0a6c65a1f9d786c6116c2698cde922ae887e8dd5
[linux-3.10.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)
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                 ret = nfs_wait_on_request(req);
245                 nfs_release_request(req);
246                 if (ret != 0)
247                         return ERR_PTR(ret);
248                 spin_lock(&inode->i_lock);
249         }
250         spin_unlock(&inode->i_lock);
251         return req;
252 }
253
254 /*
255  * Find an associated nfs write request, and prepare to flush it out
256  * May return an error if the user signalled nfs_wait_on_request().
257  */
258 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
259                                 struct page *page)
260 {
261         struct nfs_page *req;
262         int ret = 0;
263
264         req = nfs_find_and_lock_request(page);
265         if (!req)
266                 goto out;
267         ret = PTR_ERR(req);
268         if (IS_ERR(req))
269                 goto out;
270
271         ret = nfs_set_page_writeback(page);
272         BUG_ON(ret != 0);
273         BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
274
275         if (!nfs_pageio_add_request(pgio, req)) {
276                 nfs_redirty_request(req);
277                 ret = pgio->pg_error;
278         }
279 out:
280         return ret;
281 }
282
283 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
284 {
285         struct inode *inode = page->mapping->host;
286
287         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
288         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
289
290         nfs_pageio_cond_complete(pgio, page->index);
291         return nfs_page_async_flush(pgio, page);
292 }
293
294 /*
295  * Write an mmapped page to the server.
296  */
297 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
298 {
299         struct nfs_pageio_descriptor pgio;
300         int err;
301
302         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
303         err = nfs_do_writepage(page, wbc, &pgio);
304         nfs_pageio_complete(&pgio);
305         if (err < 0)
306                 return err;
307         if (pgio.pg_error < 0)
308                 return pgio.pg_error;
309         return 0;
310 }
311
312 int nfs_writepage(struct page *page, struct writeback_control *wbc)
313 {
314         int ret;
315
316         ret = nfs_writepage_locked(page, wbc);
317         unlock_page(page);
318         return ret;
319 }
320
321 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
322 {
323         int ret;
324
325         ret = nfs_do_writepage(page, wbc, data);
326         unlock_page(page);
327         return ret;
328 }
329
330 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
331 {
332         struct inode *inode = mapping->host;
333         unsigned long *bitlock = &NFS_I(inode)->flags;
334         struct nfs_pageio_descriptor pgio;
335         int err;
336
337         /* Stop dirtying of new pages while we sync */
338         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
339                         nfs_wait_bit_killable, TASK_KILLABLE);
340         if (err)
341                 goto out_err;
342
343         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
344
345         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
346         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
347         nfs_pageio_complete(&pgio);
348
349         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
350         smp_mb__after_clear_bit();
351         wake_up_bit(bitlock, NFS_INO_FLUSHING);
352
353         if (err < 0)
354                 goto out_err;
355         err = pgio.pg_error;
356         if (err < 0)
357                 goto out_err;
358         return 0;
359 out_err:
360         return err;
361 }
362
363 /*
364  * Insert a write request into an inode
365  */
366 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
367 {
368         struct nfs_inode *nfsi = NFS_I(inode);
369         int error;
370
371         error = radix_tree_preload(GFP_NOFS);
372         if (error != 0)
373                 goto out;
374
375         /* Lock the request! */
376         nfs_lock_request_dontget(req);
377
378         spin_lock(&inode->i_lock);
379         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
380         BUG_ON(error);
381         if (!nfsi->npages) {
382                 igrab(inode);
383                 if (nfs_have_delegation(inode, FMODE_WRITE))
384                         nfsi->change_attr++;
385         }
386         SetPagePrivate(req->wb_page);
387         set_page_private(req->wb_page, (unsigned long)req);
388         nfsi->npages++;
389         kref_get(&req->wb_kref);
390         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
391                                 NFS_PAGE_TAG_LOCKED);
392         spin_unlock(&inode->i_lock);
393         radix_tree_preload_end();
394 out:
395         return error;
396 }
397
398 /*
399  * Remove a write request from an inode
400  */
401 static void nfs_inode_remove_request(struct nfs_page *req)
402 {
403         struct inode *inode = req->wb_context->path.dentry->d_inode;
404         struct nfs_inode *nfsi = NFS_I(inode);
405
406         BUG_ON (!NFS_WBACK_BUSY(req));
407
408         spin_lock(&inode->i_lock);
409         set_page_private(req->wb_page, 0);
410         ClearPagePrivate(req->wb_page);
411         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
412         nfsi->npages--;
413         if (!nfsi->npages) {
414                 spin_unlock(&inode->i_lock);
415                 iput(inode);
416         } else
417                 spin_unlock(&inode->i_lock);
418         nfs_clear_request(req);
419         nfs_release_request(req);
420 }
421
422 static void
423 nfs_mark_request_dirty(struct nfs_page *req)
424 {
425         __set_page_dirty_nobuffers(req->wb_page);
426         __mark_inode_dirty(req->wb_page->mapping->host, I_DIRTY_DATASYNC);
427 }
428
429 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
430 /*
431  * Add a request to the inode's commit list.
