121218d4e5edf4cc91dcc73d5322b77d9cd0b40e
[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                         continue;
509                 while (!nfs_lock_request(subreq)) {
510                         /*
511                          * Unlock page to allow nfs_page_group_sync_on_bit()
512                          * to succeed
513                          */
514                         nfs_page_group_unlock(head);
515                         ret = nfs_wait_on_request(subreq);
516                         if (!ret)
517                                 ret = nfs_page_group_lock(head);
518                         if (ret < 0) {
519                                 nfs_unroll_locks(inode, head, subreq);
520                                 nfs_release_request(subreq);
521                                 nfs_unlock_and_release_request(head);
522                                 return ERR_PTR(ret);
523                         }
524                 }
525                 /*
526                  * Subrequests are always contiguous, non overlapping
527                  * and in order - but may be repeated (mirrored writes).
528                  */
529                 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
530                         /* keep track of how many bytes this group covers */
531                         total_bytes += subreq->wb_bytes;
532                 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
533                             ((subreq->wb_offset + subreq->wb_bytes) >
534                              (head->wb_offset + total_bytes)))) {
535                         nfs_unroll_locks(inode, head, subreq);
536                         nfs_unlock_and_release_request(subreq);
537                         nfs_page_group_unlock(head);
538                         nfs_unlock_and_release_request(head);
539                         return ERR_PTR(-EIO);
540                 }
541         }
542
543         /* Now that all requests are locked, make sure they aren't on any list.
544          * Commit list removal accounting is done after locks are dropped */
545         subreq = head;
546         do {
547                 nfs_clear_request_commit(subreq);
548                 subreq = subreq->wb_this_page;
549         } while (subreq != head);
550
551         /* unlink subrequests from head, destroy them later */
552         if (head->wb_this_page != head) {
553                 /* destroy list will be terminated by head */
554                 destroy_list = head->wb_this_page;
555                 head->wb_this_page = head;
556
557                 /* change head request to cover whole range that
558                  * the former page group covered */
559                 head->wb_bytes = total_bytes;
560         }
561
562         /* Postpone destruction of this request */
563         if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
564                 set_bit(PG_INODE_REF, &head->wb_flags);
565                 kref_get(&head->wb_kref);
566                 atomic_long_inc(&NFS_I(inode)->nrequests);
567         }
568
569         nfs_page_group_unlock(head);
570
571         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
572
573         /* Did we lose a race with nfs_inode_remove_request()? */
574         if (!(PagePrivate(page) || PageSwapCache(page))) {
575                 nfs_unlock_and_release_request(head);
576                 return NULL;
577         }
578
579         /* still holds ref on head from nfs_page_find_head_request
580          * and still has lock on head from lock loop */
581         return head;
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 = 0;
625         /* If there is a fatal error that covers this write, just exit */
626         if (nfs_error_is_fatal_on_server(req->wb_context->error))
627                 goto out_launder;
628
629         if (!nfs_pageio_add_request(pgio, req)) {
630                 ret = pgio->pg_error;
631                 /*
632                  * Remove the problematic req upon fatal errors on the server
633                  */
634                 if (nfs_error_is_fatal(ret)) {
635                         nfs_context_set_write_error(req->wb_context, ret);
636                         if (nfs_error_is_fatal_on_server(ret))
637                                 goto out_launder;
638                 }
639                 nfs_redirty_request(req);
640                 ret = -EAGAIN;
641         } else
642                 nfs_add_stats(page_file_mapping(page)->host,
643                                 NFSIOS_WRITEPAGES, 1);
644 out:
645         return ret;
646 out_launder:
647         nfs_write_error_remove_page(req);
648         return ret;
649 }
650
651 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
652                             struct nfs_pageio_descriptor *pgio)
653 {
654         int ret;
655
656         nfs_pageio_cond_complete(pgio, page_index(page));
657         ret = nfs_page_async_flush(pgio, page);
658         if (ret == -EAGAIN) {
659                 redirty_page_for_writepage(wbc, page);
660                 ret = 0;
661         }
662         return ret;
663 }
664
665 /*
666  * Write an mmapped page to the server.
667  */
668 static int nfs_writepage_locked(struct page *page,
669                                 struct writeback_control *wbc)
670 {
671         struct nfs_pageio_descriptor pgio;
672         struct inode *inode = page_file_mapping(page)->host;
673         int err;
674
675         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
676         nfs_pageio_init_write(&pgio, inode, 0,
677                                 false, &nfs_async_write_completion_ops);
678         err = nfs_do_writepage(page, wbc, &pgio);
679         nfs_pageio_complete(&pgio);
680         if (err < 0)
681                 return err;
682         if (pgio.pg_error < 0)
683                 return pgio.pg_error;
684         return 0;
685 }
686
687 int nfs_writepage(struct page *page, struct writeback_control *wbc)
688 {
689         int ret;
690
691         ret = nfs_writepage_locked(page, wbc);
692         unlock_page(page);
693         return ret;
694 }
695
696 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
697 {
698         int ret;
699
700         ret = nfs_do_writepage(page, wbc, data);
701         unlock_page(page);
702         return ret;
703 }
704
705 static void nfs_io_completion_commit(void *inode)
706 {
707         nfs_commit_inode(inode, 0);
708 }
709
710 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
711 {
712         struct inode *inode = mapping->host;
713         struct nfs_pageio_descriptor pgio;
714         struct nfs_io_completion *ioc = nfs_io_completion_alloc(GFP_NOFS);
715         int err;
716
717         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
718
719         if (ioc)
720                 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
721
722         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
723                                 &nfs_async_write_completion_ops);
724         pgio.pg_io_completion = ioc;
725         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
726         nfs_pageio_complete(&pgio);
727         nfs_io_completion_put(ioc);
728
729         if (err < 0)
730                 goto out_err;
731         err = pgio.pg_error;
732         if (err < 0)
733                 goto out_err;
734         return 0;
735 out_err:
736         return err;
737 }
738
739 /*
740  * Insert a write request into an inode
741  */
742 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
743 {
744         struct address_space *mapping = page_file_mapping(req->wb_page);
745         struct nfs_inode *nfsi = NFS_I(inode);
746
747         WARN_ON_ONCE(req->wb_this_page != req);
748
749         /* Lock the request! */
750         nfs_lock_request(req);
751
752         /*
753          * Swap-space should not get truncated. Hence no need to plug the race
754          * with invalidate/truncate.
