Merge branch 'bugfixes' of git://git.linux-nfs.org/projects/trondmy/nfs-2.6
[linux-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
24 #include <asm/uaccess.h>
25
26 #include "delegation.h"
27 #include "internal.h"
28 #include "iostat.h"
29 #include "nfs4_fs.h"
30 #include "fscache.h"
31
32 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
33
34 #define MIN_POOL_WRITE          (32)
35 #define MIN_POOL_COMMIT         (4)
36
37 /*
38  * Local function declarations
39  */
40 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
41                                   struct inode *inode, int ioflags);
42 static void nfs_redirty_request(struct nfs_page *req);
43 static const struct rpc_call_ops nfs_write_partial_ops;
44 static const struct rpc_call_ops nfs_write_full_ops;
45 static const struct rpc_call_ops nfs_commit_ops;
46
47 static struct kmem_cache *nfs_wdata_cachep;
48 static mempool_t *nfs_wdata_mempool;
49 static mempool_t *nfs_commit_mempool;
50
51 struct nfs_write_data *nfs_commitdata_alloc(void)
52 {
53         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
54
55         if (p) {
56                 memset(p, 0, sizeof(*p));
57                 INIT_LIST_HEAD(&p->pages);
58                 p->res.seq_res.sr_slotid = NFS4_MAX_SLOT_TABLE;
59         }
60         return p;
61 }
62
63 void nfs_commit_free(struct nfs_write_data *p)
64 {
65         if (p && (p->pagevec != &p->page_array[0]))
66                 kfree(p->pagevec);
67         mempool_free(p, nfs_commit_mempool);
68 }
69
70 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
71 {
72         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
73
74         if (p) {
75                 memset(p, 0, sizeof(*p));
76                 INIT_LIST_HEAD(&p->pages);
77                 p->npages = pagecount;
78                 p->res.seq_res.sr_slotid = NFS4_MAX_SLOT_TABLE;
79                 if (pagecount <= ARRAY_SIZE(p->page_array))
80                         p->pagevec = p->page_array;
81                 else {
82                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
83                         if (!p->pagevec) {
84                                 mempool_free(p, nfs_wdata_mempool);
85                                 p = NULL;
86                         }
87                 }
88         }
89         return p;
90 }
91
92 void nfs_writedata_free(struct nfs_write_data *p)
93 {
94         if (p && (p->pagevec != &p->page_array[0]))
95                 kfree(p->pagevec);
96         mempool_free(p, nfs_wdata_mempool);
97 }
98
99 static void nfs_writedata_release(struct nfs_write_data *wdata)
100 {
101         put_nfs_open_context(wdata->args.context);
102         nfs_writedata_free(wdata);
103 }
104
105 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
106 {
107         ctx->error = error;
108         smp_wmb();
109         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
110 }
111
112 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
113 {
114         struct nfs_page *req = NULL;
115
116         if (PagePrivate(page)) {
117                 req = (struct nfs_page *)page_private(page);
118                 if (req != NULL)
119                         kref_get(&req->wb_kref);
120         }
121         return req;
122 }
123
124 static struct nfs_page *nfs_page_find_request(struct page *page)
125 {
126         struct inode *inode = page->mapping->host;
127         struct nfs_page *req = NULL;
128
129         spin_lock(&inode->i_lock);
130         req = nfs_page_find_request_locked(page);
131         spin_unlock(&inode->i_lock);
132         return req;
133 }
134
135 /* Adjust the file length if we're writing beyond the end */
136 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
137 {
138         struct inode *inode = page->mapping->host;
139         loff_t end, i_size;
140         pgoff_t end_index;
141
142         spin_lock(&inode->i_lock);
143         i_size = i_size_read(inode);
144         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
145         if (i_size > 0 && page->index < end_index)
146                 goto out;
147         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
148         if (i_size >= end)
149                 goto out;
150         i_size_write(inode, end);
151         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
152 out:
153         spin_unlock(&inode->i_lock);
154 }
155
156 /* A writeback failed: mark the page as bad, and invalidate the page cache */
157 static void nfs_set_pageerror(struct page *page)
158 {
159         SetPageError(page);
160         nfs_zap_mapping(page->mapping->host, page->mapping);
161 }
162
163 /* We can set the PG_uptodate flag if we see that a write request
164  * covers the full page.
165  */
166 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
167 {
168         if (PageUptodate(page))
169                 return;
170         if (base != 0)
171                 return;
172         if (count != nfs_page_length(page))
173                 return;
174         SetPageUptodate(page);
175 }
176
177 static int wb_priority(struct writeback_control *wbc)
178 {
179         if (wbc->for_reclaim)
180                 return FLUSH_HIGHPRI | FLUSH_STABLE;
181         if (wbc->for_kupdate || wbc->for_background)
182                 return FLUSH_LOWPRI;
183         return 0;
184 }
185
186 /*
187  * NFS congestion control
188  */
189
190 int nfs_congestion_kb;
191
192 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
193 #define NFS_CONGESTION_OFF_THRESH       \
194         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
195
196 static int nfs_set_page_writeback(struct page *page)
197 {
198         int ret = test_set_page_writeback(page);
199
200         if (!ret) {
201                 struct inode *inode = page->mapping->host;
202                 struct nfs_server *nfss = NFS_SERVER(inode);
203
204                 page_cache_get(page);
205                 if (atomic_long_inc_return(&nfss->writeback) >
206                                 NFS_CONGESTION_ON_THRESH) {
207                         set_bdi_congested(&nfss->backing_dev_info,
208                                                 BLK_RW_ASYNC);
209                 }
210         }
211         return ret;
212 }
213
214 static void nfs_end_page_writeback(struct page *page)
215 {
216         struct inode *inode = page->mapping->host;
217         struct nfs_server *nfss = NFS_SERVER(inode);
218
219         end_page_writeback(page);
220         page_cache_release(page);
221         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
222                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
223 }
224
225 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
226 {
227         struct inode *inode = page->mapping->host;
228         struct nfs_page *req;
229         int ret;
230
231         spin_lock(&inode->i_lock);
232         for (;;) {
233                 req = nfs_page_find_request_locked(page);
234                 if (req == NULL)
235                         break;
236                 if (nfs_set_page_tag_locked(req))
237                         break;
238                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
239                  *       then the call to nfs_set_page_tag_locked() will always
240                  *       succeed provided that someone hasn't already marked the
241                  *       request as dirty (in which case we don't care).
