]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * linux/fs/nfs/write.c | |
3 | * | |
4 | * Write file data over NFS. | |
5 | * | |
6 | * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de> | |
7 | */ | |
8 | ||
9 | #include <linux/types.h> | |
10 | #include <linux/slab.h> | |
11 | #include <linux/mm.h> | |
12 | #include <linux/pagemap.h> | |
13 | #include <linux/file.h> | |
14 | #include <linux/writeback.h> | |
15 | #include <linux/swap.h> | |
16 | #include <linux/migrate.h> | |
17 | ||
18 | #include <linux/sunrpc/clnt.h> | |
19 | #include <linux/nfs_fs.h> | |
20 | #include <linux/nfs_mount.h> | |
21 | #include <linux/nfs_page.h> | |
22 | #include <linux/backing-dev.h> | |
23 | #include <linux/export.h> | |
24 | ||
25 | #include <asm/uaccess.h> | |
26 | ||
27 | #include "delegation.h" | |
28 | #include "internal.h" | |
29 | #include "iostat.h" | |
30 | #include "nfs4_fs.h" | |
31 | #include "fscache.h" | |
32 | #include "pnfs.h" | |
33 | ||
34 | #include "nfstrace.h" | |
35 | ||
36 | #define NFSDBG_FACILITY NFSDBG_PAGECACHE | |
37 | ||
38 | #define MIN_POOL_WRITE (32) | |
39 | #define MIN_POOL_COMMIT (4) | |
40 | ||
41 | /* | |
42 | * Local function declarations | |
43 | */ | |
44 | static void nfs_redirty_request(struct nfs_page *req); | |
45 | static const struct rpc_call_ops nfs_commit_ops; | |
46 | static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops; | |
47 | static const struct nfs_commit_completion_ops nfs_commit_completion_ops; | |
48 | static const struct nfs_rw_ops nfs_rw_write_ops; | |
49 | static void nfs_clear_request_commit(struct nfs_page *req); | |
50 | static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo, | |
51 | struct inode *inode); | |
52 | static struct nfs_page * | |
53 | nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi, | |
54 | struct page *page); | |
55 | ||
56 | static struct kmem_cache *nfs_wdata_cachep; | |
57 | static mempool_t *nfs_wdata_mempool; | |
58 | static struct kmem_cache *nfs_cdata_cachep; | |
59 | static mempool_t *nfs_commit_mempool; | |
60 | ||
61 | struct nfs_commit_data *nfs_commitdata_alloc(void) | |
62 | { | |
63 | struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO); | |
64 | ||
65 | if (p) { | |
66 | memset(p, 0, sizeof(*p)); | |
67 | INIT_LIST_HEAD(&p->pages); | |
68 | } | |
69 | return p; | |
70 | } | |
71 | EXPORT_SYMBOL_GPL(nfs_commitdata_alloc); | |
72 | ||
73 | void nfs_commit_free(struct nfs_commit_data *p) | |
74 | { | |
75 | mempool_free(p, nfs_commit_mempool); | |
76 | } | |
77 | EXPORT_SYMBOL_GPL(nfs_commit_free); | |
78 | ||
79 | static struct nfs_pgio_header *nfs_writehdr_alloc(void) | |
80 | { | |
81 | struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO); | |
82 | ||
83 | if (p) | |
84 | memset(p, 0, sizeof(*p)); | |
85 | return p; | |
86 | } | |
87 | ||
88 | static void nfs_writehdr_free(struct nfs_pgio_header *hdr) | |
89 | { | |
90 | mempool_free(hdr, nfs_wdata_mempool); | |
91 | } | |
92 | ||
93 | static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error) | |
94 | { | |
95 | ctx->error = error; | |
96 | smp_wmb(); | |
97 | set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags); | |
98 | } | |
99 | ||
100 | /* | |
101 | * nfs_page_find_head_request_locked - find head request associated with @page | |
102 | * | |
103 | * must be called while holding the inode lock. | |
104 | * | |
105 | * returns matching head request with reference held, or NULL if not found. | |
106 | */ | |
107 | static struct nfs_page * | |
108 | nfs_page_find_head_request_locked(struct nfs_inode *nfsi, struct page *page) | |
109 | { | |
110 | struct nfs_page *req = NULL; | |
111 | ||
112 | if (PagePrivate(page)) | |
113 | req = (struct nfs_page *)page_private(page); | |
114 | else if (unlikely(PageSwapCache(page))) | |
115 | req = nfs_page_search_commits_for_head_request_locked(nfsi, | |
116 | page); | |
117 | ||
118 | if (req) { | |
119 | WARN_ON_ONCE(req->wb_head != req); | |
120 | kref_get(&req->wb_kref); | |
121 | } | |
122 | ||
123 | return req; | |
124 | } | |
125 | ||
126 | /* | |
127 | * nfs_page_find_head_request - find head request associated with @page | |
128 | * | |
129 | * returns matching head request with reference held, or NULL if not found. | |
130 | */ | |
131 | static struct nfs_page *nfs_page_find_head_request(struct page *page) | |
132 | { | |
133 | struct inode *inode = page_file_mapping(page)->host; | |
134 | struct nfs_page *req = NULL; | |
135 | ||
136 | spin_lock(&inode->i_lock); | |
137 | req = nfs_page_find_head_request_locked(NFS_I(inode), page); | |
138 | spin_unlock(&inode->i_lock); | |
139 | return req; | |
140 | } | |
141 | ||
142 | /* Adjust the file length if we're writing beyond the end */ | |
143 | static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count) | |
144 | { | |
145 | struct inode *inode = page_file_mapping(page)->host; | |
146 | loff_t end, i_size; | |
147 | pgoff_t end_index; | |
148 | ||
149 | spin_lock(&inode->i_lock); | |
150 | i_size = i_size_read(inode); | |
151 | end_index = (i_size - 1) >> PAGE_CACHE_SHIFT; | |
152 | if (i_size > 0 && page_file_index(page) < end_index) | |
153 | goto out; | |
154 | end = page_file_offset(page) + ((loff_t)offset+count); | |
155 | if (i_size >= end) | |
156 | goto out; | |
157 | i_size_write(inode, end); | |
158 | nfs_inc_stats(inode, NFSIOS_EXTENDWRITE); | |
159 | out: | |
160 | spin_unlock(&inode->i_lock); | |
161 | } | |
162 | ||
163 | /* A writeback failed: mark the page as bad, and invalidate the page cache */ | |
164 | static void nfs_set_pageerror(struct page *page) | |
165 | { | |
166 | nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page)); | |
167 | } | |
168 | ||
169 | /* | |
170 | * nfs_page_group_search_locked | |
171 | * @head - head request of page group | |
172 | * @page_offset - offset into page | |
173 | * | |
174 | * Search page group with head @head to find a request that contains the | |
175 | * page offset @page_offset. | |
176 | * | |
177 | * Returns a pointer to the first matching nfs request, or NULL if no | |
178 | * match is found. | |
179 | * | |
180 | * Must be called with the page group lock held | |
181 | */ | |
182 | static struct nfs_page * | |
183 | nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset) | |
184 | { | |
185 | struct nfs_page *req; | |
186 | ||
187 | WARN_ON_ONCE(head != head->wb_head); | |
188 | WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_head->wb_flags)); | |
189 | ||
190 | req = head; | |
191 | do { | |
192 | if (page_offset >= req->wb_pgbase && | |
193 | page_offset < (req->wb_pgbase + req->wb_bytes)) | |
194 | return req; | |
195 | ||
196 | req = req->wb_this_page; | |
197 | } while (req != head); | |
198 | ||
199 | return NULL; | |
200 | } | |
201 | ||
202 | /* | |
203 | * nfs_page_group_covers_page | |
204 | * @head - head request of page group | |
205 | * | |
206 | * Return true if the page group with head @head covers the whole page, | |
207 | * returns false otherwise | |
208 | */ | |
209 | static bool nfs_page_group_covers_page(struct nfs_page *req) | |
210 | { | |
211 | struct nfs_page *tmp; | |
212 | unsigned int pos = 0; | |
213 | unsigned int len = nfs_page_length(req->wb_page); | |
214 | ||
215 | nfs_page_group_lock(req, false); | |
216 | ||
217 | do { | |
218 | tmp = nfs_page_group_search_locked(req->wb_head, pos); | |
219 | if (tmp) { | |
220 | /* no way this should happen */ | |
221 | WARN_ON_ONCE(tmp->wb_pgbase != pos); | |
222 | pos += tmp->wb_bytes - (pos - tmp->wb_pgbase); | |
223 | } | |
224 | } while (tmp && pos < len); | |
225 | ||
226 | nfs_page_group_unlock(req); | |
227 | WARN_ON_ONCE(pos > len); | |
228 | return pos == len; | |
229 | } | |
230 | ||
231 | /* We can set the PG_uptodate flag if we see that a write request | |
232 | * covers the full page. | |
233 | */ | |
234 | static void nfs_mark_uptodate(struct nfs_page *req) | |
235 | { | |
236 | if (PageUptodate(req->wb_page)) | |
237 | return; | |
238 | if (!nfs_page_group_covers_page(req)) | |
239 | return; | |
240 | SetPageUptodate(req->wb_page); | |
241 | } | |
242 | ||
243 | static int wb_priority(struct writeback_control *wbc) | |
244 | { | |
245 | int ret = 0; | |
246 | if (wbc->for_reclaim) | |
247 | return FLUSH_HIGHPRI | FLUSH_STABLE; | |
248 | if (wbc->sync_mode == WB_SYNC_ALL) | |
249 | ret = FLUSH_COND_STABLE; | |
250 | if (wbc->for_kupdate || wbc->for_background) | |
251 | ret |= FLUSH_LOWPRI; | |
252 | return ret; | |
253 | } | |
254 | ||
255 | /* | |
256 | * NFS congestion control | |
257 | */ | |
258 | ||
259 | int nfs_congestion_kb; | |
260 | ||
261 | #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10)) | |
262 | #define NFS_CONGESTION_OFF_THRESH \ | |
263 | (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2)) | |
264 | ||
265 | static void nfs_set_page_writeback(struct page *page) | |
266 | { | |
267 | struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host); | |
268 | int ret = test_set_page_writeback(page); | |
