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. If not, it's a programming error.
477 */
478 WARN_ON_ONCE(subreq->wb_offset !=
479 (head->wb_offset + total_bytes));
480
481 /* keep track of how many bytes this group covers */
482 total_bytes += subreq->wb_bytes;
483
484 if (!nfs_lock_request(subreq)) {
485 /* releases page group bit lock and
486 * inode spin lock and all references */
487 ret = nfs_unroll_locks_and_wait(inode, head,
488 subreq, nonblock);
489
490 if (ret == 0)
491 goto try_again;
492
493 return ERR_PTR(ret);
494 }
495
496 subreq = subreq->wb_this_page;
497 } while (subreq != head);
498
499 /* Now that all requests are locked, make sure they aren't on any list.
500 * Commit list removal accounting is done after locks are dropped */
501 subreq = head;
502 do {
503 nfs_clear_request_commit(subreq);
504 subreq = subreq->wb_this_page;
505 } while (subreq != head);
506
507 /* unlink subrequests from head, destroy them later */
508 if (head->wb_this_page != head) {
509 /* destroy list will be terminated by head */
510 destroy_list = head->wb_this_page;
511 head->wb_this_page = head;
512
513 /* change head request to cover whole range that
514 * the former page group covered */
515 head->wb_bytes = total_bytes;
516 }
517
518 /*
519 * prepare head request to be added to new pgio descriptor
520 */
521 nfs_page_group_clear_bits(head);
522
523 /*
524 * some part of the group was still on the inode list - otherwise
525 * the group wouldn't be involved in async write.
526 * grab a reference for the head request, iff it needs one.
527 */
528 if (!test_and_set_bit(PG_INODE_REF, &head->wb_flags))
529 kref_get(&head->wb_kref);
530
531 nfs_page_group_unlock(head);
532
533 /* drop lock to clean uprequests on destroy list */
534 spin_unlock(&inode->i_lock);
535
536 nfs_destroy_unlinked_subrequests(destroy_list, head);
537
538 /* still holds ref on head from nfs_page_find_head_request_locked
539 * and still has lock on head from lock loop */
540 return head;
541 }
542
543 /*
544 * Find an associated nfs write request, and prepare to flush it out
545 * May return an error if the user signalled nfs_wait_on_request().
546 */
547 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
548 struct page *page, bool nonblock)
549 {
550 struct nfs_page *req;
551 int ret = 0;
552
553 req = nfs_lock_and_join_requests(page, nonblock);
554 if (!req)
555 goto out;
556 ret = PTR_ERR(req);
557 if (IS_ERR(req))
558 goto out;
559
560 nfs_set_page_writeback(page);
561 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
562
563 ret = 0;
564 if (!nfs_pageio_add_request(pgio, req)) {
565 nfs_redirty_request(req);
566 ret = pgio->pg_error;
567 }
568 out:
569 return ret;
570 }
571
572 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
573 {
574 struct inode *inode = page_file_mapping(page)->host;
575 int ret;
576
577 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
578 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
579
580 nfs_pageio_cond_complete(pgio, page_file_index(page));
581 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
582 if (ret == -EAGAIN) {
583 redirty_page_for_writepage(wbc, page);
584 ret = 0;
585 }
586 return ret;
587 }
588
589 /*
590 * Write an mmapped page to the server.
591 */
592 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
593 {
594 struct nfs_pageio_descriptor pgio;
595 int err;
596
597 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
598 false, &nfs_async_write_completion_ops);
599 err = nfs_do_writepage(page, wbc, &pgio);
600 nfs_pageio_complete(&pgio);
601 if (err < 0)
602 return err;
603 if (pgio.pg_error < 0)
604 return pgio.pg_error;
605 return 0;
606 }
607
608 int nfs_writepage(struct page *page, struct writeback_control *wbc)
609 {
610 int ret;
611
612 ret = nfs_writepage_locked(page, wbc);
613 unlock_page(page);
614 return ret;
615 }
616
617 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
618 {
619 int ret;
620
621 ret = nfs_do_writepage(page, wbc, data);
622 unlock_page(page);
623 return ret;
624 }
625
626 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
627 {
628 struct inode *inode = mapping->host;
629 unsigned long *bitlock = &NFS_I(inode)->flags;
630 struct nfs_pageio_descriptor pgio;
631 int err;
632
633 /* Stop dirtying of new pages while we sync */
634 err = wait_on_bit_lock_action(bitlock, NFS_INO_FLUSHING,
635 nfs_wait_bit_killable, TASK_KILLABLE);
636 if (err)
637 goto out_err;
638
639 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
640
641 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
642 &nfs_async_write_completion_ops);
643 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
644 nfs_pageio_complete(&pgio);
645
646 clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
647 smp_mb__after_atomic();
648 wake_up_bit(bitlock, NFS_INO_FLUSHING);
649
650 if (err < 0)
651 goto out_err;
652 err = pgio.pg_error;
653 if (err < 0)
654 goto out_err;
655 return 0;
656 out_err:
657 return err;
658 }
659
660 /*
661 * Insert a write request into an inode
662 */
663 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
664 {
665 struct nfs_inode *nfsi = NFS_I(inode);
666
667 WARN_ON_ONCE(req->wb_this_page != req);
668
669 /* Lock the request! */
670 nfs_lock_request(req);
671
672 spin_lock(&inode->i_lock);
673 if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
674 inode->i_version++;
675 /*
676 * Swap-space should not get truncated. Hence no need to plug the race
677 * with invalidate/truncate.
