1 | // SPDX-License-Identifier: GPL-2.0-only |
2 | /* |
3 | * linux/fs/nfs/write.c |
4 | * |
5 | * Write file data over NFS. |
6 | * |
7 | * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de> |
8 | */ |
9 | |
10 | #include <linux/types.h> |
11 | #include <linux/slab.h> |
12 | #include <linux/mm.h> |
13 | #include <linux/pagemap.h> |
14 | #include <linux/file.h> |
15 | #include <linux/writeback.h> |
16 | #include <linux/swap.h> |
17 | #include <linux/migrate.h> |
18 | |
19 | #include <linux/sunrpc/clnt.h> |
20 | #include <linux/nfs_fs.h> |
21 | #include <linux/nfs_mount.h> |
22 | #include <linux/nfs_page.h> |
23 | #include <linux/backing-dev.h> |
24 | #include <linux/export.h> |
25 | #include <linux/freezer.h> |
26 | #include <linux/wait.h> |
27 | #include <linux/iversion.h> |
28 | #include <linux/filelock.h> |
29 | |
30 | #include <linux/uaccess.h> |
31 | #include <linux/sched/mm.h> |
32 | |
33 | #include "delegation.h" |
34 | #include "internal.h" |
35 | #include "iostat.h" |
36 | #include "nfs4_fs.h" |
37 | #include "fscache.h" |
38 | #include "pnfs.h" |
39 | |
40 | #include "nfstrace.h" |
41 | |
42 | #define NFSDBG_FACILITY NFSDBG_PAGECACHE |
43 | |
44 | #define MIN_POOL_WRITE (32) |
45 | #define MIN_POOL_COMMIT (4) |
46 | |
47 | struct nfs_io_completion { |
48 | void (*complete)(void *data); |
49 | void *data; |
50 | struct kref refcount; |
51 | }; |
52 | |
53 | /* |
54 | * Local function declarations |
55 | */ |
56 | static void nfs_redirty_request(struct nfs_page *req); |
57 | static const struct rpc_call_ops nfs_commit_ops; |
58 | static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops; |
59 | static const struct nfs_commit_completion_ops nfs_commit_completion_ops; |
60 | static const struct nfs_rw_ops nfs_rw_write_ops; |
61 | static void nfs_inode_remove_request(struct nfs_page *req); |
62 | static void nfs_clear_request_commit(struct nfs_commit_info *cinfo, |
63 | struct nfs_page *req); |
64 | static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo, |
65 | struct inode *inode); |
66 | static struct nfs_page * |
67 | nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi, |
68 | struct folio *folio); |
69 | |
70 | static struct kmem_cache *nfs_wdata_cachep; |
71 | static mempool_t *nfs_wdata_mempool; |
72 | static struct kmem_cache *nfs_cdata_cachep; |
73 | static mempool_t *nfs_commit_mempool; |
74 | |
75 | struct nfs_commit_data *nfs_commitdata_alloc(void) |
76 | { |
77 | struct nfs_commit_data *p; |
78 | |
79 | p = kmem_cache_zalloc(k: nfs_cdata_cachep, flags: nfs_io_gfp_mask()); |
80 | if (!p) { |
81 | p = mempool_alloc(pool: nfs_commit_mempool, GFP_NOWAIT); |
82 | if (!p) |
83 | return NULL; |
84 | memset(p, 0, sizeof(*p)); |
85 | } |
86 | INIT_LIST_HEAD(list: &p->pages); |
87 | return p; |
88 | } |
89 | EXPORT_SYMBOL_GPL(nfs_commitdata_alloc); |
90 | |
91 | void nfs_commit_free(struct nfs_commit_data *p) |
92 | { |
93 | mempool_free(element: p, pool: nfs_commit_mempool); |
94 | } |
95 | EXPORT_SYMBOL_GPL(nfs_commit_free); |
96 | |
97 | static struct nfs_pgio_header *nfs_writehdr_alloc(void) |
98 | { |
99 | struct nfs_pgio_header *p; |
100 | |
101 | p = kmem_cache_zalloc(k: nfs_wdata_cachep, flags: nfs_io_gfp_mask()); |
102 | if (!p) { |
103 | p = mempool_alloc(pool: nfs_wdata_mempool, GFP_NOWAIT); |
104 | if (!p) |
105 | return NULL; |
106 | memset(p, 0, sizeof(*p)); |
107 | } |
108 | p->rw_mode = FMODE_WRITE; |
109 | return p; |
110 | } |
111 | |
112 | static void nfs_writehdr_free(struct nfs_pgio_header *hdr) |
113 | { |
114 | mempool_free(element: hdr, pool: nfs_wdata_mempool); |
115 | } |
116 | |
117 | static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags) |
118 | { |
119 | return kmalloc(size: sizeof(struct nfs_io_completion), flags: gfp_flags); |
120 | } |
121 | |
122 | static void nfs_io_completion_init(struct nfs_io_completion *ioc, |
123 | void (*complete)(void *), void *data) |
124 | { |
125 | ioc->complete = complete; |
126 | ioc->data = data; |
127 | kref_init(kref: &ioc->refcount); |
128 | } |
129 | |
130 | static void nfs_io_completion_release(struct kref *kref) |
131 | { |
132 | struct nfs_io_completion *ioc = container_of(kref, |
133 | struct nfs_io_completion, refcount); |
134 | ioc->complete(ioc->data); |
135 | kfree(objp: ioc); |
136 | } |
137 | |
138 | static void nfs_io_completion_get(struct nfs_io_completion *ioc) |
139 | { |
140 | if (ioc != NULL) |
141 | kref_get(kref: &ioc->refcount); |
142 | } |
143 | |
144 | static void nfs_io_completion_put(struct nfs_io_completion *ioc) |
145 | { |
146 | if (ioc != NULL) |
147 | kref_put(kref: &ioc->refcount, release: nfs_io_completion_release); |
148 | } |
149 | |
150 | static void |
151 | nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode) |
152 | { |
153 | if (!test_and_set_bit(nr: PG_INODE_REF, addr: &req->wb_flags)) { |
154 | kref_get(kref: &req->wb_kref); |
155 | atomic_long_inc(v: &NFS_I(inode)->nrequests); |
156 | } |
157 | } |
158 | |
159 | static int |
160 | nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode) |
161 | { |
162 | int ret; |
163 | |
164 | if (!test_bit(PG_REMOVE, &req->wb_flags)) |
165 | return 0; |
166 | ret = nfs_page_group_lock(req); |
167 | if (ret) |
168 | return ret; |
169 | if (test_and_clear_bit(nr: PG_REMOVE, addr: &req->wb_flags)) |
170 | nfs_page_set_inode_ref(req, inode); |
171 | nfs_page_group_unlock(req); |
172 | return 0; |
173 | } |
174 | |
175 | static struct nfs_page *nfs_folio_private_request(struct folio *folio) |
176 | { |
177 | return folio_get_private(folio); |
178 | } |
179 | |
180 | /** |
181 | * nfs_folio_find_private_request - find head request associated with a folio |
182 | * @folio: pointer to folio |
183 | * |
184 | * must be called while holding the inode lock. |
185 | * |
186 | * returns matching head request with reference held, or NULL if not found. |
187 | */ |
188 | static struct nfs_page *nfs_folio_find_private_request(struct folio *folio) |
189 | { |
190 | struct address_space *mapping = folio_file_mapping(folio); |
191 | struct nfs_page *req; |
192 | |
193 | if (!folio_test_private(folio)) |
194 | return NULL; |
195 | spin_lock(lock: &mapping->private_lock); |
196 | req = nfs_folio_private_request(folio); |
197 | if (req) { |
198 | WARN_ON_ONCE(req->wb_head != req); |
199 | kref_get(kref: &req->wb_kref); |
200 | } |
201 | spin_unlock(lock: &mapping->private_lock); |
202 | return req; |
203 | } |
204 | |
205 | static struct nfs_page *nfs_folio_find_swap_request(struct folio *folio) |
206 | { |
207 | struct inode *inode = folio_file_mapping(folio)->host; |
208 | struct nfs_inode *nfsi = NFS_I(inode); |
209 | struct nfs_page *req = NULL; |
210 | if (!folio_test_swapcache(folio)) |
211 | return NULL; |
212 | mutex_lock(&nfsi->commit_mutex); |
213 | if (folio_test_swapcache(folio)) { |
214 | req = nfs_page_search_commits_for_head_request_locked(nfsi, |
215 | folio); |
216 | if (req) { |
217 | WARN_ON_ONCE(req->wb_head != req); |
218 | kref_get(kref: &req->wb_kref); |
219 | } |
220 | } |
221 | mutex_unlock(lock: &nfsi->commit_mutex); |
222 | return req; |
223 | } |
224 | |
225 | /** |
226 | * nfs_folio_find_head_request - find head request associated with a folio |
227 | * @folio: pointer to folio |
228 | * |
229 | * returns matching head request with reference held, or NULL if not found. |
230 | */ |
231 | static struct nfs_page *nfs_folio_find_head_request(struct folio *folio) |
232 | { |
233 | struct nfs_page *req; |
234 | |
235 | req = nfs_folio_find_private_request(folio); |
236 | if (!req) |
237 | req = nfs_folio_find_swap_request(folio); |
238 | return req; |
239 | } |
240 | |
241 | static struct nfs_page *nfs_folio_find_and_lock_request(struct folio *folio) |
242 | { |
243 | struct inode *inode = folio_file_mapping(folio)->host; |
244 | struct nfs_page *req, *head; |
245 | int ret; |
246 | |
247 | for (;;) { |
248 | req = nfs_folio_find_head_request(folio); |
249 | if (!req) |
250 | return req; |
251 | head = nfs_page_group_lock_head(req); |
252 | if (head != req) |
253 | nfs_release_request(req); |
254 | if (IS_ERR(ptr: head)) |
255 | return head; |
256 | ret = nfs_cancel_remove_inode(req: head, inode); |
257 | if (ret < 0) { |
258 | nfs_unlock_and_release_request(head); |
259 | return ERR_PTR(error: ret); |
260 | } |
261 | /* Ensure that nobody removed the request before we locked it */ |
262 | if (head == nfs_folio_private_request(folio)) |
263 | break; |
264 | if (folio_test_swapcache(folio)) |
265 | break; |
266 | nfs_unlock_and_release_request(head); |
267 | } |
268 | return head; |
269 | } |
270 | |
271 | /* Adjust the file length if we're writing beyond the end */ |
272 | static void nfs_grow_file(struct folio *folio, unsigned int offset, |
273 | unsigned int count) |
274 | { |
275 | struct inode *inode = folio_file_mapping(folio)->host; |
276 | loff_t end, i_size; |
277 | pgoff_t end_index; |
278 | |
279 | spin_lock(lock: &inode->i_lock); |
280 | i_size = i_size_read(inode); |
281 | end_index = ((i_size - 1) >> folio_shift(folio)) << folio_order(folio); |
282 | if (i_size > 0 && folio_index(folio) < end_index) |
283 | goto out; |
284 | end = folio_file_pos(folio) + (loff_t)offset + (loff_t)count; |
285 | if (i_size >= end) |
286 | goto out; |
287 | trace_nfs_size_grow(inode, new_size: end); |
288 | i_size_write(inode, i_size: end); |
289 | NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE; |
290 | nfs_inc_stats(inode, stat: NFSIOS_EXTENDWRITE); |
291 | out: |
292 | spin_unlock(lock: &inode->i_lock); |
293 | nfs_fscache_invalidate(inode, flags: 0); |
294 | } |
295 | |
296 | /* A writeback failed: mark the page as bad, and invalidate the page cache */ |
