overlap. */ |
425 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
426 | long seg_from; |
427 | long seg_to; | |
84afd99b AW |
428 | |
429 | if (rg->to <= f) | |
430 | continue; | |
431 | if (rg->from >= t) | |
432 | break; | |
433 | ||
434 | seg_from = max(rg->from, f); | |
435 | seg_to = min(rg->to, t); | |
436 | ||
437 | chg += seg_to - seg_from; | |
438 | } | |
7b24d861 | 439 | spin_unlock(&resv->lock); |
84afd99b AW |
440 | |
441 | return chg; | |
442 | } | |
443 | ||
e7c4b0bf AW |
444 | /* |
445 | * Convert the address within this vma to the page offset within | |
446 | * the mapping, in pagecache page units; huge pages here. | |
447 | */ | |
a5516438 AK |
448 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
449 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 450 | { |
a5516438 AK |
451 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
452 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
453 | } |
454 | ||
0fe6e20b NH |
455 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
456 | unsigned long address) | |
457 | { | |
458 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
459 | } | |
460 | ||
08fba699 MG |
461 | /* |
462 | * Return the size of the pages allocated when backing a VMA. In the majority | |
463 | * cases this will be same size as used by the page table entries. | |
464 | */ | |
465 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
466 | { | |
467 | struct hstate *hstate; | |
468 | ||
469 | if (!is_vm_hugetlb_page(vma)) | |
470 | return PAGE_SIZE; | |
471 | ||
472 | hstate = hstate_vma(vma); | |
473 | ||
2415cf12 | 474 | return 1UL << huge_page_shift(hstate); |
08fba699 | 475 | } |
f340ca0f | 476 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 477 | |
3340289d MG |
478 | /* |
479 | * Return the page size being used by the MMU to back a VMA. In the majority | |
480 | * of cases, the page size used by the kernel matches the MMU size. On | |
481 | * architectures where it differs, an architecture-specific version of this | |
482 | * function is required. | |
483 | */ | |
484 | #ifndef vma_mmu_pagesize | |
485 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
486 | { | |
487 | return vma_kernel_pagesize(vma); | |
488 | } | |
489 | #endif | |
490 | ||
84afd99b AW |
491 | /* |
492 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
493 | * bits of the reservation map pointer, which are always clear due to | |
494 | * alignment. | |
495 | */ | |
496 | #define HPAGE_RESV_OWNER (1UL << 0) | |
497 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 498 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 499 | |
a1e78772 MG |
500 | /* |
501 | * These helpers are used to track how many pages are reserved for | |
502 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
503 | * is guaranteed to have their future faults succeed. | |
504 | * | |
505 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
506 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
507 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
508 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
509 | * |
510 | * The private mapping reservation is represented in a subtly different | |
511 | * manner to a shared mapping. A shared mapping has a region map associated | |
512 | * with the underlying file, this region map represents the backing file | |
513 | * pages which have ever had a reservation assigned which this persists even | |
514 | * after the page is instantiated. A private mapping has a region map | |
515 | * associated with the original mmap which is attached to all VMAs which | |
516 | * reference it, this region map represents those offsets which have consumed | |
517 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 518 | */ |
e7c4b0bf AW |
519 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
520 | { | |
521 | return (unsigned long)vma->vm_private_data; | |
522 | } | |
523 | ||
524 | static void set_vma_private_data(struct vm_area_struct *vma, | |
525 | unsigned long value) | |
526 | { | |
527 | vma->vm_private_data = (void *)value; | |
528 | } | |
529 | ||
9119a41e | 530 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
531 | { |
532 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
533 | if (!resv_map) | |
534 | return NULL; | |
535 | ||
536 | kref_init(&resv_map->refs); | |
7b24d861 | 537 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
538 | INIT_LIST_HEAD(&resv_map->regions); |
539 | ||
540 | return resv_map; | |
541 | } | |
542 | ||
9119a41e | 543 | void resv_map_release(struct kref *ref) |
84afd99b AW |
544 | { |
545 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
546 | ||
547 | /* Clear out any active regions before we release the map. */ | |
1406ec9b | 548 | region_truncate(resv_map, 0); |
84afd99b AW |
549 | kfree(resv_map); |
550 | } | |
551 | ||
4e35f483 JK |
552 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
553 | { | |
554 | return inode->i_mapping->private_data; | |
555 | } | |
556 | ||
84afd99b | 557 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 558 | { |
81d1b09c | 559 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
560 | if (vma->vm_flags & VM_MAYSHARE) { |
561 | struct address_space *mapping = vma->vm_file->f_mapping; | |
562 | struct inode *inode = mapping->host; | |
563 | ||
564 | return inode_resv_map(inode); | |
565 | ||
566 | } else { | |
84afd99b AW |
567 | return (struct resv_map *)(get_vma_private_data(vma) & |
568 | ~HPAGE_RESV_MASK); | |
4e35f483 | 569 | } |
a1e78772 MG |
570 | } |
571 | ||
84afd99b | 572 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 573 | { |
81d1b09c SL |
574 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
575 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 576 | |
84afd99b AW |
577 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
578 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
579 | } |
580 | ||
581 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
582 | { | |
81d1b09c SL |
583 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
584 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
585 | |
586 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
587 | } |
588 | ||
589 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
590 | { | |
81d1b09c | 591 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
592 | |
593 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
594 | } |
595 | ||
04f2cbe3 | 596 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
597 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
598 | { | |
81d1b09c | 599 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 600 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
601 | vma->vm_private_data = (void *)0; |
602 | } | |
603 | ||
604 | /* Returns true if the VMA has associated reserve pages */ | |
af0ed73e | 605 | static int vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 606 | { |
af0ed73e JK |
607 | if (vma->vm_flags & VM_NORESERVE) { |
608 | /* | |
609 | * This address is already reserved by other process(chg == 0), | |
610 | * so, we should decrement reserved count. Without decrementing, | |
611 | * reserve count remains after releasing inode, because this | |
612 | * allocated page will go into page cache and is regarded as | |
613 | * coming from reserved pool in releasing step. Currently, we | |
614 | * don't have any other solution to deal with this situation | |
615 | * properly, so add work-around here. | |
616 | */ | |
617 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
618 | return 1; | |
619 | else | |
620 | return 0; | |
621 | } | |
a63884e9 JK |
622 | |
623 | /* Shared mappings always use reserves */ | |
f83a275d | 624 | if (vma->vm_flags & VM_MAYSHARE) |
7f09ca51 | 625 | return 1; |
a63884e9 JK |
626 | |
627 | /* | |
628 | * Only the process that called mmap() has reserves for | |
629 | * private mappings. | |
630 | */ | |
7f09ca51 MG |
631 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
632 | return 1; | |
a63884e9 | 633 | |
7f09ca51 | 634 | return 0; |
a1e78772 MG |
635 | } |
636 | ||
a5516438 | 637 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
638 | { |
639 | int nid = page_to_nid(page); | |
0edaecfa | 640 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
641 | h->free_huge_pages++; |
642 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
643 | } |
644 | ||
bf50bab2 NH |
645 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
646 | { | |
647 | struct page *page; | |
648 | ||
c8721bbb NH |
649 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
650 | if (!is_migrate_isolate_page(page)) | |
651 | break; | |
652 | /* | |
653 | * if 'non-isolated free hugepage' not found on the list, | |
654 | * the allocation fails. | |
655 | */ | |
656 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 657 | return NULL; |
0edaecfa | 658 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 659 | set_page_refcounted(page); |
bf50bab2 NH |
660 | h->free_huge_pages--; |
661 | h->free_huge_pages_node[nid]--; | |
662 | return page; | |
663 | } | |
664 | ||
86cdb465 NH |
665 | /* Movability of hugepages depends on migration support. */ |
666 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
667 | { | |
100873d7 | 668 | if (hugepages_treat_as_movable || hugepage_migration_supported(h)) |
86cdb465 NH |
669 | return GFP_HIGHUSER_MOVABLE; |
670 | else | |
671 | return GFP_HIGHUSER; | |
672 | } | |
673 | ||
a5516438 AK |
674 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
675 | struct vm_area_struct *vma, | |
af0ed73e JK |
676 | unsigned long address, int avoid_reserve, |
677 | long chg) | |
1da177e4 | 678 | { |
b1c12cbc | 679 | struct page *page = NULL; |
480eccf9 | 680 | struct mempolicy *mpol; |
19770b32 | 681 | nodemask_t *nodemask; |
c0ff7453 | 682 | struct zonelist *zonelist; |
dd1a239f MG |
683 | struct zone *zone; |
684 | struct zoneref *z; | |
cc9a6c87 | 685 | unsigned int cpuset_mems_cookie; |
1da177e4 | 686 | |
a1e78772 MG |
687 | /* |
688 | * A child process with MAP_PRIVATE mappings created by their parent | |
689 | * have no page reserves. This check ensures that reservations are | |
690 | * not "stolen". The child may still get SIGKILLed | |
691 | */ | |
af0ed73e | 692 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 693 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 694 | goto err; |
a1e78772 | 695 | |
04f2cbe3 | 696 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 697 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 698 | goto err; |
04f2cbe3 | 699 | |
9966c4bb | 700 | retry_cpuset: |
d26914d1 | 701 | cpuset_mems_cookie = read_mems_allowed_begin(); |
9966c4bb | 702 | zonelist = huge_zonelist(vma, address, |
86cdb465 | 703 | htlb_alloc_mask(h), &mpol, &nodemask); |
9966c4bb | 704 | |
19770b32 MG |
705 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
706 | MAX_NR_ZONES - 1, nodemask) { | |
344736f2 | 707 | if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { |
bf50bab2 NH |
708 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
709 | if (page) { | |
af0ed73e JK |
710 | if (avoid_reserve) |
711 | break; | |
712 | if (!vma_has_reserves(vma, chg)) | |
713 | break; | |
714 | ||
07443a85 | 715 | SetPagePrivate(page); |
af0ed73e | 716 | h->resv_huge_pages--; |
bf50bab2 NH |
717 | break; |
718 | } | |
3abf7afd | 719 | } |
1da177e4 | 720 | } |
cc9a6c87 | 721 | |
52cd3b07 | 722 | mpol_cond_put(mpol); |
d26914d1 | 723 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
cc9a6c87 | 724 | goto retry_cpuset; |
1da177e4 | 725 | return page; |
cc9a6c87 MG |
726 | |
727 | err: | |
cc9a6c87 | 728 | return NULL; |
1da177e4 LT |
729 | } |
730 | ||
1cac6f2c LC |
731 | /* |
732 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
733 | * We may have allocated or freed a huge page based on a different | |
734 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
735 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
736 | * node for alloc or free. | |
737 | */ | |
738 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
739 | { | |
740 | nid = next_node(nid, *nodes_allowed); | |
741 | if (nid == MAX_NUMNODES) | |
742 | nid = first_node(*nodes_allowed); | |
743 | VM_BUG_ON(nid >= MAX_NUMNODES); | |
744 | ||
745 | return nid; | |
746 | } | |
747 | ||
748 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
749 | { | |
750 | if (!node_isset(nid, *nodes_allowed)) | |
751 | nid = next_node_allowed(nid, nodes_allowed); | |
752 | return nid; | |
753 | } | |
754 | ||
755 | /* | |
756 | * returns the previously saved node ["this node"] from which to | |
757 | * allocate a persistent huge page for the pool and advance the | |
758 | * next node from which to allocate, handling wrap at end of node | |
759 | * mask. | |
760 | */ | |
761 | static int hstate_next_node_to_alloc(struct hstate *h, | |
762 | nodemask_t *nodes_allowed) | |
763 | { | |
764 | int nid; | |
765 | ||
766 | VM_BUG_ON(!nodes_allowed); | |
767 | ||
768 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
769 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
770 | ||
771 | return nid; | |
772 | } | |
773 | ||
774 | /* | |
775 | * helper for free_pool_huge_page() - return the previously saved | |
776 | * node ["this node"] from which to free a huge page. Advance the | |
777 | * next node id whether or not we find a free huge page to free so | |
778 | * that the next attempt to free addresses the next node. | |
779 | */ | |
780 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
781 | { | |
782 | int nid; | |
783 | ||
784 | VM_BUG_ON(!nodes_allowed); | |
785 | ||
786 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
787 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
788 | ||
789 | return nid; | |
790 | } | |
791 | ||
792 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
793 | for (nr_nodes = nodes_weight(*mask); \ | |
794 | nr_nodes > 0 && \ | |
795 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
796 | nr_nodes--) | |
797 | ||
798 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
799 | for (nr_nodes = nodes_weight(*mask); \ | |
800 | nr_nodes > 0 && \ | |
801 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
802 | nr_nodes--) | |
803 | ||
944d9fec LC |
804 | #if defined(CONFIG_CMA) && defined(CONFIG_X86_64) |
805 | static void destroy_compound_gigantic_page(struct page *page, | |
806 | unsigned long order) | |
807 | { | |
808 | int i; | |
809 | int nr_pages = 1 << order; | |
810 | struct page *p = page + 1; | |
811 | ||
812 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
813 | __ClearPageTail(p); | |
814 | set_page_refcounted(p); | |
815 | p->first_page = NULL; | |
816 | } | |
817 | ||
818 | set_compound_order(page, 0); | |
819 | __ClearPageHead(page); | |
820 | } | |
821 | ||
822 | static void free_gigantic_page(struct page *page, unsigned order) | |
823 | { | |
824 | free_contig_range(page_to_pfn(page), 1 << order); | |
825 | } | |
826 | ||
827 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
828 | unsigned long nr_pages) | |
829 | { | |
830 | unsigned long end_pfn = start_pfn + nr_pages; | |
831 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); | |
832 | } | |
833 | ||
834 | static bool pfn_range_valid_gigantic(unsigned long start_pfn, | |
835 | unsigned long nr_pages) | |
836 | { | |
837 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
838 | struct page *page; | |
839 | ||
840 | for (i = start_pfn; i < end_pfn; i++) { | |
841 | if (!pfn_valid(i)) | |
842 | return false; | |
843 | ||
844 | page = pfn_to_page(i); | |
845 | ||
846 | if (PageReserved(page)) | |
847 | return false; | |
848 | ||
849 | if (page_count(page) > 0) | |
850 | return false; | |
851 | ||
852 | if (PageHuge(page)) | |
853 | return false; | |
854 | } | |
855 | ||
856 | return true; | |
857 | } | |
858 | ||
859 | static bool zone_spans_last_pfn(const struct zone *zone, | |
860 | unsigned long start_pfn, unsigned long nr_pages) | |
861 | { | |
862 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
863 | return zone_spans_pfn(zone, last_pfn); | |
864 | } | |
865 | ||
866 | static struct page *alloc_gigantic_page(int nid, unsigned order) | |
867 | { | |
868 | unsigned long nr_pages = 1 << order; | |
869 | unsigned long ret, pfn, flags; | |
870 | struct zone *z; | |
871 | ||
872 | z = NODE_DATA(nid)->node_zones; | |
873 | for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { | |
874 | spin_lock_irqsave(&z->lock, flags); | |
875 | ||
876 | pfn = ALIGN(z->zone_start_pfn, nr_pages); | |
877 | while (zone_spans_last_pfn(z, pfn, nr_pages)) { | |
878 | if (pfn_range_valid_gigantic(pfn, nr_pages)) { | |
879 | /* | |
880 | * We release the zone lock here because | |
881 | * alloc_contig_range() will also lock the zone | |
882 | * at some point. If there's an allocation | |
883 | * spinning on this lock, it may win the race | |
884 | * and cause alloc_contig_range() to fail... | |
885 | */ | |
886 | spin_unlock_irqrestore(&z->lock, flags); | |
887 | ret = __alloc_gigantic_page(pfn, nr_pages); | |
888 | if (!ret) | |
889 | return pfn_to_page(pfn); | |
890 | spin_lock_irqsave(&z->lock, flags); | |