678 */
679 if (likely(!PageSwapCache(req->wb_page))) {
680 set_bit(PG_MAPPED, &req->wb_flags);
681 SetPagePrivate(req->wb_page);
682 set_page_private(req->wb_page, (unsigned long)req);
683 }
684 nfsi->npages++;
685 /* this a head request for a page group - mark it as having an
686 * extra reference so sub groups can follow suit */
687 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
688 kref_get(&req->wb_kref);
689 spin_unlock(&inode->i_lock);
690 }
691
692 /*
693 * Remove a write request from an inode
694 */
695 static void nfs_inode_remove_request(struct nfs_page *req)
696 {
697 struct inode *inode = req->wb_context->dentry->d_inode;
698 struct nfs_inode *nfsi = NFS_I(inode);
699 struct nfs_page *head;
700
701 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
702 head = req->wb_head;
703
704 spin_lock(&inode->i_lock);
705 if (likely(!PageSwapCache(head->wb_page))) {
706 set_page_private(head->wb_page, 0);
707 ClearPagePrivate(head->wb_page);
708 smp_mb__after_atomic();
709 wake_up_page(head->wb_page, PG_private);
710 clear_bit(PG_MAPPED, &head->wb_flags);
711 }
712 nfsi->npages--;
713 spin_unlock(&inode->i_lock);
714 }
715
716 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
717 nfs_release_request(req);
718 }
719
720 static void
721 nfs_mark_request_dirty(struct nfs_page *req)
722 {
723 __set_page_dirty_nobuffers(req->wb_page);
724 }
725
726 /*
727 * nfs_page_search_commits_for_head_request_locked
728 *
729 * Search through commit lists on @inode for the head request for @page.
730 * Must be called while holding the inode (which is cinfo) lock.
731 *
732 * Returns the head request if found, or NULL if not found.
733 */
734 static struct nfs_page *
735 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
736 struct page *page)
737 {
738 struct nfs_page *freq, *t;
739 struct nfs_commit_info cinfo;
740 struct inode *inode = &nfsi->vfs_inode;
741
742 nfs_init_cinfo_from_inode(&cinfo, inode);
743
744 /* search through pnfs commit lists */
745 freq = pnfs_search_commit_reqs(inode, &cinfo, page);
746 if (freq)
747 return freq->wb_head;
748
749 /* Linearly search the commit list for the correct request */
750 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
751 if (freq->wb_page == page)
752 return freq->wb_head;
753 }
754
755 return NULL;
756 }
757
758 /**
759 * nfs_request_add_commit_list - add request to a commit list
760 * @req: pointer to a struct nfs_page
761 * @dst: commit list head
762 * @cinfo: holds list lock and accounting info
763 *
764 * This sets the PG_CLEAN bit, updates the cinfo count of
765 * number of outstanding requests requiring a commit as well as
766 * the MM page stats.
767 *
768 * The caller must _not_ hold the cinfo->lock, but must be
769 * holding the nfs_page lock.
770 */
771 void
772 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
773 struct nfs_commit_info *cinfo)
774 {
775 set_bit(PG_CLEAN, &(req)->wb_flags);
776 spin_lock(cinfo->lock);
777 nfs_list_add_request(req, dst);
778 cinfo->mds->ncommit++;
779 spin_unlock(cinfo->lock);
780 if (!cinfo->dreq) {
781 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
782 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
783 BDI_RECLAIMABLE);
784 __mark_inode_dirty(req->wb_context->dentry->d_inode,
785 I_DIRTY_DATASYNC);
786 }
787 }
788 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
789
790 /**
791 * nfs_request_remove_commit_list - Remove request from a commit list
792 * @req: pointer to a nfs_page
793 * @cinfo: holds list lock and accounting info
794 *
795 * This clears the PG_CLEAN bit, and updates the cinfo's count of
796 * number of outstanding requests requiring a commit
797 * It does not update the MM page stats.
798 *
799 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
800 */
801 void
802 nfs_request_remove_commit_list(struct nfs_page *req,
803 struct nfs_commit_info *cinfo)
804 {
805 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
806 return;
807 nfs_list_remove_request(req);
808 cinfo->mds->ncommit--;
809 }
810 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
811
812 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
813 struct inode *inode)
814 {
815 cinfo->lock = &inode->i_lock;
816 cinfo->mds = &NFS_I(inode)->commit_info;
817 cinfo->ds = pnfs_get_ds_info(inode);
818 cinfo->dreq = NULL;
819 cinfo->completion_ops = &nfs_commit_completion_ops;
820 }
821
822 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
823 struct inode *inode,
824 struct nfs_direct_req *dreq)
825 {
826 if (dreq)
827 nfs_init_cinfo_from_dreq(cinfo, dreq);
828 else
829 nfs_init_cinfo_from_inode(cinfo, inode);
830 }
831 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
832
833 /*
834 * Add a request to the inode's commit list.
