Merge tag 'nfs-for-4.14-2' of git://git.linux-nfs.org/projects/trondmy/linux-nfs
[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 struct nfs_page *
149 nfs_page_private_request(struct page *page)
150 {
151         if (!PagePrivate(page))
152                 return NULL;
153         return (struct nfs_page *)page_private(page);
154 }
155
156 /*
157  * nfs_page_find_head_request_locked - find head request associated with @page
158  *
159  * must be called while holding the inode lock.
160  *
161  * returns matching head request with reference held, or NULL if not found.
162  */
163 static struct nfs_page *
164 nfs_page_find_private_request(struct page *page)
165 {
166         struct address_space *mapping = page_file_mapping(page);
167         struct nfs_page *req;
168
169         if (!PagePrivate(page))
170                 return NULL;
171         spin_lock(&mapping->private_lock);
172         req = nfs_page_private_request(page);
173         if (req) {
174                 WARN_ON_ONCE(req->wb_head != req);
175                 kref_get(&req->wb_kref);
176         }
177         spin_unlock(&mapping->private_lock);
178         return req;
179 }
180
181 static struct nfs_page *
182 nfs_page_find_swap_request(struct page *page)
183 {
184         struct inode *inode = page_file_mapping(page)->host;
185         struct nfs_inode *nfsi = NFS_I(inode);
186         struct nfs_page *req = NULL;
187         if (!PageSwapCache(page))
188                 return NULL;
189         mutex_lock(&nfsi->commit_mutex);
190         if (PageSwapCache(page)) {
191                 req = nfs_page_search_commits_for_head_request_locked(nfsi,
192                         page);
193                 if (req) {
194                         WARN_ON_ONCE(req->wb_head != req);
195                         kref_get(&req->wb_kref);
196                 }
197         }
198         mutex_unlock(&nfsi->commit_mutex);
199         return req;
200 }
201
202 /*
203  * nfs_page_find_head_request - find head request associated with @page
204  *
205  * returns matching head request with reference held, or NULL if not found.
206  */
207 static struct nfs_page *nfs_page_find_head_request(struct page *page)
208 {
209         struct nfs_page *req;
210
211         req = nfs_page_find_private_request(page);
212         if (!req)
213                 req = nfs_page_find_swap_request(page);
214         return req;
215 }
216
217 /* Adjust the file length if we're writing beyond the end */
218 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
219 {
220         struct inode *inode = page_file_mapping(page)->host;
221         loff_t end, i_size;
222         pgoff_t end_index;
223
224         spin_lock(&inode->i_lock);
225         i_size = i_size_read(inode);
226         end_index = (i_size - 1) >> PAGE_SHIFT;
227         if (i_size > 0 && page_index(page) < end_index)
228                 goto out;
229         end = page_file_offset(page) + ((loff_t)offset+count);
230         if (i_size >= end)
231                 goto out;
232         i_size_write(inode, end);
233         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
234 out:
235         spin_unlock(&inode->i_lock);
236 }
237
238 /* A writeback failed: mark the page as bad, and invalidate the page cache */
239 static void nfs_set_pageerror(struct page *page)
240 {
241         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
242 }
243
244 /*
245  * nfs_page_group_search_locked
246  * @head - head request of page group
247  * @page_offset - offset into page
248  *
249  * Search page group with head @head to find a request that contains the
250  * page offset @page_offset.
251  *
252  * Returns a pointer to the first matching nfs request, or NULL if no
253  * match is found.
254  *
255  * Must be called with the page group lock held
256  */
257 static struct nfs_page *
258 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
259 {
260         struct nfs_page *req;
261
262         req = head;
263         do {
264                 if (page_offset >= req->wb_pgbase &&
265                     page_offset < (req->wb_pgbase + req->wb_bytes))
266                         return req;
267
268                 req = req->wb_this_page;
269         } while (req != head);
270
271         return NULL;
272 }
273
274 /*
275  * nfs_page_group_covers_page
276  * @head - head request of page group
277  *
278  * Return true if the page group with head @head covers the whole page,
279  * returns false otherwise
280  */
281 static bool nfs_page_group_covers_page(struct nfs_page *req)
282 {
283         struct nfs_page *tmp;
284         unsigned int pos = 0;
285         unsigned int len = nfs_page_length(req->wb_page);
286
287         nfs_page_group_lock(req);
288
289         for (;;) {
290                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
291                 if (!tmp)
292                         break;
293                 pos = tmp->wb_pgbase + tmp->wb_bytes;
294         }
295
296         nfs_page_group_unlock(req);
297         return pos >= len;
298 }
299
300 /* We can set the PG_uptodate flag if we see that a write request
301  * covers the full page.
302  */
303 static void nfs_mark_uptodate(struct nfs_page *req)
304 {
305         if (PageUptodate(req->wb_page))
306                 return;
307         if (!nfs_page_group_covers_page(req))
308                 return;
309         SetPageUptodate(req->wb_page);
310 }
311
312 static int wb_priority(struct writeback_control *wbc)
313 {
314         int ret = 0;
315
316         if (wbc->sync_mode == WB_SYNC_ALL)
317                 ret = FLUSH_COND_STABLE;
318         return ret;
319 }
320
321 /*
322  * NFS congestion control
323  */
324
325 int nfs_congestion_kb;
326
327 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
328 #define NFS_CONGESTION_OFF_THRESH       \
329         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
330
331 static void nfs_set_page_writeback(struct page *page)
332 {
333         struct inode *inode = page_file_mapping(page)->host;
334         struct nfs_server *nfss = NFS_SERVER(inode);
335         int ret = test_set_page_writeback(page);
336
337         WARN_ON_ONCE(ret != 0);
338
339         if (atomic_long_inc_return(&nfss->writeback) >
340                         NFS_CONGESTION_ON_THRESH)
341                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
342 }
343
344 static void nfs_end_page_writeback(struct nfs_page *req)
345 {
346         struct inode *inode = page_file_mapping(req->wb_page)->host;
347         struct nfs_server *nfss = NFS_SERVER(inode);
348         bool is_done;
349
350         is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
351         nfs_unlock_request(req);
352         if (!is_done)
353                 return;
354
355         end_page_writeback(req->wb_page);
356         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
357                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
358 }
359
360 /*
361  * nfs_unroll_locks_and_wait -  unlock all newly locked reqs and wait on @req
362  *
363  * this is a helper function for nfs_lock_and_join_requests
364  *
365  * @inode - inode associated with request page group, must be holding inode lock
366  * @head  - head request of page group, must be holding head lock
367  * @req   - request that couldn't lock and needs to wait on the req bit lock
368  *
369  * NOTE: this must be called holding page_group bit lock
370  *       which will be released before returning.