432  */
433 static void
434 nfs_mark_request_commit(struct nfs_page *req)
435 {
436         struct inode *inode = req->wb_context->path.dentry->d_inode;
437         struct nfs_inode *nfsi = NFS_I(inode);
438
439         spin_lock(&inode->i_lock);
440         set_bit(PG_CLEAN, &(req)->wb_flags);
441         radix_tree_tag_set(&nfsi->nfs_page_tree,
442                         req->wb_index,
443                         NFS_PAGE_TAG_COMMIT);
444         nfsi->ncommit++;
445         spin_unlock(&inode->i_lock);
446         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
447         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
448         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
449 }
450
451 static int
452 nfs_clear_request_commit(struct nfs_page *req)
453 {
454         struct page *page = req->wb_page;
455
456         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
457                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
458                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
459                 return 1;
460         }
461         return 0;
462 }
463
464 static inline
465 int nfs_write_need_commit(struct nfs_write_data *data)
466 {
467         return data->verf.committed != NFS_FILE_SYNC;
468 }
469
470 static inline
471 int nfs_reschedule_unstable_write(struct nfs_page *req)
472 {
473         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
474                 nfs_mark_request_commit(req);
475                 return 1;
476         }
477         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
478                 nfs_mark_request_dirty(req);
479                 return 1;
480         }
481         return 0;
482 }
483 #else
484 static inline void
485 nfs_mark_request_commit(struct nfs_page *req)
486 {
487 }
488
489 static inline int
490 nfs_clear_request_commit(struct nfs_page *req)
491 {
492         return 0;
493 }
494
495 static inline
496 int nfs_write_need_commit(struct nfs_write_data *data)
497 {
498         return 0;
499 }
500
501 static inline
502 int nfs_reschedule_unstable_write(struct nfs_page *req)
503 {
504         return 0;
505 }
506 #endif
507
508 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
509 static int
510 nfs_need_commit(struct nfs_inode *nfsi)
511 {
512         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
513 }
514
515 /*
516  * nfs_scan_commit - Scan an inode for commit requests
517  * @inode: NFS inode to scan
518  * @dst: destination list
519  * @idx_start: lower bound of page->index to scan.
520  * @npages: idx_start + npages sets the upper bound to scan.
521  *
522  * Moves requests from the inode's 'commit' request list.
523  * The requests are *not* checked to ensure that they form a contiguous set.
524  */
525 static int
526 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
527 {
528         struct nfs_inode *nfsi = NFS_I(inode);
529         int ret;
530
531         if (!nfs_need_commit(nfsi))
532                 return 0;
533
534         ret = nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
535         if (ret > 0)
536                 nfsi->ncommit -= ret;
537         if (nfs_need_commit(NFS_I(inode)))
538                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
539         return ret;
540 }
541 #else
542 static inline int nfs_need_commit(struct nfs_inode *nfsi)
543 {
544         return 0;
545 }
546
547 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
548 {
549         return 0;
550 }
551 #endif
552
553 /*
554  * Search for an existing write request, and attempt to update
555  * it to reflect a new dirty region on a given page.
556  *
557  * If the attempt fails, then the existing request is flushed out
558  * to disk.
559  */
560 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
561                 struct page *page,
562                 unsigned int offset,
563                 unsigned int bytes)
564 {
565         struct nfs_page *req;
566         unsigned int rqend;
567         unsigned int end;
568         int error;
569
570         if (!PagePrivate(page))
571                 return NULL;
572
573         end = offset + bytes;
574         spin_lock(&inode->i_lock);
575
576         for (;;) {
577                 req = nfs_page_find_request_locked(page);
578                 if (req == NULL)
579                         goto out_unlock;
580
581                 rqend = req->wb_offset + req->wb_bytes;
582                 /*
583                  * Tell the caller to flush out the request if
584                  * the offsets are non-contiguous.
585                  * Note: nfs_flush_incompatible() will already
586                  * have flushed out requests having wrong owners.
587                  */
588                 if (offset > rqend
589                     || end < req->wb_offset)
590                         goto out_flushme;
591
592                 if (nfs_set_page_tag_locked(req))
593                         break;
594
595                 /* The request is locked, so wait and then retry */
596                 spin_unlock(&inode->i_lock);
597                 error = nfs_wait_on_request(req);
598                 nfs_release_request(req);
599                 if (error != 0)
600                         goto out_err;
601                 spin_lock(&inode->i_lock);
602         }
603
604         if (nfs_clear_request_commit(req) &&
605                         radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
606                                 req->wb_index, NFS_PAGE_TAG_COMMIT) != NULL)
607                 NFS_I(inode)->ncommit--;
608
609         /* Okay, the request matches. Update the region */
610         if (offset < req->wb_offset) {
611                 req->wb_offset = offset;
612                 req->wb_pgbase = offset;
613         }
614         if (end > rqend)
615                 req->wb_bytes = end - req->wb_offset;
616         else
617                 req->wb_bytes = rqend - req->wb_offset;
618 out_unlock:
619         spin_unlock(&inode->i_lock);
620         return req;
621 out_flushme:
622         spin_unlock(&inode->i_lock);
623         nfs_release_request(req);
624         error = nfs_wb_page(inode, page);
625 out_err:
626         return ERR_PTR(error);
627 }
628
629 /*
630  * Try to update an existing write request, or create one if there is none.
631  *
632  * Note: Should always be called with the Page Lock held to prevent races
633  * if we have to add a new request. Also assumes that the caller has
634  * already called nfs_flush_incompatible() if necessary.