835 */
836 void
837 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
838 struct nfs_commit_info *cinfo)
839 {
840 if (pnfs_mark_request_commit(req, lseg, cinfo))
841 return;
842 nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
843 }
844
845 static void
846 nfs_clear_page_commit(struct page *page)
847 {
848 dec_zone_page_state(page, NR_UNSTABLE_NFS);
849 dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
850 }
851
852 /* Called holding inode (/cinfo) lock */
853 static void
854 nfs_clear_request_commit(struct nfs_page *req)
855 {
856 if (test_bit(PG_CLEAN, &req->wb_flags)) {
857 struct inode *inode = req->wb_context->dentry->d_inode;
858 struct nfs_commit_info cinfo;
859
860 nfs_init_cinfo_from_inode(&cinfo, inode);
861 if (!pnfs_clear_request_commit(req, &cinfo)) {
862 nfs_request_remove_commit_list(req, &cinfo);
863 }
864 nfs_clear_page_commit(req->wb_page);
865 }
866 }
867
868 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
869 {
870 if (hdr->verf.committed == NFS_DATA_SYNC)
871 return hdr->lseg == NULL;
872 return hdr->verf.committed != NFS_FILE_SYNC;
873 }
874
875 static void nfs_write_completion(struct nfs_pgio_header *hdr)
876 {
877 struct nfs_commit_info cinfo;
878 unsigned long bytes = 0;
879
880 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
881 goto out;
882 nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
883 while (!list_empty(&hdr->pages)) {
884 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
885
886 bytes += req->wb_bytes;
887 nfs_list_remove_request(req);
888 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
889 (hdr->good_bytes < bytes)) {
890 nfs_set_pageerror(req->wb_page);
891 nfs_context_set_write_error(req->wb_context, hdr->error);
892 goto remove_req;
893 }
894 if (nfs_write_need_commit(hdr)) {
895 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
896 nfs_mark_request_commit(req, hdr->lseg, &cinfo);
897 goto next;
898 }
899 remove_req:
900 nfs_inode_remove_request(req);
901 next:
902 nfs_unlock_request(req);
903 nfs_end_page_writeback(req);
904 nfs_release_request(req);
905 }
906 out:
907 hdr->release(hdr);
908 }
909
910 unsigned long
911 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
912 {
913 return cinfo->mds->ncommit;
914 }
915
916 /* cinfo->lock held by caller */
917 int
918 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
919 struct nfs_commit_info *cinfo, int max)
920 {
921 struct nfs_page *req, *tmp;
922 int ret = 0;
923
924 list_for_each_entry_safe(req, tmp, src, wb_list) {
925 if (!nfs_lock_request(req))
926 continue;
927 kref_get(&req->wb_kref);
928 if (cond_resched_lock(cinfo->lock))
929 list_safe_reset_next(req, tmp, wb_list);
930 nfs_request_remove_commit_list(req, cinfo);
931 nfs_list_add_request(req, dst);
932 ret++;
933 if ((ret == max) && !cinfo->dreq)
934 break;
935 }
936 return ret;
937 }
938
939 /*
940 * nfs_scan_commit - Scan an inode for commit requests
941 * @inode: NFS inode to scan
942 * @dst: mds destination list
943 * @cinfo: mds and ds lists of reqs ready to commit
944 *
945 * Moves requests from the inode's 'commit' request list.
946 * The requests are *not* checked to ensure that they form a contiguous set.
947 */
948 int
949 nfs_scan_commit(struct inode *inode, struct list_head *dst,
950 struct nfs_commit_info *cinfo)
951 {
952 int ret = 0;
953
954 spin_lock(cinfo->lock);
955 if (cinfo->mds->ncommit > 0) {
956 const int max = INT_MAX;
957
958 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
959 cinfo, max);
960 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
961 }
962 spin_unlock(cinfo->lock);
963 return ret;
964 }
965
966 /*
967 * Search for an existing write request, and attempt to update
968 * it to reflect a new dirty region on a given page.
969 *
970 * If the attempt fails, then the existing request is flushed out
971 * to disk.
972 */
973 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
974 struct page *page,
975 unsigned int offset,
976 unsigned int bytes)
977 {
978 struct nfs_page *req;
979 unsigned int rqend;
980 unsigned int end;
981 int error;
982
983 if (!PagePrivate(page))
984 return NULL;
985
986 end = offset + bytes;
987 spin_lock(&inode->i_lock);
988
989 for (;;) {
990 req = nfs_page_find_head_request_locked(NFS_I(inode), page);
991 if (req == NULL)
992 goto out_unlock;
993
994 /* should be handled by nfs_flush_incompatible */
995 WARN_ON_ONCE(req->wb_head != req);
996 WARN_ON_ONCE(req->wb_this_page != req);
997
998 rqend = req->wb_offset + req->wb_bytes;
999 /*
1000 * Tell the caller to flush out the request if
1001 * the offsets are non-contiguous.
1002 * Note: nfs_flush_incompatible() will already
1003 * have flushed out requests having wrong owners.
1004 */
1005 if (offset > rqend
1006 || end < req->wb_offset)
1007 goto out_flushme;
1008
1009 if (nfs_lock_request(req))
1010 break;
1011
1012 /* The request is locked, so wait and then retry */
1013 spin_unlock(&inode->i_lock);
1014 error = nfs_wait_on_request(req);
1015 nfs_release_request(req);
1016 if (error != 0)
1017 goto out_err;
1018 spin_lock(&inode->i_lock);
1019 }
1020
1021 /* Okay, the request matches. Update the region */
1022 if (offset < req->wb_offset) {
1023 req->wb_offset = offset;
1024 req->wb_pgbase = offset;
1025 }
1026 if (end > rqend)
1027 req->wb_bytes = end - req->wb_offset;
1028 else
1029 req->wb_bytes = rqend - req->wb_offset;
1030 out_unlock:
1031 if (req)
1032 nfs_clear_request_commit(req);
1033 spin_unlock(&inode->i_lock);
1034 return req;
1035 out_flushme:
1036 spin_unlock(&inode->i_lock);
1037 nfs_release_request(req);
1038 error = nfs_wb_page(inode, page);
1039 out_err:
1040 return ERR_PTR(error);
1041 }
1042
1043 /*
1044 * Try to update an existing write request, or create one if there is none.
1045 *
1046 * Note: Should always be called with the Page Lock held to prevent races
1047 * if we have to add a new request. Also assumes that the caller has
1048 * already called nfs_flush_incompatible() if necessary.
