]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - fs/nfs/write.c
nfs: rename pgio header ds_idx to ds_commit_idx
[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 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
794 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
795 BDI_RECLAIMABLE);
796 __mark_inode_dirty(req->wb_context->dentry->d_inode,
797 I_DIRTY_DATASYNC);
798 }
799 }
800 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
801
802 /**
803 * nfs_request_remove_commit_list - Remove request from a commit list
804 * @req: pointer to a nfs_page
805 * @cinfo: holds list lock and accounting info
806 *
807 * This clears the PG_CLEAN bit, and updates the cinfo's count of
808 * number of outstanding requests requiring a commit
809 * It does not update the MM page stats.
810 *
811 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
812 */
813 void
814 nfs_request_remove_commit_list(struct nfs_page *req,
815 struct nfs_commit_info *cinfo)
816 {
817 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
818 return;
819 nfs_list_remove_request(req);
820 cinfo->mds->ncommit--;
821 }
822 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
823
824 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
825 struct inode *inode)
826 {
827 cinfo->lock = &inode->i_lock;
828 cinfo->mds = &NFS_I(inode)->commit_info;
829 cinfo->ds = pnfs_get_ds_info(inode);
830 cinfo->dreq = NULL;
831 cinfo->completion_ops = &nfs_commit_completion_ops;
832 }
833
834 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
835 struct inode *inode,
836 struct nfs_direct_req *dreq)
837 {
838 if (dreq)
839 nfs_init_cinfo_from_dreq(cinfo, dreq);
840 else
841 nfs_init_cinfo_from_inode(cinfo, inode);
842 }
843 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
844
845 /*
846 * Add a request to the inode's commit list.
847 */
848 void
849 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
850 struct nfs_commit_info *cinfo)
851 {
852 if (pnfs_mark_request_commit(req, lseg, cinfo))
853 return;
854 nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
855 }
856
857 static void
858 nfs_clear_page_commit(struct page *page)
859 {
860 dec_zone_page_state(page, NR_UNSTABLE_NFS);
861 dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
862 }
863
864 /* Called holding inode (/cinfo) lock */
865 static void
866 nfs_clear_request_commit(struct nfs_page *req)
867 {
868 if (test_bit(PG_CLEAN, &req->wb_flags)) {
869 struct inode *inode = req->wb_context->dentry->d_inode;
870 struct nfs_commit_info cinfo;
871
872 nfs_init_cinfo_from_inode(&cinfo, inode);
873 if (!pnfs_clear_request_commit(req, &cinfo)) {
874 nfs_request_remove_commit_list(req, &cinfo);
875 }
876 nfs_clear_page_commit(req->wb_page);
877 }
878 }
879
880 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
881 {
882 if (hdr->verf.committed == NFS_DATA_SYNC)
883 return hdr->lseg == NULL;
884 return hdr->verf.committed != NFS_FILE_SYNC;
885 }
886
887 static void nfs_write_completion(struct nfs_pgio_header *hdr)
888 {
889 struct nfs_commit_info cinfo;
890 unsigned long bytes = 0;
891
892 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
893 goto out;
894 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
895 while (!list_empty(&hdr->pages)) {
896 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
897
898 bytes += req->wb_bytes;
899 nfs_list_remove_request(req);
900 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
901 (hdr->good_bytes < bytes)) {
902 nfs_set_pageerror(req->wb_page);
903 nfs_context_set_write_error(req->wb_context, hdr->error);
904 goto remove_req;
905 }
906 if (nfs_write_need_commit(hdr)) {
907 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
908 nfs_mark_request_commit(req, hdr->lseg, &cinfo);
909 goto next;
910 }
911 remove_req:
912 nfs_inode_remove_request(req);
913 next:
914 nfs_unlock_request(req);
915 nfs_end_page_writeback(req);
916 nfs_release_request(req);
917 }
918 out:
919 hdr->release(hdr);
920 }
921
922 unsigned long
923 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
924 {
925 return cinfo->mds->ncommit;
926 }
927
928 /* cinfo->lock held by caller */
929 int
930 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
931 struct nfs_commit_info *cinfo, int max)
932 {
933 struct nfs_page *req, *tmp;
934 int ret = 0;
935
936 list_for_each_entry_safe(req, tmp, src, wb_list) {
937 if (!nfs_lock_request(req))
938 continue;
939 kref_get(&req->wb_kref);
940 if (cond_resched_lock(cinfo->lock))
941 list_safe_reset_next(req, tmp, wb_list);
942 nfs_request_remove_commit_list(req, cinfo);
943 nfs_list_add_request(req, dst);
944 ret++;
945 if ((ret == max) && !cinfo->dreq)
946 break;
947 }
948 return ret;
949 }
950
951 /*
952 * nfs_scan_commit - Scan an inode for commit requests
953 * @inode: NFS inode to scan
954 * @dst: mds destination list
955 * @cinfo: mds and ds lists of reqs ready to commit
956 *
957 * Moves requests from the inode's 'commit' request list.