269 | ||
270 | WARN_ON_ONCE(ret != 0); | |
271 | ||
272 | if (atomic_long_inc_return(&nfss->writeback) > | |
273 | NFS_CONGESTION_ON_THRESH) { | |
274 | set_bdi_congested(&nfss->backing_dev_info, | |
275 | BLK_RW_ASYNC); | |
276 | } | |
277 | } | |
278 | ||
279 | static void nfs_end_page_writeback(struct nfs_page *req) | |
280 | { | |
281 | struct inode *inode = page_file_mapping(req->wb_page)->host; | |
282 | struct nfs_server *nfss = NFS_SERVER(inode); | |
283 | ||
284 | if (!nfs_page_group_sync_on_bit(req, PG_WB_END)) | |
285 | return; | |
286 | ||
287 | end_page_writeback(req->wb_page); | |
288 | if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) | |
289 | clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC); | |
290 | } | |
291 | ||
292 | ||
293 | /* nfs_page_group_clear_bits | |
294 | * @req - an nfs request | |
295 | * clears all page group related bits from @req | |
296 | */ | |
297 | static void | |
298 | nfs_page_group_clear_bits(struct nfs_page *req) | |
299 | { | |
300 | clear_bit(PG_TEARDOWN, &req->wb_flags); | |
301 | clear_bit(PG_UNLOCKPAGE, &req->wb_flags); | |
302 | clear_bit(PG_UPTODATE, &req->wb_flags); | |
303 | clear_bit(PG_WB_END, &req->wb_flags); | |
304 | clear_bit(PG_REMOVE, &req->wb_flags); | |
305 | } | |
306 | ||
307 | ||
308 | /* | |
309 | * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req | |
310 | * | |
311 | * this is a helper function for nfs_lock_and_join_requests | |
312 | * | |
313 | * @inode - inode associated with request page group, must be holding inode lock | |
314 | * @head - head request of page group, must be holding head lock | |
315 | * @req - request that couldn't lock and needs to wait on the req bit lock | |
316 | * @nonblock - if true, don't actually wait | |
317 | * | |
318 | * NOTE: this must be called holding page_group bit lock and inode spin lock | |
319 | * and BOTH will be released before returning. | |
320 | * | |
321 | * returns 0 on success, < 0 on error. | |
322 | */ | |
323 | static int | |
324 | nfs_unroll_locks_and_wait(struct inode *inode, struct nfs_page *head, | |
325 | struct nfs_page *req, bool nonblock) | |
326 | __releases(&inode->i_lock) | |
327 | { | |
328 | struct nfs_page *tmp; | |
329 | int ret; | |
330 | ||
331 | /* relinquish all the locks successfully grabbed this run */ | |
332 | for (tmp = head ; tmp != req; tmp = tmp->wb_this_page) | |
333 | nfs_unlock_request(tmp); | |
334 | ||
335 | WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags)); | |
336 | ||
337 | /* grab a ref on the request that will be waited on */ | |
338 | kref_get(&req->wb_kref); | |
339 | ||
340 | nfs_page_group_unlock(head); | |
341 | spin_unlock(&inode->i_lock); | |
342 | ||
343 | /* release ref from nfs_page_find_head_request_locked */ | |
344 | nfs_release_request(head); | |
345 | ||
346 | if (!nonblock) | |
347 | ret = nfs_wait_on_request(req); | |
348 | else | |
349 | ret = -EAGAIN; | |
350 | nfs_release_request(req); | |
351 | ||
352 | return ret; | |
353 | } | |
354 | ||
355 | /* | |
356 | * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests | |
357 | * | |
358 | * @destroy_list - request list (using wb_this_page) terminated by @old_head | |
359 | * @old_head - the old head of the list | |
360 | * | |
361 | * All subrequests must be locked and removed from all lists, so at this point | |
362 | * they are only "active" in this function, and possibly in nfs_wait_on_request | |
363 | * with a reference held by some other context. | |
364 | */ | |
365 | static void | |
366 | nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list, | |
367 | struct nfs_page *old_head) | |
368 | { | |
369 | while (destroy_list) { | |
370 | struct nfs_page *subreq = destroy_list; | |
371 | ||
372 | destroy_list = (subreq->wb_this_page == old_head) ? | |
373 | NULL : subreq->wb_this_page; | |
374 | ||
375 | WARN_ON_ONCE(old_head != subreq->wb_head); | |
376 | ||
377 | /* make sure old group is not used */ | |
378 | subreq->wb_head = subreq; | |
379 | subreq->wb_this_page = subreq; | |
380 | ||
381 | /* subreq is now totally disconnected from page group or any | |
382 | * write / commit lists. last chance to wake any waiters */ | |
383 | nfs_unlock_request(subreq); | |
384 | ||
385 | if (!test_bit(PG_TEARDOWN, &subreq->wb_flags)) { | |
386 | /* release ref on old head request */ | |
387 | nfs_release_request(old_head); | |
388 | ||
389 | nfs_page_group_clear_bits(subreq); | |
390 | ||
391 | /* release the PG_INODE_REF reference */ | |
392 | if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) | |
393 | nfs_release_request(subreq); | |
394 | else | |
395 | WARN_ON_ONCE(1); | |
396 | } else { | |
397 | WARN_ON_ONCE(test_bit(PG_CLEAN, &subreq->wb_flags)); | |
398 | /* zombie requests have already released the last | |
399 | * reference and were waiting on the rest of the | |
400 | * group to complete. Since it's no longer part of a | |
401 | * group, simply free the request */ | |
402 | nfs_page_group_clear_bits(subreq); | |
403 | nfs_free_request(subreq); | |
404 | } | |
405 | } | |
406 | } | |
407 | ||
408 | /* | |
409 | * nfs_lock_and_join_requests - join all subreqs to the head req and return | |
410 | * a locked reference, cancelling any pending | |
411 | * operations for this page. | |
412 | * | |
413 | * @page - the page used to lookup the "page group" of nfs_page structures | |
414 | * @nonblock - if true, don't block waiting for request locks | |
415 | * | |
416 | * This function joins all sub requests to the head request by first | |
417 | * locking all requests in the group, cancelling any pending operations | |
418 | * and finally updating the head request to cover the whole range covered by | |
419 | * the (former) group. All subrequests are removed from any write or commit | |
420 | * lists, unlinked from the group and destroyed. | |
421 | * | |
422 | * Returns a locked, referenced pointer to the head request - which after | |
423 | * this call is guaranteed to be the only request associated with the page. | |
424 | * Returns NULL if no requests are found for @page, or a ERR_PTR if an | |
425 | * error was encountered. | |
426 | */ | |
427 | static struct nfs_page * | |
428 | nfs_lock_and_join_requests(struct page *page, bool nonblock) | |
429 | { | |
430 | struct inode *inode = page_file_mapping(page)->host; | |
431 | struct nfs_page *head, *subreq; | |
432 | struct nfs_page *destroy_list = NULL; | |
433 | unsigned int total_bytes; | |
434 | int ret; | |
435 | ||
436 | try_again: | |
437 | total_bytes = 0; | |
438 | ||
439 | WARN_ON_ONCE(destroy_list); | |
440 | ||
441 | spin_lock(&inode->i_lock); | |
442 | ||
443 | /* | |
444 | * A reference is taken only on the head request which acts as a | |
445 | * reference to the whole page group - the group will not be destroyed | |
446 | * until the head reference is released. | |
447 | */ | |
448 | head = nfs_page_find_head_request_locked(NFS_I(inode), page); | |
449 | ||
450 | if (!head) { | |
451 | spin_unlock(&inode->i_lock); | |
452 | return NULL; | |
453 | } | |
454 | ||
455 | /* holding inode lock, so always make a non-blocking call to try the | |
456 | * page group lock */ | |
457 | ret = nfs_page_group_lock(head, true); | |
458 | if (ret < 0) { | |
459 | spin_unlock(&inode->i_lock); | |
460 | ||
461 | if (!nonblock && ret == -EAGAIN) { | |
462 | nfs_page_group_lock_wait(head); | |
463 | nfs_release_request(head); | |
464 | goto try_again; | |
465 | } | |
466 | ||
467 | nfs_release_request(head); | |
468 | return ERR_PTR(ret); | |
469 | } | |
470 | ||
471 | /* lock each request in the page group */ | |
472 | subreq = head; | |
473 | do { | |
474 | /* | |
475 | * Subrequests are always contiguous, non overlapping | |
476 | * and in order - but may be repeated (mirrored writes). | |
477 | */ | |
478 | if (subreq->wb_offset == (head->wb_offset + total_bytes)) { | |
479 | /* keep track of how many bytes this group covers */ | |
480 | total_bytes += subreq->wb_bytes; | |
481 | } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset || | |
482 | ((subreq->wb_offset + subreq->wb_bytes) > | |
483 | (head->wb_offset + total_bytes)))) { | |
484 | nfs_page_group_unlock(head); | |
485 | spin_unlock(&inode->i_lock); | |
486 | return ERR_PTR(-EIO); | |
487 | } | |
488 | ||
489 | if (!nfs_lock_request(subreq)) { | |
490 | /* releases page group bit lock and | |
491 | * inode spin lock and all references */ | |
492 | ret = nfs_unroll_locks_and_wait(inode, head, | |
493 | subreq, nonblock); | |
494 | ||
495 | if (ret == 0) | |
496 | goto try_again; | |
497 | ||
498 | return ERR_PTR(ret); | |
499 | } | |
500 | ||
501 | subreq = subreq->wb_this_page; | |
502 | } while (subreq != head); | |
503 | ||
504 | /* Now that all requests are locked, make sure they aren't on any list. | |
505 | * Commit list removal accounting is done after locks are dropped */ | |
506 | subreq = head; | |
507 | do { | |
508 | nfs_clear_request_commit(subreq); | |
