Merge branch 'work.misc' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[sfrench/cifs-2.6.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
27 #include <linux/uaccess.h>
28
29 #include "delegation.h"
30 #include "internal.h"
31 #include "iostat.h"
32 #include "nfs4_fs.h"
33 #include "fscache.h"
34 #include "pnfs.h"
35
36 #include "nfstrace.h"
37
38 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
39
40 #define MIN_POOL_WRITE          (32)
41 #define MIN_POOL_COMMIT         (4)
42
43 struct nfs_io_completion {
44         void (*complete)(void *data);
45         void *data;
46         struct kref refcount;
47 };
48
49 /*
50  * Local function declarations
51  */
52 static void nfs_redirty_request(struct nfs_page *req);
53 static const struct rpc_call_ops nfs_commit_ops;
54 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
55 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
56 static const struct nfs_rw_ops nfs_rw_write_ops;
57 static void nfs_clear_request_commit(struct nfs_page *req);
58 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
59                                       struct inode *inode);
60 static struct nfs_page *
61 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
62                                                 struct page *page);
63
64 static struct kmem_cache *nfs_wdata_cachep;
65 static mempool_t *nfs_wdata_mempool;
66 static struct kmem_cache *nfs_cdata_cachep;
67 static mempool_t *nfs_commit_mempool;
68
69 struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
70 {
71         struct nfs_commit_data *p;
72
73         if (never_fail)
74                 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
75         else {
76                 /* It is OK to do some reclaim, not no safe to wait
77                  * for anything to be returned to the pool.
78                  * mempool_alloc() cannot handle that particular combination,
79                  * so we need two separate attempts.
80                  */
81                 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
82                 if (!p)
83                         p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
84                                              __GFP_NOWARN | __GFP_NORETRY);
85                 if (!p)
86                         return NULL;
87         }
88
89         memset(p, 0, sizeof(*p));
90         INIT_LIST_HEAD(&p->pages);
91         return p;
92 }
93 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
94
95 void nfs_commit_free(struct nfs_commit_data *p)
96 {
97         mempool_free(p, nfs_commit_mempool);
98 }
99 EXPORT_SYMBOL_GPL(nfs_commit_free);
100
101 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
102 {
103         struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
104
105         memset(p, 0, sizeof(*p));
106         p->rw_mode = FMODE_WRITE;
107         return p;
108 }
109
110 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
111 {
112         mempool_free(hdr, nfs_wdata_mempool);
113 }
114
115 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
116 {
117         return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
118 }
119
120 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
121                 void (*complete)(void *), void *data)
122 {
123         ioc->complete = complete;
124         ioc->data = data;
125         kref_init(&ioc->refcount);
126 }
127
128 static void nfs_io_completion_release(struct kref *kref)
129 {
130         struct nfs_io_completion *ioc = container_of(kref,
131                         struct nfs_io_completion, refcount);
132         ioc->complete(ioc->data);
133         kfree(ioc);
134 }
135
136 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
137 {
138         if (ioc != NULL)
139                 kref_get(&ioc->refcount);
140 }
141
142 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
143 {
144         if (ioc != NULL)
145                 kref_put(&ioc->refcount, nfs_io_completion_release);
146 }
147
148 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
149 {
150         ctx->error = error;
151         smp_wmb();
152         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
153 }
154
155 static struct nfs_page *
156 nfs_page_private_request(struct page *page)
157 {
158         if (!PagePrivate(page))
159                 return NULL;
160         return (struct nfs_page *)page_private(page);
161 }
162
163 /*
164  * nfs_page_find_head_request_locked - find head request associated with @page
165  *
166  * must be called while holding the inode lock.
167  *
168  * returns matching head request with reference held, or NULL if not found.
169  */
170 static struct nfs_page *
171 nfs_page_find_private_request(struct page *page)
172 {
173         struct address_space *mapping = page_file_mapping(page);
174         struct nfs_page *req;
175
176         if (!PagePrivate(page))
177                 return NULL;
178         spin_lock(&mapping->private_lock);
179         req = nfs_page_private_request(page);
180         if (req) {
181                 WARN_ON_ONCE(req->wb_head != req);
182                 kref_get(&req->wb_kref);
183         }
184         spin_unlock(&mapping->private_lock);
185         return req;
186 }
187
188 static struct nfs_page *
189 nfs_page_find_swap_request(struct page *page)
190 {
191         struct inode *inode = page_file_mapping(page)->host;
192         struct nfs_inode *nfsi = NFS_I(inode);
193         struct nfs_page *req = NULL;
194         if (!PageSwapCache(page))
195                 return NULL;
196         mutex_lock(&nfsi->commit_mutex);
197         if (PageSwapCache(page)) {
198                 req = nfs_page_search_commits_for_head_request_locked(nfsi,
199                         page);
200                 if (req) {
201                         WARN_ON_ONCE(req->wb_head != req);
202                         kref_get(&req->wb_kref);
203                 }
204         }
205         mutex_unlock(&nfsi->commit_mutex);
206         return req;
207 }
208
209 /*
210  * nfs_page_find_head_request - find head request associated with @page
211  *
212  * returns matching head request with reference held, or NULL if not found.
213  */
214 static struct nfs_page *nfs_page_find_head_request(struct page *page)
215 {
216         struct nfs_page *req;
217
218         req = nfs_page_find_private_request(page);
219         if (!req)
220                 req = nfs_page_find_swap_request(page);
221         return req;
222 }
223
224 /* Adjust the file length if we're writing beyond the end */
225 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
226 {
227         struct inode *inode = page_file_mapping(page)->host;
228         loff_t end, i_size;
229         pgoff_t end_index;
230
231         spin_lock(&inode->i_lock);
232         i_size = i_size_read(inode);
233         end_index = (i_size - 1) >> PAGE_SHIFT;
234         if (i_size > 0 && page_index(page) < end_index)
235                 goto out;
236         end = page_file_offset(page) + ((loff_t)offset+count);
237         if (i_size >= end)
238                 goto out;
239         i_size_write(inode, end);
240         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
241 out:
242         spin_unlock(&inode->i_lock);
243 }
244
245 /* A writeback failed: mark the page as bad, and invalidate the page cache */
246 static void nfs_set_pageerror(struct page *page)
247 {
248         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
249 }
250
251 /*
252  * nfs_page_group_search_locked
253  * @head - head request of page group
254  * @page_offset - offset into page
255  *
256  * Search page group with head @head to find a request that contains the
257  * page offset @page_offset.
258  *
259  * Returns a pointer to the first matching nfs request, or NULL if no
260  * match is found.
261  *
262  * Must be called with the page group lock held
263  */
264 static struct nfs_page *
265 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
266 {
267         struct nfs_page *req;
268
269         req = head;
270         do {
271                 if (page_offset >= req->wb_pgbase &&
272                     page_offset < (req->wb_pgbase + req->wb_bytes))
273                         return req;
274
275                 req = req->wb_this_page;
276         } while (req != head);
277
278         return NULL;
279 }
280
281 /*
282  * nfs_page_group_covers_page
283  * @head - head request of page group
284  *
285  * Return true if the page group with head @head covers the whole page,
286  * returns false otherwise
287  */
288 static bool nfs_page_group_covers_page(struct nfs_page *req)
289 {
290         struct nfs_page *tmp;
291         unsigned int pos = 0;
292         unsigned int len = nfs_page_length(req->wb_page);
293
294         nfs_page_group_lock(req);
295
296         for (;;) {
297                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
298                 if (!tmp)
299                         break;
300                 pos = tmp->wb_pgbase + tmp->wb_bytes;
301         }
302
303         nfs_page_group_unlock(req);
304         return pos >= len;
305 }
306
307 /* We can set the PG_uptodate flag if we see that a write request
308  * covers the full page.
