Merge tag 'nfs-for-4.21-1' of git://git.linux-nfs.org/projects/anna/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 #include <linux/iversion.h>
27
28 #include <linux/uaccess.h>
29
30 #include "delegation.h"
31 #include "internal.h"
32 #include "iostat.h"
33 #include "nfs4_fs.h"
34 #include "fscache.h"
35 #include "pnfs.h"
36
37 #include "nfstrace.h"
38
39 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
40
41 #define MIN_POOL_WRITE          (32)
42 #define MIN_POOL_COMMIT         (4)
43
44 struct nfs_io_completion {
45         void (*complete)(void *data);
46         void *data;
47         struct kref refcount;
48 };
49
50 /*
51  * Local function declarations
52  */
53 static void nfs_redirty_request(struct nfs_page *req);
54 static const struct rpc_call_ops nfs_commit_ops;
55 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
56 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
57 static const struct nfs_rw_ops nfs_rw_write_ops;
58 static void nfs_clear_request_commit(struct nfs_page *req);
59 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
60                                       struct inode *inode);
61 static struct nfs_page *
62 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
63                                                 struct page *page);
64
65 static struct kmem_cache *nfs_wdata_cachep;
66 static mempool_t *nfs_wdata_mempool;
67 static struct kmem_cache *nfs_cdata_cachep;
68 static mempool_t *nfs_commit_mempool;
69
70 struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
71 {
72         struct nfs_commit_data *p;
73
74         if (never_fail)
75                 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
76         else {
77                 /* It is OK to do some reclaim, not no safe to wait
78                  * for anything to be returned to the pool.
79                  * mempool_alloc() cannot handle that particular combination,
80                  * so we need two separate attempts.
81                  */
82                 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
83                 if (!p)
84                         p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
85                                              __GFP_NOWARN | __GFP_NORETRY);
86                 if (!p)
87                         return NULL;
88         }
89
90         memset(p, 0, sizeof(*p));
91         INIT_LIST_HEAD(&p->pages);
92         return p;
93 }
94 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
95
96 void nfs_commit_free(struct nfs_commit_data *p)
97 {
98         mempool_free(p, nfs_commit_mempool);
99 }
100 EXPORT_SYMBOL_GPL(nfs_commit_free);
101
102 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
103 {
104         struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
105
106         memset(p, 0, sizeof(*p));
107         p->rw_mode = FMODE_WRITE;
108         return p;
109 }
110
111 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
112 {
113         mempool_free(hdr, nfs_wdata_mempool);
114 }
115
116 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
117 {
118         return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
119 }
120
121 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
122                 void (*complete)(void *), void *data)
123 {
124         ioc->complete = complete;
125         ioc->data = data;
126         kref_init(&ioc->refcount);
127 }
128
129 static void nfs_io_completion_release(struct kref *kref)
130 {
131         struct nfs_io_completion *ioc = container_of(kref,
132                         struct nfs_io_completion, refcount);
133         ioc->complete(ioc->data);
134         kfree(ioc);
135 }
136
137 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
138 {
139         if (ioc != NULL)
140                 kref_get(&ioc->refcount);
141 }
142
143 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
144 {
145         if (ioc != NULL)
146                 kref_put(&ioc->refcount, nfs_io_completion_release);
147 }
148
149 static struct nfs_page *
150 nfs_page_private_request(struct page *page)
151 {
152         if (!PagePrivate(page))
153                 return NULL;
154         return (struct nfs_page *)page_private(page);
155 }
156
157 /*
158  * nfs_page_find_head_request_locked - find head request associated with @page
159  *
160  * must be called while holding the inode lock.
161  *
162  * returns matching head request with reference held, or NULL if not found.
163  */
164 static struct nfs_page *
165 nfs_page_find_private_request(struct page *page)
166 {
167         struct address_space *mapping = page_file_mapping(page);
168         struct nfs_page *req;
169
170         if (!PagePrivate(page))
171                 return NULL;
172         spin_lock(&mapping->private_lock);
173         req = nfs_page_private_request(page);
174         if (req) {
175                 WARN_ON_ONCE(req->wb_head != req);
176                 kref_get(&req->wb_kref);
177         }
178         spin_unlock(&mapping->private_lock);
179         return req;
180 }
181
182 static struct nfs_page *
183 nfs_page_find_swap_request(struct page *page)
184 {
185         struct inode *inode = page_file_mapping(page)->host;
186         struct nfs_inode *nfsi = NFS_I(inode);
187         struct nfs_page *req = NULL;
188         if (!PageSwapCache(page))
189                 return NULL;
190         mutex_lock(&nfsi->commit_mutex);
191         if (PageSwapCache(page)) {
192                 req = nfs_page_search_commits_for_head_request_locked(nfsi,
193                         page);
194                 if (req) {
195                         WARN_ON_ONCE(req->wb_head != req);
196                         kref_get(&req->wb_kref);
197                 }
198         }
199         mutex_unlock(&nfsi->commit_mutex);
200         return req;
201 }
202
203 /*
204  * nfs_page_find_head_request - find head request associated with @page
205  *
206  * returns matching head request with reference held, or NULL if not found.
207  */
208 static struct nfs_page *nfs_page_find_head_request(struct page *page)
209 {
210         struct nfs_page *req;
211
212         req = nfs_page_find_private_request(page);
213         if (!req)
214                 req = nfs_page_find_swap_request(page);
215         return req;
216 }
217
218 /* Adjust the file length if we're writing beyond the end */
219 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
220 {
221         struct inode *inode = page_file_mapping(page)->host;
222         loff_t end, i_size;
223         pgoff_t end_index;
224
225         spin_lock(&inode->i_lock);
226         i_size = i_size_read(inode);
227         end_index = (i_size - 1) >> PAGE_SHIFT;
228         if (i_size > 0 && page_index(page) < end_index)
229                 goto out;
230         end = page_file_offset(page) + ((loff_t)offset+count);
231         if (i_size >= end)
232                 goto out;
233         i_size_write(inode, end);
234         NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
235         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
236 out:
237         spin_unlock(&inode->i_lock);
238 }
239
240 /* A writeback failed: mark the page as bad, and invalidate the page cache */
241 static void nfs_set_pageerror(struct page *page)
242 {
243         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
244 }
245
246 /*
247  * nfs_page_group_search_locked
248  * @head - head request of page group
249  * @page_offset - offset into page
250  *
251  * Search page group with head @head to find a request that contains the
252  * page offset @page_offset.