509 | subreq = subreq->wb_this_page; | |
510 | } while (subreq != head); | |
511 | ||
512 | /* unlink subrequests from head, destroy them later */ | |
513 | if (head->wb_this_page != head) { | |
514 | /* destroy list will be terminated by head */ | |
515 | destroy_list = head->wb_this_page; | |
516 | head->wb_this_page = head; | |
517 | ||
518 | /* change head request to cover whole range that | |
519 | * the former page group covered */ | |
520 | head->wb_bytes = total_bytes; | |
521 | } | |
522 | ||
523 | /* | |
524 | * prepare head request to be added to new pgio descriptor | |
525 | */ | |
526 | nfs_page_group_clear_bits(head); | |
527 | ||
528 | /* | |
529 | * some part of the group was still on the inode list - otherwise | |
530 | * the group wouldn't be involved in async write. | |
531 | * grab a reference for the head request, iff it needs one. | |
532 | */ | |
533 | if (!test_and_set_bit(PG_INODE_REF, &head->wb_flags)) | |
534 | kref_get(&head->wb_kref); | |
535 | ||
536 | nfs_page_group_unlock(head); | |
537 | ||
538 | /* drop lock to clean uprequests on destroy list */ | |
539 | spin_unlock(&inode->i_lock); | |
540 | ||
541 | nfs_destroy_unlinked_subrequests(destroy_list, head); | |
542 | ||
543 | /* still holds ref on head from nfs_page_find_head_request_locked | |
544 | * and still has lock on head from lock loop */ | |
545 | return head; | |
546 | } | |
547 | ||
548 | /* | |
549 | * Find an associated nfs write request, and prepare to flush it out | |
550 | * May return an error if the user signalled nfs_wait_on_request(). | |
551 | */ | |
552 | static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio, | |
553 | struct page *page, bool nonblock) | |
554 | { | |
555 | struct nfs_page *req; | |
556 | int ret = 0; | |
557 | ||
558 | req = nfs_lock_and_join_requests(page, nonblock); | |
559 | if (!req) | |
560 | goto out; | |
561 | ret = PTR_ERR(req); | |
562 | if (IS_ERR(req)) | |
563 | goto out; | |
564 | ||
565 | nfs_set_page_writeback(page); | |
566 | WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags)); | |
567 | ||
568 | ret = 0; | |
569 | if (!nfs_pageio_add_request(pgio, req)) { | |
570 | nfs_redirty_request(req); | |
571 | ret = pgio->pg_error; | |
572 | } | |
573 | out: | |
574 | return ret; | |
575 | } | |
576 | ||
577 | static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio) | |
578 | { | |
579 | struct inode *inode = page_file_mapping(page)->host; | |
580 | int ret; | |
581 | ||
582 | nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE); | |
583 | nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1); | |
584 | ||
585 | nfs_pageio_cond_complete(pgio, page_file_index(page)); | |
586 | ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE); | |
587 | if (ret == -EAGAIN) { | |
588 | redirty_page_for_writepage(wbc, page); | |
589 | ret = 0; | |
590 | } | |
591 | return ret; | |
592 | } | |
593 | ||
594 | /* | |
595 | * Write an mmapped page to the server. | |
596 | */ | |
597 | static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc) | |
598 | { | |
599 | struct nfs_pageio_descriptor pgio; | |
600 | int err; | |
601 | ||
602 | nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc), | |
603 | false, &nfs_async_write_completion_ops); | |
604 | err = nfs_do_writepage(page, wbc, &pgio); | |
605 | nfs_pageio_complete(&pgio); | |
606 | if (err < 0) | |
607 | return err; | |
608 | if (pgio.pg_error < 0) | |
609 | return pgio.pg_error; | |
610 | return 0; | |
611 | } | |
612 | ||
613 | int nfs_writepage(struct page *page, struct writeback_control *wbc) | |
614 | { | |
615 | int ret; | |
616 | ||
617 | ret = nfs_writepage_locked(page, wbc); | |
618 | unlock_page(page); | |
619 | return ret; | |
620 | } | |
621 | ||
622 | static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data) | |
623 | { | |
624 | int ret; | |
625 | ||
626 | ret = nfs_do_writepage(page, wbc, data); | |
627 | unlock_page(page); | |
628 | return ret; | |
629 | } | |
630 | ||
631 | int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc) | |
632 | { | |
633 | struct inode *inode = mapping->host; | |
634 | unsigned long *bitlock = &NFS_I(inode)->flags; | |
635 | struct nfs_pageio_descriptor pgio; | |
636 | int err; | |
637 | ||
638 | /* Stop dirtying of new pages while we sync */ | |
639 | err = wait_on_bit_lock_action(bitlock, NFS_INO_FLUSHING, | |
640 | nfs_wait_bit_killable, TASK_KILLABLE); | |
641 | if (err) | |
642 | goto out_err; | |
643 | ||
644 | nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES); | |
645 | ||
646 | nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false, | |
647 | &nfs_async_write_completion_ops); | |
648 | err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio); | |
649 | nfs_pageio_complete(&pgio); | |
650 | ||
651 | clear_bit_unlock(NFS_INO_FLUSHING, bitlock); | |
652 | smp_mb__after_atomic(); | |
653 | wake_up_bit(bitlock, NFS_INO_FLUSHING); | |
654 | ||
655 | if (err < 0) | |
656 | goto out_err; | |
657 | err = pgio.pg_error; | |
658 | if (err < 0) | |
659 | goto out_err; | |
660 | return 0; | |
661 | out_err: | |
662 | return err; | |
663 | } | |
664 | ||
665 | /* | |
666 | * Insert a write request into an inode | |
667 | */ | |
668 | static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req) | |
669 | { | |
670 | struct nfs_inode *nfsi = NFS_I(inode); | |
671 | ||
672 | WARN_ON_ONCE(req->wb_this_page != req); | |
673 | ||
674 | /* Lock the request! */ | |
675 | nfs_lock_request(req); | |
676 | ||
677 | spin_lock(&inode->i_lock); | |
678 | if (!nfsi->nrequests && | |
679 | NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE)) | |
680 | inode->i_version++; | |
681 | /* | |
682 | * Swap-space should not get truncated. Hence no need to plug the race | |
683 | * with invalidate/truncate. | |
684 | */ | |
685 | if (likely(!PageSwapCache(req->wb_page))) { | |
686 | set_bit(PG_MAPPED, &req->wb_flags); | |
687 | SetPagePrivate(req->wb_page); | |
688 | set_page_private(req->wb_page, (unsigned long)req); | |
689 | } | |
690 | nfsi->nrequests++; | |
691 | /* this a head request for a page group - mark it as having an | |
692 | * extra reference so sub groups can follow suit. | |
693 | * This flag also informs pgio layer when to bump nrequests when | |
694 | * adding subrequests. */ | |
695 | WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags)); | |
696 | kref_get(&req->wb_kref); | |
697 | spin_unlock(&inode->i_lock); | |
698 | } | |
699 | ||
700 | /* | |
701 | * Remove a write request from an inode | |
702 | */ | |
703 | static void nfs_inode_remove_request(struct nfs_page *req) | |
704 | { | |
705 | struct inode *inode = d_inode(req->wb_context->dentry); | |
706 | struct nfs_inode *nfsi = NFS_I(inode); | |
707 | struct nfs_page *head; | |
708 | ||
709 | if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) { | |
710 | head = req->wb_head; | |
711 | ||
712 | spin_lock(&inode->i_lock); | |
713 | if (likely(!PageSwapCache(head->wb_page))) { | |
714 | set_page_private(head->wb_page, 0); | |
715 | ClearPagePrivate(head->wb_page); | |
716 | smp_mb__after_atomic(); | |
717 | wake_up_page(head->wb_page, PG_private); | |
718 | clear_bit(PG_MAPPED, &head->wb_flags); | |
719 | } | |
720 | nfsi->nrequests--; | |
721 | spin_unlock(&inode->i_lock); | |
722 | } else { | |
723 | spin_lock(&inode->i_lock); | |
724 | nfsi->nrequests--; | |
725 | spin_unlock(&inode->i_lock); | |
726 | } | |
727 | ||
728 | if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) | |
729 | nfs_release_request(req); | |
730 | } | |
731 | ||
732 | static void | |
733 | nfs_mark_request_dirty(struct nfs_page *req) | |
734 | { | |
735 | __set_page_dirty_nobuffers(req->wb_page); | |
736 | } | |
737 | ||
738 | /* | |
739 | * nfs_page_search_commits_for_head_request_locked | |
740 | * | |
741 | * Search through commit lists on @inode for the head request for @page. | |
742 | * Must be called while holding the inode (which is cinfo) lock. | |
743 | * | |
744 | * Returns the head request if found, or NULL if not found. | |
745 | */ | |
746 | static struct nfs_page * | |
747 | nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi, | |
748 | struct page *page) | |
749 | { | |
750 | struct nfs_page *freq, *t; | |
751 | struct nfs_commit_info cinfo; | |
752 | struct inode *inode = &nfsi->vfs_inode; | |
753 | ||
754 | nfs_init_cinfo_from_inode(&cinfo, inode); | |
755 | ||
756 | /* search through pnfs commit lists */ | |
757 | freq = pnfs_search_commit_reqs(inode, &cinfo, page); | |
758 | if (freq) | |
759 | return freq->wb_head; | |
760 | ||
761 | /* Linearly search the commit list for the correct request */ | |
762 | list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) { | |
763 | if (freq->wb_page == page) | |
764 | return freq->wb_head; | |
765 | } | |
766 | ||
767 | return NULL; | |
768 | } | |
769 | ||
770 | /** | |
771 | * nfs_request_add_commit_list - add request to a commit list | |
772 | * @req: pointer to a struct nfs_page | |
