]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - fs/nfs/write.c
NFS: Set an attribute barrier on all updates
[mirror_ubuntu-bionic-kernel.git] / fs / nfs / write.c
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_inc_stats(inode, NFSIOS_WRITEPAGES);
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 = req->wb_context->dentry->d_inode;
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_bdi_stat(inode_to_bdi(page_file_mapping(page)->host), BDI_RECLAIMABLE);
857 }
858
859 /* Called holding inode (/cinfo) lock */
860 static void
861 nfs_clear_request_commit(struct nfs_page *req)
862 {
863 if (test_bit(PG_CLEAN, &req->wb_flags)) {
864 struct inode *inode = req->wb_context->dentry->d_inode;
865 struct nfs_commit_info cinfo;
866
867 nfs_init_cinfo_from_inode(&cinfo, inode);
868 if (!pnfs_clear_request_commit(req, &cinfo)) {
869 nfs_request_remove_commit_list(req, &cinfo);
870 }
871 nfs_clear_page_commit(req->wb_page);
872 }
873 }
874
875 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
876 {
877 if (hdr->verf.committed == NFS_DATA_SYNC)
878 return hdr->lseg == NULL;
879 return hdr->verf.committed != NFS_FILE_SYNC;
880 }
881
882 static void nfs_write_completion(struct nfs_pgio_header *hdr)
883 {
884 struct nfs_commit_info cinfo;
885 unsigned long bytes = 0;
886
887 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
888 goto out;
889 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
890 while (!list_empty(&hdr->pages)) {
891 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
892
893 bytes += req->wb_bytes;
894 nfs_list_remove_request(req);
895 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
896 (hdr->good_bytes < bytes)) {
897 nfs_set_pageerror(req->wb_page);
898 nfs_context_set_write_error(req->wb_context, hdr->error);
899 goto remove_req;
900 }
901 if (nfs_write_need_commit(hdr)) {
902 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
903 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
904 hdr->pgio_mirror_idx);
905 goto next;
906 }
907 remove_req:
908 nfs_inode_remove_request(req);
909 next:
910 nfs_unlock_request(req);
911 nfs_end_page_writeback(req);
912 nfs_release_request(req);
913 }
914 out:
915 hdr->release(hdr);
916 }
917
918 unsigned long
919 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
920 {
921 return cinfo->mds->ncommit;
922 }
923
924 /* cinfo->lock held by caller */
925 int
926 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
927 struct nfs_commit_info *cinfo, int max)
928 {
929 struct nfs_page *req, *tmp;
930 int ret = 0;
931
932 list_for_each_entry_safe(req, tmp, src, wb_list) {
933 if (!nfs_lock_request(req))
934 continue;
935 kref_get(&req->wb_kref);
936 if (cond_resched_lock(cinfo->lock))
937 list_safe_reset_next(req, tmp, wb_list);
938 nfs_request_remove_commit_list(req, cinfo);
939 nfs_list_add_request(req, dst);
940 ret++;
941 if ((ret == max) && !cinfo->dreq)
942 break;
943 }
944 return ret;
945 }
946
947 /*
948 * nfs_scan_commit - Scan an inode for commit requests
949 * @inode: NFS inode to scan
950 * @dst: mds destination list
951 * @cinfo: mds and ds lists of reqs ready to commit
952 *
953 * Moves requests from the inode's 'commit' request list.
954 * The requests are *not* checked to ensure that they form a contiguous set.
955 */
956 int
957 nfs_scan_commit(struct inode *inode, struct list_head *dst,
958 struct nfs_commit_info *cinfo)
959 {
960 int ret = 0;
961
962 spin_lock(cinfo->lock);
963 if (cinfo->mds->ncommit > 0) {
964 const int max = INT_MAX;
965
966 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
967 cinfo, max);
968 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
969 }
970 spin_unlock(cinfo->lock);
971 return ret;
972 }
973
974 /*
975 * Search for an existing write request, and attempt to update
976 * it to reflect a new dirty region on a given page.
977 *
978 * If the attempt fails, then the existing request is flushed out
979 * to disk.
980 */
981 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
982 struct page *page,
983 unsigned int offset,
984 unsigned int bytes)
985 {
986 struct nfs_page *req;
987 unsigned int rqend;
988 unsigned int end;
989 int error;
990
991 if (!PagePrivate(page))
992 return NULL;
993
994 end = offset + bytes;
995 spin_lock(&inode->i_lock);
996
997 for (;;) {
998 req = nfs_page_find_head_request_locked(NFS_I(inode), page);
999 if (req == NULL)
1000 goto out_unlock;
1001
1002 /* should be handled by nfs_flush_incompatible */
1003 WARN_ON_ONCE(req->wb_head != req);
1004 WARN_ON_ONCE(req->wb_this_page != req);
1005
1006 rqend = req->wb_offset + req->wb_bytes;
1007 /*
1008 * Tell the caller to flush out the request if
1009 * the offsets are non-contiguous.
