704e67d392e5a8bd01c13e0c8b8010dea10691f5
[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 static void nfs_cancel_commit_list(struct list_head *head)
544 {
545         struct nfs_page *req;
546
547         while(!list_empty(head)) {
548                 req = nfs_list_entry(head->next);
549                 nfs_list_remove_request(req);
550                 nfs_clear_request_commit(req);
551                 nfs_inode_remove_request(req);
552                 nfs_unlock_request(req);
553         }
554 }
555
556 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
557 static int
558 nfs_need_commit(struct nfs_inode *nfsi)
559 {
560         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
561 }
562
563 /*
564  * nfs_scan_commit - Scan an inode for commit requests
565  * @inode: NFS inode to scan
566  * @dst: destination list
567  * @idx_start: lower bound of page->index to scan.
568  * @npages: idx_start + npages sets the upper bound to scan.
569  *
570  * Moves requests from the inode's 'commit' request list.
571  * The requests are *not* checked to ensure that they form a contiguous set.
572  */
573 static int
574 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
575 {
576         struct nfs_inode *nfsi = NFS_I(inode);
577         int ret;
578
579         if (!nfs_need_commit(nfsi))
580                 return 0;
581
582         ret = nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
583         if (ret > 0)
584                 nfsi->ncommit -= ret;
585         return ret;
586 }
587 #else
588 static inline int nfs_need_commit(struct nfs_inode *nfsi)
589 {
590         return 0;
591 }
592
593 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
594 {
595         return 0;
596 }
597 #endif
598
599 /*
600  * Search for an existing write request, and attempt to update
601  * it to reflect a new dirty region on a given page.
602  *
603  * If the attempt fails, then the existing request is flushed out
604  * to disk.
605  */
606 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
607                 struct page *page,
608                 unsigned int offset,
609                 unsigned int bytes)
610 {
611         struct nfs_page *req;
612         unsigned int rqend;
613         unsigned int end;
614         int error;
615
616         if (!PagePrivate(page))
617                 return NULL;
618
619         end = offset + bytes;
620         spin_lock(&inode->i_lock);
621
622         for (;;) {
623                 req = nfs_page_find_request_locked(page);
624                 if (req == NULL)
625                         goto out_unlock;
626
627                 rqend = req->wb_offset + req->wb_bytes;
628                 /*
629                  * Tell the caller to flush out the request if
630                  * the offsets are non-contiguous.
631                  * Note: nfs_flush_incompatible() will already
632                  * have flushed out requests having wrong owners.
633                  */
634                 if (offset > rqend
635                     || end < req->wb_offset)
636                         goto out_flushme;
637
638                 if (nfs_set_page_tag_locked(req))
639                         break;
640
641                 /* The request is locked, so wait and then retry */
642                 spin_unlock(&inode->i_lock);
643                 error = nfs_wait_on_request(req);
644                 nfs_release_request(req);
645                 if (error != 0)
646                         goto out_err;
647                 spin_lock(&inode->i_lock);
648         }
649
650         if (nfs_clear_request_commit(req) &&
651                         radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
652                                 req->wb_index, NFS_PAGE_TAG_COMMIT) != NULL)
653                 NFS_I(inode)->ncommit--;
654
655         /* Okay, the request matches. Update the region */
656         if (offset < req->wb_offset) {
657                 req->wb_offset = offset;
658                 req->wb_pgbase = offset;
659         }
660         if (end > rqend)
661                 req->wb_bytes = end - req->wb_offset;
662         else
663                 req->wb_bytes = rqend - req->wb_offset;
664 out_unlock:
665         spin_unlock(&inode->i_lock);
666         return req;
667 out_flushme:
668         spin_unlock(&inode->i_lock);
669         nfs_release_request(req);
670         error = nfs_wb_page(inode, page);
671 out_err:
672         return ERR_PTR(error);
673 }
674
675 /*
676  * Try to update an existing write request, or create one if there is none.
677  *
678  * Note: Should always be called with the Page Lock held to prevent races
679  * if we have to add a new request. Also assumes that the caller has
680  * already called nfs_flush_incompatible() if necessary.