297 | static void nfs_set_pageerror(struct address_space *mapping) |
298 | { |
299 | struct inode *inode = mapping->host; |
300 | |
301 | nfs_zap_mapping(inode: mapping->host, mapping); |
302 | /* Force file size revalidation */ |
303 | spin_lock(lock: &inode->i_lock); |
304 | nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED | |
305 | NFS_INO_INVALID_CHANGE | |
306 | NFS_INO_INVALID_SIZE); |
307 | spin_unlock(lock: &inode->i_lock); |
308 | } |
309 | |
310 | static void nfs_mapping_set_error(struct folio *folio, int error) |
311 | { |
312 | struct address_space *mapping = folio_file_mapping(folio); |
313 | |
314 | folio_set_error(folio); |
315 | filemap_set_wb_err(mapping, err: error); |
316 | if (mapping->host) |
317 | errseq_set(eseq: &mapping->host->i_sb->s_wb_err, |
318 | err: error == -ENOSPC ? -ENOSPC : -EIO); |
319 | nfs_set_pageerror(mapping); |
320 | } |
321 | |
322 | /* |
323 | * nfs_page_group_search_locked |
324 | * @head - head request of page group |
325 | * @page_offset - offset into page |
326 | * |
327 | * Search page group with head @head to find a request that contains the |
328 | * page offset @page_offset. |
329 | * |
330 | * Returns a pointer to the first matching nfs request, or NULL if no |
331 | * match is found. |
332 | * |
333 | * Must be called with the page group lock held |
334 | */ |
335 | static struct nfs_page * |
336 | nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset) |
337 | { |
338 | struct nfs_page *req; |
339 | |
340 | req = head; |
341 | do { |
342 | if (page_offset >= req->wb_pgbase && |
343 | page_offset < (req->wb_pgbase + req->wb_bytes)) |
344 | return req; |
345 | |
346 | req = req->wb_this_page; |
347 | } while (req != head); |
348 | |
349 | return NULL; |
350 | } |
351 | |
352 | /* |
353 | * nfs_page_group_covers_page |
354 | * @head - head request of page group |
355 | * |
356 | * Return true if the page group with head @head covers the whole page, |
357 | * returns false otherwise |
358 | */ |
359 | static bool nfs_page_group_covers_page(struct nfs_page *req) |
360 | { |
361 | unsigned int len = nfs_folio_length(folio: nfs_page_to_folio(req)); |
362 | struct nfs_page *tmp; |
363 | unsigned int pos = 0; |
364 | |
365 | nfs_page_group_lock(req); |
366 | |
367 | for (;;) { |
368 | tmp = nfs_page_group_search_locked(head: req->wb_head, page_offset: pos); |
369 | if (!tmp) |
370 | break; |
371 | pos = tmp->wb_pgbase + tmp->wb_bytes; |
372 | } |
373 | |
374 | nfs_page_group_unlock(req); |
375 | return pos >= len; |
376 | } |
377 | |
378 | /* We can set the PG_uptodate flag if we see that a write request |
379 | * covers the full page. |
380 | */ |
381 | static void nfs_mark_uptodate(struct nfs_page *req) |
382 | { |
383 | struct folio *folio = nfs_page_to_folio(req); |
384 | |
385 | if (folio_test_uptodate(folio)) |
386 | return; |
387 | if (!nfs_page_group_covers_page(req)) |
388 | return; |
389 | folio_mark_uptodate(folio); |
390 | } |
391 | |
392 | static int wb_priority(struct writeback_control *wbc) |
393 | { |
394 | int ret = 0; |
395 | |
396 | if (wbc->sync_mode == WB_SYNC_ALL) |
397 | ret = FLUSH_COND_STABLE; |
398 | return ret; |
399 | } |
400 | |
401 | /* |
402 | * NFS congestion control |
403 | */ |
404 | |
405 | int nfs_congestion_kb; |
406 | |
407 | #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10)) |
408 | #define NFS_CONGESTION_OFF_THRESH \ |
409 | (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2)) |
410 | |
411 | static void nfs_folio_set_writeback(struct folio *folio) |
412 | { |
413 | struct nfs_server *nfss = NFS_SERVER(inode: folio_file_mapping(folio)->host); |
414 | |
415 | folio_start_writeback(folio); |
416 | if (atomic_long_inc_return(v: &nfss->writeback) > NFS_CONGESTION_ON_THRESH) |
417 | nfss->write_congested = 1; |
418 | } |
419 | |
420 | static void nfs_folio_end_writeback(struct folio *folio) |
421 | { |
422 | struct nfs_server *nfss = NFS_SERVER(inode: folio_file_mapping(folio)->host); |
423 | |
424 | folio_end_writeback(folio); |
425 | if (atomic_long_dec_return(v: &nfss->writeback) < |
426 | NFS_CONGESTION_OFF_THRESH) |
427 | nfss->write_congested = 0; |
428 | } |
429 | |
430 | static void nfs_page_end_writeback(struct nfs_page *req) |
431 | { |
432 | if (nfs_page_group_sync_on_bit(req, PG_WB_END)) { |
433 | nfs_unlock_request(req); |
434 | nfs_folio_end_writeback(folio: nfs_page_to_folio(req)); |
435 | } else |
436 | nfs_unlock_request(req); |
437 | } |
438 | |
439 | /* |
440 | * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests |
441 | * |
442 | * @destroy_list - request list (using wb_this_page) terminated by @old_head |
443 | * @old_head - the old head of the list |
444 | * |
445 | * All subrequests must be locked and removed from all lists, so at this point |
446 | * they are only "active" in this function, and possibly in nfs_wait_on_request |
447 | * with a reference held by some other context. |
448 | */ |
449 | static void |
450 | nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list, |
451 | struct nfs_page *old_head, |
452 | struct inode *inode) |
453 | { |
454 | while (destroy_list) { |
455 | struct nfs_page *subreq = destroy_list; |
456 | |
457 | destroy_list = (subreq->wb_this_page == old_head) ? |
458 | NULL : subreq->wb_this_page; |
459 | |
460 | /* Note: lock subreq in order to change subreq->wb_head */ |
461 | nfs_page_set_headlock(req: subreq); |
462 | WARN_ON_ONCE(old_head != subreq->wb_head); |
463 | |
464 | /* make sure old group is not used */ |
465 | subreq->wb_this_page = subreq; |
466 | subreq->wb_head = subreq; |
467 | |
468 | clear_bit(nr: PG_REMOVE, addr: &subreq->wb_flags); |
469 | |
470 | /* Note: races with nfs_page_group_destroy() */ |
471 | if (!kref_read(kref: &subreq->wb_kref)) { |
472 | /* Check if we raced with nfs_page_group_destroy() */ |
473 | if (test_and_clear_bit(nr: PG_TEARDOWN, addr: &subreq->wb_flags)) { |
474 | nfs_page_clear_headlock(req: subreq); |
475 | nfs_free_request(req: subreq); |
476 | } else |
477 | nfs_page_clear_headlock(req: subreq); |
478 | continue; |
479 | } |
480 | nfs_page_clear_headlock(req: subreq); |
481 | |
482 | nfs_release_request(old_head); |
483 | |
484 | if (test_and_clear_bit(nr: PG_INODE_REF, addr: &subreq->wb_flags)) { |
485 | nfs_release_request(subreq); |
486 | atomic_long_dec(v: &NFS_I(inode)->nrequests); |
487 | } |
488 | |
489 | /* subreq is now totally disconnected from page group or any |
490 | * write / commit lists. last chance to wake any waiters */ |
491 | nfs_unlock_and_release_request(subreq); |
492 | } |
493 | } |
494 | |
495 | /* |
496 | * nfs_join_page_group - destroy subrequests of the head req |
497 | * @head: the page used to lookup the "page group" of nfs_page structures |
498 | * @inode: Inode to which the request belongs. |
499 | * |
500 | * This function joins all sub requests to the head request by first |
501 | * locking all requests in the group, cancelling any pending operations |
502 | * and finally updating the head request to cover the whole range covered by |
503 | * the (former) group. All subrequests are removed from any write or commit |
504 | * lists, unlinked from the group and destroyed. |
505 | */ |
506 | void nfs_join_page_group(struct nfs_page *head, struct nfs_commit_info *cinfo, |
507 | struct inode *inode) |
508 | { |
509 | struct nfs_page *subreq; |
510 | struct nfs_page *destroy_list = NULL; |
511 | unsigned int pgbase, off, bytes; |
512 | |
513 | pgbase = head->wb_pgbase; |
514 | bytes = head->wb_bytes; |
515 | off = head->wb_offset; |
516 | for (subreq = head->wb_this_page; subreq != head; |
517 | subreq = subreq->wb_this_page) { |
518 | /* Subrequests should always form a contiguous range */ |
519 | if (pgbase > subreq->wb_pgbase) { |
520 | off -= pgbase - subreq->wb_pgbase; |
521 | bytes += pgbase - subreq->wb_pgbase; |
522 | pgbase = subreq->wb_pgbase; |
523 | } |
524 | bytes = max(subreq->wb_pgbase + subreq->wb_bytes |
525 | - pgbase, bytes); |
526 | } |
527 | |
528 | /* Set the head request's range to cover the former page group */ |
529 | head->wb_pgbase = pgbase; |
530 | head->wb_bytes = bytes; |
531 | head->wb_offset = off; |
532 | |
533 | /* Now that all requests are locked, make sure they aren't on any list. |
534 | * Commit list removal accounting is done after locks are dropped */ |
535 | subreq = head; |
536 | do { |
537 | nfs_clear_request_commit(cinfo, req: subreq); |
538 | subreq = subreq->wb_this_page; |
539 | } while (subreq != head); |
540 | |
541 | /* unlink subrequests from head, destroy them later */ |
542 | if (head->wb_this_page != head) { |
543 | /* destroy list will be terminated by head */ |
544 | destroy_list = head->wb_this_page; |
545 | head->wb_this_page = head; |
546 | } |
547 | |
548 | nfs_destroy_unlinked_subrequests(destroy_list, old_head: head, inode); |
549 | } |
550 | |
551 | /* |
552 | * nfs_lock_and_join_requests - join all subreqs to the head req |
553 | * @folio: the folio used to lookup the "page group" of nfs_page structures |
554 | * |
555 | * This function joins all sub requests to the head request by first |
556 | * locking all requests in the group, cancelling any pending operations |
557 | * and finally updating the head request to cover the whole range covered by |
558 | * the (former) group. All subrequests are removed from any write or commit |
559 | * lists, unlinked from the group and destroyed. |
560 | * |
561 | * Returns a locked, referenced pointer to the head request - which after |
562 | * this call is guaranteed to be the only request associated with the page. |