1010 * Note: nfs_flush_incompatible() will already
1011 * have flushed out requests having wrong owners.
1012 */
1013 if (offset > rqend
1014 || end < req->wb_offset)
1015 goto out_flushme;
1016
1017 if (nfs_lock_request(req))
1018 break;
1019
1020 /* The request is locked, so wait and then retry */
1021 spin_unlock(&inode->i_lock);
1022 error = nfs_wait_on_request(req);
1023 nfs_release_request(req);
1024 if (error != 0)
1025 goto out_err;
1026 spin_lock(&inode->i_lock);
1027 }
1028
1029 /* Okay, the request matches. Update the region */
1030 if (offset < req->wb_offset) {
1031 req->wb_offset = offset;
1032 req->wb_pgbase = offset;
1033 }
1034 if (end > rqend)
1035 req->wb_bytes = end - req->wb_offset;
1036 else
1037 req->wb_bytes = rqend - req->wb_offset;
1038 out_unlock:
1039 if (req)
1040 nfs_clear_request_commit(req);
1041 spin_unlock(&inode->i_lock);
1042 return req;
1043 out_flushme:
1044 spin_unlock(&inode->i_lock);
1045 nfs_release_request(req);
1046 error = nfs_wb_page(inode, page);
1047 out_err:
1048 return ERR_PTR(error);
1049 }
1050
1051 /*
1052 * Try to update an existing write request, or create one if there is none.
1053 *
1054 * Note: Should always be called with the Page Lock held to prevent races
1055 * if we have to add a new request. Also assumes that the caller has
1056 * already called nfs_flush_incompatible() if necessary.
1057 */
1058 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1059 struct page *page, unsigned int offset, unsigned int bytes)
1060 {
1061 struct inode *inode = page_file_mapping(page)->host;
1062 struct nfs_page *req;
1063
1064 req = nfs_try_to_update_request(inode, page, offset, bytes);
1065 if (req != NULL)
1066 goto out;
1067 req = nfs_create_request(ctx, page, NULL, offset, bytes);
1068 if (IS_ERR(req))
1069 goto out;
1070 nfs_inode_add_request(inode, req);
1071 out:
1072 return req;
1073 }
1074
1075 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1076 unsigned int offset, unsigned int count)
1077 {
1078 struct nfs_page *req;
1079
1080 req = nfs_setup_write_request(ctx, page, offset, count);
1081 if (IS_ERR(req))
1082 return PTR_ERR(req);
1083 /* Update file length */
1084 nfs_grow_file(page, offset, count);
1085 nfs_mark_uptodate(req);
1086 nfs_mark_request_dirty(req);
1087 nfs_unlock_and_release_request(req);
1088 return 0;
1089 }
1090
1091 int nfs_flush_incompatible(struct file *file, struct page *page)
1092 {
1093 struct nfs_open_context *ctx = nfs_file_open_context(file);
1094 struct nfs_lock_context *l_ctx;
1095 struct file_lock_context *flctx = file_inode(file)->i_flctx;
1096 struct nfs_page *req;
1097 int do_flush, status;
1098 /*
1099 * Look for a request corresponding to this page. If there
1100 * is one, and it belongs to another file, we flush it out
1101 * before we try to copy anything into the page. Do this
1102 * due to the lack of an ACCESS-type call in NFSv2.
1103 * Also do the same if we find a request from an existing
1104 * dropped page.
1105 */
1106 do {
1107 req = nfs_page_find_head_request(page);
1108 if (req == NULL)
1109 return 0;
1110 l_ctx = req->wb_lock_context;
1111 do_flush = req->wb_page != page || req->wb_context != ctx;
1112 /* for now, flush if more than 1 request in page_group */
1113 do_flush |= req->wb_this_page != req;
1114 if (l_ctx && flctx &&
1115 !(list_empty_careful(&flctx->flc_posix) &&
1116 list_empty_careful(&flctx->flc_flock))) {
1117 do_flush |= l_ctx->lockowner.l_owner != current->files
1118 || l_ctx->lockowner.l_pid != current->tgid;
1119 }
1120 nfs_release_request(req);
1121 if (!do_flush)
1122 return 0;
1123 status = nfs_wb_page(page_file_mapping(page)->host, page);
1124 } while (status == 0);
1125 return status;
1126 }
1127
1128 /*
1129 * Avoid buffered writes when a open context credential's key would
1130 * expire soon.