309  */
310 static void nfs_mark_uptodate(struct nfs_page *req)
311 {
312         if (PageUptodate(req->wb_page))
313                 return;
314         if (!nfs_page_group_covers_page(req))
315                 return;
316         SetPageUptodate(req->wb_page);
317 }
318
319 static int wb_priority(struct writeback_control *wbc)
320 {
321         int ret = 0;
322
323         if (wbc->sync_mode == WB_SYNC_ALL)
324                 ret = FLUSH_COND_STABLE;
325         return ret;
326 }
327
328 /*
329  * NFS congestion control
330  */
331
332 int nfs_congestion_kb;
333
334 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
335 #define NFS_CONGESTION_OFF_THRESH       \
336         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
337
338 static void nfs_set_page_writeback(struct page *page)
339 {
340         struct inode *inode = page_file_mapping(page)->host;
341         struct nfs_server *nfss = NFS_SERVER(inode);
342         int ret = test_set_page_writeback(page);
343
344         WARN_ON_ONCE(ret != 0);
345
346         if (atomic_long_inc_return(&nfss->writeback) >
347                         NFS_CONGESTION_ON_THRESH)
348                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
349 }
350
351 static void nfs_end_page_writeback(struct nfs_page *req)
352 {
353         struct inode *inode = page_file_mapping(req->wb_page)->host;
354         struct nfs_server *nfss = NFS_SERVER(inode);
355         bool is_done;
356
357         is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
358         nfs_unlock_request(req);
359         if (!is_done)
360                 return;
361
362         end_page_writeback(req->wb_page);
363         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
364                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
365 }
366
367 /*
368  * nfs_unroll_locks_and_wait -  unlock all newly locked reqs and wait on @req
369  *
370  * this is a helper function for nfs_lock_and_join_requests
371  *
372  * @inode - inode associated with request page group, must be holding inode lock
373  * @head  - head request of page group, must be holding head lock
374  * @req   - request that couldn't lock and needs to wait on the req bit lock
375  *
376  * NOTE: this must be called holding page_group bit lock
377  *       which will be released before returning.
378  *
379  * returns 0 on success, < 0 on error.
380  */
381 static void
382 nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
383                           struct nfs_page *req)
384 {
385         struct nfs_page *tmp;
386
387         /* relinquish all the locks successfully grabbed this run */
388         for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
389                 if (!kref_read(&tmp->wb_kref))
390                         continue;
391                 nfs_unlock_and_release_request(tmp);
392         }
393 }
394
395 /*
396  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
397  *
398  * @destroy_list - request list (using wb_this_page) terminated by @old_head
399  * @old_head - the old head of the list
400  *
401  * All subrequests must be locked and removed from all lists, so at this point
402  * they are only "active" in this function, and possibly in nfs_wait_on_request
403  * with a reference held by some other context.
404  */
405 static void
406 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
407                                  struct nfs_page *old_head,
408                                  struct inode *inode)
409 {
410         while (destroy_list) {
411                 struct nfs_page *subreq = destroy_list;
412
413                 destroy_list = (subreq->wb_this_page == old_head) ?
414                                    NULL : subreq->wb_this_page;
415
416                 WARN_ON_ONCE(old_head != subreq->wb_head);
417
418                 /* make sure old group is not used */
419                 subreq->wb_this_page = subreq;
420
421                 clear_bit(PG_REMOVE, &subreq->wb_flags);
422
423                 /* Note: races with nfs_page_group_destroy() */
424                 if (!kref_read(&subreq->wb_kref)) {
425                         /* Check if we raced with nfs_page_group_destroy() */
426                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags))
427                                 nfs_free_request(subreq);
428                         continue;
429                 }
430
431                 subreq->wb_head = subreq;
432
433                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
434                         nfs_release_request(subreq);
435                         atomic_long_dec(&NFS_I(inode)->nrequests);
436                 }
437
438                 /* subreq is now totally disconnected from page group or any
439                  * write / commit lists. last chance to wake any waiters */
440                 nfs_unlock_and_release_request(subreq);
441         }
442 }
443
444 /*
445  * nfs_lock_and_join_requests - join all subreqs to the head req and return
446  *                              a locked reference, cancelling any pending
447  *                              operations for this page.
448  *
449  * @page - the page used to lookup the "page group" of nfs_page structures
450  *
451  * This function joins all sub requests to the head request by first
452  * locking all requests in the group, cancelling any pending operations
453  * and finally updating the head request to cover the whole range covered by
454  * the (former) group.  All subrequests are removed from any write or commit
455  * lists, unlinked from the group and destroyed.
456  *
457  * Returns a locked, referenced pointer to the head request - which after
458  * this call is guaranteed to be the only request associated with the page.
459  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
460  * error was encountered.
461  */
462 static struct nfs_page *
463 nfs_lock_and_join_requests(struct page *page)
464 {
465         struct inode *inode = page_file_mapping(page)->host;
466         struct nfs_page *head, *subreq;
467         struct nfs_page *destroy_list = NULL;
468         unsigned int total_bytes;
469         int ret;
470
471 try_again:
472         /*
473          * A reference is taken only on the head request which acts as a
474          * reference to the whole page group - the group will not be destroyed
475          * until the head reference is released.
476          */
477         head = nfs_page_find_head_request(page);
478         if (!head)
479                 return NULL;
480
481         /* lock the page head first in order to avoid an ABBA inefficiency */
482         if (!nfs_lock_request(head)) {
483                 ret = nfs_wait_on_request(head);
484                 nfs_release_request(head);
485                 if (ret < 0)
486                         return ERR_PTR(ret);
487                 goto try_again;
488         }
489
490         /* Ensure that nobody removed the request before we locked it */
491         if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
492                 nfs_unlock_and_release_request(head);
493                 goto try_again;
494         }
495
496         ret = nfs_page_group_lock(head);
497         if (ret < 0) {
498                 nfs_unlock_and_release_request(head);
499                 return ERR_PTR(ret);
500         }
501
502         /* lock each request in the page group */
503         total_bytes = head->wb_bytes;
504         for (subreq = head->wb_this_page; subreq != head;
505                         subreq = subreq->wb_this_page) {
506
507                 if (!kref_get_unless_zero(&subreq->wb_kref)) {
508                         if (subreq->wb_offset == head->wb_offset + total_bytes)
509                                 total_bytes += subreq->wb_bytes;
510                         continue;
511                 }
512
513                 while (!nfs_lock_request(subreq)) {
514                         /*
515                          * Unlock page to allow nfs_page_group_sync_on_bit()
516                          * to succeed
517                          */
518                         nfs_page_group_unlock(head);
519                         ret = nfs_wait_on_request(subreq);
520                         if (!ret)
521                                 ret = nfs_page_group_lock(head);
522                         if (ret < 0) {
523                                 nfs_unroll_locks(inode, head, subreq);
524                                 nfs_release_request(subreq);
525                                 nfs_unlock_and_release_request(head);
526                                 return ERR_PTR(ret);
527                         }
528                 }
529                 /*
530                  * Subrequests are always contiguous, non overlapping
531                  * and in order - but may be repeated (mirrored writes).
532                  */
533                 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
534                         /* keep track of how many bytes this group covers */
535                         total_bytes += subreq->wb_bytes;
536                 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
537                             ((subreq->wb_offset + subreq->wb_bytes) >
538                              (head->wb_offset + total_bytes)))) {
539                         nfs_page_group_unlock(head);
540                         nfs_unroll_locks(inode, head, subreq);
541                         nfs_unlock_and_release_request(subreq);
542                         nfs_unlock_and_release_request(head);
543                         return ERR_PTR(-EIO);
544                 }
545         }
546
547         /* Now that all requests are locked, make sure they aren't on any list.