371  *
372  * returns 0 on success, < 0 on error.
373  */
374 static void
375 nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
376                           struct nfs_page *req)
377 {
378         struct nfs_page *tmp;
379
380         /* relinquish all the locks successfully grabbed this run */
381         for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
382                 if (!kref_read(&tmp->wb_kref))
383                         continue;
384                 nfs_unlock_and_release_request(tmp);
385         }
386 }
387
388 /*
389  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
390  *
391  * @destroy_list - request list (using wb_this_page) terminated by @old_head
392  * @old_head - the old head of the list
393  *
394  * All subrequests must be locked and removed from all lists, so at this point
395  * they are only "active" in this function, and possibly in nfs_wait_on_request
396  * with a reference held by some other context.
397  */
398 static void
399 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
400                                  struct nfs_page *old_head,
401                                  struct inode *inode)
402 {
403         while (destroy_list) {
404                 struct nfs_page *subreq = destroy_list;
405
406                 destroy_list = (subreq->wb_this_page == old_head) ?
407                                    NULL : subreq->wb_this_page;
408
409                 WARN_ON_ONCE(old_head != subreq->wb_head);
410
411                 /* make sure old group is not used */
412                 subreq->wb_this_page = subreq;
413
414                 clear_bit(PG_REMOVE, &subreq->wb_flags);
415
416                 /* Note: races with nfs_page_group_destroy() */
417                 if (!kref_read(&subreq->wb_kref)) {
418                         /* Check if we raced with nfs_page_group_destroy() */
419                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags))
420                                 nfs_free_request(subreq);
421                         continue;
422                 }
423
424                 subreq->wb_head = subreq;
425
426                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
427                         nfs_release_request(subreq);
428                         atomic_long_dec(&NFS_I(inode)->nrequests);
429                 }
430
431                 /* subreq is now totally disconnected from page group or any
432                  * write / commit lists. last chance to wake any waiters */
433                 nfs_unlock_and_release_request(subreq);
434         }
435 }
436
437 /*
438  * nfs_lock_and_join_requests - join all subreqs to the head req and return
439  *                              a locked reference, cancelling any pending
440  *                              operations for this page.
441  *
442  * @page - the page used to lookup the "page group" of nfs_page structures
443  *
444  * This function joins all sub requests to the head request by first
445  * locking all requests in the group, cancelling any pending operations
446  * and finally updating the head request to cover the whole range covered by
447  * the (former) group.  All subrequests are removed from any write or commit
448  * lists, unlinked from the group and destroyed.
449  *
450  * Returns a locked, referenced pointer to the head request - which after
451  * this call is guaranteed to be the only request associated with the page.
452  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
453  * error was encountered.
454  */
455 static struct nfs_page *
456 nfs_lock_and_join_requests(struct page *page)
457 {
458         struct inode *inode = page_file_mapping(page)->host;
459         struct nfs_page *head, *subreq;
460         struct nfs_page *destroy_list = NULL;
461         unsigned int total_bytes;
462         int ret;
463
464 try_again:
465         /*
466          * A reference is taken only on the head request which acts as a
467          * reference to the whole page group - the group will not be destroyed
468          * until the head reference is released.
469          */
470         head = nfs_page_find_head_request(page);
471         if (!head)
472                 return NULL;
473
474         /* lock the page head first in order to avoid an ABBA inefficiency */
475         if (!nfs_lock_request(head)) {
476                 ret = nfs_wait_on_request(head);
477                 nfs_release_request(head);
478                 if (ret < 0)
479                         return ERR_PTR(ret);
480                 goto try_again;
481         }
482
483         /* Ensure that nobody removed the request before we locked it */
484         if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
485                 nfs_unlock_and_release_request(head);
486                 goto try_again;
487         }
488
489         ret = nfs_page_group_lock(head);
490         if (ret < 0) {
491                 nfs_unlock_and_release_request(head);
492                 return ERR_PTR(ret);
493         }
494
495         /* lock each request in the page group */
496         total_bytes = head->wb_bytes;
497         for (subreq = head->wb_this_page; subreq != head;
498                         subreq = subreq->wb_this_page) {
499
500                 if (!kref_get_unless_zero(&subreq->wb_kref)) {
501                         if (subreq->wb_offset == head->wb_offset + total_bytes)
502                                 total_bytes += subreq->wb_bytes;
503                         continue;
504                 }
505
506                 while (!nfs_lock_request(subreq)) {
507                         /*
508                          * Unlock page to allow nfs_page_group_sync_on_bit()
509                          * to succeed
510                          */
511                         nfs_page_group_unlock(head);
512                         ret = nfs_wait_on_request(subreq);
513                         if (!ret)
514                                 ret = nfs_page_group_lock(head);
515                         if (ret < 0) {
516                                 nfs_unroll_locks(inode, head, subreq);
517                                 nfs_release_request(subreq);
518                                 nfs_unlock_and_release_request(head);
519                                 return ERR_PTR(ret);
520                         }
521                 }
522                 /*
523                  * Subrequests are always contiguous, non overlapping
524                  * and in order - but may be repeated (mirrored writes).
525                  */
526                 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
527                         /* keep track of how many bytes this group covers */
528                         total_bytes += subreq->wb_bytes;
529                 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
530                             ((subreq->wb_offset + subreq->wb_bytes) >
531                              (head->wb_offset + total_bytes)))) {
532                         nfs_page_group_unlock(head);
533                         nfs_unroll_locks(inode, head, subreq);
534                         nfs_unlock_and_release_request(subreq);
535                         nfs_unlock_and_release_request(head);
536                         return ERR_PTR(-EIO);
537                 }
538         }
539
540         /* Now that all requests are locked, make sure they aren't on any list.
