NFS: Simplify nfs_wb_page_cancel()
[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                 if (atomic_long_inc_return(&nfss->writeback) >
205                                 NFS_CONGESTION_ON_THRESH) {
206                         set_bdi_congested(&nfss->backing_dev_info,
207                                                 BLK_RW_ASYNC);
208                 }
209         }
210         return ret;
211 }
212
213 static void nfs_end_page_writeback(struct page *page)
214 {
215         struct inode *inode = page->mapping->host;
216         struct nfs_server *nfss = NFS_SERVER(inode);
217
218         end_page_writeback(page);
219         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
220                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
221 }
222
223 static struct nfs_page *nfs_find_and_lock_request(struct page *page)
224 {
225         struct inode *inode = page->mapping->host;
226         struct nfs_page *req;
227         int ret;
228
229         spin_lock(&inode->i_lock);
230         for (;;) {
231                 req = nfs_page_find_request_locked(page);
232                 if (req == NULL)
233                         break;
234                 if (nfs_set_page_tag_locked(req))
235                         break;
236                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
237                  *       then the call to nfs_set_page_tag_locked() will always
238                  *       succeed provided that someone hasn't already marked the
239                  *       request as dirty (in which case we don't care).
240                  */
241                 spin_unlock(&inode->i_lock);
242                 ret = nfs_wait_on_request(req);
243                 nfs_release_request(req);
244                 if (ret != 0)
245                         return ERR_PTR(ret);
246                 spin_lock(&inode->i_lock);
247         }
248         spin_unlock(&inode->i_lock);
249         return req;
250 }
251
252 /*
253  * Find an associated nfs write request, and prepare to flush it out
254  * May return an error if the user signalled nfs_wait_on_request().
255  */
256 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
257                                 struct page *page)
258 {
259         struct nfs_page *req;
260         int ret = 0;
261
262         req = nfs_find_and_lock_request(page);
263         if (!req)
264                 goto out;
265         ret = PTR_ERR(req);
266         if (IS_ERR(req))
267                 goto out;
268
269         ret = nfs_set_page_writeback(page);
270         BUG_ON(ret != 0);
271         BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
272
273         if (!nfs_pageio_add_request(pgio, req)) {
274                 nfs_redirty_request(req);
275                 ret = pgio->pg_error;
276         }
277 out:
278         return ret;
279 }
280
281 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
282 {
283         struct inode *inode = page->mapping->host;
284
285         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
286         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
287
288         nfs_pageio_cond_complete(pgio, page->index);
289         return nfs_page_async_flush(pgio, page);
290 }
291
292 /*
293  * Write an mmapped page to the server.
294  */
295 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
296 {
297         struct nfs_pageio_descriptor pgio;
298         int err;
299
300         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
301         err = nfs_do_writepage(page, wbc, &pgio);
302         nfs_pageio_complete(&pgio);
303         if (err < 0)
304                 return err;
305         if (pgio.pg_error < 0)
306                 return pgio.pg_error;
307         return 0;
308 }
309
310 int nfs_writepage(struct page *page, struct writeback_control *wbc)
311 {
312         int ret;
313
314         ret = nfs_writepage_locked(page, wbc);
315         unlock_page(page);
316         return ret;
317 }
318
319 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
320 {
321         int ret;
322
323         ret = nfs_do_writepage(page, wbc, data);
324         unlock_page(page);
325         return ret;
326 }
327
328 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
329 {
330         struct inode *inode = mapping->host;
331         unsigned long *bitlock = &NFS_I(inode)->flags;
332         struct nfs_pageio_descriptor pgio;
333         int err;
334
335         /* Stop dirtying of new pages while we sync */
336         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
337                         nfs_wait_bit_killable, TASK_KILLABLE);
338         if (err)
339                 goto out_err;
340
341         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
342
343         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
344         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
345         nfs_pageio_complete(&pgio);
346
347         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
348         smp_mb__after_clear_bit();
349         wake_up_bit(bitlock, NFS_INO_FLUSHING);
350
351         if (err < 0)
352                 goto out_err;
353         err = pgio.pg_error;
354         if (err < 0)
355                 goto out_err;
356         return 0;
357 out_err:
358         return err;
359 }
360
361 /*
362  * Insert a write request into an inode
363  */
364 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
365 {
366         struct nfs_inode *nfsi = NFS_I(inode);
367         int error;
368
369         error = radix_tree_preload(GFP_NOFS);
370         if (error != 0)
371                 goto out;
372
373         /* Lock the request! */
374         nfs_lock_request_dontget(req);
375
376         spin_lock(&inode->i_lock);
377         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
378         BUG_ON(error);
379         if (!nfsi->npages) {
380                 igrab(inode);
381                 if (nfs_have_delegation(inode, FMODE_WRITE))
382                         nfsi->change_attr++;
383         }
384         SetPagePrivate(req->wb_page);
385         set_page_private(req->wb_page, (unsigned long)req);
386         nfsi->npages++;
387         kref_get(&req->wb_kref);
388         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
389                                 NFS_PAGE_TAG_LOCKED);
390         spin_unlock(&inode->i_lock);
391         radix_tree_preload_end();
392 out:
393         return error;
394 }
395
396 /*
397  * Remove a write request from an inode
398  */
399 static void nfs_inode_remove_request(struct nfs_page *req)
400 {
401         struct inode *inode = req->wb_context->path.dentry->d_inode;
402         struct nfs_inode *nfsi = NFS_I(inode);
403
404         BUG_ON (!NFS_WBACK_BUSY(req));
405
406         spin_lock(&inode->i_lock);
407         set_page_private(req->wb_page, 0);
408         ClearPagePrivate(req->wb_page);
409         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
410         nfsi->npages--;
411         if (!nfsi->npages) {
412                 spin_unlock(&inode->i_lock);
413                 iput(inode);
414         } else
415                 spin_unlock(&inode->i_lock);
416         nfs_clear_request(req);
417         nfs_release_request(req);
418 }
419
420 static void
421 nfs_mark_request_dirty(struct nfs_page *req)
422 {
423         __set_page_dirty_nobuffers(req->wb_page);
424 }
425
426 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
427 /*
428  * Add a request to the inode's commit list.
