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