1049 */
1050 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1051 struct page *page, unsigned int offset, unsigned int bytes)
1052 {
1053 struct inode *inode = page_file_mapping(page)->host;
1054 struct nfs_page *req;
1055
1056 req = nfs_try_to_update_request(inode, page, offset, bytes);
1057 if (req != NULL)
1058 goto out;
1059 req = nfs_create_request(ctx, page, NULL, offset, bytes);
1060 if (IS_ERR(req))
1061 goto out;
1062 nfs_inode_add_request(inode, req);
1063 out:
1064 return req;
1065 }
1066
1067 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1068 unsigned int offset, unsigned int count)
1069 {
1070 struct nfs_page *req;
1071
1072 req = nfs_setup_write_request(ctx, page, offset, count);
1073 if (IS_ERR(req))
1074 return PTR_ERR(req);
1075 /* Update file length */
1076 nfs_grow_file(page, offset, count);
1077 nfs_mark_uptodate(req);
1078 nfs_mark_request_dirty(req);
1079 nfs_unlock_and_release_request(req);
1080 return 0;
1081 }
1082
1083 int nfs_flush_incompatible(struct file *file, struct page *page)
1084 {
1085 struct nfs_open_context *ctx = nfs_file_open_context(file);
1086 struct nfs_lock_context *l_ctx;
1087 struct nfs_page *req;
1088 int do_flush, status;
1089 /*
1090 * Look for a request corresponding to this page. If there
1091 * is one, and it belongs to another file, we flush it out
1092 * before we try to copy anything into the page. Do this
1093 * due to the lack of an ACCESS-type call in NFSv2.
1094 * Also do the same if we find a request from an existing
1095 * dropped page.
1096 */
1097 do {
1098 req = nfs_page_find_head_request(page);
1099 if (req == NULL)
1100 return 0;
1101 l_ctx = req->wb_lock_context;
1102 do_flush = req->wb_page != page || req->wb_context != ctx;
1103 /* for now, flush if more than 1 request in page_group */
1104 do_flush |= req->wb_this_page != req;
1105 if (l_ctx && ctx->dentry->d_inode->i_flock != NULL) {
1106 do_flush |= l_ctx->lockowner.l_owner != current->files
1107 || l_ctx->lockowner.l_pid != current->tgid;
1108 }
1109 nfs_release_request(req);
1110 if (!do_flush)
1111 return 0;
1112 status = nfs_wb_page(page_file_mapping(page)->host, page);
1113 } while (status == 0);
1114 return status;
1115 }
1116
1117 /*
1118 * Avoid buffered writes when a open context credential's key would
1119 * expire soon.
1120 *
1121 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1122 *
1123 * Return 0 and set a credential flag which triggers the inode to flush
1124 * and performs NFS_FILE_SYNC writes if the key will expired within
1125 * RPC_KEY_EXPIRE_TIMEO.
1126 */
1127 int
1128 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1129 {
1130 struct nfs_open_context *ctx = nfs_file_open_context(filp);
1131 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1132
1133 return rpcauth_key_timeout_notify(auth, ctx->cred);
1134 }
1135
1136 /*
1137 * Test if the open context credential key is marked to expire soon.
1138 */
1139 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
1140 {
1141 return rpcauth_cred_key_to_expire(ctx->cred);
1142 }
1143
1144 /*
1145 * If the page cache is marked as unsafe or invalid, then we can't rely on
1146 * the PageUptodate() flag. In this case, we will need to turn off
1147 * write optimisations that depend on the page contents being correct.
1148 */
1149 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1150 {
1151 struct nfs_inode *nfsi = NFS_I(inode);
1152
1153 if (nfs_have_delegated_attributes(inode))
1154 goto out;
1155 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1156 return false;
1157 smp_rmb();
1158 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1159 return false;
1160 out:
1161 if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1162 return false;
1163 return PageUptodate(page) != 0;
1164 }
1165
1166 /* If we know the page is up to date, and we're not using byte range locks (or
1167 * if we have the whole file locked for writing), it may be more efficient to
1168 * extend the write to cover the entire page in order to avoid fragmentation
1169 * inefficiencies.
1170 *
1171 * If the file is opened for synchronous writes then we can just skip the rest
1172 * of the checks.
1173 */
1174 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1175 {
1176 if (file->f_flags & O_DSYNC)
1177 return 0;
1178 if (!nfs_write_pageuptodate(page, inode))
1179 return 0;
1180 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1181 return 1;
1182 if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 &&
1183 inode->i_flock->fl_end == OFFSET_MAX &&
1184 inode->i_flock->fl_type != F_RDLCK))
1185 return 1;
1186 return 0;
1187 }
1188
1189 /*
1190 * Update and possibly write a cached page of an NFS file.
1191 *
1192 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1193 * things with a page scheduled for an RPC call (e.g. invalidate it).
1194 */
1195 int nfs_updatepage(struct file *file, struct page *page,
1196 unsigned int offset, unsigned int count)
1197 {
1198 struct nfs_open_context *ctx = nfs_file_open_context(file);
1199 struct inode *inode = page_file_mapping(page)->host;
1200 int status = 0;
1201
1202 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1203
1204 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1205 file, count, (long long)(page_file_offset(page) + offset));
1206
1207 if (!count)
1208 goto out;
1209
1210 if (nfs_can_extend_write(file, page, inode)) {
1211 count = max(count + offset, nfs_page_length(page));
1212 offset = 0;
1213 }
1214
1215 status = nfs_writepage_setup(ctx, page, offset, count);
1216 if (status < 0)
1217 nfs_set_pageerror(page);
1218 else
1219 __set_page_dirty_nobuffers(page);
1220 out:
1221 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1222 status, (long long)i_size_read(inode));
1223 return status;
1224 }
1225
1226 static int flush_task_priority(int how)
1227 {
1228 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1229 case FLUSH_HIGHPRI:
1230 return RPC_PRIORITY_HIGH;
1231 case FLUSH_LOWPRI:
1232 return RPC_PRIORITY_LOW;
1233 }
1234 return RPC_PRIORITY_NORMAL;
1235 }
1236
1237 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1238 struct rpc_message *msg,
1239 struct rpc_task_setup *task_setup_data, int how)
1240 {
1241 struct inode *inode = hdr->inode;
1242 int priority = flush_task_priority(how);
1243
1244 task_setup_data->priority = priority;
1245 NFS_PROTO(inode)->write_setup(hdr, msg);
1246
1247 nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client,
1248 &task_setup_data->rpc_client, msg, hdr);
1249 }
1250
1251 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1252 * call this on each, which will prepare them to be retried on next
1253 * writeback using standard nfs.