681  */
682 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
683                 struct page *page, unsigned int offset, unsigned int bytes)
684 {
685         struct inode *inode = page->mapping->host;
686         struct nfs_page *req;
687         int error;
688
689         req = nfs_try_to_update_request(inode, page, offset, bytes);
690         if (req != NULL)
691                 goto out;
692         req = nfs_create_request(ctx, inode, page, offset, bytes);
693         if (IS_ERR(req))
694                 goto out;
695         error = nfs_inode_add_request(inode, req);
696         if (error != 0) {
697                 nfs_release_request(req);
698                 req = ERR_PTR(error);
699         }
700 out:
701         return req;
702 }
703
704 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
705                 unsigned int offset, unsigned int count)
706 {
707         struct nfs_page *req;
708
709         req = nfs_setup_write_request(ctx, page, offset, count);
710         if (IS_ERR(req))
711                 return PTR_ERR(req);
712         /* Update file length */
713         nfs_grow_file(page, offset, count);
714         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
715         nfs_clear_page_tag_locked(req);
716         return 0;
717 }
718
719 int nfs_flush_incompatible(struct file *file, struct page *page)
720 {
721         struct nfs_open_context *ctx = nfs_file_open_context(file);
722         struct nfs_page *req;
723         int do_flush, status;
724         /*
725          * Look for a request corresponding to this page. If there
726          * is one, and it belongs to another file, we flush it out
727          * before we try to copy anything into the page. Do this
728          * due to the lack of an ACCESS-type call in NFSv2.
729          * Also do the same if we find a request from an existing
730          * dropped page.
731          */
732         do {
733                 req = nfs_page_find_request(page);
734                 if (req == NULL)
735                         return 0;
736                 do_flush = req->wb_page != page || req->wb_context != ctx;
737                 nfs_release_request(req);
738                 if (!do_flush)
739                         return 0;
740                 status = nfs_wb_page(page->mapping->host, page);
741         } while (status == 0);
742         return status;
743 }
744
745 /*
746  * If the page cache is marked as unsafe or invalid, then we can't rely on
747  * the PageUptodate() flag. In this case, we will need to turn off
748  * write optimisations that depend on the page contents being correct.
749  */
750 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
751 {
752         return PageUptodate(page) &&
753                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
754 }
755
756 /*
757  * Update and possibly write a cached page of an NFS file.
758  *
759  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
760  * things with a page scheduled for an RPC call (e.g. invalidate it).
761  */
762 int nfs_updatepage(struct file *file, struct page *page,
763                 unsigned int offset, unsigned int count)
764 {
765         struct nfs_open_context *ctx = nfs_file_open_context(file);
766         struct inode    *inode = page->mapping->host;
767         int             status = 0;
768
769         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
770
771         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
772                 file->f_path.dentry->d_parent->d_name.name,
773                 file->f_path.dentry->d_name.name, count,
774                 (long long)(page_offset(page) + offset));
775
776         /* If we're not using byte range locks, and we know the page
777          * is up to date, it may be more efficient to extend the write
778          * to cover the entire page in order to avoid fragmentation
779          * inefficiencies.
780          */
781         if (nfs_write_pageuptodate(page, inode) &&
782                         inode->i_flock == NULL &&
783                         !(file->f_flags & O_DSYNC)) {
784                 count = max(count + offset, nfs_page_length(page));
785                 offset = 0;
786         }
787
788         status = nfs_writepage_setup(ctx, page, offset, count);
789         if (status < 0)
790                 nfs_set_pageerror(page);
791         else
792                 __set_page_dirty_nobuffers(page);
793
794         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
795                         status, (long long)i_size_read(inode));
796         return status;
797 }
798
799 static void nfs_writepage_release(struct nfs_page *req)
800 {
801
802         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
803                 nfs_end_page_writeback(req->wb_page);
804                 nfs_inode_remove_request(req);
805         } else
806                 nfs_end_page_writeback(req->wb_page);
807         nfs_clear_page_tag_locked(req);
808 }
809
810 static int flush_task_priority(int how)
811 {
812         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
813                 case FLUSH_HIGHPRI:
814                         return RPC_PRIORITY_HIGH;
815                 case FLUSH_LOWPRI:
816                         return RPC_PRIORITY_LOW;
817         }
818         return RPC_PRIORITY_NORMAL;
819 }
820
821 /*
822  * Set up the argument/result storage required for the RPC call.