755          */
756         spin_lock(&mapping->private_lock);
757         if (!nfs_have_writebacks(inode) &&
758             NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE)) {
759                 spin_lock(&inode->i_lock);
760                 inode->i_version++;
761                 spin_unlock(&inode->i_lock);
762         }
763         if (likely(!PageSwapCache(req->wb_page))) {
764                 set_bit(PG_MAPPED, &req->wb_flags);
765                 SetPagePrivate(req->wb_page);
766                 set_page_private(req->wb_page, (unsigned long)req);
767         }
768         spin_unlock(&mapping->private_lock);
769         atomic_long_inc(&nfsi->nrequests);
770         /* this a head request for a page group - mark it as having an
771          * extra reference so sub groups can follow suit.
772          * This flag also informs pgio layer when to bump nrequests when
773          * adding subrequests. */
774         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
775         kref_get(&req->wb_kref);
776 }
777
778 /*
779  * Remove a write request from an inode
780  */
781 static void nfs_inode_remove_request(struct nfs_page *req)
782 {
783         struct address_space *mapping = page_file_mapping(req->wb_page);
784         struct inode *inode = mapping->host;
785         struct nfs_inode *nfsi = NFS_I(inode);
786         struct nfs_page *head;
787
788         atomic_long_dec(&nfsi->nrequests);
789         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
790                 head = req->wb_head;
791
792                 spin_lock(&mapping->private_lock);
793                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
794                         set_page_private(head->wb_page, 0);
795                         ClearPagePrivate(head->wb_page);
796                         clear_bit(PG_MAPPED, &head->wb_flags);
797                 }
798                 spin_unlock(&mapping->private_lock);
799         }
800
801         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
802                 nfs_release_request(req);
803 }
804
805 static void
806 nfs_mark_request_dirty(struct nfs_page *req)
807 {
808         if (req->wb_page)
809                 __set_page_dirty_nobuffers(req->wb_page);
810 }
811
812 /*
813  * nfs_page_search_commits_for_head_request_locked
814  *
815  * Search through commit lists on @inode for the head request for @page.
816  * Must be called while holding the inode (which is cinfo) lock.
817  *
818  * Returns the head request if found, or NULL if not found.
819  */
820 static struct nfs_page *
821 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
822                                                 struct page *page)
823 {
824         struct nfs_page *freq, *t;
825         struct nfs_commit_info cinfo;
826         struct inode *inode = &nfsi->vfs_inode;
827
828         nfs_init_cinfo_from_inode(&cinfo, inode);
829
830         /* search through pnfs commit lists */
831         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
832         if (freq)
833                 return freq->wb_head;
834
835         /* Linearly search the commit list for the correct request */
836         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
837                 if (freq->wb_page == page)
838                         return freq->wb_head;
839         }
840
841         return NULL;
842 }
843
844 /**
845  * nfs_request_add_commit_list_locked - add request to a commit list
846  * @req: pointer to a struct nfs_page
847  * @dst: commit list head
848  * @cinfo: holds list lock and accounting info
849  *
850  * This sets the PG_CLEAN bit, updates the cinfo count of
851  * number of outstanding requests requiring a commit as well as
852  * the MM page stats.
853  *
854  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
855  * nfs_page lock.
856  */
857 void
858 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
859                             struct nfs_commit_info *cinfo)
860 {
861         set_bit(PG_CLEAN, &req->wb_flags);
862         nfs_list_add_request(req, dst);
863         atomic_long_inc(&cinfo->mds->ncommit);
864 }
865 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
866
867 /**
868  * nfs_request_add_commit_list - add request to a commit list
869  * @req: pointer to a struct nfs_page
870  * @dst: commit list head
871  * @cinfo: holds list lock and accounting info
872  *
873  * This sets the PG_CLEAN bit, updates the cinfo count of
874  * number of outstanding requests requiring a commit as well as
875  * the MM page stats.
876  *
877  * The caller must _not_ hold the cinfo->lock, but must be
878  * holding the nfs_page lock.
879  */
880 void
881 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
882 {
883         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
884         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
885         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
886         if (req->wb_page)
887                 nfs_mark_page_unstable(req->wb_page, cinfo);
888 }
889 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
890
891 /**
892  * nfs_request_remove_commit_list - Remove request from a commit list
893  * @req: pointer to a nfs_page
894  * @cinfo: holds list lock and accounting info
895  *
896  * This clears the PG_CLEAN bit, and updates the cinfo's count of
897  * number of outstanding requests requiring a commit
898  * It does not update the MM page stats.
899  *
900  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
901  */
902 void
903 nfs_request_remove_commit_list(struct nfs_page *req,
904                                struct nfs_commit_info *cinfo)
905 {
906         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
907                 return;
908         nfs_list_remove_request(req);
909         atomic_long_dec(&cinfo->mds->ncommit);
910 }
911 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
912
913 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
914                                       struct inode *inode)
915 {
916         cinfo->inode = inode;
917         cinfo->mds = &NFS_I(inode)->commit_info;
918         cinfo->ds = pnfs_get_ds_info(inode);
919         cinfo->dreq = NULL;
920         cinfo->completion_ops = &nfs_commit_completion_ops;
921 }
922
923 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
924                     struct inode *inode,
925                     struct nfs_direct_req *dreq)
926 {
927         if (dreq)
928                 nfs_init_cinfo_from_dreq(cinfo, dreq);
929         else
930                 nfs_init_cinfo_from_inode(cinfo, inode);
931 }
932 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
933
934 /*
935  * Add a request to the inode's commit list.