563 | * Returns NULL if no requests are found for @folio, or a ERR_PTR if an |
564 | * error was encountered. |
565 | */ |
566 | static struct nfs_page *nfs_lock_and_join_requests(struct folio *folio) |
567 | { |
568 | struct inode *inode = folio_file_mapping(folio)->host; |
569 | struct nfs_page *head; |
570 | struct nfs_commit_info cinfo; |
571 | int ret; |
572 | |
573 | nfs_init_cinfo_from_inode(cinfo: &cinfo, inode); |
574 | /* |
575 | * A reference is taken only on the head request which acts as a |
576 | * reference to the whole page group - the group will not be destroyed |
577 | * until the head reference is released. |
578 | */ |
579 | head = nfs_folio_find_and_lock_request(folio); |
580 | if (IS_ERR_OR_NULL(ptr: head)) |
581 | return head; |
582 | |
583 | /* lock each request in the page group */ |
584 | ret = nfs_page_group_lock_subrequests(head); |
585 | if (ret < 0) { |
586 | nfs_unlock_and_release_request(head); |
587 | return ERR_PTR(error: ret); |
588 | } |
589 | |
590 | nfs_join_page_group(head, cinfo: &cinfo, inode); |
591 | |
592 | return head; |
593 | } |
594 | |
595 | static void nfs_write_error(struct nfs_page *req, int error) |
596 | { |
597 | trace_nfs_write_error(inode: nfs_page_to_inode(req), req, error); |
598 | nfs_mapping_set_error(folio: nfs_page_to_folio(req), error); |
599 | nfs_inode_remove_request(req); |
600 | nfs_page_end_writeback(req); |
601 | nfs_release_request(req); |
602 | } |
603 | |
604 | /* |
605 | * Find an associated nfs write request, and prepare to flush it out |
606 | * May return an error if the user signalled nfs_wait_on_request(). |
607 | */ |
608 | static int nfs_page_async_flush(struct folio *folio, |
609 | struct writeback_control *wbc, |
610 | struct nfs_pageio_descriptor *pgio) |
611 | { |
612 | struct nfs_page *req; |
613 | int ret = 0; |
614 | |
615 | req = nfs_lock_and_join_requests(folio); |
616 | if (!req) |
617 | goto out; |
618 | ret = PTR_ERR(ptr: req); |
619 | if (IS_ERR(ptr: req)) |
620 | goto out; |
621 | |
622 | nfs_folio_set_writeback(folio); |
623 | WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags)); |
624 | |
625 | /* If there is a fatal error that covers this write, just exit */ |
626 | ret = pgio->pg_error; |
627 | if (nfs_error_is_fatal_on_server(err: ret)) |
628 | goto out_launder; |
629 | |
630 | ret = 0; |
631 | if (!nfs_pageio_add_request(pgio, req)) { |
632 | ret = pgio->pg_error; |
633 | /* |
634 | * Remove the problematic req upon fatal errors on the server |
635 | */ |
636 | if (nfs_error_is_fatal_on_server(err: ret)) |
637 | goto out_launder; |
638 | if (wbc->sync_mode == WB_SYNC_NONE) |
639 | ret = AOP_WRITEPAGE_ACTIVATE; |
640 | folio_redirty_for_writepage(wbc, folio); |
641 | nfs_redirty_request(req); |
642 | pgio->pg_error = 0; |
643 | } else |
644 | nfs_add_stats(inode: folio_file_mapping(folio)->host, |
645 | stat: NFSIOS_WRITEPAGES, addend: 1); |
646 | out: |
647 | return ret; |
648 | out_launder: |
649 | nfs_write_error(req, error: ret); |
650 | return 0; |
651 | } |
652 | |
653 | static int nfs_do_writepage(struct folio *folio, struct writeback_control *wbc, |
654 | struct nfs_pageio_descriptor *pgio) |
655 | { |
656 | nfs_pageio_cond_complete(pgio, folio_index(folio)); |
657 | return nfs_page_async_flush(folio, wbc, pgio); |
658 | } |
659 | |
660 | /* |
661 | * Write an mmapped page to the server. |
662 | */ |
663 | static int nfs_writepage_locked(struct folio *folio, |
664 | struct writeback_control *wbc) |
665 | { |
666 | struct nfs_pageio_descriptor pgio; |
667 | struct inode *inode = folio_file_mapping(folio)->host; |
668 | int err; |
669 | |
670 | if (wbc->sync_mode == WB_SYNC_NONE && |
671 | NFS_SERVER(inode)->write_congested) |
672 | return AOP_WRITEPAGE_ACTIVATE; |
673 | |
674 | nfs_inc_stats(inode, stat: NFSIOS_VFSWRITEPAGE); |
675 | nfs_pageio_init_write(pgio: &pgio, inode, ioflags: 0, force_mds: false, |
676 | compl_ops: &nfs_async_write_completion_ops); |
677 | err = nfs_do_writepage(folio, wbc, pgio: &pgio); |
678 | pgio.pg_error = 0; |
679 | nfs_pageio_complete(desc: &pgio); |
680 | return err; |
681 | } |
682 | |
683 | int nfs_writepage(struct page *page, struct writeback_control *wbc) |
684 | { |
685 | struct folio *folio = page_folio(page); |
686 | int ret; |
687 | |
688 | ret = nfs_writepage_locked(folio, wbc); |
689 | if (ret != AOP_WRITEPAGE_ACTIVATE) |
690 | unlock_page(page); |
691 | return ret; |
692 | } |
693 | |
694 | static int nfs_writepages_callback(struct folio *folio, |
695 | struct writeback_control *wbc, void *data) |
696 | { |
697 | int ret; |
698 | |
699 | ret = nfs_do_writepage(folio, wbc, pgio: data); |
700 | if (ret != AOP_WRITEPAGE_ACTIVATE) |
701 | folio_unlock(folio); |
702 | return ret; |
703 | } |
704 | |
705 | static void nfs_io_completion_commit(void *inode) |
706 | { |
707 | nfs_commit_inode(inode, 0); |
708 | } |
709 | |
710 | int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc) |
711 | { |
712 | struct inode *inode = mapping->host; |
713 | struct nfs_pageio_descriptor pgio; |
714 | struct nfs_io_completion *ioc = NULL; |
715 | unsigned int mntflags = NFS_SERVER(inode)->flags; |
716 | int priority = 0; |
717 | int err; |
718 | |
719 | if (wbc->sync_mode == WB_SYNC_NONE && |
720 | NFS_SERVER(inode)->write_congested) |
721 | return 0; |
722 | |
723 | nfs_inc_stats(inode, stat: NFSIOS_VFSWRITEPAGES); |
724 | |
725 | if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate || |
726 | wbc->for_background || wbc->for_sync || wbc->for_reclaim) { |
727 | ioc = nfs_io_completion_alloc(GFP_KERNEL); |
728 | if (ioc) |
729 | nfs_io_completion_init(ioc, complete: nfs_io_completion_commit, |
730 | data: inode); |
731 | priority = wb_priority(wbc); |
732 | } |
733 | |
734 | do { |
735 | nfs_pageio_init_write(pgio: &pgio, inode, ioflags: priority, force_mds: false, |
736 | compl_ops: &nfs_async_write_completion_ops); |
737 | pgio.pg_io_completion = ioc; |
738 | err = write_cache_pages(mapping, wbc, writepage: nfs_writepages_callback, |
739 | data: &pgio); |
740 | pgio.pg_error = 0; |
741 | nfs_pageio_complete(desc: &pgio); |
742 | if (err == -EAGAIN && mntflags & NFS_MOUNT_SOFTERR) |
743 | break; |
744 | } while (err < 0 && !nfs_error_is_fatal(err)); |
745 | nfs_io_completion_put(ioc); |
746 | |
747 | if (err < 0) |
748 | goto out_err; |
749 | return 0; |
750 | out_err: |
751 | return err; |
752 | } |
753 | |
754 | /* |
755 | * Insert a write request into an inode |
756 | */ |
757 | static void nfs_inode_add_request(struct nfs_page *req) |
758 | { |
759 | struct folio *folio = nfs_page_to_folio(req); |
760 | struct address_space *mapping = folio_file_mapping(folio); |
761 | struct nfs_inode *nfsi = NFS_I(inode: mapping->host); |
762 | |
763 | WARN_ON_ONCE(req->wb_this_page != req); |
764 | |
765 | /* Lock the request! */ |
766 | nfs_lock_request(req); |
767 | |
768 | /* |
769 | * Swap-space should not get truncated. Hence no need to plug the race |
770 | * with invalidate/truncate. |
771 | */ |
772 | spin_lock(lock: &mapping->private_lock); |
773 | if (likely(!folio_test_swapcache(folio))) { |
774 | set_bit(nr: PG_MAPPED, addr: &req->wb_flags); |
775 | folio_set_private(folio); |
776 | folio->private = req; |
777 | } |
778 | spin_unlock(lock: &mapping->private_lock); |
779 | atomic_long_inc(v: &nfsi->nrequests); |
780 | /* this a head request for a page group - mark it as having an |
781 | * extra reference so sub groups can follow suit. |
782 | * This flag also informs pgio layer when to bump nrequests when |
783 | * adding subrequests. */ |
784 | WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags)); |
785 | kref_get(kref: &req->wb_kref); |
786 | } |
787 | |
788 | /* |
789 | * Remove a write request from an inode |
790 | */ |
791 | static void nfs_inode_remove_request(struct nfs_page *req) |
792 | { |
793 | struct nfs_inode *nfsi = NFS_I(inode: nfs_page_to_inode(req)); |
794 | |
795 | if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) { |
796 | struct folio *folio = nfs_page_to_folio(req: req->wb_head); |
797 | struct address_space *mapping = folio_file_mapping(folio); |
798 | |
799 | spin_lock(lock: &mapping->private_lock); |
800 | if (likely(folio && !folio_test_swapcache(folio))) { |
801 | folio->private = NULL; |
802 | folio_clear_private(folio); |
803 | clear_bit(nr: PG_MAPPED, addr: &req->wb_head->wb_flags); |
804 | } |
805 | spin_unlock(lock: &mapping->private_lock); |
806 | } |
807 | |
808 | if (test_and_clear_bit(nr: PG_INODE_REF, addr: &req->wb_flags)) { |
809 | atomic_long_dec(v: &nfsi->nrequests); |
810 | nfs_release_request(req); |
811 | } |
812 | } |
813 | |
814 | static void nfs_mark_request_dirty(struct nfs_page *req) |
815 | { |
816 | struct folio *folio = nfs_page_to_folio(req); |
817 | if (folio) |
818 | filemap_dirty_folio(mapping: folio_mapping(folio), folio); |
819 | } |
820 | |
821 | /* |
822 | * nfs_page_search_commits_for_head_request_locked |
823 | * |
824 | * Search through commit lists on @inode for the head request for @folio. |
825 | * Must be called while holding the inode (which is cinfo) lock. |
826 | * |
827 | * Returns the head request if found, or NULL if not found. |
828 | */ |
829 | static struct nfs_page * |
830 | nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi, |
831 | struct folio *folio) |
832 | { |
833 | struct nfs_page *freq, *t; |
834 | struct nfs_commit_info cinfo; |
835 | struct inode *inode = &nfsi->vfs_inode; |
836 | |
837 | nfs_init_cinfo_from_inode(cinfo: &cinfo, inode); |
838 | |
839 | /* search through pnfs commit lists */ |
840 | freq = pnfs_search_commit_reqs(inode, cinfo: &cinfo, folio); |
841 | if (freq) |
842 | return freq->wb_head; |