823  */
824 static int nfs_write_rpcsetup(struct nfs_page *req,
825                 struct nfs_write_data *data,
826                 const struct rpc_call_ops *call_ops,
827                 unsigned int count, unsigned int offset,
828                 int how)
829 {
830         struct inode *inode = req->wb_context->path.dentry->d_inode;
831         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
832         int priority = flush_task_priority(how);
833         struct rpc_task *task;
834         struct rpc_message msg = {
835                 .rpc_argp = &data->args,
836                 .rpc_resp = &data->res,
837                 .rpc_cred = req->wb_context->cred,
838         };
839         struct rpc_task_setup task_setup_data = {
840                 .rpc_client = NFS_CLIENT(inode),
841                 .task = &data->task,
842                 .rpc_message = &msg,
843                 .callback_ops = call_ops,
844                 .callback_data = data,
845                 .workqueue = nfsiod_workqueue,
846                 .flags = flags,
847                 .priority = priority,
848         };
849
850         /* Set up the RPC argument and reply structs
851          * NB: take care not to mess about with data->commit et al. */
852
853         data->req = req;
854         data->inode = inode = req->wb_context->path.dentry->d_inode;
855         data->cred = msg.rpc_cred;
856
857         data->args.fh     = NFS_FH(inode);
858         data->args.offset = req_offset(req) + offset;
859         data->args.pgbase = req->wb_pgbase + offset;
860         data->args.pages  = data->pagevec;
861         data->args.count  = count;
862         data->args.context = get_nfs_open_context(req->wb_context);
863         data->args.stable  = NFS_UNSTABLE;
864         if (how & FLUSH_STABLE) {
865                 data->args.stable = NFS_DATA_SYNC;
866                 if (!nfs_need_commit(NFS_I(inode)))
867                         data->args.stable = NFS_FILE_SYNC;
868         }
869
870         data->res.fattr   = &data->fattr;
871         data->res.count   = count;
872         data->res.verf    = &data->verf;
873         nfs_fattr_init(&data->fattr);
874
875         /* Set up the initial task struct.  */
876         NFS_PROTO(inode)->write_setup(data, &msg);
877
878         dprintk("NFS: %5u initiated write call "
879                 "(req %s/%lld, %u bytes @ offset %llu)\n",
880                 data->task.tk_pid,
881                 inode->i_sb->s_id,
882                 (long long)NFS_FILEID(inode),
883                 count,
884                 (unsigned long long)data->args.offset);
885
886         task = rpc_run_task(&task_setup_data);
887         if (IS_ERR(task))
888                 return PTR_ERR(task);
889         rpc_put_task(task);
890         return 0;
891 }
892
893 /* If a nfs_flush_* function fails, it should remove reqs from @head and
894  * call this on each, which will prepare them to be retried on next
895  * writeback using standard nfs.
896  */
897 static void nfs_redirty_request(struct nfs_page *req)
898 {
899         nfs_mark_request_dirty(req);
900         nfs_end_page_writeback(req->wb_page);
901         nfs_clear_page_tag_locked(req);
902 }
903
904 /*
905  * Generate multiple small requests to write out a single
906  * contiguous dirty area on one page.