936  */
937 void
938 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
939                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
940 {
941         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
942                 return;
943         nfs_request_add_commit_list(req, cinfo);
944 }
945
946 static void
947 nfs_clear_page_commit(struct page *page)
948 {
949         dec_node_page_state(page, NR_UNSTABLE_NFS);
950         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
951                     WB_RECLAIMABLE);
952 }
953
954 /* Called holding the request lock on @req */
955 static void
956 nfs_clear_request_commit(struct nfs_page *req)
957 {
958         if (test_bit(PG_CLEAN, &req->wb_flags)) {
959                 struct inode *inode = d_inode(req->wb_context->dentry);
960                 struct nfs_commit_info cinfo;
961
962                 nfs_init_cinfo_from_inode(&cinfo, inode);
963                 mutex_lock(&NFS_I(inode)->commit_mutex);
964                 if (!pnfs_clear_request_commit(req, &cinfo)) {
965                         nfs_request_remove_commit_list(req, &cinfo);
966                 }
967                 mutex_unlock(&NFS_I(inode)->commit_mutex);
968                 nfs_clear_page_commit(req->wb_page);
969         }
970 }
971
972 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
973 {
974         if (hdr->verf.committed == NFS_DATA_SYNC)
975                 return hdr->lseg == NULL;
976         return hdr->verf.committed != NFS_FILE_SYNC;
977 }
978
979 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
980 {
981         nfs_io_completion_get(hdr->io_completion);
982 }
983
984 static void nfs_write_completion(struct nfs_pgio_header *hdr)
985 {
986         struct nfs_commit_info cinfo;
987         unsigned long bytes = 0;
988
989         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
990                 goto out;
991         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
992         while (!list_empty(&hdr->pages)) {
993                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
994
995                 bytes += req->wb_bytes;
996                 nfs_list_remove_request(req);
997                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
998                     (hdr->good_bytes < bytes)) {
999                         nfs_set_pageerror(req->wb_page);
1000                         nfs_context_set_write_error(req->wb_context, hdr->error);
1001                         goto remove_req;
1002                 }
1003                 if (nfs_write_need_commit(hdr)) {
1004                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1005                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1006                                 hdr->pgio_mirror_idx);
1007                         goto next;
1008                 }
1009 remove_req:
1010                 nfs_inode_remove_request(req);
1011 next:
1012                 nfs_end_page_writeback(req);
1013                 nfs_release_request(req);
1014         }
1015 out:
1016         nfs_io_completion_put(hdr->io_completion);
1017         hdr->release(hdr);
1018 }
1019
1020 unsigned long
1021 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1022 {
1023         return atomic_long_read(&cinfo->mds->ncommit);
1024 }
1025
1026 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1027 int
1028 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1029                      struct nfs_commit_info *cinfo, int max)
1030 {
1031         struct nfs_page *req, *tmp;
1032         int ret = 0;
1033
1034 restart:
1035         list_for_each_entry_safe(req, tmp, src, wb_list) {
1036                 kref_get(&req->wb_kref);
1037                 if (!nfs_lock_request(req)) {
1038                         int status;
1039
1040                         /* Prevent deadlock with nfs_lock_and_join_requests */
1041                         if (!list_empty(dst)) {
1042                                 nfs_release_request(req);
1043                                 continue;
1044                         }
1045                         /* Ensure we make progress to prevent livelock */
1046                         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1047                         status = nfs_wait_on_request(req);
1048                         nfs_release_request(req);
1049                         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1050                         if (status < 0)
1051                                 break;
1052                         goto restart;
1053                 }
1054                 nfs_request_remove_commit_list(req, cinfo);
1055                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1056                 nfs_list_add_request(req, dst);
1057                 ret++;
1058                 if ((ret == max) && !cinfo->dreq)
1059                         break;
1060                 cond_resched();
1061         }
1062         return ret;
1063 }
1064 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1065
1066 /*
1067  * nfs_scan_commit - Scan an inode for commit requests
1068  * @inode: NFS inode to scan
1069  * @dst: mds destination list
1070  * @cinfo: mds and ds lists of reqs ready to commit
1071  *
1072  * Moves requests from the inode's 'commit' request list.
1073  * The requests are *not* checked to ensure that they form a contiguous set.
1074  */
1075 int
1076 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1077                 struct nfs_commit_info *cinfo)
1078 {
1079         int ret = 0;
1080
1081         if (!atomic_long_read(&cinfo->mds->ncommit))
1082                 return 0;
1083         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1084         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1085                 const int max = INT_MAX;
1086
1087                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1088                                            cinfo, max);
1089                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1090         }
1091         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1092         return ret;
1093 }
1094
1095 /*
1096  * Search for an existing write request, and attempt to update
1097  * it to reflect a new dirty region on a given page.
1098  *
1099  * If the attempt fails, then the existing request is flushed out
1100  * to disk.
1101  */
1102 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1103                 struct page *page,
1104                 unsigned int offset,
1105                 unsigned int bytes)
1106 {
1107         struct nfs_page *req;
1108         unsigned int rqend;
1109         unsigned int end;
1110         int error;
1111
1112         end = offset + bytes;
1113
1114         req = nfs_lock_and_join_requests(page);
1115         if (IS_ERR_OR_NULL(req))
1116                 return req;
1117
1118         rqend = req->wb_offset + req->wb_bytes;
1119         /*
1120          * Tell the caller to flush out the request if
1121          * the offsets are non-contiguous.
1122          * Note: nfs_flush_incompatible() will already
1123          * have flushed out requests having wrong owners.
1124          */
1125         if (offset > rqend || end < req->wb_offset)
1126                 goto out_flushme;
1127
1128         /* Okay, the request matches. Update the region */
1129         if (offset < req->wb_offset) {
1130                 req->wb_offset = offset;
1131                 req->wb_pgbase = offset;
1132         }
1133         if (end > rqend)
1134                 req->wb_bytes = end - req->wb_offset;
1135         else
1136                 req->wb_bytes = rqend - req->wb_offset;
1137         return req;
1138 out_flushme:
1139         /*
1140          * Note: we mark the request dirty here because
1141          * nfs_lock_and_join_requests() cannot preserve
1142          * commit flags, so we have to replay the write.
1143          */
1144         nfs_mark_request_dirty(req);
1145         nfs_unlock_and_release_request(req);
1146         error = nfs_wb_page(inode, page);
1147         return (error < 0) ? ERR_PTR(error) : NULL;
1148 }
1149
1150 /*
1151  * Try to update an existing write request, or create one if there is none.
1152  *
1153  * Note: Should always be called with the Page Lock held to prevent races
1154  * if we have to add a new request. Also assumes that the caller has
1155  * already called nfs_flush_incompatible() if necessary.