429  */
430 static void
431 nfs_mark_request_commit(struct nfs_page *req)
432 {
433         struct inode *inode = req->wb_context->path.dentry->d_inode;
434         struct nfs_inode *nfsi = NFS_I(inode);
435
436         spin_lock(&inode->i_lock);
437         set_bit(PG_CLEAN, &(req)->wb_flags);
438         radix_tree_tag_set(&nfsi->nfs_page_tree,
439                         req->wb_index,
440                         NFS_PAGE_TAG_COMMIT);
441         nfsi->ncommit++;
442         spin_unlock(&inode->i_lock);
443         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
444         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
445         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
446 }
447
448 static int
449 nfs_clear_request_commit(struct nfs_page *req)
450 {
451         struct page *page = req->wb_page;
452
453         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
454                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
455                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
456                 return 1;
457         }
458         return 0;
459 }
460
461 static inline
462 int nfs_write_need_commit(struct nfs_write_data *data)
463 {
464         return data->verf.committed != NFS_FILE_SYNC;
465 }
466
467 static inline
468 int nfs_reschedule_unstable_write(struct nfs_page *req)
469 {
470         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
471                 nfs_mark_request_commit(req);
472                 return 1;
473         }
474         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
475                 nfs_mark_request_dirty(req);
476                 return 1;
477         }
478         return 0;
479 }
480 #else
481 static inline void
482 nfs_mark_request_commit(struct nfs_page *req)
483 {
484 }
485
486 static inline int
487 nfs_clear_request_commit(struct nfs_page *req)
488 {
489         return 0;
490 }
491
492 static inline
493 int nfs_write_need_commit(struct nfs_write_data *data)
494 {
495         return 0;
496 }
497
498 static inline
499 int nfs_reschedule_unstable_write(struct nfs_page *req)
500 {
501         return 0;
502 }
503 #endif
504
505 /*
506  * Wait for a request to complete.
507  *
508  * Interruptible by fatal signals only.
509  */
510 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
511 {
512         struct nfs_inode *nfsi = NFS_I(inode);
513         struct nfs_page *req;
514         pgoff_t idx_end, next;
515         unsigned int            res = 0;
516         int                     error;
517
518         if (npages == 0)
519                 idx_end = ~0;
520         else
521                 idx_end = idx_start + npages - 1;
522
523         next = idx_start;
524         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
525                 if (req->wb_index > idx_end)
526                         break;
527
528                 next = req->wb_index + 1;
529                 BUG_ON(!NFS_WBACK_BUSY(req));
530
531                 kref_get(&req->wb_kref);
532                 spin_unlock(&inode->i_lock);
533                 error = nfs_wait_on_request(req);
534                 nfs_release_request(req);
535                 spin_lock(&inode->i_lock);
536                 if (error < 0)
537                         return error;
538                 res++;
539         }
540         return res;
541 }
542
543 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
544 static int
545 nfs_need_commit(struct nfs_inode *nfsi)
546 {
547         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
548 }
549
550 /*
551  * nfs_scan_commit - Scan an inode for commit requests
552  * @inode: NFS inode to scan
553  * @dst: destination list
554  * @idx_start: lower bound of page->index to scan.
555  * @npages: idx_start + npages sets the upper bound to scan.
556  *
557  * Moves requests from the inode's 'commit' request list.
558  * The requests are *not* checked to ensure that they form a contiguous set.
559  */
560 static int
561 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
562 {
563         struct nfs_inode *nfsi = NFS_I(inode);
564         int ret;
565
566         if (!nfs_need_commit(nfsi))
567                 return 0;
568
569         ret = nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
570         if (ret > 0)
571                 nfsi->ncommit -= ret;
572         if (nfs_need_commit(NFS_I(inode)))
573                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
574         return ret;
575 }
576 #else
577 static inline int nfs_need_commit(struct nfs_inode *nfsi)
578 {
579         return 0;
580 }
581
582 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
583 {
584         return 0;
585 }
586 #endif
587
588 /*
589  * Search for an existing write request, and attempt to update
590  * it to reflect a new dirty region on a given page.
591  *
592  * If the attempt fails, then the existing request is flushed out
593  * to disk.
594  */
595 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
596                 struct page *page,
597                 unsigned int offset,
598                 unsigned int bytes)
599 {
600         struct nfs_page *req;
601         unsigned int rqend;
602         unsigned int end;
603         int error;
604
605         if (!PagePrivate(page))
606                 return NULL;
607
608         end = offset + bytes;
609         spin_lock(&inode->i_lock);
610
611         for (;;) {
612                 req = nfs_page_find_request_locked(page);
613                 if (req == NULL)
614                         goto out_unlock;
615
616                 rqend = req->wb_offset + req->wb_bytes;
617                 /*
618                  * Tell the caller to flush out the request if
619                  * the offsets are non-contiguous.
620                  * Note: nfs_flush_incompatible() will already
621                  * have flushed out requests having wrong owners.