253  *
254  * Returns a pointer to the first matching nfs request, or NULL if no
255  * match is found.
256  *
257  * Must be called with the page group lock held
258  */
259 static struct nfs_page *
260 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
261 {
262         struct nfs_page *req;
263
264         req = head;
265         do {
266                 if (page_offset >= req->wb_pgbase &&
267                     page_offset < (req->wb_pgbase + req->wb_bytes))
268                         return req;
269
270                 req = req->wb_this_page;
271         } while (req != head);
272
273         return NULL;
274 }
275
276 /*
277  * nfs_page_group_covers_page
278  * @head - head request of page group
279  *
280  * Return true if the page group with head @head covers the whole page,
281  * returns false otherwise
282  */
283 static bool nfs_page_group_covers_page(struct nfs_page *req)
284 {
285         struct nfs_page *tmp;
286         unsigned int pos = 0;
287         unsigned int len = nfs_page_length(req->wb_page);
288
289         nfs_page_group_lock(req);
290
291         for (;;) {
292                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
293                 if (!tmp)
294                         break;
295                 pos = tmp->wb_pgbase + tmp->wb_bytes;
296         }
297
298         nfs_page_group_unlock(req);
299         return pos >= len;
300 }
301
302 /* We can set the PG_uptodate flag if we see that a write request
303  * covers the full page.
304  */
305 static void nfs_mark_uptodate(struct nfs_page *req)
306 {
307         if (PageUptodate(req->wb_page))
308                 return;
309         if (!nfs_page_group_covers_page(req))
310                 return;
311         SetPageUptodate(req->wb_page);
312 }
313
314 static int wb_priority(struct writeback_control *wbc)
315 {
316         int ret = 0;
317
318         if (wbc->sync_mode == WB_SYNC_ALL)
319                 ret = FLUSH_COND_STABLE;
320         return ret;
321 }
322
323 /*
324  * NFS congestion control
325  */
326
327 int nfs_congestion_kb;
328
329 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
330 #define NFS_CONGESTION_OFF_THRESH       \
331         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
332
333 static void nfs_set_page_writeback(struct page *page)
334 {
335         struct inode *inode = page_file_mapping(page)->host;
336         struct nfs_server *nfss = NFS_SERVER(inode);
337         int ret = test_set_page_writeback(page);
338
339         WARN_ON_ONCE(ret != 0);
340
341         if (atomic_long_inc_return(&nfss->writeback) >
342                         NFS_CONGESTION_ON_THRESH)
343                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
344 }
345
346 static void nfs_end_page_writeback(struct nfs_page *req)
347 {
348         struct inode *inode = page_file_mapping(req->wb_page)->host;
349         struct nfs_server *nfss = NFS_SERVER(inode);
350         bool is_done;
351
352         is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
353         nfs_unlock_request(req);
354         if (!is_done)
355                 return;
356
357         end_page_writeback(req->wb_page);
358         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
359                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
360 }
361
362 /*
363  * nfs_unroll_locks_and_wait -  unlock all newly locked reqs and wait on @req
364  *
365  * this is a helper function for nfs_lock_and_join_requests
366  *
367  * @inode - inode associated with request page group, must be holding inode lock
368  * @head  - head request of page group, must be holding head lock
369  * @req   - request that couldn't lock and needs to wait on the req bit lock
370  *
371  * NOTE: this must be called holding page_group bit lock
372  *       which will be released before returning.
373  *
374  * returns 0 on success, < 0 on error.
375  */
376 static void
377 nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
378                           struct nfs_page *req)
379 {
380         struct nfs_page *tmp;
381
382         /* relinquish all the locks successfully grabbed this run */
383         for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
384                 if (!kref_read(&tmp->wb_kref))
385                         continue;
386                 nfs_unlock_and_release_request(tmp);
387         }
388 }
389
390 /*
391  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
392  *
393  * @destroy_list - request list (using wb_this_page) terminated by @old_head
394  * @old_head - the old head of the list
395  *
396  * All subrequests must be locked and removed from all lists, so at this point
397  * they are only "active" in this function, and possibly in nfs_wait_on_request
398  * with a reference held by some other context.
399  */
400 static void
401 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
402                                  struct nfs_page *old_head,
403                                  struct inode *inode)
404 {
405         while (destroy_list) {
406                 struct nfs_page *subreq = destroy_list;
407
408                 destroy_list = (subreq->wb_this_page == old_head) ?
409                                    NULL : subreq->wb_this_page;
410
411                 WARN_ON_ONCE(old_head != subreq->wb_head);
412
413                 /* make sure old group is not used */
414                 subreq->wb_this_page = subreq;
415
416                 clear_bit(PG_REMOVE, &subreq->wb_flags);
417
418                 /* Note: races with nfs_page_group_destroy() */
419                 if (!kref_read(&subreq->wb_kref)) {
420                         /* Check if we raced with nfs_page_group_destroy() */
421                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags))
422                                 nfs_free_request(subreq);
423                         continue;
424                 }
425
426                 subreq->wb_head = subreq;
427
428                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
429                         nfs_release_request(subreq);
430                         atomic_long_dec(&NFS_I(inode)->nrequests);
431                 }
432
433                 /* subreq is now totally disconnected from page group or any
434                  * write / commit lists. last chance to wake any waiters */
435                 nfs_unlock_and_release_request(subreq);
436         }
437 }
438
439 /*
440  * nfs_lock_and_join_requests - join all subreqs to the head req and return
441  *                              a locked reference, cancelling any pending
442  *                              operations for this page.