635  */
636 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
637                 struct page *page, unsigned int offset, unsigned int bytes)
638 {
639         struct inode *inode = page->mapping->host;
640         struct nfs_page *req;
641         int error;
642
643         req = nfs_try_to_update_request(inode, page, offset, bytes);
644         if (req != NULL)
645                 goto out;
646         req = nfs_create_request(ctx, inode, page, offset, bytes);
647         if (IS_ERR(req))
648                 goto out;
649         error = nfs_inode_add_request(inode, req);
650         if (error != 0) {
651                 nfs_release_request(req);
652                 req = ERR_PTR(error);
653         }
654 out:
655         return req;
656 }
657
658 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
659                 unsigned int offset, unsigned int count)
660 {
661         struct nfs_page *req;
662
663         req = nfs_setup_write_request(ctx, page, offset, count);
664         if (IS_ERR(req))
665                 return PTR_ERR(req);
666         nfs_mark_request_dirty(req);
667         /* Update file length */
668         nfs_grow_file(page, offset, count);
669         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
670         nfs_mark_request_dirty(req);
671         nfs_clear_page_tag_locked(req);
672         return 0;
673 }
674
675 int nfs_flush_incompatible(struct file *file, struct page *page)
676 {
677         struct nfs_open_context *ctx = nfs_file_open_context(file);
678         struct nfs_page *req;
679         int do_flush, status;
680         /*
681          * Look for a request corresponding to this page. If there
682          * is one, and it belongs to another file, we flush it out
683          * before we try to copy anything into the page. Do this
684          * due to the lack of an ACCESS-type call in NFSv2.
685          * Also do the same if we find a request from an existing
686          * dropped page.
687          */
688         do {
689                 req = nfs_page_find_request(page);
690                 if (req == NULL)
691                         return 0;
692                 do_flush = req->wb_page != page || req->wb_context != ctx;
693                 nfs_release_request(req);
694                 if (!do_flush)
695                         return 0;
696                 status = nfs_wb_page(page->mapping->host, page);
697         } while (status == 0);
698         return status;
699 }
700
701 /*
702  * If the page cache is marked as unsafe or invalid, then we can't rely on
703  * the PageUptodate() flag. In this case, we will need to turn off
704  * write optimisations that depend on the page contents being correct.
705  */
706 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
707 {
708         return PageUptodate(page) &&
709                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
710 }
711
712 /*
713  * Update and possibly write a cached page of an NFS file.
714  *
715  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
716  * things with a page scheduled for an RPC call (e.g. invalidate it).
717  */
718 int nfs_updatepage(struct file *file, struct page *page,
719                 unsigned int offset, unsigned int count)
720 {
721         struct nfs_open_context *ctx = nfs_file_open_context(file);
722         struct inode    *inode = page->mapping->host;
723         int             status = 0;
724
725         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
726
727         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
728                 file->f_path.dentry->d_parent->d_name.name,
729                 file->f_path.dentry->d_name.name, count,
730                 (long long)(page_offset(page) + offset));
731
732         /* If we're not using byte range locks, and we know the page
733          * is up to date, it may be more efficient to extend the write
734          * to cover the entire page in order to avoid fragmentation
735          * inefficiencies.
736          */
737         if (nfs_write_pageuptodate(page, inode) &&
738                         inode->i_flock == NULL &&
739                         !(file->f_flags & O_DSYNC)) {
740                 count = max(count + offset, nfs_page_length(page));
741                 offset = 0;
742         }
743
744         status = nfs_writepage_setup(ctx, page, offset, count);
745         if (status < 0)
746                 nfs_set_pageerror(page);
747
748         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
749                         status, (long long)i_size_read(inode));
750         return status;
751 }
752
753 static void nfs_writepage_release(struct nfs_page *req)
754 {
755         struct page *page = req->wb_page;
756
757         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req))
758                 nfs_inode_remove_request(req);
759         nfs_clear_page_tag_locked(req);
760         nfs_end_page_writeback(page);
761 }
762
763 static int flush_task_priority(int how)
764 {
765         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
766                 case FLUSH_HIGHPRI:
767                         return RPC_PRIORITY_HIGH;
768                 case FLUSH_LOWPRI:
769                         return RPC_PRIORITY_LOW;
770         }
771         return RPC_PRIORITY_NORMAL;
772 }
773
774 /*
775  * Set up the argument/result storage required for the RPC call.