622                  */
623                 if (offset > rqend
624                     || end < req->wb_offset)
625                         goto out_flushme;
626
627                 if (nfs_set_page_tag_locked(req))
628                         break;
629
630                 /* The request is locked, so wait and then retry */
631                 spin_unlock(&inode->i_lock);
632                 error = nfs_wait_on_request(req);
633                 nfs_release_request(req);
634                 if (error != 0)
635                         goto out_err;
636                 spin_lock(&inode->i_lock);
637         }
638
639         if (nfs_clear_request_commit(req) &&
640                         radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
641                                 req->wb_index, NFS_PAGE_TAG_COMMIT) != NULL)
642                 NFS_I(inode)->ncommit--;
643
644         /* Okay, the request matches. Update the region */
645         if (offset < req->wb_offset) {
646                 req->wb_offset = offset;
647                 req->wb_pgbase = offset;
648         }
649         if (end > rqend)
650                 req->wb_bytes = end - req->wb_offset;
651         else
652                 req->wb_bytes = rqend - req->wb_offset;
653 out_unlock:
654         spin_unlock(&inode->i_lock);
655         return req;
656 out_flushme:
657         spin_unlock(&inode->i_lock);
658         nfs_release_request(req);
659         error = nfs_wb_page(inode, page);
660 out_err:
661         return ERR_PTR(error);
662 }
663
664 /*
665  * Try to update an existing write request, or create one if there is none.
666  *
667  * Note: Should always be called with the Page Lock held to prevent races
668  * if we have to add a new request. Also assumes that the caller has
669  * already called nfs_flush_incompatible() if necessary.
670  */
671 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
672                 struct page *page, unsigned int offset, unsigned int bytes)
673 {
674         struct inode *inode = page->mapping->host;
675         struct nfs_page *req;
676         int error;
677
678         req = nfs_try_to_update_request(inode, page, offset, bytes);
679         if (req != NULL)
680                 goto out;
681         req = nfs_create_request(ctx, inode, page, offset, bytes);
682         if (IS_ERR(req))
683                 goto out;
684         error = nfs_inode_add_request(inode, req);
685         if (error != 0) {
686                 nfs_release_request(req);
687                 req = ERR_PTR(error);
688         }
689 out:
690         return req;
691 }
692
693 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
694                 unsigned int offset, unsigned int count)
695 {
696         struct nfs_page *req;
697
698         req = nfs_setup_write_request(ctx, page, offset, count);
699         if (IS_ERR(req))
700                 return PTR_ERR(req);
701         /* Update file length */
702         nfs_grow_file(page, offset, count);
703         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
704         nfs_clear_page_tag_locked(req);
705         return 0;
706 }
707
708 int nfs_flush_incompatible(struct file *file, struct page *page)
709 {
710         struct nfs_open_context *ctx = nfs_file_open_context(file);
711         struct nfs_page *req;
712         int do_flush, status;
713         /*
714          * Look for a request corresponding to this page. If there
715          * is one, and it belongs to another file, we flush it out
716          * before we try to copy anything into the page. Do this
717          * due to the lack of an ACCESS-type call in NFSv2.
718          * Also do the same if we find a request from an existing
719          * dropped page.
720          */
721         do {
722                 req = nfs_page_find_request(page);
723                 if (req == NULL)
724                         return 0;
725                 do_flush = req->wb_page != page || req->wb_context != ctx;
726                 nfs_release_request(req);
727                 if (!do_flush)
728                         return 0;
729                 status = nfs_wb_page(page->mapping->host, page);
730         } while (status == 0);
731         return status;
732 }
733
734 /*
735  * If the page cache is marked as unsafe or invalid, then we can't rely on
736  * the PageUptodate() flag. In this case, we will need to turn off
737  * write optimisations that depend on the page contents being correct.
738  */
739 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
740 {
741         return PageUptodate(page) &&
742                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
743 }
744
745 /*
746  * Update and possibly write a cached page of an NFS file.
747  *
748  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
749  * things with a page scheduled for an RPC call (e.g. invalidate it).
750  */
751 int nfs_updatepage(struct file *file, struct page *page,
752                 unsigned int offset, unsigned int count)
753 {
754         struct nfs_open_context *ctx = nfs_file_open_context(file);
755         struct inode    *inode = page->mapping->host;
756         int             status = 0;
757
758         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
759
760         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
761                 file->f_path.dentry->d_parent->d_name.name,
762                 file->f_path.dentry->d_name.name, count,
763                 (long long)(page_offset(page) + offset));
764
765         /* If we're not using byte range locks, and we know the page
766          * is up to date, it may be more efficient to extend the write
767          * to cover the entire page in order to avoid fragmentation
768          * inefficiencies.
769          */
770         if (nfs_write_pageuptodate(page, inode) &&
771                         inode->i_flock == NULL &&
772                         !(file->f_flags & O_DSYNC)) {
773                 count = max(count + offset, nfs_page_length(page));
774                 offset = 0;
775         }
776
777         status = nfs_writepage_setup(ctx, page, offset, count);
778         if (status < 0)
779                 nfs_set_pageerror(page);
780         else
781                 __set_page_dirty_nobuffers(page);
782
783         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
784                         status, (long long)i_size_read(inode));
785         return status;
786 }
787
788 static void nfs_writepage_release(struct nfs_page *req)
789 {
790
791         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
792                 nfs_end_page_writeback(req->wb_page);
793                 nfs_inode_remove_request(req);
794         } else
795                 nfs_end_page_writeback(req->wb_page);
796         nfs_clear_page_tag_locked(req);
797 }
798
799 static int flush_task_priority(int how)
800 {
801         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
802                 case FLUSH_HIGHPRI:
803                         return RPC_PRIORITY_HIGH;
804                 case FLUSH_LOWPRI:
805                         return RPC_PRIORITY_LOW;
806         }
807         return RPC_PRIORITY_NORMAL;
808 }
809
810 /*
811  * Set up the argument/result storage required for the RPC call.