1131 *
1132 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1133 *
1134 * Return 0 and set a credential flag which triggers the inode to flush
1135 * and performs NFS_FILE_SYNC writes if the key will expired within
1136 * RPC_KEY_EXPIRE_TIMEO.
1137 */
1138 int
1139 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1140 {
1141 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1142 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1143
1144 return rpcauth_key_timeout_notify(auth, ctx->cred);
1145 }
1146
1147 /*
1148 * Test if the open context credential key is marked to expire soon.
1149 */
1150 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
1151 {
1152 return rpcauth_cred_key_to_expire(ctx->cred);
1153 }
1154
1155 /*
1156 * If the page cache is marked as unsafe or invalid, then we can't rely on
1157 * the PageUptodate() flag. In this case, we will need to turn off
1158 * write optimisations that depend on the page contents being correct.
1159 */
1160 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1161 {
1162 struct nfs_inode *nfsi = NFS_I(inode);
1163
1164 if (nfs_have_delegated_attributes(inode))
1165 goto out;
1166 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1167 return false;
1168 smp_rmb();
1169 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1170 return false;
1171 out:
1172 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1173 return false;
1174 return PageUptodate(page) != 0;
1175 }
1176
1177 static bool
1178 is_whole_file_wrlock(struct file_lock *fl)
1179 {
1180 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1181 fl->fl_type == F_WRLCK;
1182 }
1183
1184 /* If we know the page is up to date, and we're not using byte range locks (or
1185 * if we have the whole file locked for writing), it may be more efficient to
1186 * extend the write to cover the entire page in order to avoid fragmentation
1187 * inefficiencies.
1188 *
1189 * If the file is opened for synchronous writes then we can just skip the rest
1190 * of the checks.
1191 */
1192 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1193 {
1194 int ret;
1195 struct file_lock_context *flctx = inode->i_flctx;
1196 struct file_lock *fl;
1197
1198 if (file->f_flags & O_DSYNC)
1199 return 0;
1200 if (!nfs_write_pageuptodate(page, inode))
1201 return 0;
1202 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1203 return 1;
1204 if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1205 list_empty_careful(&flctx->flc_posix)))
1206 return 0;
1207
1208 /* Check to see if there are whole file write locks */
1209 ret = 0;
1210 spin_lock(&flctx->flc_lock);
1211 if (!list_empty(&flctx->flc_posix)) {
1212 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1213 fl_list);
1214 if (is_whole_file_wrlock(fl))
1215 ret = 1;
1216 } else if (!list_empty(&flctx->flc_flock)) {
1217 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1218 fl_list);
1219 if (fl->fl_type == F_WRLCK)
1220 ret = 1;
1221 }
1222 spin_unlock(&flctx->flc_lock);
1223 return ret;
1224 }
1225
1226 /*
1227 * Update and possibly write a cached page of an NFS file.
1228 *
1229 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1230 * things with a page scheduled for an RPC call (e.g. invalidate it).
1231 */
1232 int nfs_updatepage(struct file *file, struct page *page,
1233 unsigned int offset, unsigned int count)
1234 {
1235 struct nfs_open_context *ctx = nfs_file_open_context(file);
1236 struct inode *inode = page_file_mapping(page)->host;
1237 int status = 0;
1238
1239 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1240
1241 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1242 file, count, (long long)(page_file_offset(page) + offset));
1243
1244 if (nfs_can_extend_write(file, page, inode)) {
1245 count = max(count + offset, nfs_page_length(page));
1246 offset = 0;
1247 }
1248
1249 status = nfs_writepage_setup(ctx, page, offset, count);
1250 if (status < 0)
1251 nfs_set_pageerror(page);
1252 else
1253 __set_page_dirty_nobuffers(page);
1254
1255 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1256 status, (long long)i_size_read(inode));
1257 return status;
1258 }
1259
1260 static int flush_task_priority(int how)
1261 {
1262 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1263 case FLUSH_HIGHPRI:
1264 return RPC_PRIORITY_HIGH;
1265 case FLUSH_LOWPRI:
1266 return RPC_PRIORITY_LOW;
1267 }
1268 return RPC_PRIORITY_NORMAL;
1269 }
1270
1271 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1272 struct rpc_message *msg,
1273 const struct nfs_rpc_ops *rpc_ops,
1274 struct rpc_task_setup *task_setup_data, int how)
1275 {
1276 int priority = flush_task_priority(how);
1277
1278 task_setup_data->priority = priority;
1279 rpc_ops->write_setup(hdr, msg);
1280
1281 nfs4_state_protect_write(NFS_SERVER(hdr->inode)->nfs_client,
1282 &task_setup_data->rpc_client, msg, hdr);
1283 }
1284
1285 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1286 * call this on each, which will prepare them to be retried on next
1287 * writeback using standard nfs.