776  */
777 static int nfs_write_rpcsetup(struct nfs_page *req,
778                 struct nfs_write_data *data,
779                 const struct rpc_call_ops *call_ops,
780                 unsigned int count, unsigned int offset,
781                 int how)
782 {
783         struct inode *inode = req->wb_context->path.dentry->d_inode;
784         int priority = flush_task_priority(how);
785         struct rpc_task *task;
786         struct rpc_message msg = {
787                 .rpc_argp = &data->args,
788                 .rpc_resp = &data->res,
789                 .rpc_cred = req->wb_context->cred,
790         };
791         struct rpc_task_setup task_setup_data = {
792                 .rpc_client = NFS_CLIENT(inode),
793                 .task = &data->task,
794                 .rpc_message = &msg,
795                 .callback_ops = call_ops,
796                 .callback_data = data,
797                 .workqueue = nfsiod_workqueue,
798                 .flags = RPC_TASK_ASYNC,
799                 .priority = priority,
800         };
801         int ret = 0;
802
803         /* Set up the RPC argument and reply structs
804          * NB: take care not to mess about with data->commit et al. */
805
806         data->req = req;
807         data->inode = inode = req->wb_context->path.dentry->d_inode;
808         data->cred = msg.rpc_cred;
809
810         data->args.fh     = NFS_FH(inode);
811         data->args.offset = req_offset(req) + offset;
812         data->args.pgbase = req->wb_pgbase + offset;
813         data->args.pages  = data->pagevec;
814         data->args.count  = count;
815         data->args.context = get_nfs_open_context(req->wb_context);
816         data->args.stable  = NFS_UNSTABLE;
817         if (how & FLUSH_STABLE) {
818                 data->args.stable = NFS_DATA_SYNC;
819                 if (!nfs_need_commit(NFS_I(inode)))
820                         data->args.stable = NFS_FILE_SYNC;
821         }
822
823         data->res.fattr   = &data->fattr;
824         data->res.count   = count;
825         data->res.verf    = &data->verf;
826         nfs_fattr_init(&data->fattr);
827
828         /* Set up the initial task struct.  */
829         NFS_PROTO(inode)->write_setup(data, &msg);
830
831         dprintk("NFS: %5u initiated write call "
832                 "(req %s/%lld, %u bytes @ offset %llu)\n",
833                 data->task.tk_pid,
834                 inode->i_sb->s_id,
835                 (long long)NFS_FILEID(inode),
836                 count,
837                 (unsigned long long)data->args.offset);
838
839         task = rpc_run_task(&task_setup_data);
840         if (IS_ERR(task)) {
841                 ret = PTR_ERR(task);
842                 goto out;
843         }
844         if (how & FLUSH_SYNC) {
845                 ret = rpc_wait_for_completion_task(task);
846                 if (ret == 0)
847                         ret = task->tk_status;
848         }
849         rpc_put_task(task);
850 out:
851         return ret;
852 }
853
854 /* If a nfs_flush_* function fails, it should remove reqs from @head and
855  * call this on each, which will prepare them to be retried on next
856  * writeback using standard nfs.
857  */
858 static void nfs_redirty_request(struct nfs_page *req)
859 {
860         struct page *page = req->wb_page;
861
862         nfs_mark_request_dirty(req);
863         nfs_clear_page_tag_locked(req);
864         nfs_end_page_writeback(page);
865 }
866
867 /*
868  * Generate multiple small requests to write out a single
869  * contiguous dirty area on one page.
870  */
871 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
872 {
873         struct nfs_page *req = nfs_list_entry(head->next);
874         struct page *page = req->wb_page;
875         struct nfs_write_data *data;
876         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
877         unsigned int offset;
878         int requests = 0;
879         int ret = 0;
880         LIST_HEAD(list);
881
882         nfs_list_remove_request(req);
883
884         nbytes = count;
885         do {
886                 size_t len = min(nbytes, wsize);
887
888                 data = nfs_writedata_alloc(1);
889                 if (!data)
890                         goto out_bad;
891                 list_add(&data->pages, &list);
892                 requests++;
893                 nbytes -= len;
894         } while (nbytes != 0);
895         atomic_set(&req->wb_complete, requests);
896
897         ClearPageError(page);
898         offset = 0;
899         nbytes = count;
900         do {
901                 int ret2;
902
903                 data = list_entry(list.next, struct nfs_write_data, pages);
904                 list_del_init(&data->pages);
905
906                 data->pagevec[0] = page;
907
908                 if (nbytes < wsize)
909                         wsize = nbytes;
910                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
911                                    wsize, offset, how);
912                 if (ret == 0)
913                         ret = ret2;
914                 offset += wsize;
915                 nbytes -= wsize;
916         } while (nbytes != 0);
917
918         return ret;
919
920 out_bad:
921         while (!list_empty(&list)) {
922                 data = list_entry(list.next, struct nfs_write_data, pages);
923                 list_del(&data->pages);
924                 nfs_writedata_release(data);
925         }
926         nfs_redirty_request(req);
927         return -ENOMEM;
928 }
929
930 /*
931  * Create an RPC task for the given write request and kick it.
932  * The page must have been locked by the caller.
933  *
934  * It may happen that the page we're passed is not marked dirty.
935  * This is the case if nfs_updatepage detects a conflicting request
936  * that has been written but not committed.
937  */
938 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
939 {
940         struct nfs_page         *req;
941         struct page             **pages;
942         struct nfs_write_data   *data;
943
944         data = nfs_writedata_alloc(npages);
945         if (!data)
946                 goto out_bad;
947
948         pages = data->pagevec;
949         while (!list_empty(head)) {
950                 req = nfs_list_entry(head->next);
951                 nfs_list_remove_request(req);
952                 nfs_list_add_request(req, &data->pages);
953                 ClearPageError(req->wb_page);
954                 *pages++ = req->wb_page;
955         }
956         req = nfs_list_entry(data->pages.next);
957
958         /* Set up the argument struct */
959         return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
960  out_bad:
961         while (!list_empty(head)) {
962                 req = nfs_list_entry(head->next);
963                 nfs_list_remove_request(req);
964                 nfs_redirty_request(req);
965         }
966         return -ENOMEM;
967 }
968
969 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
970                                   struct inode *inode, int ioflags)
971 {
972         size_t wsize = NFS_SERVER(inode)->wsize;
973
974         if (wsize < PAGE_CACHE_SIZE)
975                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
976         else
977                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
978 }
979
980 /*
981  * Handle a write reply that flushed part of a page.