891 | } | |
892 | pfn += nr_pages; | |
893 | } | |
894 | ||
895 | spin_unlock_irqrestore(&z->lock, flags); | |
896 | } | |
897 | ||
898 | return NULL; | |
899 | } | |
900 | ||
901 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
902 | static void prep_compound_gigantic_page(struct page *page, unsigned long order); | |
903 | ||
904 | static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) | |
905 | { | |
906 | struct page *page; | |
907 | ||
908 | page = alloc_gigantic_page(nid, huge_page_order(h)); | |
909 | if (page) { | |
910 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
911 | prep_new_huge_page(h, page, nid); | |
912 | } | |
913 | ||
914 | return page; | |
915 | } | |
916 | ||
917 | static int alloc_fresh_gigantic_page(struct hstate *h, | |
918 | nodemask_t *nodes_allowed) | |
919 | { | |
920 | struct page *page = NULL; | |
921 | int nr_nodes, node; | |
922 | ||
923 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
924 | page = alloc_fresh_gigantic_page_node(h, node); | |
925 | if (page) | |
926 | return 1; | |
927 | } | |
928 | ||
929 | return 0; | |
930 | } | |
931 | ||
932 | static inline bool gigantic_page_supported(void) { return true; } | |
933 | #else | |
934 | static inline bool gigantic_page_supported(void) { return false; } | |
935 | static inline void free_gigantic_page(struct page *page, unsigned order) { } | |
936 | static inline void destroy_compound_gigantic_page(struct page *page, | |
937 | unsigned long order) { } | |
938 | static inline int alloc_fresh_gigantic_page(struct hstate *h, | |
939 | nodemask_t *nodes_allowed) { return 0; } | |
940 | #endif | |
941 | ||
a5516438 | 942 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
943 | { |
944 | int i; | |
a5516438 | 945 | |
944d9fec LC |
946 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
947 | return; | |
18229df5 | 948 | |
a5516438 AK |
949 | h->nr_huge_pages--; |
950 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
951 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
952 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
953 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
954 | 1 << PG_active | 1 << PG_private | |
955 | 1 << PG_writeback); | |
6af2acb6 | 956 | } |
309381fe | 957 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
6af2acb6 AL |
958 | set_compound_page_dtor(page, NULL); |
959 | set_page_refcounted(page); | |
944d9fec LC |
960 | if (hstate_is_gigantic(h)) { |
961 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
962 | free_gigantic_page(page, huge_page_order(h)); | |
963 | } else { | |
964 | arch_release_hugepage(page); | |
965 | __free_pages(page, huge_page_order(h)); | |
966 | } | |
6af2acb6 AL |
967 | } |
968 | ||
e5ff2159 AK |
969 | struct hstate *size_to_hstate(unsigned long size) |
970 | { | |
971 | struct hstate *h; | |
972 | ||
973 | for_each_hstate(h) { | |
974 | if (huge_page_size(h) == size) | |
975 | return h; | |
976 | } | |
977 | return NULL; | |
978 | } | |
979 | ||
bcc54222 NH |
980 | /* |
981 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
982 | * to hstate->hugepage_activelist.) | |
983 | * | |
984 | * This function can be called for tail pages, but never returns true for them. | |
985 | */ | |
986 | bool page_huge_active(struct page *page) | |
987 | { | |
988 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
989 | return PageHead(page) && PagePrivate(&page[1]); | |
990 | } | |
991 | ||
992 | /* never called for tail page */ | |
993 | static void set_page_huge_active(struct page *page) | |
994 | { | |
995 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
996 | SetPagePrivate(&page[1]); | |
997 | } | |
998 | ||
999 | static void clear_page_huge_active(struct page *page) | |
1000 | { | |
1001 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1002 | ClearPagePrivate(&page[1]); | |
1003 | } | |
1004 | ||
8f1d26d0 | 1005 | void free_huge_page(struct page *page) |
27a85ef1 | 1006 | { |
a5516438 AK |
1007 | /* |
1008 | * Can't pass hstate in here because it is called from the | |
1009 | * compound page destructor. | |
1010 | */ | |
e5ff2159 | 1011 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1012 | int nid = page_to_nid(page); |
90481622 DG |
1013 | struct hugepage_subpool *spool = |
1014 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1015 | bool restore_reserve; |
27a85ef1 | 1016 | |
e5df70ab | 1017 | set_page_private(page, 0); |
23be7468 | 1018 | page->mapping = NULL; |
7893d1d5 | 1019 | BUG_ON(page_count(page)); |
0fe6e20b | 1020 | BUG_ON(page_mapcount(page)); |
07443a85 | 1021 | restore_reserve = PagePrivate(page); |
16c794b4 | 1022 | ClearPagePrivate(page); |
27a85ef1 | 1023 | |
1c5ecae3 MK |
1024 | /* |
1025 | * A return code of zero implies that the subpool will be under its | |
1026 | * minimum size if the reservation is not restored after page is free. | |
1027 | * Therefore, force restore_reserve operation. | |
1028 | */ | |
1029 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1030 | restore_reserve = true; | |
1031 | ||
27a85ef1 | 1032 | spin_lock(&hugetlb_lock); |
bcc54222 | 1033 | clear_page_huge_active(page); |
6d76dcf4 AK |
1034 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1035 | pages_per_huge_page(h), page); | |
07443a85 JK |
1036 | if (restore_reserve) |
1037 | h->resv_huge_pages++; | |
1038 | ||
944d9fec | 1039 | if (h->surplus_huge_pages_node[nid]) { |
0edaecfa AK |
1040 | /* remove the page from active list */ |
1041 | list_del(&page->lru); | |
a5516438 AK |
1042 | update_and_free_page(h, page); |
1043 | h->surplus_huge_pages--; | |
1044 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1045 | } else { |
5d3a551c | 1046 | arch_clear_hugepage_flags(page); |
a5516438 | 1047 | enqueue_huge_page(h, page); |
7893d1d5 | 1048 | } |
27a85ef1 DG |
1049 | spin_unlock(&hugetlb_lock); |
1050 | } | |
1051 | ||
a5516438 | 1052 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1053 | { |
0edaecfa | 1054 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
1055 | set_compound_page_dtor(page, free_huge_page); |
1056 | spin_lock(&hugetlb_lock); | |
9dd540e2 | 1057 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1058 | h->nr_huge_pages++; |
1059 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
1060 | spin_unlock(&hugetlb_lock); |
1061 | put_page(page); /* free it into the hugepage allocator */ | |
1062 | } | |
1063 | ||
2906dd52 | 1064 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
20a0307c WF |
1065 | { |
1066 | int i; | |
1067 | int nr_pages = 1 << order; | |
1068 | struct page *p = page + 1; | |
1069 | ||
1070 | /* we rely on prep_new_huge_page to set the destructor */ | |
1071 | set_compound_order(page, order); | |
1072 | __SetPageHead(page); | |
ef5a22be | 1073 | __ClearPageReserved(page); |
20a0307c | 1074 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1075 | /* |
1076 | * For gigantic hugepages allocated through bootmem at | |
1077 | * boot, it's safer to be consistent with the not-gigantic | |
1078 | * hugepages and clear the PG_reserved bit from all tail pages | |
1079 | * too. Otherwse drivers using get_user_pages() to access tail | |
1080 | * pages may get the reference counting wrong if they see | |
1081 | * PG_reserved set on a tail page (despite the head page not | |
1082 | * having PG_reserved set). Enforcing this consistency between | |
1083 | * head and tail pages allows drivers to optimize away a check | |
1084 | * on the head page when they need know if put_page() is needed | |
1085 | * after get_user_pages(). | |
1086 | */ | |
1087 | __ClearPageReserved(p); | |
58a84aa9 | 1088 | set_page_count(p, 0); |
20a0307c | 1089 | p->first_page = page; |
44fc8057 DR |
1090 | /* Make sure p->first_page is always valid for PageTail() */ |
1091 | smp_wmb(); | |
1092 | __SetPageTail(p); | |
20a0307c WF |
1093 | } |
1094 | } | |
1095 | ||
7795912c AM |
1096 | /* |
1097 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1098 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1099 | * details. | |
1100 | */ | |
20a0307c WF |
1101 | int PageHuge(struct page *page) |
1102 | { | |
20a0307c WF |
1103 | if (!PageCompound(page)) |
1104 | return 0; | |
1105 | ||
1106 | page = compound_head(page); | |
758f66a2 | 1107 | return get_compound_page_dtor(page) == free_huge_page; |
20a0307c | 1108 | } |
43131e14 NH |
1109 | EXPORT_SYMBOL_GPL(PageHuge); |
1110 | ||
27c73ae7 AA |
1111 | /* |
1112 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1113 | * normal or transparent huge pages. | |
1114 | */ | |
1115 | int PageHeadHuge(struct page *page_head) | |
1116 | { | |
27c73ae7 AA |
1117 | if (!PageHead(page_head)) |
1118 | return 0; | |
1119 | ||
758f66a2 | 1120 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1121 | } |
27c73ae7 | 1122 | |
13d60f4b ZY |
1123 | pgoff_t __basepage_index(struct page *page) |
1124 | { | |
1125 | struct page *page_head = compound_head(page); | |
1126 | pgoff_t index = page_index(page_head); | |
1127 | unsigned long compound_idx; | |
1128 | ||
1129 | if (!PageHuge(page_head)) | |
1130 | return page_index(page); | |
1131 | ||
1132 | if (compound_order(page_head) >= MAX_ORDER) | |
1133 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1134 | else | |
1135 | compound_idx = page - page_head; | |
1136 | ||
1137 | return (index << compound_order(page_head)) + compound_idx; | |
1138 | } | |
1139 | ||
a5516438 | 1140 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 1141 | { |
1da177e4 | 1142 | struct page *page; |
f96efd58 | 1143 | |
6484eb3e | 1144 | page = alloc_pages_exact_node(nid, |
86cdb465 | 1145 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
551883ae | 1146 | __GFP_REPEAT|__GFP_NOWARN, |
a5516438 | 1147 | huge_page_order(h)); |
1da177e4 | 1148 | if (page) { |
7f2e9525 | 1149 | if (arch_prepare_hugepage(page)) { |
caff3a2c | 1150 | __free_pages(page, huge_page_order(h)); |
7b8ee84d | 1151 | return NULL; |
7f2e9525 | 1152 | } |
a5516438 | 1153 | prep_new_huge_page(h, page, nid); |
1da177e4 | 1154 | } |
63b4613c NA |
1155 | |
1156 | return page; | |
1157 | } | |
1158 | ||
b2261026 JK |
1159 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
1160 | { | |
1161 | struct page *page; | |
1162 | int nr_nodes, node; | |
1163 | int ret = 0; | |
1164 | ||
1165 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1166 | page = alloc_fresh_huge_page_node(h, node); | |
1167 | if (page) { | |
1168 | ret = 1; | |
1169 | break; | |
1170 | } | |
1171 | } | |
1172 | ||
1173 | if (ret) | |
1174 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
1175 | else | |
1176 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1177 | ||
1178 | return ret; | |
1179 | } | |
1180 | ||
e8c5c824 LS |
1181 | /* |
1182 | * Free huge page from pool from next node to free. | |
1183 | * Attempt to keep persistent huge pages more or less | |
1184 | * balanced over allowed nodes. | |
1185 | * Called with hugetlb_lock locked. | |
1186 | */ | |
6ae11b27 LS |
1187 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1188 | bool acct_surplus) | |
e8c5c824 | 1189 | { |
b2261026 | 1190 | int nr_nodes, node; |
e8c5c824 LS |
1191 | int ret = 0; |
1192 | ||
b2261026 | 1193 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1194 | /* |
1195 | * If we're returning unused surplus pages, only examine | |
1196 | * nodes with surplus pages. | |
1197 | */ | |
b2261026 JK |
1198 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1199 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1200 | struct page *page = |
b2261026 | 1201 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1202 | struct page, lru); |
1203 | list_del(&page->lru); | |
1204 | h->free_huge_pages--; | |
b2261026 | 1205 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1206 | if (acct_surplus) { |
1207 | h->surplus_huge_pages--; | |
b2261026 | 1208 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1209 | } |
e8c5c824 LS |
1210 | update_and_free_page(h, page); |
1211 | ret = 1; | |
9a76db09 | 1212 | break; |
e8c5c824 | 1213 | } |
b2261026 | 1214 | } |
e8c5c824 LS |
1215 | |
1216 | return ret; | |
1217 | } | |
1218 | ||
c8721bbb NH |
1219 | /* |
1220 | * Dissolve a given free hugepage into free buddy pages. This function does | |
1221 | * nothing for in-use (including surplus) hugepages. | |
1222 | */ | |
1223 | static void dissolve_free_huge_page(struct page *page) | |
1224 | { | |
1225 | spin_lock(&hugetlb_lock); | |
1226 | if (PageHuge(page) && !page_count(page)) { | |
1227 | struct hstate *h = page_hstate(page); | |
1228 | int nid = page_to_nid(page); | |
1229 | list_del(&page->lru); | |
1230 | h->free_huge_pages--; | |
1231 | h->free_huge_pages_node[nid]--; | |
1232 | update_and_free_page(h, page); | |
1233 | } | |
1234 | spin_unlock(&hugetlb_lock); | |
1235 | } | |
1236 | ||
1237 | /* | |
1238 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1239 | * make specified memory blocks removable from the system. | |
1240 | * Note that start_pfn should aligned with (minimum) hugepage size. | |
1241 | */ | |
1242 | void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
1243 | { | |
c8721bbb | 1244 | unsigned long pfn; |
c8721bbb | 1245 | |
d0177639 LZ |
1246 | if (!hugepages_supported()) |
1247 | return; | |
1248 | ||
641844f5 NH |
1249 | VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order)); |
1250 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) | |
c8721bbb NH |
1251 | dissolve_free_huge_page(pfn_to_page(pfn)); |
1252 | } | |
1253 | ||
bf50bab2 | 1254 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
1255 | { |
1256 | struct page *page; | |
bf50bab2 | 1257 | unsigned int r_nid; |
7893d1d5 | 1258 | |
bae7f4ae | 1259 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1260 | return NULL; |
1261 | ||
d1c3fb1f NA |
1262 | /* |
1263 | * Assume we will successfully allocate the surplus page to | |
1264 | * prevent racing processes from causing the surplus to exceed | |
1265 | * overcommit | |
1266 | * | |
1267 | * This however introduces a different race, where a process B | |
1268 | * tries to grow the static hugepage pool while alloc_pages() is | |
1269 | * called by process A. B will only examine the per-node | |
1270 | * counters in determining if surplus huge pages can be | |
1271 | * converted to normal huge pages in adjust_pool_surplus(). A | |
1272 | * won't be able to increment the per-node counter, until the | |
1273 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
1274 | * no more huge pages can be converted from surplus to normal | |
1275 | * state (and doesn't try to convert again). Thus, we have a | |
1276 | * case where a surplus huge page exists, the pool is grown, and | |
1277 | * the surplus huge page still exists after, even though it | |
1278 | * should just have been converted to a normal huge page. This | |
1279 | * does not leak memory, though, as the hugepage will be freed | |
1280 | * once it is out of use. It also does not allow the counters to | |
1281 | * go out of whack in adjust_pool_surplus() as we don't modify | |
1282 | * the node values until we've gotten the hugepage and only the | |
1283 | * per-node value is checked there. | |
1284 | */ | |
1285 | spin_lock(&hugetlb_lock); | |
a5516438 | 1286 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
1287 | spin_unlock(&hugetlb_lock); |
1288 | return NULL; | |
1289 | } else { | |
a5516438 AK |
1290 | h->nr_huge_pages++; |
1291 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
1292 | } |
1293 | spin_unlock(&hugetlb_lock); | |
1294 | ||
bf50bab2 | 1295 | if (nid == NUMA_NO_NODE) |
86cdb465 | 1296 | page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| |
bf50bab2 NH |
1297 | __GFP_REPEAT|__GFP_NOWARN, |
1298 | huge_page_order(h)); | |
1299 | else | |
1300 | page = alloc_pages_exact_node(nid, | |
86cdb465 | 1301 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
bf50bab2 | 1302 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); |
d1c3fb1f | 1303 | |
caff3a2c GS |
1304 | if (page && arch_prepare_hugepage(page)) { |
1305 | __free_pages(page, huge_page_order(h)); | |
ea5768c7 | 1306 | page = NULL; |
caff3a2c GS |
1307 | } |
1308 | ||
d1c3fb1f | 1309 | spin_lock(&hugetlb_lock); |
7893d1d5 | 1310 | if (page) { |
0edaecfa | 1311 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1312 | r_nid = page_to_nid(page); |
7893d1d5 | 1313 | set_compound_page_dtor(page, free_huge_page); |
9dd540e2 | 1314 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1315 | /* |
1316 | * We incremented the global counters already | |
1317 | */ | |
bf50bab2 NH |
1318 | h->nr_huge_pages_node[r_nid]++; |
1319 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1320 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1321 | } else { |
a5516438 AK |
1322 | h->nr_huge_pages--; |
1323 | h->surplus_huge_pages--; | |
3b116300 | 1324 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1325 | } |
d1c3fb1f | 1326 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1327 | |
1328 | return page; | |
1329 | } | |
1330 | ||
bf50bab2 NH |
1331 | /* |