773 | * @dst: commit list head | |
774 | * @cinfo: holds list lock and accounting info | |
775 | * | |
776 | * This sets the PG_CLEAN bit, updates the cinfo count of | |
777 | * number of outstanding requests requiring a commit as well as | |
778 | * the MM page stats. | |
779 | * | |
780 | * The caller must _not_ hold the cinfo->lock, but must be | |
781 | * holding the nfs_page lock. | |
782 | */ | |
783 | void | |
784 | nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst, | |
785 | struct nfs_commit_info *cinfo) | |
786 | { | |
787 | set_bit(PG_CLEAN, &(req)->wb_flags); | |
788 | spin_lock(cinfo->lock); | |
789 | nfs_list_add_request(req, dst); | |
790 | cinfo->mds->ncommit++; | |
791 | spin_unlock(cinfo->lock); | |
792 | if (!cinfo->dreq) | |
793 | nfs_mark_page_unstable(req->wb_page); | |
794 | } | |
795 | EXPORT_SYMBOL_GPL(nfs_request_add_commit_list); | |
796 | ||
797 | /** | |
798 | * nfs_request_remove_commit_list - Remove request from a commit list | |
799 | * @req: pointer to a nfs_page | |
800 | * @cinfo: holds list lock and accounting info | |
801 | * | |
802 | * This clears the PG_CLEAN bit, and updates the cinfo's count of | |
803 | * number of outstanding requests requiring a commit | |
804 | * It does not update the MM page stats. | |
805 | * | |
806 | * The caller _must_ hold the cinfo->lock and the nfs_page lock. | |
807 | */ | |
808 | void | |
809 | nfs_request_remove_commit_list(struct nfs_page *req, | |
810 | struct nfs_commit_info *cinfo) | |
811 | { | |
812 | if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) | |
813 | return; | |
814 | nfs_list_remove_request(req); | |
815 | cinfo->mds->ncommit--; | |
816 | } | |
817 | EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list); | |
818 | ||
819 | static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo, | |
820 | struct inode *inode) | |
821 | { | |
822 | cinfo->lock = &inode->i_lock; | |
823 | cinfo->mds = &NFS_I(inode)->commit_info; | |
824 | cinfo->ds = pnfs_get_ds_info(inode); | |
825 | cinfo->dreq = NULL; | |
826 | cinfo->completion_ops = &nfs_commit_completion_ops; | |
827 | } | |
828 | ||
829 | void nfs_init_cinfo(struct nfs_commit_info *cinfo, | |
830 | struct inode *inode, | |
831 | struct nfs_direct_req *dreq) | |
832 | { | |
833 | if (dreq) | |
834 | nfs_init_cinfo_from_dreq(cinfo, dreq); | |
835 | else | |
836 | nfs_init_cinfo_from_inode(cinfo, inode); | |
837 | } | |
838 | EXPORT_SYMBOL_GPL(nfs_init_cinfo); | |
839 | ||
840 | /* | |
841 | * Add a request to the inode's commit list. | |
842 | */ | |
843 | void | |
844 | nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg, | |
845 | struct nfs_commit_info *cinfo, u32 ds_commit_idx) | |
846 | { | |
847 | if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx)) | |
848 | return; | |
849 | nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo); | |
850 | } | |
851 | ||
852 | static void | |
853 | nfs_clear_page_commit(struct page *page) | |
854 | { | |
855 | dec_zone_page_state(page, NR_UNSTABLE_NFS); | |
856 | dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb, | |
857 | WB_RECLAIMABLE); | |
858 | } | |
859 | ||
860 | /* Called holding inode (/cinfo) lock */ | |
861 | static void | |
862 | nfs_clear_request_commit(struct nfs_page *req) | |
863 | { | |
864 | if (test_bit(PG_CLEAN, &req->wb_flags)) { | |
865 | struct inode *inode = d_inode(req->wb_context->dentry); | |
866 | struct nfs_commit_info cinfo; | |
867 | ||
868 | nfs_init_cinfo_from_inode(&cinfo, inode); | |
869 | if (!pnfs_clear_request_commit(req, &cinfo)) { | |
870 | nfs_request_remove_commit_list(req, &cinfo); | |
871 | } | |
872 | nfs_clear_page_commit(req->wb_page); | |
873 | } | |
874 | } | |
875 | ||
876 | int nfs_write_need_commit(struct nfs_pgio_header *hdr) | |
877 | { | |
878 | if (hdr->verf.committed == NFS_DATA_SYNC) | |
879 | return hdr->lseg == NULL; | |
880 | return hdr->verf.committed != NFS_FILE_SYNC; | |
881 | } | |
882 | ||
883 | static void nfs_write_completion(struct nfs_pgio_header *hdr) | |
884 | { | |
885 | struct nfs_commit_info cinfo; | |
886 | unsigned long bytes = 0; | |
887 | ||
888 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) | |
889 | goto out; | |
890 | nfs_init_cinfo_from_inode(&cinfo, hdr->inode); | |
891 | while (!list_empty(&hdr->pages)) { | |
892 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
893 | ||
894 | bytes += req->wb_bytes; | |
895 | nfs_list_remove_request(req); | |
896 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && | |
897 | (hdr->good_bytes < bytes)) { | |
898 | nfs_set_pageerror(req->wb_page); | |
899 | nfs_context_set_write_error(req->wb_context, hdr->error); | |
900 | goto remove_req; | |
901 | } | |
902 | if (nfs_write_need_commit(hdr)) { | |
903 | memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf)); | |
904 | nfs_mark_request_commit(req, hdr->lseg, &cinfo, | |
905 | hdr->pgio_mirror_idx); | |
906 | goto next; | |
907 | } | |
908 | remove_req: | |
909 | nfs_inode_remove_request(req); | |
910 | next: | |
911 | nfs_unlock_request(req); | |
912 | nfs_end_page_writeback(req); | |
913 | nfs_release_request(req); | |
914 | } | |
915 | out: | |
916 | hdr->release(hdr); | |
917 | } | |
918 | ||
919 | unsigned long | |
920 | nfs_reqs_to_commit(struct nfs_commit_info *cinfo) | |
921 | { | |
922 | return cinfo->mds->ncommit; | |
923 | } | |
924 | ||
925 | /* cinfo->lock held by caller */ | |
926 | int | |
927 | nfs_scan_commit_list(struct list_head *src, struct list_head *dst, | |
928 | struct nfs_commit_info *cinfo, int max) | |
929 | { | |
930 | struct nfs_page *req, *tmp; | |
931 | int ret = 0; | |
932 | ||
933 | list_for_each_entry_safe(req, tmp, src, wb_list) { | |
934 | if (!nfs_lock_request(req)) | |
935 | continue; | |
936 | kref_get(&req->wb_kref); | |
937 | if (cond_resched_lock(cinfo->lock)) | |
938 | list_safe_reset_next(req, tmp, wb_list); | |
939 | nfs_request_remove_commit_list(req, cinfo); | |
940 | nfs_list_add_request(req, dst); | |
941 | ret++; | |
942 | if ((ret == max) && !cinfo->dreq) | |
943 | break; | |
944 | } | |
945 | return ret; | |
946 | } | |
947 | ||
948 | /* | |
949 | * nfs_scan_commit - Scan an inode for commit requests | |
950 | * @inode: NFS inode to scan | |
951 | * @dst: mds destination list | |
952 | * @cinfo: mds and ds lists of reqs ready to commit | |
953 | * | |
954 | * Moves requests from the inode's 'commit' request list. | |
955 | * The requests are *not* checked to ensure that they form a contiguous set. | |
956 | */ | |
957 | int | |
958 | nfs_scan_commit(struct inode *inode, struct list_head *dst, | |
959 | struct nfs_commit_info *cinfo) | |
960 | { | |
961 | int ret = 0; | |
962 | ||
963 | spin_lock(cinfo->lock); | |
964 | if (cinfo->mds->ncommit > 0) { | |
965 | const int max = INT_MAX; | |
966 | ||
967 | ret = nfs_scan_commit_list(&cinfo->mds->list, dst, | |
968 | cinfo, max); | |
969 | ret += pnfs_scan_commit_lists(inode, cinfo, max - ret); | |
970 | } | |
971 | spin_unlock(cinfo->lock); | |
972 | return ret; | |
973 | } | |
974 | ||
975 | /* | |
976 | * Search for an existing write request, and attempt to update | |
977 | * it to reflect a new dirty region on a given page. | |
978 | * | |
979 | * If the attempt fails, then the existing request is flushed out | |
980 | * to disk. | |
981 | */ | |
982 | static struct nfs_page *nfs_try_to_update_request(struct inode *inode, | |
983 | struct page *page, | |
984 | unsigned int offset, | |
985 | unsigned int bytes) | |
986 | { | |
987 | struct nfs_page *req; | |
988 | unsigned int rqend; | |
989 | unsigned int end; | |
990 | int error; | |
991 | ||
992 | if (!PagePrivate(page)) | |
993 | return NULL; | |
994 | ||
995 | end = offset + bytes; | |
996 | spin_lock(&inode->i_lock); | |
997 | ||
998 | for (;;) { | |
999 | req = nfs_page_find_head_request_locked(NFS_I(inode), page); | |
1000 | if (req == NULL) | |
1001 | goto out_unlock; | |
1002 | ||
1003 | /* should be handled by nfs_flush_incompatible */ | |
1004 | WARN_ON_ONCE(req->wb_head != req); | |
1005 | WARN_ON_ONCE(req->wb_this_page != req); | |
1006 | ||
1007 | rqend = req->wb_offset + req->wb_bytes; | |
1008 | /* | |
1009 | * Tell the caller to flush out the request if | |
1010 | * the offsets are non-contiguous. | |
1011 | * Note: nfs_flush_incompatible() will already | |
1012 | * have flushed out requests having wrong owners. | |
1013 | */ | |
1014 | if (offset > rqend | |
1015 | || end < req->wb_offset) | |
1016 | goto out_flushme; | |
1017 | ||
1018 | if (nfs_lock_request(req)) | |
1019 | break; | |
1020 | ||
1021 | /* The request is locked, so wait and then retry */ | |
1022 | spin_unlock(&inode->i_lock); | |
1023 | error = nfs_wait_on_request(req); | |
1024 | nfs_release_request(req); | |
1025 | if (error != 0) | |
1026 | goto out_err; | |
1027 | spin_lock(&inode->i_lock); | |
1028 | } | |
1029 | ||
1030 | /* Okay, the request matches. Update the region */ | |
1031 | if (offset < req->wb_offset) { | |