958 * The requests are *not* checked to ensure that they form a contiguous set.
959 */
960 int
961 nfs_scan_commit(struct inode *inode, struct list_head *dst,
962 struct nfs_commit_info *cinfo)
963 {
964 int ret = 0;
965
966 spin_lock(cinfo->lock);
967 if (cinfo->mds->ncommit > 0) {
968 const int max = INT_MAX;
969
970 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
971 cinfo, max);
972 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
973 }
974 spin_unlock(cinfo->lock);
975 return ret;
976 }
977
978 /*
979 * Search for an existing write request, and attempt to update
980 * it to reflect a new dirty region on a given page.
981 *
982 * If the attempt fails, then the existing request is flushed out
983 * to disk.
984 */
985 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
986 struct page *page,
987 unsigned int offset,
988 unsigned int bytes)
989 {
990 struct nfs_page *req;
991 unsigned int rqend;
992 unsigned int end;
993 int error;
994
995 if (!PagePrivate(page))
996 return NULL;
997
998 end = offset + bytes;
999 spin_lock(&inode->i_lock);
1000
1001 for (;;) {
1002 req = nfs_page_find_head_request_locked(NFS_I(inode), page);
1003 if (req == NULL)
1004 goto out_unlock;
1005
1006 /* should be handled by nfs_flush_incompatible */
1007 WARN_ON_ONCE(req->wb_head != req);
1008 WARN_ON_ONCE(req->wb_this_page != req);
1009
1010 rqend = req->wb_offset + req->wb_bytes;
1011 /*
1012 * Tell the caller to flush out the request if
1013 * the offsets are non-contiguous.
1014 * Note: nfs_flush_incompatible() will already
1015 * have flushed out requests having wrong owners.
1016 */
1017 if (offset > rqend
1018 || end < req->wb_offset)
1019 goto out_flushme;
1020
1021 if (nfs_lock_request(req))
1022 break;
1023
1024 /* The request is locked, so wait and then retry */
1025 spin_unlock(&inode->i_lock);
1026 error = nfs_wait_on_request(req);
1027 nfs_release_request(req);
1028 if (error != 0)
1029 goto out_err;
1030 spin_lock(&inode->i_lock);
1031 }
1032
1033 /* Okay, the request matches. Update the region */
1034 if (offset < req->wb_offset) {
1035 req->wb_offset = offset;
1036 req->wb_pgbase = offset;
1037 }
1038 if (end > rqend)
1039 req->wb_bytes = end - req->wb_offset;
1040 else
1041 req->wb_bytes = rqend - req->wb_offset;
1042 out_unlock:
1043 if (req)
1044 nfs_clear_request_commit(req);
1045 spin_unlock(&inode->i_lock);
1046 return req;
1047 out_flushme:
1048 spin_unlock(&inode->i_lock);
1049 nfs_release_request(req);
1050 error = nfs_wb_page(inode, page);
1051 out_err:
1052 return ERR_PTR(error);
1053 }
1054
1055 /*
1056 * Try to update an existing write request, or create one if there is none.
1057 *
1058 * Note: Should always be called with the Page Lock held to prevent races
1059 * if we have to add a new request. Also assumes that the caller has
1060 * already called nfs_flush_incompatible() if necessary.
1061 */
1062 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1063 struct page *page, unsigned int offset, unsigned int bytes)
1064 {
1065 struct inode *inode = page_file_mapping(page)->host;
1066 struct nfs_page *req;
1067
1068 req = nfs_try_to_update_request(inode, page, offset, bytes);
1069 if (req != NULL)
1070 goto out;
1071 req = nfs_create_request(ctx, page, NULL, offset, bytes);
1072 if (IS_ERR(req))
1073 goto out;
1074 nfs_inode_add_request(inode, req);
1075 out:
1076 return req;
1077 }
1078
1079 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1080 unsigned int offset, unsigned int count)
1081 {
1082 struct nfs_page *req;
1083
1084 req = nfs_setup_write_request(ctx, page, offset, count);
1085 if (IS_ERR(req))
1086 return PTR_ERR(req);
1087 /* Update file length */
1088 nfs_grow_file(page, offset, count);
1089 nfs_mark_uptodate(req);
1090 nfs_mark_request_dirty(req);
1091 nfs_unlock_and_release_request(req);
1092 return 0;
1093 }
1094
1095 int nfs_flush_incompatible(struct file *file, struct page *page)
1096 {
1097 struct nfs_open_context *ctx = nfs_file_open_context(file);
1098 struct nfs_lock_context *l_ctx;
1099 struct nfs_page *req;
1100 int do_flush, status;
1101 /*
1102 * Look for a request corresponding to this page. If there
1103 * is one, and it belongs to another file, we flush it out
1104 * before we try to copy anything into the page. Do this
1105 * due to the lack of an ACCESS-type call in NFSv2.