1332 | * This allocation function is useful in the context where vma is irrelevant. | |
1333 | * E.g. soft-offlining uses this function because it only cares physical | |
1334 | * address of error page. | |
1335 | */ | |
1336 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1337 | { | |
4ef91848 | 1338 | struct page *page = NULL; |
bf50bab2 NH |
1339 | |
1340 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1341 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1342 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1343 | spin_unlock(&hugetlb_lock); |
1344 | ||
94ae8ba7 | 1345 | if (!page) |
bf50bab2 NH |
1346 | page = alloc_buddy_huge_page(h, nid); |
1347 | ||
1348 | return page; | |
1349 | } | |
1350 | ||
e4e574b7 | 1351 | /* |
25985edc | 1352 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1353 | * of size 'delta'. |
1354 | */ | |
a5516438 | 1355 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1356 | { |
1357 | struct list_head surplus_list; | |
1358 | struct page *page, *tmp; | |
1359 | int ret, i; | |
1360 | int needed, allocated; | |
28073b02 | 1361 | bool alloc_ok = true; |
e4e574b7 | 1362 | |
a5516438 | 1363 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1364 | if (needed <= 0) { |
a5516438 | 1365 | h->resv_huge_pages += delta; |
e4e574b7 | 1366 | return 0; |
ac09b3a1 | 1367 | } |
e4e574b7 AL |
1368 | |
1369 | allocated = 0; | |
1370 | INIT_LIST_HEAD(&surplus_list); | |
1371 | ||
1372 | ret = -ENOMEM; | |
1373 | retry: | |
1374 | spin_unlock(&hugetlb_lock); | |
1375 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 1376 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
1377 | if (!page) { |
1378 | alloc_ok = false; | |
1379 | break; | |
1380 | } | |
e4e574b7 AL |
1381 | list_add(&page->lru, &surplus_list); |
1382 | } | |
28073b02 | 1383 | allocated += i; |
e4e574b7 AL |
1384 | |
1385 | /* | |
1386 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1387 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1388 | */ | |
1389 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1390 | needed = (h->resv_huge_pages + delta) - |
1391 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1392 | if (needed > 0) { |
1393 | if (alloc_ok) | |
1394 | goto retry; | |
1395 | /* | |
1396 | * We were not able to allocate enough pages to | |
1397 | * satisfy the entire reservation so we free what | |
1398 | * we've allocated so far. | |
1399 | */ | |
1400 | goto free; | |
1401 | } | |
e4e574b7 AL |
1402 | /* |
1403 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1404 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1405 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1406 | * allocator. Commit the entire reservation here to prevent another |
1407 | * process from stealing the pages as they are added to the pool but | |
1408 | * before they are reserved. | |
e4e574b7 AL |
1409 | */ |
1410 | needed += allocated; | |
a5516438 | 1411 | h->resv_huge_pages += delta; |
e4e574b7 | 1412 | ret = 0; |
a9869b83 | 1413 | |
19fc3f0a | 1414 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1415 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1416 | if ((--needed) < 0) |
1417 | break; | |
a9869b83 NH |
1418 | /* |
1419 | * This page is now managed by the hugetlb allocator and has | |
1420 | * no users -- drop the buddy allocator's reference. | |
1421 | */ | |
1422 | put_page_testzero(page); | |
309381fe | 1423 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1424 | enqueue_huge_page(h, page); |
19fc3f0a | 1425 | } |
28073b02 | 1426 | free: |
b0365c8d | 1427 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1428 | |
1429 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1430 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1431 | put_page(page); | |
a9869b83 | 1432 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1433 | |
1434 | return ret; | |
1435 | } | |
1436 | ||
1437 | /* | |
1438 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1439 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1440 | * never used. | |
685f3457 | 1441 | * Called with hugetlb_lock held. |
e4e574b7 | 1442 | */ |
a5516438 AK |
1443 | static void return_unused_surplus_pages(struct hstate *h, |
1444 | unsigned long unused_resv_pages) | |
e4e574b7 | 1445 | { |
e4e574b7 AL |
1446 | unsigned long nr_pages; |
1447 | ||
ac09b3a1 | 1448 | /* Uncommit the reservation */ |
a5516438 | 1449 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1450 | |
aa888a74 | 1451 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1452 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1453 | return; |
1454 | ||
a5516438 | 1455 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1456 | |
685f3457 LS |
1457 | /* |
1458 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1459 | * evenly across all nodes with memory. Iterate across these nodes |
1460 | * until we can no longer free unreserved surplus pages. This occurs | |
1461 | * when the nodes with surplus pages have no free pages. | |
1462 | * free_pool_huge_page() will balance the the freed pages across the | |
1463 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1464 | */ |
1465 | while (nr_pages--) { | |
8cebfcd0 | 1466 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
685f3457 | 1467 | break; |
7848a4bf | 1468 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 AL |
1469 | } |
1470 | } | |
1471 | ||
c37f9fb1 AW |
1472 | /* |
1473 | * Determine if the huge page at addr within the vma has an associated | |
1474 | * reservation. Where it does not we will need to logically increase | |
90481622 DG |
1475 | * reservation and actually increase subpool usage before an allocation |
1476 | * can occur. Where any new reservation would be required the | |
1477 | * reservation change is prepared, but not committed. Once the page | |
1478 | * has been allocated from the subpool and instantiated the change should | |
1479 | * be committed via vma_commit_reservation. No action is required on | |
1480 | * failure. | |
c37f9fb1 | 1481 | */ |
e2f17d94 | 1482 | static long vma_needs_reservation(struct hstate *h, |
a5516438 | 1483 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1484 | { |
4e35f483 JK |
1485 | struct resv_map *resv; |
1486 | pgoff_t idx; | |
1487 | long chg; | |
c37f9fb1 | 1488 | |
4e35f483 JK |
1489 | resv = vma_resv_map(vma); |
1490 | if (!resv) | |
84afd99b | 1491 | return 1; |
c37f9fb1 | 1492 | |
4e35f483 JK |
1493 | idx = vma_hugecache_offset(h, vma, addr); |
1494 | chg = region_chg(resv, idx, idx + 1); | |
84afd99b | 1495 | |
4e35f483 JK |
1496 | if (vma->vm_flags & VM_MAYSHARE) |
1497 | return chg; | |
1498 | else | |
1499 | return chg < 0 ? chg : 0; | |
c37f9fb1 | 1500 | } |
a5516438 AK |
1501 | static void vma_commit_reservation(struct hstate *h, |
1502 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 | 1503 | { |
4e35f483 JK |
1504 | struct resv_map *resv; |
1505 | pgoff_t idx; | |
84afd99b | 1506 | |
4e35f483 JK |
1507 | resv = vma_resv_map(vma); |
1508 | if (!resv) | |
1509 | return; | |
84afd99b | 1510 | |
4e35f483 JK |
1511 | idx = vma_hugecache_offset(h, vma, addr); |
1512 | region_add(resv, idx, idx + 1); | |
c37f9fb1 AW |
1513 | } |
1514 | ||
a1e78772 | 1515 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1516 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1517 | { |
90481622 | 1518 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1519 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1520 | struct page *page; |
e2f17d94 | 1521 | long chg; |
6d76dcf4 AK |
1522 | int ret, idx; |
1523 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1524 | |
6d76dcf4 | 1525 | idx = hstate_index(h); |
a1e78772 | 1526 | /* |
90481622 DG |
1527 | * Processes that did not create the mapping will have no |
1528 | * reserves and will not have accounted against subpool | |
1529 | * limit. Check that the subpool limit can be made before | |
1530 | * satisfying the allocation MAP_NORESERVE mappings may also | |
1531 | * need pages and subpool limit allocated allocated if no reserve | |
1532 | * mapping overlaps. | |
a1e78772 | 1533 | */ |
a5516438 | 1534 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 | 1535 | if (chg < 0) |
76dcee75 | 1536 | return ERR_PTR(-ENOMEM); |
8bb3f12e | 1537 | if (chg || avoid_reserve) |
1c5ecae3 | 1538 | if (hugepage_subpool_get_pages(spool, 1) < 0) |
76dcee75 | 1539 | return ERR_PTR(-ENOSPC); |
1da177e4 | 1540 | |
6d76dcf4 | 1541 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
1542 | if (ret) |
1543 | goto out_subpool_put; | |
1544 | ||
1da177e4 | 1545 | spin_lock(&hugetlb_lock); |
af0ed73e | 1546 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); |
81a6fcae | 1547 | if (!page) { |
94ae8ba7 | 1548 | spin_unlock(&hugetlb_lock); |
bf50bab2 | 1549 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
8f34af6f JZ |
1550 | if (!page) |
1551 | goto out_uncharge_cgroup; | |
1552 | ||
79dbb236 AK |
1553 | spin_lock(&hugetlb_lock); |
1554 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 1555 | /* Fall through */ |
68842c9b | 1556 | } |
81a6fcae JK |
1557 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
1558 | spin_unlock(&hugetlb_lock); | |
348ea204 | 1559 | |
90481622 | 1560 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1561 | |
a5516438 | 1562 | vma_commit_reservation(h, vma, addr); |
90d8b7e6 | 1563 | return page; |
8f34af6f JZ |
1564 | |
1565 | out_uncharge_cgroup: | |
1566 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
1567 | out_subpool_put: | |
1568 | if (chg || avoid_reserve) | |
1569 | hugepage_subpool_put_pages(spool, 1); | |
1570 | return ERR_PTR(-ENOSPC); | |
b45b5bd6 DG |
1571 | } |
1572 | ||
74060e4d NH |
1573 | /* |
1574 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
1575 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
1576 | * where no ERR_VALUE is expected to be returned. | |
1577 | */ | |
1578 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
1579 | unsigned long addr, int avoid_reserve) | |
1580 | { | |
1581 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
1582 | if (IS_ERR(page)) | |
1583 | page = NULL; | |
1584 | return page; | |
1585 | } | |
1586 | ||
91f47662 | 1587 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1588 | { |
1589 | struct huge_bootmem_page *m; | |
b2261026 | 1590 | int nr_nodes, node; |
aa888a74 | 1591 | |
b2261026 | 1592 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
1593 | void *addr; |
1594 | ||
8b89a116 GS |
1595 | addr = memblock_virt_alloc_try_nid_nopanic( |
1596 | huge_page_size(h), huge_page_size(h), | |
1597 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
1598 | if (addr) { |
1599 | /* | |
1600 | * Use the beginning of the huge page to store the | |
1601 | * huge_bootmem_page struct (until gather_bootmem | |
1602 | * puts them into the mem_map). | |
1603 | */ | |
1604 | m = addr; | |
91f47662 | 1605 | goto found; |
aa888a74 | 1606 | } |
aa888a74 AK |
1607 | } |
1608 | return 0; | |
1609 | ||
1610 | found: | |
df994ead | 1611 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
1612 | /* Put them into a private list first because mem_map is not up yet */ |
1613 | list_add(&m->list, &huge_boot_pages); | |
1614 | m->hstate = h; | |
1615 | return 1; | |
1616 | } | |
1617 | ||
f412c97a | 1618 | static void __init prep_compound_huge_page(struct page *page, int order) |
18229df5 AW |
1619 | { |
1620 | if (unlikely(order > (MAX_ORDER - 1))) | |
1621 | prep_compound_gigantic_page(page, order); | |
1622 | else | |
1623 | prep_compound_page(page, order); | |
1624 | } | |
1625 | ||
aa888a74 AK |
1626 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1627 | static void __init gather_bootmem_prealloc(void) | |
1628 | { | |
1629 | struct huge_bootmem_page *m; | |
1630 | ||
1631 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1632 | struct hstate *h = m->hstate; |
ee8f248d BB |
1633 | struct page *page; |
1634 | ||
1635 | #ifdef CONFIG_HIGHMEM | |
1636 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
1637 | memblock_free_late(__pa(m), |
1638 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
1639 | #else |
1640 | page = virt_to_page(m); | |
1641 | #endif | |
aa888a74 | 1642 | WARN_ON(page_count(page) != 1); |
18229df5 | 1643 | prep_compound_huge_page(page, h->order); |
ef5a22be | 1644 | WARN_ON(PageReserved(page)); |
aa888a74 | 1645 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1646 | /* |
1647 | * If we had gigantic hugepages allocated at boot time, we need | |
1648 | * to restore the 'stolen' pages to totalram_pages in order to | |
1649 | * fix confusing memory reports from free(1) and another | |
1650 | * side-effects, like CommitLimit going negative. | |
1651 | */ | |
bae7f4ae | 1652 | if (hstate_is_gigantic(h)) |
3dcc0571 | 1653 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
1654 | } |
1655 | } | |
1656 | ||
8faa8b07 | 1657 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1658 | { |
1659 | unsigned long i; | |
a5516438 | 1660 | |
e5ff2159 | 1661 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 1662 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
1663 | if (!alloc_bootmem_huge_page(h)) |
1664 | break; | |
9b5e5d0f | 1665 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 1666 | &node_states[N_MEMORY])) |
1da177e4 | 1667 | break; |
1da177e4 | 1668 | } |
8faa8b07 | 1669 | h->max_huge_pages = i; |
e5ff2159 AK |
1670 | } |
1671 | ||
1672 | static void __init hugetlb_init_hstates(void) | |
1673 | { | |
1674 | struct hstate *h; | |
1675 | ||
1676 | for_each_hstate(h) { | |
641844f5 NH |
1677 | if (minimum_order > huge_page_order(h)) |
1678 | minimum_order = huge_page_order(h); | |
1679 | ||
8faa8b07 | 1680 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 1681 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 1682 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 1683 | } |
641844f5 | 1684 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
1685 | } |
1686 | ||
4abd32db AK |
1687 | static char * __init memfmt(char *buf, unsigned long n) |
1688 | { | |
1689 | if (n >= (1UL << 30)) | |
1690 | sprintf(buf, "%lu GB", n >> 30); | |
1691 | else if (n >= (1UL << 20)) | |
1692 | sprintf(buf, "%lu MB", n >> 20); | |
1693 | else | |
1694 | sprintf(buf, "%lu KB", n >> 10); | |
1695 | return buf; | |
1696 | } | |
1697 | ||
e5ff2159 AK |
1698 | static void __init report_hugepages(void) |
1699 | { | |
1700 | struct hstate *h; | |
1701 | ||
1702 | for_each_hstate(h) { | |
4abd32db | 1703 | char buf[32]; |
ffb22af5 | 1704 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
1705 | memfmt(buf, huge_page_size(h)), |
1706 | h->free_huge_pages); | |
e5ff2159 AK |
1707 | } |
1708 | } | |
1709 | ||
1da177e4 | 1710 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1711 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1712 | nodemask_t *nodes_allowed) | |
1da177e4 | 1713 | { |
4415cc8d CL |
1714 | int i; |
1715 | ||
bae7f4ae | 1716 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1717 | return; |
1718 | ||
6ae11b27 | 1719 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1720 | struct page *page, *next; |
a5516438 AK |
1721 | struct list_head *freel = &h->hugepage_freelists[i]; |
1722 | list_for_each_entry_safe(page, next, freel, lru) { | |
1723 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1724 | return; |
1da177e4 LT |
1725 | if (PageHighMem(page)) |
1726 | continue; | |
1727 | list_del(&page->lru); | |
e5ff2159 | 1728 | update_and_free_page(h, page); |
a5516438 AK |
1729 | h->free_huge_pages--; |
1730 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1731 | } |
1732 | } | |
1733 | } | |
1734 | #else | |
6ae11b27 LS |
1735 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1736 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1737 | { |
1738 | } | |
1739 | #endif | |
1740 | ||
20a0307c WF |
1741 | /* |
1742 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1743 | * balanced by operating on them in a round-robin fashion. | |
1744 | * Returns 1 if an adjustment was made. | |
1745 | */ | |
6ae11b27 LS |
1746 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1747 | int delta) | |
20a0307c | 1748 | { |
b2261026 | 1749 | int nr_nodes, node; |
20a0307c WF |
1750 | |
1751 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1752 | |
b2261026 JK |
1753 | if (delta < 0) { |
1754 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1755 | if (h->surplus_huge_pages_node[node]) | |
1756 | goto found; | |
e8c5c824 | 1757 | } |
b2261026 JK |
1758 | } else { |
1759 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
1760 | if (h->surplus_huge_pages_node[node] < | |
1761 | h->nr_huge_pages_node[node]) | |
1762 | goto found; | |
e8c5c824 | 1763 | } |
b2261026 JK |
1764 | } |
1765 | return 0; | |
20a0307c | 1766 | |
b2261026 JK |
1767 | found: |
1768 | h->surplus_huge_pages += delta; | |
1769 | h->surplus_huge_pages_node[node] += delta; | |
1770 | return 1; | |
20a0307c WF |
1771 | } |
1772 | ||
a5516438 | 1773 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1774 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1775 | nodemask_t *nodes_allowed) | |