443  *
444  * @page - the page used to lookup the "page group" of nfs_page structures
445  *
446  * This function joins all sub requests to the head request by first
447  * locking all requests in the group, cancelling any pending operations
448  * and finally updating the head request to cover the whole range covered by
449  * the (former) group.  All subrequests are removed from any write or commit
450  * lists, unlinked from the group and destroyed.
451  *
452  * Returns a locked, referenced pointer to the head request - which after
453  * this call is guaranteed to be the only request associated with the page.
454  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
455  * error was encountered.
456  */
457 static struct nfs_page *
458 nfs_lock_and_join_requests(struct page *page)
459 {
460         struct inode *inode = page_file_mapping(page)->host;
461         struct nfs_page *head, *subreq;
462         struct nfs_page *destroy_list = NULL;
463         unsigned int total_bytes;
464         int ret;
465
466 try_again:
467         /*
468          * A reference is taken only on the head request which acts as a
469          * reference to the whole page group - the group will not be destroyed
470          * until the head reference is released.
471          */
472         head = nfs_page_find_head_request(page);
473         if (!head)
474                 return NULL;
475
476         /* lock the page head first in order to avoid an ABBA inefficiency */
477         if (!nfs_lock_request(head)) {
478                 ret = nfs_wait_on_request(head);
479                 nfs_release_request(head);
480                 if (ret < 0)
481                         return ERR_PTR(ret);
482                 goto try_again;
483         }
484
485         /* Ensure that nobody removed the request before we locked it */
486         if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
487                 nfs_unlock_and_release_request(head);
488                 goto try_again;
489         }
490
491         ret = nfs_page_group_lock(head);
492         if (ret < 0)
493                 goto release_request;
494
495         /* lock each request in the page group */
496         total_bytes = head->wb_bytes;
497         for (subreq = head->wb_this_page; subreq != head;
498                         subreq = subreq->wb_this_page) {
499
500                 if (!kref_get_unless_zero(&subreq->wb_kref)) {
501                         if (subreq->wb_offset == head->wb_offset + total_bytes)
502                                 total_bytes += subreq->wb_bytes;
503                         continue;
504                 }
505
506                 while (!nfs_lock_request(subreq)) {
507                         /*
508                          * Unlock page to allow nfs_page_group_sync_on_bit()
509                          * to succeed
510                          */
511                         nfs_page_group_unlock(head);
512                         ret = nfs_wait_on_request(subreq);
513                         if (!ret)
514                                 ret = nfs_page_group_lock(head);
515                         if (ret < 0) {
516                                 nfs_unroll_locks(inode, head, subreq);
517                                 nfs_release_request(subreq);
518                                 goto release_request;
519                         }
520                 }
521                 /*
522                  * Subrequests are always contiguous, non overlapping
523                  * and in order - but may be repeated (mirrored writes).
524                  */
525                 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
526                         /* keep track of how many bytes this group covers */
527                         total_bytes += subreq->wb_bytes;
528                 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
529                             ((subreq->wb_offset + subreq->wb_bytes) >
530                              (head->wb_offset + total_bytes)))) {
531                         nfs_page_group_unlock(head);
532                         nfs_unroll_locks(inode, head, subreq);
533                         nfs_unlock_and_release_request(subreq);
534                         ret = -EIO;
535                         goto release_request;
536                 }
537         }
538
539         /* Now that all requests are locked, make sure they aren't on any list.
540          * Commit list removal accounting is done after locks are dropped */
541         subreq = head;
542         do {
543                 nfs_clear_request_commit(subreq);
544                 subreq = subreq->wb_this_page;
545         } while (subreq != head);
546
547         /* unlink subrequests from head, destroy them later */
548         if (head->wb_this_page != head) {
549                 /* destroy list will be terminated by head */
550                 destroy_list = head->wb_this_page;
551                 head->wb_this_page = head;
552
553                 /* change head request to cover whole range that
554                  * the former page group covered */
555                 head->wb_bytes = total_bytes;
556         }
557
558         /* Postpone destruction of this request */
559         if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
560                 set_bit(PG_INODE_REF, &head->wb_flags);
561                 kref_get(&head->wb_kref);
562                 atomic_long_inc(&NFS_I(inode)->nrequests);
563         }
564
565         nfs_page_group_unlock(head);
566
567         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
568
569         /* Did we lose a race with nfs_inode_remove_request()? */
570         if (!(PagePrivate(page) || PageSwapCache(page))) {
571                 nfs_unlock_and_release_request(head);
572                 return NULL;
573         }
574
575         /* still holds ref on head from nfs_page_find_head_request
576          * and still has lock on head from lock loop */
577         return head;
578
579 release_request:
580         nfs_unlock_and_release_request(head);
581         return ERR_PTR(ret);
582 }
583
584 static void nfs_write_error_remove_page(struct nfs_page *req)
585 {
586         nfs_end_page_writeback(req);
587         generic_error_remove_page(page_file_mapping(req->wb_page),
588                                   req->wb_page);
589         nfs_release_request(req);
590 }
591
592 static bool
593 nfs_error_is_fatal_on_server(int err)
594 {
595         switch (err) {
596         case 0:
597         case -ERESTARTSYS:
598         case -EINTR:
599                 return false;
600         }
601         return nfs_error_is_fatal(err);
602 }
603
604 /*
605  * Find an associated nfs write request, and prepare to flush it out
606  * May return an error if the user signalled nfs_wait_on_request().