1106 * Also do the same if we find a request from an existing
1107 * dropped page.
1108 */
1109 do {
1110 req = nfs_page_find_head_request(page);
1111 if (req == NULL)
1112 return 0;
1113 l_ctx = req->wb_lock_context;
1114 do_flush = req->wb_page != page || req->wb_context != ctx;
1115 /* for now, flush if more than 1 request in page_group */
1116 do_flush |= req->wb_this_page != req;
1117 if (l_ctx && ctx->dentry->d_inode->i_flock != NULL) {
1118 do_flush |= l_ctx->lockowner.l_owner != current->files
1119 || l_ctx->lockowner.l_pid != current->tgid;
1120 }
1121 nfs_release_request(req);
1122 if (!do_flush)
1123 return 0;
1124 status = nfs_wb_page(page_file_mapping(page)->host, page);
1125 } while (status == 0);
1126 return status;
1127 }
1128
1129 /*
1130 * Avoid buffered writes when a open context credential's key would
1131 * expire soon.
1132 *
1133 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1134 *
1135 * Return 0 and set a credential flag which triggers the inode to flush
1136 * and performs NFS_FILE_SYNC writes if the key will expired within
1137 * RPC_KEY_EXPIRE_TIMEO.
1138 */
1139 int
1140 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1141 {
1142 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1143 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1144
1145 return rpcauth_key_timeout_notify(auth, ctx->cred);
1146 }
1147
1148 /*
1149 * Test if the open context credential key is marked to expire soon.
1150 */
1151 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
1152 {
1153 return rpcauth_cred_key_to_expire(ctx->cred);
1154 }
1155
1156 /*
1157 * If the page cache is marked as unsafe or invalid, then we can't rely on
1158 * the PageUptodate() flag. In this case, we will need to turn off
1159 * write optimisations that depend on the page contents being correct.
1160 */
1161 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1162 {
1163 struct nfs_inode *nfsi = NFS_I(inode);
1164
1165 if (nfs_have_delegated_attributes(inode))
1166 goto out;
1167 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1168 return false;
1169 smp_rmb();
1170 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1171 return false;
1172 out:
1173 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1174 return false;
1175 return PageUptodate(page) != 0;
1176 }
1177
1178 /* If we know the page is up to date, and we're not using byte range locks (or
1179 * if we have the whole file locked for writing), it may be more efficient to
1180 * extend the write to cover the entire page in order to avoid fragmentation
1181 * inefficiencies.
1182 *
1183 * If the file is opened for synchronous writes then we can just skip the rest
1184 * of the checks.
1185 */
1186 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1187 {
1188 if (file->f_flags & O_DSYNC)
1189 return 0;
1190 if (!nfs_write_pageuptodate(page, inode))
1191 return 0;
1192 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1193 return 1;
1194 if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 &&
1195 inode->i_flock->fl_end == OFFSET_MAX &&
1196 inode->i_flock->fl_type != F_RDLCK))
1197 return 1;
1198 return 0;
1199 }
1200
1201 /*
1202 * Update and possibly write a cached page of an NFS file.
1203 *
1204 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1205 * things with a page scheduled for an RPC call (e.g. invalidate it).
1206 */
1207 int nfs_updatepage(struct file *file, struct page *page,
1208 unsigned int offset, unsigned int count)
1209 {
1210 struct nfs_open_context *ctx = nfs_file_open_context(file);
1211 struct inode *inode = page_file_mapping(page)->host;
1212 int status = 0;
1213
1214 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1215
1216 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1217 file, count, (long long)(page_file_offset(page) + offset));
1218
1219 if (nfs_can_extend_write(file, page, inode)) {
1220 count = max(count + offset, nfs_page_length(page));
1221 offset = 0;
1222 }
1223
1224 status = nfs_writepage_setup(ctx, page, offset, count);
1225 if (status < 0)
1226 nfs_set_pageerror(page);
1227 else
1228 __set_page_dirty_nobuffers(page);
1229
1230 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1231 status, (long long)i_size_read(inode));
1232 return status;
1233 }
1234
1235 static int flush_task_priority(int how)
1236 {
1237 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1238 case FLUSH_HIGHPRI:
1239 return RPC_PRIORITY_HIGH;
1240 case FLUSH_LOWPRI:
1241 return RPC_PRIORITY_LOW;
1242 }
1243 return RPC_PRIORITY_NORMAL;
1244 }
1245
1246 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1247 struct rpc_message *msg,
1248 const struct nfs_rpc_ops *rpc_ops,
1249 struct rpc_task_setup *task_setup_data, int how)
1250 {
1251 int priority = flush_task_priority(how);
1252
1253 task_setup_data->priority = priority;
1254 rpc_ops->write_setup(hdr, msg);
1255
1256 nfs4_state_protect_write(NFS_SERVER(hdr->inode)->nfs_client,
1257 &task_setup_data->rpc_client, msg, hdr);
1258 }
1259
1260 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1261 * call this on each, which will prepare them to be retried on next
1262 * writeback using standard nfs.