843 | |
844 | /* Linearly search the commit list for the correct request */ |
845 | list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) { |
846 | if (nfs_page_to_folio(req: freq) == folio) |
847 | return freq->wb_head; |
848 | } |
849 | |
850 | return NULL; |
851 | } |
852 | |
853 | /** |
854 | * nfs_request_add_commit_list_locked - add request to a commit list |
855 | * @req: pointer to a struct nfs_page |
856 | * @dst: commit list head |
857 | * @cinfo: holds list lock and accounting info |
858 | * |
859 | * This sets the PG_CLEAN bit, updates the cinfo count of |
860 | * number of outstanding requests requiring a commit as well as |
861 | * the MM page stats. |
862 | * |
863 | * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the |
864 | * nfs_page lock. |
865 | */ |
866 | void |
867 | nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst, |
868 | struct nfs_commit_info *cinfo) |
869 | { |
870 | set_bit(nr: PG_CLEAN, addr: &req->wb_flags); |
871 | nfs_list_add_request(req, head: dst); |
872 | atomic_long_inc(v: &cinfo->mds->ncommit); |
873 | } |
874 | EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked); |
875 | |
876 | /** |
877 | * nfs_request_add_commit_list - add request to a commit list |
878 | * @req: pointer to a struct nfs_page |
879 | * @cinfo: holds list lock and accounting info |
880 | * |
881 | * This sets the PG_CLEAN bit, updates the cinfo count of |
882 | * number of outstanding requests requiring a commit as well as |
883 | * the MM page stats. |
884 | * |
885 | * The caller must _not_ hold the cinfo->lock, but must be |
886 | * holding the nfs_page lock. |
887 | */ |
888 | void |
889 | nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo) |
890 | { |
891 | mutex_lock(&NFS_I(cinfo->inode)->commit_mutex); |
892 | nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo); |
893 | mutex_unlock(lock: &NFS_I(inode: cinfo->inode)->commit_mutex); |
894 | nfs_folio_mark_unstable(folio: nfs_page_to_folio(req), cinfo); |
895 | } |
896 | EXPORT_SYMBOL_GPL(nfs_request_add_commit_list); |
897 | |
898 | /** |
899 | * nfs_request_remove_commit_list - Remove request from a commit list |
900 | * @req: pointer to a nfs_page |
901 | * @cinfo: holds list lock and accounting info |
902 | * |
903 | * This clears the PG_CLEAN bit, and updates the cinfo's count of |
904 | * number of outstanding requests requiring a commit |
905 | * It does not update the MM page stats. |
906 | * |
907 | * The caller _must_ hold the cinfo->lock and the nfs_page lock. |
908 | */ |
909 | void |
910 | nfs_request_remove_commit_list(struct nfs_page *req, |
911 | struct nfs_commit_info *cinfo) |
912 | { |
913 | if (!test_and_clear_bit(nr: PG_CLEAN, addr: &(req)->wb_flags)) |
914 | return; |
915 | nfs_list_remove_request(req); |
916 | atomic_long_dec(v: &cinfo->mds->ncommit); |
917 | } |
918 | EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list); |
919 | |
920 | static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo, |
921 | struct inode *inode) |
922 | { |
923 | cinfo->inode = inode; |
924 | cinfo->mds = &NFS_I(inode)->commit_info; |
925 | cinfo->ds = pnfs_get_ds_info(inode); |
926 | cinfo->dreq = NULL; |
927 | cinfo->completion_ops = &nfs_commit_completion_ops; |
928 | } |
929 | |
930 | void nfs_init_cinfo(struct nfs_commit_info *cinfo, |
931 | struct inode *inode, |
932 | struct nfs_direct_req *dreq) |
933 | { |
934 | if (dreq) |
935 | nfs_init_cinfo_from_dreq(cinfo, dreq); |
936 | else |
937 | nfs_init_cinfo_from_inode(cinfo, inode); |
938 | } |
939 | EXPORT_SYMBOL_GPL(nfs_init_cinfo); |
940 | |
941 | /* |
942 | * Add a request to the inode's commit list. |
943 | */ |
944 | void |
945 | nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg, |
946 | struct nfs_commit_info *cinfo, u32 ds_commit_idx) |
947 | { |
948 | if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx)) |
949 | return; |
950 | nfs_request_add_commit_list(req, cinfo); |
951 | } |
952 | |
953 | static void nfs_folio_clear_commit(struct folio *folio) |
954 | { |
955 | if (folio) { |
956 | long nr = folio_nr_pages(folio); |
957 | |
958 | node_stat_mod_folio(folio, item: NR_WRITEBACK, nr: -nr); |
959 | wb_stat_mod(wb: &inode_to_bdi(inode: folio_file_mapping(folio)->host)->wb, |
960 | item: WB_WRITEBACK, amount: -nr); |
961 | } |
962 | } |
963 | |
964 | /* Called holding the request lock on @req */ |
965 | static void nfs_clear_request_commit(struct nfs_commit_info *cinfo, |
966 | struct nfs_page *req) |
967 | { |
968 | if (test_bit(PG_CLEAN, &req->wb_flags)) { |
969 | struct nfs_open_context *ctx = nfs_req_openctx(req); |
970 | struct inode *inode = d_inode(dentry: ctx->dentry); |
971 | |
972 | mutex_lock(&NFS_I(inode)->commit_mutex); |
973 | if (!pnfs_clear_request_commit(req, cinfo)) { |
974 | nfs_request_remove_commit_list(req, cinfo); |
975 | } |
976 | mutex_unlock(lock: &NFS_I(inode)->commit_mutex); |
977 | nfs_folio_clear_commit(folio: nfs_page_to_folio(req)); |
978 | } |
979 | } |
980 | |
981 | int nfs_write_need_commit(struct nfs_pgio_header *hdr) |
982 | { |
983 | if (hdr->verf.committed == NFS_DATA_SYNC) |
984 | return hdr->lseg == NULL; |
985 | return hdr->verf.committed != NFS_FILE_SYNC; |
986 | } |
987 | |
988 | static void nfs_async_write_init(struct nfs_pgio_header *hdr) |
989 | { |
990 | nfs_io_completion_get(ioc: hdr->io_completion); |
991 | } |
992 | |
993 | static void nfs_write_completion(struct nfs_pgio_header *hdr) |
994 | { |
995 | struct nfs_commit_info cinfo; |
996 | unsigned long bytes = 0; |
997 | |
998 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) |
999 | goto out; |
1000 | nfs_init_cinfo_from_inode(cinfo: &cinfo, inode: hdr->inode); |
1001 | while (!list_empty(head: &hdr->pages)) { |
1002 | struct nfs_page *req = nfs_list_entry(head: hdr->pages.next); |
1003 | |
1004 | bytes += req->wb_bytes; |
1005 | nfs_list_remove_request(req); |
1006 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && |
1007 | (hdr->good_bytes < bytes)) { |
1008 | trace_nfs_comp_error(inode: hdr->inode, req, error: hdr->error); |
1009 | nfs_mapping_set_error(folio: nfs_page_to_folio(req), |
1010 | error: hdr->error); |
1011 | goto remove_req; |
1012 | } |
1013 | if (nfs_write_need_commit(hdr)) { |
1014 | /* Reset wb_nio, since the write was successful. */ |
1015 | req->wb_nio = 0; |
1016 | memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf)); |
1017 | nfs_mark_request_commit(req, lseg: hdr->lseg, cinfo: &cinfo, |
1018 | ds_commit_idx: hdr->pgio_mirror_idx); |
1019 | goto next; |
1020 | } |
1021 | remove_req: |
1022 | nfs_inode_remove_request(req); |
1023 | next: |
1024 | nfs_page_end_writeback(req); |
1025 | nfs_release_request(req); |
1026 | } |
1027 | out: |
1028 | nfs_io_completion_put(ioc: hdr->io_completion); |
1029 | hdr->release(hdr); |
1030 | } |
1031 | |
1032 | unsigned long |
1033 | nfs_reqs_to_commit(struct nfs_commit_info *cinfo) |
1034 | { |
1035 | return atomic_long_read(v: &cinfo->mds->ncommit); |
1036 | } |
1037 | |
1038 | /* NFS_I(cinfo->inode)->commit_mutex held by caller */ |
1039 | int |
1040 | nfs_scan_commit_list(struct list_head *src, struct list_head *dst, |
1041 | struct nfs_commit_info *cinfo, int max) |
1042 | { |
1043 | struct nfs_page *req, *tmp; |
1044 | int ret = 0; |
1045 | |
1046 | list_for_each_entry_safe(req, tmp, src, wb_list) { |
1047 | kref_get(kref: &req->wb_kref); |
1048 | if (!nfs_lock_request(req)) { |
1049 | nfs_release_request(req); |
1050 | continue; |
1051 | } |
1052 | nfs_request_remove_commit_list(req, cinfo); |
1053 | clear_bit(nr: PG_COMMIT_TO_DS, addr: &req->wb_flags); |
1054 | nfs_list_add_request(req, head: dst); |
1055 | ret++; |
1056 | if ((ret == max) && !cinfo->dreq) |
1057 | break; |
1058 | cond_resched(); |
1059 | } |
1060 | return ret; |
1061 | } |
1062 | EXPORT_SYMBOL_GPL(nfs_scan_commit_list); |
1063 | |
1064 | /* |
1065 | * nfs_scan_commit - Scan an inode for commit requests |
1066 | * @inode: NFS inode to scan |
1067 | * @dst: mds destination list |
1068 | * @cinfo: mds and ds lists of reqs ready to commit |
1069 | * |
1070 | * Moves requests from the inode's 'commit' request list. |
1071 | * The requests are *not* checked to ensure that they form a contiguous set. |
1072 | */ |
1073 | int |
1074 | nfs_scan_commit(struct inode *inode, struct list_head *dst, |
1075 | struct nfs_commit_info *cinfo) |
1076 | { |
1077 | int ret = 0; |
1078 | |
1079 | if (!atomic_long_read(v: &cinfo->mds->ncommit)) |
1080 | return 0; |
1081 | mutex_lock(&NFS_I(cinfo->inode)->commit_mutex); |
1082 | if (atomic_long_read(v: &cinfo->mds->ncommit) > 0) { |
1083 | const int max = INT_MAX; |
1084 | |
1085 | ret = nfs_scan_commit_list(&cinfo->mds->list, dst, |
1086 | cinfo, max); |
1087 | ret += pnfs_scan_commit_lists(inode, cinfo, max: max - ret); |
1088 | } |
1089 | mutex_unlock(lock: &NFS_I(inode: cinfo->inode)->commit_mutex); |
1090 | return ret; |
1091 | } |
1092 | |
1093 | /* |
1094 | * Search for an existing write request, and attempt to update |
1095 | * it to reflect a new dirty region on a given page. |
1096 | * |
1097 | * If the attempt fails, then the existing request is flushed out |
1098 | * to disk. |
1099 | */ |
1100 | static struct nfs_page *nfs_try_to_update_request(struct folio *folio, |
1101 | unsigned int offset, |
1102 | unsigned int bytes) |
1103 | { |
1104 | struct nfs_page *req; |
1105 | unsigned int rqend; |
1106 | unsigned int end; |
1107 | int error; |
1108 | |
1109 | end = offset + bytes; |
1110 | |
1111 | req = nfs_lock_and_join_requests(folio); |
1112 | if (IS_ERR_OR_NULL(ptr: req)) |
1113 | return req; |
1114 | |
1115 | rqend = req->wb_offset + req->wb_bytes; |
