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