1254 */
1255 static void nfs_redirty_request(struct nfs_page *req)
1256 {
1257 nfs_mark_request_dirty(req);
1258 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1259 nfs_unlock_request(req);
1260 nfs_end_page_writeback(req);
1261 nfs_release_request(req);
1262 }
1263
1264 static void nfs_async_write_error(struct list_head *head)
1265 {
1266 struct nfs_page *req;
1267
1268 while (!list_empty(head)) {
1269 req = nfs_list_entry(head->next);
1270 nfs_list_remove_request(req);
1271 nfs_redirty_request(req);
1272 }
1273 }
1274
1275 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1276 .error_cleanup = nfs_async_write_error,
1277 .completion = nfs_write_completion,
1278 };
1279
1280 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1281 struct inode *inode, int ioflags, bool force_mds,
1282 const struct nfs_pgio_completion_ops *compl_ops)
1283 {
1284 struct nfs_server *server = NFS_SERVER(inode);
1285 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1286
1287 #ifdef CONFIG_NFS_V4_1
1288 if (server->pnfs_curr_ld && !force_mds)
1289 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1290 #endif
1291 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1292 server->wsize, ioflags);
1293 }
1294 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1295
1296 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1297 {
1298 pgio->pg_ops = &nfs_pgio_rw_ops;
1299 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1300 }
1301 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1302
1303
1304 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1305 {
1306 struct nfs_commit_data *data = calldata;
1307
1308 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1309 }
1310
1311 static void nfs_writeback_release_common(struct nfs_pgio_header *hdr)
1312 {
1313 /* do nothing! */
1314 }
1315
1316 /*
1317 * Special version of should_remove_suid() that ignores capabilities.
1318 */
1319 static int nfs_should_remove_suid(const struct inode *inode)
1320 {
1321 umode_t mode = inode->i_mode;
1322 int kill = 0;
1323
1324 /* suid always must be killed */
1325 if (unlikely(mode & S_ISUID))
1326 kill = ATTR_KILL_SUID;
1327
1328 /*
1329 * sgid without any exec bits is just a mandatory locking mark; leave
1330 * it alone. If some exec bits are set, it's a real sgid; kill it.
1331 */
1332 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1333 kill |= ATTR_KILL_SGID;
1334
1335 if (unlikely(kill && S_ISREG(mode)))
1336 return kill;
1337
1338 return 0;
1339 }
1340
1341 /*
1342 * This function is called when the WRITE call is complete.
1343 */
1344 static int nfs_writeback_done(struct rpc_task *task,
1345 struct nfs_pgio_header *hdr,
1346 struct inode *inode)
1347 {
1348 int status;
1349
1350 /*
1351 * ->write_done will attempt to use post-op attributes to detect
1352 * conflicting writes by other clients. A strict interpretation
1353 * of close-to-open would allow us to continue caching even if
1354 * another writer had changed the file, but some applications
1355 * depend on tighter cache coherency when writing.
1356 */
1357 status = NFS_PROTO(inode)->write_done(task, hdr);
1358 if (status != 0)
1359 return status;
1360 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1361
1362 if (hdr->res.verf->committed < hdr->args.stable &&
1363 task->tk_status >= 0) {
1364 /* We tried a write call, but the server did not
1365 * commit data to stable storage even though we
1366 * requested it.
1367 * Note: There is a known bug in Tru64 < 5.0 in which
1368 * the server reports NFS_DATA_SYNC, but performs
1369 * NFS_FILE_SYNC. We therefore implement this checking
1370 * as a dprintk() in order to avoid filling syslog.
1371 */
1372 static unsigned long complain;
1373
1374 /* Note this will print the MDS for a DS write */
1375 if (time_before(complain, jiffies)) {
1376 dprintk("NFS: faulty NFS server %s:"
1377 " (committed = %d) != (stable = %d)\n",
1378 NFS_SERVER(inode)->nfs_client->cl_hostname,
1379 hdr->res.verf->committed, hdr->args.stable);
1380 complain = jiffies + 300 * HZ;
1381 }
1382 }
1383
1384 /* Deal with the suid/sgid bit corner case */
1385 if (nfs_should_remove_suid(inode))
1386 nfs_mark_for_revalidate(inode);
1387 return 0;
1388 }
1389
1390 /*
1391 * This function is called when the WRITE call is complete.
1392 */
1393 static void nfs_writeback_result(struct rpc_task *task,
1394 struct nfs_pgio_header *hdr)
1395 {
1396 struct nfs_pgio_args *argp = &hdr->args;
1397 struct nfs_pgio_res *resp = &hdr->res;
1398
1399 if (resp->count < argp->count) {
1400 static unsigned long complain;
1401
1402 /* This a short write! */
1403 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1404
1405 /* Has the server at least made some progress? */
1406 if (resp->count == 0) {
1407 if (time_before(complain, jiffies)) {
1408 printk(KERN_WARNING
1409 "NFS: Server wrote zero bytes, expected %u.\n",
1410 argp->count);
1411 complain = jiffies + 300 * HZ;
1412 }
1413 nfs_set_pgio_error(hdr, -EIO, argp->offset);
1414 task->tk_status = -EIO;
1415 return;
1416 }
1417 /* Was this an NFSv2 write or an NFSv3 stable write? */
1418 if (resp->verf->committed != NFS_UNSTABLE) {
1419 /* Resend from where the server left off */
1420 hdr->mds_offset += resp->count;
1421 argp->offset += resp->count;
1422 argp->pgbase += resp->count;
1423 argp->count -= resp->count;
1424 } else {
1425 /* Resend as a stable write in order to avoid
1426 * headaches in the case of a server crash.