548          * Commit list removal accounting is done after locks are dropped */
549         subreq = head;
550         do {
551                 nfs_clear_request_commit(subreq);
552                 subreq = subreq->wb_this_page;
553         } while (subreq != head);
554
555         /* unlink subrequests from head, destroy them later */
556         if (head->wb_this_page != head) {
557                 /* destroy list will be terminated by head */
558                 destroy_list = head->wb_this_page;
559                 head->wb_this_page = head;
560
561                 /* change head request to cover whole range that
562                  * the former page group covered */
563                 head->wb_bytes = total_bytes;
564         }
565
566         /* Postpone destruction of this request */
567         if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
568                 set_bit(PG_INODE_REF, &head->wb_flags);
569                 kref_get(&head->wb_kref);
570                 atomic_long_inc(&NFS_I(inode)->nrequests);
571         }
572
573         nfs_page_group_unlock(head);
574
575         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
576
577         /* Did we lose a race with nfs_inode_remove_request()? */
578         if (!(PagePrivate(page) || PageSwapCache(page))) {
579                 nfs_unlock_and_release_request(head);
580                 return NULL;
581         }
582
583         /* still holds ref on head from nfs_page_find_head_request
584          * and still has lock on head from lock loop */
585         return head;
586 }
587
588 static void nfs_write_error_remove_page(struct nfs_page *req)
589 {
590         nfs_end_page_writeback(req);
591         generic_error_remove_page(page_file_mapping(req->wb_page),
592                                   req->wb_page);
593         nfs_release_request(req);
594 }
595
596 static bool
597 nfs_error_is_fatal_on_server(int err)
598 {
599         switch (err) {
600         case 0:
601         case -ERESTARTSYS:
602         case -EINTR:
603                 return false;
604         }
605         return nfs_error_is_fatal(err);
606 }
607
608 /*
609  * Find an associated nfs write request, and prepare to flush it out
610  * May return an error if the user signalled nfs_wait_on_request().
611  */
612 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
613                                 struct page *page)
614 {
615         struct nfs_page *req;
616         int ret = 0;
617
618         req = nfs_lock_and_join_requests(page);
619         if (!req)
620                 goto out;
621         ret = PTR_ERR(req);
622         if (IS_ERR(req))
623                 goto out;
624
625         nfs_set_page_writeback(page);
626         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
627
628         ret = 0;
629         /* If there is a fatal error that covers this write, just exit */
630         if (nfs_error_is_fatal_on_server(req->wb_context->error))
631                 goto out_launder;
632
633         if (!nfs_pageio_add_request(pgio, req)) {
634                 ret = pgio->pg_error;
635                 /*
636                  * Remove the problematic req upon fatal errors on the server
637                  */
638                 if (nfs_error_is_fatal(ret)) {
639                         nfs_context_set_write_error(req->wb_context, ret);
640                         if (nfs_error_is_fatal_on_server(ret))
641                                 goto out_launder;
642                 }
643                 nfs_redirty_request(req);
644                 ret = -EAGAIN;
645         } else
646                 nfs_add_stats(page_file_mapping(page)->host,
647                                 NFSIOS_WRITEPAGES, 1);
648 out:
649         return ret;
650 out_launder:
651         nfs_write_error_remove_page(req);
652         return ret;
653 }
654
655 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
656                             struct nfs_pageio_descriptor *pgio)
657 {
658         int ret;
659
660         nfs_pageio_cond_complete(pgio, page_index(page));
661         ret = nfs_page_async_flush(pgio, page);
662         if (ret == -EAGAIN) {
663                 redirty_page_for_writepage(wbc, page);
664                 ret = 0;
665         }
666         return ret;
667 }
668
669 /*
670  * Write an mmapped page to the server.
671  */
672 static int nfs_writepage_locked(struct page *page,
673                                 struct writeback_control *wbc)
674 {
675         struct nfs_pageio_descriptor pgio;
676         struct inode *inode = page_file_mapping(page)->host;
677         int err;
678
679         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
680         nfs_pageio_init_write(&pgio, inode, 0,
681                                 false, &nfs_async_write_completion_ops);
682         err = nfs_do_writepage(page, wbc, &pgio);
683         nfs_pageio_complete(&pgio);
684         if (err < 0)
685                 return err;
686         if (pgio.pg_error < 0)
687                 return pgio.pg_error;
688         return 0;
689 }
690
691 int nfs_writepage(struct page *page, struct writeback_control *wbc)
692 {
693         int ret;
694
695         ret = nfs_writepage_locked(page, wbc);
696         unlock_page(page);
697         return ret;
698 }
699
700 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
701 {
702         int ret;
703
704         ret = nfs_do_writepage(page, wbc, data);
705         unlock_page(page);
706         return ret;
707 }
708
709 static void nfs_io_completion_commit(void *inode)
710 {
711         nfs_commit_inode(inode, 0);
712 }
713
714 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
715 {
716         struct inode *inode = mapping->host;
717         struct nfs_pageio_descriptor pgio;
718         struct nfs_io_completion *ioc = nfs_io_completion_alloc(GFP_NOFS);
719         int err;
720
721         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
722
723         if (ioc)
724                 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
725
726         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
727                                 &nfs_async_write_completion_ops);
728         pgio.pg_io_completion = ioc;
729         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
730         nfs_pageio_complete(&pgio);
731         nfs_io_completion_put(ioc);
732
733         if (err < 0)
734                 goto out_err;
735         err = pgio.pg_error;
736         if (err < 0)
737                 goto out_err;
738         return 0;
739 out_err:
740         return err;
741 }
742
743 /*
744  * Insert a write request into an inode
745  */
746 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
747 {
748         struct address_space *mapping = page_file_mapping(req->wb_page);
749         struct nfs_inode *nfsi = NFS_I(inode);
750
751         WARN_ON_ONCE(req->wb_this_page != req);
752
753         /* Lock the request! */
754         nfs_lock_request(req);
755
756         /*
757          * Swap-space should not get truncated. Hence no need to plug the race
758          * with invalidate/truncate.
759          */
760         spin_lock(&mapping->private_lock);
761         if (!nfs_have_writebacks(inode) &&
762             NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE)) {
763                 spin_lock(&inode->i_lock);
764                 inode->i_version++;
765                 spin_unlock(&inode->i_lock);
766         }
767         if (likely(!PageSwapCache(req->wb_page))) {
768                 set_bit(PG_MAPPED, &req->wb_flags);
769                 SetPagePrivate(req->wb_page);
770                 set_page_private(req->wb_page, (unsigned long)req);
771         }
772         spin_unlock(&mapping->private_lock);
773         atomic_long_inc(&nfsi->nrequests);
774         /* this a head request for a page group - mark it as having an
775          * extra reference so sub groups can follow suit.
776          * This flag also informs pgio layer when to bump nrequests when
777          * adding subrequests. */
778         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
779         kref_get(&req->wb_kref);
780 }
781
782 /*
783  * Remove a write request from an inode
784  */
785 static void nfs_inode_remove_request(struct nfs_page *req)
786 {
787         struct address_space *mapping = page_file_mapping(req->wb_page);
788         struct inode *inode = mapping->host;
789         struct nfs_inode *nfsi = NFS_I(inode);
790         struct nfs_page *head;
791
792         atomic_long_dec(&nfsi->nrequests);
793         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
794                 head = req->wb_head;
795
796                 spin_lock(&mapping->private_lock);
797                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
798                         set_page_private(head->wb_page, 0);
799                         ClearPagePrivate(head->wb_page);
800                         clear_bit(PG_MAPPED, &head->wb_flags);
801                 }
802                 spin_unlock(&mapping->private_lock);
803         }
804
805         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
806                 nfs_release_request(req);
807 }
808
809 static void
810 nfs_mark_request_dirty(struct nfs_page *req)
811 {
812         if (req->wb_page)
813                 __set_page_dirty_nobuffers(req->wb_page);
814 }
815
816 /*
817  * nfs_page_search_commits_for_head_request_locked
818  *
819  * Search through commit lists on @inode for the head request for @page.
820  * Must be called while holding the inode (which is cinfo) lock.
821  *
822  * Returns the head request if found, or NULL if not found.
823  */
824 static struct nfs_page *
825 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
826                                                 struct page *page)
827 {
828         struct nfs_page *freq, *t;
829         struct nfs_commit_info cinfo;
830         struct inode *inode = &nfsi->vfs_inode;
831
832         nfs_init_cinfo_from_inode(&cinfo, inode);
833
834         /* search through pnfs commit lists */
835         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
836         if (freq)
837                 return freq->wb_head;
838
839         /* Linearly search the commit list for the correct request */
840         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
841                 if (freq->wb_page == page)
842                         return freq->wb_head;
843         }
844
845         return NULL;
846 }
847
848 /**
849  * nfs_request_add_commit_list_locked - add request to a commit list
850  * @req: pointer to a struct nfs_page
851  * @dst: commit list head
852  * @cinfo: holds list lock and accounting info
853  *
854  * This sets the PG_CLEAN bit, updates the cinfo count of
855  * number of outstanding requests requiring a commit as well as
856  * the MM page stats.