982  */
983 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
984 {
985         struct nfs_write_data   *data = calldata;
986
987         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
988                 task->tk_pid,
989                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
990                 (long long)
991                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
992                 data->req->wb_bytes, (long long)req_offset(data->req));
993
994         nfs_writeback_done(task, data);
995 }
996
997 static void nfs_writeback_release_partial(void *calldata)
998 {
999         struct nfs_write_data   *data = calldata;
1000         struct nfs_page         *req = data->req;
1001         struct page             *page = req->wb_page;
1002         int status = data->task.tk_status;
1003
1004         if (status < 0) {
1005                 nfs_set_pageerror(page);
1006                 nfs_context_set_write_error(req->wb_context, status);
1007                 dprintk(", error = %d\n", status);
1008                 goto out;
1009         }
1010
1011         if (nfs_write_need_commit(data)) {
1012                 struct inode *inode = page->mapping->host;
1013
1014                 spin_lock(&inode->i_lock);
1015                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1016                         /* Do nothing we need to resend the writes */
1017                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1018                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1019                         dprintk(" defer commit\n");
1020                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1021                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1022                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1023                         dprintk(" server reboot detected\n");
1024                 }
1025                 spin_unlock(&inode->i_lock);
1026         } else
1027                 dprintk(" OK\n");
1028
1029 out:
1030         if (atomic_dec_and_test(&req->wb_complete))
1031                 nfs_writepage_release(req);
1032         nfs_writedata_release(calldata);
1033 }
1034
1035 #if defined(CONFIG_NFS_V4_1)
1036 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1037 {
1038         struct nfs_write_data *data = calldata;
1039         struct nfs_client *clp = (NFS_SERVER(data->inode))->nfs_client;
1040
1041         if (nfs4_setup_sequence(clp, &data->args.seq_args,
1042                                 &data->res.seq_res, 1, task))
1043                 return;
1044         rpc_call_start(task);
1045 }
1046 #endif /* CONFIG_NFS_V4_1 */
1047
1048 static const struct rpc_call_ops nfs_write_partial_ops = {
1049 #if defined(CONFIG_NFS_V4_1)
1050         .rpc_call_prepare = nfs_write_prepare,
1051 #endif /* CONFIG_NFS_V4_1 */
1052         .rpc_call_done = nfs_writeback_done_partial,
1053         .rpc_release = nfs_writeback_release_partial,
1054 };
1055
1056 /*
1057  * Handle a write reply that flushes a whole page.
1058  *
1059  * FIXME: There is an inherent race with invalidate_inode_pages and
1060  *        writebacks since the page->count is kept > 1 for as long
1061  *        as the page has a write request pending.
1062  */
1063 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1064 {
1065         struct nfs_write_data   *data = calldata;
1066
1067         nfs_writeback_done(task, data);
1068 }
1069
1070 static void nfs_writeback_release_full(void *calldata)
1071 {
1072         struct nfs_write_data   *data = calldata;
1073         int status = data->task.tk_status;
1074
1075         /* Update attributes as result of writeback. */
1076         while (!list_empty(&data->pages)) {
1077                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1078                 struct page *page = req->wb_page;
1079
1080                 nfs_list_remove_request(req);
1081
1082                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1083                         data->task.tk_pid,
1084                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1085                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1086                         req->wb_bytes,
1087                         (long long)req_offset(req));
1088
1089                 if (status < 0) {
1090                         nfs_set_pageerror(page);
1091                         nfs_context_set_write_error(req->wb_context, status);
1092                         dprintk(", error = %d\n", status);
1093                         goto remove_request;
1094                 }
1095
1096                 if (nfs_write_need_commit(data)) {
1097                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1098                         nfs_mark_request_commit(req);
1099                         dprintk(" marked for commit\n");
1100                         goto next;
1101                 }
1102                 dprintk(" OK\n");
1103 remove_request:
1104                 nfs_inode_remove_request(req);
1105         next:
1106                 nfs_clear_page_tag_locked(req);
1107                 nfs_end_page_writeback(page);
1108         }
1109         nfs_writedata_release(calldata);
1110 }
1111
1112 static const struct rpc_call_ops nfs_write_full_ops = {
1113 #if defined(CONFIG_NFS_V4_1)
1114         .rpc_call_prepare = nfs_write_prepare,
1115 #endif /* CONFIG_NFS_V4_1 */
1116         .rpc_call_done = nfs_writeback_done_full,
1117         .rpc_release = nfs_writeback_release_full,
1118 };
1119
1120
1121 /*
1122  * This function is called when the WRITE call is complete.
1123  */
1124 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1125 {
1126         struct nfs_writeargs    *argp = &data->args;
1127         struct nfs_writeres     *resp = &data->res;
1128         struct nfs_server       *server = NFS_SERVER(data->inode);
1129         int status;
1130
1131         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1132                 task->tk_pid, task->tk_status);
1133
1134         /*
1135          * ->write_done will attempt to use post-op attributes to detect
1136          * conflicting writes by other clients.  A strict interpretation
1137          * of close-to-open would allow us to continue caching even if
1138          * another writer had changed the file, but some applications
1139          * depend on tighter cache coherency when writing.