242                  */
243                 spin_unlock(&inode->i_lock);
244                 if (!nonblock)
245                         ret = nfs_wait_on_request(req);
246                 else
247                         ret = -EAGAIN;
248                 nfs_release_request(req);
249                 if (ret != 0)
250                         return ERR_PTR(ret);
251                 spin_lock(&inode->i_lock);
252         }
253         spin_unlock(&inode->i_lock);
254         return req;
255 }
256
257 /*
258  * Find an associated nfs write request, and prepare to flush it out
259  * May return an error if the user signalled nfs_wait_on_request().
260  */
261 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
262                                 struct page *page, bool nonblock)
263 {
264         struct nfs_page *req;
265         int ret = 0;
266
267         req = nfs_find_and_lock_request(page, nonblock);
268         if (!req)
269                 goto out;
270         ret = PTR_ERR(req);
271         if (IS_ERR(req))
272                 goto out;
273
274         ret = nfs_set_page_writeback(page);
275         BUG_ON(ret != 0);
276         BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
277
278         if (!nfs_pageio_add_request(pgio, req)) {
279                 nfs_redirty_request(req);
280                 ret = pgio->pg_error;
281         }
282 out:
283         return ret;
284 }
285
286 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
287 {
288         struct inode *inode = page->mapping->host;
289         int ret;
290
291         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
292         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
293
294         nfs_pageio_cond_complete(pgio, page->index);
295         ret = nfs_page_async_flush(pgio, page,
296                         wbc->sync_mode == WB_SYNC_NONE ||
297                         wbc->nonblocking != 0);
298         if (ret == -EAGAIN) {
299                 redirty_page_for_writepage(wbc, page);
300                 ret = 0;
301         }
302         return ret;
303 }
304
305 /*
306  * Write an mmapped page to the server.
307  */
308 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
309 {
310         struct nfs_pageio_descriptor pgio;
311         int err;
312
313         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
314         err = nfs_do_writepage(page, wbc, &pgio);
315         nfs_pageio_complete(&pgio);
316         if (err < 0)
317                 return err;
318         if (pgio.pg_error < 0)
319                 return pgio.pg_error;
320         return 0;
321 }
322
323 int nfs_writepage(struct page *page, struct writeback_control *wbc)
324 {
325         int ret;
326
327         ret = nfs_writepage_locked(page, wbc);
328         unlock_page(page);
329         return ret;
330 }
331
332 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
333 {
334         int ret;
335
336         ret = nfs_do_writepage(page, wbc, data);
337         unlock_page(page);
338         return ret;
339 }
340
341 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
342 {
343         struct inode *inode = mapping->host;
344         unsigned long *bitlock = &NFS_I(inode)->flags;
345         struct nfs_pageio_descriptor pgio;
346         int err;
347
348         /* Stop dirtying of new pages while we sync */
349         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
350                         nfs_wait_bit_killable, TASK_KILLABLE);
351         if (err)
352                 goto out_err;
353
354         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
355
356         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
357         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
358         nfs_pageio_complete(&pgio);
359
360         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
361         smp_mb__after_clear_bit();
362         wake_up_bit(bitlock, NFS_INO_FLUSHING);
363
364         if (err < 0)
365                 goto out_err;
366         err = pgio.pg_error;
367         if (err < 0)
368                 goto out_err;
369         return 0;
370 out_err:
371         return err;
372 }
373
374 /*
375  * Insert a write request into an inode
376  */
377 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
378 {
379         struct nfs_inode *nfsi = NFS_I(inode);
380         int error;
381
382         error = radix_tree_preload(GFP_NOFS);
383         if (error != 0)
384                 goto out;
385
386         /* Lock the request! */
387         nfs_lock_request_dontget(req);
388
389         spin_lock(&inode->i_lock);
390         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
391         BUG_ON(error);
392         if (!nfsi->npages) {
393                 igrab(inode);
394                 if (nfs_have_delegation(inode, FMODE_WRITE))
395                         nfsi->change_attr++;
396         }
397         SetPagePrivate(req->wb_page);
398         set_page_private(req->wb_page, (unsigned long)req);
399         nfsi->npages++;
400         kref_get(&req->wb_kref);
401         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
402                                 NFS_PAGE_TAG_LOCKED);
403         spin_unlock(&inode->i_lock);
404         radix_tree_preload_end();
405 out:
406         return error;
407 }
408
409 /*
410  * Remove a write request from an inode
411  */
412 static void nfs_inode_remove_request(struct nfs_page *req)
413 {
414         struct inode *inode = req->wb_context->path.dentry->d_inode;
415         struct nfs_inode *nfsi = NFS_I(inode);
416
417         BUG_ON (!NFS_WBACK_BUSY(req));
418
419         spin_lock(&inode->i_lock);
420         set_page_private(req->wb_page, 0);
421         ClearPagePrivate(req->wb_page);
422         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
423         nfsi->npages--;
424         if (!nfsi->npages) {
425                 spin_unlock(&inode->i_lock);
426                 iput(inode);
427         } else
428                 spin_unlock(&inode->i_lock);
429         nfs_clear_request(req);
430         nfs_release_request(req);
431 }
432
433 static void
434 nfs_mark_request_dirty(struct nfs_page *req)
435 {
436         __set_page_dirty_nobuffers(req->wb_page);
437         __mark_inode_dirty(req->wb_page->mapping->host, I_DIRTY_DATASYNC);
438 }
439
440 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
441 /*
442  * Add a request to the inode's commit list.