541          * Commit list removal accounting is done after locks are dropped */
542         subreq = head;
543         do {
544                 nfs_clear_request_commit(subreq);
545                 subreq = subreq->wb_this_page;
546         } while (subreq != head);
547
548         /* unlink subrequests from head, destroy them later */
549         if (head->wb_this_page != head) {
550                 /* destroy list will be terminated by head */
551                 destroy_list = head->wb_this_page;
552                 head->wb_this_page = head;
553
554                 /* change head request to cover whole range that
555                  * the former page group covered */
556                 head->wb_bytes = total_bytes;
557         }
558
559         /* Postpone destruction of this request */
560         if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
561                 set_bit(PG_INODE_REF, &head->wb_flags);
562                 kref_get(&head->wb_kref);
563                 atomic_long_inc(&NFS_I(inode)->nrequests);
564         }
565
566         nfs_page_group_unlock(head);
567
568         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
569
570         /* Did we lose a race with nfs_inode_remove_request()? */
571         if (!(PagePrivate(page) || PageSwapCache(page))) {
572                 nfs_unlock_and_release_request(head);
573                 return NULL;
574         }
575
576         /* still holds ref on head from nfs_page_find_head_request
577          * and still has lock on head from lock loop */
578         return head;
579 }
580
581 static void nfs_write_error_remove_page(struct nfs_page *req)
582 {
583         nfs_end_page_writeback(req);
584         generic_error_remove_page(page_file_mapping(req->wb_page),
585                                   req->wb_page);
586         nfs_release_request(req);
587 }
588
589 static bool
590 nfs_error_is_fatal_on_server(int err)
591 {
592         switch (err) {
593         case 0:
594         case -ERESTARTSYS:
595         case -EINTR:
596                 return false;
597         }
598         return nfs_error_is_fatal(err);
599 }
600
601 /*
602  * Find an associated nfs write request, and prepare to flush it out
603  * May return an error if the user signalled nfs_wait_on_request().
604  */
605 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
606                                 struct page *page)
607 {
608         struct nfs_page *req;
609         int ret = 0;
610
611         req = nfs_lock_and_join_requests(page);
612         if (!req)
613                 goto out;
614         ret = PTR_ERR(req);
615         if (IS_ERR(req))
616                 goto out;
617
618         nfs_set_page_writeback(page);
619         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
620
621         ret = 0;
622         /* If there is a fatal error that covers this write, just exit */
623         if (nfs_error_is_fatal_on_server(req->wb_context->error))
624                 goto out_launder;
625
626         if (!nfs_pageio_add_request(pgio, req)) {
627                 ret = pgio->pg_error;
628                 /*
629                  * Remove the problematic req upon fatal errors on the server
630                  */
631                 if (nfs_error_is_fatal(ret)) {
632                         nfs_context_set_write_error(req->wb_context, ret);
633                         if (nfs_error_is_fatal_on_server(ret))
634                                 goto out_launder;
635                 }
636                 nfs_redirty_request(req);
637                 ret = -EAGAIN;
638         } else
639                 nfs_add_stats(page_file_mapping(page)->host,
640                                 NFSIOS_WRITEPAGES, 1);
641 out:
642         return ret;
643 out_launder:
644         nfs_write_error_remove_page(req);
645         return ret;
646 }
647
648 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
649                             struct nfs_pageio_descriptor *pgio)
650 {
651         int ret;
652
653         nfs_pageio_cond_complete(pgio, page_index(page));
654         ret = nfs_page_async_flush(pgio, page);
655         if (ret == -EAGAIN) {
656                 redirty_page_for_writepage(wbc, page);
657                 ret = 0;
658         }
659         return ret;
660 }
661
662 /*
663  * Write an mmapped page to the server.
664  */
665 static int nfs_writepage_locked(struct page *page,
666                                 struct writeback_control *wbc)
667 {
668         struct nfs_pageio_descriptor pgio;
669         struct inode *inode = page_file_mapping(page)->host;
670         int err;
671
672         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
673         nfs_pageio_init_write(&pgio, inode, 0,
674                                 false, &nfs_async_write_completion_ops);
675         err = nfs_do_writepage(page, wbc, &pgio);
676         nfs_pageio_complete(&pgio);
677         if (err < 0)
678                 return err;
679         if (pgio.pg_error < 0)
680                 return pgio.pg_error;
681         return 0;
682 }
683
684 int nfs_writepage(struct page *page, struct writeback_control *wbc)
685 {
686         int ret;
687
688         ret = nfs_writepage_locked(page, wbc);
689         unlock_page(page);
690         return ret;
691 }
692
693 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
694 {
695         int ret;
696
697         ret = nfs_do_writepage(page, wbc, data);
698         unlock_page(page);
699         return ret;
700 }
701
702 static void nfs_io_completion_commit(void *inode)
703 {
704         nfs_commit_inode(inode, 0);
705 }
706
707 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
708 {
709         struct inode *inode = mapping->host;
710         struct nfs_pageio_descriptor pgio;
711         struct nfs_io_completion *ioc = nfs_io_completion_alloc(GFP_NOFS);
712         int err;
713
714         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
715
716         if (ioc)
717                 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
718
719         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
720                                 &nfs_async_write_completion_ops);
721         pgio.pg_io_completion = ioc;
722         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
723         nfs_pageio_complete(&pgio);
724         nfs_io_completion_put(ioc);
725
726         if (err < 0)
727                 goto out_err;
728         err = pgio.pg_error;
729         if (err < 0)
730                 goto out_err;
731         return 0;
732 out_err:
733         return err;
734 }
735
736 /*
737  * Insert a write request into an inode
738  */
739 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
740 {
741         struct address_space *mapping = page_file_mapping(req->wb_page);
742         struct nfs_inode *nfsi = NFS_I(inode);
743
744         WARN_ON_ONCE(req->wb_this_page != req);
745
746         /* Lock the request! */
747         nfs_lock_request(req);
748
749         /*
750          * Swap-space should not get truncated. Hence no need to plug the race
751          * with invalidate/truncate.
752          */
753         spin_lock(&mapping->private_lock);
754         if (!nfs_have_writebacks(inode) &&
755             NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE)) {
756                 spin_lock(&inode->i_lock);
757                 inode->i_version++;
758                 spin_unlock(&inode->i_lock);
759         }
760         if (likely(!PageSwapCache(req->wb_page))) {
761                 set_bit(PG_MAPPED, &req->wb_flags);
762                 SetPagePrivate(req->wb_page);
763                 set_page_private(req->wb_page, (unsigned long)req);
764         }
765         spin_unlock(&mapping->private_lock);
766         atomic_long_inc(&nfsi->nrequests);
767         /* this a head request for a page group - mark it as having an
768          * extra reference so sub groups can follow suit.