1140          */
1141         status = NFS_PROTO(data->inode)->write_done(task, data);
1142         if (status != 0)
1143                 return status;
1144         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1145
1146 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1147         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1148                 /* We tried a write call, but the server did not
1149                  * commit data to stable storage even though we
1150                  * requested it.
1151                  * Note: There is a known bug in Tru64 < 5.0 in which
1152                  *       the server reports NFS_DATA_SYNC, but performs
1153                  *       NFS_FILE_SYNC. We therefore implement this checking
1154                  *       as a dprintk() in order to avoid filling syslog.
1155                  */
1156                 static unsigned long    complain;
1157
1158                 if (time_before(complain, jiffies)) {
1159                         dprintk("NFS:       faulty NFS server %s:"
1160                                 " (committed = %d) != (stable = %d)\n",
1161                                 server->nfs_client->cl_hostname,
1162                                 resp->verf->committed, argp->stable);
1163                         complain = jiffies + 300 * HZ;
1164                 }
1165         }
1166 #endif
1167         /* Is this a short write? */
1168         if (task->tk_status >= 0 && resp->count < argp->count) {
1169                 static unsigned long    complain;
1170
1171                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1172
1173                 /* Has the server at least made some progress? */
1174                 if (resp->count != 0) {
1175                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1176                         if (resp->verf->committed != NFS_UNSTABLE) {
1177                                 /* Resend from where the server left off */
1178                                 argp->offset += resp->count;
1179                                 argp->pgbase += resp->count;
1180                                 argp->count -= resp->count;
1181                         } else {
1182                                 /* Resend as a stable write in order to avoid
1183                                  * headaches in the case of a server crash.
1184                                  */
1185                                 argp->stable = NFS_FILE_SYNC;
1186                         }
1187                         nfs_restart_rpc(task, server->nfs_client);
1188                         return -EAGAIN;
1189                 }
1190                 if (time_before(complain, jiffies)) {
1191                         printk(KERN_WARNING
1192                                "NFS: Server wrote zero bytes, expected %u.\n",
1193                                         argp->count);
1194                         complain = jiffies + 300 * HZ;
1195                 }
1196                 /* Can't do anything about it except throw an error. */
1197                 task->tk_status = -EIO;
1198         }
1199         return 0;
1200 }
1201
1202
1203 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1204 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1205 {
1206         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1207                 return 1;
1208         if (may_wait && !out_of_line_wait_on_bit_lock(&nfsi->flags,
1209                                 NFS_INO_COMMIT, nfs_wait_bit_killable,
1210                                 TASK_KILLABLE))
1211                 return 1;
1212         return 0;
1213 }
1214
1215 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1216 {
1217         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1218         smp_mb__after_clear_bit();
1219         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1220 }
1221
1222
1223 static void nfs_commitdata_release(void *data)
1224 {
1225         struct nfs_write_data *wdata = data;
1226
1227         put_nfs_open_context(wdata->args.context);
1228         nfs_commit_free(wdata);
1229 }
1230
1231 /*
1232  * Set up the argument/result storage required for the RPC call.
1233  */
1234 static int nfs_commit_rpcsetup(struct list_head *head,
1235                 struct nfs_write_data *data,
1236                 int how)
1237 {
1238         struct nfs_page *first = nfs_list_entry(head->next);
1239         struct inode *inode = first->wb_context->path.dentry->d_inode;
1240         int priority = flush_task_priority(how);
1241         struct rpc_task *task;
1242         struct rpc_message msg = {
1243                 .rpc_argp = &data->args,
1244                 .rpc_resp = &data->res,
1245                 .rpc_cred = first->wb_context->cred,
1246         };
1247         struct rpc_task_setup task_setup_data = {
1248                 .task = &data->task,
1249                 .rpc_client = NFS_CLIENT(inode),
1250                 .rpc_message = &msg,
1251                 .callback_ops = &nfs_commit_ops,
1252                 .callback_data = data,
1253                 .workqueue = nfsiod_workqueue,
1254                 .flags = RPC_TASK_ASYNC,
1255                 .priority = priority,
1256         };
1257
1258         /* Set up the RPC argument and reply structs
1259          * NB: take care not to mess about with data->commit et al. */
1260
1261         list_splice_init(head, &data->pages);
1262
1263         data->inode       = inode;
1264         data->cred        = msg.rpc_cred;
1265
1266         data->args.fh     = NFS_FH(data->inode);
1267         /* Note: we always request a commit of the entire inode */
1268         data->args.offset = 0;
1269         data->args.count  = 0;
1270         data->args.context = get_nfs_open_context(first->wb_context);
1271         data->res.count   = 0;
1272         data->res.fattr   = &data->fattr;
1273         data->res.verf    = &data->verf;
1274         nfs_fattr_init(&data->fattr);
1275
1276         /* Set up the initial task struct.  */
1277         NFS_PROTO(inode)->commit_setup(data, &msg);
1278
1279         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1280
1281         task = rpc_run_task(&task_setup_data);
1282         if (IS_ERR(task))
1283                 return PTR_ERR(task);
1284         rpc_put_task(task);
1285         return 0;
1286 }
1287
1288 /*
1289  * Commit dirty pages
1290  */
1291 static int
1292 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1293 {