1288 */
1289 static void nfs_redirty_request(struct nfs_page *req)
1290 {
1291 nfs_mark_request_dirty(req);
1292 nfs_unlock_request(req);
1293 nfs_end_page_writeback(req);
1294 nfs_release_request(req);
1295 }
1296
1297 static void nfs_async_write_error(struct list_head *head)
1298 {
1299 struct nfs_page *req;
1300
1301 while (!list_empty(head)) {
1302 req = nfs_list_entry(head->next);
1303 nfs_list_remove_request(req);
1304 nfs_redirty_request(req);
1305 }
1306 }
1307
1308 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1309 .error_cleanup = nfs_async_write_error,
1310 .completion = nfs_write_completion,
1311 };
1312
1313 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1314 struct inode *inode, int ioflags, bool force_mds,
1315 const struct nfs_pgio_completion_ops *compl_ops)
1316 {
1317 struct nfs_server *server = NFS_SERVER(inode);
1318 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1319
1320 #ifdef CONFIG_NFS_V4_1
1321 if (server->pnfs_curr_ld && !force_mds)
1322 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1323 #endif
1324 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1325 server->wsize, ioflags);
1326 }
1327 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1328
1329 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1330 {
1331 struct nfs_pgio_mirror *mirror;
1332
1333 pgio->pg_ops = &nfs_pgio_rw_ops;
1334
1335 nfs_pageio_stop_mirroring(pgio);
1336
1337 mirror = &pgio->pg_mirrors[0];
1338 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1339 }
1340 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1341
1342
1343 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1344 {
1345 struct nfs_commit_data *data = calldata;
1346
1347 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1348 }
1349
1350 static void nfs_writeback_release_common(struct nfs_pgio_header *hdr)
1351 {
1352 /* do nothing! */
1353 }
1354
1355 /*
1356 * Special version of should_remove_suid() that ignores capabilities.
1357 */
1358 static int nfs_should_remove_suid(const struct inode *inode)
1359 {
1360 umode_t mode = inode->i_mode;
1361 int kill = 0;
1362
1363 /* suid always must be killed */
1364 if (unlikely(mode & S_ISUID))
1365 kill = ATTR_KILL_SUID;
1366
1367 /*
1368 * sgid without any exec bits is just a mandatory locking mark; leave
1369 * it alone. If some exec bits are set, it's a real sgid; kill it.
1370 */
1371 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1372 kill |= ATTR_KILL_SGID;
1373
1374 if (unlikely(kill && S_ISREG(mode)))
1375 return kill;
1376
1377 return 0;
1378 }
1379
1380 /*
1381 * This function is called when the WRITE call is complete.
1382 */
1383 static int nfs_writeback_done(struct rpc_task *task,
1384 struct nfs_pgio_header *hdr,
1385 struct inode *inode)
1386 {
1387 int status;
1388
1389 /*
1390 * ->write_done will attempt to use post-op attributes to detect
1391 * conflicting writes by other clients. A strict interpretation
1392 * of close-to-open would allow us to continue caching even if
1393 * another writer had changed the file, but some applications
1394 * depend on tighter cache coherency when writing.
1395 */
1396 status = NFS_PROTO(inode)->write_done(task, hdr);
1397 if (status != 0)
1398 return status;
1399 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1400
1401 if (hdr->res.verf->committed < hdr->args.stable &&
1402 task->tk_status >= 0) {
1403 /* We tried a write call, but the server did not
1404 * commit data to stable storage even though we
1405 * requested it.
1406 * Note: There is a known bug in Tru64 < 5.0 in which
1407 * the server reports NFS_DATA_SYNC, but performs
1408 * NFS_FILE_SYNC. We therefore implement this checking
1409 * as a dprintk() in order to avoid filling syslog.
1410 */
1411 static unsigned long complain;
1412
1413 /* Note this will print the MDS for a DS write */
1414 if (time_before(complain, jiffies)) {
1415 dprintk("NFS: faulty NFS server %s:"
1416 " (committed = %d) != (stable = %d)\n",
1417 NFS_SERVER(inode)->nfs_client->cl_hostname,
1418 hdr->res.verf->committed, hdr->args.stable);
1419 complain = jiffies + 300 * HZ;
1420 }
1421 }
1422
1423 /* Deal with the suid/sgid bit corner case */
1424 if (nfs_should_remove_suid(inode))
1425 nfs_mark_for_revalidate(inode);
1426 return 0;
1427 }
1428
1429 /*
1430 * This function is called when the WRITE call is complete.
