GNU Linux-libre 4.19.211-gnu1
[releases.git] / fs / nfs / write.c
1 /*
2  * linux/fs/nfs/write.c
3  *
4  * Write file data over NFS.
5  *
6  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
7  */
8
9 #include <linux/types.h>
10 #include <linux/slab.h>
11 #include <linux/mm.h>
12 #include <linux/pagemap.h>
13 #include <linux/file.h>
14 #include <linux/writeback.h>
15 #include <linux/swap.h>
16 #include <linux/migrate.h>
17
18 #include <linux/sunrpc/clnt.h>
19 #include <linux/nfs_fs.h>
20 #include <linux/nfs_mount.h>
21 #include <linux/nfs_page.h>
22 #include <linux/backing-dev.h>
23 #include <linux/export.h>
24 #include <linux/freezer.h>
25 #include <linux/wait.h>
26 #include <linux/iversion.h>
27
28 #include <linux/uaccess.h>
29
30 #include "delegation.h"
31 #include "internal.h"
32 #include "iostat.h"
33 #include "nfs4_fs.h"
34 #include "fscache.h"
35 #include "pnfs.h"
36
37 #include "nfstrace.h"
38
39 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
40
41 #define MIN_POOL_WRITE          (32)
42 #define MIN_POOL_COMMIT         (4)
43
44 struct nfs_io_completion {
45         void (*complete)(void *data);
46         void *data;
47         struct kref refcount;
48 };
49
50 /*
51  * Local function declarations
52  */
53 static void nfs_redirty_request(struct nfs_page *req);
54 static const struct rpc_call_ops nfs_commit_ops;
55 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
56 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
57 static const struct nfs_rw_ops nfs_rw_write_ops;
58 static void nfs_clear_request_commit(struct nfs_page *req);
59 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
60                                       struct inode *inode);
61 static struct nfs_page *
62 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
63                                                 struct page *page);
64
65 static struct kmem_cache *nfs_wdata_cachep;
66 static mempool_t *nfs_wdata_mempool;
67 static struct kmem_cache *nfs_cdata_cachep;
68 static mempool_t *nfs_commit_mempool;
69
70 struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
71 {
72         struct nfs_commit_data *p;
73
74         if (never_fail)
75                 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
76         else {
77                 /* It is OK to do some reclaim, not no safe to wait
78                  * for anything to be returned to the pool.
79                  * mempool_alloc() cannot handle that particular combination,
80                  * so we need two separate attempts.
81                  */
82                 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
83                 if (!p)
84                         p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
85                                              __GFP_NOWARN | __GFP_NORETRY);
86                 if (!p)
87                         return NULL;
88         }
89
90         memset(p, 0, sizeof(*p));
91         INIT_LIST_HEAD(&p->pages);
92         return p;
93 }
94 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
95
96 void nfs_commit_free(struct nfs_commit_data *p)
97 {
98         mempool_free(p, nfs_commit_mempool);
99 }
100 EXPORT_SYMBOL_GPL(nfs_commit_free);
101
102 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
103 {
104         struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
105
106         memset(p, 0, sizeof(*p));
107         p->rw_mode = FMODE_WRITE;
108         return p;
109 }
110
111 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
112 {
113         mempool_free(hdr, nfs_wdata_mempool);
114 }
115
116 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
117 {
118         return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
119 }
120
121 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
122                 void (*complete)(void *), void *data)
123 {
124         ioc->complete = complete;
125         ioc->data = data;
126         kref_init(&ioc->refcount);
127 }
128
129 static void nfs_io_completion_release(struct kref *kref)
130 {
131         struct nfs_io_completion *ioc = container_of(kref,
132                         struct nfs_io_completion, refcount);
133         ioc->complete(ioc->data);
134         kfree(ioc);
135 }
136
137 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
138 {
139         if (ioc != NULL)
140                 kref_get(&ioc->refcount);
141 }
142
143 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
144 {
145         if (ioc != NULL)
146                 kref_put(&ioc->refcount, nfs_io_completion_release);
147 }
148
149 static struct nfs_page *
150 nfs_page_private_request(struct page *page)
151 {
152         if (!PagePrivate(page))
153                 return NULL;
154         return (struct nfs_page *)page_private(page);
155 }
156
157 /*
158  * nfs_page_find_head_request_locked - find head request associated with @page
159  *
160  * must be called while holding the inode lock.
161  *
162  * returns matching head request with reference held, or NULL if not found.
163  */
164 static struct nfs_page *
165 nfs_page_find_private_request(struct page *page)
166 {
167         struct address_space *mapping = page_file_mapping(page);
168         struct nfs_page *req;
169
170         if (!PagePrivate(page))
171                 return NULL;
172         spin_lock(&mapping->private_lock);
173         req = nfs_page_private_request(page);
174         if (req) {
175                 WARN_ON_ONCE(req->wb_head != req);
176                 kref_get(&req->wb_kref);
177         }
178         spin_unlock(&mapping->private_lock);
179         return req;
180 }
181
182 static struct nfs_page *
183 nfs_page_find_swap_request(struct page *page)
184 {
185         struct inode *inode = page_file_mapping(page)->host;
186         struct nfs_inode *nfsi = NFS_I(inode);
187         struct nfs_page *req = NULL;
188         if (!PageSwapCache(page))
189                 return NULL;
190         mutex_lock(&nfsi->commit_mutex);
191         if (PageSwapCache(page)) {
192                 req = nfs_page_search_commits_for_head_request_locked(nfsi,
193                         page);
194                 if (req) {
195                         WARN_ON_ONCE(req->wb_head != req);
196                         kref_get(&req->wb_kref);
197                 }
198         }
199         mutex_unlock(&nfsi->commit_mutex);
200         return req;
201 }
202
203 /*
204  * nfs_page_find_head_request - find head request associated with @page
205  *
206  * returns matching head request with reference held, or NULL if not found.
207  */
208 static struct nfs_page *nfs_page_find_head_request(struct page *page)
209 {
210         struct nfs_page *req;
211
212         req = nfs_page_find_private_request(page);
213         if (!req)
214                 req = nfs_page_find_swap_request(page);
215         return req;
216 }
217
218 /* Adjust the file length if we're writing beyond the end */
219 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
220 {
221         struct inode *inode = page_file_mapping(page)->host;
222         loff_t end, i_size;
223         pgoff_t end_index;
224
225         spin_lock(&inode->i_lock);
226         i_size = i_size_read(inode);
227         end_index = (i_size - 1) >> PAGE_SHIFT;
228         if (i_size > 0 && page_index(page) < end_index)
229                 goto out;
230         end = page_file_offset(page) + ((loff_t)offset+count);
231         if (i_size >= end)
232                 goto out;
233         i_size_write(inode, end);
234         NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
235         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
236 out:
237         spin_unlock(&inode->i_lock);
238 }
239
240 /* A writeback failed: mark the page as bad, and invalidate the page cache */
241 static void nfs_set_pageerror(struct address_space *mapping)
242 {
243         struct inode *inode = mapping->host;
244
245         nfs_zap_mapping(mapping->host, mapping);
246         /* Force file size revalidation */
247         spin_lock(&inode->i_lock);
248         NFS_I(inode)->cache_validity |= NFS_INO_REVAL_FORCED |
249                                         NFS_INO_REVAL_PAGECACHE |
250                                         NFS_INO_INVALID_SIZE;
251         spin_unlock(&inode->i_lock);
252 }
253
254 /*
255  * nfs_page_group_search_locked
256  * @head - head request of page group
257  * @page_offset - offset into page
258  *
259  * Search page group with head @head to find a request that contains the
260  * page offset @page_offset.
261  *
262  * Returns a pointer to the first matching nfs request, or NULL if no
263  * match is found.