812  */
813 static int nfs_write_rpcsetup(struct nfs_page *req,
814                 struct nfs_write_data *data,
815                 const struct rpc_call_ops *call_ops,
816                 unsigned int count, unsigned int offset,
817                 int how)
818 {
819         struct inode *inode = req->wb_context->path.dentry->d_inode;
820         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
821         int priority = flush_task_priority(how);
822         struct rpc_task *task;
823         struct rpc_message msg = {
824                 .rpc_argp = &data->args,
825                 .rpc_resp = &data->res,
826                 .rpc_cred = req->wb_context->cred,
827         };
828         struct rpc_task_setup task_setup_data = {
829                 .rpc_client = NFS_CLIENT(inode),
830                 .task = &data->task,
831                 .rpc_message = &msg,
832                 .callback_ops = call_ops,
833                 .callback_data = data,
834                 .workqueue = nfsiod_workqueue,
835                 .flags = flags,
836                 .priority = priority,
837         };
838
839         /* Set up the RPC argument and reply structs
840          * NB: take care not to mess about with data->commit et al. */
841
842         data->req = req;
843         data->inode = inode = req->wb_context->path.dentry->d_inode;
844         data->cred = msg.rpc_cred;
845
846         data->args.fh     = NFS_FH(inode);
847         data->args.offset = req_offset(req) + offset;
848         data->args.pgbase = req->wb_pgbase + offset;
849         data->args.pages  = data->pagevec;
850         data->args.count  = count;
851         data->args.context = get_nfs_open_context(req->wb_context);
852         data->args.stable  = NFS_UNSTABLE;
853         if (how & FLUSH_STABLE) {
854                 data->args.stable = NFS_DATA_SYNC;
855                 if (!nfs_need_commit(NFS_I(inode)))
856                         data->args.stable = NFS_FILE_SYNC;
857         }
858
859         data->res.fattr   = &data->fattr;
860         data->res.count   = count;
861         data->res.verf    = &data->verf;
862         nfs_fattr_init(&data->fattr);
863
864         /* Set up the initial task struct.  */
865         NFS_PROTO(inode)->write_setup(data, &msg);
866
867         dprintk("NFS: %5u initiated write call "
868                 "(req %s/%lld, %u bytes @ offset %llu)\n",
869                 data->task.tk_pid,
870                 inode->i_sb->s_id,
871                 (long long)NFS_FILEID(inode),
872                 count,
873                 (unsigned long long)data->args.offset);
874
875         task = rpc_run_task(&task_setup_data);
876         if (IS_ERR(task))
877                 return PTR_ERR(task);
878         rpc_put_task(task);
879         return 0;
880 }
881
882 /* If a nfs_flush_* function fails, it should remove reqs from @head and
883  * call this on each, which will prepare them to be retried on next
884  * writeback using standard nfs.
885  */
886 static void nfs_redirty_request(struct nfs_page *req)
887 {
888         nfs_mark_request_dirty(req);
889         nfs_end_page_writeback(req->wb_page);
890         nfs_clear_page_tag_locked(req);
891 }
892
893 /*
894  * Generate multiple small requests to write out a single
895  * contiguous dirty area on one page.
896  */
897 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
898 {
899         struct nfs_page *req = nfs_list_entry(head->next);
900         struct page *page = req->wb_page;
901         struct nfs_write_data *data;
902         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
903         unsigned int offset;
904         int requests = 0;
905         int ret = 0;
906         LIST_HEAD(list);
907
908         nfs_list_remove_request(req);
909
910         nbytes = count;
911         do {
912                 size_t len = min(nbytes, wsize);
913
914                 data = nfs_writedata_alloc(1);
915                 if (!data)
916                         goto out_bad;
917                 list_add(&data->pages, &list);
918                 requests++;
919                 nbytes -= len;
920         } while (nbytes != 0);
921         atomic_set(&req->wb_complete, requests);
922
923         ClearPageError(page);
924         offset = 0;
925         nbytes = count;
926         do {
927                 int ret2;
928
929                 data = list_entry(list.next, struct nfs_write_data, pages);
930                 list_del_init(&data->pages);
931
932                 data->pagevec[0] = page;
933
934                 if (nbytes < wsize)
935                         wsize = nbytes;
936                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
937                                    wsize, offset, how);
938                 if (ret == 0)
939                         ret = ret2;
940                 offset += wsize;
941                 nbytes -= wsize;
942         } while (nbytes != 0);
943
944         return ret;
945
946 out_bad:
947         while (!list_empty(&list)) {
948                 data = list_entry(list.next, struct nfs_write_data, pages);
949                 list_del(&data->pages);
950                 nfs_writedata_release(data);
951         }
952         nfs_redirty_request(req);
953         return -ENOMEM;
954 }
955
956 /*
957  * Create an RPC task for the given write request and kick it.
958  * The page must have been locked by the caller.
959  *
960  * It may happen that the page we're passed is not marked dirty.
961  * This is the case if nfs_updatepage detects a conflicting request
962  * that has been written but not committed.