1263 */
1264 static void nfs_redirty_request(struct nfs_page *req)
1265 {
1266 nfs_mark_request_dirty(req);
1267 nfs_unlock_request(req);
1268 nfs_end_page_writeback(req);
1269 nfs_release_request(req);
1270 }
1271
1272 static void nfs_async_write_error(struct list_head *head)
1273 {
1274 struct nfs_page *req;
1275
1276 while (!list_empty(head)) {
1277 req = nfs_list_entry(head->next);
1278 nfs_list_remove_request(req);
1279 nfs_redirty_request(req);
1280 }
1281 }
1282
1283 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1284 .error_cleanup = nfs_async_write_error,
1285 .completion = nfs_write_completion,
1286 };
1287
1288 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1289 struct inode *inode, int ioflags, bool force_mds,
1290 const struct nfs_pgio_completion_ops *compl_ops)
1291 {
1292 struct nfs_server *server = NFS_SERVER(inode);
1293 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1294
1295 #ifdef CONFIG_NFS_V4_1
1296 if (server->pnfs_curr_ld && !force_mds)
1297 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1298 #endif
1299 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1300 server->wsize, ioflags);
1301 }
1302 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1303
1304 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1305 {
1306 pgio->pg_ops = &nfs_pgio_rw_ops;
1307 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1308 }
1309 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1310
1311
1312 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1313 {
1314 struct nfs_commit_data *data = calldata;
1315
1316 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1317 }
1318
1319 static void nfs_writeback_release_common(struct nfs_pgio_header *hdr)
1320 {
1321 /* do nothing! */
1322 }
1323
1324 /*
1325 * Special version of should_remove_suid() that ignores capabilities.
1326 */
1327 static int nfs_should_remove_suid(const struct inode *inode)
1328 {
1329 umode_t mode = inode->i_mode;
1330 int kill = 0;
1331
1332 /* suid always must be killed */
1333 if (unlikely(mode & S_ISUID))
1334 kill = ATTR_KILL_SUID;
1335
1336 /*
1337 * sgid without any exec bits is just a mandatory locking mark; leave
1338 * it alone. If some exec bits are set, it's a real sgid; kill it.
1339 */
1340 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1341 kill |= ATTR_KILL_SGID;
1342
1343 if (unlikely(kill && S_ISREG(mode)))
1344 return kill;
1345
1346 return 0;
1347 }
1348
1349 /*
1350 * This function is called when the WRITE call is complete.
1351 */
1352 static int nfs_writeback_done(struct rpc_task *task,
1353 struct nfs_pgio_header *hdr,
1354 struct inode *inode)
1355 {
1356 int status;
1357
1358 /*
1359 * ->write_done will attempt to use post-op attributes to detect
1360 * conflicting writes by other clients. A strict interpretation
1361 * of close-to-open would allow us to continue caching even if
1362 * another writer had changed the file, but some applications
1363 * depend on tighter cache coherency when writing.
1364 */
1365 status = NFS_PROTO(inode)->write_done(task, hdr);
1366 if (status != 0)
1367 return status;
1368 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1369
1370 if (hdr->res.verf->committed < hdr->args.stable &&
1371 task->tk_status >= 0) {
1372 /* We tried a write call, but the server did not
1373 * commit data to stable storage even though we
1374 * requested it.
1375 * Note: There is a known bug in Tru64 < 5.0 in which
1376 * the server reports NFS_DATA_SYNC, but performs
1377 * NFS_FILE_SYNC. We therefore implement this checking
1378 * as a dprintk() in order to avoid filling syslog.