1da177e4 | 1776 | { |
7893d1d5 | 1777 | unsigned long min_count, ret; |
1da177e4 | 1778 | |
944d9fec | 1779 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
1780 | return h->max_huge_pages; |
1781 | ||
7893d1d5 AL |
1782 | /* |
1783 | * Increase the pool size | |
1784 | * First take pages out of surplus state. Then make up the | |
1785 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1786 | * |
1787 | * We might race with alloc_buddy_huge_page() here and be unable | |
1788 | * to convert a surplus huge page to a normal huge page. That is | |
1789 | * not critical, though, it just means the overall size of the | |
1790 | * pool might be one hugepage larger than it needs to be, but | |
1791 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1792 | */ |
1da177e4 | 1793 | spin_lock(&hugetlb_lock); |
a5516438 | 1794 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 1795 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
1796 | break; |
1797 | } | |
1798 | ||
a5516438 | 1799 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1800 | /* |
1801 | * If this allocation races such that we no longer need the | |
1802 | * page, free_huge_page will handle it by freeing the page | |
1803 | * and reducing the surplus. | |
1804 | */ | |
1805 | spin_unlock(&hugetlb_lock); | |
944d9fec LC |
1806 | if (hstate_is_gigantic(h)) |
1807 | ret = alloc_fresh_gigantic_page(h, nodes_allowed); | |
1808 | else | |
1809 | ret = alloc_fresh_huge_page(h, nodes_allowed); | |
7893d1d5 AL |
1810 | spin_lock(&hugetlb_lock); |
1811 | if (!ret) | |
1812 | goto out; | |
1813 | ||
536240f2 MG |
1814 | /* Bail for signals. Probably ctrl-c from user */ |
1815 | if (signal_pending(current)) | |
1816 | goto out; | |
7893d1d5 | 1817 | } |
7893d1d5 AL |
1818 | |
1819 | /* | |
1820 | * Decrease the pool size | |
1821 | * First return free pages to the buddy allocator (being careful | |
1822 | * to keep enough around to satisfy reservations). Then place | |
1823 | * pages into surplus state as needed so the pool will shrink | |
1824 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1825 | * |
1826 | * By placing pages into the surplus state independent of the | |
1827 | * overcommit value, we are allowing the surplus pool size to | |
1828 | * exceed overcommit. There are few sane options here. Since | |
1829 | * alloc_buddy_huge_page() is checking the global counter, | |
1830 | * though, we'll note that we're not allowed to exceed surplus | |
1831 | * and won't grow the pool anywhere else. Not until one of the | |
1832 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1833 | */ |
a5516438 | 1834 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1835 | min_count = max(count, min_count); |
6ae11b27 | 1836 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 1837 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 1838 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 1839 | break; |
55f67141 | 1840 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 1841 | } |
a5516438 | 1842 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 1843 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
1844 | break; |
1845 | } | |
1846 | out: | |
a5516438 | 1847 | ret = persistent_huge_pages(h); |
1da177e4 | 1848 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1849 | return ret; |
1da177e4 LT |
1850 | } |
1851 | ||
a3437870 NA |
1852 | #define HSTATE_ATTR_RO(_name) \ |
1853 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1854 | ||
1855 | #define HSTATE_ATTR(_name) \ | |
1856 | static struct kobj_attribute _name##_attr = \ | |
1857 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1858 | ||
1859 | static struct kobject *hugepages_kobj; | |
1860 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1861 | ||
9a305230 LS |
1862 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
1863 | ||
1864 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
1865 | { |
1866 | int i; | |
9a305230 | 1867 | |
a3437870 | 1868 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
1869 | if (hstate_kobjs[i] == kobj) { |
1870 | if (nidp) | |
1871 | *nidp = NUMA_NO_NODE; | |
a3437870 | 1872 | return &hstates[i]; |
9a305230 LS |
1873 | } |
1874 | ||
1875 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
1876 | } |
1877 | ||
06808b08 | 1878 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
1879 | struct kobj_attribute *attr, char *buf) |
1880 | { | |
9a305230 LS |
1881 | struct hstate *h; |
1882 | unsigned long nr_huge_pages; | |
1883 | int nid; | |
1884 | ||
1885 | h = kobj_to_hstate(kobj, &nid); | |
1886 | if (nid == NUMA_NO_NODE) | |
1887 | nr_huge_pages = h->nr_huge_pages; | |
1888 | else | |
1889 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
1890 | ||
1891 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 1892 | } |
adbe8726 | 1893 | |
238d3c13 DR |
1894 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
1895 | struct hstate *h, int nid, | |
1896 | unsigned long count, size_t len) | |
a3437870 NA |
1897 | { |
1898 | int err; | |
bad44b5b | 1899 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 1900 | |
944d9fec | 1901 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
1902 | err = -EINVAL; |
1903 | goto out; | |
1904 | } | |
1905 | ||
9a305230 LS |
1906 | if (nid == NUMA_NO_NODE) { |
1907 | /* | |
1908 | * global hstate attribute | |
1909 | */ | |
1910 | if (!(obey_mempolicy && | |
1911 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1912 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 1913 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
1914 | } |
1915 | } else if (nodes_allowed) { | |
1916 | /* | |
1917 | * per node hstate attribute: adjust count to global, | |
1918 | * but restrict alloc/free to the specified node. | |
1919 | */ | |
1920 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
1921 | init_nodemask_of_node(nodes_allowed, nid); | |
1922 | } else | |
8cebfcd0 | 1923 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 1924 | |
06808b08 | 1925 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 1926 | |
8cebfcd0 | 1927 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
1928 | NODEMASK_FREE(nodes_allowed); |
1929 | ||
1930 | return len; | |
adbe8726 EM |
1931 | out: |
1932 | NODEMASK_FREE(nodes_allowed); | |
1933 | return err; | |
06808b08 LS |
1934 | } |
1935 | ||
238d3c13 DR |
1936 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
1937 | struct kobject *kobj, const char *buf, | |
1938 | size_t len) | |
1939 | { | |
1940 | struct hstate *h; | |
1941 | unsigned long count; | |
1942 | int nid; | |
1943 | int err; | |
1944 | ||
1945 | err = kstrtoul(buf, 10, &count); | |
1946 | if (err) | |
1947 | return err; | |
1948 | ||
1949 | h = kobj_to_hstate(kobj, &nid); | |
1950 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
1951 | } | |
1952 | ||
06808b08 LS |
1953 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
1954 | struct kobj_attribute *attr, char *buf) | |
1955 | { | |
1956 | return nr_hugepages_show_common(kobj, attr, buf); | |
1957 | } | |
1958 | ||
1959 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1960 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1961 | { | |
238d3c13 | 1962 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
1963 | } |
1964 | HSTATE_ATTR(nr_hugepages); | |
1965 | ||
06808b08 LS |
1966 | #ifdef CONFIG_NUMA |
1967 | ||
1968 | /* | |
1969 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
1970 | * huge page alloc/free. | |
1971 | */ | |
1972 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
1973 | struct kobj_attribute *attr, char *buf) | |
1974 | { | |
1975 | return nr_hugepages_show_common(kobj, attr, buf); | |
1976 | } | |
1977 | ||
1978 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
1979 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1980 | { | |
238d3c13 | 1981 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
1982 | } |
1983 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
1984 | #endif | |
1985 | ||
1986 | ||
a3437870 NA |
1987 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
1988 | struct kobj_attribute *attr, char *buf) | |
1989 | { | |
9a305230 | 1990 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1991 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
1992 | } | |
adbe8726 | 1993 | |
a3437870 NA |
1994 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
1995 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1996 | { | |
1997 | int err; | |
1998 | unsigned long input; | |
9a305230 | 1999 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2000 | |
bae7f4ae | 2001 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2002 | return -EINVAL; |
2003 | ||
3dbb95f7 | 2004 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2005 | if (err) |
73ae31e5 | 2006 | return err; |
a3437870 NA |
2007 | |
2008 | spin_lock(&hugetlb_lock); | |
2009 | h->nr_overcommit_huge_pages = input; | |
2010 | spin_unlock(&hugetlb_lock); | |
2011 | ||
2012 | return count; | |
2013 | } | |
2014 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2015 | ||
2016 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2017 | struct kobj_attribute *attr, char *buf) | |
2018 | { | |
9a305230 LS |
2019 | struct hstate *h; |
2020 | unsigned long free_huge_pages; | |
2021 | int nid; | |
2022 | ||
2023 | h = kobj_to_hstate(kobj, &nid); | |
2024 | if (nid == NUMA_NO_NODE) | |
2025 | free_huge_pages = h->free_huge_pages; | |
2026 | else | |
2027 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2028 | ||
2029 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2030 | } |
2031 | HSTATE_ATTR_RO(free_hugepages); | |
2032 | ||
2033 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2034 | struct kobj_attribute *attr, char *buf) | |
2035 | { | |
9a305230 | 2036 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2037 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2038 | } | |
2039 | HSTATE_ATTR_RO(resv_hugepages); | |
2040 | ||
2041 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2042 | struct kobj_attribute *attr, char *buf) | |
2043 | { | |
9a305230 LS |
2044 | struct hstate *h; |
2045 | unsigned long surplus_huge_pages; | |
2046 | int nid; | |
2047 | ||
2048 | h = kobj_to_hstate(kobj, &nid); | |
2049 | if (nid == NUMA_NO_NODE) | |
2050 | surplus_huge_pages = h->surplus_huge_pages; | |
2051 | else | |
2052 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2053 | ||
2054 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2055 | } |
2056 | HSTATE_ATTR_RO(surplus_hugepages); | |
2057 | ||
2058 | static struct attribute *hstate_attrs[] = { | |
2059 | &nr_hugepages_attr.attr, | |
2060 | &nr_overcommit_hugepages_attr.attr, | |
2061 | &free_hugepages_attr.attr, | |
2062 | &resv_hugepages_attr.attr, | |
2063 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2064 | #ifdef CONFIG_NUMA |
2065 | &nr_hugepages_mempolicy_attr.attr, | |
2066 | #endif | |
a3437870 NA |
2067 | NULL, |
2068 | }; | |
2069 | ||
2070 | static struct attribute_group hstate_attr_group = { | |
2071 | .attrs = hstate_attrs, | |
2072 | }; | |
2073 | ||
094e9539 JM |
2074 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2075 | struct kobject **hstate_kobjs, | |
2076 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
2077 | { |
2078 | int retval; | |
972dc4de | 2079 | int hi = hstate_index(h); |
a3437870 | 2080 | |
9a305230 LS |
2081 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2082 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2083 | return -ENOMEM; |
2084 | ||
9a305230 | 2085 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2086 | if (retval) |
9a305230 | 2087 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2088 | |
2089 | return retval; | |
2090 | } | |
2091 | ||
2092 | static void __init hugetlb_sysfs_init(void) | |
2093 | { | |
2094 | struct hstate *h; | |
2095 | int err; | |
2096 | ||
2097 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2098 | if (!hugepages_kobj) | |
2099 | return; | |
2100 | ||
2101 | for_each_hstate(h) { | |
9a305230 LS |
2102 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2103 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2104 | if (err) |
ffb22af5 | 2105 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2106 | } |
2107 | } | |
2108 | ||
9a305230 LS |
2109 | #ifdef CONFIG_NUMA |
2110 | ||
2111 | /* | |
2112 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2113 | * with node devices in node_devices[] using a parallel array. The array |
2114 | * index of a node device or _hstate == node id. | |
2115 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2116 | * the base kernel, on the hugetlb module. |
2117 | */ | |
2118 | struct node_hstate { | |
2119 | struct kobject *hugepages_kobj; | |
2120 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2121 | }; | |
2122 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
2123 | ||
2124 | /* | |
10fbcf4c | 2125 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2126 | */ |
2127 | static struct attribute *per_node_hstate_attrs[] = { | |
2128 | &nr_hugepages_attr.attr, | |
2129 | &free_hugepages_attr.attr, | |
2130 | &surplus_hugepages_attr.attr, | |
2131 | NULL, | |
2132 | }; | |
2133 | ||
2134 | static struct attribute_group per_node_hstate_attr_group = { | |
2135 | .attrs = per_node_hstate_attrs, | |
2136 | }; | |
2137 | ||
2138 | /* | |
10fbcf4c | 2139 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2140 | * Returns node id via non-NULL nidp. |
2141 | */ | |
2142 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2143 | { | |
2144 | int nid; | |
2145 | ||
2146 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2147 | struct node_hstate *nhs = &node_hstates[nid]; | |
2148 | int i; | |
2149 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2150 | if (nhs->hstate_kobjs[i] == kobj) { | |
2151 | if (nidp) | |
2152 | *nidp = nid; | |
2153 | return &hstates[i]; | |
2154 | } | |
2155 | } | |
2156 | ||
2157 | BUG(); | |
2158 | return NULL; | |
2159 | } | |
2160 | ||
2161 | /* | |
10fbcf4c | 2162 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2163 | * No-op if no hstate attributes attached. |
2164 | */ | |
3cd8b44f | 2165 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2166 | { |
2167 | struct hstate *h; | |
10fbcf4c | 2168 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2169 | |
2170 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2171 | return; /* no hstate attributes */ |
9a305230 | 2172 | |
972dc4de AK |
2173 | for_each_hstate(h) { |
2174 | int idx = hstate_index(h); | |
2175 | if (nhs->hstate_kobjs[idx]) { | |
2176 | kobject_put(nhs->hstate_kobjs[idx]); | |
2177 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2178 | } |
972dc4de | 2179 | } |
9a305230 LS |
2180 | |
2181 | kobject_put(nhs->hugepages_kobj); | |
2182 | nhs->hugepages_kobj = NULL; | |
2183 | } | |
2184 | ||
2185 | /* | |
10fbcf4c | 2186 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
2187 | * that have them. |
2188 | */ | |
2189 | static void hugetlb_unregister_all_nodes(void) | |
2190 | { | |
2191 | int nid; | |
2192 | ||
2193 | /* | |
10fbcf4c | 2194 | * disable node device registrations. |
9a305230 LS |
2195 | */ |
2196 | register_hugetlbfs_with_node(NULL, NULL); | |
2197 | ||
2198 | /* | |
2199 | * remove hstate attributes from any nodes that have them. | |
2200 | */ | |
2201 | for (nid = 0; nid < nr_node_ids; nid++) | |
8732794b | 2202 | hugetlb_unregister_node(node_devices[nid]); |
9a305230 LS |
2203 | } |
2204 | ||
2205 | /* | |
10fbcf4c | 2206 | * Register hstate attributes for a single node device. |
9a305230 LS |
2207 | * No-op if attributes already registered. |
2208 | */ | |
3cd8b44f | 2209 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2210 | { |
2211 | struct hstate *h; | |
10fbcf4c | 2212 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2213 | int err; |
2214 | ||
2215 | if (nhs->hugepages_kobj) | |
2216 | return; /* already allocated */ | |
2217 | ||
2218 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2219 | &node->dev.kobj); |
9a305230 LS |
2220 | if (!nhs->hugepages_kobj) |
2221 | return; | |
2222 | ||
2223 | for_each_hstate(h) { | |
2224 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2225 | nhs->hstate_kobjs, | |
2226 | &per_node_hstate_attr_group); | |
2227 | if (err) { | |
ffb22af5 AM |
2228 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2229 | h->name, node->dev.id); | |
9a305230 LS |
2230 | hugetlb_unregister_node(node); |
2231 | break; | |
2232 | } | |
2233 | } | |
2234 | } | |
2235 | ||
2236 | /* | |
9b5e5d0f | 2237 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2238 | * devices of nodes that have memory. All on-line nodes should have |
2239 | * registered their associated device by this time. | |
9a305230 | 2240 | */ |
7d9ca000 | 2241 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2242 | { |
2243 | int nid; | |
2244 | ||
8cebfcd0 | 2245 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2246 | struct node *node = node_devices[nid]; |
10fbcf4c | 2247 | if (node->dev.id == nid) |
9a305230 LS |