1427 */
1428 argp->stable = NFS_FILE_SYNC;
1429 }
1430 rpc_restart_call_prepare(task);
1431 }
1432 }
1433
1434
1435 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1436 {
1437 int ret;
1438
1439 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1440 return 1;
1441 if (!may_wait)
1442 return 0;
1443 ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1444 NFS_INO_COMMIT,
1445 nfs_wait_bit_killable,
1446 TASK_KILLABLE);
1447 return (ret < 0) ? ret : 1;
1448 }
1449
1450 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1451 {
1452 clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1453 smp_mb__after_atomic();
1454 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1455 }
1456
1457 void nfs_commitdata_release(struct nfs_commit_data *data)
1458 {
1459 put_nfs_open_context(data->context);
1460 nfs_commit_free(data);
1461 }
1462 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1463
1464 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1465 const struct rpc_call_ops *call_ops,
1466 int how, int flags)
1467 {
1468 struct rpc_task *task;
1469 int priority = flush_task_priority(how);
1470 struct rpc_message msg = {
1471 .rpc_argp = &data->args,
1472 .rpc_resp = &data->res,
1473 .rpc_cred = data->cred,
1474 };
1475 struct rpc_task_setup task_setup_data = {
1476 .task = &data->task,
1477 .rpc_client = clnt,
1478 .rpc_message = &msg,
1479 .callback_ops = call_ops,
1480 .callback_data = data,
1481 .workqueue = nfsiod_workqueue,
1482 .flags = RPC_TASK_ASYNC | flags,
1483 .priority = priority,
1484 };
1485 /* Set up the initial task struct. */
1486 NFS_PROTO(data->inode)->commit_setup(data, &msg);
1487
1488 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1489
1490 nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1491 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1492
1493 task = rpc_run_task(&task_setup_data);
1494 if (IS_ERR(task))
1495 return PTR_ERR(task);
1496 if (how & FLUSH_SYNC)
1497 rpc_wait_for_completion_task(task);
1498 rpc_put_task(task);
1499 return 0;
1500 }
1501 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1502
1503 static loff_t nfs_get_lwb(struct list_head *head)
1504 {
1505 loff_t lwb = 0;
1506 struct nfs_page *req;
1507
1508 list_for_each_entry(req, head, wb_list)
1509 if (lwb < (req_offset(req) + req->wb_bytes))
1510 lwb = req_offset(req) + req->wb_bytes;
1511
1512 return lwb;
1513 }
1514
1515 /*
1516 * Set up the argument/result storage required for the RPC call.
1517 */
1518 void nfs_init_commit(struct nfs_commit_data *data,
1519 struct list_head *head,
1520 struct pnfs_layout_segment *lseg,
1521 struct nfs_commit_info *cinfo)
1522 {
1523 struct nfs_page *first = nfs_list_entry(head->next);
1524 struct inode *inode = first->wb_context->dentry->d_inode;
1525
1526 /* Set up the RPC argument and reply structs
1527 * NB: take care not to mess about with data->commit et al. */
1528
1529 list_splice_init(head, &data->pages);
1530
1531 data->inode = inode;
1532 data->cred = first->wb_context->cred;
1533 data->lseg = lseg; /* reference transferred */
1534 /* only set lwb for pnfs commit */
1535 if (lseg)
1536 data->lwb = nfs_get_lwb(&data->pages);
1537 data->mds_ops = &nfs_commit_ops;
1538 data->completion_ops = cinfo->completion_ops;
1539 data->dreq = cinfo->dreq;
1540
1541 data->args.fh = NFS_FH(data->inode);
1542 /* Note: we always request a commit of the entire inode */
1543 data->args.offset = 0;
1544 data->args.count = 0;
1545 data->context = get_nfs_open_context(first->wb_context);
1546 data->res.fattr = &data->fattr;
1547 data->res.verf = &data->verf;
1548 nfs_fattr_init(&data->fattr);
1549 }
1550 EXPORT_SYMBOL_GPL(nfs_init_commit);
1551
1552 void nfs_retry_commit(struct list_head *page_list,
1553 struct pnfs_layout_segment *lseg,
1554 struct nfs_commit_info *cinfo)
1555 {
1556 struct nfs_page *req;
1557
1558 while (!list_empty(page_list)) {
1559 req = nfs_list_entry(page_list->next);
1560 nfs_list_remove_request(req);
1561 nfs_mark_request_commit(req, lseg, cinfo);
1562 if (!cinfo->dreq) {
1563 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1564 dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1565 BDI_RECLAIMABLE);
1566 }
1567 nfs_unlock_and_release_request(req);
1568 }
1569 }
1570 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1571
1572 /*
1573 * Commit dirty pages
1574 */
1575 static int
1576 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1577 struct nfs_commit_info *cinfo)
1578 {
1579 struct nfs_commit_data *data;
1580
1581 data = nfs_commitdata_alloc();
1582
1583 if (!data)
1584 goto out_bad;
1585
1586 /* Set up the argument struct */
1587 nfs_init_commit(data, head, NULL, cinfo);
1588 atomic_inc(&cinfo->mds->rpcs_out);
1589 return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1590 how, 0);
1591 out_bad:
1592 nfs_retry_commit(head, NULL, cinfo);
1593 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1594 return -ENOMEM;
1595 }
1596
1597 /*
1598 * COMMIT call returned
1599 */
1600 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1601 {
1602 struct nfs_commit_data *data = calldata;