963  */
964 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
965 {
966         struct nfs_page         *req;
967         struct page             **pages;
968         struct nfs_write_data   *data;
969
970         data = nfs_writedata_alloc(npages);
971         if (!data)
972                 goto out_bad;
973
974         pages = data->pagevec;
975         while (!list_empty(head)) {
976                 req = nfs_list_entry(head->next);
977                 nfs_list_remove_request(req);
978                 nfs_list_add_request(req, &data->pages);
979                 ClearPageError(req->wb_page);
980                 *pages++ = req->wb_page;
981         }
982         req = nfs_list_entry(data->pages.next);
983
984         /* Set up the argument struct */
985         return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
986  out_bad:
987         while (!list_empty(head)) {
988                 req = nfs_list_entry(head->next);
989                 nfs_list_remove_request(req);
990                 nfs_redirty_request(req);
991         }
992         return -ENOMEM;
993 }
994
995 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
996                                   struct inode *inode, int ioflags)
997 {
998         size_t wsize = NFS_SERVER(inode)->wsize;
999
1000         if (wsize < PAGE_CACHE_SIZE)
1001                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
1002         else
1003                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
1004 }
1005
1006 /*
1007  * Handle a write reply that flushed part of a page.
1008  */
1009 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1010 {
1011         struct nfs_write_data   *data = calldata;
1012
1013         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1014                 task->tk_pid,
1015                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1016                 (long long)
1017                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1018                 data->req->wb_bytes, (long long)req_offset(data->req));
1019
1020         nfs_writeback_done(task, data);
1021 }
1022
1023 static void nfs_writeback_release_partial(void *calldata)
1024 {
1025         struct nfs_write_data   *data = calldata;
1026         struct nfs_page         *req = data->req;
1027         struct page             *page = req->wb_page;
1028         int status = data->task.tk_status;
1029
1030         if (status < 0) {
1031                 nfs_set_pageerror(page);
1032                 nfs_context_set_write_error(req->wb_context, status);
1033                 dprintk(", error = %d\n", status);
1034                 goto out;
1035         }
1036
1037         if (nfs_write_need_commit(data)) {
1038                 struct inode *inode = page->mapping->host;
1039
1040                 spin_lock(&inode->i_lock);
1041                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1042                         /* Do nothing we need to resend the writes */
1043                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1044                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1045                         dprintk(" defer commit\n");
1046                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1047                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1048                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1049                         dprintk(" server reboot detected\n");
1050                 }
1051                 spin_unlock(&inode->i_lock);
1052         } else
1053                 dprintk(" OK\n");
1054
1055 out:
1056         if (atomic_dec_and_test(&req->wb_complete))
1057                 nfs_writepage_release(req);
1058         nfs_writedata_release(calldata);
1059 }
1060
1061 #if defined(CONFIG_NFS_V4_1)
1062 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1063 {
1064         struct nfs_write_data *data = calldata;
1065         struct nfs_client *clp = (NFS_SERVER(data->inode))->nfs_client;
1066
1067         if (nfs4_setup_sequence(clp, &data->args.seq_args,
1068                                 &data->res.seq_res, 1, task))
1069                 return;
1070         rpc_call_start(task);
1071 }
1072 #endif /* CONFIG_NFS_V4_1 */
1073
1074 static const struct rpc_call_ops nfs_write_partial_ops = {
1075 #if defined(CONFIG_NFS_V4_1)
1076         .rpc_call_prepare = nfs_write_prepare,
1077 #endif /* CONFIG_NFS_V4_1 */
1078         .rpc_call_done = nfs_writeback_done_partial,
1079         .rpc_release = nfs_writeback_release_partial,
1080 };
1081
1082 /*
1083  * Handle a write reply that flushes a whole page.
1084  *
1085  * FIXME: There is an inherent race with invalidate_inode_pages and
1086  *        writebacks since the page->count is kept > 1 for as long
1087  *        as the page has a write request pending.
1088  */
1089 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1090 {
1091         struct nfs_write_data   *data = calldata;
1092
1093         nfs_writeback_done(task, data);
1094 }
1095
1096 static void nfs_writeback_release_full(void *calldata)
1097 {
1098         struct nfs_write_data   *data = calldata;
1099         int status = data->task.tk_status;
1100
1101         /* Update attributes as result of writeback. */
1102         while (!list_empty(&data->pages)) {
1103                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1104                 struct page *page = req->wb_page;
1105
1106                 nfs_list_remove_request(req);
1107
1108                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1109                         data->task.tk_pid,
1110                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1111                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1112                         req->wb_bytes,
1113                         (long long)req_offset(req));
1114
1115                 if (status < 0) {
1116                         nfs_set_pageerror(page);
1117                         nfs_context_set_write_error(req->wb_context, status);
1118                         dprintk(", error = %d\n", status);
1119                         goto remove_request;
1120                 }
1121
1122                 if (nfs_write_need_commit(data)) {
1123                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1124                         nfs_mark_request_commit(req);
1125                         nfs_end_page_writeback(page);
1126                         dprintk(" marked for commit\n");
1127                         goto next;
1128                 }
1129                 dprintk(" OK\n");
1130 remove_request:
1131                 nfs_end_page_writeback(page);
1132                 nfs_inode_remove_request(req);
1133         next:
1134                 nfs_clear_page_tag_locked(req);
1135         }
1136         nfs_writedata_release(calldata);
1137 }
1138
1139 static const struct rpc_call_ops nfs_write_full_ops = {
1140 #if defined(CONFIG_NFS_V4_1)
1141         .rpc_call_prepare = nfs_write_prepare,
1142 #endif /* CONFIG_NFS_V4_1 */
1143         .rpc_call_done = nfs_writeback_done_full,
1144         .rpc_release = nfs_writeback_release_full,
1145 };
1146
1147
1148 /*
1149  * This function is called when the WRITE call is complete.