607  */
608 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
609                                 struct page *page)
610 {
611         struct nfs_page *req;
612         int ret = 0;
613
614         req = nfs_lock_and_join_requests(page);
615         if (!req)
616                 goto out;
617         ret = PTR_ERR(req);
618         if (IS_ERR(req))
619                 goto out;
620
621         nfs_set_page_writeback(page);
622         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
623
624         ret = 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                 inode_inc_iversion_raw(inode);
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
1237         if (nfs_ctx_key_to_expire(ctx, inode) &&
1238             !ctx->ll_cred)
1239                 /* Already expired! */
1240                 return -EACCES;
1241         return 0;
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         struct rpc_cred *cred = ctx->ll_cred;
1251         struct auth_cred acred = {
1252                 .cred = ctx->cred,
1253         };
1254
1255         if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1256                 put_rpccred(cred);
1257                 ctx->ll_cred = NULL;
1258                 cred = NULL;
1259         }
1260         if (!cred)
1261                 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1262         if (!cred || IS_ERR(cred))
1263                 return true;
1264         ctx->ll_cred = cred;
1265         return !!(cred->cr_ops->crkey_timeout &&
1266                   cred->cr_ops->crkey_timeout(cred));
1267 }
1268
1269 /*
1270  * If the page cache is marked as unsafe or invalid, then we can't rely on
1271  * the PageUptodate() flag. In this case, we will need to turn off
1272  * write optimisations that depend on the page contents being correct.
1273  */
1274 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1275 {
1276         struct nfs_inode *nfsi = NFS_I(inode);
1277
1278         if (nfs_have_delegated_attributes(inode))
1279                 goto out;
1280         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1281                 return false;
1282         smp_rmb();
1283         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1284                 return false;
1285 out:
1286         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1287                 return false;
1288         return PageUptodate(page) != 0;
1289 }
1290
1291 static bool
1292 is_whole_file_wrlock(struct file_lock *fl)
1293 {
1294         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1295                         fl->fl_type == F_WRLCK;
1296 }
1297
1298 /* If we know the page is up to date, and we're not using byte range locks (or
1299  * if we have the whole file locked for writing), it may be more efficient to
1300  * extend the write to cover the entire page in order to avoid fragmentation
1301  * inefficiencies.
1302  *
1303  * If the file is opened for synchronous writes then we can just skip the rest
1304  * of the checks.
1305  */
1306 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1307 {
1308         int ret;
1309         struct file_lock_context *flctx = inode->i_flctx;
1310         struct file_lock *fl;
1311
1312         if (file->f_flags & O_DSYNC)
1313                 return 0;
1314         if (!nfs_write_pageuptodate(page, inode))
1315                 return 0;
1316         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1317                 return 1;
1318         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1319                        list_empty_careful(&flctx->flc_posix)))
1320                 return 1;
1321
1322         /* Check to see if there are whole file write locks */
1323         ret = 0;
1324         spin_lock(&flctx->flc_lock);
1325         if (!list_empty(&flctx->flc_posix)) {
1326                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1327                                         fl_list);
1328                 if (is_whole_file_wrlock(fl))
1329                         ret = 1;
1330         } else if (!list_empty(&flctx->flc_flock)) {
1331                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1332                                         fl_list);
1333                 if (fl->fl_type == F_WRLCK)
1334                         ret = 1;
1335         }
1336         spin_unlock(&flctx->flc_lock);
1337         return ret;
1338 }
1339
1340 /*
1341  * Update and possibly write a cached page of an NFS file.
1342  *
1343  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1344  * things with a page scheduled for an RPC call (e.g. invalidate it).
1345  */
1346 int nfs_updatepage(struct file *file, struct page *page,
1347                 unsigned int offset, unsigned int count)
1348 {
1349         struct nfs_open_context *ctx = nfs_file_open_context(file);
1350         struct inode    *inode = page_file_mapping(page)->host;
1351         int             status = 0;
1352
1353         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1354
1355         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1356                 file, count, (long long)(page_file_offset(page) + offset));
1357
1358         if (!count)
1359                 goto out;
1360
1361         if (nfs_can_extend_write(file, page, inode)) {
1362                 count = max(count + offset, nfs_page_length(page));
1363                 offset = 0;
1364         }
1365
1366         status = nfs_writepage_setup(ctx, page, offset, count);
1367         if (status < 0)
1368                 nfs_set_pageerror(page);
1369         else
1370                 __set_page_dirty_nobuffers(page);
1371 out:
1372         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1373                         status, (long long)i_size_read(inode));
1374         return status;
1375 }
1376
1377 static int flush_task_priority(int how)
1378 {
1379         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1380                 case FLUSH_HIGHPRI:
1381                         return RPC_PRIORITY_HIGH;
1382                 case FLUSH_LOWPRI:
1383                         return RPC_PRIORITY_LOW;
1384         }
1385         return RPC_PRIORITY_NORMAL;
1386 }
1387
1388 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1389                                struct rpc_message *msg,
1390                                const struct nfs_rpc_ops *rpc_ops,
1391                                struct rpc_task_setup *task_setup_data, int how)
1392 {
1393         int priority = flush_task_priority(how);
1394
1395         task_setup_data->priority = priority;
1396         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1397         trace_nfs_initiate_write(hdr->inode, hdr->io_start, hdr->good_bytes,
1398                                  hdr->args.stable);
1399 }
1400
1401 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1402  * call this on each, which will prepare them to be retried on next
1403  * writeback using standard nfs.