1603
1604 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1605 task->tk_pid, task->tk_status);
1606
1607 /* Call the NFS version-specific code */
1608 NFS_PROTO(data->inode)->commit_done(task, data);
1609 }
1610
1611 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1612 {
1613 struct nfs_page *req;
1614 int status = data->task.tk_status;
1615 struct nfs_commit_info cinfo;
1616 struct nfs_server *nfss;
1617
1618 while (!list_empty(&data->pages)) {
1619 req = nfs_list_entry(data->pages.next);
1620 nfs_list_remove_request(req);
1621 nfs_clear_page_commit(req->wb_page);
1622
1623 dprintk("NFS: commit (%s/%llu %d@%lld)",
1624 req->wb_context->dentry->d_sb->s_id,
1625 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1626 req->wb_bytes,
1627 (long long)req_offset(req));
1628 if (status < 0) {
1629 nfs_context_set_write_error(req->wb_context, status);
1630 nfs_inode_remove_request(req);
1631 dprintk(", error = %d\n", status);
1632 goto next;
1633 }
1634
1635 /* Okay, COMMIT succeeded, apparently. Check the verifier
1636 * returned by the server against all stored verfs. */
1637 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1638 /* We have a match */
1639 nfs_inode_remove_request(req);
1640 dprintk(" OK\n");
1641 goto next;
1642 }
1643 /* We have a mismatch. Write the page again */
1644 dprintk(" mismatch\n");
1645 nfs_mark_request_dirty(req);
1646 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1647 next:
1648 nfs_unlock_and_release_request(req);
1649 /* Latency breaker */
1650 cond_resched();
1651 }
1652 nfss = NFS_SERVER(data->inode);
1653 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1654 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
1655
1656 nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1657 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1658 nfs_commit_clear_lock(NFS_I(data->inode));
1659 }
1660
1661 static void nfs_commit_release(void *calldata)
1662 {
1663 struct nfs_commit_data *data = calldata;
1664
1665 data->completion_ops->completion(data);
1666 nfs_commitdata_release(calldata);
1667 }
1668
1669 static const struct rpc_call_ops nfs_commit_ops = {
1670 .rpc_call_prepare = nfs_commit_prepare,
1671 .rpc_call_done = nfs_commit_done,
1672 .rpc_release = nfs_commit_release,
1673 };
1674
1675 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1676 .completion = nfs_commit_release_pages,
1677 .error_cleanup = nfs_commit_clear_lock,
1678 };
1679
1680 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1681 int how, struct nfs_commit_info *cinfo)
1682 {
1683 int status;
1684
1685 status = pnfs_commit_list(inode, head, how, cinfo);
1686 if (status == PNFS_NOT_ATTEMPTED)
1687 status = nfs_commit_list(inode, head, how, cinfo);
1688 return status;
1689 }
1690
1691 int nfs_commit_inode(struct inode *inode, int how)
1692 {
1693 LIST_HEAD(head);
1694 struct nfs_commit_info cinfo;
1695 int may_wait = how & FLUSH_SYNC;
1696 int res;
1697
1698 res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1699 if (res <= 0)
1700 goto out_mark_dirty;
1701 nfs_init_cinfo_from_inode(&cinfo, inode);
1702 res = nfs_scan_commit(inode, &head, &cinfo);
1703 if (res) {
1704 int error;
1705
1706 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1707 if (error < 0)
1708 return error;
1709 if (!may_wait)
1710 goto out_mark_dirty;
1711 error = wait_on_bit_action(&NFS_I(inode)->flags,
1712 NFS_INO_COMMIT,
1713 nfs_wait_bit_killable,
1714 TASK_KILLABLE);
1715 if (error < 0)
1716 return error;
1717 } else
1718 nfs_commit_clear_lock(NFS_I(inode));
1719 return res;
1720 /* Note: If we exit without ensuring that the commit is complete,
1721 * we must mark the inode as dirty. Otherwise, future calls to
1722 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1723 * that the data is on the disk.
1724 */
1725 out_mark_dirty:
1726 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1727 return res;
1728 }
1729
1730 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1731 {
1732 struct nfs_inode *nfsi = NFS_I(inode);
1733 int flags = FLUSH_SYNC;
1734 int ret = 0;
1735
1736 /* no commits means nothing needs to be done */
1737 if (!nfsi->commit_info.ncommit)
1738 return ret;
1739
1740 if (wbc->sync_mode == WB_SYNC_NONE) {
1741 /* Don't commit yet if this is a non-blocking flush and there
1742 * are a lot of outstanding writes for this mapping.
1743 */
1744 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1745 goto out_mark_dirty;
1746
1747 /* don't wait for the COMMIT response */
1748 flags = 0;
1749 }
1750
1751 ret = nfs_commit_inode(inode, flags);
1752 if (ret >= 0) {
1753 if (wbc->sync_mode == WB_SYNC_NONE) {
1754 if (ret < wbc->nr_to_write)
1755 wbc->nr_to_write -= ret;
1756 else
1757 wbc->nr_to_write = 0;
1758 }
1759 return 0;
1760 }
1761 out_mark_dirty:
1762 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1763 return ret;
1764 }
1765
1766 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1767 {
1768 return nfs_commit_unstable_pages(inode, wbc);
1769 }
1770 EXPORT_SYMBOL_GPL(nfs_write_inode);
1771
1772 /*
1773 * flush the inode to disk.