2248 | hugetlb_register_node(node); |
2249 | } | |
2250 | ||
2251 | /* | |
10fbcf4c | 2252 | * Let the node device driver know we're here so it can |
9a305230 LS |
2253 | * [un]register hstate attributes on node hotplug. |
2254 | */ | |
2255 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2256 | hugetlb_unregister_node); | |
2257 | } | |
2258 | #else /* !CONFIG_NUMA */ | |
2259 | ||
2260 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2261 | { | |
2262 | BUG(); | |
2263 | if (nidp) | |
2264 | *nidp = -1; | |
2265 | return NULL; | |
2266 | } | |
2267 | ||
2268 | static void hugetlb_unregister_all_nodes(void) { } | |
2269 | ||
2270 | static void hugetlb_register_all_nodes(void) { } | |
2271 | ||
2272 | #endif | |
2273 | ||
a3437870 NA |
2274 | static void __exit hugetlb_exit(void) |
2275 | { | |
2276 | struct hstate *h; | |
2277 | ||
9a305230 LS |
2278 | hugetlb_unregister_all_nodes(); |
2279 | ||
a3437870 | 2280 | for_each_hstate(h) { |
972dc4de | 2281 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
2282 | } |
2283 | ||
2284 | kobject_put(hugepages_kobj); | |
8382d914 | 2285 | kfree(htlb_fault_mutex_table); |
a3437870 NA |
2286 | } |
2287 | module_exit(hugetlb_exit); | |
2288 | ||
2289 | static int __init hugetlb_init(void) | |
2290 | { | |
8382d914 DB |
2291 | int i; |
2292 | ||
457c1b27 | 2293 | if (!hugepages_supported()) |
0ef89d25 | 2294 | return 0; |
a3437870 | 2295 | |
e11bfbfc NP |
2296 | if (!size_to_hstate(default_hstate_size)) { |
2297 | default_hstate_size = HPAGE_SIZE; | |
2298 | if (!size_to_hstate(default_hstate_size)) | |
2299 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2300 | } |
972dc4de | 2301 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
2302 | if (default_hstate_max_huge_pages) |
2303 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
2304 | |
2305 | hugetlb_init_hstates(); | |
aa888a74 | 2306 | gather_bootmem_prealloc(); |
a3437870 NA |
2307 | report_hugepages(); |
2308 | ||
2309 | hugetlb_sysfs_init(); | |
9a305230 | 2310 | hugetlb_register_all_nodes(); |
7179e7bf | 2311 | hugetlb_cgroup_file_init(); |
9a305230 | 2312 | |
8382d914 DB |
2313 | #ifdef CONFIG_SMP |
2314 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2315 | #else | |
2316 | num_fault_mutexes = 1; | |
2317 | #endif | |
2318 | htlb_fault_mutex_table = | |
2319 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); | |
2320 | BUG_ON(!htlb_fault_mutex_table); | |
2321 | ||
2322 | for (i = 0; i < num_fault_mutexes; i++) | |
2323 | mutex_init(&htlb_fault_mutex_table[i]); | |
a3437870 NA |
2324 | return 0; |
2325 | } | |
2326 | module_init(hugetlb_init); | |
2327 | ||
2328 | /* Should be called on processing a hugepagesz=... option */ | |
2329 | void __init hugetlb_add_hstate(unsigned order) | |
2330 | { | |
2331 | struct hstate *h; | |
8faa8b07 AK |
2332 | unsigned long i; |
2333 | ||
a3437870 | 2334 | if (size_to_hstate(PAGE_SIZE << order)) { |
ffb22af5 | 2335 | pr_warning("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2336 | return; |
2337 | } | |
47d38344 | 2338 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2339 | BUG_ON(order == 0); |
47d38344 | 2340 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2341 | h->order = order; |
2342 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2343 | h->nr_huge_pages = 0; |
2344 | h->free_huge_pages = 0; | |
2345 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2346 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2347 | INIT_LIST_HEAD(&h->hugepage_activelist); |
8cebfcd0 LJ |
2348 | h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); |
2349 | h->next_nid_to_free = first_node(node_states[N_MEMORY]); | |
a3437870 NA |
2350 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2351 | huge_page_size(h)/1024); | |
8faa8b07 | 2352 | |
a3437870 NA |
2353 | parsed_hstate = h; |
2354 | } | |
2355 | ||
e11bfbfc | 2356 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2357 | { |
2358 | unsigned long *mhp; | |
8faa8b07 | 2359 | static unsigned long *last_mhp; |
a3437870 NA |
2360 | |
2361 | /* | |
47d38344 | 2362 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2363 | * so this hugepages= parameter goes to the "default hstate". |
2364 | */ | |
47d38344 | 2365 | if (!hugetlb_max_hstate) |
a3437870 NA |
2366 | mhp = &default_hstate_max_huge_pages; |
2367 | else | |
2368 | mhp = &parsed_hstate->max_huge_pages; | |
2369 | ||
8faa8b07 | 2370 | if (mhp == last_mhp) { |
ffb22af5 AM |
2371 | pr_warning("hugepages= specified twice without " |
2372 | "interleaving hugepagesz=, ignoring\n"); | |
8faa8b07 AK |
2373 | return 1; |
2374 | } | |
2375 | ||
a3437870 NA |
2376 | if (sscanf(s, "%lu", mhp) <= 0) |
2377 | *mhp = 0; | |
2378 | ||
8faa8b07 AK |
2379 | /* |
2380 | * Global state is always initialized later in hugetlb_init. | |
2381 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2382 | * use the bootmem allocator. | |
2383 | */ | |
47d38344 | 2384 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2385 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2386 | ||
2387 | last_mhp = mhp; | |
2388 | ||
a3437870 NA |
2389 | return 1; |
2390 | } | |
e11bfbfc NP |
2391 | __setup("hugepages=", hugetlb_nrpages_setup); |
2392 | ||
2393 | static int __init hugetlb_default_setup(char *s) | |
2394 | { | |
2395 | default_hstate_size = memparse(s, &s); | |
2396 | return 1; | |
2397 | } | |
2398 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2399 | |
8a213460 NA |
2400 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2401 | { | |
2402 | int node; | |
2403 | unsigned int nr = 0; | |
2404 | ||
2405 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2406 | nr += array[node]; | |
2407 | ||
2408 | return nr; | |
2409 | } | |
2410 | ||
2411 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2412 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2413 | struct ctl_table *table, int write, | |
2414 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2415 | { |
e5ff2159 | 2416 | struct hstate *h = &default_hstate; |
238d3c13 | 2417 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2418 | int ret; |
e5ff2159 | 2419 | |
457c1b27 NA |
2420 | if (!hugepages_supported()) |
2421 | return -ENOTSUPP; | |
2422 | ||
e5ff2159 AK |
2423 | table->data = &tmp; |
2424 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2425 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2426 | if (ret) | |
2427 | goto out; | |
e5ff2159 | 2428 | |
238d3c13 DR |
2429 | if (write) |
2430 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2431 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2432 | out: |
2433 | return ret; | |
1da177e4 | 2434 | } |
396faf03 | 2435 | |
06808b08 LS |
2436 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2437 | void __user *buffer, size_t *length, loff_t *ppos) | |
2438 | { | |
2439 | ||
2440 | return hugetlb_sysctl_handler_common(false, table, write, | |
2441 | buffer, length, ppos); | |
2442 | } | |
2443 | ||
2444 | #ifdef CONFIG_NUMA | |
2445 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2446 | void __user *buffer, size_t *length, loff_t *ppos) | |
2447 | { | |
2448 | return hugetlb_sysctl_handler_common(true, table, write, | |
2449 | buffer, length, ppos); | |
2450 | } | |
2451 | #endif /* CONFIG_NUMA */ | |
2452 | ||
a3d0c6aa | 2453 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2454 | void __user *buffer, |
a3d0c6aa NA |
2455 | size_t *length, loff_t *ppos) |
2456 | { | |
a5516438 | 2457 | struct hstate *h = &default_hstate; |
e5ff2159 | 2458 | unsigned long tmp; |
08d4a246 | 2459 | int ret; |
e5ff2159 | 2460 | |
457c1b27 NA |
2461 | if (!hugepages_supported()) |
2462 | return -ENOTSUPP; | |
2463 | ||
c033a93c | 2464 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2465 | |
bae7f4ae | 2466 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2467 | return -EINVAL; |
2468 | ||
e5ff2159 AK |
2469 | table->data = &tmp; |
2470 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2471 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2472 | if (ret) | |
2473 | goto out; | |
e5ff2159 AK |
2474 | |
2475 | if (write) { | |
2476 | spin_lock(&hugetlb_lock); | |
2477 | h->nr_overcommit_huge_pages = tmp; | |
2478 | spin_unlock(&hugetlb_lock); | |
2479 | } | |
08d4a246 MH |
2480 | out: |
2481 | return ret; | |
a3d0c6aa NA |
2482 | } |
2483 | ||
1da177e4 LT |
2484 | #endif /* CONFIG_SYSCTL */ |
2485 | ||
e1759c21 | 2486 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2487 | { |
a5516438 | 2488 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2489 | if (!hugepages_supported()) |
2490 | return; | |
e1759c21 | 2491 | seq_printf(m, |
4f98a2fe RR |
2492 | "HugePages_Total: %5lu\n" |
2493 | "HugePages_Free: %5lu\n" | |
2494 | "HugePages_Rsvd: %5lu\n" | |
2495 | "HugePages_Surp: %5lu\n" | |
2496 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2497 | h->nr_huge_pages, |
2498 | h->free_huge_pages, | |
2499 | h->resv_huge_pages, | |
2500 | h->surplus_huge_pages, | |
2501 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2502 | } |
2503 | ||
2504 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2505 | { | |
a5516438 | 2506 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2507 | if (!hugepages_supported()) |
2508 | return 0; | |
1da177e4 LT |
2509 | return sprintf(buf, |
2510 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2511 | "Node %d HugePages_Free: %5u\n" |
2512 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2513 | nid, h->nr_huge_pages_node[nid], |
2514 | nid, h->free_huge_pages_node[nid], | |
2515 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2516 | } |
2517 | ||
949f7ec5 DR |
2518 | void hugetlb_show_meminfo(void) |
2519 | { | |
2520 | struct hstate *h; | |
2521 | int nid; | |
2522 | ||
457c1b27 NA |
2523 | if (!hugepages_supported()) |
2524 | return; | |
2525 | ||
949f7ec5 DR |
2526 | for_each_node_state(nid, N_MEMORY) |
2527 | for_each_hstate(h) | |
2528 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
2529 | nid, | |
2530 | h->nr_huge_pages_node[nid], | |
2531 | h->free_huge_pages_node[nid], | |
2532 | h->surplus_huge_pages_node[nid], | |
2533 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
2534 | } | |
2535 | ||
1da177e4 LT |
2536 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2537 | unsigned long hugetlb_total_pages(void) | |
2538 | { | |
d0028588 WL |
2539 | struct hstate *h; |
2540 | unsigned long nr_total_pages = 0; | |
2541 | ||
2542 | for_each_hstate(h) | |
2543 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
2544 | return nr_total_pages; | |
1da177e4 | 2545 | } |
1da177e4 | 2546 | |
a5516438 | 2547 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2548 | { |
2549 | int ret = -ENOMEM; | |
2550 | ||
2551 | spin_lock(&hugetlb_lock); | |
2552 | /* | |
2553 | * When cpuset is configured, it breaks the strict hugetlb page | |
2554 | * reservation as the accounting is done on a global variable. Such | |
2555 | * reservation is completely rubbish in the presence of cpuset because | |
2556 | * the reservation is not checked against page availability for the | |
2557 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2558 | * with lack of free htlb page in cpuset that the task is in. | |
2559 | * Attempt to enforce strict accounting with cpuset is almost | |
2560 | * impossible (or too ugly) because cpuset is too fluid that | |
2561 | * task or memory node can be dynamically moved between cpusets. | |
2562 | * | |
2563 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2564 | * undesirable. However, in order to preserve some of the semantics, | |
2565 | * we fall back to check against current free page availability as | |
2566 | * a best attempt and hopefully to minimize the impact of changing | |
2567 | * semantics that cpuset has. | |
2568 | */ | |
2569 | if (delta > 0) { | |
a5516438 | 2570 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2571 | goto out; |
2572 | ||
a5516438 AK |
2573 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2574 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2575 | goto out; |
2576 | } | |
2577 | } | |
2578 | ||
2579 | ret = 0; | |
2580 | if (delta < 0) | |
a5516438 | 2581 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2582 | |
2583 | out: | |
2584 | spin_unlock(&hugetlb_lock); | |
2585 | return ret; | |
2586 | } | |
2587 | ||
84afd99b AW |
2588 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2589 | { | |
f522c3ac | 2590 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
2591 | |
2592 | /* | |
2593 | * This new VMA should share its siblings reservation map if present. | |
2594 | * The VMA will only ever have a valid reservation map pointer where | |
2595 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2596 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2597 | * after this open call completes. It is therefore safe to take a |
2598 | * new reference here without additional locking. | |
2599 | */ | |
4e35f483 | 2600 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 2601 | kref_get(&resv->refs); |
84afd99b AW |
2602 | } |
2603 | ||
a1e78772 MG |
2604 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2605 | { | |
a5516438 | 2606 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 2607 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 2608 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 2609 | unsigned long reserve, start, end; |
1c5ecae3 | 2610 | long gbl_reserve; |
84afd99b | 2611 | |
4e35f483 JK |
2612 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
2613 | return; | |
84afd99b | 2614 | |
4e35f483 JK |
2615 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2616 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 2617 | |
4e35f483 | 2618 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 2619 | |
4e35f483 JK |
2620 | kref_put(&resv->refs, resv_map_release); |
2621 | ||
2622 | if (reserve) { | |
1c5ecae3 MK |
2623 | /* |
2624 | * Decrement reserve counts. The global reserve count may be | |
2625 | * adjusted if the subpool has a minimum size. | |
2626 | */ | |
2627 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
2628 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 2629 | } |
a1e78772 MG |
2630 | } |
2631 | ||
1da177e4 LT |
2632 | /* |
2633 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2634 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2635 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2636 | * this far. | |
2637 | */ | |
d0217ac0 | 2638 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2639 | { |
2640 | BUG(); | |
d0217ac0 | 2641 | return 0; |
1da177e4 LT |
2642 | } |
2643 | ||
f0f37e2f | 2644 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2645 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2646 | .open = hugetlb_vm_op_open, |
a1e78772 | 2647 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2648 | }; |
2649 | ||
1e8f889b DG |
2650 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2651 | int writable) | |
63551ae0 DG |
2652 | { |
2653 | pte_t entry; | |
2654 | ||
1e8f889b | 2655 | if (writable) { |
106c992a GS |
2656 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
2657 | vma->vm_page_prot))); | |
63551ae0 | 2658 | } else { |
106c992a GS |
2659 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
2660 | vma->vm_page_prot)); | |
63551ae0 DG |
2661 | } |
2662 | entry = pte_mkyoung(entry); | |
2663 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2664 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2665 | |
2666 | return entry; | |
2667 | } | |
2668 | ||
1e8f889b DG |
2669 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2670 | unsigned long address, pte_t *ptep) | |
2671 | { | |
2672 | pte_t entry; | |
2673 | ||
106c992a | 2674 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2675 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2676 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2677 | } |
2678 | ||
4a705fef NH |
2679 | static int is_hugetlb_entry_migration(pte_t pte) |
2680 | { | |
2681 | swp_entry_t swp; | |
2682 | ||
2683 | if (huge_pte_none(pte) || pte_present(pte)) | |
2684 | return 0; | |
2685 | swp = pte_to_swp_entry(pte); | |
2686 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
2687 | return 1; | |
2688 | else | |
2689 | return 0; | |
2690 | } | |
2691 | ||
2692 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
2693 | { | |
2694 | swp_entry_t swp; | |
2695 | ||
2696 | if (huge_pte_none(pte) || pte_present(pte)) | |
2697 | return 0; | |
2698 | swp = pte_to_swp_entry(pte); | |
2699 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
2700 | return 1; | |
2701 | else | |
2702 | return 0; | |
2703 | } | |
1e8f889b | 2704 | |
63551ae0 DG |
2705 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2706 | struct vm_area_struct *vma) | |
2707 | { | |
2708 | pte_t *src_pte, *dst_pte, entry; | |
2709 | struct page *ptepage; | |
1c59827d | 2710 | unsigned long addr; |
1e8f889b | 2711 | int cow; |
a5516438 AK |
2712 | struct hstate *h = hstate_vma(vma); |
2713 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