264  *
265  * Must be called with the page group lock held
266  */
267 static struct nfs_page *
268 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
269 {
270         struct nfs_page *req;
271
272         req = head;
273         do {
274                 if (page_offset >= req->wb_pgbase &&
275                     page_offset < (req->wb_pgbase + req->wb_bytes))
276                         return req;
277
278                 req = req->wb_this_page;
279         } while (req != head);
280
281         return NULL;
282 }
283
284 /*
285  * nfs_page_group_covers_page
286  * @head - head request of page group
287  *
288  * Return true if the page group with head @head covers the whole page,
289  * returns false otherwise
290  */
291 static bool nfs_page_group_covers_page(struct nfs_page *req)
292 {
293         struct nfs_page *tmp;
294         unsigned int pos = 0;
295         unsigned int len = nfs_page_length(req->wb_page);
296
297         nfs_page_group_lock(req);
298
299         for (;;) {
300                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
301                 if (!tmp)
302                         break;
303                 pos = tmp->wb_pgbase + tmp->wb_bytes;
304         }
305
306         nfs_page_group_unlock(req);
307         return pos >= len;
308 }
309
310 /* We can set the PG_uptodate flag if we see that a write request
311  * covers the full page.
312  */
313 static void nfs_mark_uptodate(struct nfs_page *req)
314 {
315         if (PageUptodate(req->wb_page))
316                 return;
317         if (!nfs_page_group_covers_page(req))
318                 return;
319         SetPageUptodate(req->wb_page);
320 }
321
322 static int wb_priority(struct writeback_control *wbc)
323 {
324         int ret = 0;
325
326         if (wbc->sync_mode == WB_SYNC_ALL)
327                 ret = FLUSH_COND_STABLE;
328         return ret;
329 }
330
331 /*
332  * NFS congestion control
333  */
334
335 int nfs_congestion_kb;
336
337 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
338 #define NFS_CONGESTION_OFF_THRESH       \
339         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
340
341 static void nfs_set_page_writeback(struct page *page)
342 {
343         struct inode *inode = page_file_mapping(page)->host;
344         struct nfs_server *nfss = NFS_SERVER(inode);
345         int ret = test_set_page_writeback(page);
346
347         WARN_ON_ONCE(ret != 0);
348
349         if (atomic_long_inc_return(&nfss->writeback) >
350                         NFS_CONGESTION_ON_THRESH)
351                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
352 }
353
354 static void nfs_end_page_writeback(struct nfs_page *req)
355 {
356         struct inode *inode = page_file_mapping(req->wb_page)->host;
357         struct nfs_server *nfss = NFS_SERVER(inode);
358         bool is_done;
359
360         is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
361         nfs_unlock_request(req);
362         if (!is_done)
363                 return;
364
365         end_page_writeback(req->wb_page);
366         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
367                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
368 }
369
370 /*
371  * nfs_unroll_locks_and_wait -  unlock all newly locked reqs and wait on @req
372  *
373  * this is a helper function for nfs_lock_and_join_requests
374  *
375  * @inode - inode associated with request page group, must be holding inode lock
376  * @head  - head request of page group, must be holding head lock
377  * @req   - request that couldn't lock and needs to wait on the req bit lock
378  *
379  * NOTE: this must be called holding page_group bit lock
380  *       which will be released before returning.
381  *
382  * returns 0 on success, < 0 on error.
383  */
384 static void
385 nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
386                           struct nfs_page *req)
387 {
388         struct nfs_page *tmp;
389
390         /* relinquish all the locks successfully grabbed this run */
391         for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
392                 if (!kref_read(&tmp->wb_kref))
393                         continue;
394                 nfs_unlock_and_release_request(tmp);
395         }
396 }
397
398 /*
399  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
400  *
401  * @destroy_list - request list (using wb_this_page) terminated by @old_head
402  * @old_head - the old head of the list
403  *
404  * All subrequests must be locked and removed from all lists, so at this point
405  * they are only "active" in this function, and possibly in nfs_wait_on_request
406  * with a reference held by some other context.
407  */
408 static void
409 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
410                                  struct nfs_page *old_head,
411                                  struct inode *inode)
412 {
413         while (destroy_list) {
414                 struct nfs_page *subreq = destroy_list;
415
416                 destroy_list = (subreq->wb_this_page == old_head) ?
417                                    NULL : subreq->wb_this_page;
418
419                 /* Note: lock subreq in order to change subreq->wb_head */
420                 nfs_page_set_headlock(subreq);
421                 WARN_ON_ONCE(old_head != subreq->wb_head);
422
423                 /* make sure old group is not used */
424                 subreq->wb_this_page = subreq;
425                 subreq->wb_head = subreq;
426
427                 clear_bit(PG_REMOVE, &subreq->wb_flags);
428
429                 /* Note: races with nfs_page_group_destroy() */
430                 if (!kref_read(&subreq->wb_kref)) {
431                         /* Check if we raced with nfs_page_group_destroy() */
432                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
433                                 nfs_page_clear_headlock(subreq);
434                                 nfs_free_request(subreq);
435                         } else
436                                 nfs_page_clear_headlock(subreq);
437                         continue;
438                 }
439                 nfs_page_clear_headlock(subreq);
440
441                 nfs_release_request(old_head);
442
443                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
444                         nfs_release_request(subreq);
445                         atomic_long_dec(&NFS_I(inode)->nrequests);
446                 }
447
448                 /* subreq is now totally disconnected from page group or any
449                  * write / commit lists. last chance to wake any waiters */
450                 nfs_unlock_and_release_request(subreq);
451         }
452 }
453
454 /*
455  * nfs_lock_and_join_requests - join all subreqs to the head req and return
456  *                              a locked reference, cancelling any pending
457  *                              operations for this page.
458  *
459  * @page - the page used to lookup the "page group" of nfs_page structures
460  *
461  * This function joins all sub requests to the head request by first
462  * locking all requests in the group, cancelling any pending operations
463  * and finally updating the head request to cover the whole range covered by
464  * the (former) group.  All subrequests are removed from any write or commit
465  * lists, unlinked from the group and destroyed.
466  *
467  * Returns a locked, referenced pointer to the head request - which after
468  * this call is guaranteed to be the only request associated with the page.
469  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
470  * error was encountered.
471  */
472 static struct nfs_page *
473 nfs_lock_and_join_requests(struct page *page)
474 {
475         struct inode *inode = page_file_mapping(page)->host;
476         struct nfs_page *head, *subreq;
477         struct nfs_page *destroy_list = NULL;
478         unsigned int total_bytes;
479         int ret;
480
481 try_again:
482         /*
483          * A reference is taken only on the head request which acts as a
484          * reference to the whole page group - the group will not be destroyed
485          * until the head reference is released.
486          */
487         head = nfs_page_find_head_request(page);
488         if (!head)
489                 return NULL;
490
491         /* lock the page head first in order to avoid an ABBA inefficiency */
492         if (!nfs_lock_request(head)) {
493                 ret = nfs_wait_on_request(head);
494                 nfs_release_request(head);
495                 if (ret < 0)
496                         return ERR_PTR(ret);
497                 goto try_again;
498         }
499
500         /* Ensure that nobody removed the request before we locked it */
501         if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
502                 nfs_unlock_and_release_request(head);
503                 goto try_again;
504         }
505
506         ret = nfs_page_group_lock(head);
507         if (ret < 0)
508                 goto release_request;
509
510         /* lock each request in the page group */
511         total_bytes = head->wb_bytes;
512         for (subreq = head->wb_this_page; subreq != head;
513                         subreq = subreq->wb_this_page) {
514
515                 if (!kref_get_unless_zero(&subreq->wb_kref)) {
516                         if (subreq->wb_offset == head->wb_offset + total_bytes)
517                                 total_bytes += subreq->wb_bytes;
518                         continue;
519                 }
520
521                 while (!nfs_lock_request(subreq)) {
522                         /*
523                          * Unlock page to allow nfs_page_group_sync_on_bit()
524                          * to succeed
525                          */
526                         nfs_page_group_unlock(head);
527                         ret = nfs_wait_on_request(subreq);
528                         if (!ret)
529                                 ret = nfs_page_group_lock(head);
530                         if (ret < 0) {
531                                 nfs_unroll_locks(inode, head, subreq);
532                                 nfs_release_request(subreq);
533                                 goto release_request;
534                         }
535                 }
536                 /*
537                  * Subrequests are always contiguous, non overlapping
538                  * and in order - but may be repeated (mirrored writes).