1150  */
1151 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1152 {
1153         struct nfs_writeargs    *argp = &data->args;
1154         struct nfs_writeres     *resp = &data->res;
1155         struct nfs_server       *server = NFS_SERVER(data->inode);
1156         int status;
1157
1158         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1159                 task->tk_pid, task->tk_status);
1160
1161         /*
1162          * ->write_done will attempt to use post-op attributes to detect
1163          * conflicting writes by other clients.  A strict interpretation
1164          * of close-to-open would allow us to continue caching even if
1165          * another writer had changed the file, but some applications
1166          * depend on tighter cache coherency when writing.
1167          */
1168         status = NFS_PROTO(data->inode)->write_done(task, data);
1169         if (status != 0)
1170                 return status;
1171         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1172
1173 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1174         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1175                 /* We tried a write call, but the server did not
1176                  * commit data to stable storage even though we
1177                  * requested it.
1178                  * Note: There is a known bug in Tru64 < 5.0 in which
1179                  *       the server reports NFS_DATA_SYNC, but performs
1180                  *       NFS_FILE_SYNC. We therefore implement this checking
1181                  *       as a dprintk() in order to avoid filling syslog.
1182                  */
1183                 static unsigned long    complain;
1184
1185                 if (time_before(complain, jiffies)) {
1186                         dprintk("NFS:       faulty NFS server %s:"
1187                                 " (committed = %d) != (stable = %d)\n",
1188                                 server->nfs_client->cl_hostname,
1189                                 resp->verf->committed, argp->stable);
1190                         complain = jiffies + 300 * HZ;
1191                 }
1192         }
1193 #endif
1194         /* Is this a short write? */
1195         if (task->tk_status >= 0 && resp->count < argp->count) {
1196                 static unsigned long    complain;
1197
1198                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1199
1200                 /* Has the server at least made some progress? */
1201                 if (resp->count != 0) {
1202                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1203                         if (resp->verf->committed != NFS_UNSTABLE) {
1204                                 /* Resend from where the server left off */
1205                                 argp->offset += resp->count;
1206                                 argp->pgbase += resp->count;
1207                                 argp->count -= resp->count;
1208                         } else {
1209                                 /* Resend as a stable write in order to avoid
1210                                  * headaches in the case of a server crash.
1211                                  */
1212                                 argp->stable = NFS_FILE_SYNC;
1213                         }
1214                         nfs_restart_rpc(task, server->nfs_client);
1215                         return -EAGAIN;
1216                 }
1217                 if (time_before(complain, jiffies)) {
1218                         printk(KERN_WARNING
1219                                "NFS: Server wrote zero bytes, expected %u.\n",
1220                                         argp->count);
1221                         complain = jiffies + 300 * HZ;
1222                 }
1223                 /* Can't do anything about it except throw an error. */
1224                 task->tk_status = -EIO;
1225         }
1226         return 0;
1227 }
1228
1229
1230 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1231 static void nfs_commitdata_release(void *data)
1232 {
1233         struct nfs_write_data *wdata = data;
1234
1235         put_nfs_open_context(wdata->args.context);
1236         nfs_commit_free(wdata);
1237 }
1238
1239 /*
1240  * Set up the argument/result storage required for the RPC call.
1241  */
1242 static int nfs_commit_rpcsetup(struct list_head *head,
1243                 struct nfs_write_data *data,
1244                 int how)
1245 {
1246         struct nfs_page *first = nfs_list_entry(head->next);
1247         struct inode *inode = first->wb_context->path.dentry->d_inode;
1248         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1249         int priority = flush_task_priority(how);
1250         struct rpc_task *task;
1251         struct rpc_message msg = {
1252                 .rpc_argp = &data->args,
1253                 .rpc_resp = &data->res,
1254                 .rpc_cred = first->wb_context->cred,
1255         };
1256         struct rpc_task_setup task_setup_data = {
1257                 .task = &data->task,
1258                 .rpc_client = NFS_CLIENT(inode),
1259                 .rpc_message = &msg,
1260                 .callback_ops = &nfs_commit_ops,
1261                 .callback_data = data,
1262                 .workqueue = nfsiod_workqueue,
1263                 .flags = flags,
1264                 .priority = priority,
1265         };
1266
1267         /* Set up the RPC argument and reply structs
1268          * NB: take care not to mess about with data->commit et al. */
1269
1270         list_splice_init(head, &data->pages);
1271
1272         data->inode       = inode;
1273         data->cred        = msg.rpc_cred;
1274
1275         data->args.fh     = NFS_FH(data->inode);
1276         /* Note: we always request a commit of the entire inode */
1277         data->args.offset = 0;
1278         data->args.count  = 0;
1279         data->args.context = get_nfs_open_context(first->wb_context);
1280         data->res.count   = 0;
1281         data->res.fattr   = &data->fattr;
1282         data->res.verf    = &data->verf;
1283         nfs_fattr_init(&data->fattr);
1284
1285         /* Set up the initial task struct.  */
1286         NFS_PROTO(inode)->commit_setup(data, &msg);
1287
1288         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1289
1290         task = rpc_run_task(&task_setup_data);
1291         if (IS_ERR(task))
1292                 return PTR_ERR(task);
1293         rpc_put_task(task);
1294         return 0;
1295 }
1296
1297 /*
1298  * Commit dirty pages
1299  */
1300 static int
1301 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1302 {
1303         struct nfs_write_data   *data;
1304         struct nfs_page         *req;
1305
1306         data = nfs_commitdata_alloc();
1307
1308         if (!data)
1309                 goto out_bad;
1310
1311         /* Set up the argument struct */
1312         return nfs_commit_rpcsetup(head, data, how);
1313  out_bad:
1314         while (!list_empty(head)) {
1315                 req = nfs_list_entry(head->next);
1316                 nfs_list_remove_request(req);
1317                 nfs_mark_request_commit(req);
1318                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1319                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1320                                 BDI_RECLAIMABLE);