443  */
444 static void
445 nfs_mark_request_commit(struct nfs_page *req)
446 {
447         struct inode *inode = req->wb_context->path.dentry->d_inode;
448         struct nfs_inode *nfsi = NFS_I(inode);
449
450         spin_lock(&inode->i_lock);
451         set_bit(PG_CLEAN, &(req)->wb_flags);
452         radix_tree_tag_set(&nfsi->nfs_page_tree,
453                         req->wb_index,
454                         NFS_PAGE_TAG_COMMIT);
455         nfsi->ncommit++;
456         spin_unlock(&inode->i_lock);
457         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
458         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
459         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
460 }
461
462 static int
463 nfs_clear_request_commit(struct nfs_page *req)
464 {
465         struct page *page = req->wb_page;
466
467         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
468                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
469                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
470                 return 1;
471         }
472         return 0;
473 }
474
475 static inline
476 int nfs_write_need_commit(struct nfs_write_data *data)
477 {
478         return data->verf.committed != NFS_FILE_SYNC;
479 }
480
481 static inline
482 int nfs_reschedule_unstable_write(struct nfs_page *req)
483 {
484         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
485                 nfs_mark_request_commit(req);
486                 return 1;
487         }
488         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
489                 nfs_mark_request_dirty(req);
490                 return 1;
491         }
492         return 0;
493 }
494 #else
495 static inline void
496 nfs_mark_request_commit(struct nfs_page *req)
497 {
498 }
499
500 static inline int
501 nfs_clear_request_commit(struct nfs_page *req)
502 {
503         return 0;
504 }
505
506 static inline
507 int nfs_write_need_commit(struct nfs_write_data *data)
508 {
509         return 0;
510 }
511
512 static inline
513 int nfs_reschedule_unstable_write(struct nfs_page *req)
514 {
515         return 0;
516 }
517 #endif
518
519 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
520 static int
521 nfs_need_commit(struct nfs_inode *nfsi)
522 {
523         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
524 }
525
526 /*
527  * nfs_scan_commit - Scan an inode for commit requests
528  * @inode: NFS inode to scan
529  * @dst: destination list
530  * @idx_start: lower bound of page->index to scan.
531  * @npages: idx_start + npages sets the upper bound to scan.
532  *
533  * Moves requests from the inode's 'commit' request list.
534  * The requests are *not* checked to ensure that they form a contiguous set.
535  */
536 static int
537 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
538 {
539         struct nfs_inode *nfsi = NFS_I(inode);
540         int ret;
541
542         if (!nfs_need_commit(nfsi))
543                 return 0;
544
545         ret = nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
546         if (ret > 0)
547                 nfsi->ncommit -= ret;
548         if (nfs_need_commit(NFS_I(inode)))
549                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
550         return ret;
551 }
552 #else
553 static inline int nfs_need_commit(struct nfs_inode *nfsi)
554 {
555         return 0;
556 }
557
558 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
559 {
560         return 0;
561 }
562 #endif
563
564 /*
565  * Search for an existing write request, and attempt to update
566  * it to reflect a new dirty region on a given page.
567  *
568  * If the attempt fails, then the existing request is flushed out
569  * to disk.
570  */
571 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
572                 struct page *page,
573                 unsigned int offset,
574                 unsigned int bytes)
575 {
576         struct nfs_page *req;
577         unsigned int rqend;
578         unsigned int end;
579         int error;
580
581         if (!PagePrivate(page))
582                 return NULL;
583
584         end = offset + bytes;
585         spin_lock(&inode->i_lock);
586
587         for (;;) {
588                 req = nfs_page_find_request_locked(page);
589                 if (req == NULL)
590                         goto out_unlock;
591
592                 rqend = req->wb_offset + req->wb_bytes;
593                 /*
594                  * Tell the caller to flush out the request if
595                  * the offsets are non-contiguous.
596                  * Note: nfs_flush_incompatible() will already
597                  * have flushed out requests having wrong owners.
598                  */
599                 if (offset > rqend
600                     || end < req->wb_offset)
601                         goto out_flushme;
602
603                 if (nfs_set_page_tag_locked(req))
604                         break;
605
606                 /* The request is locked, so wait and then retry */
607                 spin_unlock(&inode->i_lock);
608                 error = nfs_wait_on_request(req);
609                 nfs_release_request(req);
610                 if (error != 0)
611                         goto out_err;
612                 spin_lock(&inode->i_lock);
613         }
614
615         if (nfs_clear_request_commit(req) &&
616                         radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
617                                 req->wb_index, NFS_PAGE_TAG_COMMIT) != NULL)
618                 NFS_I(inode)->ncommit--;
619
620         /* Okay, the request matches. Update the region */
621         if (offset < req->wb_offset) {
622                 req->wb_offset = offset;
623                 req->wb_pgbase = offset;
624         }
625         if (end > rqend)
626                 req->wb_bytes = end - req->wb_offset;
627         else
628                 req->wb_bytes = rqend - req->wb_offset;
629 out_unlock:
630         spin_unlock(&inode->i_lock);
631         return req;
632 out_flushme:
633         spin_unlock(&inode->i_lock);
634         nfs_release_request(req);
635         error = nfs_wb_page(inode, page);
636 out_err:
637         return ERR_PTR(error);
638 }
639
640 /*
641  * Try to update an existing write request, or create one if there is none.
642  *
643  * Note: Should always be called with the Page Lock held to prevent races
644  * if we have to add a new request. Also assumes that the caller has
645  * already called nfs_flush_incompatible() if necessary.
646  */
647 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
648                 struct page *page, unsigned int offset, unsigned int bytes)
649 {
650         struct inode *inode = page->mapping->host;
651         struct nfs_page *req;
652         int error;
653
654         req = nfs_try_to_update_request(inode, page, offset, bytes);
655         if (req != NULL)
656                 goto out;
657         req = nfs_create_request(ctx, inode, page, offset, bytes);
658         if (IS_ERR(req))
659                 goto out;
660         error = nfs_inode_add_request(inode, req);
661         if (error != 0) {
662                 nfs_release_request(req);
663                 req = ERR_PTR(error);
664         }
665 out:
666         return req;
667 }
668
669 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
670                 unsigned int offset, unsigned int count)
671 {
672         struct nfs_page *req;
673
674         req = nfs_setup_write_request(ctx, page, offset, count);
675         if (IS_ERR(req))
676                 return PTR_ERR(req);
677         nfs_mark_request_dirty(req);
678         /* Update file length */
679         nfs_grow_file(page, offset, count);
680         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
681         nfs_mark_request_dirty(req);
682         nfs_clear_page_tag_locked(req);
683         return 0;
684 }
685
686 int nfs_flush_incompatible(struct file *file, struct page *page)
687 {
688         struct nfs_open_context *ctx = nfs_file_open_context(file);
689         struct nfs_page *req;
690         int do_flush, status;
691         /*
692          * Look for a request corresponding to this page. If there
693          * is one, and it belongs to another file, we flush it out
694          * before we try to copy anything into the page. Do this
695          * due to the lack of an ACCESS-type call in NFSv2.