1379 */
1380 static unsigned long complain;
1381
1382 /* Note this will print the MDS for a DS write */
1383 if (time_before(complain, jiffies)) {
1384 dprintk("NFS: faulty NFS server %s:"
1385 " (committed = %d) != (stable = %d)\n",
1386 NFS_SERVER(inode)->nfs_client->cl_hostname,
1387 hdr->res.verf->committed, hdr->args.stable);
1388 complain = jiffies + 300 * HZ;
1389 }
1390 }
1391
1392 /* Deal with the suid/sgid bit corner case */
1393 if (nfs_should_remove_suid(inode))
1394 nfs_mark_for_revalidate(inode);
1395 return 0;
1396 }
1397
1398 /*
1399 * This function is called when the WRITE call is complete.
1400 */
1401 static void nfs_writeback_result(struct rpc_task *task,
1402 struct nfs_pgio_header *hdr)
1403 {
1404 struct nfs_pgio_args *argp = &hdr->args;
1405 struct nfs_pgio_res *resp = &hdr->res;
1406
1407 if (resp->count < argp->count) {
1408 static unsigned long complain;
1409
1410 /* This a short write! */
1411 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1412
1413 /* Has the server at least made some progress? */
1414 if (resp->count == 0) {
1415 if (time_before(complain, jiffies)) {
1416 printk(KERN_WARNING
1417 "NFS: Server wrote zero bytes, expected %u.\n",
1418 argp->count);
1419 complain = jiffies + 300 * HZ;
1420 }
1421 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1422 task->tk_status = -EIO;
1423 return;
1424 }
1425 /* Was this an NFSv2 write or an NFSv3 stable write? */
1426 if (resp->verf->committed != NFS_UNSTABLE) {
1427 /* Resend from where the server left off */
1428 hdr->mds_offset += resp->count;
1429 argp->offset += resp->count;
1430 argp->pgbase += resp->count;
1431 argp->count -= resp->count;
1432 } else {
1433 /* Resend as a stable write in order to avoid
1434 * headaches in the case of a server crash.
1435 */
1436 argp->stable = NFS_FILE_SYNC;
1437 }
1438 rpc_restart_call_prepare(task);
1439 }
1440 }
1441
1442
1443 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1444 {
1445 int ret;
1446
1447 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1448 return 1;
1449 if (!may_wait)
1450 return 0;
1451 ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1452 NFS_INO_COMMIT,
1453 nfs_wait_bit_killable,
1454 TASK_KILLABLE);
1455 return (ret < 0) ? ret : 1;
1456 }
1457
1458 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1459 {
1460 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1461 smp_mb__after_atomic();
1462 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1463 }
1464
1465 void nfs_commitdata_release(struct nfs_commit_data *data)
1466 {
1467 put_nfs_open_context(data->context);
1468 nfs_commit_free(data);
1469 }
1470 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1471
1472 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1473 const struct nfs_rpc_ops *nfs_ops,
1474 const struct rpc_call_ops *call_ops,
1475 int how, int flags)
1476 {
1477 struct rpc_task *task;
1478 int priority = flush_task_priority(how);
1479 struct rpc_message msg = {
1480 .rpc_argp = &data->args,
1481 .rpc_resp = &data->res,
1482 .rpc_cred = data->cred,
1483 };
1484 struct rpc_task_setup task_setup_data = {
1485 .task = &data->task,
1486 .rpc_client = clnt,
1487 .rpc_message = &msg,
1488 .callback_ops = call_ops,
1489 .callback_data = data,
1490 .workqueue = nfsiod_workqueue,
1491 .flags = RPC_TASK_ASYNC | flags,
1492 .priority = priority,
1493 };
1494 /* Set up the initial task struct. */
1495 nfs_ops->commit_setup(data, &msg);
1496
1497 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1498
1499 nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1500 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1501
1502 task = rpc_run_task(&task_setup_data);
1503 if (IS_ERR(task))
1504 return PTR_ERR(task);
1505 if (how & FLUSH_SYNC)
1506 rpc_wait_for_completion_task(task);
1507 rpc_put_task(task);
1508 return 0;
1509 }
1510 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1511
1512 static loff_t nfs_get_lwb(struct list_head *head)
1513 {
1514 loff_t lwb = 0;
1515 struct nfs_page *req;
1516
1517 list_for_each_entry(req, head, wb_list)
1518 if (lwb < (req_offset(req) + req->wb_bytes))
1519 lwb = req_offset(req) + req->wb_bytes;
1520
1521 return lwb;
1522 }
1523
1524 /*
1525 * Set up the argument/result storage required for the RPC call.