539                  */
540                 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
541                         /* keep track of how many bytes this group covers */
542                         total_bytes += subreq->wb_bytes;
543                 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
544                             ((subreq->wb_offset + subreq->wb_bytes) >
545                              (head->wb_offset + total_bytes)))) {
546                         nfs_page_group_unlock(head);
547                         nfs_unroll_locks(inode, head, subreq);
548                         nfs_unlock_and_release_request(subreq);
549                         ret = -EIO;
550                         goto release_request;
551                 }
552         }
553
554         /* Now that all requests are locked, make sure they aren't on any list.
555          * Commit list removal accounting is done after locks are dropped */
556         subreq = head;
557         do {
558                 nfs_clear_request_commit(subreq);
559                 subreq = subreq->wb_this_page;
560         } while (subreq != head);
561
562         /* unlink subrequests from head, destroy them later */
563         if (head->wb_this_page != head) {
564                 /* destroy list will be terminated by head */
565                 destroy_list = head->wb_this_page;
566                 head->wb_this_page = head;
567
568                 /* change head request to cover whole range that
569                  * the former page group covered */
570                 head->wb_bytes = total_bytes;
571         }
572
573         /* Postpone destruction of this request */
574         if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
575                 set_bit(PG_INODE_REF, &head->wb_flags);
576                 kref_get(&head->wb_kref);
577                 atomic_long_inc(&NFS_I(inode)->nrequests);
578         }
579
580         nfs_page_group_unlock(head);
581
582         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
583
584         /* Did we lose a race with nfs_inode_remove_request()? */
585         if (!(PagePrivate(page) || PageSwapCache(page))) {
586                 nfs_unlock_and_release_request(head);
587                 return NULL;
588         }
589
590         /* still holds ref on head from nfs_page_find_head_request
591          * and still has lock on head from lock loop */
592         return head;
593
594 release_request:
595         nfs_unlock_and_release_request(head);
596         return ERR_PTR(ret);
597 }
598
599 static void nfs_write_error_remove_page(struct nfs_page *req)
600 {
601         nfs_end_page_writeback(req);
602         generic_error_remove_page(page_file_mapping(req->wb_page),
603                                   req->wb_page);
604         nfs_release_request(req);
605 }
606
607 static bool
608 nfs_error_is_fatal_on_server(int err)
609 {
610         switch (err) {
611         case 0:
612         case -ERESTARTSYS:
613         case -EINTR:
614                 return false;
615         }
616         return nfs_error_is_fatal(err);
617 }
618
619 /*
620  * Find an associated nfs write request, and prepare to flush it out
621  * May return an error if the user signalled nfs_wait_on_request().
622  */
623 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
624                                 struct page *page)
625 {
626         struct nfs_page *req;
627         int ret = 0;
628
629         req = nfs_lock_and_join_requests(page);
630         if (!req)
631                 goto out;
632         ret = PTR_ERR(req);
633         if (IS_ERR(req))
634                 goto out;
635
636         nfs_set_page_writeback(page);
637         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
638
639         ret = req->wb_context->error;
640         /* If there is a fatal error that covers this write, just exit */
641         if (nfs_error_is_fatal_on_server(ret))
642                 goto out_launder;
643
644         ret = 0;
645         if (!nfs_pageio_add_request(pgio, req)) {
646                 ret = pgio->pg_error;
647                 /*
648                  * Remove the problematic req upon fatal errors on the server
649                  */
650                 if (nfs_error_is_fatal(ret)) {
651                         nfs_context_set_write_error(req->wb_context, ret);
652                         if (nfs_error_is_fatal_on_server(ret))
653                                 goto out_launder;
654                 } else
655                         ret = -EAGAIN;
656                 nfs_redirty_request(req);
657         } else
658                 nfs_add_stats(page_file_mapping(page)->host,
659                                 NFSIOS_WRITEPAGES, 1);
660 out:
661         return ret;
662 out_launder:
663         nfs_write_error_remove_page(req);
664         return 0;
665 }
666
667 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
668                             struct nfs_pageio_descriptor *pgio)
669 {
670         int ret;
671
672         nfs_pageio_cond_complete(pgio, page_index(page));
673         ret = nfs_page_async_flush(pgio, page);
674         if (ret == -EAGAIN) {
675                 redirty_page_for_writepage(wbc, page);
676                 ret = 0;
677         }
678         return ret;
679 }
680
681 /*
682  * Write an mmapped page to the server.
683  */
684 static int nfs_writepage_locked(struct page *page,
685                                 struct writeback_control *wbc)
686 {
687         struct nfs_pageio_descriptor pgio;
688         struct inode *inode = page_file_mapping(page)->host;
689         int err;
690
691         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
692         nfs_pageio_init_write(&pgio, inode, 0,
693                                 false, &nfs_async_write_completion_ops);
694         err = nfs_do_writepage(page, wbc, &pgio);
695         nfs_pageio_complete(&pgio);
696         if (err < 0)
697                 return err;
698         if (pgio.pg_error < 0)
699                 return pgio.pg_error;
700         return 0;
701 }
702
703 int nfs_writepage(struct page *page, struct writeback_control *wbc)
704 {
705         int ret;
706
707         ret = nfs_writepage_locked(page, wbc);
708         unlock_page(page);
709         return ret;
710 }
711
712 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
713 {
714         int ret;
715
716         ret = nfs_do_writepage(page, wbc, data);
717         unlock_page(page);
718         return ret;
719 }
720
721 static void nfs_io_completion_commit(void *inode)
722 {
723         nfs_commit_inode(inode, 0);
724 }
725
726 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
727 {
728         struct inode *inode = mapping->host;
729         struct nfs_pageio_descriptor pgio;
730         struct nfs_io_completion *ioc = nfs_io_completion_alloc(GFP_NOFS);
731         int err;
732
733         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
734
735         if (ioc)
736                 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
737
738         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
739                                 &nfs_async_write_completion_ops);
740         pgio.pg_io_completion = ioc;
741         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
742         nfs_pageio_complete(&pgio);
743         nfs_io_completion_put(ioc);
744
745         if (err < 0)
746                 goto out_err;
747         err = pgio.pg_error;
748         if (err < 0)
749                 goto out_err;
750         return 0;
751 out_err:
752         return err;
753 }
754
755 /*
756  * Insert a write request into an inode
757  */
758 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
759 {
760         struct address_space *mapping = page_file_mapping(req->wb_page);
761         struct nfs_inode *nfsi = NFS_I(inode);
762
763         WARN_ON_ONCE(req->wb_this_page != req);
764
765         /* Lock the request! */
766         nfs_lock_request(req);
767
768         /*
769          * Swap-space should not get truncated. Hence no need to plug the race
770          * with invalidate/truncate.
771          */
772         spin_lock(&mapping->private_lock);
773         if (!nfs_have_writebacks(inode) &&
774             NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
775                 inode_inc_iversion_raw(inode);
776         if (likely(!PageSwapCache(req->wb_page))) {
777                 set_bit(PG_MAPPED, &req->wb_flags);
778                 SetPagePrivate(req->wb_page);
779                 set_page_private(req->wb_page, (unsigned long)req);
780         }
781         spin_unlock(&mapping->private_lock);
782         atomic_long_inc(&nfsi->nrequests);
783         /* this a head request for a page group - mark it as having an
784          * extra reference so sub groups can follow suit.
785          * This flag also informs pgio layer when to bump nrequests when
786          * adding subrequests. */
787         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
788         kref_get(&req->wb_kref);
789 }
790
791 /*
792  * Remove a write request from an inode
793  */
794 static void nfs_inode_remove_request(struct nfs_page *req)
795 {
796         struct address_space *mapping = page_file_mapping(req->wb_page);
797         struct inode *inode = mapping->host;
798         struct nfs_inode *nfsi = NFS_I(inode);
799         struct nfs_page *head;
800
801         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
802                 head = req->wb_head;
803
804                 spin_lock(&mapping->private_lock);
805                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
806                         set_page_private(head->wb_page, 0);
807                         ClearPagePrivate(head->wb_page);
808                         clear_bit(PG_MAPPED, &head->wb_flags);
809                 }
810                 spin_unlock(&mapping->private_lock);
811         }
812
813         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
814                 nfs_release_request(req);
815                 atomic_long_dec(&nfsi->nrequests);
816         }
817 }
818
819 static void
820 nfs_mark_request_dirty(struct nfs_page *req)
821 {
822         if (req->wb_page)
823                 __set_page_dirty_nobuffers(req->wb_page);
824 }
825
826 /*
827  * nfs_page_search_commits_for_head_request_locked
828  *
829  * Search through commit lists on @inode for the head request for @page.