1431 */
1432 static void nfs_writeback_result(struct rpc_task *task,
1433 struct nfs_pgio_header *hdr)
1434 {
1435 struct nfs_pgio_args *argp = &hdr->args;
1436 struct nfs_pgio_res *resp = &hdr->res;
1437
1438 if (resp->count < argp->count) {
1439 static unsigned long complain;
1440
1441 /* This a short write! */
1442 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1443
1444 /* Has the server at least made some progress? */
1445 if (resp->count == 0) {
1446 if (time_before(complain, jiffies)) {
1447 printk(KERN_WARNING
1448 "NFS: Server wrote zero bytes, expected %u.\n",
1449 argp->count);
1450 complain = jiffies + 300 * HZ;
1451 }
1452 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1453 task->tk_status = -EIO;
1454 return;
1455 }
1456 /* Was this an NFSv2 write or an NFSv3 stable write? */
1457 if (resp->verf->committed != NFS_UNSTABLE) {
1458 /* Resend from where the server left off */
1459 hdr->mds_offset += resp->count;
1460 argp->offset += resp->count;
1461 argp->pgbase += resp->count;
1462 argp->count -= resp->count;
1463 } else {
1464 /* Resend as a stable write in order to avoid
1465 * headaches in the case of a server crash.
1466 */
1467 argp->stable = NFS_FILE_SYNC;
1468 }
1469 rpc_restart_call_prepare(task);
1470 }
1471 }
1472
1473
1474 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1475 {
1476 int ret;
1477
1478 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1479 return 1;
1480 if (!may_wait)
1481 return 0;
1482 ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1483 NFS_INO_COMMIT,
1484 nfs_wait_bit_killable,
1485 TASK_KILLABLE);
1486 return (ret < 0) ? ret : 1;
1487 }
1488
1489 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1490 {
1491 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1492 smp_mb__after_atomic();
1493 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1494 }
1495
1496 void nfs_commitdata_release(struct nfs_commit_data *data)
1497 {
1498 put_nfs_open_context(data->context);
1499 nfs_commit_free(data);
1500 }
1501 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1502
1503 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1504 const struct nfs_rpc_ops *nfs_ops,
1505 const struct rpc_call_ops *call_ops,
1506 int how, int flags)
1507 {
1508 struct rpc_task *task;
1509 int priority = flush_task_priority(how);
1510 struct rpc_message msg = {
1511 .rpc_argp = &data->args,
1512 .rpc_resp = &data->res,
1513 .rpc_cred = data->cred,
1514 };
1515 struct rpc_task_setup task_setup_data = {
1516 .task = &data->task,
1517 .rpc_client = clnt,
1518 .rpc_message = &msg,
1519 .callback_ops = call_ops,
1520 .callback_data = data,
1521 .workqueue = nfsiod_workqueue,
1522 .flags = RPC_TASK_ASYNC | flags,
1523 .priority = priority,
1524 };
1525 /* Set up the initial task struct. */
1526 nfs_ops->commit_setup(data, &msg);
1527
1528 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1529
1530 nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1531 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1532
1533 task = rpc_run_task(&task_setup_data);
1534 if (IS_ERR(task))
1535 return PTR_ERR(task);
1536 if (how & FLUSH_SYNC)
1537 rpc_wait_for_completion_task(task);
1538 rpc_put_task(task);
1539 return 0;
1540 }
1541 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1542
1543 static loff_t nfs_get_lwb(struct list_head *head)
1544 {
1545 loff_t lwb = 0;
1546 struct nfs_page *req;
1547
1548 list_for_each_entry(req, head, wb_list)
1549 if (lwb < (req_offset(req) + req->wb_bytes))
1550 lwb = req_offset(req) + req->wb_bytes;
1551
1552 return lwb;
1553 }
1554
1555 /*
1556 * Set up the argument/result storage required for the RPC call.