696          * Also do the same if we find a request from an existing
697          * dropped page.
698          */
699         do {
700                 req = nfs_page_find_request(page);
701                 if (req == NULL)
702                         return 0;
703                 do_flush = req->wb_page != page || req->wb_context != ctx;
704                 nfs_release_request(req);
705                 if (!do_flush)
706                         return 0;
707                 status = nfs_wb_page(page->mapping->host, page);
708         } while (status == 0);
709         return status;
710 }
711
712 /*
713  * If the page cache is marked as unsafe or invalid, then we can't rely on
714  * the PageUptodate() flag. In this case, we will need to turn off
715  * write optimisations that depend on the page contents being correct.
716  */
717 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
718 {
719         return PageUptodate(page) &&
720                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
721 }
722
723 /*
724  * Update and possibly write a cached page of an NFS file.
725  *
726  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
727  * things with a page scheduled for an RPC call (e.g. invalidate it).
728  */
729 int nfs_updatepage(struct file *file, struct page *page,
730                 unsigned int offset, unsigned int count)
731 {
732         struct nfs_open_context *ctx = nfs_file_open_context(file);
733         struct inode    *inode = page->mapping->host;
734         int             status = 0;
735
736         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
737
738         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
739                 file->f_path.dentry->d_parent->d_name.name,
740                 file->f_path.dentry->d_name.name, count,
741                 (long long)(page_offset(page) + offset));
742
743         /* If we're not using byte range locks, and we know the page
744          * is up to date, it may be more efficient to extend the write
745          * to cover the entire page in order to avoid fragmentation
746          * inefficiencies.
747          */
748         if (nfs_write_pageuptodate(page, inode) &&
749                         inode->i_flock == NULL &&
750                         !(file->f_flags & O_DSYNC)) {
751                 count = max(count + offset, nfs_page_length(page));
752                 offset = 0;
753         }
754
755         status = nfs_writepage_setup(ctx, page, offset, count);
756         if (status < 0)
757                 nfs_set_pageerror(page);
758
759         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
760                         status, (long long)i_size_read(inode));
761         return status;
762 }
763
764 static void nfs_writepage_release(struct nfs_page *req)
765 {
766         struct page *page = req->wb_page;
767
768         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req))
769                 nfs_inode_remove_request(req);
770         nfs_clear_page_tag_locked(req);
771         nfs_end_page_writeback(page);
772 }
773
774 static int flush_task_priority(int how)
775 {
776         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
777                 case FLUSH_HIGHPRI:
778                         return RPC_PRIORITY_HIGH;
779                 case FLUSH_LOWPRI:
780                         return RPC_PRIORITY_LOW;
781         }
782         return RPC_PRIORITY_NORMAL;
783 }
784
785 /*
786  * Set up the argument/result storage required for the RPC call.
787  */
788 static int nfs_write_rpcsetup(struct nfs_page *req,
789                 struct nfs_write_data *data,
790                 const struct rpc_call_ops *call_ops,
791                 unsigned int count, unsigned int offset,
792                 int how)
793 {
794         struct inode *inode = req->wb_context->path.dentry->d_inode;
795         int priority = flush_task_priority(how);
796         struct rpc_task *task;
797         struct rpc_message msg = {
798                 .rpc_argp = &data->args,
799                 .rpc_resp = &data->res,
800                 .rpc_cred = req->wb_context->cred,
801         };
802         struct rpc_task_setup task_setup_data = {
803                 .rpc_client = NFS_CLIENT(inode),
804                 .task = &data->task,
805                 .rpc_message = &msg,
806                 .callback_ops = call_ops,
807                 .callback_data = data,
808                 .workqueue = nfsiod_workqueue,
809                 .flags = RPC_TASK_ASYNC,
810                 .priority = priority,
811         };
812         int ret = 0;
813
814         /* Set up the RPC argument and reply structs
815          * NB: take care not to mess about with data->commit et al. */
816
817         data->req = req;
818         data->inode = inode = req->wb_context->path.dentry->d_inode;
819         data->cred = msg.rpc_cred;
820
821         data->args.fh     = NFS_FH(inode);
822         data->args.offset = req_offset(req) + offset;
823         data->args.pgbase = req->wb_pgbase + offset;
824         data->args.pages  = data->pagevec;
825         data->args.count  = count;
826         data->args.context = get_nfs_open_context(req->wb_context);
827         data->args.stable  = NFS_UNSTABLE;
828         if (how & FLUSH_STABLE) {
829                 data->args.stable = NFS_DATA_SYNC;
830                 if (!nfs_need_commit(NFS_I(inode)))
831                         data->args.stable = NFS_FILE_SYNC;
832         }
833
834         data->res.fattr   = &data->fattr;
835         data->res.count   = count;
836         data->res.verf    = &data->verf;
837         nfs_fattr_init(&data->fattr);
838
839         /* Set up the initial task struct.  */
840         NFS_PROTO(inode)->write_setup(data, &msg);
841
842         dprintk("NFS: %5u initiated write call "
843                 "(req %s/%lld, %u bytes @ offset %llu)\n",
844                 data->task.tk_pid,
845                 inode->i_sb->s_id,
846                 (long long)NFS_FILEID(inode),
847                 count,
848                 (unsigned long long)data->args.offset);
849
850         task = rpc_run_task(&task_setup_data);
851         if (IS_ERR(task)) {
852                 ret = PTR_ERR(task);
853                 goto out;
854         }
855         if (how & FLUSH_SYNC) {
856                 ret = rpc_wait_for_completion_task(task);
857                 if (ret == 0)
858                         ret = task->tk_status;
859         }
860         rpc_put_task(task);
861 out:
862         return ret;
863 }
864
865 /* If a nfs_flush_* function fails, it should remove reqs from @head and
866  * call this on each, which will prepare them to be retried on next
867  * writeback using standard nfs.