1156  */
1157 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1158                 struct page *page, unsigned int offset, unsigned int bytes)
1159 {
1160         struct inode *inode = page_file_mapping(page)->host;
1161         struct nfs_page *req;
1162
1163         req = nfs_try_to_update_request(inode, page, offset, bytes);
1164         if (req != NULL)
1165                 goto out;
1166         req = nfs_create_request(ctx, page, NULL, offset, bytes);
1167         if (IS_ERR(req))
1168                 goto out;
1169         nfs_inode_add_request(inode, req);
1170 out:
1171         return req;
1172 }
1173
1174 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1175                 unsigned int offset, unsigned int count)
1176 {
1177         struct nfs_page *req;
1178
1179         req = nfs_setup_write_request(ctx, page, offset, count);
1180         if (IS_ERR(req))
1181                 return PTR_ERR(req);
1182         /* Update file length */
1183         nfs_grow_file(page, offset, count);
1184         nfs_mark_uptodate(req);
1185         nfs_mark_request_dirty(req);
1186         nfs_unlock_and_release_request(req);
1187         return 0;
1188 }
1189
1190 int nfs_flush_incompatible(struct file *file, struct page *page)
1191 {
1192         struct nfs_open_context *ctx = nfs_file_open_context(file);
1193         struct nfs_lock_context *l_ctx;
1194         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1195         struct nfs_page *req;
1196         int do_flush, status;
1197         /*
1198          * Look for a request corresponding to this page. If there
1199          * is one, and it belongs to another file, we flush it out
1200          * before we try to copy anything into the page. Do this
1201          * due to the lack of an ACCESS-type call in NFSv2.
1202          * Also do the same if we find a request from an existing
1203          * dropped page.
1204          */
1205         do {
1206                 req = nfs_page_find_head_request(page);
1207                 if (req == NULL)
1208                         return 0;
1209                 l_ctx = req->wb_lock_context;
1210                 do_flush = req->wb_page != page ||
1211                         !nfs_match_open_context(req->wb_context, ctx);
1212                 if (l_ctx && flctx &&
1213                     !(list_empty_careful(&flctx->flc_posix) &&
1214                       list_empty_careful(&flctx->flc_flock))) {
1215                         do_flush |= l_ctx->lockowner != current->files;
1216                 }
1217                 nfs_release_request(req);
1218                 if (!do_flush)
1219                         return 0;
1220                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1221         } while (status == 0);
1222         return status;
1223 }
1224
1225 /*
1226  * Avoid buffered writes when a open context credential's key would
1227  * expire soon.
1228  *
1229  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1230  *
1231  * Return 0 and set a credential flag which triggers the inode to flush
1232  * and performs  NFS_FILE_SYNC writes if the key will expired within
1233  * RPC_KEY_EXPIRE_TIMEO.
1234  */
1235 int
1236 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1237 {
1238         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1239         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1240
1241         return rpcauth_key_timeout_notify(auth, ctx->cred);
1242 }
1243
1244 /*
1245  * Test if the open context credential key is marked to expire soon.
1246  */
1247 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1248 {
1249         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1250
1251         return rpcauth_cred_key_to_expire(auth, ctx->cred);
1252 }
1253
1254 /*
1255  * If the page cache is marked as unsafe or invalid, then we can't rely on
1256  * the PageUptodate() flag. In this case, we will need to turn off
1257  * write optimisations that depend on the page contents being correct.
1258  */
1259 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1260 {
1261         struct nfs_inode *nfsi = NFS_I(inode);
1262
1263         if (nfs_have_delegated_attributes(inode))
1264                 goto out;
1265         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1266                 return false;
1267         smp_rmb();
1268         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1269                 return false;
1270 out:
1271         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1272                 return false;
1273         return PageUptodate(page) != 0;
1274 }
1275
1276 static bool
1277 is_whole_file_wrlock(struct file_lock *fl)
1278 {
1279         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1280                         fl->fl_type == F_WRLCK;
1281 }
1282
1283 /* If we know the page is up to date, and we're not using byte range locks (or
1284  * if we have the whole file locked for writing), it may be more efficient to
1285  * extend the write to cover the entire page in order to avoid fragmentation
1286  * inefficiencies.
1287  *
1288  * If the file is opened for synchronous writes then we can just skip the rest
1289  * of the checks.
1290  */
1291 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1292 {
1293         int ret;
1294         struct file_lock_context *flctx = inode->i_flctx;
1295         struct file_lock *fl;
1296
1297         if (file->f_flags & O_DSYNC)
1298                 return 0;
1299         if (!nfs_write_pageuptodate(page, inode))
1300                 return 0;
1301         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1302                 return 1;
1303         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1304                        list_empty_careful(&flctx->flc_posix)))
1305                 return 1;
1306
1307         /* Check to see if there are whole file write locks */
1308         ret = 0;
1309         spin_lock(&flctx->flc_lock);
1310         if (!list_empty(&flctx->flc_posix)) {
1311                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1312                                         fl_list);
1313                 if (is_whole_file_wrlock(fl))
1314                         ret = 1;
1315         } else if (!list_empty(&flctx->flc_flock)) {
1316                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1317                                         fl_list);
1318                 if (fl->fl_type == F_WRLCK)
1319                         ret = 1;
1320         }
1321         spin_unlock(&flctx->flc_lock);
1322         return ret;
1323 }
1324
1325 /*
1326  * Update and possibly write a cached page of an NFS file.
1327  *
1328  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1329  * things with a page scheduled for an RPC call (e.g. invalidate it).
1330  */
1331 int nfs_updatepage(struct file *file, struct page *page,
1332                 unsigned int offset, unsigned int count)
1333 {
1334         struct nfs_open_context *ctx = nfs_file_open_context(file);
1335         struct inode    *inode = page_file_mapping(page)->host;
1336         int             status = 0;
1337
1338         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1339
1340         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1341                 file, count, (long long)(page_file_offset(page) + offset));
1342
1343         if (!count)
1344                 goto out;
1345
1346         if (nfs_can_extend_write(file, page, inode)) {
1347                 count = max(count + offset, nfs_page_length(page));
1348                 offset = 0;
1349         }
1350
1351         status = nfs_writepage_setup(ctx, page, offset, count);
1352         if (status < 0)
1353                 nfs_set_pageerror(page);
1354         else
1355                 __set_page_dirty_nobuffers(page);
1356 out:
1357         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1358                         status, (long long)i_size_read(inode));
1359         return status;
1360 }
1361
1362 static int flush_task_priority(int how)
1363 {
1364         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1365                 case FLUSH_HIGHPRI:
1366                         return RPC_PRIORITY_HIGH;
1367                 case FLUSH_LOWPRI:
1368                         return RPC_PRIORITY_LOW;
1369         }
1370         return RPC_PRIORITY_NORMAL;
1371 }
1372
1373 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1374                                struct rpc_message *msg,
1375                                const struct nfs_rpc_ops *rpc_ops,
1376                                struct rpc_task_setup *task_setup_data, int how)
1377 {
1378         int priority = flush_task_priority(how);
1379
1380         task_setup_data->priority = priority;
1381         rpc_ops->write_setup(hdr, msg);
1382
1383         nfs4_state_protect_write(NFS_SERVER(hdr->inode)->nfs_client,
1384                                  &task_setup_data->rpc_client, msg, hdr);
1385 }
1386
1387 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1388  * call this on each, which will prepare them to be retried on next
1389  * writeback using standard nfs.