1526 */
1527 void nfs_init_commit(struct nfs_commit_data *data,
1528 struct list_head *head,
1529 struct pnfs_layout_segment *lseg,
1530 struct nfs_commit_info *cinfo)
1531 {
1532 struct nfs_page *first = nfs_list_entry(head->next);
1533 struct inode *inode = first->wb_context->dentry->d_inode;
1534
1535 /* Set up the RPC argument and reply structs
1536 * NB: take care not to mess about with data->commit et al. */
1537
1538 list_splice_init(head, &data->pages);
1539
1540 data->inode = inode;
1541 data->cred = first->wb_context->cred;
1542 data->lseg = lseg; /* reference transferred */
1543 /* only set lwb for pnfs commit */
1544 if (lseg)
1545 data->lwb = nfs_get_lwb(&data->pages);
1546 data->mds_ops = &nfs_commit_ops;
1547 data->completion_ops = cinfo->completion_ops;
1548 data->dreq = cinfo->dreq;
1549
1550 data->args.fh = NFS_FH(data->inode);
1551 /* Note: we always request a commit of the entire inode */
1552 data->args.offset = 0;
1553 data->args.count = 0;
1554 data->context = get_nfs_open_context(first->wb_context);
1555 data->res.fattr = &data->fattr;
1556 data->res.verf = &data->verf;
1557 nfs_fattr_init(&data->fattr);
1558 }
1559 EXPORT_SYMBOL_GPL(nfs_init_commit);
1560
1561 void nfs_retry_commit(struct list_head *page_list,
1562 struct pnfs_layout_segment *lseg,
1563 struct nfs_commit_info *cinfo)
1564 {
1565 struct nfs_page *req;
1566
1567 while (!list_empty(page_list)) {
1568 req = nfs_list_entry(page_list->next);
1569 nfs_list_remove_request(req);
1570 nfs_mark_request_commit(req, lseg, cinfo);
1571 if (!cinfo->dreq) {
1572 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1573 dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1574 BDI_RECLAIMABLE);
1575 }
1576 nfs_unlock_and_release_request(req);
1577 }
1578 }
1579 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1580
1581 /*
1582 * Commit dirty pages
1583 */
1584 static int
1585 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1586 struct nfs_commit_info *cinfo)
1587 {
1588 struct nfs_commit_data *data;
1589
1590 data = nfs_commitdata_alloc();
1591
1592 if (!data)
1593 goto out_bad;
1594
1595 /* Set up the argument struct */
1596 nfs_init_commit(data, head, NULL, cinfo);
1597 atomic_inc(&cinfo->mds->rpcs_out);
1598 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1599 data->mds_ops, how, 0);
1600 out_bad:
1601 nfs_retry_commit(head, NULL, cinfo);
1602 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1603 return -ENOMEM;
1604 }
1605
1606 /*
1607 * COMMIT call returned
1608 */
1609 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1610 {
1611 struct nfs_commit_data *data = calldata;
1612
1613 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1614 task->tk_pid, task->tk_status);
1615
1616 /* Call the NFS version-specific code */
1617 NFS_PROTO(data->inode)->commit_done(task, data);
1618 }
1619
1620 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1621 {
1622 struct nfs_page *req;
1623 int status = data->task.tk_status;
1624 struct nfs_commit_info cinfo;
1625 struct nfs_server *nfss;
1626
1627 while (!list_empty(&data->pages)) {
1628 req = nfs_list_entry(data->pages.next);
1629 nfs_list_remove_request(req);
1630 nfs_clear_page_commit(req->wb_page);
1631
1632 dprintk("NFS: commit (%s/%llu %d@%lld)",
1633 req->wb_context->dentry->d_sb->s_id,
1634 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1635 req->wb_bytes,
1636 (long long)req_offset(req));
1637 if (status < 0) {
1638 nfs_context_set_write_error(req->wb_context, status);
1639 nfs_inode_remove_request(req);
1640 dprintk(", error = %d\n", status);
1641 goto next;
1642 }
1643
1644 /* Okay, COMMIT succeeded, apparently. Check the verifier
1645 * returned by the server against all stored verfs. */
1646 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1647 /* We have a match */
1648 nfs_inode_remove_request(req);
1649 dprintk(" OK\n");
1650 goto next;
1651 }
1652 /* We have a mismatch. Write the page again */
1653 dprintk(" mismatch\n");
1654 nfs_mark_request_dirty(req);
1655 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1656 next:
1657 nfs_unlock_and_release_request(req);
1658 }
1659 nfss = NFS_SERVER(data->inode);
1660 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1661 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
1662
1663 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1664 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1665 nfs_commit_clear_lock(NFS_I(data->inode));
1666 }
1667
1668 static void nfs_commit_release(void *calldata)