868  */
869 static void nfs_redirty_request(struct nfs_page *req)
870 {
871         struct page *page = req->wb_page;
872
873         nfs_mark_request_dirty(req);
874         nfs_clear_page_tag_locked(req);
875         nfs_end_page_writeback(page);
876 }
877
878 /*
879  * Generate multiple small requests to write out a single
880  * contiguous dirty area on one page.
881  */
882 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
883 {
884         struct nfs_page *req = nfs_list_entry(head->next);
885         struct page *page = req->wb_page;
886         struct nfs_write_data *data;
887         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
888         unsigned int offset;
889         int requests = 0;
890         int ret = 0;
891         LIST_HEAD(list);
892
893         nfs_list_remove_request(req);
894
895         nbytes = count;
896         do {
897                 size_t len = min(nbytes, wsize);
898
899                 data = nfs_writedata_alloc(1);
900                 if (!data)
901                         goto out_bad;
902                 list_add(&data->pages, &list);
903                 requests++;
904                 nbytes -= len;
905         } while (nbytes != 0);
906         atomic_set(&req->wb_complete, requests);
907
908         ClearPageError(page);
909         offset = 0;
910         nbytes = count;
911         do {
912                 int ret2;
913
914                 data = list_entry(list.next, struct nfs_write_data, pages);
915                 list_del_init(&data->pages);
916
917                 data->pagevec[0] = page;
918
919                 if (nbytes < wsize)
920                         wsize = nbytes;
921                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
922                                    wsize, offset, how);
923                 if (ret == 0)
924                         ret = ret2;
925                 offset += wsize;
926                 nbytes -= wsize;
927         } while (nbytes != 0);
928
929         return ret;
930
931 out_bad:
932         while (!list_empty(&list)) {
933                 data = list_entry(list.next, struct nfs_write_data, pages);
934                 list_del(&data->pages);
935                 nfs_writedata_release(data);
936         }
937         nfs_redirty_request(req);
938         return -ENOMEM;
939 }
940
941 /*
942  * Create an RPC task for the given write request and kick it.
943  * The page must have been locked by the caller.
944  *
945  * It may happen that the page we're passed is not marked dirty.
946  * This is the case if nfs_updatepage detects a conflicting request
947  * that has been written but not committed.
948  */
949 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
950 {
951         struct nfs_page         *req;
952         struct page             **pages;
953         struct nfs_write_data   *data;
954
955         data = nfs_writedata_alloc(npages);
956         if (!data)
957                 goto out_bad;
958
959         pages = data->pagevec;
960         while (!list_empty(head)) {
961                 req = nfs_list_entry(head->next);
962                 nfs_list_remove_request(req);
963                 nfs_list_add_request(req, &data->pages);
964                 ClearPageError(req->wb_page);
965                 *pages++ = req->wb_page;
966         }
967         req = nfs_list_entry(data->pages.next);
968
969         /* Set up the argument struct */
970         return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
971  out_bad:
972         while (!list_empty(head)) {
973                 req = nfs_list_entry(head->next);
974                 nfs_list_remove_request(req);
975                 nfs_redirty_request(req);
976         }
977         return -ENOMEM;
978 }
979
980 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
981                                   struct inode *inode, int ioflags)
982 {
983         size_t wsize = NFS_SERVER(inode)->wsize;
984
985         if (wsize < PAGE_CACHE_SIZE)
986                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
987         else
988                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
989 }
990
991 /*
992  * Handle a write reply that flushed part of a page.
993  */
994 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
995 {
996         struct nfs_write_data   *data = calldata;
997
998         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
999                 task->tk_pid,
1000                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1001                 (long long)
1002                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1003                 data->req->wb_bytes, (long long)req_offset(data->req));
1004
1005         nfs_writeback_done(task, data);
1006 }
1007
1008 static void nfs_writeback_release_partial(void *calldata)
1009 {
1010         struct nfs_write_data   *data = calldata;
1011         struct nfs_page         *req = data->req;
1012         struct page             *page = req->wb_page;
1013         int status = data->task.tk_status;
1014
1015         if (status < 0) {
1016                 nfs_set_pageerror(page);
1017                 nfs_context_set_write_error(req->wb_context, status);
1018                 dprintk(", error = %d\n", status);
1019                 goto out;
1020         }
1021
1022         if (nfs_write_need_commit(data)) {
1023                 struct inode *inode = page->mapping->host;
1024
1025                 spin_lock(&inode->i_lock);
1026                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1027                         /* Do nothing we need to resend the writes */
1028                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1029                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1030                         dprintk(" defer commit\n");
1031                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1032                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1033                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1034                         dprintk(" server reboot detected\n");
1035                 }
1036                 spin_unlock(&inode->i_lock);
1037         } else
1038                 dprintk(" OK\n");
1039
1040 out:
1041         if (atomic_dec_and_test(&req->wb_complete))
1042                 nfs_writepage_release(req);
1043         nfs_writedata_release(calldata);
1044 }
1045
1046 #if defined(CONFIG_NFS_V4_1)
1047 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1048 {
1049         struct nfs_write_data *data = calldata;
1050         struct nfs_client *clp = (NFS_SERVER(data->inode))->nfs_client;
1051
1052         if (nfs4_setup_sequence(clp, &data->args.seq_args,
1053                                 &data->res.seq_res, 1, task))
1054                 return;
1055         rpc_call_start(task);
1056 }
1057 #endif /* CONFIG_NFS_V4_1 */
1058
1059 static const struct rpc_call_ops nfs_write_partial_ops = {
1060 #if defined(CONFIG_NFS_V4_1)
1061         .rpc_call_prepare = nfs_write_prepare,
1062 #endif /* CONFIG_NFS_V4_1 */
1063         .rpc_call_done = nfs_writeback_done_partial,
1064         .rpc_release = nfs_writeback_release_partial,
1065 };
1066
1067 /*
1068  * Handle a write reply that flushes a whole page.