1390  */
1391 static void nfs_redirty_request(struct nfs_page *req)
1392 {
1393         nfs_mark_request_dirty(req);
1394         set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1395         nfs_end_page_writeback(req);
1396         nfs_release_request(req);
1397 }
1398
1399 static void nfs_async_write_error(struct list_head *head)
1400 {
1401         struct nfs_page *req;
1402
1403         while (!list_empty(head)) {
1404                 req = nfs_list_entry(head->next);
1405                 nfs_list_remove_request(req);
1406                 nfs_redirty_request(req);
1407         }
1408 }
1409
1410 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1411 {
1412         nfs_async_write_error(&hdr->pages);
1413 }
1414
1415 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1416         .init_hdr = nfs_async_write_init,
1417         .error_cleanup = nfs_async_write_error,
1418         .completion = nfs_write_completion,
1419         .reschedule_io = nfs_async_write_reschedule_io,
1420 };
1421
1422 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1423                                struct inode *inode, int ioflags, bool force_mds,
1424                                const struct nfs_pgio_completion_ops *compl_ops)
1425 {
1426         struct nfs_server *server = NFS_SERVER(inode);
1427         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1428
1429 #ifdef CONFIG_NFS_V4_1
1430         if (server->pnfs_curr_ld && !force_mds)
1431                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1432 #endif
1433         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1434                         server->wsize, ioflags);
1435 }
1436 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1437
1438 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1439 {
1440         struct nfs_pgio_mirror *mirror;
1441
1442         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1443                 pgio->pg_ops->pg_cleanup(pgio);
1444
1445         pgio->pg_ops = &nfs_pgio_rw_ops;
1446
1447         nfs_pageio_stop_mirroring(pgio);
1448
1449         mirror = &pgio->pg_mirrors[0];
1450         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1451 }
1452 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1453
1454
1455 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1456 {
1457         struct nfs_commit_data *data = calldata;
1458
1459         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1460 }
1461
1462 /*
1463  * Special version of should_remove_suid() that ignores capabilities.
1464  */
1465 static int nfs_should_remove_suid(const struct inode *inode)
1466 {
1467         umode_t mode = inode->i_mode;
1468         int kill = 0;
1469
1470         /* suid always must be killed */
1471         if (unlikely(mode & S_ISUID))
1472                 kill = ATTR_KILL_SUID;
1473
1474         /*
1475          * sgid without any exec bits is just a mandatory locking mark; leave
1476          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1477          */
1478         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1479                 kill |= ATTR_KILL_SGID;
1480
1481         if (unlikely(kill && S_ISREG(mode)))
1482                 return kill;
1483
1484         return 0;
1485 }
1486
1487 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1488                 struct nfs_fattr *fattr)
1489 {
1490         struct nfs_pgio_args *argp = &hdr->args;
1491         struct nfs_pgio_res *resp = &hdr->res;
1492         u64 size = argp->offset + resp->count;
1493
1494         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1495                 fattr->size = size;
1496         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1497                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1498                 return;
1499         }
1500         if (size != fattr->size)
1501                 return;
1502         /* Set attribute barrier */
1503         nfs_fattr_set_barrier(fattr);
1504         /* ...and update size */
1505         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1506 }
1507
1508 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1509 {
1510         struct nfs_fattr *fattr = &hdr->fattr;
1511         struct inode *inode = hdr->inode;
1512
1513         spin_lock(&inode->i_lock);
1514         nfs_writeback_check_extend(hdr, fattr);
1515         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1516         spin_unlock(&inode->i_lock);
1517 }
1518 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1519
1520 /*
1521  * This function is called when the WRITE call is complete.
1522  */
1523 static int nfs_writeback_done(struct rpc_task *task,
1524                               struct nfs_pgio_header *hdr,
1525                               struct inode *inode)
1526 {
1527         int status;
1528
1529         /*
1530          * ->write_done will attempt to use post-op attributes to detect
1531          * conflicting writes by other clients.  A strict interpretation
1532          * of close-to-open would allow us to continue caching even if
1533          * another writer had changed the file, but some applications
1534          * depend on tighter cache coherency when writing.
1535          */
1536         status = NFS_PROTO(inode)->write_done(task, hdr);
1537         if (status != 0)
1538                 return status;
1539         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1540
1541         if (hdr->res.verf->committed < hdr->args.stable &&
1542             task->tk_status >= 0) {
1543                 /* We tried a write call, but the server did not
1544                  * commit data to stable storage even though we
1545                  * requested it.
1546                  * Note: There is a known bug in Tru64 < 5.0 in which
1547                  *       the server reports NFS_DATA_SYNC, but performs
1548                  *       NFS_FILE_SYNC. We therefore implement this checking
1549                  *       as a dprintk() in order to avoid filling syslog.
1550                  */
1551                 static unsigned long    complain;
1552
1553                 /* Note this will print the MDS for a DS write */
1554                 if (time_before(complain, jiffies)) {
1555                         dprintk("NFS:       faulty NFS server %s:"
1556                                 " (committed = %d) != (stable = %d)\n",
1557                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1558                                 hdr->res.verf->committed, hdr->args.stable);
1559                         complain = jiffies + 300 * HZ;
1560                 }
1561         }
1562
1563         /* Deal with the suid/sgid bit corner case */
1564         if (nfs_should_remove_suid(inode))
1565                 nfs_mark_for_revalidate(inode);
1566         return 0;
1567 }
1568
1569 /*
1570  * This function is called when the WRITE call is complete.