1669 {
1670 struct nfs_commit_data *data = calldata;
1671
1672 data->completion_ops->completion(data);
1673 nfs_commitdata_release(calldata);
1674 }
1675
1676 static const struct rpc_call_ops nfs_commit_ops = {
1677 .rpc_call_prepare = nfs_commit_prepare,
1678 .rpc_call_done = nfs_commit_done,
1679 .rpc_release = nfs_commit_release,
1680 };
1681
1682 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1683 .completion = nfs_commit_release_pages,
1684 .error_cleanup = nfs_commit_clear_lock,
1685 };
1686
1687 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1688 int how, struct nfs_commit_info *cinfo)
1689 {
1690 int status;
1691
1692 status = pnfs_commit_list(inode, head, how, cinfo);
1693 if (status == PNFS_NOT_ATTEMPTED)
1694 status = nfs_commit_list(inode, head, how, cinfo);
1695 return status;
1696 }
1697
1698 int nfs_commit_inode(struct inode *inode, int how)
1699 {
1700 LIST_HEAD(head);
1701 struct nfs_commit_info cinfo;
1702 int may_wait = how & FLUSH_SYNC;
1703 int res;
1704
1705 res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1706 if (res <= 0)
1707 goto out_mark_dirty;
1708 nfs_init_cinfo_from_inode(&cinfo, inode);
1709 res = nfs_scan_commit(inode, &head, &cinfo);
1710 if (res) {
1711 int error;
1712
1713 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1714 if (error < 0)
1715 return error;
1716 if (!may_wait)
1717 goto out_mark_dirty;
1718 error = wait_on_bit_action(&NFS_I(inode)->flags,
1719 NFS_INO_COMMIT,
1720 nfs_wait_bit_killable,
1721 TASK_KILLABLE);
1722 if (error < 0)
1723 return error;
1724 } else
1725 nfs_commit_clear_lock(NFS_I(inode));
1726 return res;
1727 /* Note: If we exit without ensuring that the commit is complete,
1728 * we must mark the inode as dirty. Otherwise, future calls to
1729 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1730 * that the data is on the disk.
1731 */
1732 out_mark_dirty:
1733 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1734 return res;
1735 }
1736
1737 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1738 {
1739 struct nfs_inode *nfsi = NFS_I(inode);
1740 int flags = FLUSH_SYNC;
1741 int ret = 0;
1742
1743 /* no commits means nothing needs to be done */
1744 if (!nfsi->commit_info.ncommit)
1745 return ret;
1746
1747 if (wbc->sync_mode == WB_SYNC_NONE) {
1748 /* Don't commit yet if this is a non-blocking flush and there
1749 * are a lot of outstanding writes for this mapping.
1750 */
1751 if (nfsi->commit_info.ncommit <= (nfsi->nrequests >> 1))
1752 goto out_mark_dirty;
1753
1754 /* don't wait for the COMMIT response */
1755 flags = 0;
1756 }
1757
1758 ret = nfs_commit_inode(inode, flags);
1759 if (ret >= 0) {
1760 if (wbc->sync_mode == WB_SYNC_NONE) {
1761 if (ret < wbc->nr_to_write)
1762 wbc->nr_to_write -= ret;
1763 else
1764 wbc->nr_to_write = 0;
1765 }
1766 return 0;
1767 }
1768 out_mark_dirty:
1769 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1770 return ret;
1771 }
1772
1773 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1774 {
1775 return nfs_commit_unstable_pages(inode, wbc);
1776 }
1777 EXPORT_SYMBOL_GPL(nfs_write_inode);
1778
1779 /*
1780 * flush the inode to disk.
1781 */
1782 int nfs_wb_all(struct inode *inode)
1783 {
1784 struct writeback_control wbc = {
1785 .sync_mode = WB_SYNC_ALL,
1786 .nr_to_write = LONG_MAX,
1787 .range_start = 0,
1788 .range_end = LLONG_MAX,
1789 };
1790 int ret;
1791
1792 trace_nfs_writeback_inode_enter(inode);
1793
1794 ret = sync_inode(inode, &wbc);
1795
1796 trace_nfs_writeback_inode_exit(inode, ret);
1797 return ret;
1798 }
1799 EXPORT_SYMBOL_GPL(nfs_wb_all);
1800
1801 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1802 {
1803 struct nfs_page *req;
1804 int ret = 0;
1805
1806 wait_on_page_writeback(page);
1807
1808 /* blocking call to cancel all requests and join to a single (head)
1809 * request */
1810 req = nfs_lock_and_join_requests(page, false);
1811
1812 if (IS_ERR(req)) {
1813 ret = PTR_ERR(req);
1814 } else if (req) {
1815 /* all requests from this page have been cancelled by
1816 * nfs_lock_and_join_requests, so just remove the head
1817 * request from the inode / page_private pointer and
1818 * release it */
1819 nfs_inode_remove_request(req);
1820 /*
1821 * In case nfs_inode_remove_request has marked the
1822 * page as being dirty
1823 */
1824 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1825 nfs_unlock_and_release_request(req);
1826 }
1827
1828 return ret;
1829 }
1830
1831 /*
1832 * Write back all requests on one page - we do this before reading it.