2714 | unsigned long mmun_start; /* For mmu_notifiers */ |
2715 | unsigned long mmun_end; /* For mmu_notifiers */ | |
2716 | int ret = 0; | |
1e8f889b DG |
2717 | |
2718 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2719 | |
e8569dd2 AS |
2720 | mmun_start = vma->vm_start; |
2721 | mmun_end = vma->vm_end; | |
2722 | if (cow) | |
2723 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
2724 | ||
a5516438 | 2725 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 2726 | spinlock_t *src_ptl, *dst_ptl; |
c74df32c HD |
2727 | src_pte = huge_pte_offset(src, addr); |
2728 | if (!src_pte) | |
2729 | continue; | |
a5516438 | 2730 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
2731 | if (!dst_pte) { |
2732 | ret = -ENOMEM; | |
2733 | break; | |
2734 | } | |
c5c99429 LW |
2735 | |
2736 | /* If the pagetables are shared don't copy or take references */ | |
2737 | if (dst_pte == src_pte) | |
2738 | continue; | |
2739 | ||
cb900f41 KS |
2740 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
2741 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
2742 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
2743 | entry = huge_ptep_get(src_pte); |
2744 | if (huge_pte_none(entry)) { /* skip none entry */ | |
2745 | ; | |
2746 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
2747 | is_hugetlb_entry_hwpoisoned(entry))) { | |
2748 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
2749 | ||
2750 | if (is_write_migration_entry(swp_entry) && cow) { | |
2751 | /* | |
2752 | * COW mappings require pages in both | |
2753 | * parent and child to be set to read. | |
2754 | */ | |
2755 | make_migration_entry_read(&swp_entry); | |
2756 | entry = swp_entry_to_pte(swp_entry); | |
2757 | set_huge_pte_at(src, addr, src_pte, entry); | |
2758 | } | |
2759 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
2760 | } else { | |
34ee645e | 2761 | if (cow) { |
7f2e9525 | 2762 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e JR |
2763 | mmu_notifier_invalidate_range(src, mmun_start, |
2764 | mmun_end); | |
2765 | } | |
0253d634 | 2766 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
2767 | ptepage = pte_page(entry); |
2768 | get_page(ptepage); | |
0fe6e20b | 2769 | page_dup_rmap(ptepage); |
1c59827d HD |
2770 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2771 | } | |
cb900f41 KS |
2772 | spin_unlock(src_ptl); |
2773 | spin_unlock(dst_ptl); | |
63551ae0 | 2774 | } |
63551ae0 | 2775 | |
e8569dd2 AS |
2776 | if (cow) |
2777 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
2778 | ||
2779 | return ret; | |
63551ae0 DG |
2780 | } |
2781 | ||
24669e58 AK |
2782 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
2783 | unsigned long start, unsigned long end, | |
2784 | struct page *ref_page) | |
63551ae0 | 2785 | { |
24669e58 | 2786 | int force_flush = 0; |
63551ae0 DG |
2787 | struct mm_struct *mm = vma->vm_mm; |
2788 | unsigned long address; | |
c7546f8f | 2789 | pte_t *ptep; |
63551ae0 | 2790 | pte_t pte; |
cb900f41 | 2791 | spinlock_t *ptl; |
63551ae0 | 2792 | struct page *page; |
a5516438 AK |
2793 | struct hstate *h = hstate_vma(vma); |
2794 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
2795 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
2796 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 2797 | |
63551ae0 | 2798 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
2799 | BUG_ON(start & ~huge_page_mask(h)); |
2800 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 2801 | |
24669e58 | 2802 | tlb_start_vma(tlb, vma); |
2ec74c3e | 2803 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 2804 | address = start; |
24669e58 | 2805 | again: |
569f48b8 | 2806 | for (; address < end; address += sz) { |
c7546f8f | 2807 | ptep = huge_pte_offset(mm, address); |
4c887265 | 2808 | if (!ptep) |
c7546f8f DG |
2809 | continue; |
2810 | ||
cb900f41 | 2811 | ptl = huge_pte_lock(h, mm, ptep); |
39dde65c | 2812 | if (huge_pmd_unshare(mm, &address, ptep)) |
cb900f41 | 2813 | goto unlock; |
39dde65c | 2814 | |
6629326b HD |
2815 | pte = huge_ptep_get(ptep); |
2816 | if (huge_pte_none(pte)) | |
cb900f41 | 2817 | goto unlock; |
6629326b HD |
2818 | |
2819 | /* | |
9fbc1f63 NH |
2820 | * Migrating hugepage or HWPoisoned hugepage is already |
2821 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 2822 | */ |
9fbc1f63 | 2823 | if (unlikely(!pte_present(pte))) { |
106c992a | 2824 | huge_pte_clear(mm, address, ptep); |
cb900f41 | 2825 | goto unlock; |
8c4894c6 | 2826 | } |
6629326b HD |
2827 | |
2828 | page = pte_page(pte); | |
04f2cbe3 MG |
2829 | /* |
2830 | * If a reference page is supplied, it is because a specific | |
2831 | * page is being unmapped, not a range. Ensure the page we | |
2832 | * are about to unmap is the actual page of interest. | |
2833 | */ | |
2834 | if (ref_page) { | |
04f2cbe3 | 2835 | if (page != ref_page) |
cb900f41 | 2836 | goto unlock; |
04f2cbe3 MG |
2837 | |
2838 | /* | |
2839 | * Mark the VMA as having unmapped its page so that | |
2840 | * future faults in this VMA will fail rather than | |
2841 | * looking like data was lost | |
2842 | */ | |
2843 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
2844 | } | |
2845 | ||
c7546f8f | 2846 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 2847 | tlb_remove_tlb_entry(tlb, ptep, address); |
106c992a | 2848 | if (huge_pte_dirty(pte)) |
6649a386 | 2849 | set_page_dirty(page); |
9e81130b | 2850 | |
24669e58 AK |
2851 | page_remove_rmap(page); |
2852 | force_flush = !__tlb_remove_page(tlb, page); | |
cb900f41 | 2853 | if (force_flush) { |
569f48b8 | 2854 | address += sz; |
cb900f41 | 2855 | spin_unlock(ptl); |
24669e58 | 2856 | break; |
cb900f41 | 2857 | } |
9e81130b | 2858 | /* Bail out after unmapping reference page if supplied */ |
cb900f41 KS |
2859 | if (ref_page) { |
2860 | spin_unlock(ptl); | |
9e81130b | 2861 | break; |
cb900f41 KS |
2862 | } |
2863 | unlock: | |
2864 | spin_unlock(ptl); | |
63551ae0 | 2865 | } |
24669e58 AK |
2866 | /* |
2867 | * mmu_gather ran out of room to batch pages, we break out of | |
2868 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
2869 | * and page-free while holding it. | |
2870 | */ | |
2871 | if (force_flush) { | |
2872 | force_flush = 0; | |
2873 | tlb_flush_mmu(tlb); | |
2874 | if (address < end && !ref_page) | |
2875 | goto again; | |
fe1668ae | 2876 | } |
2ec74c3e | 2877 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 2878 | tlb_end_vma(tlb, vma); |
1da177e4 | 2879 | } |
63551ae0 | 2880 | |
d833352a MG |
2881 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
2882 | struct vm_area_struct *vma, unsigned long start, | |
2883 | unsigned long end, struct page *ref_page) | |
2884 | { | |
2885 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
2886 | ||
2887 | /* | |
2888 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
2889 | * test will fail on a vma being torn down, and not grab a page table | |
2890 | * on its way out. We're lucky that the flag has such an appropriate | |
2891 | * name, and can in fact be safely cleared here. We could clear it | |
2892 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 2893 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 2894 | * because in the context this is called, the VMA is about to be |
c8c06efa | 2895 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
2896 | */ |
2897 | vma->vm_flags &= ~VM_MAYSHARE; | |
2898 | } | |
2899 | ||
502717f4 | 2900 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2901 | unsigned long end, struct page *ref_page) |
502717f4 | 2902 | { |
24669e58 AK |
2903 | struct mm_struct *mm; |
2904 | struct mmu_gather tlb; | |
2905 | ||
2906 | mm = vma->vm_mm; | |
2907 | ||
2b047252 | 2908 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
2909 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
2910 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 KC |
2911 | } |
2912 | ||
04f2cbe3 MG |
2913 | /* |
2914 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
2915 | * mappping it owns the reserve page for. The intention is to unmap the page | |
2916 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
2917 | * same region. | |
2918 | */ | |
2f4612af DB |
2919 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
2920 | struct page *page, unsigned long address) | |
04f2cbe3 | 2921 | { |
7526674d | 2922 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
2923 | struct vm_area_struct *iter_vma; |
2924 | struct address_space *mapping; | |
04f2cbe3 MG |
2925 | pgoff_t pgoff; |
2926 | ||
2927 | /* | |
2928 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
2929 | * from page cache lookup which is in HPAGE_SIZE units. | |
2930 | */ | |
7526674d | 2931 | address = address & huge_page_mask(h); |
36e4f20a MH |
2932 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
2933 | vma->vm_pgoff; | |
496ad9aa | 2934 | mapping = file_inode(vma->vm_file)->i_mapping; |
04f2cbe3 | 2935 | |
4eb2b1dc MG |
2936 | /* |
2937 | * Take the mapping lock for the duration of the table walk. As | |
2938 | * this mapping should be shared between all the VMAs, | |
2939 | * __unmap_hugepage_range() is called as the lock is already held | |
2940 | */ | |
83cde9e8 | 2941 | i_mmap_lock_write(mapping); |
6b2dbba8 | 2942 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
2943 | /* Do not unmap the current VMA */ |
2944 | if (iter_vma == vma) | |
2945 | continue; | |
2946 | ||
2947 | /* | |
2948 | * Unmap the page from other VMAs without their own reserves. | |
2949 | * They get marked to be SIGKILLed if they fault in these | |
2950 | * areas. This is because a future no-page fault on this VMA | |
2951 | * could insert a zeroed page instead of the data existing | |
2952 | * from the time of fork. This would look like data corruption | |
2953 | */ | |
2954 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
2955 | unmap_hugepage_range(iter_vma, address, |
2956 | address + huge_page_size(h), page); | |
04f2cbe3 | 2957 | } |
83cde9e8 | 2958 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
2959 | } |
2960 | ||
0fe6e20b NH |
2961 | /* |
2962 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
2963 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
2964 | * cannot race with other handlers or page migration. | |
2965 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 2966 | */ |
1e8f889b | 2967 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 | 2968 | unsigned long address, pte_t *ptep, pte_t pte, |
cb900f41 | 2969 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 2970 | { |
a5516438 | 2971 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 2972 | struct page *old_page, *new_page; |
ad4404a2 | 2973 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
2974 | unsigned long mmun_start; /* For mmu_notifiers */ |
2975 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
2976 | |
2977 | old_page = pte_page(pte); | |
2978 | ||
04f2cbe3 | 2979 | retry_avoidcopy: |
1e8f889b DG |
2980 | /* If no-one else is actually using this page, avoid the copy |
2981 | * and just make the page writable */ | |
37a2140d JK |
2982 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
2983 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 2984 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 2985 | return 0; |
1e8f889b DG |
2986 | } |
2987 | ||
04f2cbe3 MG |
2988 | /* |
2989 | * If the process that created a MAP_PRIVATE mapping is about to | |
2990 | * perform a COW due to a shared page count, attempt to satisfy | |
2991 | * the allocation without using the existing reserves. The pagecache | |
2992 | * page is used to determine if the reserve at this address was | |
2993 | * consumed or not. If reserves were used, a partial faulted mapping | |
2994 | * at the time of fork() could consume its reserves on COW instead | |
2995 | * of the full address range. | |
2996 | */ | |
5944d011 | 2997 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
2998 | old_page != pagecache_page) |
2999 | outside_reserve = 1; | |
3000 | ||
1e8f889b | 3001 | page_cache_get(old_page); |
b76c8cfb | 3002 | |
ad4404a2 DB |
3003 | /* |
3004 | * Drop page table lock as buddy allocator may be called. It will | |
3005 | * be acquired again before returning to the caller, as expected. | |
3006 | */ | |
cb900f41 | 3007 | spin_unlock(ptl); |
04f2cbe3 | 3008 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 3009 | |
2fc39cec | 3010 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3011 | /* |
3012 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3013 | * it is due to references held by a child and an insufficient | |
3014 | * huge page pool. To guarantee the original mappers | |
3015 | * reliability, unmap the page from child processes. The child | |
3016 | * may get SIGKILLed if it later faults. | |
3017 | */ | |
3018 | if (outside_reserve) { | |
ad4404a2 | 3019 | page_cache_release(old_page); |
04f2cbe3 | 3020 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
3021 | unmap_ref_private(mm, vma, old_page, address); |
3022 | BUG_ON(huge_pte_none(pte)); | |
3023 | spin_lock(ptl); | |
3024 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); | |
3025 | if (likely(ptep && | |
3026 | pte_same(huge_ptep_get(ptep), pte))) | |
3027 | goto retry_avoidcopy; | |
3028 | /* | |
3029 | * race occurs while re-acquiring page table | |
3030 | * lock, and our job is done. | |
3031 | */ | |
3032 | return 0; | |
04f2cbe3 MG |
3033 | } |
3034 | ||
ad4404a2 DB |
3035 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
3036 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
3037 | goto out_release_old; | |
1e8f889b DG |
3038 | } |
3039 | ||
0fe6e20b NH |
3040 | /* |
3041 | * When the original hugepage is shared one, it does not have | |
3042 | * anon_vma prepared. | |
3043 | */ | |
44e2aa93 | 3044 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3045 | ret = VM_FAULT_OOM; |
3046 | goto out_release_all; | |
44e2aa93 | 3047 | } |
0fe6e20b | 3048 | |
47ad8475 AA |
3049 | copy_user_huge_page(new_page, old_page, address, vma, |
3050 | pages_per_huge_page(h)); | |
0ed361de | 3051 | __SetPageUptodate(new_page); |
bcc54222 | 3052 | set_page_huge_active(new_page); |
1e8f889b | 3053 | |
2ec74c3e SG |
3054 | mmun_start = address & huge_page_mask(h); |
3055 | mmun_end = mmun_start + huge_page_size(h); | |
3056 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 3057 | |
b76c8cfb | 3058 | /* |
cb900f41 | 3059 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3060 | * before the page tables are altered |
3061 | */ | |
cb900f41 | 3062 | spin_lock(ptl); |
a5516438 | 3063 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
a9af0c5d | 3064 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3065 | ClearPagePrivate(new_page); |
3066 | ||
1e8f889b | 3067 | /* Break COW */ |
8fe627ec | 3068 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 3069 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
3070 | set_huge_pte_at(mm, address, ptep, |
3071 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 3072 | page_remove_rmap(old_page); |
cd67f0d2 | 3073 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
3074 | /* Make the old page be freed below */ |
3075 | new_page = old_page; | |
3076 | } | |
cb900f41 | 3077 | spin_unlock(ptl); |
2ec74c3e | 3078 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 3079 | out_release_all: |
1e8f889b | 3080 | page_cache_release(new_page); |
ad4404a2 | 3081 | out_release_old: |
1e8f889b | 3082 | page_cache_release(old_page); |
8312034f | 3083 | |
ad4404a2 DB |
3084 | spin_lock(ptl); /* Caller expects lock to be held */ |
3085 | return ret; | |
1e8f889b DG |
3086 | } |
3087 | ||
04f2cbe3 | 3088 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3089 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3090 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3091 | { |
3092 | struct address_space *mapping; | |
e7c4b0bf | 3093 | pgoff_t idx; |
04f2cbe3 MG |
3094 | |
3095 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3096 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3097 | |
3098 | return find_lock_page(mapping, idx); | |
3099 | } | |
3100 | ||
3ae77f43 HD |
3101 | /* |
3102 | * Return whether there is a pagecache page to back given address within VMA. | |
3103 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3104 | */ | |
3105 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3106 | struct vm_area_struct *vma, unsigned long address) |
3107 | { | |
3108 | struct address_space *mapping; | |
3109 | pgoff_t idx; | |
3110 | struct page *page; | |
3111 | ||
3112 | mapping = vma->vm_file->f_mapping; | |
3113 | idx = vma_hugecache_offset(h, vma, address); | |
3114 | ||
3115 | page = find_get_page(mapping, idx); | |
3116 | if (page) | |
3117 | put_page(page); | |
3118 | return page != NULL; | |
3119 | } | |
3120 | ||
a1ed3dda | 3121 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
3122 | struct address_space *mapping, pgoff_t idx, |
3123 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3124 | { |