1571  */
1572 static void nfs_writeback_result(struct rpc_task *task,
1573                                  struct nfs_pgio_header *hdr)
1574 {
1575         struct nfs_pgio_args    *argp = &hdr->args;
1576         struct nfs_pgio_res     *resp = &hdr->res;
1577
1578         if (resp->count < argp->count) {
1579                 static unsigned long    complain;
1580
1581                 /* This a short write! */
1582                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1583
1584                 /* Has the server at least made some progress? */
1585                 if (resp->count == 0) {
1586                         if (time_before(complain, jiffies)) {
1587                                 printk(KERN_WARNING
1588                                        "NFS: Server wrote zero bytes, expected %u.\n",
1589                                        argp->count);
1590                                 complain = jiffies + 300 * HZ;
1591                         }
1592                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1593                         task->tk_status = -EIO;
1594                         return;
1595                 }
1596
1597                 /* For non rpc-based layout drivers, retry-through-MDS */
1598                 if (!task->tk_ops) {
1599                         hdr->pnfs_error = -EAGAIN;
1600                         return;
1601                 }
1602
1603                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1604                 if (resp->verf->committed != NFS_UNSTABLE) {
1605                         /* Resend from where the server left off */
1606                         hdr->mds_offset += resp->count;
1607                         argp->offset += resp->count;
1608                         argp->pgbase += resp->count;
1609                         argp->count -= resp->count;
1610                 } else {
1611                         /* Resend as a stable write in order to avoid
1612                          * headaches in the case of a server crash.
1613                          */
1614                         argp->stable = NFS_FILE_SYNC;
1615                 }
1616                 rpc_restart_call_prepare(task);
1617         }
1618 }
1619
1620 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1621 {
1622         return wait_on_atomic_t(&cinfo->rpcs_out,
1623                         nfs_wait_atomic_killable, TASK_KILLABLE);
1624 }
1625
1626 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1627 {
1628         atomic_inc(&cinfo->rpcs_out);
1629 }
1630
1631 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1632 {
1633         if (atomic_dec_and_test(&cinfo->rpcs_out))
1634                 wake_up_atomic_t(&cinfo->rpcs_out);
1635 }
1636
1637 void nfs_commitdata_release(struct nfs_commit_data *data)
1638 {
1639         put_nfs_open_context(data->context);
1640         nfs_commit_free(data);
1641 }
1642 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1643
1644 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1645                         const struct nfs_rpc_ops *nfs_ops,
1646                         const struct rpc_call_ops *call_ops,
1647                         int how, int flags)
1648 {
1649         struct rpc_task *task;
1650         int priority = flush_task_priority(how);
1651         struct rpc_message msg = {
1652                 .rpc_argp = &data->args,
1653                 .rpc_resp = &data->res,
1654                 .rpc_cred = data->cred,
1655         };
1656         struct rpc_task_setup task_setup_data = {
1657                 .task = &data->task,
1658                 .rpc_client = clnt,
1659                 .rpc_message = &msg,
1660                 .callback_ops = call_ops,
1661                 .callback_data = data,
1662                 .workqueue = nfsiod_workqueue,
1663                 .flags = RPC_TASK_ASYNC | flags,
1664                 .priority = priority,
1665         };
1666         /* Set up the initial task struct.  */
1667         nfs_ops->commit_setup(data, &msg);
1668
1669         dprintk("NFS: initiated commit call\n");
1670
1671         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1672                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1673
1674         task = rpc_run_task(&task_setup_data);
1675         if (IS_ERR(task))
1676                 return PTR_ERR(task);
1677         if (how & FLUSH_SYNC)
1678                 rpc_wait_for_completion_task(task);
1679         rpc_put_task(task);
1680         return 0;
1681 }
1682 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1683
1684 static loff_t nfs_get_lwb(struct list_head *head)
1685 {
1686         loff_t lwb = 0;
1687         struct nfs_page *req;
1688
1689         list_for_each_entry(req, head, wb_list)
1690                 if (lwb < (req_offset(req) + req->wb_bytes))
1691                         lwb = req_offset(req) + req->wb_bytes;
1692
1693         return lwb;
1694 }
1695
1696 /*
1697  * Set up the argument/result storage required for the RPC call.
1698  */
1699 void nfs_init_commit(struct nfs_commit_data *data,
1700                      struct list_head *head,
1701                      struct pnfs_layout_segment *lseg,
1702                      struct nfs_commit_info *cinfo)
1703 {
1704         struct nfs_page *first = nfs_list_entry(head->next);
1705         struct inode *inode = d_inode(first->wb_context->dentry);
1706
1707         /* Set up the RPC argument and reply structs
1708          * NB: take care not to mess about with data->commit et al. */
1709
1710         list_splice_init(head, &data->pages);
1711
1712         data->inode       = inode;
1713         data->cred        = first->wb_context->cred;
1714         data->lseg        = lseg; /* reference transferred */
1715         /* only set lwb for pnfs commit */
1716         if (lseg)
1717                 data->lwb = nfs_get_lwb(&data->pages);
1718         data->mds_ops     = &nfs_commit_ops;
1719         data->completion_ops = cinfo->completion_ops;
1720         data->dreq        = cinfo->dreq;
1721
1722         data->args.fh     = NFS_FH(data->inode);
1723         /* Note: we always request a commit of the entire inode */
1724         data->args.offset = 0;
1725         data->args.count  = 0;
1726         data->context     = get_nfs_open_context(first->wb_context);
1727         data->res.fattr   = &data->fattr;
1728         data->res.verf    = &data->verf;
1729         nfs_fattr_init(&data->fattr);
1730 }
1731 EXPORT_SYMBOL_GPL(nfs_init_commit);
1732
1733 void nfs_retry_commit(struct list_head *page_list,
1734                       struct pnfs_layout_segment *lseg,
1735                       struct nfs_commit_info *cinfo,
1736                       u32 ds_commit_idx)
1737 {
1738         struct nfs_page *req;
1739
1740         while (!list_empty(page_list)) {
1741                 req = nfs_list_entry(page_list->next);
1742                 nfs_list_remove_request(req);
1743                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1744                 if (!cinfo->dreq)
1745                         nfs_clear_page_commit(req->wb_page);
1746                 nfs_unlock_and_release_request(req);
1747         }
1748 }
1749 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1750
1751 static void
1752 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1753                 struct nfs_page *req)
1754 {
1755         __set_page_dirty_nobuffers(req->wb_page);