1032 | req->wb_offset = offset; | |
1033 | req->wb_pgbase = offset; | |
1034 | } | |
1035 | if (end > rqend) | |
1036 | req->wb_bytes = end - req->wb_offset; | |
1037 | else | |
1038 | req->wb_bytes = rqend - req->wb_offset; | |
1039 | out_unlock: | |
1040 | if (req) | |
1041 | nfs_clear_request_commit(req); | |
1042 | spin_unlock(&inode->i_lock); | |
1043 | return req; | |
1044 | out_flushme: | |
1045 | spin_unlock(&inode->i_lock); | |
1046 | nfs_release_request(req); | |
1047 | error = nfs_wb_page(inode, page); | |
1048 | out_err: | |
1049 | return ERR_PTR(error); | |
1050 | } | |
1051 | ||
1052 | /* | |
1053 | * Try to update an existing write request, or create one if there is none. | |
1054 | * | |
1055 | * Note: Should always be called with the Page Lock held to prevent races | |
1056 | * if we have to add a new request. Also assumes that the caller has | |
1057 | * already called nfs_flush_incompatible() if necessary. | |
1058 | */ | |
1059 | static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx, | |
1060 | struct page *page, unsigned int offset, unsigned int bytes) | |
1061 | { | |
1062 | struct inode *inode = page_file_mapping(page)->host; | |
1063 | struct nfs_page *req; | |
1064 | ||
1065 | req = nfs_try_to_update_request(inode, page, offset, bytes); | |
1066 | if (req != NULL) | |
1067 | goto out; | |
1068 | req = nfs_create_request(ctx, page, NULL, offset, bytes); | |
1069 | if (IS_ERR(req)) | |
1070 | goto out; | |
1071 | nfs_inode_add_request(inode, req); | |
1072 | out: | |
1073 | return req; | |
1074 | } | |
1075 | ||
1076 | static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page, | |
1077 | unsigned int offset, unsigned int count) | |
1078 | { | |
1079 | struct nfs_page *req; | |
1080 | ||
1081 | req = nfs_setup_write_request(ctx, page, offset, count); | |
1082 | if (IS_ERR(req)) | |
1083 | return PTR_ERR(req); | |
1084 | /* Update file length */ | |
1085 | nfs_grow_file(page, offset, count); | |
1086 | nfs_mark_uptodate(req); | |
1087 | nfs_mark_request_dirty(req); | |
1088 | nfs_unlock_and_release_request(req); | |
1089 | return 0; | |
1090 | } | |
1091 | ||
1092 | int nfs_flush_incompatible(struct file *file, struct page *page) | |
1093 | { | |
1094 | struct nfs_open_context *ctx = nfs_file_open_context(file); | |
1095 | struct nfs_lock_context *l_ctx; | |
1096 | struct file_lock_context *flctx = file_inode(file)->i_flctx; | |
1097 | struct nfs_page *req; | |
1098 | int do_flush, status; | |
1099 | /* | |
1100 | * Look for a request corresponding to this page. If there | |
1101 | * is one, and it belongs to another file, we flush it out | |
1102 | * before we try to copy anything into the page. Do this | |
1103 | * due to the lack of an ACCESS-type call in NFSv2. | |
1104 | * Also do the same if we find a request from an existing | |
1105 | * dropped page. | |
1106 | */ | |
1107 | do { | |
1108 | req = nfs_page_find_head_request(page); | |
1109 | if (req == NULL) | |
1110 | return 0; | |
1111 | l_ctx = req->wb_lock_context; | |
1112 | do_flush = req->wb_page != page || req->wb_context != ctx; | |
1113 | /* for now, flush if more than 1 request in page_group */ | |
1114 | do_flush |= req->wb_this_page != req; | |
1115 | if (l_ctx && flctx && | |
1116 | !(list_empty_careful(&flctx->flc_posix) && | |
1117 | list_empty_careful(&flctx->flc_flock))) { | |
1118 | do_flush |= l_ctx->lockowner.l_owner != current->files | |
1119 | || l_ctx->lockowner.l_pid != current->tgid; | |
1120 | } | |
1121 | nfs_release_request(req); | |
1122 | if (!do_flush) | |
1123 | return 0; | |
1124 | status = nfs_wb_page(page_file_mapping(page)->host, page); | |
1125 | } while (status == 0); | |
1126 | return status; | |
1127 | } | |
1128 | ||
1129 | /* | |
1130 | * Avoid buffered writes when a open context credential's key would | |
1131 | * expire soon. | |
1132 | * | |
1133 | * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL. | |
1134 | * | |
1135 | * Return 0 and set a credential flag which triggers the inode to flush | |
1136 | * and performs NFS_FILE_SYNC writes if the key will expired within | |
1137 | * RPC_KEY_EXPIRE_TIMEO. | |
1138 | */ | |
1139 | int | |
1140 | nfs_key_timeout_notify(struct file *filp, struct inode *inode) | |
1141 | { | |
1142 | struct nfs_open_context *ctx = nfs_file_open_context(filp); | |
1143 | struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth; | |
1144 | ||
1145 | return rpcauth_key_timeout_notify(auth, ctx->cred); | |
1146 | } | |
1147 | ||
1148 | /* | |
1149 | * Test if the open context credential key is marked to expire soon. | |
1150 | */ | |
1151 | bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx) | |
1152 | { | |
1153 | return rpcauth_cred_key_to_expire(ctx->cred); | |
1154 | } | |
1155 | ||
1156 | /* | |
1157 | * If the page cache is marked as unsafe or invalid, then we can't rely on | |
1158 | * the PageUptodate() flag. In this case, we will need to turn off | |
1159 | * write optimisations that depend on the page contents being correct. | |
1160 | */ | |
1161 | static bool nfs_write_pageuptodate(struct page *page, struct inode *inode) | |
1162 | { | |
1163 | struct nfs_inode *nfsi = NFS_I(inode); | |
1164 | ||
1165 | if (nfs_have_delegated_attributes(inode)) | |
1166 | goto out; | |
1167 | if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE) | |
1168 | return false; | |
1169 | smp_rmb(); | |
1170 | if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags)) | |
1171 | return false; | |
1172 | out: | |
1173 | if (nfsi->cache_validity & NFS_INO_INVALID_DATA) | |
1174 | return false; | |
1175 | return PageUptodate(page) != 0; | |
1176 | } | |
1177 | ||
1178 | static bool | |
1179 | is_whole_file_wrlock(struct file_lock *fl) | |
1180 | { | |
1181 | return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX && | |
1182 | fl->fl_type == F_WRLCK; | |
1183 | } | |
1184 | ||
1185 | /* If we know the page is up to date, and we're not using byte range locks (or | |
1186 | * if we have the whole file locked for writing), it may be more efficient to | |
1187 | * extend the write to cover the entire page in order to avoid fragmentation | |
1188 | * inefficiencies. | |
1189 | * | |
1190 | * If the file is opened for synchronous writes then we can just skip the rest | |
1191 | * of the checks. | |
1192 | */ | |
1193 | static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode) | |
1194 | { | |
1195 | int ret; | |
1196 | struct file_lock_context *flctx = inode->i_flctx; | |
1197 | struct file_lock *fl; | |
1198 | ||
1199 | if (file->f_flags & O_DSYNC) | |
1200 | return 0; | |
1201 | if (!nfs_write_pageuptodate(page, inode)) | |
1202 | return 0; | |
1203 | if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE)) | |
1204 | return 1; | |
1205 | if (!flctx || (list_empty_careful(&flctx->flc_flock) && | |
1206 | list_empty_careful(&flctx->flc_posix))) | |
1207 | return 0; | |
1208 | ||
1209 | /* Check to see if there are whole file write locks */ | |
1210 | ret = 0; | |
1211 | spin_lock(&flctx->flc_lock); | |
1212 | if (!list_empty(&flctx->flc_posix)) { | |
1213 | fl = list_first_entry(&flctx->flc_posix, struct file_lock, | |
1214 | fl_list); | |
1215 | if (is_whole_file_wrlock(fl)) | |
1216 | ret = 1; | |
1217 | } else if (!list_empty(&flctx->flc_flock)) { | |
1218 | fl = list_first_entry(&flctx->flc_flock, struct file_lock, | |
1219 | fl_list); | |
1220 | if (fl->fl_type == F_WRLCK) | |
1221 | ret = 1; | |
1222 | } | |
1223 | spin_unlock(&flctx->flc_lock); | |
1224 | return ret; | |
1225 | } | |
1226 | ||
1227 | /* | |
1228 | * Update and possibly write a cached page of an NFS file. | |
1229 | * | |
1230 | * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad | |
1231 | * things with a page scheduled for an RPC call (e.g. invalidate it). | |
1232 | */ | |
1233 | int nfs_updatepage(struct file *file, struct page *page, | |
1234 | unsigned int offset, unsigned int count) | |
1235 | { | |
1236 | struct nfs_open_context *ctx = nfs_file_open_context(file); | |
1237 | struct inode *inode = page_file_mapping(page)->host; | |
1238 | int status = 0; | |
1239 | ||
1240 | nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE); | |
1241 | ||
1242 | dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n", | |
1243 | file, count, (long long)(page_file_offset(page) + offset)); | |
1244 | ||
1245 | if (nfs_can_extend_write(file, page, inode)) { | |
1246 | count = max(count + offset, nfs_page_length(page)); | |
1247 | offset = 0; | |
1248 | } | |
1249 | ||
1250 | status = nfs_writepage_setup(ctx, page, offset, count); | |
1251 | if (status < 0) | |
1252 | nfs_set_pageerror(page); | |
1253 | else | |
1254 | __set_page_dirty_nobuffers(page); | |
1255 | ||
1256 | dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n", | |
1257 | status, (long long)i_size_read(inode)); | |
1258 | return status; | |
1259 | } | |
1260 | ||
1261 | static int flush_task_priority(int how) | |
1262 | { | |
1263 | switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) { | |