1116 | /* |
1117 | * Tell the caller to flush out the request if |
1118 | * the offsets are non-contiguous. |
1119 | * Note: nfs_flush_incompatible() will already |
1120 | * have flushed out requests having wrong owners. |
1121 | */ |
1122 | if (offset > rqend || end < req->wb_offset) |
1123 | goto out_flushme; |
1124 | |
1125 | /* Okay, the request matches. Update the region */ |
1126 | if (offset < req->wb_offset) { |
1127 | req->wb_offset = offset; |
1128 | req->wb_pgbase = offset; |
1129 | } |
1130 | if (end > rqend) |
1131 | req->wb_bytes = end - req->wb_offset; |
1132 | else |
1133 | req->wb_bytes = rqend - req->wb_offset; |
1134 | req->wb_nio = 0; |
1135 | return req; |
1136 | out_flushme: |
1137 | /* |
1138 | * Note: we mark the request dirty here because |
1139 | * nfs_lock_and_join_requests() cannot preserve |
1140 | * commit flags, so we have to replay the write. |
1141 | */ |
1142 | nfs_mark_request_dirty(req); |
1143 | nfs_unlock_and_release_request(req); |
1144 | error = nfs_wb_folio(inode: folio_file_mapping(folio)->host, folio); |
1145 | return (error < 0) ? ERR_PTR(error) : NULL; |
1146 | } |
1147 | |
1148 | /* |
1149 | * Try to update an existing write request, or create one if there is none. |
1150 | * |
1151 | * Note: Should always be called with the Page Lock held to prevent races |
1152 | * if we have to add a new request. Also assumes that the caller has |
1153 | * already called nfs_flush_incompatible() if necessary. |
1154 | */ |
1155 | static struct nfs_page *nfs_setup_write_request(struct nfs_open_context *ctx, |
1156 | struct folio *folio, |
1157 | unsigned int offset, |
1158 | unsigned int bytes) |
1159 | { |
1160 | struct nfs_page *req; |
1161 | |
1162 | req = nfs_try_to_update_request(folio, offset, bytes); |
1163 | if (req != NULL) |
1164 | goto out; |
1165 | req = nfs_page_create_from_folio(ctx, folio, offset, count: bytes); |
1166 | if (IS_ERR(ptr: req)) |
1167 | goto out; |
1168 | nfs_inode_add_request(req); |
1169 | out: |
1170 | return req; |
1171 | } |
1172 | |
1173 | static int nfs_writepage_setup(struct nfs_open_context *ctx, |
1174 | struct folio *folio, unsigned int offset, |
1175 | unsigned int count) |
1176 | { |
1177 | struct nfs_page *req; |
1178 | |
1179 | req = nfs_setup_write_request(ctx, folio, offset, bytes: count); |
1180 | if (IS_ERR(ptr: req)) |
1181 | return PTR_ERR(ptr: req); |
1182 | /* Update file length */ |
1183 | nfs_grow_file(folio, offset, count); |
1184 | nfs_mark_uptodate(req); |
1185 | nfs_mark_request_dirty(req); |
1186 | nfs_unlock_and_release_request(req); |
1187 | return 0; |
1188 | } |
1189 | |
1190 | int nfs_flush_incompatible(struct file *file, struct folio *folio) |
1191 | { |
1192 | struct nfs_open_context *ctx = nfs_file_open_context(filp: file); |
1193 | struct nfs_lock_context *l_ctx; |
1194 | struct file_lock_context *flctx = locks_inode_context(inode: file_inode(f: file)); |
1195 | struct nfs_page *req; |
1196 | int do_flush, status; |
1197 | /* |
1198 | * Look for a request corresponding to this page. If there |
1199 | * is one, and it belongs to another file, we flush it out |
1200 | * before we try to copy anything into the page. Do this |
1201 | * due to the lack of an ACCESS-type call in NFSv2. |
1202 | * Also do the same if we find a request from an existing |
1203 | * dropped page. |
1204 | */ |
1205 | do { |
1206 | req = nfs_folio_find_head_request(folio); |
1207 | if (req == NULL) |
1208 | return 0; |
1209 | l_ctx = req->wb_lock_context; |
1210 | do_flush = nfs_page_to_folio(req) != folio || |
1211 | !nfs_match_open_context(ctx1: nfs_req_openctx(req), ctx2: ctx); |
1212 | if (l_ctx && flctx && |
1213 | !(list_empty_careful(head: &flctx->flc_posix) && |
1214 | list_empty_careful(head: &flctx->flc_flock))) { |
1215 | do_flush |= l_ctx->lockowner != current->files; |
1216 | } |
1217 | nfs_release_request(req); |
1218 | if (!do_flush) |
1219 | return 0; |
1220 | status = nfs_wb_folio(inode: folio_file_mapping(folio)->host, folio); |
1221 | } while (status == 0); |
1222 | return status; |
1223 | } |
1224 | |
1225 | /* |
1226 | * Avoid buffered writes when a open context credential's key would |
1227 | * expire soon. |
1228 | * |
1229 | * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL. |
1230 | * |
1231 | * Return 0 and set a credential flag which triggers the inode to flush |
1232 | * and performs NFS_FILE_SYNC writes if the key will expired within |
1233 | * RPC_KEY_EXPIRE_TIMEO. |
1234 | */ |
1235 | int |
1236 | nfs_key_timeout_notify(struct file *filp, struct inode *inode) |
1237 | { |
1238 | struct nfs_open_context *ctx = nfs_file_open_context(filp); |
1239 | |
1240 | if (nfs_ctx_key_to_expire(ctx, inode) && |
1241 | !rcu_access_pointer(ctx->ll_cred)) |
1242 | /* Already expired! */ |
1243 | return -EACCES; |
1244 | return 0; |
1245 | } |
1246 | |
1247 | /* |
1248 | * Test if the open context credential key is marked to expire soon. |
1249 | */ |
1250 | bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode) |
1251 | { |
1252 | struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth; |
1253 | struct rpc_cred *cred, *new, *old = NULL; |
1254 | struct auth_cred acred = { |
1255 | .cred = ctx->cred, |
1256 | }; |
1257 | bool ret = false; |
1258 | |
1259 | rcu_read_lock(); |
1260 | cred = rcu_dereference(ctx->ll_cred); |
1261 | if (cred && !(cred->cr_ops->crkey_timeout && |
1262 | cred->cr_ops->crkey_timeout(cred))) |
1263 | goto out; |
1264 | rcu_read_unlock(); |
1265 | |
1266 | new = auth->au_ops->lookup_cred(auth, &acred, 0); |
1267 | if (new == cred) { |
1268 | put_rpccred(new); |
1269 | return true; |
1270 | } |
1271 | if (IS_ERR_OR_NULL(ptr: new)) { |
1272 | new = NULL; |
1273 | ret = true; |
1274 | } else if (new->cr_ops->crkey_timeout && |
1275 | new->cr_ops->crkey_timeout(new)) |
1276 | ret = true; |
1277 | |
1278 | rcu_read_lock(); |
1279 | old = rcu_dereference_protected(xchg(&ctx->ll_cred, |
1280 | RCU_INITIALIZER(new)), 1); |
1281 | out: |
1282 | rcu_read_unlock(); |
1283 | put_rpccred(old); |
1284 | return ret; |
1285 | } |
1286 | |
1287 | /* |
1288 | * If the page cache is marked as unsafe or invalid, then we can't rely on |
1289 | * the PageUptodate() flag. In this case, we will need to turn off |
1290 | * write optimisations that depend on the page contents being correct. |
1291 | */ |
1292 | static bool nfs_folio_write_uptodate(struct folio *folio, unsigned int pagelen) |
1293 | { |
1294 | struct inode *inode = folio_file_mapping(folio)->host; |
1295 | struct nfs_inode *nfsi = NFS_I(inode); |
1296 | |
1297 | if (nfs_have_delegated_attributes(inode)) |
1298 | goto out; |
1299 | if (nfsi->cache_validity & |
1300 | (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE)) |
1301 | return false; |
1302 | smp_rmb(); |
1303 | if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0) |
1304 | return false; |
1305 | out: |
1306 | if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0) |
1307 | return false; |
1308 | return folio_test_uptodate(folio) != 0; |
1309 | } |
1310 | |
1311 | static bool |
1312 | is_whole_file_wrlock(struct file_lock *fl) |
1313 | { |
1314 | return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX && |
1315 | fl->fl_type == F_WRLCK; |
1316 | } |
1317 | |
1318 | /* If we know the page is up to date, and we're not using byte range locks (or |
1319 | * if we have the whole file locked for writing), it may be more efficient to |
1320 | * extend the write to cover the entire page in order to avoid fragmentation |
1321 | * inefficiencies. |
1322 | * |
1323 | * If the file is opened for synchronous writes then we can just skip the rest |
1324 | * of the checks. |
1325 | */ |
1326 | static int nfs_can_extend_write(struct file *file, struct folio *folio, |
1327 | unsigned int pagelen) |
1328 | { |
1329 | struct inode *inode = file_inode(f: file); |
1330 | struct file_lock_context *flctx = locks_inode_context(inode); |
1331 | struct file_lock *fl; |
1332 | int ret; |
1333 | |
1334 | if (file->f_flags & O_DSYNC) |
1335 | return 0; |
1336 | if (!nfs_folio_write_uptodate(folio, pagelen)) |
1337 | return 0; |
1338 | if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE)) |
1339 | return 1; |
1340 | if (!flctx || (list_empty_careful(head: &flctx->flc_flock) && |
1341 | list_empty_careful(head: &flctx->flc_posix))) |
1342 | return 1; |
1343 | |
1344 | /* Check to see if there are whole file write locks */ |
1345 | ret = 0; |
1346 | spin_lock(lock: &flctx->flc_lock); |
1347 | if (!list_empty(head: &flctx->flc_posix)) { |
1348 | fl = list_first_entry(&flctx->flc_posix, struct file_lock, |
1349 | fl_list); |
1350 | if (is_whole_file_wrlock(fl)) |
1351 | ret = 1; |
1352 | } else if (!list_empty(head: &flctx->flc_flock)) { |
1353 | fl = list_first_entry(&flctx->flc_flock, struct file_lock, |
1354 | fl_list); |
1355 | if (fl->fl_type == F_WRLCK) |
1356 | ret = 1; |
1357 | } |
1358 | spin_unlock(lock: &flctx->flc_lock); |
1359 | return ret; |
1360 | } |
1361 | |
1362 | /* |
1363 | * Update and possibly write a cached page of an NFS file. |
1364 | * |
1365 | * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad |
1366 | * things with a page scheduled for an RPC call (e.g. invalidate it). |
1367 | */ |
1368 | int nfs_update_folio(struct file *file, struct folio *folio, |
1369 | unsigned int offset, unsigned int count) |
1370 | { |
1371 | struct nfs_open_context *ctx = nfs_file_open_context(filp: file); |
1372 | struct address_space *mapping = folio_file_mapping(folio); |
1373 | struct inode *inode = mapping->host; |
1374 | unsigned int pagelen = nfs_folio_length(folio); |
1375 | int status = 0; |