1404  */
1405 static void nfs_redirty_request(struct nfs_page *req)
1406 {
1407         nfs_mark_request_dirty(req);
1408         set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1409         nfs_end_page_writeback(req);
1410         nfs_release_request(req);
1411 }
1412
1413 static void nfs_async_write_error(struct list_head *head)
1414 {
1415         struct nfs_page *req;
1416
1417         while (!list_empty(head)) {
1418                 req = nfs_list_entry(head->next);
1419                 nfs_list_remove_request(req);
1420                 nfs_redirty_request(req);
1421         }
1422 }
1423
1424 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1425 {
1426         nfs_async_write_error(&hdr->pages);
1427         filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1428                         hdr->args.offset + hdr->args.count - 1);
1429 }
1430
1431 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1432         .init_hdr = nfs_async_write_init,
1433         .error_cleanup = nfs_async_write_error,
1434         .completion = nfs_write_completion,
1435         .reschedule_io = nfs_async_write_reschedule_io,
1436 };
1437
1438 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1439                                struct inode *inode, int ioflags, bool force_mds,
1440                                const struct nfs_pgio_completion_ops *compl_ops)
1441 {
1442         struct nfs_server *server = NFS_SERVER(inode);
1443         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1444
1445 #ifdef CONFIG_NFS_V4_1
1446         if (server->pnfs_curr_ld && !force_mds)
1447                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1448 #endif
1449         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1450                         server->wsize, ioflags);
1451 }
1452 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1453
1454 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1455 {
1456         struct nfs_pgio_mirror *mirror;
1457
1458         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1459                 pgio->pg_ops->pg_cleanup(pgio);
1460
1461         pgio->pg_ops = &nfs_pgio_rw_ops;
1462
1463         nfs_pageio_stop_mirroring(pgio);
1464
1465         mirror = &pgio->pg_mirrors[0];
1466         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1467 }
1468 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1469
1470
1471 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1472 {
1473         struct nfs_commit_data *data = calldata;
1474
1475         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1476 }
1477
1478 /*
1479  * Special version of should_remove_suid() that ignores capabilities.
1480  */
1481 static int nfs_should_remove_suid(const struct inode *inode)
1482 {
1483         umode_t mode = inode->i_mode;
1484         int kill = 0;
1485
1486         /* suid always must be killed */
1487         if (unlikely(mode & S_ISUID))
1488                 kill = ATTR_KILL_SUID;
1489
1490         /*
1491          * sgid without any exec bits is just a mandatory locking mark; leave
1492          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1493          */
1494         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1495                 kill |= ATTR_KILL_SGID;
1496
1497         if (unlikely(kill && S_ISREG(mode)))
1498                 return kill;
1499
1500         return 0;
1501 }
1502
1503 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1504                 struct nfs_fattr *fattr)
1505 {
1506         struct nfs_pgio_args *argp = &hdr->args;
1507         struct nfs_pgio_res *resp = &hdr->res;
1508         u64 size = argp->offset + resp->count;
1509
1510         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1511                 fattr->size = size;
1512         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1513                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1514                 return;
1515         }
1516         if (size != fattr->size)
1517                 return;
1518         /* Set attribute barrier */
1519         nfs_fattr_set_barrier(fattr);
1520         /* ...and update size */
1521         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1522 }
1523
1524 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1525 {
1526         struct nfs_fattr *fattr = &hdr->fattr;
1527         struct inode *inode = hdr->inode;
1528
1529         spin_lock(&inode->i_lock);
1530         nfs_writeback_check_extend(hdr, fattr);
1531         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1532         spin_unlock(&inode->i_lock);
1533 }
1534 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1535
1536 /*
1537  * This function is called when the WRITE call is complete.
1538  */
1539 static int nfs_writeback_done(struct rpc_task *task,
1540                               struct nfs_pgio_header *hdr,
1541                               struct inode *inode)
1542 {
1543         int status;
1544
1545         /*
1546          * ->write_done will attempt to use post-op attributes to detect
1547          * conflicting writes by other clients.  A strict interpretation
1548          * of close-to-open would allow us to continue caching even if
1549          * another writer had changed the file, but some applications
1550          * depend on tighter cache coherency when writing.
1551          */
1552         status = NFS_PROTO(inode)->write_done(task, hdr);
1553         if (status != 0)
1554                 return status;
1555
1556         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1557         trace_nfs_writeback_done(inode, task->tk_status,
1558                                  hdr->args.offset, hdr->res.verf);
1559
1560         if (hdr->res.verf->committed < hdr->args.stable &&
1561             task->tk_status >= 0) {
1562                 /* We tried a write call, but the server did not
1563                  * commit data to stable storage even though we
1564                  * requested it.
1565                  * Note: There is a known bug in Tru64 < 5.0 in which
1566                  *       the server reports NFS_DATA_SYNC, but performs
1567                  *       NFS_FILE_SYNC. We therefore implement this checking
1568                  *       as a dprintk() in order to avoid filling syslog.
1569                  */
1570                 static unsigned long    complain;
1571
1572                 /* Note this will print the MDS for a DS write */
1573                 if (time_before(complain, jiffies)) {
1574                         dprintk("NFS:       faulty NFS server %s:"
1575                                 " (committed = %d) != (stable = %d)\n",
1576                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1577                                 hdr->res.verf->committed, hdr->args.stable);
1578                         complain = jiffies + 300 * HZ;
1579                 }
1580         }
1581
1582         /* Deal with the suid/sgid bit corner case */
1583         if (nfs_should_remove_suid(inode)) {
1584                 spin_lock(&inode->i_lock);
1585                 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1586                 spin_unlock(&inode->i_lock);
1587         }
1588         return 0;
1589 }
1590
1591 /*
1592  * This function is called when the WRITE call is complete.
1593  */
1594 static void nfs_writeback_result(struct rpc_task *task,
1595                                  struct nfs_pgio_header *hdr)
1596 {
1597         struct nfs_pgio_args    *argp = &hdr->args;
1598         struct nfs_pgio_res     *resp = &hdr->res;
1599
1600         if (resp->count < argp->count) {
1601                 static unsigned long    complain;
1602
1603                 /* This a short write! */
1604                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1605
1606                 /* Has the server at least made some progress? */
1607                 if (resp->count == 0) {
1608                         if (time_before(complain, jiffies)) {
1609                                 printk(KERN_WARNING
1610                                        "NFS: Server wrote zero bytes, expected %u.\n",
1611                                        argp->count);
1612                                 complain = jiffies + 300 * HZ;
1613                         }
1614                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1615                         task->tk_status = -EIO;
1616                         return;
1617                 }
1618
1619                 /* For non rpc-based layout drivers, retry-through-MDS */
1620                 if (!task->tk_ops) {
1621                         hdr->pnfs_error = -EAGAIN;
1622                         return;
1623                 }
1624
1625                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1626                 if (resp->verf->committed != NFS_UNSTABLE) {
1627                         /* Resend from where the server left off */
1628                         hdr->mds_offset += resp->count;
1629                         argp->offset += resp->count;
1630                         argp->pgbase += resp->count;
1631                         argp->count -= resp->count;
1632                 } else {
1633                         /* Resend as a stable write in order to avoid
1634                          * headaches in the case of a server crash.