830  * Must be called while holding the inode (which is cinfo) lock.
831  *
832  * Returns the head request if found, or NULL if not found.
833  */
834 static struct nfs_page *
835 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
836                                                 struct page *page)
837 {
838         struct nfs_page *freq, *t;
839         struct nfs_commit_info cinfo;
840         struct inode *inode = &nfsi->vfs_inode;
841
842         nfs_init_cinfo_from_inode(&cinfo, inode);
843
844         /* search through pnfs commit lists */
845         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
846         if (freq)
847                 return freq->wb_head;
848
849         /* Linearly search the commit list for the correct request */
850         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
851                 if (freq->wb_page == page)
852                         return freq->wb_head;
853         }
854
855         return NULL;
856 }
857
858 /**
859  * nfs_request_add_commit_list_locked - add request to a commit list
860  * @req: pointer to a struct nfs_page
861  * @dst: commit list head
862  * @cinfo: holds list lock and accounting info
863  *
864  * This sets the PG_CLEAN bit, updates the cinfo count of
865  * number of outstanding requests requiring a commit as well as
866  * the MM page stats.
867  *
868  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
869  * nfs_page lock.
870  */
871 void
872 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
873                             struct nfs_commit_info *cinfo)
874 {
875         set_bit(PG_CLEAN, &req->wb_flags);
876         nfs_list_add_request(req, dst);
877         atomic_long_inc(&cinfo->mds->ncommit);
878 }
879 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
880
881 /**
882  * nfs_request_add_commit_list - add request to a commit list
883  * @req: pointer to a struct nfs_page
884  * @dst: commit list head
885  * @cinfo: holds list lock and accounting info
886  *
887  * This sets the PG_CLEAN bit, updates the cinfo count of
888  * number of outstanding requests requiring a commit as well as
889  * the MM page stats.
890  *
891  * The caller must _not_ hold the cinfo->lock, but must be
892  * holding the nfs_page lock.
893  */
894 void
895 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
896 {
897         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
898         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
899         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
900         if (req->wb_page)
901                 nfs_mark_page_unstable(req->wb_page, cinfo);
902 }
903 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
904
905 /**
906  * nfs_request_remove_commit_list - Remove request from a commit list
907  * @req: pointer to a nfs_page
908  * @cinfo: holds list lock and accounting info
909  *
910  * This clears the PG_CLEAN bit, and updates the cinfo's count of
911  * number of outstanding requests requiring a commit
912  * It does not update the MM page stats.
913  *
914  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
915  */
916 void
917 nfs_request_remove_commit_list(struct nfs_page *req,
918                                struct nfs_commit_info *cinfo)
919 {
920         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
921                 return;
922         nfs_list_remove_request(req);
923         atomic_long_dec(&cinfo->mds->ncommit);
924 }
925 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
926
927 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
928                                       struct inode *inode)
929 {
930         cinfo->inode = inode;
931         cinfo->mds = &NFS_I(inode)->commit_info;
932         cinfo->ds = pnfs_get_ds_info(inode);
933         cinfo->dreq = NULL;
934         cinfo->completion_ops = &nfs_commit_completion_ops;
935 }
936
937 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
938                     struct inode *inode,
939                     struct nfs_direct_req *dreq)
940 {
941         if (dreq)
942                 nfs_init_cinfo_from_dreq(cinfo, dreq);
943         else
944                 nfs_init_cinfo_from_inode(cinfo, inode);
945 }
946 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
947
948 /*
949  * Add a request to the inode's commit list.
950  */
951 void
952 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
953                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
954 {
955         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
956                 return;
957         nfs_request_add_commit_list(req, cinfo);
958 }
959
960 static void
961 nfs_clear_page_commit(struct page *page)
962 {
963         dec_node_page_state(page, NR_UNSTABLE_NFS);
964         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
965                     WB_RECLAIMABLE);
966 }
967
968 /* Called holding the request lock on @req */
969 static void
970 nfs_clear_request_commit(struct nfs_page *req)
971 {
972         if (test_bit(PG_CLEAN, &req->wb_flags)) {
973                 struct inode *inode = d_inode(req->wb_context->dentry);
974                 struct nfs_commit_info cinfo;
975
976                 nfs_init_cinfo_from_inode(&cinfo, inode);
977                 mutex_lock(&NFS_I(inode)->commit_mutex);
978                 if (!pnfs_clear_request_commit(req, &cinfo)) {
979                         nfs_request_remove_commit_list(req, &cinfo);
980                 }
981                 mutex_unlock(&NFS_I(inode)->commit_mutex);
982                 nfs_clear_page_commit(req->wb_page);
983         }
984 }
985
986 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
987 {
988         if (hdr->verf.committed == NFS_DATA_SYNC)
989                 return hdr->lseg == NULL;
990         return hdr->verf.committed != NFS_FILE_SYNC;
991 }
992
993 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
994 {
995         nfs_io_completion_get(hdr->io_completion);
996 }
997
998 static void nfs_write_completion(struct nfs_pgio_header *hdr)
999 {
1000         struct nfs_commit_info cinfo;
1001         unsigned long bytes = 0;
1002
1003         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
1004                 goto out;
1005         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
1006         while (!list_empty(&hdr->pages)) {
1007                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1008
1009                 bytes += req->wb_bytes;
1010                 nfs_list_remove_request(req);
1011                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1012                     (hdr->good_bytes < bytes)) {
1013                         nfs_set_pageerror(page_file_mapping(req->wb_page));
1014                         nfs_context_set_write_error(req->wb_context, hdr->error);
1015                         goto remove_req;
1016                 }
1017                 if (nfs_write_need_commit(hdr)) {
1018                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1019                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1020                                 hdr->pgio_mirror_idx);
1021                         goto next;
1022                 }
1023 remove_req:
1024                 nfs_inode_remove_request(req);
1025 next:
1026                 nfs_end_page_writeback(req);
1027                 nfs_release_request(req);
1028         }
1029 out:
1030         nfs_io_completion_put(hdr->io_completion);
1031         hdr->release(hdr);
1032 }
1033
1034 unsigned long
1035 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1036 {
1037         return atomic_long_read(&cinfo->mds->ncommit);
1038 }
1039
1040 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1041 int
1042 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1043                      struct nfs_commit_info *cinfo, int max)
1044 {
1045         struct nfs_page *req, *tmp;
1046         int ret = 0;
1047
1048 restart:
1049         list_for_each_entry_safe(req, tmp, src, wb_list) {
1050                 kref_get(&req->wb_kref);
1051                 if (!nfs_lock_request(req)) {
1052                         int status;
1053
1054                         /* Prevent deadlock with nfs_lock_and_join_requests */
1055                         if (!list_empty(dst)) {
1056                                 nfs_release_request(req);
1057                                 continue;
1058                         }
1059                         /* Ensure we make progress to prevent livelock */
1060                         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1061                         status = nfs_wait_on_request(req);
1062                         nfs_release_request(req);
1063                         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1064                         if (status < 0)
1065                                 break;
1066                         goto restart;
1067                 }
1068                 nfs_request_remove_commit_list(req, cinfo);
1069                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1070                 nfs_list_add_request(req, dst);
1071                 ret++;
1072                 if ((ret == max) && !cinfo->dreq)
1073                         break;
1074                 cond_resched();
1075         }
1076         return ret;
1077 }
1078 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1079
1080 /*
1081  * nfs_scan_commit - Scan an inode for commit requests
1082  * @inode: NFS inode to scan
1083  * @dst: mds destination list
1084  * @cinfo: mds and ds lists of reqs ready to commit
1085  *
1086  * Moves requests from the inode's 'commit' request list.
1087  * The requests are *not* checked to ensure that they form a contiguous set.