1756 }
1757
1758 /*
1759  * Commit dirty pages
1760  */
1761 static int
1762 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1763                 struct nfs_commit_info *cinfo)
1764 {
1765         struct nfs_commit_data  *data;
1766
1767         /* another commit raced with us */
1768         if (list_empty(head))
1769                 return 0;
1770
1771         data = nfs_commitdata_alloc(true);
1772
1773         /* Set up the argument struct */
1774         nfs_init_commit(data, head, NULL, cinfo);
1775         atomic_inc(&cinfo->mds->rpcs_out);
1776         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1777                                    data->mds_ops, how, 0);
1778 }
1779
1780 /*
1781  * COMMIT call returned
1782  */
1783 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1784 {
1785         struct nfs_commit_data  *data = calldata;
1786
1787         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1788                                 task->tk_pid, task->tk_status);
1789
1790         /* Call the NFS version-specific code */
1791         NFS_PROTO(data->inode)->commit_done(task, data);
1792 }
1793
1794 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1795 {
1796         struct nfs_page *req;
1797         int status = data->task.tk_status;
1798         struct nfs_commit_info cinfo;
1799         struct nfs_server *nfss;
1800
1801         while (!list_empty(&data->pages)) {
1802                 req = nfs_list_entry(data->pages.next);
1803                 nfs_list_remove_request(req);
1804                 if (req->wb_page)
1805                         nfs_clear_page_commit(req->wb_page);
1806
1807                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1808                         req->wb_context->dentry->d_sb->s_id,
1809                         (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
1810                         req->wb_bytes,
1811                         (long long)req_offset(req));
1812                 if (status < 0) {
1813                         nfs_context_set_write_error(req->wb_context, status);
1814                         if (req->wb_page)
1815                                 nfs_inode_remove_request(req);
1816                         dprintk_cont(", error = %d\n", status);
1817                         goto next;
1818                 }
1819
1820                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1821                  * returned by the server against all stored verfs. */
1822                 if (!nfs_write_verifier_cmp(&req->wb_verf, &data->verf.verifier)) {
1823                         /* We have a match */
1824                         if (req->wb_page)
1825                                 nfs_inode_remove_request(req);
1826                         dprintk_cont(" OK\n");
1827                         goto next;
1828                 }
1829                 /* We have a mismatch. Write the page again */
1830                 dprintk_cont(" mismatch\n");
1831                 nfs_mark_request_dirty(req);
1832                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1833         next:
1834                 nfs_unlock_and_release_request(req);
1835         }
1836         nfss = NFS_SERVER(data->inode);
1837         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1838                 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1839
1840         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1841         nfs_commit_end(cinfo.mds);
1842 }
1843
1844 static void nfs_commit_release(void *calldata)
1845 {
1846         struct nfs_commit_data *data = calldata;
1847
1848         data->completion_ops->completion(data);
1849         nfs_commitdata_release(calldata);
1850 }
1851
1852 static const struct rpc_call_ops nfs_commit_ops = {
1853         .rpc_call_prepare = nfs_commit_prepare,
1854         .rpc_call_done = nfs_commit_done,
1855         .rpc_release = nfs_commit_release,
1856 };
1857
1858 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1859         .completion = nfs_commit_release_pages,
1860         .resched_write = nfs_commit_resched_write,
1861 };
1862
1863 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1864                             int how, struct nfs_commit_info *cinfo)
1865 {
1866         int status;
1867
1868         status = pnfs_commit_list(inode, head, how, cinfo);
1869         if (status == PNFS_NOT_ATTEMPTED)
1870                 status = nfs_commit_list(inode, head, how, cinfo);
1871         return status;
1872 }
1873
1874 int nfs_commit_inode(struct inode *inode, int how)
1875 {
1876         LIST_HEAD(head);
1877         struct nfs_commit_info cinfo;
1878         int may_wait = how & FLUSH_SYNC;
1879         int error = 0;
1880         int res;
1881
1882         nfs_init_cinfo_from_inode(&cinfo, inode);
1883         nfs_commit_begin(cinfo.mds);
1884         res = nfs_scan_commit(inode, &head, &cinfo);
1885         if (res)
1886                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1887         nfs_commit_end(cinfo.mds);
1888         if (error < 0)
1889                 goto out_error;
1890         if (!may_wait)
1891                 goto out_mark_dirty;
1892         error = wait_on_commit(cinfo.mds);
1893         if (error < 0)
1894                 return error;
1895         return res;
1896 out_error:
1897         res = error;
1898         /* Note: If we exit without ensuring that the commit is complete,
1899          * we must mark the inode as dirty. Otherwise, future calls to
1900          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1901          * that the data is on the disk.
1902          */
1903 out_mark_dirty:
1904         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1905         return res;
1906 }
1907 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1908
1909 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1910 {
1911         struct nfs_inode *nfsi = NFS_I(inode);
1912         int flags = FLUSH_SYNC;
1913         int ret = 0;
1914
1915         /* no commits means nothing needs to be done */
1916         if (!atomic_long_read(&nfsi->commit_info.ncommit))
1917                 return ret;
1918
1919         if (wbc->sync_mode == WB_SYNC_NONE) {
1920                 /* Don't commit yet if this is a non-blocking flush and there
1921                  * are a lot of outstanding writes for this mapping.
1922                  */
1923                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1924                         goto out_mark_dirty;
1925
1926                 /* don't wait for the COMMIT response */
1927                 flags = 0;
1928         }
1929
1930         ret = nfs_commit_inode(inode, flags);
1931         if (ret >= 0) {
1932                 if (wbc->sync_mode == WB_SYNC_NONE) {
1933                         if (ret < wbc->nr_to_write)
1934                                 wbc->nr_to_write -= ret;
1935                         else
1936                                 wbc->nr_to_write = 0;
1937                 }
1938                 return 0;
1939         }
1940 out_mark_dirty:
1941         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1942         return ret;
1943 }
1944 EXPORT_SYMBOL_GPL(nfs_write_inode);
1945
1946 /*
1947  * Wrapper for filemap_write_and_wait_range()
1948  *
1949  * Needed for pNFS in order to ensure data becomes visible to the
1950  * client.