1635                          */
1636                         argp->stable = NFS_FILE_SYNC;
1637                 }
1638                 rpc_restart_call_prepare(task);
1639         }
1640 }
1641
1642 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1643 {
1644         return wait_var_event_killable(&cinfo->rpcs_out,
1645                                        !atomic_read(&cinfo->rpcs_out));
1646 }
1647
1648 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1649 {
1650         atomic_inc(&cinfo->rpcs_out);
1651 }
1652
1653 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1654 {
1655         if (atomic_dec_and_test(&cinfo->rpcs_out))
1656                 wake_up_var(&cinfo->rpcs_out);
1657 }
1658
1659 void nfs_commitdata_release(struct nfs_commit_data *data)
1660 {
1661         put_nfs_open_context(data->context);
1662         nfs_commit_free(data);
1663 }
1664 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1665
1666 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1667                         const struct nfs_rpc_ops *nfs_ops,
1668                         const struct rpc_call_ops *call_ops,
1669                         int how, int flags)
1670 {
1671         struct rpc_task *task;
1672         int priority = flush_task_priority(how);
1673         struct rpc_message msg = {
1674                 .rpc_argp = &data->args,
1675                 .rpc_resp = &data->res,
1676                 .rpc_cred = data->cred,
1677         };
1678         struct rpc_task_setup task_setup_data = {
1679                 .task = &data->task,
1680                 .rpc_client = clnt,
1681                 .rpc_message = &msg,
1682                 .callback_ops = call_ops,
1683                 .callback_data = data,
1684                 .workqueue = nfsiod_workqueue,
1685                 .flags = RPC_TASK_ASYNC | flags,
1686                 .priority = priority,
1687         };
1688         /* Set up the initial task struct.  */
1689         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1690         trace_nfs_initiate_commit(data);
1691
1692         dprintk("NFS: initiated commit call\n");
1693
1694         task = rpc_run_task(&task_setup_data);
1695         if (IS_ERR(task))
1696                 return PTR_ERR(task);
1697         if (how & FLUSH_SYNC)
1698                 rpc_wait_for_completion_task(task);
1699         rpc_put_task(task);
1700         return 0;
1701 }
1702 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1703
1704 static loff_t nfs_get_lwb(struct list_head *head)
1705 {
1706         loff_t lwb = 0;
1707         struct nfs_page *req;
1708
1709         list_for_each_entry(req, head, wb_list)
1710                 if (lwb < (req_offset(req) + req->wb_bytes))
1711                         lwb = req_offset(req) + req->wb_bytes;
1712
1713         return lwb;
1714 }
1715
1716 /*
1717  * Set up the argument/result storage required for the RPC call.
1718  */
1719 void nfs_init_commit(struct nfs_commit_data *data,
1720                      struct list_head *head,
1721                      struct pnfs_layout_segment *lseg,
1722                      struct nfs_commit_info *cinfo)
1723 {
1724         struct nfs_page *first = nfs_list_entry(head->next);
1725         struct inode *inode = d_inode(first->wb_context->dentry);
1726
1727         /* Set up the RPC argument and reply structs
1728          * NB: take care not to mess about with data->commit et al. */
1729
1730         list_splice_init(head, &data->pages);
1731
1732         data->inode       = inode;
1733         data->cred        = first->wb_context->cred;
1734         data->lseg        = lseg; /* reference transferred */
1735         /* only set lwb for pnfs commit */
1736         if (lseg)
1737                 data->lwb = nfs_get_lwb(&data->pages);
1738         data->mds_ops     = &nfs_commit_ops;
1739         data->completion_ops = cinfo->completion_ops;
1740         data->dreq        = cinfo->dreq;
1741
1742         data->args.fh     = NFS_FH(data->inode);
1743         /* Note: we always request a commit of the entire inode */
1744         data->args.offset = 0;
1745         data->args.count  = 0;
1746         data->context     = get_nfs_open_context(first->wb_context);
1747         data->res.fattr   = &data->fattr;
1748         data->res.verf    = &data->verf;
1749         nfs_fattr_init(&data->fattr);
1750 }
1751 EXPORT_SYMBOL_GPL(nfs_init_commit);
1752
1753 void nfs_retry_commit(struct list_head *page_list,
1754                       struct pnfs_layout_segment *lseg,
1755                       struct nfs_commit_info *cinfo,
1756                       u32 ds_commit_idx)
1757 {
1758         struct nfs_page *req;
1759
1760         while (!list_empty(page_list)) {
1761                 req = nfs_list_entry(page_list->next);
1762                 nfs_list_remove_request(req);
1763                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1764                 if (!cinfo->dreq)
1765                         nfs_clear_page_commit(req->wb_page);
1766                 nfs_unlock_and_release_request(req);
1767         }
1768 }
1769 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1770
1771 static void
1772 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1773                 struct nfs_page *req)
1774 {
1775         __set_page_dirty_nobuffers(req->wb_page);
1776 }
1777
1778 /*
1779  * Commit dirty pages
1780  */
1781 static int
1782 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1783                 struct nfs_commit_info *cinfo)
1784 {
1785         struct nfs_commit_data  *data;
1786
1787         /* another commit raced with us */
1788         if (list_empty(head))
1789                 return 0;
1790
1791         data = nfs_commitdata_alloc(true);
1792
1793         /* Set up the argument struct */
1794         nfs_init_commit(data, head, NULL, cinfo);
1795         atomic_inc(&cinfo->mds->rpcs_out);
1796         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1797                                    data->mds_ops, how, 0);
1798 }
1799
1800 /*
1801  * COMMIT call returned
1802  */
1803 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1804 {
1805         struct nfs_commit_data  *data = calldata;
1806
1807         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1808                                 task->tk_pid, task->tk_status);
1809
1810         /* Call the NFS version-specific code */
1811         NFS_PROTO(data->inode)->commit_done(task, data);
1812         trace_nfs_commit_done(data);
1813 }
1814
1815 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1816 {
1817         struct nfs_page *req;
1818         int status = data->task.tk_status;
1819         struct nfs_commit_info cinfo;
1820         struct nfs_server *nfss;
1821