1088  */
1089 int
1090 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1091                 struct nfs_commit_info *cinfo)
1092 {
1093         int ret = 0;
1094
1095         if (!atomic_long_read(&cinfo->mds->ncommit))
1096                 return 0;
1097         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1098         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1099                 const int max = INT_MAX;
1100
1101                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1102                                            cinfo, max);
1103                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1104         }
1105         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1106         return ret;
1107 }
1108
1109 /*
1110  * Search for an existing write request, and attempt to update
1111  * it to reflect a new dirty region on a given page.
1112  *
1113  * If the attempt fails, then the existing request is flushed out
1114  * to disk.
1115  */
1116 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1117                 struct page *page,
1118                 unsigned int offset,
1119                 unsigned int bytes)
1120 {
1121         struct nfs_page *req;
1122         unsigned int rqend;
1123         unsigned int end;
1124         int error;
1125
1126         end = offset + bytes;
1127
1128         req = nfs_lock_and_join_requests(page);
1129         if (IS_ERR_OR_NULL(req))
1130                 return req;
1131
1132         rqend = req->wb_offset + req->wb_bytes;
1133         /*
1134          * Tell the caller to flush out the request if
1135          * the offsets are non-contiguous.
1136          * Note: nfs_flush_incompatible() will already
1137          * have flushed out requests having wrong owners.
1138          */
1139         if (offset > rqend || end < req->wb_offset)
1140                 goto out_flushme;
1141
1142         /* Okay, the request matches. Update the region */
1143         if (offset < req->wb_offset) {
1144                 req->wb_offset = offset;
1145                 req->wb_pgbase = offset;
1146         }
1147         if (end > rqend)
1148                 req->wb_bytes = end - req->wb_offset;
1149         else
1150                 req->wb_bytes = rqend - req->wb_offset;
1151         return req;
1152 out_flushme:
1153         /*
1154          * Note: we mark the request dirty here because
1155          * nfs_lock_and_join_requests() cannot preserve
1156          * commit flags, so we have to replay the write.
1157          */
1158         nfs_mark_request_dirty(req);
1159         nfs_unlock_and_release_request(req);
1160         error = nfs_wb_page(inode, page);
1161         return (error < 0) ? ERR_PTR(error) : NULL;
1162 }
1163
1164 /*
1165  * Try to update an existing write request, or create one if there is none.
1166  *
1167  * Note: Should always be called with the Page Lock held to prevent races
1168  * if we have to add a new request. Also assumes that the caller has
1169  * already called nfs_flush_incompatible() if necessary.
1170  */
1171 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1172                 struct page *page, unsigned int offset, unsigned int bytes)
1173 {
1174         struct inode *inode = page_file_mapping(page)->host;
1175         struct nfs_page *req;
1176
1177         req = nfs_try_to_update_request(inode, page, offset, bytes);
1178         if (req != NULL)
1179                 goto out;
1180         req = nfs_create_request(ctx, page, NULL, offset, bytes);
1181         if (IS_ERR(req))
1182                 goto out;
1183         nfs_inode_add_request(inode, req);
1184 out:
1185         return req;
1186 }
1187
1188 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1189                 unsigned int offset, unsigned int count)
1190 {
1191         struct nfs_page *req;
1192
1193         req = nfs_setup_write_request(ctx, page, offset, count);
1194         if (IS_ERR(req))
1195                 return PTR_ERR(req);
1196         /* Update file length */
1197         nfs_grow_file(page, offset, count);
1198         nfs_mark_uptodate(req);
1199         nfs_mark_request_dirty(req);
1200         nfs_unlock_and_release_request(req);
1201         return 0;
1202 }
1203
1204 int nfs_flush_incompatible(struct file *file, struct page *page)
1205 {
1206         struct nfs_open_context *ctx = nfs_file_open_context(file);
1207         struct nfs_lock_context *l_ctx;
1208         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1209         struct nfs_page *req;
1210         int do_flush, status;
1211         /*
1212          * Look for a request corresponding to this page. If there
1213          * is one, and it belongs to another file, we flush it out
1214          * before we try to copy anything into the page. Do this
1215          * due to the lack of an ACCESS-type call in NFSv2.
1216          * Also do the same if we find a request from an existing
1217          * dropped page.
1218          */
1219         do {
1220                 req = nfs_page_find_head_request(page);
1221                 if (req == NULL)
1222                         return 0;
1223                 l_ctx = req->wb_lock_context;
1224                 do_flush = req->wb_page != page ||
1225                         !nfs_match_open_context(req->wb_context, ctx);
1226                 if (l_ctx && flctx &&
1227                     !(list_empty_careful(&flctx->flc_posix) &&
1228                       list_empty_careful(&flctx->flc_flock))) {
1229                         do_flush |= l_ctx->lockowner != current->files;
1230                 }
1231                 nfs_release_request(req);
1232                 if (!do_flush)
1233                         return 0;
1234                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1235         } while (status == 0);
1236         return status;
1237 }
1238
1239 /*
1240  * Avoid buffered writes when a open context credential's key would
1241  * expire soon.
1242  *
1243  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1244  *
1245  * Return 0 and set a credential flag which triggers the inode to flush
1246  * and performs  NFS_FILE_SYNC writes if the key will expired within
1247  * RPC_KEY_EXPIRE_TIMEO.
1248  */
1249 int
1250 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1251 {
1252         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1253         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1254
1255         return rpcauth_key_timeout_notify(auth, ctx->cred);
1256 }
1257
1258 /*
1259  * Test if the open context credential key is marked to expire soon.
1260  */
1261 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1262 {
1263         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1264
1265         return rpcauth_cred_key_to_expire(auth, ctx->cred);
1266 }
1267
1268 /*
1269  * If the page cache is marked as unsafe or invalid, then we can't rely on
1270  * the PageUptodate() flag. In this case, we will need to turn off
1271  * write optimisations that depend on the page contents being correct.
1272  */
1273 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1274 {
1275         struct nfs_inode *nfsi = NFS_I(inode);
1276
1277         if (nfs_have_delegated_attributes(inode))
1278                 goto out;
1279         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1280                 return false;
1281         smp_rmb();
1282         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1283                 return false;
1284 out:
1285         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1286                 return false;
1287         return PageUptodate(page) != 0;
1288 }
1289
1290 static bool
1291 is_whole_file_wrlock(struct file_lock *fl)
1292 {
1293         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1294                         fl->fl_type == F_WRLCK;
1295 }
1296
1297 /* If we know the page is up to date, and we're not using byte range locks (or
1298  * if we have the whole file locked for writing), it may be more efficient to
1299  * extend the write to cover the entire page in order to avoid fragmentation
1300  * inefficiencies.
1301  *
1302  * If the file is opened for synchronous writes then we can just skip the rest
1303  * of the checks.
1304  */
1305 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1306 {
1307         int ret;
1308         struct file_lock_context *flctx = inode->i_flctx;
1309         struct file_lock *fl;
1310
1311         if (file->f_flags & O_DSYNC)
1312                 return 0;
1313         if (!nfs_write_pageuptodate(page, inode))
1314                 return 0;
1315         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1316                 return 1;
1317         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1318                        list_empty_careful(&flctx->flc_posix)))
1319                 return 1;
1320
1321         /* Check to see if there are whole file write locks */
1322         ret = 0;
1323         spin_lock(&flctx->flc_lock);
1324         if (!list_empty(&flctx->flc_posix)) {
1325                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1326                                         fl_list);
1327                 if (is_whole_file_wrlock(fl))
1328                         ret = 1;
1329         } else if (!list_empty(&flctx->flc_flock)) {
1330                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1331                                         fl_list);
1332                 if (fl->fl_type == F_WRLCK)
1333                         ret = 1;
1334         }
1335         spin_unlock(&flctx->flc_lock);
1336         return ret;
1337 }
1338
1339 /*
1340  * Update and possibly write a cached page of an NFS file.
1341  *
1342  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1343  * things with a page scheduled for an RPC call (e.g. invalidate it).