1069  *
1070  * FIXME: There is an inherent race with invalidate_inode_pages and
1071  *        writebacks since the page->count is kept > 1 for as long
1072  *        as the page has a write request pending.
1073  */
1074 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1075 {
1076         struct nfs_write_data   *data = calldata;
1077
1078         nfs_writeback_done(task, data);
1079 }
1080
1081 static void nfs_writeback_release_full(void *calldata)
1082 {
1083         struct nfs_write_data   *data = calldata;
1084         int status = data->task.tk_status;
1085
1086         /* Update attributes as result of writeback. */
1087         while (!list_empty(&data->pages)) {
1088                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1089                 struct page *page = req->wb_page;
1090
1091                 nfs_list_remove_request(req);
1092
1093                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1094                         data->task.tk_pid,
1095                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1096                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1097                         req->wb_bytes,
1098                         (long long)req_offset(req));
1099
1100                 if (status < 0) {
1101                         nfs_set_pageerror(page);
1102                         nfs_context_set_write_error(req->wb_context, status);
1103                         dprintk(", error = %d\n", status);
1104                         goto remove_request;
1105                 }
1106
1107                 if (nfs_write_need_commit(data)) {
1108                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1109                         nfs_mark_request_commit(req);
1110                         dprintk(" marked for commit\n");
1111                         goto next;
1112                 }
1113                 dprintk(" OK\n");
1114 remove_request:
1115                 nfs_inode_remove_request(req);
1116         next:
1117                 nfs_clear_page_tag_locked(req);
1118                 nfs_end_page_writeback(page);
1119         }
1120         nfs_writedata_release(calldata);
1121 }
1122
1123 static const struct rpc_call_ops nfs_write_full_ops = {
1124 #if defined(CONFIG_NFS_V4_1)
1125         .rpc_call_prepare = nfs_write_prepare,
1126 #endif /* CONFIG_NFS_V4_1 */
1127         .rpc_call_done = nfs_writeback_done_full,
1128         .rpc_release = nfs_writeback_release_full,
1129 };
1130
1131
1132 /*
1133  * This function is called when the WRITE call is complete.
1134  */
1135 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1136 {
1137         struct nfs_writeargs    *argp = &data->args;
1138         struct nfs_writeres     *resp = &data->res;
1139         struct nfs_server       *server = NFS_SERVER(data->inode);
1140         int status;
1141
1142         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1143                 task->tk_pid, task->tk_status);
1144
1145         /*
1146          * ->write_done will attempt to use post-op attributes to detect
1147          * conflicting writes by other clients.  A strict interpretation
1148          * of close-to-open would allow us to continue caching even if
1149          * another writer had changed the file, but some applications
1150          * depend on tighter cache coherency when writing.
1151          */
1152         status = NFS_PROTO(data->inode)->write_done(task, data);
1153         if (status != 0)
1154                 return status;
1155         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1156
1157 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1158         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1159                 /* We tried a write call, but the server did not
1160                  * commit data to stable storage even though we
1161                  * requested it.
1162                  * Note: There is a known bug in Tru64 < 5.0 in which
1163                  *       the server reports NFS_DATA_SYNC, but performs
1164                  *       NFS_FILE_SYNC. We therefore implement this checking
1165                  *       as a dprintk() in order to avoid filling syslog.
1166                  */
1167                 static unsigned long    complain;
1168
1169                 if (time_before(complain, jiffies)) {
1170                         dprintk("NFS:       faulty NFS server %s:"
1171                                 " (committed = %d) != (stable = %d)\n",
1172                                 server->nfs_client->cl_hostname,
1173                                 resp->verf->committed, argp->stable);
1174                         complain = jiffies + 300 * HZ;
1175                 }
1176         }
1177 #endif
1178         /* Is this a short write? */
1179         if (task->tk_status >= 0 && resp->count < argp->count) {
1180                 static unsigned long    complain;
1181
1182                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1183
1184                 /* Has the server at least made some progress? */
1185                 if (resp->count != 0) {
1186                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1187                         if (resp->verf->committed != NFS_UNSTABLE) {
1188                                 /* Resend from where the server left off */
1189                                 argp->offset += resp->count;
1190                                 argp->pgbase += resp->count;
1191                                 argp->count -= resp->count;
1192                         } else {
1193                                 /* Resend as a stable write in order to avoid
1194                                  * headaches in the case of a server crash.
1195                                  */
1196                                 argp->stable = NFS_FILE_SYNC;
1197                         }
1198                         nfs_restart_rpc(task, server->nfs_client);
1199                         return -EAGAIN;
1200                 }
1201                 if (time_before(complain, jiffies)) {
1202                         printk(KERN_WARNING
1203                                "NFS: Server wrote zero bytes, expected %u.\n",
1204                                         argp->count);
1205                         complain = jiffies + 300 * HZ;
1206                 }
1207                 /* Can't do anything about it except throw an error. */
1208                 task->tk_status = -EIO;
1209         }
1210         return 0;
1211 }
1212
1213
1214 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1215 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1216 {
1217         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1218                 return 1;
1219         if (may_wait && !out_of_line_wait_on_bit_lock(&nfsi->flags,
1220                                 NFS_INO_COMMIT, nfs_wait_bit_killable,
1221                                 TASK_KILLABLE))
1222                 return 1;
1223         return 0;
1224 }
1225
1226 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1227 {
1228         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1229         smp_mb__after_clear_bit();
1230         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1231 }
1232
1233
1234 static void nfs_commitdata_release(void *data)
1235 {
1236         struct nfs_write_data *wdata = data;
1237
1238         put_nfs_open_context(wdata->args.context);
1239         nfs_commit_free(wdata);
1240 }
1241
1242 /*
1243  * Set up the argument/result storage required for the RPC call.