1822         while (!list_empty(&data->pages)) {
1823                 req = nfs_list_entry(data->pages.next);
1824                 nfs_list_remove_request(req);
1825                 if (req->wb_page)
1826                         nfs_clear_page_commit(req->wb_page);
1827
1828                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1829                         req->wb_context->dentry->d_sb->s_id,
1830                         (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
1831                         req->wb_bytes,
1832                         (long long)req_offset(req));
1833                 if (status < 0) {
1834                         nfs_context_set_write_error(req->wb_context, status);
1835                         if (req->wb_page)
1836                                 nfs_inode_remove_request(req);
1837                         dprintk_cont(", error = %d\n", status);
1838                         goto next;
1839                 }
1840
1841                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1842                  * returned by the server against all stored verfs. */
1843                 if (!nfs_write_verifier_cmp(&req->wb_verf, &data->verf.verifier)) {
1844                         /* We have a match */
1845                         if (req->wb_page)
1846                                 nfs_inode_remove_request(req);
1847                         dprintk_cont(" OK\n");
1848                         goto next;
1849                 }
1850                 /* We have a mismatch. Write the page again */
1851                 dprintk_cont(" mismatch\n");
1852                 nfs_mark_request_dirty(req);
1853                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1854         next:
1855                 nfs_unlock_and_release_request(req);
1856                 /* Latency breaker */
1857                 cond_resched();
1858         }
1859         nfss = NFS_SERVER(data->inode);
1860         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1861                 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1862
1863         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1864         nfs_commit_end(cinfo.mds);
1865 }
1866
1867 static void nfs_commit_release(void *calldata)
1868 {
1869         struct nfs_commit_data *data = calldata;
1870
1871         data->completion_ops->completion(data);
1872         nfs_commitdata_release(calldata);
1873 }
1874
1875 static const struct rpc_call_ops nfs_commit_ops = {
1876         .rpc_call_prepare = nfs_commit_prepare,
1877         .rpc_call_done = nfs_commit_done,
1878         .rpc_release = nfs_commit_release,
1879 };
1880
1881 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1882         .completion = nfs_commit_release_pages,
1883         .resched_write = nfs_commit_resched_write,
1884 };
1885
1886 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1887                             int how, struct nfs_commit_info *cinfo)
1888 {
1889         int status;
1890
1891         status = pnfs_commit_list(inode, head, how, cinfo);
1892         if (status == PNFS_NOT_ATTEMPTED)
1893                 status = nfs_commit_list(inode, head, how, cinfo);
1894         return status;
1895 }
1896
1897 static int __nfs_commit_inode(struct inode *inode, int how,
1898                 struct writeback_control *wbc)
1899 {
1900         LIST_HEAD(head);
1901         struct nfs_commit_info cinfo;
1902         int may_wait = how & FLUSH_SYNC;
1903         int ret, nscan;
1904
1905         nfs_init_cinfo_from_inode(&cinfo, inode);
1906         nfs_commit_begin(cinfo.mds);
1907         for (;;) {
1908                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1909                 if (ret <= 0)
1910                         break;
1911                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1912                 if (ret < 0)
1913                         break;
1914                 ret = 0;
1915                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1916                         if (nscan < wbc->nr_to_write)
1917                                 wbc->nr_to_write -= nscan;
1918                         else
1919                                 wbc->nr_to_write = 0;
1920                 }
1921                 if (nscan < INT_MAX)
1922                         break;
1923                 cond_resched();
1924         }
1925         nfs_commit_end(cinfo.mds);
1926         if (ret || !may_wait)
1927                 return ret;
1928         return wait_on_commit(cinfo.mds);
1929 }
1930
1931 int nfs_commit_inode(struct inode *inode, int how)
1932 {
1933         return __nfs_commit_inode(inode, how, NULL);
1934 }
1935 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1936
1937 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1938 {
1939         struct nfs_inode *nfsi = NFS_I(inode);
1940         int flags = FLUSH_SYNC;
1941         int ret = 0;
1942
1943         if (wbc->sync_mode == WB_SYNC_NONE) {
1944                 /* no commits means nothing needs to be done */
1945                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1946                         goto check_requests_outstanding;
1947
1948                 /* Don't commit yet if this is a non-blocking flush and there
1949                  * are a lot of outstanding writes for this mapping.
1950                  */
1951                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1952                         goto out_mark_dirty;
1953
1954                 /* don't wait for the COMMIT response */
1955                 flags = 0;
1956         }
1957
1958         ret = __nfs_commit_inode(inode, flags, wbc);
1959         if (!ret) {
1960                 if (flags & FLUSH_SYNC)
1961                         return 0;
1962         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1963                 goto out_mark_dirty;
1964
1965 check_requests_outstanding:
1966         if (!atomic_read(&nfsi->commit_info.rpcs_out))
1967                 return ret;
1968 out_mark_dirty:
1969         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1970         return ret;
1971 }
1972 EXPORT_SYMBOL_GPL(nfs_write_inode);
1973
1974 /*
1975  * Wrapper for filemap_write_and_wait_range()
1976  *
1977  * Needed for pNFS in order to ensure data becomes visible to the
1978  * client.
1979  */
1980 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1981                 loff_t lstart, loff_t lend)
1982 {
1983         int ret;
1984
1985         ret = filemap_write_and_wait_range(mapping, lstart, lend);
1986         if (ret == 0)
1987                 ret = pnfs_sync_inode(mapping->host, true);
1988         return ret;
1989 }
1990 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
1991
1992 /*
1993  * flush the inode to disk.