1344  */
1345 int nfs_updatepage(struct file *file, struct page *page,
1346                 unsigned int offset, unsigned int count)
1347 {
1348         struct nfs_open_context *ctx = nfs_file_open_context(file);
1349         struct address_space *mapping = page_file_mapping(page);
1350         struct inode    *inode = mapping->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(mapping);
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, int error)
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                 if (nfs_error_is_fatal(error)) {
1421                         nfs_context_set_write_error(req->wb_context, error);
1422                         if (nfs_error_is_fatal_on_server(error)) {
1423                                 nfs_write_error_remove_page(req);
1424                                 continue;
1425                         }
1426                 }
1427                 nfs_redirty_request(req);
1428         }
1429 }
1430
1431 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1432 {
1433         nfs_async_write_error(&hdr->pages, 0);
1434         filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1435                         hdr->args.offset + hdr->args.count - 1);
1436 }
1437
1438 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1439         .init_hdr = nfs_async_write_init,
1440         .error_cleanup = nfs_async_write_error,
1441         .completion = nfs_write_completion,
1442         .reschedule_io = nfs_async_write_reschedule_io,
1443 };
1444
1445 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1446                                struct inode *inode, int ioflags, bool force_mds,
1447                                const struct nfs_pgio_completion_ops *compl_ops)
1448 {
1449         struct nfs_server *server = NFS_SERVER(inode);
1450         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1451
1452 #ifdef CONFIG_NFS_V4_1
1453         if (server->pnfs_curr_ld && !force_mds)
1454                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1455 #endif
1456         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1457                         server->wsize, ioflags);
1458 }
1459 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1460
1461 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1462 {
1463         struct nfs_pgio_mirror *mirror;
1464
1465         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1466                 pgio->pg_ops->pg_cleanup(pgio);
1467
1468         pgio->pg_ops = &nfs_pgio_rw_ops;
1469
1470         nfs_pageio_stop_mirroring(pgio);
1471
1472         mirror = &pgio->pg_mirrors[0];
1473         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1474 }
1475 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1476
1477
1478 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1479 {
1480         struct nfs_commit_data *data = calldata;
1481
1482         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1483 }
1484
1485 /*
1486  * Special version of should_remove_suid() that ignores capabilities.
1487  */
1488 static int nfs_should_remove_suid(const struct inode *inode)
1489 {
1490         umode_t mode = inode->i_mode;
1491         int kill = 0;
1492
1493         /* suid always must be killed */
1494         if (unlikely(mode & S_ISUID))
1495                 kill = ATTR_KILL_SUID;
1496
1497         /*
1498          * sgid without any exec bits is just a mandatory locking mark; leave
1499          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1500          */
1501         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1502                 kill |= ATTR_KILL_SGID;
1503
1504         if (unlikely(kill && S_ISREG(mode)))
1505                 return kill;
1506
1507         return 0;
1508 }
1509
1510 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1511                 struct nfs_fattr *fattr)
1512 {
1513         struct nfs_pgio_args *argp = &hdr->args;
1514         struct nfs_pgio_res *resp = &hdr->res;
1515         u64 size = argp->offset + resp->count;
1516
1517         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1518                 fattr->size = size;
1519         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1520                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1521                 return;
1522         }
1523         if (size != fattr->size)
1524                 return;
1525         /* Set attribute barrier */
1526         nfs_fattr_set_barrier(fattr);
1527         /* ...and update size */
1528         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1529 }
1530
1531 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1532 {
1533         struct nfs_fattr *fattr = &hdr->fattr;
1534         struct inode *inode = hdr->inode;
1535
1536         spin_lock(&inode->i_lock);
1537         nfs_writeback_check_extend(hdr, fattr);
1538         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1539         spin_unlock(&inode->i_lock);
1540 }
1541 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1542
1543 /*
1544  * This function is called when the WRITE call is complete.
1545  */
1546 static int nfs_writeback_done(struct rpc_task *task,
1547                               struct nfs_pgio_header *hdr,
1548                               struct inode *inode)
1549 {
1550         int status;
1551
1552         /*
1553          * ->write_done will attempt to use post-op attributes to detect
1554          * conflicting writes by other clients.  A strict interpretation
1555          * of close-to-open would allow us to continue caching even if
1556          * another writer had changed the file, but some applications
1557          * depend on tighter cache coherency when writing.
1558          */
1559         status = NFS_PROTO(inode)->write_done(task, hdr);
1560         if (status != 0)
1561                 return status;
1562
1563         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1564         trace_nfs_writeback_done(inode, task->tk_status,
1565                                  hdr->args.offset, hdr->res.verf);
1566
1567         if (hdr->res.verf->committed < hdr->args.stable &&
1568             task->tk_status >= 0) {
1569                 /* We tried a write call, but the server did not
1570                  * commit data to stable storage even though we
1571                  * requested it.
1572                  * Note: There is a known bug in Tru64 < 5.0 in which
1573                  *       the server reports NFS_DATA_SYNC, but performs
1574                  *       NFS_FILE_SYNC. We therefore implement this checking
1575                  *       as a dprintk() in order to avoid filling syslog.
1576                  */
1577                 static unsigned long    complain;
1578
1579                 /* Note this will print the MDS for a DS write */
1580                 if (time_before(complain, jiffies)) {
1581                         dprintk("NFS:       faulty NFS server %s:"
1582                                 " (committed = %d) != (stable = %d)\n",
1583                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1584                                 hdr->res.verf->committed, hdr->args.stable);
1585                         complain = jiffies + 300 * HZ;
1586                 }
1587         }
1588
1589         /* Deal with the suid/sgid bit corner case */
1590         if (nfs_should_remove_suid(inode)) {
1591                 spin_lock(&inode->i_lock);
1592                 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1593                 spin_unlock(&inode->i_lock);
1594         }
1595         return 0;
1596 }
1597
1598 /*
1599  * This function is called when the WRITE call is complete.
1600  */
1601 static void nfs_writeback_result(struct rpc_task *task,
1602                                  struct nfs_pgio_header *hdr)
1603 {
1604         struct nfs_pgio_args    *argp = &hdr->args;
1605         struct nfs_pgio_res     *resp = &hdr->res;
1606
1607         if (resp->count < argp->count) {
1608                 static unsigned long    complain;
1609
1610                 /* This a short write! */
1611                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1612
1613                 /* Has the server at least made some progress? */
1614                 if (resp->count == 0) {
1615                         if (time_before(complain, jiffies)) {
1616                                 printk(KERN_WARNING
1617                                        "NFS: Server wrote zero bytes, expected %u.\n",
1618                                        argp->count);
1619                                 complain = jiffies + 300 * HZ;
1620                         }
1621                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1622                         task->tk_status = -EIO;
1623                         return;
1624                 }
1625
1626                 /* For non rpc-based layout drivers, retry-through-MDS */
1627                 if (!task->tk_ops) {
1628                         hdr->pnfs_error = -EAGAIN;
1629                         return;
1630                 }
1631
1632                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1633                 if (resp->verf->committed != NFS_UNSTABLE) {
1634                         /* Resend from where the server left off */
1635                         hdr->mds_offset += resp->count;
1636                         argp->offset += resp->count;
1637                         argp->pgbase += resp->count;
1638                         argp->count -= resp->count;
1639                 } else {
1640                         /* Resend as a stable write in order to avoid
1641                          * headaches in the case of a server crash.