1244  */
1245 static int nfs_commit_rpcsetup(struct list_head *head,
1246                 struct nfs_write_data *data,
1247                 int how)
1248 {
1249         struct nfs_page *first = nfs_list_entry(head->next);
1250         struct inode *inode = first->wb_context->path.dentry->d_inode;
1251         int priority = flush_task_priority(how);
1252         struct rpc_task *task;
1253         struct rpc_message msg = {
1254                 .rpc_argp = &data->args,
1255                 .rpc_resp = &data->res,
1256                 .rpc_cred = first->wb_context->cred,
1257         };
1258         struct rpc_task_setup task_setup_data = {
1259                 .task = &data->task,
1260                 .rpc_client = NFS_CLIENT(inode),
1261                 .rpc_message = &msg,
1262                 .callback_ops = &nfs_commit_ops,
1263                 .callback_data = data,
1264                 .workqueue = nfsiod_workqueue,
1265                 .flags = RPC_TASK_ASYNC,
1266                 .priority = priority,
1267         };
1268
1269         /* Set up the RPC argument and reply structs
1270          * NB: take care not to mess about with data->commit et al. */
1271
1272         list_splice_init(head, &data->pages);
1273
1274         data->inode       = inode;
1275         data->cred        = msg.rpc_cred;
1276
1277         data->args.fh     = NFS_FH(data->inode);
1278         /* Note: we always request a commit of the entire inode */
1279         data->args.offset = 0;
1280         data->args.count  = 0;
1281         data->args.context = get_nfs_open_context(first->wb_context);
1282         data->res.count   = 0;
1283         data->res.fattr   = &data->fattr;
1284         data->res.verf    = &data->verf;
1285         nfs_fattr_init(&data->fattr);
1286
1287         /* Set up the initial task struct.  */
1288         NFS_PROTO(inode)->commit_setup(data, &msg);
1289
1290         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1291
1292         task = rpc_run_task(&task_setup_data);
1293         if (IS_ERR(task))
1294                 return PTR_ERR(task);
1295         rpc_put_task(task);
1296         return 0;
1297 }
1298
1299 /*
1300  * Commit dirty pages
1301  */
1302 static int
1303 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1304 {
1305         struct nfs_write_data   *data;
1306         struct nfs_page         *req;
1307
1308         data = nfs_commitdata_alloc();
1309
1310         if (!data)
1311                 goto out_bad;
1312
1313         /* Set up the argument struct */
1314         return nfs_commit_rpcsetup(head, data, how);
1315  out_bad:
1316         while (!list_empty(head)) {
1317                 req = nfs_list_entry(head->next);
1318                 nfs_list_remove_request(req);
1319                 nfs_mark_request_commit(req);
1320                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1321                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1322                                 BDI_RECLAIMABLE);
1323                 nfs_clear_page_tag_locked(req);
1324         }
1325         nfs_commit_clear_lock(NFS_I(inode));
1326         return -ENOMEM;
1327 }
1328
1329 /*
1330  * COMMIT call returned
1331  */
1332 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1333 {
1334         struct nfs_write_data   *data = calldata;
1335
1336         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1337                                 task->tk_pid, task->tk_status);
1338
1339         /* Call the NFS version-specific code */
1340         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1341                 return;
1342 }
1343
1344 static void nfs_commit_release(void *calldata)
1345 {
1346         struct nfs_write_data   *data = calldata;
1347         struct nfs_page         *req;
1348         int status = data->task.tk_status;
1349
1350         while (!list_empty(&data->pages)) {
1351                 req = nfs_list_entry(data->pages.next);
1352                 nfs_list_remove_request(req);
1353                 nfs_clear_request_commit(req);
1354
1355                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1356                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1357                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1358                         req->wb_bytes,
1359                         (long long)req_offset(req));
1360                 if (status < 0) {
1361                         nfs_context_set_write_error(req->wb_context, status);
1362                         nfs_inode_remove_request(req);
1363                         dprintk(", error = %d\n", status);
1364                         goto next;
1365                 }
1366
1367                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1368                  * returned by the server against all stored verfs. */
1369                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1370                         /* We have a match */
1371                         nfs_inode_remove_request(req);
1372                         dprintk(" OK\n");
1373                         goto next;
1374                 }
1375                 /* We have a mismatch. Write the page again */
1376                 dprintk(" mismatch\n");
1377                 nfs_mark_request_dirty(req);
1378         next:
1379                 nfs_clear_page_tag_locked(req);
1380         }
1381         nfs_commit_clear_lock(NFS_I(data->inode));
1382         nfs_commitdata_release(calldata);
1383 }
1384
1385 static const struct rpc_call_ops nfs_commit_ops = {
1386 #if defined(CONFIG_NFS_V4_1)
1387         .rpc_call_prepare = nfs_write_prepare,
1388 #endif /* CONFIG_NFS_V4_1 */
1389         .rpc_call_done = nfs_commit_done,
1390         .rpc_release = nfs_commit_release,
1391 };
1392
1393 int nfs_commit_inode(struct inode *inode, int how)
1394 {
1395         LIST_HEAD(head);
1396         int may_wait = how & FLUSH_SYNC;
1397         int res = 0;
1398
1399         if (!nfs_commit_set_lock(NFS_I(inode), may_wait))
1400                 goto out_mark_dirty;
1401         spin_lock(&inode->i_lock);
1402         res = nfs_scan_commit(inode, &head, 0, 0);
1403         spin_unlock(&inode->i_lock);
1404         if (res) {
1405                 int error = nfs_commit_list(inode, &head, how);
1406                 if (error < 0)
1407                         return error;
1408                 if (may_wait)
1409                         wait_on_bit(&NFS_I(inode)->flags, NFS_INO_COMMIT,
1410                                         nfs_wait_bit_killable,
1411                                         TASK_KILLABLE);
1412                 else
1413                         goto out_mark_dirty;
1414         } else
1415                 nfs_commit_clear_lock(NFS_I(inode));
1416         return res;
1417         /* Note: If we exit without ensuring that the commit is complete,
1418          * we must mark the inode as dirty. Otherwise, future calls to
1419          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1420          * that the data is on the disk.
1421          */
1422 out_mark_dirty:
1423         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1424         return res;
1425 }
1426
1427 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1428 {
1429         struct nfs_inode *nfsi = NFS_I(inode);
1430         int flags = FLUSH_SYNC;
1431         int ret = 0;
1432
1433         /* Don't commit yet if this is a non-blocking flush and there are
1434          * lots of outstanding writes for this mapping.