907  */
908 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
909 {
910         struct nfs_page *req = nfs_list_entry(head->next);
911         struct page *page = req->wb_page;
912         struct nfs_write_data *data;
913         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
914         unsigned int offset;
915         int requests = 0;
916         int ret = 0;
917         LIST_HEAD(list);
918
919         nfs_list_remove_request(req);
920
921         nbytes = count;
922         do {
923                 size_t len = min(nbytes, wsize);
924
925                 data = nfs_writedata_alloc(1);
926                 if (!data)
927                         goto out_bad;
928                 list_add(&data->pages, &list);
929                 requests++;
930                 nbytes -= len;
931         } while (nbytes != 0);
932         atomic_set(&req->wb_complete, requests);
933
934         ClearPageError(page);
935         offset = 0;
936         nbytes = count;
937         do {
938                 int ret2;
939
940                 data = list_entry(list.next, struct nfs_write_data, pages);
941                 list_del_init(&data->pages);
942
943                 data->pagevec[0] = page;
944
945                 if (nbytes < wsize)
946                         wsize = nbytes;
947                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
948                                    wsize, offset, how);
949                 if (ret == 0)
950                         ret = ret2;
951                 offset += wsize;
952                 nbytes -= wsize;
953         } while (nbytes != 0);
954
955         return ret;
956
957 out_bad:
958         while (!list_empty(&list)) {
959                 data = list_entry(list.next, struct nfs_write_data, pages);
960                 list_del(&data->pages);
961                 nfs_writedata_release(data);
962         }
963         nfs_redirty_request(req);
964         return -ENOMEM;
965 }
966
967 /*
968  * Create an RPC task for the given write request and kick it.
969  * The page must have been locked by the caller.
970  *
971  * It may happen that the page we're passed is not marked dirty.
972  * This is the case if nfs_updatepage detects a conflicting request
973  * that has been written but not committed.
974  */
975 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
976 {
977         struct nfs_page         *req;
978         struct page             **pages;
979         struct nfs_write_data   *data;
980
981         data = nfs_writedata_alloc(npages);
982         if (!data)
983                 goto out_bad;
984
985         pages = data->pagevec;
986         while (!list_empty(head)) {
987                 req = nfs_list_entry(head->next);
988                 nfs_list_remove_request(req);
989                 nfs_list_add_request(req, &data->pages);
990                 ClearPageError(req->wb_page);
991                 *pages++ = req->wb_page;
992         }
993         req = nfs_list_entry(data->pages.next);
994
995         /* Set up the argument struct */
996         return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
997  out_bad:
998         while (!list_empty(head)) {
999                 req = nfs_list_entry(head->next);
1000                 nfs_list_remove_request(req);
1001                 nfs_redirty_request(req);
1002         }
1003         return -ENOMEM;
1004 }
1005
1006 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1007                                   struct inode *inode, int ioflags)
1008 {
1009         size_t wsize = NFS_SERVER(inode)->wsize;
1010
1011         if (wsize < PAGE_CACHE_SIZE)
1012                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
1013         else
1014                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
1015 }
1016
1017 /*
1018  * Handle a write reply that flushed part of a page.
1019  */
1020 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1021 {
1022         struct nfs_write_data   *data = calldata;
1023
1024         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1025                 task->tk_pid,
1026                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1027                 (long long)
1028                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1029                 data->req->wb_bytes, (long long)req_offset(data->req));
1030
1031         nfs_writeback_done(task, data);
1032 }
1033
1034 static void nfs_writeback_release_partial(void *calldata)
1035 {
1036         struct nfs_write_data   *data = calldata;
1037         struct nfs_page         *req = data->req;
1038         struct page             *page = req->wb_page;
1039         int status = data->task.tk_status;
1040
1041         if (status < 0) {
1042                 nfs_set_pageerror(page);
1043                 nfs_context_set_write_error(req->wb_context, status);
1044                 dprintk(", error = %d\n", status);
1045                 goto out;
1046         }
1047
1048         if (nfs_write_need_commit(data)) {
1049                 struct inode *inode = page->mapping->host;
1050
1051                 spin_lock(&inode->i_lock);
1052                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1053                         /* Do nothing we need to resend the writes */
1054                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1055                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1056                         dprintk(" defer commit\n");
1057                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1058                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1059                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1060                         dprintk(" server reboot detected\n");
1061                 }
1062                 spin_unlock(&inode->i_lock);
1063         } else
1064                 dprintk(" OK\n");
1065
1066 out:
1067         if (atomic_dec_and_test(&req->wb_complete))
1068                 nfs_writepage_release(req);
1069         nfs_writedata_release(calldata);
1070 }
1071
1072 #if defined(CONFIG_NFS_V4_1)
1073 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1074 {
1075         struct nfs_write_data *data = calldata;
1076         struct nfs_client *clp = (NFS_SERVER(data->inode))->nfs_client;
1077
1078         if (nfs4_setup_sequence(clp, &data->args.seq_args,
1079                                 &data->res.seq_res, 1, task))
1080                 return;
1081         rpc_call_start(task);
1082 }
1083 #endif /* CONFIG_NFS_V4_1 */
1084
1085 static const struct rpc_call_ops nfs_write_partial_ops = {
1086 #if defined(CONFIG_NFS_V4_1)
1087         .rpc_call_prepare = nfs_write_prepare,
1088 #endif /* CONFIG_NFS_V4_1 */
1089         .rpc_call_done = nfs_writeback_done_partial,
1090         .rpc_release = nfs_writeback_release_partial,
1091 };
1092
1093 /*
1094  * Handle a write reply that flushes a whole page.