a5516438 | 3125 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 3126 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3127 | int anon_rmap = 0; |
4c887265 | 3128 | unsigned long size; |
4c887265 | 3129 | struct page *page; |
1e8f889b | 3130 | pte_t new_pte; |
cb900f41 | 3131 | spinlock_t *ptl; |
4c887265 | 3132 | |
04f2cbe3 MG |
3133 | /* |
3134 | * Currently, we are forced to kill the process in the event the | |
3135 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3136 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3137 | */ |
3138 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
ffb22af5 AM |
3139 | pr_warning("PID %d killed due to inadequate hugepage pool\n", |
3140 | current->pid); | |
04f2cbe3 MG |
3141 | return ret; |
3142 | } | |
3143 | ||
4c887265 AL |
3144 | /* |
3145 | * Use page lock to guard against racing truncation | |
3146 | * before we get page_table_lock. | |
3147 | */ | |
6bda666a CL |
3148 | retry: |
3149 | page = find_lock_page(mapping, idx); | |
3150 | if (!page) { | |
a5516438 | 3151 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
3152 | if (idx >= size) |
3153 | goto out; | |
04f2cbe3 | 3154 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 3155 | if (IS_ERR(page)) { |
76dcee75 AK |
3156 | ret = PTR_ERR(page); |
3157 | if (ret == -ENOMEM) | |
3158 | ret = VM_FAULT_OOM; | |
3159 | else | |
3160 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3161 | goto out; |
3162 | } | |
47ad8475 | 3163 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3164 | __SetPageUptodate(page); |
bcc54222 | 3165 | set_page_huge_active(page); |
ac9b9c66 | 3166 | |
f83a275d | 3167 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 3168 | int err; |
45c682a6 | 3169 | struct inode *inode = mapping->host; |
6bda666a CL |
3170 | |
3171 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3172 | if (err) { | |
3173 | put_page(page); | |
6bda666a CL |
3174 | if (err == -EEXIST) |
3175 | goto retry; | |
3176 | goto out; | |
3177 | } | |
07443a85 | 3178 | ClearPagePrivate(page); |
45c682a6 KC |
3179 | |
3180 | spin_lock(&inode->i_lock); | |
a5516438 | 3181 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 3182 | spin_unlock(&inode->i_lock); |
23be7468 | 3183 | } else { |
6bda666a | 3184 | lock_page(page); |
0fe6e20b NH |
3185 | if (unlikely(anon_vma_prepare(vma))) { |
3186 | ret = VM_FAULT_OOM; | |
3187 | goto backout_unlocked; | |
3188 | } | |
409eb8c2 | 3189 | anon_rmap = 1; |
23be7468 | 3190 | } |
0fe6e20b | 3191 | } else { |
998b4382 NH |
3192 | /* |
3193 | * If memory error occurs between mmap() and fault, some process | |
3194 | * don't have hwpoisoned swap entry for errored virtual address. | |
3195 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3196 | */ | |
3197 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3198 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3199 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3200 | goto backout_unlocked; |
3201 | } | |
6bda666a | 3202 | } |
1e8f889b | 3203 | |
57303d80 AW |
3204 | /* |
3205 | * If we are going to COW a private mapping later, we examine the | |
3206 | * pending reservations for this page now. This will ensure that | |
3207 | * any allocations necessary to record that reservation occur outside | |
3208 | * the spinlock. | |
3209 | */ | |
788c7df4 | 3210 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
2b26736c AW |
3211 | if (vma_needs_reservation(h, vma, address) < 0) { |
3212 | ret = VM_FAULT_OOM; | |
3213 | goto backout_unlocked; | |
3214 | } | |
57303d80 | 3215 | |
cb900f41 KS |
3216 | ptl = huge_pte_lockptr(h, mm, ptep); |
3217 | spin_lock(ptl); | |
a5516438 | 3218 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3219 | if (idx >= size) |
3220 | goto backout; | |
3221 | ||
83c54070 | 3222 | ret = 0; |
7f2e9525 | 3223 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3224 | goto backout; |
3225 | ||
07443a85 JK |
3226 | if (anon_rmap) { |
3227 | ClearPagePrivate(page); | |
409eb8c2 | 3228 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3229 | } else |
409eb8c2 | 3230 | page_dup_rmap(page); |
1e8f889b DG |
3231 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3232 | && (vma->vm_flags & VM_SHARED))); | |
3233 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3234 | ||
788c7df4 | 3235 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3236 | /* Optimization, do the COW without a second fault */ |
cb900f41 | 3237 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); |
1e8f889b DG |
3238 | } |
3239 | ||
cb900f41 | 3240 | spin_unlock(ptl); |
4c887265 AL |
3241 | unlock_page(page); |
3242 | out: | |
ac9b9c66 | 3243 | return ret; |
4c887265 AL |
3244 | |
3245 | backout: | |
cb900f41 | 3246 | spin_unlock(ptl); |
2b26736c | 3247 | backout_unlocked: |
4c887265 AL |
3248 | unlock_page(page); |
3249 | put_page(page); | |
3250 | goto out; | |
ac9b9c66 HD |
3251 | } |
3252 | ||
8382d914 DB |
3253 | #ifdef CONFIG_SMP |
3254 | static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, | |
3255 | struct vm_area_struct *vma, | |
3256 | struct address_space *mapping, | |
3257 | pgoff_t idx, unsigned long address) | |
3258 | { | |
3259 | unsigned long key[2]; | |
3260 | u32 hash; | |
3261 | ||
3262 | if (vma->vm_flags & VM_SHARED) { | |
3263 | key[0] = (unsigned long) mapping; | |
3264 | key[1] = idx; | |
3265 | } else { | |
3266 | key[0] = (unsigned long) mm; | |
3267 | key[1] = address >> huge_page_shift(h); | |
3268 | } | |
3269 | ||
3270 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3271 | ||
3272 | return hash & (num_fault_mutexes - 1); | |
3273 | } | |
3274 | #else | |
3275 | /* | |
3276 | * For uniprocesor systems we always use a single mutex, so just | |
3277 | * return 0 and avoid the hashing overhead. | |
3278 | */ | |
3279 | static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, | |
3280 | struct vm_area_struct *vma, | |
3281 | struct address_space *mapping, | |
3282 | pgoff_t idx, unsigned long address) | |
3283 | { | |
3284 | return 0; | |
3285 | } | |
3286 | #endif | |
3287 | ||
86e5216f | 3288 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3289 | unsigned long address, unsigned int flags) |
86e5216f | 3290 | { |
8382d914 | 3291 | pte_t *ptep, entry; |
cb900f41 | 3292 | spinlock_t *ptl; |
1e8f889b | 3293 | int ret; |
8382d914 DB |
3294 | u32 hash; |
3295 | pgoff_t idx; | |
0fe6e20b | 3296 | struct page *page = NULL; |
57303d80 | 3297 | struct page *pagecache_page = NULL; |
a5516438 | 3298 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3299 | struct address_space *mapping; |
0f792cf9 | 3300 | int need_wait_lock = 0; |
86e5216f | 3301 | |
1e16a539 KH |
3302 | address &= huge_page_mask(h); |
3303 | ||
fd6a03ed NH |
3304 | ptep = huge_pte_offset(mm, address); |
3305 | if (ptep) { | |
3306 | entry = huge_ptep_get(ptep); | |
290408d4 | 3307 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3308 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3309 | return 0; |
3310 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3311 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3312 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
3313 | } |
3314 | ||
a5516438 | 3315 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
3316 | if (!ptep) |
3317 | return VM_FAULT_OOM; | |
3318 | ||
8382d914 DB |
3319 | mapping = vma->vm_file->f_mapping; |
3320 | idx = vma_hugecache_offset(h, vma, address); | |
3321 | ||
3935baa9 DG |
3322 | /* |
3323 | * Serialize hugepage allocation and instantiation, so that we don't | |
3324 | * get spurious allocation failures if two CPUs race to instantiate | |
3325 | * the same page in the page cache. | |
3326 | */ | |
8382d914 DB |
3327 | hash = fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3328 | mutex_lock(&htlb_fault_mutex_table[hash]); | |
3329 | ||
7f2e9525 GS |
3330 | entry = huge_ptep_get(ptep); |
3331 | if (huge_pte_none(entry)) { | |
8382d914 | 3332 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3333 | goto out_mutex; |
3935baa9 | 3334 | } |
86e5216f | 3335 | |
83c54070 | 3336 | ret = 0; |
1e8f889b | 3337 | |
0f792cf9 NH |
3338 | /* |
3339 | * entry could be a migration/hwpoison entry at this point, so this | |
3340 | * check prevents the kernel from going below assuming that we have | |
3341 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3342 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3343 | * handle it. | |
3344 | */ | |
3345 | if (!pte_present(entry)) | |
3346 | goto out_mutex; | |
3347 | ||
57303d80 AW |
3348 | /* |
3349 | * If we are going to COW the mapping later, we examine the pending | |
3350 | * reservations for this page now. This will ensure that any | |
3351 | * allocations necessary to record that reservation occur outside the | |
3352 | * spinlock. For private mappings, we also lookup the pagecache | |
3353 | * page now as it is used to determine if a reservation has been | |
3354 | * consumed. | |
3355 | */ | |
106c992a | 3356 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3357 | if (vma_needs_reservation(h, vma, address) < 0) { |
3358 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3359 | goto out_mutex; |
2b26736c | 3360 | } |
57303d80 | 3361 | |
f83a275d | 3362 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3363 | pagecache_page = hugetlbfs_pagecache_page(h, |
3364 | vma, address); | |
3365 | } | |
3366 | ||
0f792cf9 NH |
3367 | ptl = huge_pte_lock(h, mm, ptep); |
3368 | ||
3369 | /* Check for a racing update before calling hugetlb_cow */ | |
3370 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3371 | goto out_ptl; | |
3372 | ||
56c9cfb1 NH |
3373 | /* |
3374 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3375 | * pagecache_page, so here we need take the former one | |
3376 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3377 | */ |
3378 | page = pte_page(entry); | |
3379 | if (page != pagecache_page) | |
0f792cf9 NH |
3380 | if (!trylock_page(page)) { |
3381 | need_wait_lock = 1; | |
3382 | goto out_ptl; | |
3383 | } | |
b4d1d99f | 3384 | |
0f792cf9 | 3385 | get_page(page); |
b4d1d99f | 3386 | |
788c7df4 | 3387 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3388 | if (!huge_pte_write(entry)) { |
57303d80 | 3389 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
cb900f41 | 3390 | pagecache_page, ptl); |
0f792cf9 | 3391 | goto out_put_page; |
b4d1d99f | 3392 | } |
106c992a | 3393 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3394 | } |
3395 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3396 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3397 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3398 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
3399 | out_put_page: |
3400 | if (page != pagecache_page) | |
3401 | unlock_page(page); | |
3402 | put_page(page); | |
cb900f41 KS |
3403 | out_ptl: |
3404 | spin_unlock(ptl); | |
57303d80 AW |
3405 | |
3406 | if (pagecache_page) { | |
3407 | unlock_page(pagecache_page); | |
3408 | put_page(pagecache_page); | |
3409 | } | |
b4d1d99f | 3410 | out_mutex: |
8382d914 | 3411 | mutex_unlock(&htlb_fault_mutex_table[hash]); |
0f792cf9 NH |
3412 | /* |
3413 | * Generally it's safe to hold refcount during waiting page lock. But | |
3414 | * here we just wait to defer the next page fault to avoid busy loop and | |
3415 | * the page is not used after unlocked before returning from the current | |
3416 | * page fault. So we are safe from accessing freed page, even if we wait | |
3417 | * here without taking refcount. | |
3418 | */ | |
3419 | if (need_wait_lock) | |
3420 | wait_on_page_locked(page); | |
1e8f889b | 3421 | return ret; |
86e5216f AL |
3422 | } |
3423 | ||
28a35716 ML |
3424 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
3425 | struct page **pages, struct vm_area_struct **vmas, | |
3426 | unsigned long *position, unsigned long *nr_pages, | |
3427 | long i, unsigned int flags) | |
63551ae0 | 3428 | { |
d5d4b0aa KC |
3429 | unsigned long pfn_offset; |
3430 | unsigned long vaddr = *position; | |
28a35716 | 3431 | unsigned long remainder = *nr_pages; |
a5516438 | 3432 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 3433 | |
63551ae0 | 3434 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 3435 | pte_t *pte; |
cb900f41 | 3436 | spinlock_t *ptl = NULL; |
2a15efc9 | 3437 | int absent; |
4c887265 | 3438 | struct page *page; |
63551ae0 | 3439 | |
02057967 DR |
3440 | /* |
3441 | * If we have a pending SIGKILL, don't keep faulting pages and | |
3442 | * potentially allocating memory. | |
3443 | */ | |
3444 | if (unlikely(fatal_signal_pending(current))) { | |
3445 | remainder = 0; | |
3446 | break; | |
3447 | } | |
3448 | ||
4c887265 AL |
3449 | /* |
3450 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 3451 | * each hugepage. We have to make sure we get the |
4c887265 | 3452 | * first, for the page indexing below to work. |
cb900f41 KS |
3453 | * |
3454 | * Note that page table lock is not held when pte is null. | |
4c887265 | 3455 | */ |
a5516438 | 3456 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
cb900f41 KS |
3457 | if (pte) |
3458 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
3459 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
3460 | ||
3461 | /* | |
3462 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
3463 | * an error where there's an empty slot with no huge pagecache |
3464 | * to back it. This way, we avoid allocating a hugepage, and | |
3465 | * the sparse dumpfile avoids allocating disk blocks, but its | |
3466 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 3467 | */ |
3ae77f43 HD |
3468 | if (absent && (flags & FOLL_DUMP) && |
3469 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
3470 | if (pte) |
3471 | spin_unlock(ptl); | |
2a15efc9 HD |
3472 | remainder = 0; |
3473 | break; | |
3474 | } | |
63551ae0 | 3475 | |
9cc3a5bd NH |
3476 | /* |
3477 | * We need call hugetlb_fault for both hugepages under migration | |
3478 | * (in which case hugetlb_fault waits for the migration,) and | |
3479 | * hwpoisoned hugepages (in which case we need to prevent the | |
3480 | * caller from accessing to them.) In order to do this, we use | |
3481 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
3482 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
3483 | * both cases, and because we can't follow correct pages | |
3484 | * directly from any kind of swap entries. | |
3485 | */ | |
3486 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
3487 | ((flags & FOLL_WRITE) && |
3488 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 3489 | int ret; |
63551ae0 | 3490 | |
cb900f41 KS |
3491 | if (pte) |
3492 | spin_unlock(ptl); | |
2a15efc9 HD |
3493 | ret = hugetlb_fault(mm, vma, vaddr, |
3494 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
a89182c7 | 3495 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 3496 | continue; |
63551ae0 | 3497 | |
4c887265 | 3498 | remainder = 0; |
4c887265 AL |
3499 | break; |
3500 | } | |
3501 | ||
a5516438 | 3502 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 3503 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 3504 | same_page: |
d6692183 | 3505 | if (pages) { |
2a15efc9 | 3506 | pages[i] = mem_map_offset(page, pfn_offset); |
a0368d4e | 3507 | get_page_foll(pages[i]); |
d6692183 | 3508 | } |
63551ae0 DG |
3509 | |
3510 | if (vmas) | |
3511 | vmas[i] = vma; | |
3512 | ||
3513 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 3514 | ++pfn_offset; |
63551ae0 DG |
3515 | --remainder; |
3516 | ++i; | |
d5d4b0aa | 3517 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 3518 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
3519 | /* |
3520 | * We use pfn_offset to avoid touching the pageframes | |
3521 | * of this compound page. | |
3522 | */ | |
3523 | goto same_page; | |
3524 | } | |
cb900f41 | 3525 | spin_unlock(ptl); |
63551ae0 | 3526 | } |
28a35716 | 3527 | *nr_pages = remainder; |
63551ae0 DG |
3528 | *position = vaddr; |
3529 | ||
2a15efc9 | 3530 | return i ? i : -EFAULT; |
63551ae0 | 3531 | } |
8f860591 | 3532 | |
7da4d641 | 3533 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
3534 | unsigned long address, unsigned long end, pgprot_t newprot) |
3535 | { | |
3536 | struct mm_struct *mm = vma->vm_mm; | |
3537 | unsigned long start = address; | |
3538 | pte_t *ptep; | |
3539 | pte_t pte; | |
a5516438 | 3540 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 3541 | unsigned long pages = 0; |
8f860591 ZY |
3542 | |
3543 | BUG_ON(address >= end); | |
3544 | flush_cache_range(vma, address, end); | |
3545 | ||
a5338093 | 3546 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 3547 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 3548 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 3549 | spinlock_t *ptl; |
8f860591 ZY |
3550 | ptep = huge_pte_offset(mm, address); |
3551 | if (!ptep) | |
3552 | continue; | |
cb900f41 | 3553 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
3554 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3555 | pages++; | |
cb900f41 | 3556 | spin_unlock(ptl); |
39dde65c | 3557 | continue; |
7da4d641 | 3558 | } |
a8bda28d NH |