857  *
858  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
859  * nfs_page lock.
860  */
861 void
862 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
863                             struct nfs_commit_info *cinfo)
864 {
865         set_bit(PG_CLEAN, &req->wb_flags);
866         nfs_list_add_request(req, dst);
867         atomic_long_inc(&cinfo->mds->ncommit);
868 }
869 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
870
871 /**
872  * nfs_request_add_commit_list - add request to a commit list
873  * @req: pointer to a struct nfs_page
874  * @dst: commit list head
875  * @cinfo: holds list lock and accounting info
876  *
877  * This sets the PG_CLEAN bit, updates the cinfo count of
878  * number of outstanding requests requiring a commit as well as
879  * the MM page stats.
880  *
881  * The caller must _not_ hold the cinfo->lock, but must be
882  * holding the nfs_page lock.
883  */
884 void
885 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
886 {
887         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
888         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
889         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
890         if (req->wb_page)
891                 nfs_mark_page_unstable(req->wb_page, cinfo);
892 }
893 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
894
895 /**
896  * nfs_request_remove_commit_list - Remove request from a commit list
897  * @req: pointer to a nfs_page
898  * @cinfo: holds list lock and accounting info
899  *
900  * This clears the PG_CLEAN bit, and updates the cinfo's count of
901  * number of outstanding requests requiring a commit
902  * It does not update the MM page stats.
903  *
904  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
905  */
906 void
907 nfs_request_remove_commit_list(struct nfs_page *req,
908                                struct nfs_commit_info *cinfo)
909 {
910         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
911                 return;
912         nfs_list_remove_request(req);
913         atomic_long_dec(&cinfo->mds->ncommit);
914 }
915 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
916
917 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
918                                       struct inode *inode)
919 {
920         cinfo->inode = inode;
921         cinfo->mds = &NFS_I(inode)->commit_info;
922         cinfo->ds = pnfs_get_ds_info(inode);
923         cinfo->dreq = NULL;
924         cinfo->completion_ops = &nfs_commit_completion_ops;
925 }
926
927 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
928                     struct inode *inode,
929                     struct nfs_direct_req *dreq)
930 {
931         if (dreq)
932                 nfs_init_cinfo_from_dreq(cinfo, dreq);
933         else
934                 nfs_init_cinfo_from_inode(cinfo, inode);
935 }
936 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
937
938 /*
939  * Add a request to the inode's commit list.
940  */
941 void
942 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
943                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
944 {
945         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
946                 return;
947         nfs_request_add_commit_list(req, cinfo);
948 }
949
950 static void
951 nfs_clear_page_commit(struct page *page)
952 {
953         dec_node_page_state(page, NR_UNSTABLE_NFS);
954         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
955                     WB_RECLAIMABLE);
956 }
957
958 /* Called holding the request lock on @req */
959 static void
960 nfs_clear_request_commit(struct nfs_page *req)
961 {
962         if (test_bit(PG_CLEAN, &req->wb_flags)) {
963                 struct inode *inode = d_inode(req->wb_context->dentry);
964                 struct nfs_commit_info cinfo;
965
966                 nfs_init_cinfo_from_inode(&cinfo, inode);
967                 mutex_lock(&NFS_I(inode)->commit_mutex);
968                 if (!pnfs_clear_request_commit(req, &cinfo)) {
969                         nfs_request_remove_commit_list(req, &cinfo);
970                 }
971                 mutex_unlock(&NFS_I(inode)->commit_mutex);
972                 nfs_clear_page_commit(req->wb_page);
973         }
974 }
975
976 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
977 {
978         if (hdr->verf.committed == NFS_DATA_SYNC)
979                 return hdr->lseg == NULL;
980         return hdr->verf.committed != NFS_FILE_SYNC;
981 }
982
983 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
984 {
985         nfs_io_completion_get(hdr->io_completion);
986 }
987
988 static void nfs_write_completion(struct nfs_pgio_header *hdr)
989 {
990         struct nfs_commit_info cinfo;
991         unsigned long bytes = 0;
992
993         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
994                 goto out;
995         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
996         while (!list_empty(&hdr->pages)) {
997                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
998
999                 bytes += req->wb_bytes;
1000                 nfs_list_remove_request(req);
1001                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1002                     (hdr->good_bytes < bytes)) {
1003                         nfs_set_pageerror(req->wb_page);
1004                         nfs_context_set_write_error(req->wb_context, hdr->error);
1005                         goto remove_req;
1006                 }
1007                 if (nfs_write_need_commit(hdr)) {
1008                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1009                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1010                                 hdr->pgio_mirror_idx);
1011                         goto next;
1012                 }
1013 remove_req:
1014                 nfs_inode_remove_request(req);
1015 next:
1016                 nfs_end_page_writeback(req);
1017                 nfs_release_request(req);
1018         }
1019 out:
1020         nfs_io_completion_put(hdr->io_completion);
1021         hdr->release(hdr);
1022 }
1023
1024 unsigned long
1025 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1026 {
1027         return atomic_long_read(&cinfo->mds->ncommit);
1028 }
1029
1030 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1031 int
1032 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1033                      struct nfs_commit_info *cinfo, int max)
1034 {
1035         struct nfs_page *req, *tmp;
1036         int ret = 0;
1037
1038 restart:
1039         list_for_each_entry_safe(req, tmp, src, wb_list) {
1040                 kref_get(&req->wb_kref);
1041                 if (!nfs_lock_request(req)) {
1042                         int status;
1043
1044                         /* Prevent deadlock with nfs_lock_and_join_requests */
1045                         if (!list_empty(dst)) {
1046                                 nfs_release_request(req);
1047                                 continue;
1048                         }
1049                         /* Ensure we make progress to prevent livelock */
1050                         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1051                         status = nfs_wait_on_request(req);
1052                         nfs_release_request(req);
1053                         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1054                         if (status < 0)
1055                                 break;
1056                         goto restart;
1057                 }
1058                 nfs_request_remove_commit_list(req, cinfo);
1059                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1060                 nfs_list_add_request(req, dst);
1061                 ret++;
1062                 if ((ret == max) && !cinfo->dreq)
1063                         break;
1064                 cond_resched();
1065         }
1066         return ret;
1067 }
1068 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1069
1070 /*
1071  * nfs_scan_commit - Scan an inode for commit requests
1072  * @inode: NFS inode to scan
1073  * @dst: mds destination list
1074  * @cinfo: mds and ds lists of reqs ready to commit
1075  *
1076  * Moves requests from the inode's 'commit' request list.
1077  * The requests are *not* checked to ensure that they form a contiguous set.
1078  */
1079 int
1080 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1081                 struct nfs_commit_info *cinfo)
1082 {
1083         int ret = 0;
1084
1085         if (!atomic_long_read(&cinfo->mds->ncommit))
1086                 return 0;
1087         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1088         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1089                 const int max = INT_MAX;
1090
1091                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1092                                            cinfo, max);
1093                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1094         }
1095         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1096         return ret;
1097 }
1098
1099 /*
1100  * Search for an existing write request, and attempt to update
1101  * it to reflect a new dirty region on a given page.
1102  *
1103  * If the attempt fails, then the existing request is flushed out
1104  * to disk.
1105  */
1106 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1107                 struct page *page,
1108                 unsigned int offset,
1109                 unsigned int bytes)
1110 {
1111         struct nfs_page *req;
1112         unsigned int rqend;
1113         unsigned int end;
1114         int error;
1115
1116         end = offset + bytes;
1117
1118         req = nfs_lock_and_join_requests(page);
1119         if (IS_ERR_OR_NULL(req))
1120                 return req;
1121
1122         rqend = req->wb_offset + req->wb_bytes;
1123         /*
1124          * Tell the caller to flush out the request if
1125          * the offsets are non-contiguous.
1126          * Note: nfs_flush_incompatible() will already
1127          * have flushed out requests having wrong owners.