1264 | case FLUSH_HIGHPRI: | |
1265 | return RPC_PRIORITY_HIGH; | |
1266 | case FLUSH_LOWPRI: | |
1267 | return RPC_PRIORITY_LOW; | |
1268 | } | |
1269 | return RPC_PRIORITY_NORMAL; | |
1270 | } | |
1271 | ||
1272 | static void nfs_initiate_write(struct nfs_pgio_header *hdr, | |
1273 | struct rpc_message *msg, | |
1274 | const struct nfs_rpc_ops *rpc_ops, | |
1275 | struct rpc_task_setup *task_setup_data, int how) | |
1276 | { | |
1277 | int priority = flush_task_priority(how); | |
1278 | ||
1279 | task_setup_data->priority = priority; | |
1280 | rpc_ops->write_setup(hdr, msg); | |
1281 | ||
1282 | nfs4_state_protect_write(NFS_SERVER(hdr->inode)->nfs_client, | |
1283 | &task_setup_data->rpc_client, msg, hdr); | |
1284 | } | |
1285 | ||
1286 | /* If a nfs_flush_* function fails, it should remove reqs from @head and | |
1287 | * call this on each, which will prepare them to be retried on next | |
1288 | * writeback using standard nfs. | |
1289 | */ | |
1290 | static void nfs_redirty_request(struct nfs_page *req) | |
1291 | { | |
1292 | nfs_mark_request_dirty(req); | |
1293 | nfs_unlock_request(req); | |
1294 | nfs_end_page_writeback(req); | |
1295 | nfs_release_request(req); | |
1296 | } | |
1297 | ||
1298 | static void nfs_async_write_error(struct list_head *head) | |
1299 | { | |
1300 | struct nfs_page *req; | |
1301 | ||
1302 | while (!list_empty(head)) { | |
1303 | req = nfs_list_entry(head->next); | |
1304 | nfs_list_remove_request(req); | |
1305 | nfs_redirty_request(req); | |
1306 | } | |
1307 | } | |
1308 | ||
1309 | static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = { | |
1310 | .error_cleanup = nfs_async_write_error, | |
1311 | .completion = nfs_write_completion, | |
1312 | }; | |
1313 | ||
1314 | void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, | |
1315 | struct inode *inode, int ioflags, bool force_mds, | |
1316 | const struct nfs_pgio_completion_ops *compl_ops) | |
1317 | { | |
1318 | struct nfs_server *server = NFS_SERVER(inode); | |
1319 | const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops; | |
1320 | ||
1321 | #ifdef CONFIG_NFS_V4_1 | |
1322 | if (server->pnfs_curr_ld && !force_mds) | |
1323 | pg_ops = server->pnfs_curr_ld->pg_write_ops; | |
1324 | #endif | |
1325 | nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops, | |
1326 | server->wsize, ioflags); | |
1327 | } | |
1328 | EXPORT_SYMBOL_GPL(nfs_pageio_init_write); | |
1329 | ||
1330 | void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio) | |
1331 | { | |
1332 | struct nfs_pgio_mirror *mirror; | |
1333 | ||
1334 | pgio->pg_ops = &nfs_pgio_rw_ops; | |
1335 | ||
1336 | nfs_pageio_stop_mirroring(pgio); | |
1337 | ||
1338 | mirror = &pgio->pg_mirrors[0]; | |
1339 | mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize; | |
1340 | } | |
1341 | EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds); | |
1342 | ||
1343 | ||
1344 | void nfs_commit_prepare(struct rpc_task *task, void *calldata) | |
1345 | { | |
1346 | struct nfs_commit_data *data = calldata; | |
1347 | ||
1348 | NFS_PROTO(data->inode)->commit_rpc_prepare(task, data); | |
1349 | } | |
1350 | ||
1351 | static void nfs_writeback_release_common(struct nfs_pgio_header *hdr) | |
1352 | { | |
1353 | /* do nothing! */ | |
1354 | } | |
1355 | ||
1356 | /* | |
1357 | * Special version of should_remove_suid() that ignores capabilities. | |
1358 | */ | |
1359 | static int nfs_should_remove_suid(const struct inode *inode) | |
1360 | { | |
1361 | umode_t mode = inode->i_mode; | |
1362 | int kill = 0; | |
1363 | ||
1364 | /* suid always must be killed */ | |
1365 | if (unlikely(mode & S_ISUID)) | |
1366 | kill = ATTR_KILL_SUID; | |
1367 | ||
1368 | /* | |
1369 | * sgid without any exec bits is just a mandatory locking mark; leave | |
1370 | * it alone. If some exec bits are set, it's a real sgid; kill it. | |
1371 | */ | |
1372 | if (unlikely((mode & S_ISGID) && (mode & S_IXGRP))) | |
1373 | kill |= ATTR_KILL_SGID; | |
1374 | ||
1375 | if (unlikely(kill && S_ISREG(mode))) | |
1376 | return kill; | |
1377 | ||
1378 | return 0; | |
1379 | } | |
1380 | ||
1381 | static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr, | |
1382 | struct nfs_fattr *fattr) | |
1383 | { | |
1384 | struct nfs_pgio_args *argp = &hdr->args; | |
1385 | struct nfs_pgio_res *resp = &hdr->res; | |
1386 | ||
1387 | if (!(fattr->valid & NFS_ATTR_FATTR_SIZE)) | |
1388 | return; | |
1389 | if (argp->offset + resp->count != fattr->size) | |
1390 | return; | |
1391 | if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) | |
1392 | return; | |
1393 | /* Set attribute barrier */ | |
1394 | nfs_fattr_set_barrier(fattr); | |
1395 | } | |
1396 | ||
1397 | void nfs_writeback_update_inode(struct nfs_pgio_header *hdr) | |
1398 | { | |
1399 | struct nfs_fattr *fattr = hdr->res.fattr; | |
1400 | struct inode *inode = hdr->inode; | |
1401 | ||
1402 | if (fattr == NULL) | |
1403 | return; | |
1404 | spin_lock(&inode->i_lock); | |
1405 | nfs_writeback_check_extend(hdr, fattr); | |
1406 | nfs_post_op_update_inode_force_wcc_locked(inode, fattr); | |
1407 | spin_unlock(&inode->i_lock); | |
1408 | } | |
1409 | EXPORT_SYMBOL_GPL(nfs_writeback_update_inode); | |
1410 | ||
1411 | /* | |
1412 | * This function is called when the WRITE call is complete. | |
1413 | */ | |
1414 | static int nfs_writeback_done(struct rpc_task *task, | |
1415 | struct nfs_pgio_header *hdr, | |
1416 | struct inode *inode) | |
1417 | { | |
1418 | int status; | |
1419 | ||
1420 | /* | |
1421 | * ->write_done will attempt to use post-op attributes to detect | |
1422 | * conflicting writes by other clients. A strict interpretation | |
1423 | * of close-to-open would allow us to continue caching even if | |
1424 | * another writer had changed the file, but some applications | |
1425 | * depend on tighter cache coherency when writing. | |
1426 | */ | |
1427 | status = NFS_PROTO(inode)->write_done(task, hdr); | |
1428 | if (status != 0) | |
1429 | return status; | |
1430 | nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count); | |
1431 | ||
1432 | if (hdr->res.verf->committed < hdr->args.stable && | |
1433 | task->tk_status >= 0) { | |
1434 | /* We tried a write call, but the server did not | |
1435 | * commit data to stable storage even though we | |
1436 | * requested it. | |
1437 | * Note: There is a known bug in Tru64 < 5.0 in which | |
1438 | * the server reports NFS_DATA_SYNC, but performs | |
1439 | * NFS_FILE_SYNC. We therefore implement this checking | |
1440 | * as a dprintk() in order to avoid filling syslog. | |
1441 | */ | |
1442 | static unsigned long complain; | |
1443 | ||
1444 | /* Note this will print the MDS for a DS write */ | |
1445 | if (time_before(complain, jiffies)) { | |
1446 | dprintk("NFS: faulty NFS server %s:" | |
1447 | " (committed = %d) != (stable = %d)\n", | |
1448 | NFS_SERVER(inode)->nfs_client->cl_hostname, | |
1449 | hdr->res.verf->committed, hdr->args.stable); | |
1450 | complain = jiffies + 300 * HZ; | |
1451 | } | |
1452 | } | |
1453 | ||
1454 | /* Deal with the suid/sgid bit corner case */ | |
1455 | if (nfs_should_remove_suid(inode)) | |
1456 | nfs_mark_for_revalidate(inode); | |
1457 | return 0; | |
1458 | } | |
1459 | ||
1460 | /* | |
1461 | * This function is called when the WRITE call is complete. | |
1462 | */ | |
1463 | static void nfs_writeback_result(struct rpc_task *task, | |
1464 | struct nfs_pgio_header *hdr) | |
1465 | { | |
1466 | struct nfs_pgio_args *argp = &hdr->args; | |
1467 | struct nfs_pgio_res *resp = &hdr->res; | |
1468 | ||
1469 | if (resp->count < argp->count) { | |
1470 | static unsigned long complain; | |
1471 | ||
1472 | /* This a short write! */ | |
1473 | nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE); | |
1474 | ||
1475 | /* Has the server at least made some progress? */ | |
1476 | if (resp->count == 0) { | |
1477 | if (time_before(complain, jiffies)) { | |
1478 | printk(KERN_WARNING | |
1479 | "NFS: Server wrote zero bytes, expected %u.\n", | |
1480 | argp->count); | |
1481 | complain = jiffies + 300 * HZ; | |
1482 | } | |
1483 | nfs_set_pgio_error(hdr, -EIO, argp->offset); | |
1484 | task->tk_status = -EIO; | |
1485 | return; | |
1486 | } | |
1487 | /* Was this an NFSv2 write or an NFSv3 stable write? */ | |
1488 | if (resp->verf->committed != NFS_UNSTABLE) { | |
1489 | /* Resend from where the server left off */ | |
1490 | hdr->mds_offset += resp->count; | |
1491 | argp->offset += resp->count; | |
1492 | argp->pgbase += resp->count; | |
1493 | argp->count -= resp->count; | |
1494 | } else { | |
1495 | /* Resend as a stable write in order to avoid | |
1496 | * headaches in the case of a server crash. | |
1497 | */ | |
1498 | argp->stable = NFS_FILE_SYNC; | |
1499 | } | |
1500 | rpc_restart_call_prepare(task); | |
1501 | } | |
1502 | } | |
1503 | ||
1504 | ||
1505 | static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait) | |
1506 | { | |
1507 | int ret; | |
1508 | ||