1557 */
1558 void nfs_init_commit(struct nfs_commit_data *data,
1559 struct list_head *head,
1560 struct pnfs_layout_segment *lseg,
1561 struct nfs_commit_info *cinfo)
1562 {
1563 struct nfs_page *first = nfs_list_entry(head->next);
1564 struct inode *inode = first->wb_context->dentry->d_inode;
1565
1566 /* Set up the RPC argument and reply structs
1567 * NB: take care not to mess about with data->commit et al. */
1568
1569 list_splice_init(head, &data->pages);
1570
1571 data->inode = inode;
1572 data->cred = first->wb_context->cred;
1573 data->lseg = lseg; /* reference transferred */
1574 /* only set lwb for pnfs commit */
1575 if (lseg)
1576 data->lwb = nfs_get_lwb(&data->pages);
1577 data->mds_ops = &nfs_commit_ops;
1578 data->completion_ops = cinfo->completion_ops;
1579 data->dreq = cinfo->dreq;
1580
1581 data->args.fh = NFS_FH(data->inode);
1582 /* Note: we always request a commit of the entire inode */
1583 data->args.offset = 0;
1584 data->args.count = 0;
1585 data->context = get_nfs_open_context(first->wb_context);
1586 data->res.fattr = &data->fattr;
1587 data->res.verf = &data->verf;
1588 nfs_fattr_init(&data->fattr);
1589 }
1590 EXPORT_SYMBOL_GPL(nfs_init_commit);
1591
1592 void nfs_retry_commit(struct list_head *page_list,
1593 struct pnfs_layout_segment *lseg,
1594 struct nfs_commit_info *cinfo,
1595 u32 ds_commit_idx)
1596 {
1597 struct nfs_page *req;
1598
1599 while (!list_empty(page_list)) {
1600 req = nfs_list_entry(page_list->next);
1601 nfs_list_remove_request(req);
1602 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1603 if (!cinfo->dreq)
1604 nfs_clear_page_commit(req->wb_page);
1605 nfs_unlock_and_release_request(req);
1606 }
1607 }
1608 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1609
1610 /*
1611 * Commit dirty pages
1612 */
1613 static int
1614 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1615 struct nfs_commit_info *cinfo)
1616 {
1617 struct nfs_commit_data *data;
1618
1619 data = nfs_commitdata_alloc();
1620
1621 if (!data)
1622 goto out_bad;
1623
1624 /* Set up the argument struct */
1625 nfs_init_commit(data, head, NULL, cinfo);
1626 atomic_inc(&cinfo->mds->rpcs_out);
1627 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1628 data->mds_ops, how, 0);
1629 out_bad:
1630 nfs_retry_commit(head, NULL, cinfo, 0);
1631 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1632 return -ENOMEM;
1633 }
1634
1635 /*
1636 * COMMIT call returned
1637 */
1638 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1639 {
1640 struct nfs_commit_data *data = calldata;
1641
1642 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1643 task->tk_pid, task->tk_status);
1644
1645 /* Call the NFS version-specific code */
1646 NFS_PROTO(data->inode)->commit_done(task, data);
1647 }
1648
1649 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1650 {
1651 struct nfs_page *req;
1652 int status = data->task.tk_status;
1653 struct nfs_commit_info cinfo;
1654 struct nfs_server *nfss;
1655
1656 while (!list_empty(&data->pages)) {
1657 req = nfs_list_entry(data->pages.next);
1658 nfs_list_remove_request(req);
1659 nfs_clear_page_commit(req->wb_page);
1660
1661 dprintk("NFS: commit (%s/%llu %d@%lld)",
1662 req->wb_context->dentry->d_sb->s_id,
1663 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1664 req->wb_bytes,
1665 (long long)req_offset(req));
1666 if (status < 0) {
1667 nfs_context_set_write_error(req->wb_context, status);
1668 nfs_inode_remove_request(req);
1669 dprintk(", error = %d\n", status);
1670 goto next;
1671 }
1672
1673 /* Okay, COMMIT succeeded, apparently. Check the verifier
1674 * returned by the server against all stored verfs. */
1675 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1676 /* We have a match */
1677 nfs_inode_remove_request(req);
1678 dprintk(" OK\n");
1679 goto next;
1680 }
1681 /* We have a mismatch. Write the page again */
1682 dprintk(" mismatch\n");
1683 nfs_mark_request_dirty(req);
1684 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1685 next:
1686 nfs_unlock_and_release_request(req);
1687 }
1688 nfss = NFS_SERVER(data->inode);
1689 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1690 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
1691
1692 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1693 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1694 nfs_commit_clear_lock(NFS_I(data->inode));
1695 }
1696
1697 static void nfs_commit_release(void *calldata)
1698 {
1699 struct nfs_commit_data *data = calldata;