1128          */
1129         if (offset > rqend || end < req->wb_offset)
1130                 goto out_flushme;
1131
1132         /* Okay, the request matches. Update the region */
1133         if (offset < req->wb_offset) {
1134                 req->wb_offset = offset;
1135                 req->wb_pgbase = offset;
1136         }
1137         if (end > rqend)
1138                 req->wb_bytes = end - req->wb_offset;
1139         else
1140                 req->wb_bytes = rqend - req->wb_offset;
1141         return req;
1142 out_flushme:
1143         /*
1144          * Note: we mark the request dirty here because
1145          * nfs_lock_and_join_requests() cannot preserve
1146          * commit flags, so we have to replay the write.
1147          */
1148         nfs_mark_request_dirty(req);
1149         nfs_unlock_and_release_request(req);
1150         error = nfs_wb_page(inode, page);
1151         return (error < 0) ? ERR_PTR(error) : NULL;
1152 }
1153
1154 /*
1155  * Try to update an existing write request, or create one if there is none.
1156  *
1157  * Note: Should always be called with the Page Lock held to prevent races
1158  * if we have to add a new request. Also assumes that the caller has
1159  * already called nfs_flush_incompatible() if necessary.
1160  */
1161 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1162                 struct page *page, unsigned int offset, unsigned int bytes)
1163 {
1164         struct inode *inode = page_file_mapping(page)->host;
1165         struct nfs_page *req;
1166
1167         req = nfs_try_to_update_request(inode, page, offset, bytes);
1168         if (req != NULL)
1169                 goto out;
1170         req = nfs_create_request(ctx, page, NULL, offset, bytes);
1171         if (IS_ERR(req))
1172                 goto out;
1173         nfs_inode_add_request(inode, req);
1174 out:
1175         return req;
1176 }
1177
1178 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1179                 unsigned int offset, unsigned int count)
1180 {
1181         struct nfs_page *req;
1182
1183         req = nfs_setup_write_request(ctx, page, offset, count);
1184         if (IS_ERR(req))
1185                 return PTR_ERR(req);
1186         /* Update file length */
1187         nfs_grow_file(page, offset, count);
1188         nfs_mark_uptodate(req);
1189         nfs_mark_request_dirty(req);
1190         nfs_unlock_and_release_request(req);
1191         return 0;
1192 }
1193
1194 int nfs_flush_incompatible(struct file *file, struct page *page)
1195 {
1196         struct nfs_open_context *ctx = nfs_file_open_context(file);
1197         struct nfs_lock_context *l_ctx;
1198         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1199         struct nfs_page *req;
1200         int do_flush, status;
1201         /*
1202          * Look for a request corresponding to this page. If there
1203          * is one, and it belongs to another file, we flush it out
1204          * before we try to copy anything into the page. Do this
1205          * due to the lack of an ACCESS-type call in NFSv2.
1206          * Also do the same if we find a request from an existing
1207          * dropped page.
1208          */
1209         do {
1210                 req = nfs_page_find_head_request(page);
1211                 if (req == NULL)
1212                         return 0;
1213                 l_ctx = req->wb_lock_context;
1214                 do_flush = req->wb_page != page ||
1215                         !nfs_match_open_context(req->wb_context, ctx);
1216                 if (l_ctx && flctx &&
1217                     !(list_empty_careful(&flctx->flc_posix) &&
1218                       list_empty_careful(&flctx->flc_flock))) {
1219                         do_flush |= l_ctx->lockowner != current->files;
1220                 }
1221                 nfs_release_request(req);
1222                 if (!do_flush)
1223                         return 0;
1224                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1225         } while (status == 0);
1226         return status;
1227 }
1228
1229 /*
1230  * Avoid buffered writes when a open context credential's key would
1231  * expire soon.
1232  *
1233  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1234  *
1235  * Return 0 and set a credential flag which triggers the inode to flush
1236  * and performs  NFS_FILE_SYNC writes if the key will expired within
1237  * RPC_KEY_EXPIRE_TIMEO.
1238  */
1239 int
1240 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1241 {
1242         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1243         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1244
1245         return rpcauth_key_timeout_notify(auth, ctx->cred);
1246 }
1247
1248 /*
1249  * Test if the open context credential key is marked to expire soon.
1250  */
1251 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1252 {
1253         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1254
1255         return rpcauth_cred_key_to_expire(auth, ctx->cred);
1256 }
1257
1258 /*
1259  * If the page cache is marked as unsafe or invalid, then we can't rely on
1260  * the PageUptodate() flag. In this case, we will need to turn off
1261  * write optimisations that depend on the page contents being correct.
1262  */
1263 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1264 {
1265         struct nfs_inode *nfsi = NFS_I(inode);
1266
1267         if (nfs_have_delegated_attributes(inode))
1268                 goto out;
1269         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1270                 return false;
1271         smp_rmb();
1272         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1273                 return false;
1274 out:
1275         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1276                 return false;
1277         return PageUptodate(page) != 0;
1278 }
1279
1280 static bool
1281 is_whole_file_wrlock(struct file_lock *fl)
1282 {
1283         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1284                         fl->fl_type == F_WRLCK;
1285 }
1286
1287 /* If we know the page is up to date, and we're not using byte range locks (or
1288  * if we have the whole file locked for writing), it may be more efficient to
1289  * extend the write to cover the entire page in order to avoid fragmentation
1290  * inefficiencies.
1291  *
1292  * If the file is opened for synchronous writes then we can just skip the rest
1293  * of the checks.
1294  */
1295 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1296 {
1297         int ret;
1298         struct file_lock_context *flctx = inode->i_flctx;
1299         struct file_lock *fl;
1300
1301         if (file->f_flags & O_DSYNC)
1302                 return 0;
1303         if (!nfs_write_pageuptodate(page, inode))
1304                 return 0;
1305         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1306                 return 1;
1307         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1308                        list_empty_careful(&flctx->flc_posix)))
1309                 return 1;
1310
1311         /* Check to see if there are whole file write locks */
1312         ret = 0;
1313         spin_lock(&flctx->flc_lock);
1314         if (!list_empty(&flctx->flc_posix)) {
1315                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1316                                         fl_list);
1317                 if (is_whole_file_wrlock(fl))
1318                         ret = 1;
1319         } else if (!list_empty(&flctx->flc_flock)) {
1320                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1321                                         fl_list);
1322                 if (fl->fl_type == F_WRLCK)
1323                         ret = 1;
1324         }
1325         spin_unlock(&flctx->flc_lock);
1326         return ret;
1327 }
1328
1329 /*
1330  * Update and possibly write a cached page of an NFS file.
1331  *
1332  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1333  * things with a page scheduled for an RPC call (e.g. invalidate it).
1334  */
1335 int nfs_updatepage(struct file *file, struct page *page,
1336                 unsigned int offset, unsigned int count)
1337 {
1338         struct nfs_open_context *ctx = nfs_file_open_context(file);
1339         struct inode    *inode = page_file_mapping(page)->host;
1340         int             status = 0;
1341
1342         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1343
1344         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1345                 file, count, (long long)(page_file_offset(page) + offset));
1346
1347         if (!count)
1348                 goto out;
1349
1350         if (nfs_can_extend_write(file, page, inode)) {
1351                 count = max(count + offset, nfs_page_length(page));
1352                 offset = 0;
1353         }
1354
1355         status = nfs_writepage_setup(ctx, page, offset, count);
1356         if (status < 0)
1357                 nfs_set_pageerror(page);
1358         else
1359                 __set_page_dirty_nobuffers(page);
1360 out:
1361         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1362                         status, (long long)i_size_read(inode));
1363         return status;
1364 }
1365
1366 static int flush_task_priority(int how)
1367 {
1368         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1369                 case FLUSH_HIGHPRI:
1370                         return RPC_PRIORITY_HIGH;
1371                 case FLUSH_LOWPRI:
1372                         return RPC_PRIORITY_LOW;
1373         }
1374         return RPC_PRIORITY_NORMAL;
1375 }
1376
1377 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1378                                struct rpc_message *msg,
1379                                const struct nfs_rpc_ops *rpc_ops,
1380                                struct rpc_task_setup *task_setup_data, int how)
1381 {
1382         int priority = flush_task_priority(how);
1383
1384         task_setup_data->priority = priority;
1385         rpc_ops->write_setup(hdr, msg);
1386
1387         nfs4_state_protect_write(NFS_SERVER(hdr->inode)->nfs_client,
1388                                  &task_setup_data->rpc_client, msg, hdr);
1389 }
1390
1391 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1392  * call this on each, which will prepare them to be retried on next
1393  * writeback using standard nfs.