1376 | |
1377 | nfs_inc_stats(inode, stat: NFSIOS_VFSUPDATEPAGE); |
1378 | |
1379 | dprintk("NFS: nfs_update_folio(%pD2 %d@%lld)\n" , file, count, |
1380 | (long long)(folio_file_pos(folio) + offset)); |
1381 | |
1382 | if (!count) |
1383 | goto out; |
1384 | |
1385 | if (nfs_can_extend_write(file, folio, pagelen)) { |
1386 | count = max(count + offset, pagelen); |
1387 | offset = 0; |
1388 | } |
1389 | |
1390 | status = nfs_writepage_setup(ctx, folio, offset, count); |
1391 | if (status < 0) |
1392 | nfs_set_pageerror(mapping); |
1393 | out: |
1394 | dprintk("NFS: nfs_update_folio returns %d (isize %lld)\n" , |
1395 | status, (long long)i_size_read(inode)); |
1396 | return status; |
1397 | } |
1398 | |
1399 | static int flush_task_priority(int how) |
1400 | { |
1401 | switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) { |
1402 | case FLUSH_HIGHPRI: |
1403 | return RPC_PRIORITY_HIGH; |
1404 | case FLUSH_LOWPRI: |
1405 | return RPC_PRIORITY_LOW; |
1406 | } |
1407 | return RPC_PRIORITY_NORMAL; |
1408 | } |
1409 | |
1410 | static void nfs_initiate_write(struct nfs_pgio_header *hdr, |
1411 | struct rpc_message *msg, |
1412 | const struct nfs_rpc_ops *rpc_ops, |
1413 | struct rpc_task_setup *task_setup_data, int how) |
1414 | { |
1415 | int priority = flush_task_priority(how); |
1416 | |
1417 | if (IS_SWAPFILE(hdr->inode)) |
1418 | task_setup_data->flags |= RPC_TASK_SWAPPER; |
1419 | task_setup_data->priority = priority; |
1420 | rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client); |
1421 | trace_nfs_initiate_write(hdr); |
1422 | } |
1423 | |
1424 | /* If a nfs_flush_* function fails, it should remove reqs from @head and |
1425 | * call this on each, which will prepare them to be retried on next |
1426 | * writeback using standard nfs. |
1427 | */ |
1428 | static void nfs_redirty_request(struct nfs_page *req) |
1429 | { |
1430 | struct nfs_inode *nfsi = NFS_I(inode: nfs_page_to_inode(req)); |
1431 | |
1432 | /* Bump the transmission count */ |
1433 | req->wb_nio++; |
1434 | nfs_mark_request_dirty(req); |
1435 | atomic_long_inc(v: &nfsi->redirtied_pages); |
1436 | nfs_page_end_writeback(req); |
1437 | nfs_release_request(req); |
1438 | } |
1439 | |
1440 | static void nfs_async_write_error(struct list_head *head, int error) |
1441 | { |
1442 | struct nfs_page *req; |
1443 | |
1444 | while (!list_empty(head)) { |
1445 | req = nfs_list_entry(head: head->next); |
1446 | nfs_list_remove_request(req); |
1447 | if (nfs_error_is_fatal_on_server(err: error)) |
1448 | nfs_write_error(req, error); |
1449 | else |
1450 | nfs_redirty_request(req); |
1451 | } |
1452 | } |
1453 | |
1454 | static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr) |
1455 | { |
1456 | nfs_async_write_error(head: &hdr->pages, error: 0); |
1457 | } |
1458 | |
1459 | static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = { |
1460 | .init_hdr = nfs_async_write_init, |
1461 | .error_cleanup = nfs_async_write_error, |
1462 | .completion = nfs_write_completion, |
1463 | .reschedule_io = nfs_async_write_reschedule_io, |
1464 | }; |
1465 | |
1466 | void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, |
1467 | struct inode *inode, int ioflags, bool force_mds, |
1468 | const struct nfs_pgio_completion_ops *compl_ops) |
1469 | { |
1470 | struct nfs_server *server = NFS_SERVER(inode); |
1471 | const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops; |
1472 | |
1473 | #ifdef CONFIG_NFS_V4_1 |
1474 | if (server->pnfs_curr_ld && !force_mds) |
1475 | pg_ops = server->pnfs_curr_ld->pg_write_ops; |
1476 | #endif |
1477 | nfs_pageio_init(desc: pgio, inode, pg_ops, compl_ops, rw_ops: &nfs_rw_write_ops, |
1478 | bsize: server->wsize, how: ioflags); |
1479 | } |
1480 | EXPORT_SYMBOL_GPL(nfs_pageio_init_write); |
1481 | |
1482 | void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio) |
1483 | { |
1484 | struct nfs_pgio_mirror *mirror; |
1485 | |
1486 | if (pgio->pg_ops && pgio->pg_ops->pg_cleanup) |
1487 | pgio->pg_ops->pg_cleanup(pgio); |
1488 | |
1489 | pgio->pg_ops = &nfs_pgio_rw_ops; |
1490 | |
1491 | nfs_pageio_stop_mirroring(pgio); |
1492 | |
1493 | mirror = &pgio->pg_mirrors[0]; |
1494 | mirror->pg_bsize = NFS_SERVER(inode: pgio->pg_inode)->wsize; |
1495 | } |
1496 | EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds); |
1497 | |
1498 | |
1499 | void nfs_commit_prepare(struct rpc_task *task, void *calldata) |
1500 | { |
1501 | struct nfs_commit_data *data = calldata; |
1502 | |
1503 | NFS_PROTO(inode: data->inode)->commit_rpc_prepare(task, data); |
1504 | } |
1505 | |
1506 | static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr, |
1507 | struct nfs_fattr *fattr) |
1508 | { |
1509 | struct nfs_pgio_args *argp = &hdr->args; |
1510 | struct nfs_pgio_res *resp = &hdr->res; |
1511 | u64 size = argp->offset + resp->count; |
1512 | |
1513 | if (!(fattr->valid & NFS_ATTR_FATTR_SIZE)) |
1514 | fattr->size = size; |
1515 | if (nfs_size_to_loff_t(size: fattr->size) < i_size_read(inode: hdr->inode)) { |
1516 | fattr->valid &= ~NFS_ATTR_FATTR_SIZE; |
1517 | return; |
1518 | } |
1519 | if (size != fattr->size) |
1520 | return; |
1521 | /* Set attribute barrier */ |
1522 | nfs_fattr_set_barrier(fattr); |
1523 | /* ...and update size */ |
1524 | fattr->valid |= NFS_ATTR_FATTR_SIZE; |
1525 | } |
1526 | |
1527 | void nfs_writeback_update_inode(struct nfs_pgio_header *hdr) |
1528 | { |
1529 | struct nfs_fattr *fattr = &hdr->fattr; |
1530 | struct inode *inode = hdr->inode; |
1531 | |
1532 | spin_lock(lock: &inode->i_lock); |
1533 | nfs_writeback_check_extend(hdr, fattr); |
1534 | nfs_post_op_update_inode_force_wcc_locked(inode, fattr); |
1535 | spin_unlock(lock: &inode->i_lock); |
1536 | } |
1537 | EXPORT_SYMBOL_GPL(nfs_writeback_update_inode); |
1538 | |
1539 | /* |
1540 | * This function is called when the WRITE call is complete. |
1541 | */ |
1542 | static int nfs_writeback_done(struct rpc_task *task, |
1543 | struct nfs_pgio_header *hdr, |
1544 | struct inode *inode) |
1545 | { |
1546 | int status; |
1547 | |
1548 | /* |
1549 | * ->write_done will attempt to use post-op attributes to detect |
1550 | * conflicting writes by other clients. A strict interpretation |
1551 | * of close-to-open would allow us to continue caching even if |
1552 | * another writer had changed the file, but some applications |
1553 | * depend on tighter cache coherency when writing. |
1554 | */ |
1555 | status = NFS_PROTO(inode)->write_done(task, hdr); |
1556 | if (status != 0) |
1557 | return status; |
1558 | |
1559 | nfs_add_stats(inode, stat: NFSIOS_SERVERWRITTENBYTES, addend: hdr->res.count); |
1560 | trace_nfs_writeback_done(task, hdr); |
1561 | |
1562 | if (task->tk_status >= 0) { |
1563 | enum nfs3_stable_how committed = hdr->res.verf->committed; |
1564 | |
1565 | if (committed == NFS_UNSTABLE) { |
1566 | /* |
1567 | * We have some uncommitted data on the server at |
1568 | * this point, so ensure that we keep track of that |
1569 | * fact irrespective of what later writes do. |
1570 | */ |
1571 | set_bit(nr: NFS_IOHDR_UNSTABLE_WRITES, addr: &hdr->flags); |
1572 | } |
1573 | |
1574 | if (committed < hdr->args.stable) { |
1575 | /* We tried a write call, but the server did not |
1576 | * commit data to stable storage even though we |
1577 | * requested it. |
1578 | * Note: There is a known bug in Tru64 < 5.0 in which |
1579 | * the server reports NFS_DATA_SYNC, but performs |
1580 | * NFS_FILE_SYNC. We therefore implement this checking |
1581 | * as a dprintk() in order to avoid filling syslog. |
1582 | */ |
1583 | static unsigned long complain; |
1584 | |
1585 | /* Note this will print the MDS for a DS write */ |
1586 | if (time_before(complain, jiffies)) { |
1587 | dprintk("NFS: faulty NFS server %s:" |
1588 | " (committed = %d) != (stable = %d)\n" , |
1589 | NFS_SERVER(inode)->nfs_client->cl_hostname, |
1590 | committed, hdr->args.stable); |
1591 | complain = jiffies + 300 * HZ; |
1592 | } |
1593 | } |
1594 | } |
1595 | |
1596 | /* Deal with the suid/sgid bit corner case */ |
1597 | if (nfs_should_remove_suid(inode)) { |
1598 | spin_lock(lock: &inode->i_lock); |
1599 | nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE); |
1600 | spin_unlock(lock: &inode->i_lock); |
1601 | } |
1602 | return 0; |
1603 | } |
1604 | |
1605 | /* |
1606 | * This function is called when the WRITE call is complete. |
1607 | */ |
1608 | static void nfs_writeback_result(struct rpc_task *task, |
1609 | struct nfs_pgio_header *hdr) |
1610 | { |
1611 | struct nfs_pgio_args *argp = &hdr->args; |
1612 | struct nfs_pgio_res *resp = &hdr->res; |
1613 | |
1614 | if (resp->count < argp->count) { |
1615 | static unsigned long complain; |
1616 | |
1617 | /* This a short write! */ |
1618 | nfs_inc_stats(inode: hdr->inode, stat: NFSIOS_SHORTWRITE); |
1619 | |
1620 | /* Has the server at least made some progress? */ |
1621 | if (resp->count == 0) { |
1622 | if (time_before(complain, jiffies)) { |
1623 | printk(KERN_WARNING |
1624 | "NFS: Server wrote zero bytes, expected %u.\n" , |
1625 | argp->count); |
1626 | complain = jiffies + 300 * HZ; |
1627 | } |
1628 | nfs_set_pgio_error(hdr, error: -EIO, pos: argp->offset); |
1629 | task->tk_status = -EIO; |
1630 | return; |
1631 | } |
1632 | |
1633 | /* For non rpc-based layout drivers, retry-through-MDS */ |
1634 | if (!task->tk_ops) { |
1635 | hdr->pnfs_error = -EAGAIN; |
1636 | return; |
1637 | } |
1638 | |
1639 | /* Was this an NFSv2 write or an NFSv3 stable write? */ |
1640 | if (resp->verf->committed != NFS_UNSTABLE) { |
1641 | /* Resend from where the server left off */ |
1642 | hdr->mds_offset += resp->count; |