1994  */
1995 int nfs_wb_all(struct inode *inode)
1996 {
1997         int ret;
1998
1999         trace_nfs_writeback_inode_enter(inode);
2000
2001         ret = filemap_write_and_wait(inode->i_mapping);
2002         if (ret)
2003                 goto out;
2004         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2005         if (ret < 0)
2006                 goto out;
2007         pnfs_sync_inode(inode, true);
2008         ret = 0;
2009
2010 out:
2011         trace_nfs_writeback_inode_exit(inode, ret);
2012         return ret;
2013 }
2014 EXPORT_SYMBOL_GPL(nfs_wb_all);
2015
2016 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2017 {
2018         struct nfs_page *req;
2019         int ret = 0;
2020
2021         wait_on_page_writeback(page);
2022
2023         /* blocking call to cancel all requests and join to a single (head)
2024          * request */
2025         req = nfs_lock_and_join_requests(page);
2026
2027         if (IS_ERR(req)) {
2028                 ret = PTR_ERR(req);
2029         } else if (req) {
2030                 /* all requests from this page have been cancelled by
2031                  * nfs_lock_and_join_requests, so just remove the head
2032                  * request from the inode / page_private pointer and
2033                  * release it */
2034                 nfs_inode_remove_request(req);
2035                 nfs_unlock_and_release_request(req);
2036         }
2037
2038         return ret;
2039 }
2040
2041 /*
2042  * Write back all requests on one page - we do this before reading it.
2043  */
2044 int nfs_wb_page(struct inode *inode, struct page *page)
2045 {
2046         loff_t range_start = page_file_offset(page);
2047         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2048         struct writeback_control wbc = {
2049                 .sync_mode = WB_SYNC_ALL,
2050                 .nr_to_write = 0,
2051                 .range_start = range_start,
2052                 .range_end = range_end,
2053         };
2054         int ret;
2055
2056         trace_nfs_writeback_page_enter(inode);
2057
2058         for (;;) {
2059                 wait_on_page_writeback(page);
2060                 if (clear_page_dirty_for_io(page)) {
2061                         ret = nfs_writepage_locked(page, &wbc);
2062                         if (ret < 0)
2063                                 goto out_error;
2064                         continue;
2065                 }
2066                 ret = 0;
2067                 if (!PagePrivate(page))
2068                         break;
2069                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2070                 if (ret < 0)
2071                         goto out_error;
2072         }
2073 out_error:
2074         trace_nfs_writeback_page_exit(inode, ret);
2075         return ret;
2076 }
2077
2078 #ifdef CONFIG_MIGRATION
2079 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2080                 struct page *page, enum migrate_mode mode)
2081 {
2082         /*
2083          * If PagePrivate is set, then the page is currently associated with
2084          * an in-progress read or write request. Don't try to migrate it.
2085          *
2086          * FIXME: we could do this in principle, but we'll need a way to ensure
2087          *        that we can safely release the inode reference while holding
2088          *        the page lock.
2089          */
2090         if (PagePrivate(page))
2091                 return -EBUSY;
2092
2093         if (!nfs_fscache_release_page(page, GFP_KERNEL))
2094                 return -EBUSY;
2095
2096         return migrate_page(mapping, newpage, page, mode);
2097 }
2098 #endif
2099
2100 int __init nfs_init_writepagecache(void)
2101 {
2102         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2103                                              sizeof(struct nfs_pgio_header),
2104                                              0, SLAB_HWCACHE_ALIGN,
2105                                              NULL);
2106         if (nfs_wdata_cachep == NULL)
2107                 return -ENOMEM;
2108
2109         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2110                                                      nfs_wdata_cachep);
2111         if (nfs_wdata_mempool == NULL)
2112                 goto out_destroy_write_cache;
2113
2114         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2115                                              sizeof(struct nfs_commit_data),
2116                                              0, SLAB_HWCACHE_ALIGN,
2117                                              NULL);
2118         if (nfs_cdata_cachep == NULL)
2119                 goto out_destroy_write_mempool;
2120
2121         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2122                                                       nfs_cdata_cachep);
2123         if (nfs_commit_mempool == NULL)
2124                 goto out_destroy_commit_cache;
2125
2126         /*
2127          * NFS congestion size, scale with available memory.
2128          *
2129          *  64MB:    8192k
2130          * 128MB:   11585k
2131          * 256MB:   16384k
2132          * 512MB:   23170k
2133          *   1GB:   32768k
2134          *   2GB:   46340k
2135          *   4GB:   65536k
2136          *   8GB:   92681k
2137          *  16GB:  131072k
2138          *
2139          * This allows larger machines to have larger/more transfers.
2140          * Limit the default to 256M
2141          */
2142         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2143         if (nfs_congestion_kb > 256*1024)
2144                 nfs_congestion_kb = 256*1024;
2145
2146         return 0;
2147
2148 out_destroy_commit_cache:
2149         kmem_cache_destroy(nfs_cdata_cachep);
2150 out_destroy_write_mempool:
2151         mempool_destroy(nfs_wdata_mempool);
2152 out_destroy_write_cache:
2153         kmem_cache_destroy(nfs_wdata_cachep);
2154         return -ENOMEM;
2155 }
2156
2157 void nfs_destroy_writepagecache(void)
2158 {
2159         mempool_destroy(nfs_commit_mempool);
2160         kmem_cache_destroy(nfs_cdata_cachep);
2161         mempool_destroy(nfs_wdata_mempool);
2162         kmem_cache_destroy(nfs_wdata_cachep);
2163 }
2164
2165 static const struct nfs_rw_ops nfs_rw_write_ops = {
2166         .rw_alloc_header        = nfs_writehdr_alloc,
2167         .rw_free_header         = nfs_writehdr_free,
2168         .rw_done                = nfs_writeback_done,
2169         .rw_result              = nfs_writeback_result,
2170         .rw_initiate            = nfs_initiate_write,
2171 };