1394  */
1395 static void nfs_redirty_request(struct nfs_page *req)
1396 {
1397         nfs_mark_request_dirty(req);
1398         set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1399         nfs_end_page_writeback(req);
1400         nfs_release_request(req);
1401 }
1402
1403 static void nfs_async_write_error(struct list_head *head)
1404 {
1405         struct nfs_page *req;
1406
1407         while (!list_empty(head)) {
1408                 req = nfs_list_entry(head->next);
1409                 nfs_list_remove_request(req);
1410                 nfs_redirty_request(req);
1411         }
1412 }
1413
1414 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1415 {
1416         nfs_async_write_error(&hdr->pages);
1417 }
1418
1419 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1420         .init_hdr = nfs_async_write_init,
1421         .error_cleanup = nfs_async_write_error,
1422         .completion = nfs_write_completion,
1423         .reschedule_io = nfs_async_write_reschedule_io,
1424 };
1425
1426 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1427                                struct inode *inode, int ioflags, bool force_mds,
1428                                const struct nfs_pgio_completion_ops *compl_ops)
1429 {
1430         struct nfs_server *server = NFS_SERVER(inode);
1431         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1432
1433 #ifdef CONFIG_NFS_V4_1
1434         if (server->pnfs_curr_ld && !force_mds)
1435                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1436 #endif
1437         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1438                         server->wsize, ioflags);
1439 }
1440 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1441
1442 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1443 {
1444         struct nfs_pgio_mirror *mirror;
1445
1446         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1447                 pgio->pg_ops->pg_cleanup(pgio);
1448
1449         pgio->pg_ops = &nfs_pgio_rw_ops;
1450
1451         nfs_pageio_stop_mirroring(pgio);
1452
1453         mirror = &pgio->pg_mirrors[0];
1454         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1455 }
1456 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1457
1458
1459 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1460 {
1461         struct nfs_commit_data *data = calldata;
1462
1463         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1464 }
1465
1466 /*
1467  * Special version of should_remove_suid() that ignores capabilities.
1468  */
1469 static int nfs_should_remove_suid(const struct inode *inode)
1470 {
1471         umode_t mode = inode->i_mode;
1472         int kill = 0;
1473
1474         /* suid always must be killed */
1475         if (unlikely(mode & S_ISUID))
1476                 kill = ATTR_KILL_SUID;
1477
1478         /*
1479          * sgid without any exec bits is just a mandatory locking mark; leave
1480          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1481          */
1482         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1483                 kill |= ATTR_KILL_SGID;
1484
1485         if (unlikely(kill && S_ISREG(mode)))
1486                 return kill;
1487
1488         return 0;
1489 }
1490
1491 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1492                 struct nfs_fattr *fattr)
1493 {
1494         struct nfs_pgio_args *argp = &hdr->args;
1495         struct nfs_pgio_res *resp = &hdr->res;
1496         u64 size = argp->offset + resp->count;
1497
1498         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1499                 fattr->size = size;
1500         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1501                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1502                 return;
1503         }
1504         if (size != fattr->size)
1505                 return;
1506         /* Set attribute barrier */
1507         nfs_fattr_set_barrier(fattr);
1508         /* ...and update size */
1509         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1510 }
1511
1512 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1513 {
1514         struct nfs_fattr *fattr = &hdr->fattr;
1515         struct inode *inode = hdr->inode;
1516
1517         spin_lock(&inode->i_lock);
1518         nfs_writeback_check_extend(hdr, fattr);
1519         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1520         spin_unlock(&inode->i_lock);
1521 }
1522 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1523
1524 /*
1525  * This function is called when the WRITE call is complete.
1526  */
1527 static int nfs_writeback_done(struct rpc_task *task,
1528                               struct nfs_pgio_header *hdr,
1529                               struct inode *inode)
1530 {
1531         int status;
1532
1533         /*
1534          * ->write_done will attempt to use post-op attributes to detect
1535          * conflicting writes by other clients.  A strict interpretation
1536          * of close-to-open would allow us to continue caching even if
1537          * another writer had changed the file, but some applications
1538          * depend on tighter cache coherency when writing.
1539          */
1540         status = NFS_PROTO(inode)->write_done(task, hdr);
1541         if (status != 0)
1542                 return status;
1543         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1544
1545         if (hdr->res.verf->committed < hdr->args.stable &&
1546             task->tk_status >= 0) {
1547                 /* We tried a write call, but the server did not
1548                  * commit data to stable storage even though we
1549                  * requested it.
1550                  * Note: There is a known bug in Tru64 < 5.0 in which
1551                  *       the server reports NFS_DATA_SYNC, but performs
1552                  *       NFS_FILE_SYNC. We therefore implement this checking
1553                  *       as a dprintk() in order to avoid filling syslog.
1554                  */
1555                 static unsigned long    complain;
1556
1557                 /* Note this will print the MDS for a DS write */
1558                 if (time_before(complain, jiffies)) {
1559                         dprintk("NFS:       faulty NFS server %s:"
1560                                 " (committed = %d) != (stable = %d)\n",
1561                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1562                                 hdr->res.verf->committed, hdr->args.stable);
1563                         complain = jiffies + 300 * HZ;
1564                 }
1565         }
1566
1567         /* Deal with the suid/sgid bit corner case */
1568         if (nfs_should_remove_suid(inode))
1569                 nfs_mark_for_revalidate(inode);
1570         return 0;
1571 }
1572
1573 /*
1574  * This function is called when the WRITE call is complete.
1575  */
1576 static void nfs_writeback_result(struct rpc_task *task,
1577                                  struct nfs_pgio_header *hdr)
1578 {
1579         struct nfs_pgio_args    *argp = &hdr->args;
1580         struct nfs_pgio_res     *resp = &hdr->res;
1581
1582         if (resp->count < argp->count) {
1583                 static unsigned long    complain;
1584
1585                 /* This a short write! */
1586                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1587
1588                 /* Has the server at least made some progress? */
1589                 if (resp->count == 0) {
1590                         if (time_before(complain, jiffies)) {
1591                                 printk(KERN_WARNING
1592                                        "NFS: Server wrote zero bytes, expected %u.\n",
1593                                        argp->count);
1594                                 complain = jiffies + 300 * HZ;
1595                         }
1596                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1597                         task->tk_status = -EIO;
1598                         return;
1599                 }
1600
1601                 /* For non rpc-based layout drivers, retry-through-MDS */
1602                 if (!task->tk_ops) {
1603                         hdr->pnfs_error = -EAGAIN;
1604                         return;
1605                 }
1606
1607                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1608                 if (resp->verf->committed != NFS_UNSTABLE) {
1609                         /* Resend from where the server left off */
1610                         hdr->mds_offset += resp->count;
1611                         argp->offset += resp->count;
1612                         argp->pgbase += resp->count;
1613                         argp->count -= resp->count;
1614                 } else {
1615                         /* Resend as a stable write in order to avoid
1616                          * headaches in the case of a server crash.