1095  *
1096  * FIXME: There is an inherent race with invalidate_inode_pages and
1097  *        writebacks since the page->count is kept > 1 for as long
1098  *        as the page has a write request pending.
1099  */
1100 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1101 {
1102         struct nfs_write_data   *data = calldata;
1103
1104         nfs_writeback_done(task, data);
1105 }
1106
1107 static void nfs_writeback_release_full(void *calldata)
1108 {
1109         struct nfs_write_data   *data = calldata;
1110         int status = data->task.tk_status;
1111
1112         /* Update attributes as result of writeback. */
1113         while (!list_empty(&data->pages)) {
1114                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1115                 struct page *page = req->wb_page;
1116
1117                 nfs_list_remove_request(req);
1118
1119                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1120                         data->task.tk_pid,
1121                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1122                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1123                         req->wb_bytes,
1124                         (long long)req_offset(req));
1125
1126                 if (status < 0) {
1127                         nfs_set_pageerror(page);
1128                         nfs_context_set_write_error(req->wb_context, status);
1129                         dprintk(", error = %d\n", status);
1130                         goto remove_request;
1131                 }
1132
1133                 if (nfs_write_need_commit(data)) {
1134                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1135                         nfs_mark_request_commit(req);
1136                         nfs_end_page_writeback(page);
1137                         dprintk(" marked for commit\n");
1138                         goto next;
1139                 }
1140                 dprintk(" OK\n");
1141 remove_request:
1142                 nfs_end_page_writeback(page);
1143                 nfs_inode_remove_request(req);
1144         next:
1145                 nfs_clear_page_tag_locked(req);
1146         }
1147         nfs_writedata_release(calldata);
1148 }
1149
1150 static const struct rpc_call_ops nfs_write_full_ops = {
1151 #if defined(CONFIG_NFS_V4_1)
1152         .rpc_call_prepare = nfs_write_prepare,
1153 #endif /* CONFIG_NFS_V4_1 */
1154         .rpc_call_done = nfs_writeback_done_full,
1155         .rpc_release = nfs_writeback_release_full,
1156 };
1157
1158
1159 /*
1160  * This function is called when the WRITE call is complete.
1161  */
1162 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1163 {
1164         struct nfs_writeargs    *argp = &data->args;
1165         struct nfs_writeres     *resp = &data->res;
1166         struct nfs_server       *server = NFS_SERVER(data->inode);
1167         int status;
1168
1169         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1170                 task->tk_pid, task->tk_status);
1171
1172         /*
1173          * ->write_done will attempt to use post-op attributes to detect
1174          * conflicting writes by other clients.  A strict interpretation
1175          * of close-to-open would allow us to continue caching even if
1176          * another writer had changed the file, but some applications
1177          * depend on tighter cache coherency when writing.
1178          */
1179         status = NFS_PROTO(data->inode)->write_done(task, data);
1180         if (status != 0)
1181                 return status;
1182         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1183
1184 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1185         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1186                 /* We tried a write call, but the server did not
1187                  * commit data to stable storage even though we
1188                  * requested it.
1189                  * Note: There is a known bug in Tru64 < 5.0 in which
1190                  *       the server reports NFS_DATA_SYNC, but performs
1191                  *       NFS_FILE_SYNC. We therefore implement this checking
1192                  *       as a dprintk() in order to avoid filling syslog.