1642                          */
1643                         argp->stable = NFS_FILE_SYNC;
1644                 }
1645                 rpc_restart_call_prepare(task);
1646         }
1647 }
1648
1649 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1650 {
1651         return wait_var_event_killable(&cinfo->rpcs_out,
1652                                        !atomic_read(&cinfo->rpcs_out));
1653 }
1654
1655 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1656 {
1657         atomic_inc(&cinfo->rpcs_out);
1658 }
1659
1660 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1661 {
1662         if (atomic_dec_and_test(&cinfo->rpcs_out))
1663                 wake_up_var(&cinfo->rpcs_out);
1664 }
1665
1666 void nfs_commitdata_release(struct nfs_commit_data *data)
1667 {
1668         put_nfs_open_context(data->context);
1669         nfs_commit_free(data);
1670 }
1671 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1672
1673 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1674                         const struct nfs_rpc_ops *nfs_ops,
1675                         const struct rpc_call_ops *call_ops,
1676                         int how, int flags)
1677 {
1678         struct rpc_task *task;
1679         int priority = flush_task_priority(how);
1680         struct rpc_message msg = {
1681                 .rpc_argp = &data->args,
1682                 .rpc_resp = &data->res,
1683                 .rpc_cred = data->cred,
1684         };
1685         struct rpc_task_setup task_setup_data = {
1686                 .task = &data->task,
1687                 .rpc_client = clnt,
1688                 .rpc_message = &msg,
1689                 .callback_ops = call_ops,
1690                 .callback_data = data,
1691                 .workqueue = nfsiod_workqueue,
1692                 .flags = RPC_TASK_ASYNC | flags,
1693                 .priority = priority,
1694         };
1695         /* Set up the initial task struct.  */
1696         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1697         trace_nfs_initiate_commit(data);
1698
1699         dprintk("NFS: initiated commit call\n");
1700
1701         task = rpc_run_task(&task_setup_data);
1702         if (IS_ERR(task))
1703                 return PTR_ERR(task);
1704         if (how & FLUSH_SYNC)
1705                 rpc_wait_for_completion_task(task);
1706         rpc_put_task(task);
1707         return 0;
1708 }
1709 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1710
1711 static loff_t nfs_get_lwb(struct list_head *head)
1712 {
1713         loff_t lwb = 0;
1714         struct nfs_page *req;
1715
1716         list_for_each_entry(req, head, wb_list)
1717                 if (lwb < (req_offset(req) + req->wb_bytes))
1718                         lwb = req_offset(req) + req->wb_bytes;
1719
1720         return lwb;
1721 }
1722
1723 /*
1724  * Set up the argument/result storage required for the RPC call.
1725  */
1726 void nfs_init_commit(struct nfs_commit_data *data,
1727                      struct list_head *head,
1728                      struct pnfs_layout_segment *lseg,
1729                      struct nfs_commit_info *cinfo)
1730 {
1731         struct nfs_page *first = nfs_list_entry(head->next);
1732         struct inode *inode = d_inode(first->wb_context->dentry);
1733
1734         /* Set up the RPC argument and reply structs
1735          * NB: take care not to mess about with data->commit et al. */
1736
1737         list_splice_init(head, &data->pages);
1738
1739         data->inode       = inode;
1740         data->cred        = first->wb_context->cred;
1741         data->lseg        = lseg; /* reference transferred */
1742         /* only set lwb for pnfs commit */
1743         if (lseg)
1744                 data->lwb = nfs_get_lwb(&data->pages);
1745         data->mds_ops     = &nfs_commit_ops;
1746         data->completion_ops = cinfo->completion_ops;
1747         data->dreq        = cinfo->dreq;
1748
1749         data->args.fh     = NFS_FH(data->inode);
1750         /* Note: we always request a commit of the entire inode */
1751         data->args.offset = 0;
1752         data->args.count  = 0;
1753         data->context     = get_nfs_open_context(first->wb_context);
1754         data->res.fattr   = &data->fattr;
1755         data->res.verf    = &data->verf;
1756         nfs_fattr_init(&data->fattr);
1757 }
1758 EXPORT_SYMBOL_GPL(nfs_init_commit);
1759
1760 void nfs_retry_commit(struct list_head *page_list,
1761                       struct pnfs_layout_segment *lseg,
1762                       struct nfs_commit_info *cinfo,
1763                       u32 ds_commit_idx)
1764 {
1765         struct nfs_page *req;
1766
1767         while (!list_empty(page_list)) {
1768                 req = nfs_list_entry(page_list->next);
1769                 nfs_list_remove_request(req);
1770                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1771                 if (!cinfo->dreq)
1772                         nfs_clear_page_commit(req->wb_page);
1773                 nfs_unlock_and_release_request(req);
1774         }
1775 }
1776 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1777
1778 static void
1779 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1780                 struct nfs_page *req)
1781 {
1782         __set_page_dirty_nobuffers(req->wb_page);
1783 }
1784
1785 /*
1786  * Commit dirty pages
1787  */
1788 static int
1789 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1790                 struct nfs_commit_info *cinfo)
1791 {
1792         struct nfs_commit_data  *data;
1793
1794         /* another commit raced with us */
1795         if (list_empty(head))
1796                 return 0;
1797
1798         data = nfs_commitdata_alloc(true);
1799
1800         /* Set up the argument struct */
1801         nfs_init_commit(data, head, NULL, cinfo);
1802         atomic_inc(&cinfo->mds->rpcs_out);
1803         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1804                                    data->mds_ops, how, 0);
1805 }
1806
1807 /*
1808  * COMMIT call returned
1809  */
1810 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1811 {
1812         struct nfs_commit_data  *data = calldata;
1813
1814         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1815                                 task->tk_pid, task->tk_status);
1816
1817         /* Call the NFS version-specific code */
1818         NFS_PROTO(data->inode)->commit_done(task, data);
1819         trace_nfs_commit_done(data);
1820 }
1821
1822 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1823 {
1824         const struct nfs_writeverf *verf = data->res.verf;
1825         struct nfs_page *req;
1826         int status = data->task.tk_status;
1827         struct nfs_commit_info cinfo;
1828         struct nfs_server *nfss;
1829
1830         while (!list_empty(&data->pages)) {
1831                 req = nfs_list_entry(data->pages.next);
1832                 nfs_list_remove_request(req);
1833                 if (req->wb_page)
1834                         nfs_clear_page_commit(req->wb_page);
1835
1836                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1837                         req->wb_context->dentry->d_sb->s_id,
1838                         (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
1839                         req->wb_bytes,
1840                         (long long)req_offset(req));
1841                 if (status < 0) {
1842                         nfs_context_set_write_error(req->wb_context, status);
1843                         if (req->wb_page)
1844                                 nfs_inode_remove_request(req);
1845                         dprintk_cont(", error = %d\n", status);
1846                         goto next;
1847                 }
1848
1849                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1850                  * returned by the server against all stored verfs. */
1851                 if (verf->committed > NFS_UNSTABLE &&
1852                     !nfs_write_verifier_cmp(&req->wb_verf, &verf->verifier)) {
1853                         /* We have a match */
1854                         if (req->wb_page)
1855                                 nfs_inode_remove_request(req);
1856                         dprintk_cont(" OK\n");
1857                         goto next;
1858                 }
1859                 /* We have a mismatch. Write the page again */
1860                 dprintk_cont(" mismatch\n");
1861                 nfs_mark_request_dirty(req);
1862                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1863         next:
1864                 nfs_unlock_and_release_request(req);
1865                 /* Latency breaker */
1866                 cond_resched();
1867         }
1868         nfss = NFS_SERVER(data->inode);
1869         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1870                 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1871
1872         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1873         nfs_commit_end(cinfo.mds);
1874 }
1875
1876 static void nfs_commit_release(void *calldata)
1877 {
1878         struct nfs_commit_data *data = calldata;
1879
1880         data->completion_ops->completion(data);
1881         nfs_commitdata_release(calldata);
1882 }
1883
1884 static const struct rpc_call_ops nfs_commit_ops = {
1885         .rpc_call_prepare = nfs_commit_prepare,
1886         .rpc_call_done = nfs_commit_done,
1887         .rpc_release = nfs_commit_release,
1888 };
1889
1890 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1891         .completion = nfs_commit_release_pages,
1892         .resched_write = nfs_commit_resched_write,
1893 };
1894
1895 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1896                             int how, struct nfs_commit_info *cinfo)
1897 {
1898         int status;
1899
1900         status = pnfs_commit_list(inode, head, how, cinfo);
1901         if (status == PNFS_NOT_ATTEMPTED)
1902                 status = nfs_commit_list(inode, head, how, cinfo);
1903         return status;
1904 }
1905
1906 static int __nfs_commit_inode(struct inode *inode, int how,
1907                 struct writeback_control *wbc)
1908 {
1909         LIST_HEAD(head);
1910         struct nfs_commit_info cinfo;
1911         int may_wait = how & FLUSH_SYNC;
1912         int ret, nscan;
1913
1914         nfs_init_cinfo_from_inode(&cinfo, inode);
1915         nfs_commit_begin(cinfo.mds);
1916         for (;;) {
1917                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1918                 if (ret <= 0)
1919                         break;
1920                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1921                 if (ret < 0)
1922                         break;
1923                 ret = 0;
1924                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1925                         if (nscan < wbc->nr_to_write)
1926                                 wbc->nr_to_write -= nscan;
1927                         else
1928                                 wbc->nr_to_write = 0;
1929                 }
1930                 if (nscan < INT_MAX)
1931                         break;
1932                 cond_resched();
1933         }
1934         nfs_commit_end(cinfo.mds);
1935         if (ret || !may_wait)
1936                 return ret;
1937         return wait_on_commit(cinfo.mds);
1938 }
1939
1940 int nfs_commit_inode(struct inode *inode, int how)
1941 {
1942         return __nfs_commit_inode(inode, how, NULL);
1943 }
1944 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1945
1946 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1947 {
1948         struct nfs_inode *nfsi = NFS_I(inode);
1949         int flags = FLUSH_SYNC;
1950         int ret = 0;
1951
1952         if (wbc->sync_mode == WB_SYNC_NONE) {
1953                 /* no commits means nothing needs to be done */
1954                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1955                         goto check_requests_outstanding;
1956
1957                 /* Don't commit yet if this is a non-blocking flush and there
1958                  * are a lot of outstanding writes for this mapping.