769          * This flag also informs pgio layer when to bump nrequests when
770          * adding subrequests. */
771         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
772         kref_get(&req->wb_kref);
773 }
774
775 /*
776  * Remove a write request from an inode
777  */
778 static void nfs_inode_remove_request(struct nfs_page *req)
779 {
780         struct address_space *mapping = page_file_mapping(req->wb_page);
781         struct inode *inode = mapping->host;
782         struct nfs_inode *nfsi = NFS_I(inode);
783         struct nfs_page *head;
784
785         atomic_long_dec(&nfsi->nrequests);
786         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
787                 head = req->wb_head;
788
789                 spin_lock(&mapping->private_lock);
790                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
791                         set_page_private(head->wb_page, 0);
792                         ClearPagePrivate(head->wb_page);
793                         clear_bit(PG_MAPPED, &head->wb_flags);
794                 }
795                 spin_unlock(&mapping->private_lock);
796         }
797
798         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
799                 nfs_release_request(req);
800 }
801
802 static void
803 nfs_mark_request_dirty(struct nfs_page *req)
804 {
805         if (req->wb_page)
806                 __set_page_dirty_nobuffers(req->wb_page);
807 }
808
809 /*
810  * nfs_page_search_commits_for_head_request_locked
811  *
812  * Search through commit lists on @inode for the head request for @page.
813  * Must be called while holding the inode (which is cinfo) lock.
814  *
815  * Returns the head request if found, or NULL if not found.
816  */
817 static struct nfs_page *
818 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
819                                                 struct page *page)
820 {
821         struct nfs_page *freq, *t;
822         struct nfs_commit_info cinfo;
823         struct inode *inode = &nfsi->vfs_inode;
824
825         nfs_init_cinfo_from_inode(&cinfo, inode);
826
827         /* search through pnfs commit lists */
828         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
829         if (freq)
830                 return freq->wb_head;
831
832         /* Linearly search the commit list for the correct request */
833         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
834                 if (freq->wb_page == page)
835                         return freq->wb_head;
836         }
837
838         return NULL;
839 }
840
841 /**
842  * nfs_request_add_commit_list_locked - add request to a commit list
843  * @req: pointer to a struct nfs_page
844  * @dst: commit list head
845  * @cinfo: holds list lock and accounting info
846  *
847  * This sets the PG_CLEAN bit, updates the cinfo count of
848  * number of outstanding requests requiring a commit as well as
849  * the MM page stats.
850  *
851  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
852  * nfs_page lock.
853  */
854 void
855 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
856                             struct nfs_commit_info *cinfo)
857 {
858         set_bit(PG_CLEAN, &req->wb_flags);
859         nfs_list_add_request(req, dst);
860         atomic_long_inc(&cinfo->mds->ncommit);
861 }
862 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
863
864 /**
865  * nfs_request_add_commit_list - add request to a commit list
866  * @req: pointer to a struct nfs_page
867  * @dst: commit list head
868  * @cinfo: holds list lock and accounting info
869  *
870  * This sets the PG_CLEAN bit, updates the cinfo count of
871  * number of outstanding requests requiring a commit as well as
872  * the MM page stats.
873  *
874  * The caller must _not_ hold the cinfo->lock, but must be
875  * holding the nfs_page lock.
876  */
877 void
878 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
879 {
880         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
881         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
882         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
883         if (req->wb_page)
884                 nfs_mark_page_unstable(req->wb_page, cinfo);
885 }
886 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
887
888 /**
889  * nfs_request_remove_commit_list - Remove request from a commit list
890  * @req: pointer to a nfs_page
891  * @cinfo: holds list lock and accounting info
892  *
893  * This clears the PG_CLEAN bit, and updates the cinfo's count of
894  * number of outstanding requests requiring a commit
895  * It does not update the MM page stats.
896  *
897  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
898  */
899 void
900 nfs_request_remove_commit_list(struct nfs_page *req,
901                                struct nfs_commit_info *cinfo)
902 {
903         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
904                 return;
905         nfs_list_remove_request(req);
906         atomic_long_dec(&cinfo->mds->ncommit);
907 }
908 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
909
910 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
911                                       struct inode *inode)
912 {
913         cinfo->inode = inode;
914         cinfo->mds = &NFS_I(inode)->commit_info;
915         cinfo->ds = pnfs_get_ds_info(inode);
916         cinfo->dreq = NULL;
917         cinfo->completion_ops = &nfs_commit_completion_ops;
918 }
919
920 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
921                     struct inode *inode,
922                     struct nfs_direct_req *dreq)
923 {
924         if (dreq)
925                 nfs_init_cinfo_from_dreq(cinfo, dreq);
926         else
927                 nfs_init_cinfo_from_inode(cinfo, inode);
928 }
929 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
930
931 /*
932  * Add a request to the inode's commit list.