1435          */
1436         if (wbc->sync_mode == WB_SYNC_NONE &&
1437             nfsi->ncommit <= (nfsi->npages >> 1))
1438                 goto out_mark_dirty;
1439
1440         if (wbc->nonblocking || wbc->for_background)
1441                 flags = 0;
1442         ret = nfs_commit_inode(inode, flags);
1443         if (ret >= 0) {
1444                 if (wbc->sync_mode == WB_SYNC_NONE) {
1445                         if (ret < wbc->nr_to_write)
1446                                 wbc->nr_to_write -= ret;
1447                         else
1448                                 wbc->nr_to_write = 0;
1449                 }
1450                 return 0;
1451         }
1452 out_mark_dirty:
1453         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1454         return ret;
1455 }
1456 #else
1457 int nfs_commit_inode(struct inode *inode, int how)
1458 {
1459         return 0;
1460 }
1461
1462 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1463 {
1464         return 0;
1465 }
1466 #endif
1467
1468 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1469 {
1470         return nfs_commit_unstable_pages(inode, wbc);
1471 }
1472
1473 /*
1474  * flush the inode to disk.
1475  */
1476 int nfs_wb_all(struct inode *inode)
1477 {
1478         struct writeback_control wbc = {
1479                 .sync_mode = WB_SYNC_ALL,
1480                 .nr_to_write = LONG_MAX,
1481                 .range_start = 0,
1482                 .range_end = LLONG_MAX,
1483         };
1484
1485         return sync_inode(inode, &wbc);
1486 }
1487
1488 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1489 {
1490         struct nfs_page *req;
1491         int ret = 0;
1492
1493         BUG_ON(!PageLocked(page));
1494         for (;;) {
1495                 wait_on_page_writeback(page);
1496                 req = nfs_page_find_request(page);
1497                 if (req == NULL)
1498                         break;
1499                 if (nfs_lock_request_dontget(req)) {
1500                         nfs_inode_remove_request(req);
1501                         /*
1502                          * In case nfs_inode_remove_request has marked the
1503                          * page as being dirty
1504                          */
1505                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1506                         nfs_unlock_request(req);
1507                         break;
1508                 }
1509                 ret = nfs_wait_on_request(req);
1510                 nfs_release_request(req);
1511                 if (ret < 0)
1512                         break;
1513         }
1514         return ret;
1515 }
1516
1517 /*
1518  * Write back all requests on one page - we do this before reading it.
1519  */
1520 int nfs_wb_page(struct inode *inode, struct page *page)
1521 {
1522         loff_t range_start = page_offset(page);
1523         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1524         struct writeback_control wbc = {
1525                 .sync_mode = WB_SYNC_ALL,
1526                 .nr_to_write = 0,
1527                 .range_start = range_start,
1528                 .range_end = range_end,
1529         };
1530         int ret;
1531
1532         for (;;) {
1533                 wait_on_page_writeback(page);
1534                 if (clear_page_dirty_for_io(page)) {
1535                         ret = nfs_writepage_locked(page, &wbc);
1536                         if (ret < 0)
1537                                 goto out_error;
1538                         continue;
1539                 }
1540                 if (!PagePrivate(page))
1541                         break;
1542                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1543                 if (ret < 0)
1544                         goto out_error;
1545         }
1546         return 0;
1547 out_error:
1548         return ret;
1549 }
1550
1551 #ifdef CONFIG_MIGRATION
1552 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1553                 struct page *page)
1554 {
1555         struct nfs_page *req;
1556         int ret;
1557
1558         nfs_fscache_release_page(page, GFP_KERNEL);
1559
1560         req = nfs_find_and_lock_request(page, false);
1561         ret = PTR_ERR(req);
1562         if (IS_ERR(req))
1563                 goto out;
1564
1565         ret = migrate_page(mapping, newpage, page);
1566         if (!req)
1567                 goto out;
1568         if (ret)
1569                 goto out_unlock;
1570         page_cache_get(newpage);
1571         spin_lock(&mapping->host->i_lock);
1572         req->wb_page = newpage;
1573         SetPagePrivate(newpage);
1574         set_page_private(newpage, (unsigned long)req);
1575         ClearPagePrivate(page);
1576         set_page_private(page, 0);
1577         spin_unlock(&mapping->host->i_lock);
1578         page_cache_release(page);
1579 out_unlock:
1580         nfs_clear_page_tag_locked(req);
1581 out:
1582         return ret;
1583 }
1584 #endif
1585
1586 int __init nfs_init_writepagecache(void)
1587 {
1588         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1589                                              sizeof(struct nfs_write_data),
1590                                              0, SLAB_HWCACHE_ALIGN,
1591                                              NULL);
1592         if (nfs_wdata_cachep == NULL)
1593                 return -ENOMEM;
1594
1595         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1596                                                      nfs_wdata_cachep);
1597         if (nfs_wdata_mempool == NULL)
1598                 return -ENOMEM;
1599
1600         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1601                                                       nfs_wdata_cachep);
1602         if (nfs_commit_mempool == NULL)
1603                 return -ENOMEM;
1604
1605         /*
1606          * NFS congestion size, scale with available memory.
1607          *
1608          *  64MB:    8192k
1609          * 128MB:   11585k
1610          * 256MB:   16384k
1611          * 512MB:   23170k
1612          *   1GB:   32768k
1613          *   2GB:   46340k
1614          *   4GB:   65536k
1615          *   8GB:   92681k
1616          *  16GB:  131072k
1617          *
1618          * This allows larger machines to have larger/more transfers.
1619          * Limit the default to 256M
1620          */
1621         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1622         if (nfs_congestion_kb > 256*1024)
1623                 nfs_congestion_kb = 256*1024;
1624
1625         return 0;
1626 }
1627
1628 void nfs_destroy_writepagecache(void)
1629 {
1630         mempool_destroy(nfs_commit_mempool);
1631         mempool_destroy(nfs_wdata_mempool);
1632         kmem_cache_destroy(nfs_wdata_cachep);
1633 }
1634