1509 | if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags)) | |
1510 | return 1; | |
1511 | if (!may_wait) | |
1512 | return 0; | |
1513 | ret = out_of_line_wait_on_bit_lock(&nfsi->flags, | |
1514 | NFS_INO_COMMIT, | |
1515 | nfs_wait_bit_killable, | |
1516 | TASK_KILLABLE); | |
1517 | return (ret < 0) ? ret : 1; | |
1518 | } | |
1519 | ||
1520 | static void nfs_commit_clear_lock(struct nfs_inode *nfsi) | |
1521 | { | |
1522 | clear_bit(NFS_INO_COMMIT, &nfsi->flags); | |
1523 | smp_mb__after_atomic(); | |
1524 | wake_up_bit(&nfsi->flags, NFS_INO_COMMIT); | |
1525 | } | |
1526 | ||
1527 | void nfs_commitdata_release(struct nfs_commit_data *data) | |
1528 | { | |
1529 | put_nfs_open_context(data->context); | |
1530 | nfs_commit_free(data); | |
1531 | } | |
1532 | EXPORT_SYMBOL_GPL(nfs_commitdata_release); | |
1533 | ||
1534 | int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data, | |
1535 | const struct nfs_rpc_ops *nfs_ops, | |
1536 | const struct rpc_call_ops *call_ops, | |
1537 | int how, int flags) | |
1538 | { | |
1539 | struct rpc_task *task; | |
1540 | int priority = flush_task_priority(how); | |
1541 | struct rpc_message msg = { | |
1542 | .rpc_argp = &data->args, | |
1543 | .rpc_resp = &data->res, | |
1544 | .rpc_cred = data->cred, | |
1545 | }; | |
1546 | struct rpc_task_setup task_setup_data = { | |
1547 | .task = &data->task, | |
1548 | .rpc_client = clnt, | |
1549 | .rpc_message = &msg, | |
1550 | .callback_ops = call_ops, | |
1551 | .callback_data = data, | |
1552 | .workqueue = nfsiod_workqueue, | |
1553 | .flags = RPC_TASK_ASYNC | flags, | |
1554 | .priority = priority, | |
1555 | }; | |
1556 | /* Set up the initial task struct. */ | |
1557 | nfs_ops->commit_setup(data, &msg); | |
1558 | ||
1559 | dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid); | |
1560 | ||
1561 | nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client, | |
1562 | NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg); | |
1563 | ||
1564 | task = rpc_run_task(&task_setup_data); | |
1565 | if (IS_ERR(task)) | |
1566 | return PTR_ERR(task); | |
1567 | if (how & FLUSH_SYNC) | |
1568 | rpc_wait_for_completion_task(task); | |
1569 | rpc_put_task(task); | |
1570 | return 0; | |
1571 | } | |
1572 | EXPORT_SYMBOL_GPL(nfs_initiate_commit); | |
1573 | ||
1574 | static loff_t nfs_get_lwb(struct list_head *head) | |
1575 | { | |
1576 | loff_t lwb = 0; | |
1577 | struct nfs_page *req; | |
1578 | ||
1579 | list_for_each_entry(req, head, wb_list) | |
1580 | if (lwb < (req_offset(req) + req->wb_bytes)) | |
1581 | lwb = req_offset(req) + req->wb_bytes; | |
1582 | ||
1583 | return lwb; | |
1584 | } | |
1585 | ||
1586 | /* | |
1587 | * Set up the argument/result storage required for the RPC call. | |
1588 | */ | |
1589 | void nfs_init_commit(struct nfs_commit_data *data, | |
1590 | struct list_head *head, | |
1591 | struct pnfs_layout_segment *lseg, | |
1592 | struct nfs_commit_info *cinfo) | |
1593 | { | |
1594 | struct nfs_page *first = nfs_list_entry(head->next); | |
1595 | struct inode *inode = d_inode(first->wb_context->dentry); | |
1596 | ||
1597 | /* Set up the RPC argument and reply structs | |
1598 | * NB: take care not to mess about with data->commit et al. */ | |
1599 | ||
1600 | list_splice_init(head, &data->pages); | |
1601 | ||
1602 | data->inode = inode; | |
1603 | data->cred = first->wb_context->cred; | |
1604 | data->lseg = lseg; /* reference transferred */ | |
1605 | /* only set lwb for pnfs commit */ | |
1606 | if (lseg) | |
1607 | data->lwb = nfs_get_lwb(&data->pages); | |
1608 | data->mds_ops = &nfs_commit_ops; | |
1609 | data->completion_ops = cinfo->completion_ops; | |
1610 | data->dreq = cinfo->dreq; | |
1611 | ||
1612 | data->args.fh = NFS_FH(data->inode); | |
1613 | /* Note: we always request a commit of the entire inode */ | |
1614 | data->args.offset = 0; | |
1615 | data->args.count = 0; | |
1616 | data->context = get_nfs_open_context(first->wb_context); | |
1617 | data->res.fattr = &data->fattr; | |
1618 | data->res.verf = &data->verf; | |
1619 | nfs_fattr_init(&data->fattr); | |
1620 | } | |
1621 | EXPORT_SYMBOL_GPL(nfs_init_commit); | |
1622 | ||
1623 | void nfs_retry_commit(struct list_head *page_list, | |
1624 | struct pnfs_layout_segment *lseg, | |
1625 | struct nfs_commit_info *cinfo, | |
1626 | u32 ds_commit_idx) | |
1627 | { | |
1628 | struct nfs_page *req; | |
1629 | ||
1630 | while (!list_empty(page_list)) { | |
1631 | req = nfs_list_entry(page_list->next); | |
1632 | nfs_list_remove_request(req); | |
1633 | nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx); | |
1634 | if (!cinfo->dreq) | |
1635 | nfs_clear_page_commit(req->wb_page); | |
1636 | nfs_unlock_and_release_request(req); | |
1637 | } | |
1638 | } | |
1639 | EXPORT_SYMBOL_GPL(nfs_retry_commit); | |
1640 | ||
1641 | /* | |
1642 | * Commit dirty pages | |
1643 | */ | |
1644 | static int | |
1645 | nfs_commit_list(struct inode *inode, struct list_head *head, int how, | |
1646 | struct nfs_commit_info *cinfo) | |
1647 | { | |
1648 | struct nfs_commit_data *data; | |
1649 | ||
1650 | data = nfs_commitdata_alloc(); | |
1651 | ||
1652 | if (!data) | |
1653 | goto out_bad; | |
1654 | ||
1655 | /* Set up the argument struct */ | |
1656 | nfs_init_commit(data, head, NULL, cinfo); | |
1657 | atomic_inc(&cinfo->mds->rpcs_out); | |
1658 | return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode), | |
1659 | data->mds_ops, how, 0); | |
1660 | out_bad: | |
1661 | nfs_retry_commit(head, NULL, cinfo, 0); | |
1662 | cinfo->completion_ops->error_cleanup(NFS_I(inode)); | |
1663 | return -ENOMEM; | |
1664 | } | |
1665 | ||
1666 | /* | |
1667 | * COMMIT call returned | |
1668 | */ | |
1669 | static void nfs_commit_done(struct rpc_task *task, void *calldata) | |
1670 | { | |
1671 | struct nfs_commit_data *data = calldata; | |
1672 | ||
1673 | dprintk("NFS: %5u nfs_commit_done (status %d)\n", | |
1674 | task->tk_pid, task->tk_status); | |
1675 | ||
1676 | /* Call the NFS version-specific code */ | |
1677 | NFS_PROTO(data->inode)->commit_done(task, data); | |
1678 | } | |
1679 | ||
1680 | static void nfs_commit_release_pages(struct nfs_commit_data *data) | |
1681 | { | |
1682 | struct nfs_page *req; | |
1683 | int status = data->task.tk_status; | |
1684 | struct nfs_commit_info cinfo; | |
1685 | struct nfs_server *nfss; | |
1686 | ||
1687 | while (!list_empty(&data->pages)) { | |
1688 | req = nfs_list_entry(data->pages.next); | |
1689 | nfs_list_remove_request(req); | |
1690 | nfs_clear_page_commit(req->wb_page); | |
1691 | ||
1692 | dprintk("NFS: commit (%s/%llu %d@%lld)", | |
1693 | req->wb_context->dentry->d_sb->s_id, | |
1694 | (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)), | |
1695 | req->wb_bytes, | |
1696 | (long long)req_offset(req)); | |
1697 | if (status < 0) { | |
1698 | nfs_context_set_write_error(req->wb_context, status); | |
1699 | nfs_inode_remove_request(req); | |
1700 | dprintk(", error = %d\n", status); | |
1701 | goto next; | |
1702 | } | |
1703 | ||
1704 | /* Okay, COMMIT succeeded, apparently. Check the verifier | |
1705 | * returned by the server against all stored verfs. */ | |
1706 | if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) { | |
1707 | /* We have a match */ | |
1708 | nfs_inode_remove_request(req); | |
1709 | dprintk(" OK\n"); | |
1710 | goto next; | |
1711 | } | |
1712 | /* We have a mismatch. Write the page again */ | |
1713 | dprintk(" mismatch\n"); | |
1714 | nfs_mark_request_dirty(req); | |
1715 | set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags); | |
1716 | next: | |
1717 | nfs_unlock_and_release_request(req); | |
1718 | } | |
1719 | nfss = NFS_SERVER(data->inode); | |
1720 | if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) | |
1721 | clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC); | |
1722 | ||
1723 | nfs_init_cinfo(&cinfo, data->inode, data->dreq); | |
1724 | if (atomic_dec_and_test(&cinfo.mds->rpcs_out)) | |
1725 | nfs_commit_clear_lock(NFS_I(data->inode)); | |
1726 | } | |
1727 | ||
1728 | static void nfs_commit_release(void *calldata) | |
1729 | { | |
1730 | struct nfs_commit_data *data = calldata; | |
1731 | ||
1732 | data->completion_ops->completion(data); | |
1733 | nfs_commitdata_release(calldata); | |
1734 | } | |
1735 | ||
1736 | static const struct rpc_call_ops nfs_commit_ops = { | |
1737 | .rpc_call_prepare = nfs_commit_prepare, | |
1738 | .rpc_call_done = nfs_commit_done, | |
1739 | .rpc_release = nfs_commit_release, | |
1740 | }; | |
1741 | ||
1742 | static const struct nfs_commit_completion_ops nfs_commit_completion_ops = { | |
1743 | .completion = nfs_commit_release_pages, | |
1744 | .error_cleanup = nfs_commit_clear_lock, | |
1745 | }; | |
1746 | ||
1747 | int nfs_generic_commit_list(struct inode *inode, struct list_head *head, | |
1748 | int how, struct nfs_commit_info *cinfo) | |
1749 | { | |
1750 | int status; | |
1751 | ||
1752 | status = pnfs_commit_list(inode, head, how, cinfo); | |