1700
1701 data->completion_ops->completion(data);
1702 nfs_commitdata_release(calldata);
1703 }
1704
1705 static const struct rpc_call_ops nfs_commit_ops = {
1706 .rpc_call_prepare = nfs_commit_prepare,
1707 .rpc_call_done = nfs_commit_done,
1708 .rpc_release = nfs_commit_release,
1709 };
1710
1711 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1712 .completion = nfs_commit_release_pages,
1713 .error_cleanup = nfs_commit_clear_lock,
1714 };
1715
1716 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1717 int how, struct nfs_commit_info *cinfo)
1718 {
1719 int status;
1720
1721 status = pnfs_commit_list(inode, head, how, cinfo);
1722 if (status == PNFS_NOT_ATTEMPTED)
1723 status = nfs_commit_list(inode, head, how, cinfo);
1724 return status;
1725 }
1726
1727 int nfs_commit_inode(struct inode *inode, int how)
1728 {
1729 LIST_HEAD(head);
1730 struct nfs_commit_info cinfo;
1731 int may_wait = how & FLUSH_SYNC;
1732 int res;
1733
1734 res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1735 if (res <= 0)
1736 goto out_mark_dirty;
1737 nfs_init_cinfo_from_inode(&cinfo, inode);
1738 res = nfs_scan_commit(inode, &head, &cinfo);
1739 if (res) {
1740 int error;
1741
1742 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1743 if (error < 0)
1744 return error;
1745 if (!may_wait)
1746 goto out_mark_dirty;
1747 error = wait_on_bit_action(&NFS_I(inode)->flags,
1748 NFS_INO_COMMIT,
1749 nfs_wait_bit_killable,
1750 TASK_KILLABLE);
1751 if (error < 0)
1752 return error;
1753 } else
1754 nfs_commit_clear_lock(NFS_I(inode));
1755 return res;
1756 /* Note: If we exit without ensuring that the commit is complete,
1757 * we must mark the inode as dirty. Otherwise, future calls to
1758 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1759 * that the data is on the disk.
1760 */
1761 out_mark_dirty:
1762 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1763 return res;
1764 }
1765
1766 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1767 {
1768 struct nfs_inode *nfsi = NFS_I(inode);
1769 int flags = FLUSH_SYNC;
1770 int ret = 0;
1771
1772 /* no commits means nothing needs to be done */
1773 if (!nfsi->commit_info.ncommit)
1774 return ret;
1775
1776 if (wbc->sync_mode == WB_SYNC_NONE) {
1777 /* Don't commit yet if this is a non-blocking flush and there
1778 * are a lot of outstanding writes for this mapping.
1779 */
1780 if (nfsi->commit_info.ncommit <= (nfsi->nrequests >> 1))
1781 goto out_mark_dirty;
1782
1783 /* don't wait for the COMMIT response */
1784 flags = 0;
1785 }
1786
1787 ret = nfs_commit_inode(inode, flags);
1788 if (ret >= 0) {
1789 if (wbc->sync_mode == WB_SYNC_NONE) {
1790 if (ret < wbc->nr_to_write)
1791 wbc->nr_to_write -= ret;
1792 else
1793 wbc->nr_to_write = 0;
1794 }
1795 return 0;
1796 }
1797 out_mark_dirty:
1798 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1799 return ret;
1800 }
1801
1802 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1803 {
1804 return nfs_commit_unstable_pages(inode, wbc);
1805 }
1806 EXPORT_SYMBOL_GPL(nfs_write_inode);
1807
1808 /*
1809 * flush the inode to disk.
1810 */
1811 int nfs_wb_all(struct inode *inode)
1812 {
1813 struct writeback_control wbc = {
1814 .sync_mode = WB_SYNC_ALL,
1815 .nr_to_write = LONG_MAX,
1816 .range_start = 0,
1817 .range_end = LLONG_MAX,
1818 };
1819 int ret;
1820
1821 trace_nfs_writeback_inode_enter(inode);
1822
1823 ret = sync_inode(inode, &wbc);
1824
1825 trace_nfs_writeback_inode_exit(inode, ret);
1826 return ret;
1827 }
1828 EXPORT_SYMBOL_GPL(nfs_wb_all);
1829
1830 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1831 {
1832 struct nfs_page *req;
1833 int ret = 0;
1834
1835 wait_on_page_writeback(page);
1836
1837 /* blocking call to cancel all requests and join to a single (head)
1838 * request */
1839 req = nfs_lock_and_join_requests(page, false);
1840
1841 if (IS_ERR(req)) {
1842 ret = PTR_ERR(req);
1843 } else if (req) {
1844 /* all requests from this page have been cancelled by
1845 * nfs_lock_and_join_requests, so just remove the head
1846 * request from the inode / page_private pointer and
1847 * release it */
1848 nfs_inode_remove_request(req);
1849 /*
1850 * In case nfs_inode_remove_request has marked the
1851 * page as being dirty
1852 */
1853 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1854 nfs_unlock_and_release_request(req);
1855 }
1856
1857 return ret;
1858 }
1859
1860 /*
1861 * Write back all requests on one page - we do this before reading it.