1294         struct nfs_write_data   *data;
1295         struct nfs_page         *req;
1296
1297         data = nfs_commitdata_alloc();
1298
1299         if (!data)
1300                 goto out_bad;
1301
1302         /* Set up the argument struct */
1303         return nfs_commit_rpcsetup(head, data, how);
1304  out_bad:
1305         while (!list_empty(head)) {
1306                 req = nfs_list_entry(head->next);
1307                 nfs_list_remove_request(req);
1308                 nfs_mark_request_commit(req);
1309                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1310                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1311                                 BDI_RECLAIMABLE);
1312                 nfs_clear_page_tag_locked(req);
1313         }
1314         nfs_commit_clear_lock(NFS_I(inode));
1315         return -ENOMEM;
1316 }
1317
1318 /*
1319  * COMMIT call returned
1320  */
1321 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1322 {
1323         struct nfs_write_data   *data = calldata;
1324
1325         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1326                                 task->tk_pid, task->tk_status);
1327
1328         /* Call the NFS version-specific code */
1329         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1330                 return;
1331 }
1332
1333 static void nfs_commit_release(void *calldata)
1334 {
1335         struct nfs_write_data   *data = calldata;
1336         struct nfs_page         *req;
1337         int status = data->task.tk_status;
1338
1339         while (!list_empty(&data->pages)) {
1340                 req = nfs_list_entry(data->pages.next);
1341                 nfs_list_remove_request(req);
1342                 nfs_clear_request_commit(req);
1343
1344                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1345                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1346                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1347                         req->wb_bytes,
1348                         (long long)req_offset(req));
1349                 if (status < 0) {
1350                         nfs_context_set_write_error(req->wb_context, status);
1351                         nfs_inode_remove_request(req);
1352                         dprintk(", error = %d\n", status);
1353                         goto next;
1354                 }
1355
1356                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1357                  * returned by the server against all stored verfs. */
1358                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1359                         /* We have a match */
1360                         nfs_inode_remove_request(req);
1361                         dprintk(" OK\n");
1362                         goto next;
1363                 }
1364                 /* We have a mismatch. Write the page again */
1365                 dprintk(" mismatch\n");
1366                 nfs_mark_request_dirty(req);
1367         next:
1368                 nfs_clear_page_tag_locked(req);
1369         }
1370         nfs_commit_clear_lock(NFS_I(data->inode));
1371         nfs_commitdata_release(calldata);
1372 }
1373
1374 static const struct rpc_call_ops nfs_commit_ops = {
1375 #if defined(CONFIG_NFS_V4_1)
1376         .rpc_call_prepare = nfs_write_prepare,
1377 #endif /* CONFIG_NFS_V4_1 */
1378         .rpc_call_done = nfs_commit_done,
1379         .rpc_release = nfs_commit_release,
1380 };
1381
1382 int nfs_commit_inode(struct inode *inode, int how)
1383 {
1384         LIST_HEAD(head);
1385         int may_wait = how & FLUSH_SYNC;
1386         int res = 0;
1387
1388         if (!nfs_commit_set_lock(NFS_I(inode), may_wait))
1389                 goto out_mark_dirty;
1390         spin_lock(&inode->i_lock);
1391         res = nfs_scan_commit(inode, &head, 0, 0);
1392         spin_unlock(&inode->i_lock);
1393         if (res) {
1394                 int error = nfs_commit_list(inode, &head, how);
1395                 if (error < 0)
1396                         return error;
1397                 if (may_wait)
1398                         wait_on_bit(&NFS_I(inode)->flags, NFS_INO_COMMIT,
1399                                         nfs_wait_bit_killable,
1400                                         TASK_KILLABLE);
1401                 else
1402                         goto out_mark_dirty;
1403         } else
1404                 nfs_commit_clear_lock(NFS_I(inode));
1405         return res;
1406         /* Note: If we exit without ensuring that the commit is complete,
1407          * we must mark the inode as dirty. Otherwise, future calls to
1408          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1409          * that the data is on the disk.
1410          */
1411 out_mark_dirty:
1412         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1413         return res;
1414 }
1415
1416 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1417 {
1418         struct nfs_inode *nfsi = NFS_I(inode);
1419         int flags = FLUSH_SYNC;
1420         int ret = 0;
1421
1422         /* Don't commit yet if this is a non-blocking flush and there are
1423          * lots of outstanding writes for this mapping.
1424          */
1425         if (wbc->sync_mode == WB_SYNC_NONE &&
1426             nfsi->ncommit <= (nfsi->npages >> 1))
1427                 goto out_mark_dirty;
1428
1429         if (wbc->nonblocking || wbc->for_background)
1430                 flags = 0;
1431         ret = nfs_commit_inode(inode, flags);
1432         if (ret >= 0) {
1433                 if (wbc->sync_mode == WB_SYNC_NONE) {
1434                         if (ret < wbc->nr_to_write)
1435                                 wbc->nr_to_write -= ret;
1436                         else
1437                                 wbc->nr_to_write = 0;
1438                 }
1439                 return 0;
1440         }
1441 out_mark_dirty:
1442         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1443         return ret;
1444 }
1445 #else
1446 int nfs_commit_inode(struct inode *inode, int how)
1447 {
1448         return 0;
1449 }
1450
1451 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1452 {
1453         return 0;
1454 }
1455 #endif
1456
1457 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1458 {
1459         return nfs_commit_unstable_pages(inode, wbc);
1460 }
1461
1462 /*
1463  * flush the inode to disk.