1643 | argp->offset += resp->count; |
1644 | argp->pgbase += resp->count; |
1645 | argp->count -= resp->count; |
1646 | } else { |
1647 | /* Resend as a stable write in order to avoid |
1648 | * headaches in the case of a server crash. |
1649 | */ |
1650 | argp->stable = NFS_FILE_SYNC; |
1651 | } |
1652 | resp->count = 0; |
1653 | resp->verf->committed = 0; |
1654 | rpc_restart_call_prepare(task); |
1655 | } |
1656 | } |
1657 | |
1658 | static int wait_on_commit(struct nfs_mds_commit_info *cinfo) |
1659 | { |
1660 | return wait_var_event_killable(&cinfo->rpcs_out, |
1661 | !atomic_read(&cinfo->rpcs_out)); |
1662 | } |
1663 | |
1664 | static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo) |
1665 | { |
1666 | atomic_inc(v: &cinfo->rpcs_out); |
1667 | } |
1668 | |
1669 | bool nfs_commit_end(struct nfs_mds_commit_info *cinfo) |
1670 | { |
1671 | if (atomic_dec_and_test(v: &cinfo->rpcs_out)) { |
1672 | wake_up_var(var: &cinfo->rpcs_out); |
1673 | return true; |
1674 | } |
1675 | return false; |
1676 | } |
1677 | |
1678 | void nfs_commitdata_release(struct nfs_commit_data *data) |
1679 | { |
1680 | put_nfs_open_context(ctx: data->context); |
1681 | nfs_commit_free(data); |
1682 | } |
1683 | EXPORT_SYMBOL_GPL(nfs_commitdata_release); |
1684 | |
1685 | int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data, |
1686 | const struct nfs_rpc_ops *nfs_ops, |
1687 | const struct rpc_call_ops *call_ops, |
1688 | int how, int flags) |
1689 | { |
1690 | struct rpc_task *task; |
1691 | int priority = flush_task_priority(how); |
1692 | struct rpc_message msg = { |
1693 | .rpc_argp = &data->args, |
1694 | .rpc_resp = &data->res, |
1695 | .rpc_cred = data->cred, |
1696 | }; |
1697 | struct rpc_task_setup task_setup_data = { |
1698 | .task = &data->task, |
1699 | .rpc_client = clnt, |
1700 | .rpc_message = &msg, |
1701 | .callback_ops = call_ops, |
1702 | .callback_data = data, |
1703 | .workqueue = nfsiod_workqueue, |
1704 | .flags = RPC_TASK_ASYNC | flags, |
1705 | .priority = priority, |
1706 | }; |
1707 | |
1708 | if (nfs_server_capable(inode: data->inode, NFS_CAP_MOVEABLE)) |
1709 | task_setup_data.flags |= RPC_TASK_MOVEABLE; |
1710 | |
1711 | /* Set up the initial task struct. */ |
1712 | nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client); |
1713 | trace_nfs_initiate_commit(data); |
1714 | |
1715 | dprintk("NFS: initiated commit call\n" ); |
1716 | |
1717 | task = rpc_run_task(&task_setup_data); |
1718 | if (IS_ERR(ptr: task)) |
1719 | return PTR_ERR(ptr: task); |
1720 | if (how & FLUSH_SYNC) |
1721 | rpc_wait_for_completion_task(task); |
1722 | rpc_put_task(task); |
1723 | return 0; |
1724 | } |
1725 | EXPORT_SYMBOL_GPL(nfs_initiate_commit); |
1726 | |
1727 | static loff_t nfs_get_lwb(struct list_head *head) |
1728 | { |
1729 | loff_t lwb = 0; |
1730 | struct nfs_page *req; |
1731 | |
1732 | list_for_each_entry(req, head, wb_list) |
1733 | if (lwb < (req_offset(req) + req->wb_bytes)) |
1734 | lwb = req_offset(req) + req->wb_bytes; |
1735 | |
1736 | return lwb; |
1737 | } |
1738 | |
1739 | /* |
1740 | * Set up the argument/result storage required for the RPC call. |
1741 | */ |
1742 | void nfs_init_commit(struct nfs_commit_data *data, |
1743 | struct list_head *head, |
1744 | struct pnfs_layout_segment *lseg, |
1745 | struct nfs_commit_info *cinfo) |
1746 | { |
1747 | struct nfs_page *first; |
1748 | struct nfs_open_context *ctx; |
1749 | struct inode *inode; |
1750 | |
1751 | /* Set up the RPC argument and reply structs |
1752 | * NB: take care not to mess about with data->commit et al. */ |
1753 | |
1754 | if (head) |
1755 | list_splice_init(list: head, head: &data->pages); |
1756 | |
1757 | first = nfs_list_entry(head: data->pages.next); |
1758 | ctx = nfs_req_openctx(req: first); |
1759 | inode = d_inode(dentry: ctx->dentry); |
1760 | |
1761 | data->inode = inode; |
1762 | data->cred = ctx->cred; |
1763 | data->lseg = lseg; /* reference transferred */ |
1764 | /* only set lwb for pnfs commit */ |
1765 | if (lseg) |
1766 | data->lwb = nfs_get_lwb(head: &data->pages); |
1767 | data->mds_ops = &nfs_commit_ops; |
1768 | data->completion_ops = cinfo->completion_ops; |
1769 | data->dreq = cinfo->dreq; |
1770 | |
1771 | data->args.fh = NFS_FH(inode: data->inode); |
1772 | /* Note: we always request a commit of the entire inode */ |
1773 | data->args.offset = 0; |
1774 | data->args.count = 0; |
1775 | data->context = get_nfs_open_context(ctx); |
1776 | data->res.fattr = &data->fattr; |
1777 | data->res.verf = &data->verf; |
1778 | nfs_fattr_init(fattr: &data->fattr); |
1779 | nfs_commit_begin(cinfo: cinfo->mds); |
1780 | } |
1781 | EXPORT_SYMBOL_GPL(nfs_init_commit); |
1782 | |
1783 | void nfs_retry_commit(struct list_head *page_list, |
1784 | struct pnfs_layout_segment *lseg, |
1785 | struct nfs_commit_info *cinfo, |
1786 | u32 ds_commit_idx) |
1787 | { |
1788 | struct nfs_page *req; |
1789 | |
1790 | while (!list_empty(head: page_list)) { |
1791 | req = nfs_list_entry(head: page_list->next); |
1792 | nfs_list_remove_request(req); |
1793 | nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx); |
1794 | nfs_folio_clear_commit(folio: nfs_page_to_folio(req)); |
1795 | nfs_unlock_and_release_request(req); |
1796 | } |
1797 | } |
1798 | EXPORT_SYMBOL_GPL(nfs_retry_commit); |
1799 | |
1800 | static void nfs_commit_resched_write(struct nfs_commit_info *cinfo, |
1801 | struct nfs_page *req) |
1802 | { |
1803 | struct folio *folio = nfs_page_to_folio(req); |
1804 | |
1805 | filemap_dirty_folio(mapping: folio_mapping(folio), folio); |
1806 | } |
1807 | |
1808 | /* |
1809 | * Commit dirty pages |
1810 | */ |
1811 | static int |
1812 | nfs_commit_list(struct inode *inode, struct list_head *head, int how, |
1813 | struct nfs_commit_info *cinfo) |
1814 | { |
1815 | struct nfs_commit_data *data; |
1816 | unsigned short task_flags = 0; |
1817 | |
1818 | /* another commit raced with us */ |
1819 | if (list_empty(head)) |
1820 | return 0; |
1821 | |
1822 | data = nfs_commitdata_alloc(); |
1823 | if (!data) { |
1824 | nfs_retry_commit(head, NULL, cinfo, -1); |
1825 | return -ENOMEM; |
1826 | } |
1827 | |
1828 | /* Set up the argument struct */ |
1829 | nfs_init_commit(data, head, NULL, cinfo); |
1830 | if (NFS_SERVER(inode)->nfs_client->cl_minorversion) |
1831 | task_flags = RPC_TASK_MOVEABLE; |
1832 | return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode), |
1833 | data->mds_ops, how, |
1834 | RPC_TASK_CRED_NOREF | task_flags); |
1835 | } |
1836 | |
1837 | /* |
1838 | * COMMIT call returned |
1839 | */ |
1840 | static void nfs_commit_done(struct rpc_task *task, void *calldata) |
1841 | { |
1842 | struct nfs_commit_data *data = calldata; |
1843 | |
1844 | /* Call the NFS version-specific code */ |
1845 | NFS_PROTO(inode: data->inode)->commit_done(task, data); |
1846 | trace_nfs_commit_done(task, data); |
1847 | } |
1848 | |
1849 | static void nfs_commit_release_pages(struct nfs_commit_data *data) |
1850 | { |
1851 | const struct nfs_writeverf *verf = data->res.verf; |
1852 | struct nfs_page *req; |
1853 | int status = data->task.tk_status; |
1854 | struct nfs_commit_info cinfo; |
1855 | struct nfs_server *nfss; |
1856 | struct folio *folio; |
1857 | |
1858 | while (!list_empty(head: &data->pages)) { |
1859 | req = nfs_list_entry(head: data->pages.next); |
1860 | nfs_list_remove_request(req); |
1861 | folio = nfs_page_to_folio(req); |
1862 | nfs_folio_clear_commit(folio); |
1863 | |
1864 | dprintk("NFS: commit (%s/%llu %d@%lld)" , |
1865 | nfs_req_openctx(req)->dentry->d_sb->s_id, |
1866 | (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)), |
1867 | req->wb_bytes, |
1868 | (long long)req_offset(req)); |
1869 | if (status < 0) { |
1870 | if (folio) { |
1871 | trace_nfs_commit_error(inode: data->inode, req, |
1872 | error: status); |
1873 | nfs_mapping_set_error(folio, error: status); |
1874 | nfs_inode_remove_request(req); |
1875 | } |
1876 | dprintk_cont(", error = %d\n" , status); |
1877 | goto next; |
1878 | } |
1879 | |
1880 | /* Okay, COMMIT succeeded, apparently. Check the verifier |
1881 | * returned by the server against all stored verfs. */ |
1882 | if (nfs_write_match_verf(verf, req)) { |
1883 | /* We have a match */ |
1884 | if (folio) |
1885 | nfs_inode_remove_request(req); |
1886 | dprintk_cont(" OK\n" ); |
1887 | goto next; |
1888 | } |
1889 | /* We have a mismatch. Write the page again */ |
1890 | dprintk_cont(" mismatch\n" ); |
1891 | nfs_mark_request_dirty(req); |
1892 | atomic_long_inc(v: &NFS_I(inode: data->inode)->redirtied_pages); |
1893 | next: |
1894 | nfs_unlock_and_release_request(req); |
1895 | /* Latency breaker */ |
1896 | cond_resched(); |
1897 | } |
1898 | nfss = NFS_SERVER(inode: data->inode); |
1899 | if (atomic_long_read(v: &nfss->writeback) < NFS_CONGESTION_OFF_THRESH) |
1900 | nfss->write_congested = 0; |
1901 | |
1902 | nfs_init_cinfo(&cinfo, data->inode, data->dreq); |
1903 | nfs_commit_end(cinfo: cinfo.mds); |
1904 | } |
1905 | |
1906 | static void nfs_commit_release(void *calldata) |
1907 | { |
1908 | struct nfs_commit_data *data = calldata; |
1909 | |
1910 | data->completion_ops->completion(data); |
1911 | nfs_commitdata_release(calldata); |
1912 | } |
1913 | |
1914 | static const struct rpc_call_ops nfs_commit_ops = { |
1915 | .rpc_call_prepare = nfs_commit_prepare, |
1916 | .rpc_call_done = nfs_commit_done, |
1917 | .rpc_release = nfs_commit_release, |
1918 | }; |
1919 | |
1920 | static const struct nfs_commit_completion_ops nfs_commit_completion_ops = { |
1921 | .completion = nfs_commit_release_pages, |
1922 | .resched_write = nfs_commit_resched_write, |
1923 | }; |
1924 | |