933  */
934 void
935 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
936                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
937 {
938         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
939                 return;
940         nfs_request_add_commit_list(req, cinfo);
941 }
942
943 static void
944 nfs_clear_page_commit(struct page *page)
945 {
946         dec_node_page_state(page, NR_UNSTABLE_NFS);
947         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
948                     WB_RECLAIMABLE);
949 }
950
951 /* Called holding the request lock on @req */
952 static void
953 nfs_clear_request_commit(struct nfs_page *req)
954 {
955         if (test_bit(PG_CLEAN, &req->wb_flags)) {
956                 struct inode *inode = d_inode(req->wb_context->dentry);
957                 struct nfs_commit_info cinfo;
958
959                 nfs_init_cinfo_from_inode(&cinfo, inode);
960                 mutex_lock(&NFS_I(inode)->commit_mutex);
961                 if (!pnfs_clear_request_commit(req, &cinfo)) {
962                         nfs_request_remove_commit_list(req, &cinfo);
963                 }
964                 mutex_unlock(&NFS_I(inode)->commit_mutex);
965                 nfs_clear_page_commit(req->wb_page);
966         }
967 }
968
969 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
970 {
971         if (hdr->verf.committed == NFS_DATA_SYNC)
972                 return hdr->lseg == NULL;
973         return hdr->verf.committed != NFS_FILE_SYNC;
974 }
975
976 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
977 {
978         nfs_io_completion_get(hdr->io_completion);
979 }
980
981 static void nfs_write_completion(struct nfs_pgio_header *hdr)
982 {
983         struct nfs_commit_info cinfo;
984         unsigned long bytes = 0;
985
986         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
987                 goto out;
988         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
989         while (!list_empty(&hdr->pages)) {
990                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
991
992                 bytes += req->wb_bytes;
993                 nfs_list_remove_request(req);
994                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
995                     (hdr->good_bytes < bytes)) {
996                         nfs_set_pageerror(req->wb_page);
997                         nfs_context_set_write_error(req->wb_context, hdr->error);
998                         goto remove_req;
999                 }
1000                 if (nfs_write_need_commit(hdr)) {
1001                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1002                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1003                                 hdr->pgio_mirror_idx);
1004                         goto next;
1005                 }
1006 remove_req:
1007                 nfs_inode_remove_request(req);
1008 next:
1009                 nfs_end_page_writeback(req);
1010                 nfs_release_request(req);
1011         }
1012 out:
1013         nfs_io_completion_put(hdr->io_completion);
1014         hdr->release(hdr);
1015 }
1016
1017 unsigned long
1018 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1019 {
1020         return atomic_long_read(&cinfo->mds->ncommit);
1021 }
1022
1023 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1024 int
1025 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1026                      struct nfs_commit_info *cinfo, int max)
1027 {
1028         struct nfs_page *req, *tmp;
1029         int ret = 0;
1030
1031 restart:
1032         list_for_each_entry_safe(req, tmp, src, wb_list) {
1033                 kref_get(&req->wb_kref);
1034                 if (!nfs_lock_request(req)) {
1035                         int status;
1036
1037                         /* Prevent deadlock with nfs_lock_and_join_requests */
1038                         if (!list_empty(dst)) {
1039                                 nfs_release_request(req);
1040                                 continue;
1041                         }
1042                         /* Ensure we make progress to prevent livelock */
1043                         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1044                         status = nfs_wait_on_request(req);
1045                         nfs_release_request(req);
1046                         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1047                         if (status < 0)
1048                                 break;
1049                         goto restart;
1050                 }
1051                 nfs_request_remove_commit_list(req, cinfo);
1052                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1053                 nfs_list_add_request(req, dst);
1054                 ret++;
1055                 if ((ret == max) && !cinfo->dreq)
1056                         break;
1057                 cond_resched();
1058         }
1059         return ret;
1060 }
1061 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1062
1063 /*
1064  * nfs_scan_commit - Scan an inode for commit requests
1065  * @inode: NFS inode to scan
1066  * @dst: mds destination list
1067  * @cinfo: mds and ds lists of reqs ready to commit
1068  *
1069  * Moves requests from the inode's 'commit' request list.
1070  * The requests are *not* checked to ensure that they form a contiguous set.
1071  */
1072 int
1073 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1074                 struct nfs_commit_info *cinfo)
1075 {
1076         int ret = 0;
1077
1078         if (!atomic_long_read(&cinfo->mds->ncommit))
1079                 return 0;
1080         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1081         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1082                 const int max = INT_MAX;
1083
1084                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1085                                            cinfo, max);
1086                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1087         }
1088         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1089         return ret;
1090 }
1091
1092 /*
1093  * Search for an existing write request, and attempt to update
1094  * it to reflect a new dirty region on a given page.
1095  *
1096  * If the attempt fails, then the existing request is flushed out
1097  * to disk.
1098  */
1099 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1100                 struct page *page,
1101                 unsigned int offset,
1102                 unsigned int bytes)
1103 {
1104         struct nfs_page *req;
1105         unsigned int rqend;
1106         unsigned int end;
1107         int error;
1108
1109         end = offset + bytes;
1110
1111         req = nfs_lock_and_join_requests(page);
1112         if (IS_ERR_OR_NULL(req))
1113                 return req;
1114
1115         rqend = req->wb_offset + req->wb_bytes;
1116         /*
1117          * Tell the caller to flush out the request if
1118          * the offsets are non-contiguous.
1119          * Note: nfs_flush_incompatible() will already
1120          * have flushed out requests having wrong owners.
1121          */
1122         if (offset > rqend || end < req->wb_offset)
1123                 goto out_flushme;
1124
1125         /* Okay, the request matches. Update the region */
1126         if (offset < req->wb_offset) {
1127                 req->wb_offset = offset;
1128                 req->wb_pgbase = offset;
1129         }
1130         if (end > rqend)
1131                 req->wb_bytes = end - req->wb_offset;
1132         else
1133                 req->wb_bytes = rqend - req->wb_offset;
1134         return req;
1135 out_flushme:
1136         /*
1137          * Note: we mark the request dirty here because
1138          * nfs_lock_and_join_requests() cannot preserve
1139          * commit flags, so we have to replay the write.
1140          */
1141         nfs_mark_request_dirty(req);
1142         nfs_unlock_and_release_request(req);
1143         error = nfs_wb_page(inode, page);
1144         return (error < 0) ? ERR_PTR(error) : NULL;
1145 }
1146
1147 /*
1148  * Try to update an existing write request, or create one if there is none.
1149  *
1150  * Note: Should always be called with the Page Lock held to prevent races
1151  * if we have to add a new request. Also assumes that the caller has
1152  * already called nfs_flush_incompatible() if necessary.
1153  */
1154 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1155                 struct page *page, unsigned int offset, unsigned int bytes)
1156 {
1157         struct inode *inode = page_file_mapping(page)->host;
1158         struct nfs_page *req;
1159
1160         req = nfs_try_to_update_request(inode, page, offset, bytes);
1161         if (req != NULL)
1162                 goto out;
1163         req = nfs_create_request(ctx, page, NULL, offset, bytes);
1164         if (IS_ERR(req))
1165                 goto out;
1166         nfs_inode_add_request(inode, req);
1167 out:
1168         return req;
1169 }
1170
1171 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1172                 unsigned int offset, unsigned int count)
1173 {
1174         struct nfs_page *req;
1175
1176         req = nfs_setup_write_request(ctx, page, offset, count);
1177         if (IS_ERR(req))
1178                 return PTR_ERR(req);
1179         /* Update file length */
1180         nfs_grow_file(page, offset, count);
1181         nfs_mark_uptodate(req);
1182         nfs_mark_request_dirty(req);
1183         nfs_unlock_and_release_request(req);
1184         return 0;
1185 }
1186
1187 int nfs_flush_incompatible(struct file *file, struct page *page)
1188 {
1189         struct nfs_open_context *ctx = nfs_file_open_context(file);
1190         struct nfs_lock_context *l_ctx;
1191         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1192         struct nfs_page *req;
1193         int do_flush, status;
1194         /*
1195          * Look for a request corresponding to this page. If there
1196          * is one, and it belongs to another file, we flush it out
1197          * before we try to copy anything into the page. Do this
1198          * due to the lack of an ACCESS-type call in NFSv2.