1774 */
1775 int nfs_wb_all(struct inode *inode)
1776 {
1777 struct writeback_control wbc = {
1778 .sync_mode = WB_SYNC_ALL,
1779 .nr_to_write = LONG_MAX,
1780 .range_start = 0,
1781 .range_end = LLONG_MAX,
1782 };
1783 int ret;
1784
1785 trace_nfs_writeback_inode_enter(inode);
1786
1787 ret = sync_inode(inode, &wbc);
1788
1789 trace_nfs_writeback_inode_exit(inode, ret);
1790 return ret;
1791 }
1792 EXPORT_SYMBOL_GPL(nfs_wb_all);
1793
1794 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1795 {
1796 struct nfs_page *req;
1797 int ret = 0;
1798
1799 wait_on_page_writeback(page);
1800
1801 /* blocking call to cancel all requests and join to a single (head)
1802 * request */
1803 req = nfs_lock_and_join_requests(page, false);
1804
1805 if (IS_ERR(req)) {
1806 ret = PTR_ERR(req);
1807 } else if (req) {
1808 /* all requests from this page have been cancelled by
1809 * nfs_lock_and_join_requests, so just remove the head
1810 * request from the inode / page_private pointer and
1811 * release it */
1812 nfs_inode_remove_request(req);
1813 /*
1814 * In case nfs_inode_remove_request has marked the
1815 * page as being dirty
1816 */
1817 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1818 nfs_unlock_and_release_request(req);
1819 }
1820
1821 return ret;
1822 }
1823
1824 /*
1825 * Write back all requests on one page - we do this before reading it.
1826 */
1827 int nfs_wb_page(struct inode *inode, struct page *page)
1828 {
1829 loff_t range_start = page_file_offset(page);
1830 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1831 struct writeback_control wbc = {
1832 .sync_mode = WB_SYNC_ALL,
1833 .nr_to_write = 0,
1834 .range_start = range_start,
1835 .range_end = range_end,
1836 };
1837 int ret;
1838
1839 trace_nfs_writeback_page_enter(inode);
1840
1841 for (;;) {
1842 wait_on_page_writeback(page);
1843 if (clear_page_dirty_for_io(page)) {
1844 ret = nfs_writepage_locked(page, &wbc);
1845 if (ret < 0)
1846 goto out_error;
1847 continue;
1848 }
1849 ret = 0;
1850 if (!PagePrivate(page))
1851 break;
1852 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1853 if (ret < 0)
1854 goto out_error;
1855 }
1856 out_error:
1857 trace_nfs_writeback_page_exit(inode, ret);
1858 return ret;
1859 }
1860
1861 #ifdef CONFIG_MIGRATION
1862 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1863 struct page *page, enum migrate_mode mode)
1864 {
1865 /*
1866 * If PagePrivate is set, then the page is currently associated with
1867 * an in-progress read or write request. Don't try to migrate it.
1868 *
1869 * FIXME: we could do this in principle, but we'll need a way to ensure
1870 * that we can safely release the inode reference while holding
1871 * the page lock.
1872 */
1873 if (PagePrivate(page))
1874 return -EBUSY;
1875
1876 if (!nfs_fscache_release_page(page, GFP_KERNEL))
1877 return -EBUSY;
1878
1879 return migrate_page(mapping, newpage, page, mode);
1880 }
1881 #endif
1882
1883 int __init nfs_init_writepagecache(void)
1884 {
1885 nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1886 sizeof(struct nfs_pgio_header),
1887 0, SLAB_HWCACHE_ALIGN,
1888 NULL);
1889 if (nfs_wdata_cachep == NULL)
1890 return -ENOMEM;
1891
1892 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1893 nfs_wdata_cachep);
1894 if (nfs_wdata_mempool == NULL)
1895 goto out_destroy_write_cache;
1896
1897 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1898 sizeof(struct nfs_commit_data),
1899 0, SLAB_HWCACHE_ALIGN,
1900 NULL);
1901 if (nfs_cdata_cachep == NULL)
1902 goto out_destroy_write_mempool;
1903
1904 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1905 nfs_cdata_cachep);
1906 if (nfs_commit_mempool == NULL)
1907 goto out_destroy_commit_cache;
1908
1909 /*
1910 * NFS congestion size, scale with available memory.
1911 *
1912 * 64MB: 8192k
1913 * 128MB: 11585k
1914 * 256MB: 16384k
1915 * 512MB: 23170k
1916 * 1GB: 32768k
1917 * 2GB: 46340k
1918 * 4GB: 65536k
1919 * 8GB: 92681k
1920 * 16GB: 131072k
1921 *
1922 * This allows larger machines to have larger/more transfers.
1923 * Limit the default to 256M
1924 */
1925 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1926 if (nfs_congestion_kb > 256*1024)
1927 nfs_congestion_kb = 256*1024;
1928
1929 return 0;
1930
1931 out_destroy_commit_cache:
1932 kmem_cache_destroy(nfs_cdata_cachep);
1933 out_destroy_write_mempool:
1934 mempool_destroy(nfs_wdata_mempool);
1935 out_destroy_write_cache:
1936 kmem_cache_destroy(nfs_wdata_cachep);
1937 return -ENOMEM;
1938 }
1939
1940 void nfs_destroy_writepagecache(void)
1941 {
1942 mempool_destroy(nfs_commit_mempool);
1943 kmem_cache_destroy(nfs_cdata_cachep);
1944 mempool_destroy(nfs_wdata_mempool);
1945 kmem_cache_destroy(nfs_wdata_cachep);
1946 }
1947
1948 static const struct nfs_rw_ops nfs_rw_write_ops = {
1949 .rw_mode = FMODE_WRITE,
1950 .rw_alloc_header = nfs_writehdr_alloc,
1951 .rw_free_header = nfs_writehdr_free,
1952 .rw_release = nfs_writeback_release_common,
1953 .rw_done = nfs_writeback_done,
1954 .rw_result = nfs_writeback_result,
1955 .rw_initiate = nfs_initiate_write,
1956 };
1957
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.