1959                  */
1960                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1961                         goto out_mark_dirty;
1962
1963                 /* don't wait for the COMMIT response */
1964                 flags = 0;
1965         }
1966
1967         ret = __nfs_commit_inode(inode, flags, wbc);
1968         if (!ret) {
1969                 if (flags & FLUSH_SYNC)
1970                         return 0;
1971         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1972                 goto out_mark_dirty;
1973
1974 check_requests_outstanding:
1975         if (!atomic_read(&nfsi->commit_info.rpcs_out))
1976                 return ret;
1977 out_mark_dirty:
1978         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1979         return ret;
1980 }
1981 EXPORT_SYMBOL_GPL(nfs_write_inode);
1982
1983 /*
1984  * Wrapper for filemap_write_and_wait_range()
1985  *
1986  * Needed for pNFS in order to ensure data becomes visible to the
1987  * client.
1988  */
1989 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1990                 loff_t lstart, loff_t lend)
1991 {
1992         int ret;
1993
1994         ret = filemap_write_and_wait_range(mapping, lstart, lend);
1995         if (ret == 0)
1996                 ret = pnfs_sync_inode(mapping->host, true);
1997         return ret;
1998 }
1999 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2000
2001 /*
2002  * flush the inode to disk.
2003  */
2004 int nfs_wb_all(struct inode *inode)
2005 {
2006         int ret;
2007
2008         trace_nfs_writeback_inode_enter(inode);
2009
2010         ret = filemap_write_and_wait(inode->i_mapping);
2011         if (ret)
2012                 goto out;
2013         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2014         if (ret < 0)
2015                 goto out;
2016         pnfs_sync_inode(inode, true);
2017         ret = 0;
2018
2019 out:
2020         trace_nfs_writeback_inode_exit(inode, ret);
2021         return ret;
2022 }
2023 EXPORT_SYMBOL_GPL(nfs_wb_all);
2024
2025 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2026 {
2027         struct nfs_page *req;
2028         int ret = 0;
2029
2030         wait_on_page_writeback(page);
2031
2032         /* blocking call to cancel all requests and join to a single (head)
2033          * request */
2034         req = nfs_lock_and_join_requests(page);
2035
2036         if (IS_ERR(req)) {
2037                 ret = PTR_ERR(req);
2038         } else if (req) {
2039                 /* all requests from this page have been cancelled by
2040                  * nfs_lock_and_join_requests, so just remove the head
2041                  * request from the inode / page_private pointer and
2042                  * release it */
2043                 nfs_inode_remove_request(req);
2044                 nfs_unlock_and_release_request(req);
2045         }
2046
2047         return ret;
2048 }
2049
2050 /*
2051  * Write back all requests on one page - we do this before reading it.
2052  */
2053 int nfs_wb_page(struct inode *inode, struct page *page)
2054 {
2055         loff_t range_start = page_file_offset(page);
2056         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2057         struct writeback_control wbc = {
2058                 .sync_mode = WB_SYNC_ALL,
2059                 .nr_to_write = 0,
2060                 .range_start = range_start,
2061                 .range_end = range_end,
2062         };
2063         int ret;
2064
2065         trace_nfs_writeback_page_enter(inode);
2066
2067         for (;;) {
2068                 wait_on_page_writeback(page);
2069                 if (clear_page_dirty_for_io(page)) {
2070                         ret = nfs_writepage_locked(page, &wbc);
2071                         if (ret < 0)
2072                                 goto out_error;
2073                         continue;
2074                 }
2075                 ret = 0;
2076                 if (!PagePrivate(page))
2077                         break;
2078                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2079                 if (ret < 0)
2080                         goto out_error;
2081         }
2082 out_error:
2083         trace_nfs_writeback_page_exit(inode, ret);
2084         return ret;
2085 }
2086
2087 #ifdef CONFIG_MIGRATION
2088 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2089                 struct page *page, enum migrate_mode mode)
2090 {
2091         /*
2092          * If PagePrivate is set, then the page is currently associated with
2093          * an in-progress read or write request. Don't try to migrate it.
2094          *
2095          * FIXME: we could do this in principle, but we'll need a way to ensure
2096          *        that we can safely release the inode reference while holding
2097          *        the page lock.
2098          */
2099         if (PagePrivate(page))
2100                 return -EBUSY;
2101
2102         if (!nfs_fscache_release_page(page, GFP_KERNEL))
2103                 return -EBUSY;
2104
2105         return migrate_page(mapping, newpage, page, mode);
2106 }
2107 #endif
2108
2109 int __init nfs_init_writepagecache(void)
2110 {
2111         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2112                                              sizeof(struct nfs_pgio_header),
2113                                              0, SLAB_HWCACHE_ALIGN,
2114                                              NULL);
2115         if (nfs_wdata_cachep == NULL)
2116                 return -ENOMEM;
2117
2118         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2119                                                      nfs_wdata_cachep);
2120         if (nfs_wdata_mempool == NULL)
2121                 goto out_destroy_write_cache;
2122
2123         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2124                                              sizeof(struct nfs_commit_data),
2125                                              0, SLAB_HWCACHE_ALIGN,
2126                                              NULL);
2127         if (nfs_cdata_cachep == NULL)
2128                 goto out_destroy_write_mempool;
2129
2130         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2131                                                       nfs_cdata_cachep);
2132         if (nfs_commit_mempool == NULL)
2133                 goto out_destroy_commit_cache;
2134
2135         /*
2136          * NFS congestion size, scale with available memory.
2137          *
2138          *  64MB:    8192k
2139          * 128MB:   11585k
2140          * 256MB:   16384k
2141          * 512MB:   23170k
2142          *   1GB:   32768k
2143          *   2GB:   46340k
2144          *   4GB:   65536k
2145          *   8GB:   92681k
2146          *  16GB:  131072k
2147          *
2148          * This allows larger machines to have larger/more transfers.
2149          * Limit the default to 256M
2150          */
2151         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
2152         if (nfs_congestion_kb > 256*1024)
2153                 nfs_congestion_kb = 256*1024;
2154
2155         return 0;
2156
2157 out_destroy_commit_cache:
2158         kmem_cache_destroy(nfs_cdata_cachep);
2159 out_destroy_write_mempool:
2160         mempool_destroy(nfs_wdata_mempool);
2161 out_destroy_write_cache:
2162         kmem_cache_destroy(nfs_wdata_cachep);
2163         return -ENOMEM;
2164 }
2165
2166 void nfs_destroy_writepagecache(void)
2167 {
2168         mempool_destroy(nfs_commit_mempool);
2169         kmem_cache_destroy(nfs_cdata_cachep);
2170         mempool_destroy(nfs_wdata_mempool);
2171         kmem_cache_destroy(nfs_wdata_cachep);
2172 }
2173
2174 static const struct nfs_rw_ops nfs_rw_write_ops = {
2175         .rw_alloc_header        = nfs_writehdr_alloc,
2176         .rw_free_header         = nfs_writehdr_free,
2177         .rw_done                = nfs_writeback_done,
2178         .rw_result              = nfs_writeback_result,
2179         .rw_initiate            = nfs_initiate_write,
2180 };