1617                          */
1618                         argp->stable = NFS_FILE_SYNC;
1619                 }
1620                 rpc_restart_call_prepare(task);
1621         }
1622 }
1623
1624 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1625 {
1626         return wait_on_atomic_t(&cinfo->rpcs_out,
1627                         nfs_wait_atomic_killable, TASK_KILLABLE);
1628 }
1629
1630 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1631 {
1632         atomic_inc(&cinfo->rpcs_out);
1633 }
1634
1635 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1636 {
1637         if (atomic_dec_and_test(&cinfo->rpcs_out))
1638                 wake_up_atomic_t(&cinfo->rpcs_out);
1639 }
1640
1641 void nfs_commitdata_release(struct nfs_commit_data *data)
1642 {
1643         put_nfs_open_context(data->context);
1644         nfs_commit_free(data);
1645 }
1646 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1647
1648 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1649                         const struct nfs_rpc_ops *nfs_ops,
1650                         const struct rpc_call_ops *call_ops,
1651                         int how, int flags)
1652 {
1653         struct rpc_task *task;
1654         int priority = flush_task_priority(how);
1655         struct rpc_message msg = {
1656                 .rpc_argp = &data->args,
1657                 .rpc_resp = &data->res,
1658                 .rpc_cred = data->cred,
1659         };
1660         struct rpc_task_setup task_setup_data = {
1661                 .task = &data->task,
1662                 .rpc_client = clnt,
1663                 .rpc_message = &msg,
1664                 .callback_ops = call_ops,
1665                 .callback_data = data,
1666                 .workqueue = nfsiod_workqueue,
1667                 .flags = RPC_TASK_ASYNC | flags,
1668                 .priority = priority,
1669         };
1670         /* Set up the initial task struct.  */
1671         nfs_ops->commit_setup(data, &msg);
1672
1673         dprintk("NFS: initiated commit call\n");
1674
1675         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1676                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1677
1678         task = rpc_run_task(&task_setup_data);
1679         if (IS_ERR(task))
1680                 return PTR_ERR(task);
1681         if (how & FLUSH_SYNC)
1682                 rpc_wait_for_completion_task(task);
1683         rpc_put_task(task);
1684         return 0;
1685 }
1686 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1687
1688 static loff_t nfs_get_lwb(struct list_head *head)
1689 {
1690         loff_t lwb = 0;
1691         struct nfs_page *req;
1692
1693         list_for_each_entry(req, head, wb_list)
1694                 if (lwb < (req_offset(req) + req->wb_bytes))
1695                         lwb = req_offset(req) + req->wb_bytes;
1696
1697         return lwb;
1698 }
1699
1700 /*
1701  * Set up the argument/result storage required for the RPC call.
1702  */
1703 void nfs_init_commit(struct nfs_commit_data *data,
1704                      struct list_head *head,
1705                      struct pnfs_layout_segment *lseg,
1706                      struct nfs_commit_info *cinfo)
1707 {
1708         struct nfs_page *first = nfs_list_entry(head->next);
1709         struct inode *inode = d_inode(first->wb_context->dentry);
1710
1711         /* Set up the RPC argument and reply structs
1712          * NB: take care not to mess about with data->commit et al. */
1713
1714         list_splice_init(head, &data->pages);
1715
1716         data->inode       = inode;
1717         data->cred        = first->wb_context->cred;
1718         data->lseg        = lseg; /* reference transferred */
1719         /* only set lwb for pnfs commit */
1720         if (lseg)
1721                 data->lwb = nfs_get_lwb(&data->pages);
1722         data->mds_ops     = &nfs_commit_ops;
1723         data->completion_ops = cinfo->completion_ops;
1724         data->dreq        = cinfo->dreq;
1725
1726         data->args.fh     = NFS_FH(data->inode);
1727         /* Note: we always request a commit of the entire inode */
1728         data->args.offset = 0;
1729         data->args.count  = 0;
1730         data->context     = get_nfs_open_context(first->wb_context);
1731         data->res.fattr   = &data->fattr;
1732         data->res.verf    = &data->verf;
1733         nfs_fattr_init(&data->fattr);
1734 }
1735 EXPORT_SYMBOL_GPL(nfs_init_commit);
1736
1737 void nfs_retry_commit(struct list_head *page_list,
1738                       struct pnfs_layout_segment *lseg,
1739                       struct nfs_commit_info *cinfo,
1740                       u32 ds_commit_idx)
1741 {
1742         struct nfs_page *req;
1743
1744         while (!list_empty(page_list)) {
1745                 req = nfs_list_entry(page_list->next);
1746                 nfs_list_remove_request(req);
1747                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1748                 if (!cinfo->dreq)
1749                         nfs_clear_page_commit(req->wb_page);
1750                 nfs_unlock_and_release_request(req);
1751         }
1752 }
1753 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1754
1755 static void
1756 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1757                 struct nfs_page *req)
1758 {
1759         __set_page_dirty_nobuffers(req->wb_page);
1760 }
1761
1762 /*
1763  * Commit dirty pages
1764  */
1765 static int
1766 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1767                 struct nfs_commit_info *cinfo)
1768 {
1769         struct nfs_commit_data  *data;
1770
1771         /* another commit raced with us */
1772         if (list_empty(head))
1773                 return 0;
1774
1775         data = nfs_commitdata_alloc(true);
1776
1777         /* Set up the argument struct */
1778         nfs_init_commit(data, head, NULL, cinfo);
1779         atomic_inc(&cinfo->mds->rpcs_out);
1780         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1781                                    data->mds_ops, how, 0);
1782 }
1783
1784 /*
1785  * COMMIT call returned
1786  */
1787 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1788 {
1789         struct nfs_commit_data  *data = calldata;
1790
1791         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1792                                 task->tk_pid, task->tk_status);
1793
1794         /* Call the NFS version-specific code */
1795         NFS_PROTO(data->inode)->commit_done(task, data);
1796 }
1797
1798 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1799 {
1800         struct nfs_page *req;
1801         int status = data->task.tk_status;
1802         struct nfs_commit_info cinfo;
1803         struct nfs_server *nfss;
1804
1805         while (!list_empty(&data->pages)) {
1806                 req = nfs_list_entry(data->pages.next);
1807                 nfs_list_remove_request(req);
1808                 if (req->wb_page)
1809                         nfs_clear_page_commit(req->wb_page);
1810
1811                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1812                         req->wb_context->dentry->d_sb->s_id,
1813                         (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
1814                         req->wb_bytes,
1815                         (long long)req_offset(req));
1816                 if (status < 0) {
1817                         nfs_context_set_write_error(req->wb_context, status);
1818                         if (req->wb_page)
1819                                 nfs_inode_remove_request(req);
1820                         dprintk_cont(", error = %d\n", status);
1821                         goto next;
1822                 }
1823
1824                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1825                  * returned by the server against all stored verfs. */
1826                 if (!nfs_write_verifier_cmp(&req->wb_verf, &data->verf.verifier)) {
1827                         /* We have a match */
1828                         if (req->wb_page)
1829                                 nfs_inode_remove_request(req);
1830                         dprintk_cont(" OK\n");
1831                         goto next;
1832                 }
1833                 /* We have a mismatch. Write the page again */
1834                 dprintk_cont(" mismatch\n");
1835                 nfs_mark_request_dirty(req);
1836                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1837         next:
1838                 nfs_unlock_and_release_request(req);
1839         }
1840         nfss = NFS_SERVER(data->inode);
1841         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1842                 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1843
1844         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1845         nfs_commit_end(cinfo.mds);
1846 }
1847
1848 static void nfs_commit_release(void *calldata)
1849 {
1850         struct nfs_commit_data *data = calldata;
1851
1852         data->completion_ops->completion(data);
1853         nfs_commitdata_release(calldata);
1854 }
1855
1856 static const struct rpc_call_ops nfs_commit_ops = {
1857         .rpc_call_prepare = nfs_commit_prepare,
1858         .rpc_call_done = nfs_commit_done,
1859         .rpc_release = nfs_commit_release,
1860 };
1861
1862 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1863         .completion = nfs_commit_release_pages,
1864         .resched_write = nfs_commit_resched_write,
1865 };
1866
1867 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1868                             int how, struct nfs_commit_info *cinfo)
1869 {
1870         int status;
1871
1872         status = pnfs_commit_list(inode, head, how, cinfo);
1873         if (status == PNFS_NOT_ATTEMPTED)
1874                 status = nfs_commit_list(inode, head, how, cinfo);
1875         return status;
1876 }
1877
1878 int nfs_commit_inode(struct inode *inode, int how)
1879 {
1880         LIST_HEAD(head);
1881         struct nfs_commit_info cinfo;
1882         int may_wait = how & FLUSH_SYNC;
1883         int error = 0;
1884         int res;
1885
1886         nfs_init_cinfo_from_inode(&cinfo, inode);
1887         nfs_commit_begin(cinfo.mds);
1888         res = nfs_scan_commit(inode, &head, &cinfo);
1889         if (res)
1890                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1891         nfs_commit_end(cinfo.mds);
1892         if (error < 0)
1893                 goto out_error;
1894         if (!may_wait)
1895                 goto out_mark_dirty;
1896         error = wait_on_commit(cinfo.mds);
1897         if (error < 0)
1898                 return error;
1899         return res;
1900 out_error:
1901         res = error;
1902         /* Note: If we exit without ensuring that the commit is complete,
1903          * we must mark the inode as dirty. Otherwise, future calls to
1904          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1905          * that the data is on the disk.