1753 | if (status == PNFS_NOT_ATTEMPTED) | |
1754 | status = nfs_commit_list(inode, head, how, cinfo); | |
1755 | return status; | |
1756 | } | |
1757 | ||
1758 | int nfs_commit_inode(struct inode *inode, int how) | |
1759 | { | |
1760 | LIST_HEAD(head); | |
1761 | struct nfs_commit_info cinfo; | |
1762 | int may_wait = how & FLUSH_SYNC; | |
1763 | int res; | |
1764 | ||
1765 | res = nfs_commit_set_lock(NFS_I(inode), may_wait); | |
1766 | if (res <= 0) | |
1767 | goto out_mark_dirty; | |
1768 | nfs_init_cinfo_from_inode(&cinfo, inode); | |
1769 | res = nfs_scan_commit(inode, &head, &cinfo); | |
1770 | if (res) { | |
1771 | int error; | |
1772 | ||
1773 | error = nfs_generic_commit_list(inode, &head, how, &cinfo); | |
1774 | if (error < 0) | |
1775 | return error; | |
1776 | if (!may_wait) | |
1777 | goto out_mark_dirty; | |
1778 | error = wait_on_bit_action(&NFS_I(inode)->flags, | |
1779 | NFS_INO_COMMIT, | |
1780 | nfs_wait_bit_killable, | |
1781 | TASK_KILLABLE); | |
1782 | if (error < 0) | |
1783 | return error; | |
1784 | } else | |
1785 | nfs_commit_clear_lock(NFS_I(inode)); | |
1786 | return res; | |
1787 | /* Note: If we exit without ensuring that the commit is complete, | |
1788 | * we must mark the inode as dirty. Otherwise, future calls to | |
1789 | * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure | |
1790 | * that the data is on the disk. | |
1791 | */ | |
1792 | out_mark_dirty: | |
1793 | __mark_inode_dirty(inode, I_DIRTY_DATASYNC); | |
1794 | return res; | |
1795 | } | |
1796 | ||
1797 | static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc) | |
1798 | { | |
1799 | struct nfs_inode *nfsi = NFS_I(inode); | |
1800 | int flags = FLUSH_SYNC; | |
1801 | int ret = 0; | |
1802 | ||
1803 | /* no commits means nothing needs to be done */ | |
1804 | if (!nfsi->commit_info.ncommit) | |
1805 | return ret; | |
1806 | ||
1807 | if (wbc->sync_mode == WB_SYNC_NONE) { | |
1808 | /* Don't commit yet if this is a non-blocking flush and there | |
1809 | * are a lot of outstanding writes for this mapping. | |
1810 | */ | |
1811 | if (nfsi->commit_info.ncommit <= (nfsi->nrequests >> 1)) | |
1812 | goto out_mark_dirty; | |
1813 | ||
1814 | /* don't wait for the COMMIT response */ | |
1815 | flags = 0; | |
1816 | } | |
1817 | ||
1818 | ret = nfs_commit_inode(inode, flags); | |
1819 | if (ret >= 0) { | |
1820 | if (wbc->sync_mode == WB_SYNC_NONE) { | |
1821 | if (ret < wbc->nr_to_write) | |
1822 | wbc->nr_to_write -= ret; | |
1823 | else | |
1824 | wbc->nr_to_write = 0; | |
1825 | } | |
1826 | return 0; | |
1827 | } | |
1828 | out_mark_dirty: | |
1829 | __mark_inode_dirty(inode, I_DIRTY_DATASYNC); | |
1830 | return ret; | |
1831 | } | |
1832 | ||
1833 | int nfs_write_inode(struct inode *inode, struct writeback_control *wbc) | |
1834 | { | |
1835 | return nfs_commit_unstable_pages(inode, wbc); | |
1836 | } | |
1837 | EXPORT_SYMBOL_GPL(nfs_write_inode); | |
1838 | ||
1839 | /* | |
1840 | * flush the inode to disk. | |
1841 | */ | |
1842 | int nfs_wb_all(struct inode *inode) | |
1843 | { | |
1844 | int ret; | |
1845 | ||
1846 | trace_nfs_writeback_inode_enter(inode); | |
1847 | ||
1848 | ret = filemap_write_and_wait(inode->i_mapping); | |
1849 | if (!ret) { | |
1850 | ret = nfs_commit_inode(inode, FLUSH_SYNC); | |
1851 | if (!ret) | |
1852 | pnfs_sync_inode(inode, true); | |
1853 | } | |
1854 | ||
1855 | trace_nfs_writeback_inode_exit(inode, ret); | |
1856 | return ret; | |
1857 | } | |
1858 | EXPORT_SYMBOL_GPL(nfs_wb_all); | |
1859 | ||
1860 | int nfs_wb_page_cancel(struct inode *inode, struct page *page) | |
1861 | { | |
1862 | struct nfs_page *req; | |
1863 | int ret = 0; | |
1864 | ||
1865 | wait_on_page_writeback(page); | |
1866 | ||
1867 | /* blocking call to cancel all requests and join to a single (head) | |
1868 | * request */ | |
1869 | req = nfs_lock_and_join_requests(page, false); | |
1870 | ||
1871 | if (IS_ERR(req)) { | |
1872 | ret = PTR_ERR(req); | |
1873 | } else if (req) { | |
1874 | /* all requests from this page have been cancelled by | |
1875 | * nfs_lock_and_join_requests, so just remove the head | |
1876 | * request from the inode / page_private pointer and | |
1877 | * release it */ | |
1878 | nfs_inode_remove_request(req); | |
1879 | nfs_unlock_and_release_request(req); | |
1880 | } | |
1881 | ||
1882 | return ret; | |
1883 | } | |
1884 | ||
1885 | /* | |
1886 | * Write back all requests on one page - we do this before reading it. | |
1887 | */ | |
1888 | int nfs_wb_page(struct inode *inode, struct page *page) | |
1889 | { | |
1890 | loff_t range_start = page_file_offset(page); | |
1891 | loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1); | |
1892 | struct writeback_control wbc = { | |
1893 | .sync_mode = WB_SYNC_ALL, | |
1894 | .nr_to_write = 0, | |
1895 | .range_start = range_start, | |
1896 | .range_end = range_end, | |
1897 | }; | |
1898 | int ret; | |
1899 | ||
1900 | trace_nfs_writeback_page_enter(inode); | |
1901 | ||
1902 | for (;;) { | |
1903 | wait_on_page_writeback(page); | |
1904 | if (clear_page_dirty_for_io(page)) { | |
1905 | ret = nfs_writepage_locked(page, &wbc); | |
1906 | if (ret < 0) | |
1907 | goto out_error; | |
1908 | continue; | |
1909 | } | |
1910 | ret = 0; | |
1911 | if (!PagePrivate(page)) | |
1912 | break; | |
1913 | ret = nfs_commit_inode(inode, FLUSH_SYNC); | |
1914 | if (ret < 0) | |
1915 | goto out_error; | |
1916 | } | |
1917 | out_error: | |
1918 | trace_nfs_writeback_page_exit(inode, ret); | |
1919 | return ret; | |
1920 | } | |
1921 | ||
1922 | #ifdef CONFIG_MIGRATION | |
1923 | int nfs_migrate_page(struct address_space *mapping, struct page *newpage, | |
1924 | struct page *page, enum migrate_mode mode) | |
1925 | { | |
1926 | /* | |
1927 | * If PagePrivate is set, then the page is currently associated with | |
1928 | * an in-progress read or write request. Don't try to migrate it. | |
1929 | * | |
1930 | * FIXME: we could do this in principle, but we'll need a way to ensure | |
1931 | * that we can safely release the inode reference while holding | |
1932 | * the page lock. | |
1933 | */ | |
1934 | if (PagePrivate(page)) | |
1935 | return -EBUSY; | |
1936 | ||
1937 | if (!nfs_fscache_release_page(page, GFP_KERNEL)) | |
1938 | return -EBUSY; | |
1939 | ||
1940 | return migrate_page(mapping, newpage, page, mode); | |
1941 | } | |
1942 | #endif | |
1943 | ||
1944 | int __init nfs_init_writepagecache(void) | |
1945 | { | |
1946 | nfs_wdata_cachep = kmem_cache_create("nfs_write_data", | |
1947 | sizeof(struct nfs_pgio_header), | |
1948 | 0, SLAB_HWCACHE_ALIGN, | |
1949 | NULL); | |
1950 | if (nfs_wdata_cachep == NULL) | |
1951 | return -ENOMEM; | |
1952 | ||
1953 | nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE, | |
1954 | nfs_wdata_cachep); | |
1955 | if (nfs_wdata_mempool == NULL) | |
1956 | goto out_destroy_write_cache; | |
1957 | ||
1958 | nfs_cdata_cachep = kmem_cache_create("nfs_commit_data", | |
1959 | sizeof(struct nfs_commit_data), | |
1960 | 0, SLAB_HWCACHE_ALIGN, | |
1961 | NULL); | |
1962 | if (nfs_cdata_cachep == NULL) | |
1963 | goto out_destroy_write_mempool; | |
1964 | ||
1965 | nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT, | |
1966 | nfs_cdata_cachep); | |
1967 | if (nfs_commit_mempool == NULL) | |
1968 | goto out_destroy_commit_cache; | |
1969 | ||
1970 | /* | |
1971 | * NFS congestion size, scale with available memory. | |
1972 | * | |
1973 | * 64MB: 8192k | |
1974 | * 128MB: 11585k | |
1975 | * 256MB: 16384k | |
1976 | * 512MB: 23170k | |
1977 | * 1GB: 32768k | |
1978 | * 2GB: 46340k | |
1979 | * 4GB: 65536k | |
1980 | * 8GB: 92681k | |
1981 | * 16GB: 131072k | |
1982 | * | |
1983 | * This allows larger machines to have larger/more transfers. | |
1984 | * Limit the default to 256M | |
1985 | */ | |
1986 | nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10); | |
1987 | if (nfs_congestion_kb > 256*1024) | |
1988 | nfs_congestion_kb = 256*1024; | |
1989 | ||
1990 | return 0; | |
1991 | ||
1992 | out_destroy_commit_cache: | |
1993 | kmem_cache_destroy(nfs_cdata_cachep); | |
1994 | out_destroy_write_mempool: | |
1995 | mempool_destroy(nfs_wdata_mempool); | |
1996 | out_destroy_write_cache: | |
1997 | kmem_cache_destroy(nfs_wdata_cachep); | |
1998 | return -ENOMEM; | |
1999 | } | |
2000 | ||
2001 | void nfs_destroy_writepagecache(void) | |
2002 | { | |
2003 | mempool_destroy(nfs_commit_mempool); | |
2004 | kmem_cache_destroy(nfs_cdata_cachep); | |
2005 | mempool_destroy(nfs_wdata_mempool); | |
2006 | kmem_cache_destroy(nfs_wdata_cachep); | |
2007 | } | |
2008 | ||
2009 | static const struct nfs_rw_ops nfs_rw_write_ops = { | |
2010 | .rw_mode = FMODE_WRITE, | |
2011 | .rw_alloc_header = nfs_writehdr_alloc, | |
2012 | .rw_free_header = nfs_writehdr_free, | |
2013 | .rw_release = nfs_writeback_release_common, | |
2014 | .rw_done = nfs_writeback_done, | |
2015 | .rw_result = nfs_writeback_result, | |
2016 | .rw_initiate = nfs_initiate_write, | |
2017 | }; |