1833 */
1834 int nfs_wb_page(struct inode *inode, struct page *page)
1835 {
1836 loff_t range_start = page_file_offset(page);
1837 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1838 struct writeback_control wbc = {
1839 .sync_mode = WB_SYNC_ALL,
1840 .nr_to_write = 0,
1841 .range_start = range_start,
1842 .range_end = range_end,
1843 };
1844 int ret;
1845
1846 trace_nfs_writeback_page_enter(inode);
1847
1848 for (;;) {
1849 wait_on_page_writeback(page);
1850 if (clear_page_dirty_for_io(page)) {
1851 ret = nfs_writepage_locked(page, &wbc);
1852 if (ret < 0)
1853 goto out_error;
1854 continue;
1855 }
1856 ret = 0;
1857 if (!PagePrivate(page))
1858 break;
1859 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1860 if (ret < 0)
1861 goto out_error;
1862 }
1863 out_error:
1864 trace_nfs_writeback_page_exit(inode, ret);
1865 return ret;
1866 }
1867
1868 #ifdef CONFIG_MIGRATION
1869 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1870 struct page *page, enum migrate_mode mode)
1871 {
1872 /*
1873 * If PagePrivate is set, then the page is currently associated with
1874 * an in-progress read or write request. Don't try to migrate it.
1875 *
1876 * FIXME: we could do this in principle, but we'll need a way to ensure
1877 * that we can safely release the inode reference while holding
1878 * the page lock.
1879 */
1880 if (PagePrivate(page))
1881 return -EBUSY;
1882
1883 if (!nfs_fscache_release_page(page, GFP_KERNEL))
1884 return -EBUSY;
1885
1886 return migrate_page(mapping, newpage, page, mode);
1887 }
1888 #endif
1889
1890 int __init nfs_init_writepagecache(void)
1891 {
1892 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1893 sizeof(struct nfs_pgio_header),
1894 0, SLAB_HWCACHE_ALIGN,
1895 NULL);
1896 if (nfs_wdata_cachep == NULL)
1897 return -ENOMEM;
1898
1899 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1900 nfs_wdata_cachep);
1901 if (nfs_wdata_mempool == NULL)
1902 goto out_destroy_write_cache;
1903
1904 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1905 sizeof(struct nfs_commit_data),
1906 0, SLAB_HWCACHE_ALIGN,
1907 NULL);
1908 if (nfs_cdata_cachep == NULL)
1909 goto out_destroy_write_mempool;
1910
1911 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1912 nfs_cdata_cachep);
1913 if (nfs_commit_mempool == NULL)
1914 goto out_destroy_commit_cache;
1915
1916 /*
1917 * NFS congestion size, scale with available memory.
1918 *
1919 * 64MB: 8192k
1920 * 128MB: 11585k
1921 * 256MB: 16384k
1922 * 512MB: 23170k
1923 * 1GB: 32768k
1924 * 2GB: 46340k
1925 * 4GB: 65536k
1926 * 8GB: 92681k
1927 * 16GB: 131072k
1928 *
1929 * This allows larger machines to have larger/more transfers.
1930 * Limit the default to 256M
1931 */
1932 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1933 if (nfs_congestion_kb > 256*1024)
1934 nfs_congestion_kb = 256*1024;
1935
1936 return 0;
1937
1938 out_destroy_commit_cache:
1939 kmem_cache_destroy(nfs_cdata_cachep);
1940 out_destroy_write_mempool:
1941 mempool_destroy(nfs_wdata_mempool);
1942 out_destroy_write_cache:
1943 kmem_cache_destroy(nfs_wdata_cachep);
1944 return -ENOMEM;
1945 }
1946
1947 void nfs_destroy_writepagecache(void)
1948 {
1949 mempool_destroy(nfs_commit_mempool);
1950 kmem_cache_destroy(nfs_cdata_cachep);
1951 mempool_destroy(nfs_wdata_mempool);
1952 kmem_cache_destroy(nfs_wdata_cachep);
1953 }
1954
1955 static const struct nfs_rw_ops nfs_rw_write_ops = {
1956 .rw_mode = FMODE_WRITE,
1957 .rw_alloc_header = nfs_writehdr_alloc,
1958 .rw_free_header = nfs_writehdr_free,
1959 .rw_release = nfs_writeback_release_common,
1960 .rw_done = nfs_writeback_done,
1961 .rw_result = nfs_writeback_result,
1962 .rw_initiate = nfs_initiate_write,
1963 };