1464  */
1465 int nfs_wb_all(struct inode *inode)
1466 {
1467         struct writeback_control wbc = {
1468                 .sync_mode = WB_SYNC_ALL,
1469                 .nr_to_write = LONG_MAX,
1470                 .range_start = 0,
1471                 .range_end = LLONG_MAX,
1472         };
1473
1474         return sync_inode(inode, &wbc);
1475 }
1476
1477 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1478 {
1479         struct nfs_page *req;
1480         int ret = 0;
1481
1482         BUG_ON(!PageLocked(page));
1483         for (;;) {
1484                 wait_on_page_writeback(page);
1485                 req = nfs_page_find_request(page);
1486                 if (req == NULL)
1487                         break;
1488                 if (nfs_lock_request_dontget(req)) {
1489                         nfs_inode_remove_request(req);
1490                         /*
1491                          * In case nfs_inode_remove_request has marked the
1492                          * page as being dirty
1493                          */
1494                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1495                         nfs_unlock_request(req);
1496                         break;
1497                 }
1498                 ret = nfs_wait_on_request(req);
1499                 nfs_release_request(req);
1500                 if (ret < 0)
1501                         break;
1502         }
1503         return ret;
1504 }
1505
1506 /*
1507  * Write back all requests on one page - we do this before reading it.
1508  */
1509 int nfs_wb_page(struct inode *inode, struct page *page)
1510 {
1511         loff_t range_start = page_offset(page);
1512         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1513         struct writeback_control wbc = {
1514                 .sync_mode = WB_SYNC_ALL,
1515                 .nr_to_write = 0,
1516                 .range_start = range_start,
1517                 .range_end = range_end,
1518         };
1519         int ret;
1520
1521         for (;;) {
1522                 wait_on_page_writeback(page);
1523                 if (clear_page_dirty_for_io(page)) {
1524                         ret = nfs_writepage_locked(page, &wbc);
1525                         if (ret < 0)
1526                                 goto out_error;
1527                         continue;
1528                 }
1529                 if (!PagePrivate(page))
1530                         break;
1531                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1532                 if (ret < 0)
1533                         goto out_error;
1534         }
1535         return 0;
1536 out_error:
1537         return ret;
1538 }
1539
1540 #ifdef CONFIG_MIGRATION
1541 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1542                 struct page *page)
1543 {
1544         struct nfs_page *req;
1545         int ret;
1546
1547         nfs_fscache_release_page(page, GFP_KERNEL);
1548
1549         req = nfs_find_and_lock_request(page);
1550         ret = PTR_ERR(req);
1551         if (IS_ERR(req))
1552                 goto out;
1553
1554         ret = migrate_page(mapping, newpage, page);
1555         if (!req)
1556                 goto out;
1557         if (ret)
1558                 goto out_unlock;
1559         page_cache_get(newpage);
1560         spin_lock(&mapping->host->i_lock);
1561         req->wb_page = newpage;
1562         SetPagePrivate(newpage);
1563         set_page_private(newpage, (unsigned long)req);
1564         ClearPagePrivate(page);
1565         set_page_private(page, 0);
1566         spin_unlock(&mapping->host->i_lock);
1567         page_cache_release(page);
1568 out_unlock:
1569         nfs_clear_page_tag_locked(req);
1570 out:
1571         return ret;
1572 }
1573 #endif
1574
1575 int __init nfs_init_writepagecache(void)
1576 {
1577         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1578                                              sizeof(struct nfs_write_data),
1579                                              0, SLAB_HWCACHE_ALIGN,
1580                                              NULL);
1581         if (nfs_wdata_cachep == NULL)
1582                 return -ENOMEM;
1583
1584         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1585                                                      nfs_wdata_cachep);
1586         if (nfs_wdata_mempool == NULL)
1587                 return -ENOMEM;
1588
1589         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1590                                                       nfs_wdata_cachep);
1591         if (nfs_commit_mempool == NULL)
1592                 return -ENOMEM;
1593
1594         /*
1595          * NFS congestion size, scale with available memory.
1596          *
1597          *  64MB:    8192k
1598          * 128MB:   11585k
1599          * 256MB:   16384k
1600          * 512MB:   23170k
1601          *   1GB:   32768k
1602          *   2GB:   46340k
1603          *   4GB:   65536k
1604          *   8GB:   92681k
1605          *  16GB:  131072k
1606          *
1607          * This allows larger machines to have larger/more transfers.
1608          * Limit the default to 256M
1609          */
1610         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1611         if (nfs_congestion_kb > 256*1024)
1612                 nfs_congestion_kb = 256*1024;
1613
1614         return 0;
1615 }
1616
1617 void nfs_destroy_writepagecache(void)
1618 {
1619         mempool_destroy(nfs_commit_mempool);
1620         mempool_destroy(nfs_wdata_mempool);
1621         kmem_cache_destroy(nfs_wdata_cachep);
1622 }
1623