1321                 nfs_clear_page_tag_locked(req);
1322         }
1323         return -ENOMEM;
1324 }
1325
1326 /*
1327  * COMMIT call returned
1328  */
1329 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1330 {
1331         struct nfs_write_data   *data = calldata;
1332
1333         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1334                                 task->tk_pid, task->tk_status);
1335
1336         /* Call the NFS version-specific code */
1337         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1338                 return;
1339 }
1340
1341 static void nfs_commit_release(void *calldata)
1342 {
1343         struct nfs_write_data   *data = calldata;
1344         struct nfs_page         *req;
1345         int status = data->task.tk_status;
1346
1347         while (!list_empty(&data->pages)) {
1348                 req = nfs_list_entry(data->pages.next);
1349                 nfs_list_remove_request(req);
1350                 nfs_clear_request_commit(req);
1351
1352                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1353                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1354                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1355                         req->wb_bytes,
1356                         (long long)req_offset(req));
1357                 if (status < 0) {
1358                         nfs_context_set_write_error(req->wb_context, status);
1359                         nfs_inode_remove_request(req);
1360                         dprintk(", error = %d\n", status);
1361                         goto next;
1362                 }
1363
1364                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1365                  * returned by the server against all stored verfs. */
1366                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1367                         /* We have a match */
1368                         nfs_inode_remove_request(req);
1369                         dprintk(" OK\n");
1370                         goto next;
1371                 }
1372                 /* We have a mismatch. Write the page again */
1373                 dprintk(" mismatch\n");
1374                 nfs_mark_request_dirty(req);
1375         next:
1376                 nfs_clear_page_tag_locked(req);
1377         }
1378         nfs_commitdata_release(calldata);
1379 }
1380
1381 static const struct rpc_call_ops nfs_commit_ops = {
1382 #if defined(CONFIG_NFS_V4_1)
1383         .rpc_call_prepare = nfs_write_prepare,
1384 #endif /* CONFIG_NFS_V4_1 */
1385         .rpc_call_done = nfs_commit_done,
1386         .rpc_release = nfs_commit_release,
1387 };
1388
1389 static int nfs_commit_inode(struct inode *inode, int how)
1390 {
1391         LIST_HEAD(head);
1392         int res;
1393
1394         spin_lock(&inode->i_lock);
1395         res = nfs_scan_commit(inode, &head, 0, 0);
1396         spin_unlock(&inode->i_lock);
1397         if (res) {
1398                 int error = nfs_commit_list(inode, &head, how);
1399                 if (error < 0)
1400                         return error;
1401         }
1402         return res;
1403 }
1404
1405 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1406 {
1407         struct nfs_inode *nfsi = NFS_I(inode);
1408         int flags = FLUSH_SYNC;
1409         int ret = 0;
1410
1411         /* Don't commit yet if this is a non-blocking flush and there are
1412          * lots of outstanding writes for this mapping.
1413          */
1414         if (wbc->sync_mode == WB_SYNC_NONE &&
1415             nfsi->ncommit <= (nfsi->npages >> 1))
1416                 goto out_mark_dirty;
1417
1418         if (wbc->nonblocking || wbc->for_background)
1419                 flags = 0;
1420         ret = nfs_commit_inode(inode, flags);
1421         if (ret >= 0) {
1422                 if (wbc->sync_mode == WB_SYNC_NONE) {
1423                         if (ret < wbc->nr_to_write)
1424                                 wbc->nr_to_write -= ret;
1425                         else
1426                                 wbc->nr_to_write = 0;
1427                 }
1428                 return 0;
1429         }
1430 out_mark_dirty:
1431         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1432         return ret;
1433 }
1434 #else
1435 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1436 {
1437         return 0;
1438 }
1439
1440 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1441 {
1442         return 0;
1443 }
1444 #endif
1445
1446 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1447 {
1448         return nfs_commit_unstable_pages(inode, wbc);
1449 }
1450
1451 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1452 {
1453         struct inode *inode = mapping->host;
1454         pgoff_t idx_start, idx_end;
1455         unsigned int npages = 0;
1456         LIST_HEAD(head);
1457         int nocommit = how & FLUSH_NOCOMMIT;
1458         long pages, ret;
1459
1460         /* FIXME */
1461         if (wbc->range_cyclic)
1462                 idx_start = 0;
1463         else {
1464                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1465                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1466                 if (idx_end > idx_start) {
1467                         pgoff_t l_npages = 1 + idx_end - idx_start;
1468                         npages = l_npages;
1469                         if (sizeof(npages) != sizeof(l_npages) &&
1470                                         (pgoff_t)npages != l_npages)
1471                                 npages = 0;
1472                 }
1473         }
1474         how &= ~FLUSH_NOCOMMIT;
1475         spin_lock(&inode->i_lock);
1476         do {
1477                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1478                 if (ret != 0)
1479                         continue;
1480                 if (nocommit)
1481                         break;
1482                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1483                 if (pages == 0)
1484                         break;
1485                 pages += nfs_scan_commit(inode, &head, 0, 0);
1486                 spin_unlock(&inode->i_lock);
1487                 ret = nfs_commit_list(inode, &head, how);
1488                 spin_lock(&inode->i_lock);
1489
1490         } while (ret >= 0);
1491         spin_unlock(&inode->i_lock);
1492         return ret;
1493 }
1494
1495 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1496 {
1497         int ret;
1498
1499         ret = nfs_writepages(mapping, wbc);
1500         if (ret < 0)
1501                 goto out;
1502         ret = nfs_sync_mapping_wait(mapping, wbc, how);
1503         if (ret < 0)
1504                 goto out;
1505         return 0;
1506 out:
1507         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1508         return ret;
1509 }
1510
1511 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1512 static int nfs_write_mapping(struct address_space *mapping, int how)
1513 {
1514         struct writeback_control wbc = {
1515                 .bdi = mapping->backing_dev_info,
1516                 .sync_mode = WB_SYNC_ALL,
1517                 .nr_to_write = LONG_MAX,
1518                 .range_start = 0,
1519                 .range_end = LLONG_MAX,
1520         };
1521
1522         return __nfs_write_mapping(mapping, &wbc, how);
1523 }
1524
1525 /*
1526  * flush the inode to disk.