1199          * Also do the same if we find a request from an existing
1200          * dropped page.
1201          */
1202         do {
1203                 req = nfs_page_find_head_request(page);
1204                 if (req == NULL)
1205                         return 0;
1206                 l_ctx = req->wb_lock_context;
1207                 do_flush = req->wb_page != page ||
1208                         !nfs_match_open_context(req->wb_context, ctx);
1209                 if (l_ctx && flctx &&
1210                     !(list_empty_careful(&flctx->flc_posix) &&
1211                       list_empty_careful(&flctx->flc_flock))) {
1212                         do_flush |= l_ctx->lockowner != current->files;
1213                 }
1214                 nfs_release_request(req);
1215                 if (!do_flush)
1216                         return 0;
1217                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1218         } while (status == 0);
1219         return status;
1220 }
1221
1222 /*
1223  * Avoid buffered writes when a open context credential's key would
1224  * expire soon.
1225  *
1226  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1227  *
1228  * Return 0 and set a credential flag which triggers the inode to flush
1229  * and performs  NFS_FILE_SYNC writes if the key will expired within
1230  * RPC_KEY_EXPIRE_TIMEO.
1231  */
1232 int
1233 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1234 {
1235         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1236         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1237
1238         return rpcauth_key_timeout_notify(auth, ctx->cred);
1239 }
1240
1241 /*
1242  * Test if the open context credential key is marked to expire soon.
1243  */
1244 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1245 {
1246         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1247
1248         return rpcauth_cred_key_to_expire(auth, ctx->cred);
1249 }
1250
1251 /*
1252  * If the page cache is marked as unsafe or invalid, then we can't rely on
1253  * the PageUptodate() flag. In this case, we will need to turn off
1254  * write optimisations that depend on the page contents being correct.
1255  */
1256 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1257 {
1258         struct nfs_inode *nfsi = NFS_I(inode);
1259
1260         if (nfs_have_delegated_attributes(inode))
1261                 goto out;
1262         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1263                 return false;
1264         smp_rmb();
1265         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1266                 return false;
1267 out:
1268         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1269                 return false;
1270         return PageUptodate(page) != 0;
1271 }
1272
1273 static bool
1274 is_whole_file_wrlock(struct file_lock *fl)
1275 {
1276         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1277                         fl->fl_type == F_WRLCK;
1278 }
1279
1280 /* If we know the page is up to date, and we're not using byte range locks (or
1281  * if we have the whole file locked for writing), it may be more efficient to
1282  * extend the write to cover the entire page in order to avoid fragmentation
1283  * inefficiencies.
1284  *
1285  * If the file is opened for synchronous writes then we can just skip the rest
1286  * of the checks.
1287  */
1288 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1289 {
1290         int ret;
1291         struct file_lock_context *flctx = inode->i_flctx;
1292         struct file_lock *fl;
1293
1294         if (file->f_flags & O_DSYNC)
1295                 return 0;
1296         if (!nfs_write_pageuptodate(page, inode))
1297                 return 0;
1298         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1299                 return 1;
1300         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1301                        list_empty_careful(&flctx->flc_posix)))
1302                 return 1;
1303
1304         /* Check to see if there are whole file write locks */
1305         ret = 0;
1306         spin_lock(&flctx->flc_lock);
1307         if (!list_empty(&flctx->flc_posix)) {
1308                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1309                                         fl_list);
1310                 if (is_whole_file_wrlock(fl))
1311                         ret = 1;
1312         } else if (!list_empty(&flctx->flc_flock)) {
1313                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1314                                         fl_list);
1315                 if (fl->fl_type == F_WRLCK)
1316                         ret = 1;
1317         }
1318         spin_unlock(&flctx->flc_lock);
1319         return ret;
1320 }
1321
1322 /*
1323  * Update and possibly write a cached page of an NFS file.
1324  *
1325  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1326  * things with a page scheduled for an RPC call (e.g. invalidate it).
1327  */
1328 int nfs_updatepage(struct file *file, struct page *page,
1329                 unsigned int offset, unsigned int count)
1330 {
1331         struct nfs_open_context *ctx = nfs_file_open_context(file);
1332         struct inode    *inode = page_file_mapping(page)->host;
1333         int             status = 0;
1334
1335         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1336
1337         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1338                 file, count, (long long)(page_file_offset(page) + offset));
1339
1340         if (!count)
1341                 goto out;
1342
1343         if (nfs_can_extend_write(file, page, inode)) {
1344                 count = max(count + offset, nfs_page_length(page));
1345                 offset = 0;
1346         }
1347
1348         status = nfs_writepage_setup(ctx, page, offset, count);
1349         if (status < 0)
1350                 nfs_set_pageerror(page);
1351         else
1352                 __set_page_dirty_nobuffers(page);
1353 out:
1354         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1355                         status, (long long)i_size_read(inode));
1356         return status;
1357 }
1358
1359 static int flush_task_priority(int how)
1360 {
1361         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1362                 case FLUSH_HIGHPRI:
1363                         return RPC_PRIORITY_HIGH;
1364                 case FLUSH_LOWPRI:
1365                         return RPC_PRIORITY_LOW;
1366         }
1367         return RPC_PRIORITY_NORMAL;
1368 }
1369
1370 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1371                                struct rpc_message *msg,
1372                                const struct nfs_rpc_ops *rpc_ops,
1373                                struct rpc_task_setup *task_setup_data, int how)
1374 {
1375         int priority = flush_task_priority(how);
1376
1377         task_setup_data->priority = priority;
1378         rpc_ops->write_setup(hdr, msg);
1379         trace_nfs_initiate_write(hdr->inode, hdr->io_start, hdr->good_bytes,
1380                                  hdr->args.stable);
1381
1382         nfs4_state_protect_write(NFS_SERVER(hdr->inode)->nfs_client,
1383                                  &task_setup_data->rpc_client, msg, hdr);
1384 }
1385
1386 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1387  * call this on each, which will prepare them to be retried on next
1388  * writeback using standard nfs.