1862 */
1863 int nfs_wb_page(struct inode *inode, struct page *page)
1864 {
1865 loff_t range_start = page_file_offset(page);
1866 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1867 struct writeback_control wbc = {
1868 .sync_mode = WB_SYNC_ALL,
1869 .nr_to_write = 0,
1870 .range_start = range_start,
1871 .range_end = range_end,
1872 };
1873 int ret;
1874
1875 trace_nfs_writeback_page_enter(inode);
1876
1877 for (;;) {
1878 wait_on_page_writeback(page);
1879 if (clear_page_dirty_for_io(page)) {
1880 ret = nfs_writepage_locked(page, &wbc);
1881 if (ret < 0)
1882 goto out_error;
1883 continue;
1884 }
1885 ret = 0;
1886 if (!PagePrivate(page))
1887 break;
1888 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1889 if (ret < 0)
1890 goto out_error;
1891 }
1892 out_error:
1893 trace_nfs_writeback_page_exit(inode, ret);
1894 return ret;
1895 }
1896
1897 #ifdef CONFIG_MIGRATION
1898 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1899 struct page *page, enum migrate_mode mode)
1900 {
1901 /*
1902 * If PagePrivate is set, then the page is currently associated with
1903 * an in-progress read or write request. Don't try to migrate it.
1904 *
1905 * FIXME: we could do this in principle, but we'll need a way to ensure
1906 * that we can safely release the inode reference while holding
1907 * the page lock.
1908 */
1909 if (PagePrivate(page))
1910 return -EBUSY;
1911
1912 if (!nfs_fscache_release_page(page, GFP_KERNEL))
1913 return -EBUSY;
1914
1915 return migrate_page(mapping, newpage, page, mode);
1916 }
1917 #endif
1918
1919 int __init nfs_init_writepagecache(void)
1920 {
1921 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1922 sizeof(struct nfs_pgio_header),
1923 0, SLAB_HWCACHE_ALIGN,
1924 NULL);
1925 if (nfs_wdata_cachep == NULL)
1926 return -ENOMEM;
1927
1928 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1929 nfs_wdata_cachep);
1930 if (nfs_wdata_mempool == NULL)
1931 goto out_destroy_write_cache;
1932
1933 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1934 sizeof(struct nfs_commit_data),
1935 0, SLAB_HWCACHE_ALIGN,
1936 NULL);
1937 if (nfs_cdata_cachep == NULL)
1938 goto out_destroy_write_mempool;
1939
1940 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1941 nfs_cdata_cachep);
1942 if (nfs_commit_mempool == NULL)
1943 goto out_destroy_commit_cache;
1944
1945 /*
1946 * NFS congestion size, scale with available memory.
1947 *
1948 * 64MB: 8192k
1949 * 128MB: 11585k
1950 * 256MB: 16384k
1951 * 512MB: 23170k
1952 * 1GB: 32768k
1953 * 2GB: 46340k
1954 * 4GB: 65536k
1955 * 8GB: 92681k
1956 * 16GB: 131072k
1957 *
1958 * This allows larger machines to have larger/more transfers.
1959 * Limit the default to 256M
1960 */
1961 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1962 if (nfs_congestion_kb > 256*1024)
1963 nfs_congestion_kb = 256*1024;
1964
1965 return 0;
1966
1967 out_destroy_commit_cache:
1968 kmem_cache_destroy(nfs_cdata_cachep);
1969 out_destroy_write_mempool:
1970 mempool_destroy(nfs_wdata_mempool);
1971 out_destroy_write_cache:
1972 kmem_cache_destroy(nfs_wdata_cachep);
1973 return -ENOMEM;
1974 }
1975
1976 void nfs_destroy_writepagecache(void)
1977 {
1978 mempool_destroy(nfs_commit_mempool);
1979 kmem_cache_destroy(nfs_cdata_cachep);
1980 mempool_destroy(nfs_wdata_mempool);
1981 kmem_cache_destroy(nfs_wdata_cachep);
1982 }
1983
1984 static const struct nfs_rw_ops nfs_rw_write_ops = {
1985 .rw_mode = FMODE_WRITE,
1986 .rw_alloc_header = nfs_writehdr_alloc,
1987 .rw_free_header = nfs_writehdr_free,
1988 .rw_release = nfs_writeback_release_common,
1989 .rw_done = nfs_writeback_done,
1990 .rw_result = nfs_writeback_result,
1991 .rw_initiate = nfs_initiate_write,
1992 };