1527  */
1528 int nfs_wb_all(struct inode *inode)
1529 {
1530         return nfs_write_mapping(inode->i_mapping, 0);
1531 }
1532
1533 int nfs_wb_nocommit(struct inode *inode)
1534 {
1535         return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1536 }
1537
1538 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1539 {
1540         struct nfs_page *req;
1541         int ret = 0;
1542
1543         BUG_ON(!PageLocked(page));
1544         for (;;) {
1545                 req = nfs_page_find_request(page);
1546                 if (req == NULL)
1547                         break;
1548                 if (nfs_lock_request_dontget(req)) {
1549                         nfs_inode_remove_request(req);
1550                         /*
1551                          * In case nfs_inode_remove_request has marked the
1552                          * page as being dirty
1553                          */
1554                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1555                         nfs_unlock_request(req);
1556                         break;
1557                 }
1558                 ret = nfs_wait_on_request(req);
1559                 nfs_release_request(req);
1560                 if (ret < 0)
1561                         break;
1562         }
1563         return ret;
1564 }
1565
1566 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1567                                 int how)
1568 {
1569         loff_t range_start = page_offset(page);
1570         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1571         struct writeback_control wbc = {
1572                 .bdi = page->mapping->backing_dev_info,
1573                 .sync_mode = WB_SYNC_ALL,
1574                 .nr_to_write = LONG_MAX,
1575                 .range_start = range_start,
1576                 .range_end = range_end,
1577         };
1578         int ret;
1579
1580         do {
1581                 if (clear_page_dirty_for_io(page)) {
1582                         ret = nfs_writepage_locked(page, &wbc);
1583                         if (ret < 0)
1584                                 goto out_error;
1585                 } else if (!PagePrivate(page))
1586                         break;
1587                 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1588                 if (ret < 0)
1589                         goto out_error;
1590         } while (PagePrivate(page));
1591         return 0;
1592 out_error:
1593         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1594         return ret;
1595 }
1596
1597 /*
1598  * Write back all requests on one page - we do this before reading it.
1599  */
1600 int nfs_wb_page(struct inode *inode, struct page* page)
1601 {
1602         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1603 }
1604
1605 #ifdef CONFIG_MIGRATION
1606 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1607                 struct page *page)
1608 {
1609         struct nfs_page *req;
1610         int ret;
1611
1612         nfs_fscache_release_page(page, GFP_KERNEL);
1613
1614         req = nfs_find_and_lock_request(page);
1615         ret = PTR_ERR(req);
1616         if (IS_ERR(req))
1617                 goto out;
1618
1619         ret = migrate_page(mapping, newpage, page);
1620         if (!req)
1621                 goto out;
1622         if (ret)
1623                 goto out_unlock;
1624         page_cache_get(newpage);
1625         spin_lock(&mapping->host->i_lock);
1626         req->wb_page = newpage;
1627         SetPagePrivate(newpage);
1628         set_page_private(newpage, (unsigned long)req);
1629         ClearPagePrivate(page);
1630         set_page_private(page, 0);
1631         spin_unlock(&mapping->host->i_lock);
1632         page_cache_release(page);
1633 out_unlock:
1634         nfs_clear_page_tag_locked(req);
1635 out:
1636         return ret;
1637 }
1638 #endif
1639
1640 int __init nfs_init_writepagecache(void)
1641 {
1642         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1643                                              sizeof(struct nfs_write_data),
1644                                              0, SLAB_HWCACHE_ALIGN,
1645                                              NULL);
1646         if (nfs_wdata_cachep == NULL)
1647                 return -ENOMEM;
1648
1649         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1650                                                      nfs_wdata_cachep);
1651         if (nfs_wdata_mempool == NULL)
1652                 return -ENOMEM;
1653
1654         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1655                                                       nfs_wdata_cachep);
1656         if (nfs_commit_mempool == NULL)
1657                 return -ENOMEM;
1658
1659         /*
1660          * NFS congestion size, scale with available memory.
1661          *
1662          *  64MB:    8192k
1663          * 128MB:   11585k
1664          * 256MB:   16384k
1665          * 512MB:   23170k
1666          *   1GB:   32768k
1667          *   2GB:   46340k
1668          *   4GB:   65536k
1669          *   8GB:   92681k
1670          *  16GB:  131072k
1671          *
1672          * This allows larger machines to have larger/more transfers.
1673          * Limit the default to 256M
1674          */
1675         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1676         if (nfs_congestion_kb > 256*1024)
1677                 nfs_congestion_kb = 256*1024;
1678
1679         return 0;
1680 }
1681
1682 void nfs_destroy_writepagecache(void)
1683 {
1684         mempool_destroy(nfs_commit_mempool);
1685         mempool_destroy(nfs_wdata_mempool);
1686         kmem_cache_destroy(nfs_wdata_cachep);
1687 }
1688