1389  */
1390 static void nfs_redirty_request(struct nfs_page *req)
1391 {
1392         nfs_mark_request_dirty(req);
1393         set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1394         nfs_end_page_writeback(req);
1395         nfs_release_request(req);
1396 }
1397
1398 static void nfs_async_write_error(struct list_head *head)
1399 {
1400         struct nfs_page *req;
1401
1402         while (!list_empty(head)) {
1403                 req = nfs_list_entry(head->next);
1404                 nfs_list_remove_request(req);
1405                 nfs_redirty_request(req);
1406         }
1407 }
1408
1409 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1410 {
1411         nfs_async_write_error(&hdr->pages);
1412 }
1413
1414 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1415         .init_hdr = nfs_async_write_init,
1416         .error_cleanup = nfs_async_write_error,
1417         .completion = nfs_write_completion,
1418         .reschedule_io = nfs_async_write_reschedule_io,
1419 };
1420
1421 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1422                                struct inode *inode, int ioflags, bool force_mds,
1423                                const struct nfs_pgio_completion_ops *compl_ops)
1424 {
1425         struct nfs_server *server = NFS_SERVER(inode);
1426         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1427
1428 #ifdef CONFIG_NFS_V4_1
1429         if (server->pnfs_curr_ld && !force_mds)
1430                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1431 #endif
1432         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1433                         server->wsize, ioflags);
1434 }
1435 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1436
1437 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1438 {
1439         struct nfs_pgio_mirror *mirror;
1440
1441         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1442                 pgio->pg_ops->pg_cleanup(pgio);
1443
1444         pgio->pg_ops = &nfs_pgio_rw_ops;
1445
1446         nfs_pageio_stop_mirroring(pgio);
1447
1448         mirror = &pgio->pg_mirrors[0];
1449         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1450 }
1451 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1452
1453
1454 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1455 {
1456         struct nfs_commit_data *data = calldata;
1457
1458         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1459 }
1460
1461 /*
1462  * Special version of should_remove_suid() that ignores capabilities.
1463  */
1464 static int nfs_should_remove_suid(const struct inode *inode)
1465 {
1466         umode_t mode = inode->i_mode;
1467         int kill = 0;
1468
1469         /* suid always must be killed */
1470         if (unlikely(mode & S_ISUID))
1471                 kill = ATTR_KILL_SUID;
1472
1473         /*
1474          * sgid without any exec bits is just a mandatory locking mark; leave
1475          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1476          */
1477         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1478                 kill |= ATTR_KILL_SGID;
1479
1480         if (unlikely(kill && S_ISREG(mode)))
1481                 return kill;
1482
1483         return 0;
1484 }
1485
1486 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1487                 struct nfs_fattr *fattr)
1488 {
1489         struct nfs_pgio_args *argp = &hdr->args;
1490         struct nfs_pgio_res *resp = &hdr->res;
1491         u64 size = argp->offset + resp->count;
1492
1493         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1494                 fattr->size = size;
1495         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1496                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1497                 return;
1498         }
1499         if (size != fattr->size)
1500                 return;
1501         /* Set attribute barrier */
1502         nfs_fattr_set_barrier(fattr);
1503         /* ...and update size */
1504         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1505 }
1506
1507 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1508 {
1509         struct nfs_fattr *fattr = &hdr->fattr;
1510         struct inode *inode = hdr->inode;
1511
1512         spin_lock(&inode->i_lock);
1513         nfs_writeback_check_extend(hdr, fattr);
1514         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1515         spin_unlock(&inode->i_lock);
1516 }
1517 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1518
1519 /*
1520  * This function is called when the WRITE call is complete.
1521  */
1522 static int nfs_writeback_done(struct rpc_task *task,
1523                               struct nfs_pgio_header *hdr,
1524                               struct inode *inode)
1525 {
1526         int status;
1527
1528         /*
1529          * ->write_done will attempt to use post-op attributes to detect
1530          * conflicting writes by other clients.  A strict interpretation
1531          * of close-to-open would allow us to continue caching even if
1532          * another writer had changed the file, but some applications
1533          * depend on tighter cache coherency when writing.
1534          */
1535         status = NFS_PROTO(inode)->write_done(task, hdr);
1536         if (status != 0)
1537                 return status;
1538
1539         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1540         trace_nfs_writeback_done(inode, task->tk_status,
1541                                  hdr->args.offset, hdr->res.verf);
1542
1543         if (hdr->res.verf->committed < hdr->args.stable &&
1544             task->tk_status >= 0) {
1545                 /* We tried a write call, but the server did not
1546                  * commit data to stable storage even though we
1547                  * requested it.
1548                  * Note: There is a known bug in Tru64 < 5.0 in which
1549                  *       the server reports NFS_DATA_SYNC, but performs
1550                  *       NFS_FILE_SYNC. We therefore implement this checking
1551                  *       as a dprintk() in order to avoid filling syslog.
1552                  */
1553                 static unsigned long    complain;
1554
1555                 /* Note this will print the MDS for a DS write */
1556                 if (time_before(complain, jiffies)) {
1557                         dprintk("NFS:       faulty NFS server %s:"
1558                                 " (committed = %d) != (stable = %d)\n",
1559                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1560                                 hdr->res.verf->committed, hdr->args.stable);
1561                         complain = jiffies + 300 * HZ;
1562                 }
1563         }
1564
1565         /* Deal with the suid/sgid bit corner case */
1566         if (nfs_should_remove_suid(inode))
1567                 nfs_mark_for_revalidate(inode);
1568         return 0;
1569 }
1570
1571 /*
1572  * This function is called when the WRITE call is complete.
1573  */
1574 static void nfs_writeback_result(struct rpc_task *task,
1575                                  struct nfs_pgio_header *hdr)
1576 {
1577         struct nfs_pgio_args    *argp = &hdr->args;
1578         struct nfs_pgio_res     *resp = &hdr->res;
1579
1580         if (resp->count < argp->count) {
1581                 static unsigned long    complain;
1582
1583                 /* This a short write! */
1584                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1585
1586                 /* Has the server at least made some progress? */
1587                 if (resp->count == 0) {
1588                         if (time_before(complain, jiffies)) {
1589                                 printk(KERN_WARNING
1590                                        "NFS: Server wrote zero bytes, expected %u.\n",
1591                                        argp->count);
1592                                 complain = jiffies + 300 * HZ;
1593                         }
1594                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1595                         task->tk_status = -EIO;
1596                         return;
1597                 }
1598
1599                 /* For non rpc-based layout drivers, retry-through-MDS */
1600                 if (!task->tk_ops) {
1601                         hdr->pnfs_error = -EAGAIN;
1602                         return;
1603                 }
1604
1605                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1606                 if (resp->verf->committed != NFS_UNSTABLE) {
1607                         /* Resend from where the server left off */
1608                         hdr->mds_offset += resp->count;
1609                         argp->offset += resp->count;
1610                         argp->pgbase += resp->count;
1611                         argp->count -= resp->count;
1612                 } else {
1613                         /* Resend as a stable write in order to avoid
1614                          * headaches in the case of a server crash.