1925 | int nfs_generic_commit_list(struct inode *inode, struct list_head *head, |
1926 | int how, struct nfs_commit_info *cinfo) |
1927 | { |
1928 | int status; |
1929 | |
1930 | status = pnfs_commit_list(inode, mds_pages: head, how, cinfo); |
1931 | if (status == PNFS_NOT_ATTEMPTED) |
1932 | status = nfs_commit_list(inode, head, how, cinfo); |
1933 | return status; |
1934 | } |
1935 | |
1936 | static int __nfs_commit_inode(struct inode *inode, int how, |
1937 | struct writeback_control *wbc) |
1938 | { |
1939 | LIST_HEAD(head); |
1940 | struct nfs_commit_info cinfo; |
1941 | int may_wait = how & FLUSH_SYNC; |
1942 | int ret, nscan; |
1943 | |
1944 | how &= ~FLUSH_SYNC; |
1945 | nfs_init_cinfo_from_inode(cinfo: &cinfo, inode); |
1946 | nfs_commit_begin(cinfo: cinfo.mds); |
1947 | for (;;) { |
1948 | ret = nscan = nfs_scan_commit(inode, dst: &head, cinfo: &cinfo); |
1949 | if (ret <= 0) |
1950 | break; |
1951 | ret = nfs_generic_commit_list(inode, head: &head, how, cinfo: &cinfo); |
1952 | if (ret < 0) |
1953 | break; |
1954 | ret = 0; |
1955 | if (wbc && wbc->sync_mode == WB_SYNC_NONE) { |
1956 | if (nscan < wbc->nr_to_write) |
1957 | wbc->nr_to_write -= nscan; |
1958 | else |
1959 | wbc->nr_to_write = 0; |
1960 | } |
1961 | if (nscan < INT_MAX) |
1962 | break; |
1963 | cond_resched(); |
1964 | } |
1965 | nfs_commit_end(cinfo: cinfo.mds); |
1966 | if (ret || !may_wait) |
1967 | return ret; |
1968 | return wait_on_commit(cinfo: cinfo.mds); |
1969 | } |
1970 | |
1971 | int nfs_commit_inode(struct inode *inode, int how) |
1972 | { |
1973 | return __nfs_commit_inode(inode, how, NULL); |
1974 | } |
1975 | EXPORT_SYMBOL_GPL(nfs_commit_inode); |
1976 | |
1977 | int nfs_write_inode(struct inode *inode, struct writeback_control *wbc) |
1978 | { |
1979 | struct nfs_inode *nfsi = NFS_I(inode); |
1980 | int flags = FLUSH_SYNC; |
1981 | int ret = 0; |
1982 | |
1983 | if (wbc->sync_mode == WB_SYNC_NONE) { |
1984 | /* no commits means nothing needs to be done */ |
1985 | if (!atomic_long_read(v: &nfsi->commit_info.ncommit)) |
1986 | goto check_requests_outstanding; |
1987 | |
1988 | /* Don't commit yet if this is a non-blocking flush and there |
1989 | * are a lot of outstanding writes for this mapping. |
1990 | */ |
1991 | if (mapping_tagged(mapping: inode->i_mapping, PAGECACHE_TAG_WRITEBACK)) |
1992 | goto out_mark_dirty; |
1993 | |
1994 | /* don't wait for the COMMIT response */ |
1995 | flags = 0; |
1996 | } |
1997 | |
1998 | ret = __nfs_commit_inode(inode, how: flags, wbc); |
1999 | if (!ret) { |
2000 | if (flags & FLUSH_SYNC) |
2001 | return 0; |
2002 | } else if (atomic_long_read(v: &nfsi->commit_info.ncommit)) |
2003 | goto out_mark_dirty; |
2004 | |
2005 | check_requests_outstanding: |
2006 | if (!atomic_read(v: &nfsi->commit_info.rpcs_out)) |
2007 | return ret; |
2008 | out_mark_dirty: |
2009 | __mark_inode_dirty(inode, I_DIRTY_DATASYNC); |
2010 | return ret; |
2011 | } |
2012 | EXPORT_SYMBOL_GPL(nfs_write_inode); |
2013 | |
2014 | /* |
2015 | * Wrapper for filemap_write_and_wait_range() |
2016 | * |
2017 | * Needed for pNFS in order to ensure data becomes visible to the |
2018 | * client. |
2019 | */ |
2020 | int nfs_filemap_write_and_wait_range(struct address_space *mapping, |
2021 | loff_t lstart, loff_t lend) |
2022 | { |
2023 | int ret; |
2024 | |
2025 | ret = filemap_write_and_wait_range(mapping, lstart, lend); |
2026 | if (ret == 0) |
2027 | ret = pnfs_sync_inode(inode: mapping->host, datasync: true); |
2028 | return ret; |
2029 | } |
2030 | EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range); |
2031 | |
2032 | /* |
2033 | * flush the inode to disk. |
2034 | */ |
2035 | int nfs_wb_all(struct inode *inode) |
2036 | { |
2037 | int ret; |
2038 | |
2039 | trace_nfs_writeback_inode_enter(inode); |
2040 | |
2041 | ret = filemap_write_and_wait(mapping: inode->i_mapping); |
2042 | if (ret) |
2043 | goto out; |
2044 | ret = nfs_commit_inode(inode, FLUSH_SYNC); |
2045 | if (ret < 0) |
2046 | goto out; |
2047 | pnfs_sync_inode(inode, datasync: true); |
2048 | ret = 0; |
2049 | |
2050 | out: |
2051 | trace_nfs_writeback_inode_exit(inode, error: ret); |
2052 | return ret; |
2053 | } |
2054 | EXPORT_SYMBOL_GPL(nfs_wb_all); |
2055 | |
2056 | int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio) |
2057 | { |
2058 | struct nfs_page *req; |
2059 | int ret = 0; |
2060 | |
2061 | folio_wait_writeback(folio); |
2062 | |
2063 | /* blocking call to cancel all requests and join to a single (head) |
2064 | * request */ |
2065 | req = nfs_lock_and_join_requests(folio); |
2066 | |
2067 | if (IS_ERR(ptr: req)) { |
2068 | ret = PTR_ERR(ptr: req); |
2069 | } else if (req) { |
2070 | /* all requests from this folio have been cancelled by |
2071 | * nfs_lock_and_join_requests, so just remove the head |
2072 | * request from the inode / page_private pointer and |
2073 | * release it */ |
2074 | nfs_inode_remove_request(req); |
2075 | nfs_unlock_and_release_request(req); |
2076 | } |
2077 | |
2078 | return ret; |
2079 | } |
2080 | |
2081 | /** |
2082 | * nfs_wb_folio - Write back all requests on one page |
2083 | * @inode: pointer to page |
2084 | * @folio: pointer to folio |
2085 | * |
2086 | * Assumes that the folio has been locked by the caller, and will |
2087 | * not unlock it. |
2088 | */ |
2089 | int nfs_wb_folio(struct inode *inode, struct folio *folio) |
2090 | { |
2091 | loff_t range_start = folio_file_pos(folio); |
2092 | loff_t range_end = range_start + (loff_t)folio_size(folio) - 1; |
2093 | struct writeback_control wbc = { |
2094 | .sync_mode = WB_SYNC_ALL, |
2095 | .nr_to_write = 0, |
2096 | .range_start = range_start, |
2097 | .range_end = range_end, |
2098 | }; |
2099 | int ret; |
2100 | |
2101 | trace_nfs_writeback_folio(inode, folio); |
2102 | |
2103 | for (;;) { |
2104 | folio_wait_writeback(folio); |
2105 | if (folio_clear_dirty_for_io(folio)) { |
2106 | ret = nfs_writepage_locked(folio, wbc: &wbc); |
2107 | if (ret < 0) |
2108 | goto out_error; |
2109 | continue; |
2110 | } |
2111 | ret = 0; |
2112 | if (!folio_test_private(folio)) |
2113 | break; |
2114 | ret = nfs_commit_inode(inode, FLUSH_SYNC); |
2115 | if (ret < 0) |
2116 | goto out_error; |
2117 | } |
2118 | out_error: |
2119 | trace_nfs_writeback_folio_done(inode, folio, ret); |
2120 | return ret; |
2121 | } |
2122 | |
2123 | #ifdef CONFIG_MIGRATION |
2124 | int nfs_migrate_folio(struct address_space *mapping, struct folio *dst, |
2125 | struct folio *src, enum migrate_mode mode) |
2126 | { |
2127 | /* |
2128 | * If the private flag is set, the folio is currently associated with |
2129 | * an in-progress read or write request. Don't try to migrate it. |
2130 | * |
2131 | * FIXME: we could do this in principle, but we'll need a way to ensure |
2132 | * that we can safely release the inode reference while holding |
2133 | * the folio lock. |
2134 | */ |
2135 | if (folio_test_private(folio: src)) |
2136 | return -EBUSY; |
2137 | |
2138 | if (folio_test_fscache(src)) { |
2139 | if (mode == MIGRATE_ASYNC) |
2140 | return -EBUSY; |
2141 | folio_wait_fscache(folio: src); |
2142 | } |
2143 | |
2144 | return migrate_folio(mapping, dst, src, mode); |
2145 | } |
2146 | #endif |
2147 | |
2148 | int __init nfs_init_writepagecache(void) |
2149 | { |
2150 | nfs_wdata_cachep = kmem_cache_create(name: "nfs_write_data" , |
2151 | size: sizeof(struct nfs_pgio_header), |
2152 | align: 0, SLAB_HWCACHE_ALIGN, |
2153 | NULL); |
2154 | if (nfs_wdata_cachep == NULL) |
2155 | return -ENOMEM; |
2156 | |
2157 | nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE, |
2158 | kc: nfs_wdata_cachep); |
2159 | if (nfs_wdata_mempool == NULL) |
2160 | goto out_destroy_write_cache; |
2161 | |
2162 | nfs_cdata_cachep = kmem_cache_create(name: "nfs_commit_data" , |
2163 | size: sizeof(struct nfs_commit_data), |
2164 | align: 0, SLAB_HWCACHE_ALIGN, |
2165 | NULL); |
2166 | if (nfs_cdata_cachep == NULL) |
2167 | goto out_destroy_write_mempool; |
2168 | |
2169 | nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT, |
2170 | kc: nfs_cdata_cachep); |
2171 | if (nfs_commit_mempool == NULL) |
2172 | goto out_destroy_commit_cache; |
2173 | |
2174 | /* |
2175 | * NFS congestion size, scale with available memory. |
2176 | * |
2177 | * 64MB: 8192k |
2178 | * 128MB: 11585k |
2179 | * 256MB: 16384k |
2180 | * 512MB: 23170k |
2181 | * 1GB: 32768k |
2182 | * 2GB: 46340k |
2183 | * 4GB: 65536k |
2184 | * 8GB: 92681k |
2185 | * 16GB: 131072k |
2186 | * |
2187 | * This allows larger machines to have larger/more transfers. |
2188 | * Limit the default to 256M |
2189 | */ |
2190 | nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10); |
2191 | if (nfs_congestion_kb > 256*1024) |
2192 | nfs_congestion_kb = 256*1024; |
2193 | |
2194 | return 0; |
2195 | |
2196 | out_destroy_commit_cache: |
2197 | kmem_cache_destroy(s: nfs_cdata_cachep); |
2198 | out_destroy_write_mempool: |
2199 | mempool_destroy(pool: nfs_wdata_mempool); |
2200 | out_destroy_write_cache: |
2201 | kmem_cache_destroy(s: nfs_wdata_cachep); |
2202 | return -ENOMEM; |
2203 | } |
2204 | |
2205 | void nfs_destroy_writepagecache(void) |
2206 | { |
2207 | mempool_destroy(pool: nfs_commit_mempool); |
2208 | kmem_cache_destroy(s: nfs_cdata_cachep); |
2209 | mempool_destroy(pool: nfs_wdata_mempool); |
2210 | kmem_cache_destroy(s: nfs_wdata_cachep); |
2211 | } |
2212 | |
2213 | static const struct nfs_rw_ops nfs_rw_write_ops = { |
2214 | .rw_alloc_header = nfs_writehdr_alloc, |
2215 | .rw_free_header = nfs_writehdr_free, |
2216 | .rw_done = nfs_writeback_done, |
2217 | .rw_result = nfs_writeback_result, |
2218 | .rw_initiate = nfs_initiate_write, |
2219 | }; |
2220 | |