3559 | pte = huge_ptep_get(ptep); |
3560 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
3561 | spin_unlock(ptl); | |
3562 | continue; | |
3563 | } | |
3564 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
3565 | swp_entry_t entry = pte_to_swp_entry(pte); | |
3566 | ||
3567 | if (is_write_migration_entry(entry)) { | |
3568 | pte_t newpte; | |
3569 | ||
3570 | make_migration_entry_read(&entry); | |
3571 | newpte = swp_entry_to_pte(entry); | |
3572 | set_huge_pte_at(mm, address, ptep, newpte); | |
3573 | pages++; | |
3574 | } | |
3575 | spin_unlock(ptl); | |
3576 | continue; | |
3577 | } | |
3578 | if (!huge_pte_none(pte)) { | |
8f860591 | 3579 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 3580 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 3581 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 3582 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 3583 | pages++; |
8f860591 | 3584 | } |
cb900f41 | 3585 | spin_unlock(ptl); |
8f860591 | 3586 | } |
d833352a | 3587 | /* |
c8c06efa | 3588 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 3589 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 3590 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
3591 | * and that page table be reused and filled with junk. |
3592 | */ | |
8f860591 | 3593 | flush_tlb_range(vma, start, end); |
34ee645e | 3594 | mmu_notifier_invalidate_range(mm, start, end); |
83cde9e8 | 3595 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 3596 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
3597 | |
3598 | return pages << h->order; | |
8f860591 ZY |
3599 | } |
3600 | ||
a1e78772 MG |
3601 | int hugetlb_reserve_pages(struct inode *inode, |
3602 | long from, long to, | |
5a6fe125 | 3603 | struct vm_area_struct *vma, |
ca16d140 | 3604 | vm_flags_t vm_flags) |
e4e574b7 | 3605 | { |
17c9d12e | 3606 | long ret, chg; |
a5516438 | 3607 | struct hstate *h = hstate_inode(inode); |
90481622 | 3608 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 3609 | struct resv_map *resv_map; |
1c5ecae3 | 3610 | long gbl_reserve; |
e4e574b7 | 3611 | |
17c9d12e MG |
3612 | /* |
3613 | * Only apply hugepage reservation if asked. At fault time, an | |
3614 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3615 | * without using reserves |
17c9d12e | 3616 | */ |
ca16d140 | 3617 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3618 | return 0; |
3619 | ||
a1e78772 MG |
3620 | /* |
3621 | * Shared mappings base their reservation on the number of pages that | |
3622 | * are already allocated on behalf of the file. Private mappings need | |
3623 | * to reserve the full area even if read-only as mprotect() may be | |
3624 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3625 | */ | |
9119a41e | 3626 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 3627 | resv_map = inode_resv_map(inode); |
9119a41e | 3628 | |
1406ec9b | 3629 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
3630 | |
3631 | } else { | |
3632 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
3633 | if (!resv_map) |
3634 | return -ENOMEM; | |
3635 | ||
a1e78772 | 3636 | chg = to - from; |
84afd99b | 3637 | |
17c9d12e MG |
3638 | set_vma_resv_map(vma, resv_map); |
3639 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3640 | } | |
3641 | ||
c50ac050 DH |
3642 | if (chg < 0) { |
3643 | ret = chg; | |
3644 | goto out_err; | |
3645 | } | |
8a630112 | 3646 | |
1c5ecae3 MK |
3647 | /* |
3648 | * There must be enough pages in the subpool for the mapping. If | |
3649 | * the subpool has a minimum size, there may be some global | |
3650 | * reservations already in place (gbl_reserve). | |
3651 | */ | |
3652 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
3653 | if (gbl_reserve < 0) { | |
c50ac050 DH |
3654 | ret = -ENOSPC; |
3655 | goto out_err; | |
3656 | } | |
5a6fe125 MG |
3657 | |
3658 | /* | |
17c9d12e | 3659 | * Check enough hugepages are available for the reservation. |
90481622 | 3660 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3661 | */ |
1c5ecae3 | 3662 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 3663 | if (ret < 0) { |
1c5ecae3 MK |
3664 | /* put back original number of pages, chg */ |
3665 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 3666 | goto out_err; |
68842c9b | 3667 | } |
17c9d12e MG |
3668 | |
3669 | /* | |
3670 | * Account for the reservations made. Shared mappings record regions | |
3671 | * that have reservations as they are shared by multiple VMAs. | |
3672 | * When the last VMA disappears, the region map says how much | |
3673 | * the reservation was and the page cache tells how much of | |
3674 | * the reservation was consumed. Private mappings are per-VMA and | |
3675 | * only the consumed reservations are tracked. When the VMA | |
3676 | * disappears, the original reservation is the VMA size and the | |
3677 | * consumed reservations are stored in the map. Hence, nothing | |
3678 | * else has to be done for private mappings here | |
3679 | */ | |
f83a275d | 3680 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
1406ec9b | 3681 | region_add(resv_map, from, to); |
a43a8c39 | 3682 | return 0; |
c50ac050 | 3683 | out_err: |
f031dd27 JK |
3684 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3685 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 3686 | return ret; |
a43a8c39 KC |
3687 | } |
3688 | ||
3689 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
3690 | { | |
a5516438 | 3691 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 3692 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 3693 | long chg = 0; |
90481622 | 3694 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 3695 | long gbl_reserve; |
45c682a6 | 3696 | |
9119a41e | 3697 | if (resv_map) |
1406ec9b | 3698 | chg = region_truncate(resv_map, offset); |
45c682a6 | 3699 | spin_lock(&inode->i_lock); |
e4c6f8be | 3700 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
3701 | spin_unlock(&inode->i_lock); |
3702 | ||
1c5ecae3 MK |
3703 | /* |
3704 | * If the subpool has a minimum size, the number of global | |
3705 | * reservations to be released may be adjusted. | |
3706 | */ | |
3707 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
3708 | hugetlb_acct_memory(h, -gbl_reserve); | |
a43a8c39 | 3709 | } |
93f70f90 | 3710 | |
3212b535 SC |
3711 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
3712 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
3713 | struct vm_area_struct *vma, | |
3714 | unsigned long addr, pgoff_t idx) | |
3715 | { | |
3716 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
3717 | svma->vm_start; | |
3718 | unsigned long sbase = saddr & PUD_MASK; | |
3719 | unsigned long s_end = sbase + PUD_SIZE; | |
3720 | ||
3721 | /* Allow segments to share if only one is marked locked */ | |
3722 | unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; | |
3723 | unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; | |
3724 | ||
3725 | /* | |
3726 | * match the virtual addresses, permission and the alignment of the | |
3727 | * page table page. | |
3728 | */ | |
3729 | if (pmd_index(addr) != pmd_index(saddr) || | |
3730 | vm_flags != svm_flags || | |
3731 | sbase < svma->vm_start || svma->vm_end < s_end) | |
3732 | return 0; | |
3733 | ||
3734 | return saddr; | |
3735 | } | |
3736 | ||
3737 | static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) | |
3738 | { | |
3739 | unsigned long base = addr & PUD_MASK; | |
3740 | unsigned long end = base + PUD_SIZE; | |
3741 | ||
3742 | /* | |
3743 | * check on proper vm_flags and page table alignment | |
3744 | */ | |
3745 | if (vma->vm_flags & VM_MAYSHARE && | |
3746 | vma->vm_start <= base && end <= vma->vm_end) | |
3747 | return 1; | |
3748 | return 0; | |
3749 | } | |
3750 | ||
3751 | /* | |
3752 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
3753 | * and returns the corresponding pte. While this is not necessary for the | |
3754 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
3755 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 3756 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
3757 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
3758 | * bad pmd for sharing. | |
3759 | */ | |
3760 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3761 | { | |
3762 | struct vm_area_struct *vma = find_vma(mm, addr); | |
3763 | struct address_space *mapping = vma->vm_file->f_mapping; | |
3764 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
3765 | vma->vm_pgoff; | |
3766 | struct vm_area_struct *svma; | |
3767 | unsigned long saddr; | |
3768 | pte_t *spte = NULL; | |
3769 | pte_t *pte; | |
cb900f41 | 3770 | spinlock_t *ptl; |
3212b535 SC |
3771 | |
3772 | if (!vma_shareable(vma, addr)) | |
3773 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
3774 | ||
83cde9e8 | 3775 | i_mmap_lock_write(mapping); |
3212b535 SC |
3776 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
3777 | if (svma == vma) | |
3778 | continue; | |
3779 | ||
3780 | saddr = page_table_shareable(svma, vma, addr, idx); | |
3781 | if (saddr) { | |
3782 | spte = huge_pte_offset(svma->vm_mm, saddr); | |
3783 | if (spte) { | |
dc6c9a35 | 3784 | mm_inc_nr_pmds(mm); |
3212b535 SC |
3785 | get_page(virt_to_page(spte)); |
3786 | break; | |
3787 | } | |
3788 | } | |
3789 | } | |
3790 | ||
3791 | if (!spte) | |
3792 | goto out; | |
3793 | ||
cb900f41 KS |
3794 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); |
3795 | spin_lock(ptl); | |
dc6c9a35 | 3796 | if (pud_none(*pud)) { |
3212b535 SC |
3797 | pud_populate(mm, pud, |
3798 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
dc6c9a35 | 3799 | } else { |
3212b535 | 3800 | put_page(virt_to_page(spte)); |
dc6c9a35 KS |
3801 | mm_inc_nr_pmds(mm); |
3802 | } | |
cb900f41 | 3803 | spin_unlock(ptl); |
3212b535 SC |
3804 | out: |
3805 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 3806 | i_mmap_unlock_write(mapping); |
3212b535 SC |
3807 | return pte; |
3808 | } | |
3809 | ||
3810 | /* | |
3811 | * unmap huge page backed by shared pte. | |
3812 | * | |
3813 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
3814 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
3815 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
3816 | * | |
cb900f41 | 3817 | * called with page table lock held. |
3212b535 SC |
3818 | * |
3819 | * returns: 1 successfully unmapped a shared pte page | |
3820 | * 0 the underlying pte page is not shared, or it is the last user | |
3821 | */ | |
3822 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
3823 | { | |
3824 | pgd_t *pgd = pgd_offset(mm, *addr); | |
3825 | pud_t *pud = pud_offset(pgd, *addr); | |
3826 | ||
3827 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
3828 | if (page_count(virt_to_page(ptep)) == 1) | |
3829 | return 0; | |
3830 | ||
3831 | pud_clear(pud); | |
3832 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 3833 | mm_dec_nr_pmds(mm); |
3212b535 SC |
3834 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
3835 | return 1; | |
3836 | } | |
9e5fc74c SC |
3837 | #define want_pmd_share() (1) |
3838 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
3839 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3840 | { | |
3841 | return NULL; | |
3842 | } | |
e81f2d22 ZZ |
3843 | |
3844 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
3845 | { | |
3846 | return 0; | |
3847 | } | |
9e5fc74c | 3848 | #define want_pmd_share() (0) |
3212b535 SC |
3849 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
3850 | ||
9e5fc74c SC |
3851 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
3852 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
3853 | unsigned long addr, unsigned long sz) | |
3854 | { | |
3855 | pgd_t *pgd; | |
3856 | pud_t *pud; | |
3857 | pte_t *pte = NULL; | |
3858 | ||
3859 | pgd = pgd_offset(mm, addr); | |
3860 | pud = pud_alloc(mm, pgd, addr); | |
3861 | if (pud) { | |
3862 | if (sz == PUD_SIZE) { | |
3863 | pte = (pte_t *)pud; | |
3864 | } else { | |
3865 | BUG_ON(sz != PMD_SIZE); | |
3866 | if (want_pmd_share() && pud_none(*pud)) | |
3867 | pte = huge_pmd_share(mm, addr, pud); | |
3868 | else | |
3869 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3870 | } | |
3871 | } | |
3872 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); | |
3873 | ||
3874 | return pte; | |
3875 | } | |
3876 | ||
3877 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
3878 | { | |
3879 | pgd_t *pgd; | |
3880 | pud_t *pud; | |
3881 | pmd_t *pmd = NULL; | |
3882 | ||
3883 | pgd = pgd_offset(mm, addr); | |
3884 | if (pgd_present(*pgd)) { | |
3885 | pud = pud_offset(pgd, addr); | |
3886 | if (pud_present(*pud)) { | |
3887 | if (pud_huge(*pud)) | |
3888 | return (pte_t *)pud; | |
3889 | pmd = pmd_offset(pud, addr); | |
3890 | } | |
3891 | } | |
3892 | return (pte_t *) pmd; | |
3893 | } | |
3894 | ||
61f77eda NH |
3895 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
3896 | ||
3897 | /* | |
3898 | * These functions are overwritable if your architecture needs its own | |
3899 | * behavior. | |
3900 | */ | |
3901 | struct page * __weak | |
3902 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
3903 | int write) | |
3904 | { | |
3905 | return ERR_PTR(-EINVAL); | |
3906 | } | |
3907 | ||
3908 | struct page * __weak | |
9e5fc74c | 3909 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 3910 | pmd_t *pmd, int flags) |
9e5fc74c | 3911 | { |
e66f17ff NH |
3912 | struct page *page = NULL; |
3913 | spinlock_t *ptl; | |
3914 | retry: | |
3915 | ptl = pmd_lockptr(mm, pmd); | |
3916 | spin_lock(ptl); | |
3917 | /* | |
3918 | * make sure that the address range covered by this pmd is not | |
3919 | * unmapped from other threads. | |
3920 | */ | |
3921 | if (!pmd_huge(*pmd)) | |
3922 | goto out; | |
3923 | if (pmd_present(*pmd)) { | |
97534127 | 3924 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
3925 | if (flags & FOLL_GET) |
3926 | get_page(page); | |
3927 | } else { | |
3928 | if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { | |
3929 | spin_unlock(ptl); | |
3930 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
3931 | goto retry; | |
3932 | } | |
3933 | /* | |
3934 | * hwpoisoned entry is treated as no_page_table in | |
3935 | * follow_page_mask(). | |
3936 | */ | |
3937 | } | |
3938 | out: | |
3939 | spin_unlock(ptl); | |
9e5fc74c SC |
3940 | return page; |
3941 | } | |
3942 | ||
61f77eda | 3943 | struct page * __weak |
9e5fc74c | 3944 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 3945 | pud_t *pud, int flags) |
9e5fc74c | 3946 | { |
e66f17ff NH |
3947 | if (flags & FOLL_GET) |
3948 | return NULL; | |
9e5fc74c | 3949 | |
e66f17ff | 3950 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
3951 | } |
3952 | ||
d5bd9106 AK |
3953 | #ifdef CONFIG_MEMORY_FAILURE |
3954 | ||
93f70f90 NH |
3955 | /* |
3956 | * This function is called from memory failure code. | |
3957 | * Assume the caller holds page lock of the head page. | |
3958 | */ | |
6de2b1aa | 3959 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
3960 | { |
3961 | struct hstate *h = page_hstate(hpage); | |
3962 | int nid = page_to_nid(hpage); | |
6de2b1aa | 3963 | int ret = -EBUSY; |
93f70f90 NH |
3964 | |
3965 | spin_lock(&hugetlb_lock); | |
7e1f049e NH |
3966 | /* |
3967 | * Just checking !page_huge_active is not enough, because that could be | |
3968 | * an isolated/hwpoisoned hugepage (which have >0 refcount). | |
3969 | */ | |
3970 | if (!page_huge_active(hpage) && !page_count(hpage)) { | |
56f2fb14 NH |
3971 | /* |
3972 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
3973 | * but dangling hpage->lru can trigger list-debug warnings | |
3974 | * (this happens when we call unpoison_memory() on it), | |
3975 | * so let it point to itself with list_del_init(). | |
3976 | */ | |
3977 | list_del_init(&hpage->lru); | |
8c6c2ecb | 3978 | set_page_refcounted(hpage); |
6de2b1aa NH |
3979 | h->free_huge_pages--; |
3980 | h->free_huge_pages_node[nid]--; | |
3981 | ret = 0; | |
3982 | } | |
93f70f90 | 3983 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 3984 | return ret; |
93f70f90 | 3985 | } |
6de2b1aa | 3986 | #endif |
31caf665 NH |
3987 | |
3988 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
3989 | { | |
bcc54222 NH |
3990 | bool ret = true; |
3991 | ||
309381fe | 3992 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 3993 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
3994 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
3995 | ret = false; | |
3996 | goto unlock; | |
3997 | } | |
3998 | clear_page_huge_active(page); | |
31caf665 | 3999 | list_move_tail(&page->lru, list); |
bcc54222 | 4000 | unlock: |
31caf665 | 4001 | spin_unlock(&hugetlb_lock); |
bcc54222 | 4002 | return ret; |
31caf665 NH |
4003 | } |
4004 | ||
4005 | void putback_active_hugepage(struct page *page) | |
4006 | { | |
309381fe | 4007 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4008 | spin_lock(&hugetlb_lock); |
bcc54222 | 4009 | set_page_huge_active(page); |
31caf665 NH |
4010 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
4011 | spin_unlock(&hugetlb_lock); | |
4012 | put_page(page); | |
4013 | } |