1906          */
1907 out_mark_dirty:
1908         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1909         return res;
1910 }
1911 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1912
1913 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1914 {
1915         struct nfs_inode *nfsi = NFS_I(inode);
1916         int flags = FLUSH_SYNC;
1917         int ret = 0;
1918
1919         /* no commits means nothing needs to be done */
1920         if (!atomic_long_read(&nfsi->commit_info.ncommit))
1921                 return ret;
1922
1923         if (wbc->sync_mode == WB_SYNC_NONE) {
1924                 /* Don't commit yet if this is a non-blocking flush and there
1925                  * are a lot of outstanding writes for this mapping.
1926                  */
1927                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1928                         goto out_mark_dirty;
1929
1930                 /* don't wait for the COMMIT response */
1931                 flags = 0;
1932         }
1933
1934         ret = nfs_commit_inode(inode, flags);
1935         if (ret >= 0) {
1936                 if (wbc->sync_mode == WB_SYNC_NONE) {
1937                         if (ret < wbc->nr_to_write)
1938                                 wbc->nr_to_write -= ret;
1939                         else
1940                                 wbc->nr_to_write = 0;
1941                 }
1942                 return 0;
1943         }
1944 out_mark_dirty:
1945         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1946         return ret;
1947 }
1948 EXPORT_SYMBOL_GPL(nfs_write_inode);
1949
1950 /*
1951  * Wrapper for filemap_write_and_wait_range()
1952  *
1953  * Needed for pNFS in order to ensure data becomes visible to the
1954  * client.
1955  */
1956 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1957                 loff_t lstart, loff_t lend)
1958 {
1959         int ret;
1960
1961         ret = filemap_write_and_wait_range(mapping, lstart, lend);
1962         if (ret == 0)
1963                 ret = pnfs_sync_inode(mapping->host, true);
1964         return ret;
1965 }
1966 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
1967
1968 /*
1969  * flush the inode to disk.
1970  */
1971 int nfs_wb_all(struct inode *inode)
1972 {
1973         int ret;
1974
1975         trace_nfs_writeback_inode_enter(inode);
1976
1977         ret = filemap_write_and_wait(inode->i_mapping);
1978         if (ret)
1979                 goto out;
1980         ret = nfs_commit_inode(inode, FLUSH_SYNC);
1981         if (ret < 0)
1982                 goto out;
1983         pnfs_sync_inode(inode, true);
1984         ret = 0;
1985
1986 out:
1987         trace_nfs_writeback_inode_exit(inode, ret);
1988         return ret;
1989 }
1990 EXPORT_SYMBOL_GPL(nfs_wb_all);
1991
1992 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1993 {
1994         struct nfs_page *req;
1995         int ret = 0;
1996
1997         wait_on_page_writeback(page);
1998
1999         /* blocking call to cancel all requests and join to a single (head)
2000          * request */
2001         req = nfs_lock_and_join_requests(page);
2002
2003         if (IS_ERR(req)) {
2004                 ret = PTR_ERR(req);
2005         } else if (req) {
2006                 /* all requests from this page have been cancelled by
2007                  * nfs_lock_and_join_requests, so just remove the head
2008                  * request from the inode / page_private pointer and
2009                  * release it */
2010                 nfs_inode_remove_request(req);
2011                 nfs_unlock_and_release_request(req);
2012         }
2013
2014         return ret;
2015 }
2016
2017 /*
2018  * Write back all requests on one page - we do this before reading it.
2019  */
2020 int nfs_wb_page(struct inode *inode, struct page *page)
2021 {
2022         loff_t range_start = page_file_offset(page);
2023         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2024         struct writeback_control wbc = {
2025                 .sync_mode = WB_SYNC_ALL,
2026                 .nr_to_write = 0,
2027                 .range_start = range_start,
2028                 .range_end = range_end,
2029         };
2030         int ret;
2031
2032         trace_nfs_writeback_page_enter(inode);
2033
2034         for (;;) {
2035                 wait_on_page_writeback(page);
2036                 if (clear_page_dirty_for_io(page)) {
2037                         ret = nfs_writepage_locked(page, &wbc);
2038                         if (ret < 0)
2039                                 goto out_error;
2040                         continue;
2041                 }
2042                 ret = 0;
2043                 if (!PagePrivate(page))
2044                         break;
2045                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2046                 if (ret < 0)
2047                         goto out_error;
2048         }
2049 out_error:
2050         trace_nfs_writeback_page_exit(inode, ret);
2051         return ret;
2052 }
2053
2054 #ifdef CONFIG_MIGRATION
2055 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2056                 struct page *page, enum migrate_mode mode)
2057 {
2058         /*
2059          * If PagePrivate is set, then the page is currently associated with
2060          * an in-progress read or write request. Don't try to migrate it.
2061          *
2062          * FIXME: we could do this in principle, but we'll need a way to ensure
2063          *        that we can safely release the inode reference while holding
2064          *        the page lock.
2065          */
2066         if (PagePrivate(page))
2067                 return -EBUSY;
2068
2069         if (!nfs_fscache_release_page(page, GFP_KERNEL))
2070                 return -EBUSY;
2071
2072         return migrate_page(mapping, newpage, page, mode);
2073 }
2074 #endif
2075
2076 int __init nfs_init_writepagecache(void)
2077 {
2078         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2079                                              sizeof(struct nfs_pgio_header),
2080                                              0, SLAB_HWCACHE_ALIGN,
2081                                              NULL);
2082         if (nfs_wdata_cachep == NULL)
2083                 return -ENOMEM;
2084
2085         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2086                                                      nfs_wdata_cachep);
2087         if (nfs_wdata_mempool == NULL)
2088                 goto out_destroy_write_cache;
2089
2090         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2091                                              sizeof(struct nfs_commit_data),
2092                                              0, SLAB_HWCACHE_ALIGN,
2093                                              NULL);
2094         if (nfs_cdata_cachep == NULL)
2095                 goto out_destroy_write_mempool;
2096
2097         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2098                                                       nfs_cdata_cachep);
2099         if (nfs_commit_mempool == NULL)
2100                 goto out_destroy_commit_cache;
2101
2102         /*
2103          * NFS congestion size, scale with available memory.
2104          *
2105          *  64MB:    8192k
2106          * 128MB:   11585k
2107          * 256MB:   16384k
2108          * 512MB:   23170k
2109          *   1GB:   32768k
2110          *   2GB:   46340k
2111          *   4GB:   65536k
2112          *   8GB:   92681k
2113          *  16GB:  131072k
2114          *
2115          * This allows larger machines to have larger/more transfers.
2116          * Limit the default to 256M
2117          */
2118         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
2119         if (nfs_congestion_kb > 256*1024)
2120                 nfs_congestion_kb = 256*1024;
2121
2122         return 0;
2123
2124 out_destroy_commit_cache:
2125         kmem_cache_destroy(nfs_cdata_cachep);
2126 out_destroy_write_mempool:
2127         mempool_destroy(nfs_wdata_mempool);
2128 out_destroy_write_cache:
2129         kmem_cache_destroy(nfs_wdata_cachep);
2130         return -ENOMEM;
2131 }
2132
2133 void nfs_destroy_writepagecache(void)
2134 {
2135         mempool_destroy(nfs_commit_mempool);
2136         kmem_cache_destroy(nfs_cdata_cachep);
2137         mempool_destroy(nfs_wdata_mempool);
2138         kmem_cache_destroy(nfs_wdata_cachep);
2139 }
2140
2141 static const struct nfs_rw_ops nfs_rw_write_ops = {
2142         .rw_alloc_header        = nfs_writehdr_alloc,
2143         .rw_free_header         = nfs_writehdr_free,
2144         .rw_done                = nfs_writeback_done,
2145         .rw_result              = nfs_writeback_result,
2146         .rw_initiate            = nfs_initiate_write,
2147 };