1615                          */
1616                         argp->stable = NFS_FILE_SYNC;
1617                 }
1618                 rpc_restart_call_prepare(task);
1619         }
1620 }
1621
1622 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1623 {
1624         return wait_on_atomic_t(&cinfo->rpcs_out,
1625                         nfs_wait_atomic_killable, TASK_KILLABLE);
1626 }
1627
1628 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1629 {
1630         atomic_inc(&cinfo->rpcs_out);
1631 }
1632
1633 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1634 {
1635         if (atomic_dec_and_test(&cinfo->rpcs_out))
1636                 wake_up_atomic_t(&cinfo->rpcs_out);
1637 }
1638
1639 void nfs_commitdata_release(struct nfs_commit_data *data)
1640 {
1641         put_nfs_open_context(data->context);
1642         nfs_commit_free(data);
1643 }
1644 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1645
1646 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1647                         const struct nfs_rpc_ops *nfs_ops,
1648                         const struct rpc_call_ops *call_ops,
1649                         int how, int flags)
1650 {
1651         struct rpc_task *task;
1652         int priority = flush_task_priority(how);
1653         struct rpc_message msg = {
1654                 .rpc_argp = &data->args,
1655                 .rpc_resp = &data->res,
1656                 .rpc_cred = data->cred,
1657         };
1658         struct rpc_task_setup task_setup_data = {
1659                 .task = &data->task,
1660                 .rpc_client = clnt,
1661                 .rpc_message = &msg,
1662                 .callback_ops = call_ops,
1663                 .callback_data = data,
1664                 .workqueue = nfsiod_workqueue,
1665                 .flags = RPC_TASK_ASYNC | flags,
1666                 .priority = priority,
1667         };
1668         /* Set up the initial task struct.  */
1669         nfs_ops->commit_setup(data, &msg);
1670         trace_nfs_initiate_commit(data);
1671
1672         dprintk("NFS: initiated commit call\n");
1673
1674         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1675                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1676
1677         task = rpc_run_task(&task_setup_data);
1678         if (IS_ERR(task))
1679                 return PTR_ERR(task);
1680         if (how & FLUSH_SYNC)
1681                 rpc_wait_for_completion_task(task);
1682         rpc_put_task(task);
1683         return 0;
1684 }
1685 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1686
1687 static loff_t nfs_get_lwb(struct list_head *head)
1688 {
1689         loff_t lwb = 0;
1690         struct nfs_page *req;
1691
1692         list_for_each_entry(req, head, wb_list)
1693                 if (lwb < (req_offset(req) + req->wb_bytes))
1694                         lwb = req_offset(req) + req->wb_bytes;
1695
1696         return lwb;
1697 }
1698
1699 /*
1700  * Set up the argument/result storage required for the RPC call.
1701  */
1702 void nfs_init_commit(struct nfs_commit_data *data,
1703                      struct list_head *head,
1704                      struct pnfs_layout_segment *lseg,
1705                      struct nfs_commit_info *cinfo)
1706 {
1707         struct nfs_page *first = nfs_list_entry(head->next);
1708         struct inode *inode = d_inode(first->wb_context->dentry);
1709
1710         /* Set up the RPC argument and reply structs
1711          * NB: take care not to mess about with data->commit et al. */
1712
1713         list_splice_init(head, &data->pages);
1714
1715         data->inode       = inode;
1716         data->cred        = first->wb_context->cred;
1717         data->lseg        = lseg; /* reference transferred */
1718         /* only set lwb for pnfs commit */
1719         if (lseg)
1720                 data->lwb = nfs_get_lwb(&data->pages);
1721         data->mds_ops     = &nfs_commit_ops;
1722         data->completion_ops = cinfo->completion_ops;
1723         data->dreq        = cinfo->dreq;
1724
1725         data->args.fh     = NFS_FH(data->inode);
1726         /* Note: we always request a commit of the entire inode */
1727         data->args.offset = 0;
1728         data->args.count  = 0;
1729         data->context     = get_nfs_open_context(first->wb_context);
1730         data->res.fattr   = &data->fattr;
1731         data->res.verf    = &data->verf;
1732         nfs_fattr_init(&data->fattr);
1733 }
1734 EXPORT_SYMBOL_GPL(nfs_init_commit);
1735
1736 void nfs_retry_commit(struct list_head *page_list,
1737                       struct pnfs_layout_segment *lseg,
1738                       struct nfs_commit_info *cinfo,
1739                       u32 ds_commit_idx)
1740 {
1741         struct nfs_page *req;
1742
1743         while (!list_empty(page_list)) {
1744                 req = nfs_list_entry(page_list->next);
1745                 nfs_list_remove_request(req);
1746                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1747                 if (!cinfo->dreq)
1748                         nfs_clear_page_commit(req->wb_page);
1749                 nfs_unlock_and_release_request(req);
1750         }
1751 }
1752 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1753
1754 static void
1755 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1756                 struct nfs_page *req)
1757 {
1758         __set_page_dirty_nobuffers(req->wb_page);
1759 }
1760
1761 /*
1762  * Commit dirty pages
1763  */
1764 static int
1765 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1766                 struct nfs_commit_info *cinfo)
1767 {
1768         struct nfs_commit_data  *data;
1769
1770         /* another commit raced with us */
1771         if (list_empty(head))
1772                 return 0;
1773
1774         data = nfs_commitdata_alloc(true);
1775
1776         /* Set up the argument struct */
1777         nfs_init_commit(data, head, NULL, cinfo);
1778         atomic_inc(&cinfo->mds->rpcs_out);
1779         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1780                                    data->mds_ops, how, 0);
1781 }
1782
1783 /*
1784  * COMMIT call returned
1785  */
1786 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1787 {
1788         struct nfs_commit_data  *data = calldata;
1789
1790         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1791                                 task->tk_pid, task->tk_status);
1792
1793         /* Call the NFS version-specific code */
1794         NFS_PROTO(data->inode)->commit_done(task, data);
1795         trace_nfs_commit_done(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 };