1951  */
1952 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1953                 loff_t lstart, loff_t lend)
1954 {
1955         int ret;
1956
1957         ret = filemap_write_and_wait_range(mapping, lstart, lend);
1958         if (ret == 0)
1959                 ret = pnfs_sync_inode(mapping->host, true);
1960         return ret;
1961 }
1962 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
1963
1964 /*
1965  * flush the inode to disk.
1966  */
1967 int nfs_wb_all(struct inode *inode)
1968 {
1969         int ret;
1970
1971         trace_nfs_writeback_inode_enter(inode);
1972
1973         ret = filemap_write_and_wait(inode->i_mapping);
1974         if (ret)
1975                 goto out;
1976         ret = nfs_commit_inode(inode, FLUSH_SYNC);
1977         if (ret < 0)
1978                 goto out;
1979         pnfs_sync_inode(inode, true);
1980         ret = 0;
1981
1982 out:
1983         trace_nfs_writeback_inode_exit(inode, ret);
1984         return ret;
1985 }
1986 EXPORT_SYMBOL_GPL(nfs_wb_all);
1987
1988 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1989 {
1990         struct nfs_page *req;
1991         int ret = 0;
1992
1993         wait_on_page_writeback(page);
1994
1995         /* blocking call to cancel all requests and join to a single (head)
1996          * request */
1997         req = nfs_lock_and_join_requests(page);
1998
1999         if (IS_ERR(req)) {
2000                 ret = PTR_ERR(req);
2001         } else if (req) {
2002                 /* all requests from this page have been cancelled by
2003                  * nfs_lock_and_join_requests, so just remove the head
2004                  * request from the inode / page_private pointer and
2005                  * release it */
2006                 nfs_inode_remove_request(req);
2007                 nfs_unlock_and_release_request(req);
2008         }
2009
2010         return ret;
2011 }
2012
2013 /*
2014  * Write back all requests on one page - we do this before reading it.
2015  */
2016 int nfs_wb_page(struct inode *inode, struct page *page)
2017 {
2018         loff_t range_start = page_file_offset(page);
2019         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2020         struct writeback_control wbc = {
2021                 .sync_mode = WB_SYNC_ALL,
2022                 .nr_to_write = 0,
2023                 .range_start = range_start,
2024                 .range_end = range_end,
2025         };
2026         int ret;
2027
2028         trace_nfs_writeback_page_enter(inode);
2029
2030         for (;;) {
2031                 wait_on_page_writeback(page);
2032                 if (clear_page_dirty_for_io(page)) {
2033                         ret = nfs_writepage_locked(page, &wbc);
2034                         if (ret < 0)
2035                                 goto out_error;
2036                         continue;
2037                 }
2038                 ret = 0;
2039                 if (!PagePrivate(page))
2040                         break;
2041                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2042                 if (ret < 0)
2043                         goto out_error;
2044         }
2045 out_error:
2046         trace_nfs_writeback_page_exit(inode, ret);
2047         return ret;
2048 }
2049
2050 #ifdef CONFIG_MIGRATION
2051 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2052                 struct page *page, enum migrate_mode mode)
2053 {
2054         /*
2055          * If PagePrivate is set, then the page is currently associated with
2056          * an in-progress read or write request. Don't try to migrate it.
2057          *
2058          * FIXME: we could do this in principle, but we'll need a way to ensure
2059          *        that we can safely release the inode reference while holding
2060          *        the page lock.
2061          */
2062         if (PagePrivate(page))
2063                 return -EBUSY;
2064
2065         if (!nfs_fscache_release_page(page, GFP_KERNEL))
2066                 return -EBUSY;
2067
2068         return migrate_page(mapping, newpage, page, mode);
2069 }
2070 #endif
2071
2072 int __init nfs_init_writepagecache(void)
2073 {
2074         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2075                                              sizeof(struct nfs_pgio_header),
2076                                              0, SLAB_HWCACHE_ALIGN,
2077                                              NULL);
2078         if (nfs_wdata_cachep == NULL)
2079                 return -ENOMEM;
2080
2081         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2082                                                      nfs_wdata_cachep);
2083         if (nfs_wdata_mempool == NULL)
2084                 goto out_destroy_write_cache;
2085
2086         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2087                                              sizeof(struct nfs_commit_data),
2088                                              0, SLAB_HWCACHE_ALIGN,
2089                                              NULL);
2090         if (nfs_cdata_cachep == NULL)
2091                 goto out_destroy_write_mempool;
2092
2093         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2094                                                       nfs_cdata_cachep);
2095         if (nfs_commit_mempool == NULL)
2096                 goto out_destroy_commit_cache;
2097
2098         /*
2099          * NFS congestion size, scale with available memory.
2100          *
2101          *  64MB:    8192k
2102          * 128MB:   11585k
2103          * 256MB:   16384k
2104          * 512MB:   23170k
2105          *   1GB:   32768k
2106          *   2GB:   46340k
2107          *   4GB:   65536k
2108          *   8GB:   92681k
2109          *  16GB:  131072k
2110          *
2111          * This allows larger machines to have larger/more transfers.
2112          * Limit the default to 256M
2113          */
2114         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
2115         if (nfs_congestion_kb > 256*1024)
2116                 nfs_congestion_kb = 256*1024;
2117
2118         return 0;
2119
2120 out_destroy_commit_cache:
2121         kmem_cache_destroy(nfs_cdata_cachep);
2122 out_destroy_write_mempool:
2123         mempool_destroy(nfs_wdata_mempool);
2124 out_destroy_write_cache:
2125         kmem_cache_destroy(nfs_wdata_cachep);
2126         return -ENOMEM;
2127 }
2128
2129 void nfs_destroy_writepagecache(void)
2130 {
2131         mempool_destroy(nfs_commit_mempool);
2132         kmem_cache_destroy(nfs_cdata_cachep);
2133         mempool_destroy(nfs_wdata_mempool);
2134         kmem_cache_destroy(nfs_wdata_cachep);
2135 }
2136
2137 static const struct nfs_rw_ops nfs_rw_write_ops = {
2138         .rw_alloc_header        = nfs_writehdr_alloc,
2139         .rw_free_header         = nfs_writehdr_free,
2140         .rw_done                = nfs_writeback_done,
2141         .rw_result              = nfs_writeback_result,
2142         .rw_initiate            = nfs_initiate_write,
2143 };