1193                  */
1194                 static unsigned long    complain;
1195
1196                 if (time_before(complain, jiffies)) {
1197                         dprintk("NFS:       faulty NFS server %s:"
1198                                 " (committed = %d) != (stable = %d)\n",
1199                                 server->nfs_client->cl_hostname,
1200                                 resp->verf->committed, argp->stable);
1201                         complain = jiffies + 300 * HZ;
1202                 }
1203         }
1204 #endif
1205         /* Is this a short write? */
1206         if (task->tk_status >= 0 && resp->count < argp->count) {
1207                 static unsigned long    complain;
1208
1209                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1210
1211                 /* Has the server at least made some progress? */
1212                 if (resp->count != 0) {
1213                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1214                         if (resp->verf->committed != NFS_UNSTABLE) {
1215                                 /* Resend from where the server left off */
1216                                 argp->offset += resp->count;
1217                                 argp->pgbase += resp->count;
1218                                 argp->count -= resp->count;
1219                         } else {
1220                                 /* Resend as a stable write in order to avoid
1221                                  * headaches in the case of a server crash.
1222                                  */
1223                                 argp->stable = NFS_FILE_SYNC;
1224                         }
1225                         nfs_restart_rpc(task, server->nfs_client);
1226                         return -EAGAIN;
1227                 }
1228                 if (time_before(complain, jiffies)) {
1229                         printk(KERN_WARNING
1230                                "NFS: Server wrote zero bytes, expected %u.\n",
1231                                         argp->count);
1232                         complain = jiffies + 300 * HZ;
1233                 }
1234                 /* Can't do anything about it except throw an error. */
1235                 task->tk_status = -EIO;
1236         }
1237         return 0;
1238 }
1239
1240
1241 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1242 static void nfs_commitdata_release(void *data)
1243 {
1244         struct nfs_write_data *wdata = data;
1245
1246         put_nfs_open_context(wdata->args.context);
1247         nfs_commit_free(wdata);
1248 }
1249
1250 /*
1251  * Set up the argument/result storage required for the RPC call.
1252  */
1253 static int nfs_commit_rpcsetup(struct list_head *head,
1254                 struct nfs_write_data *data,
1255                 int how)
1256 {
1257         struct nfs_page *first = nfs_list_entry(head->next);
1258         struct inode *inode = first->wb_context->path.dentry->d_inode;
1259         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1260         int priority = flush_task_priority(how);
1261         struct rpc_task *task;
1262         struct rpc_message msg = {
1263                 .rpc_argp = &data->args,
1264                 .rpc_resp = &data->res,
1265                 .rpc_cred = first->wb_context->cred,
1266         };
1267         struct rpc_task_setup task_setup_data = {
1268                 .task = &data->task,
1269                 .rpc_client = NFS_CLIENT(inode),
1270                 .rpc_message = &msg,
1271                 .callback_ops = &nfs_commit_ops,
1272                 .callback_data = data,
1273                 .workqueue = nfsiod_workqueue,
1274                 .flags = flags,
1275                 .priority = priority,
1276         };
1277
1278         /* Set up the RPC argument and reply structs
1279          * NB: take care not to mess about with data->commit et al. */
1280
1281         list_splice_init(head, &data->pages);
1282
1283         data->inode       = inode;
1284         data->cred        = msg.rpc_cred;
1285
1286         data->args.fh     = NFS_FH(data->inode);
1287         /* Note: we always request a commit of the entire inode */
1288         data->args.offset = 0;
1289         data->args.count  = 0;
1290         data->args.context = get_nfs_open_context(first->wb_context);
1291         data->res.count   = 0;
1292         data->res.fattr   = &data->fattr;
1293         data->res.verf    = &data->verf;
1294         nfs_fattr_init(&data->fattr);
1295
1296         /* Set up the initial task struct.  */
1297         NFS_PROTO(inode)->commit_setup(data, &msg);
1298
1299         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1300
1301         task = rpc_run_task(&task_setup_data);
1302         if (IS_ERR(task))
1303                 return PTR_ERR(task);
1304         rpc_put_task(task);
1305         return 0;
1306 }
1307
1308 /*
1309  * Commit dirty pages
1310  */
1311 static int
1312 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1313 {
1314         struct nfs_write_data   *data;
1315         struct nfs_page         *req;
1316
1317         data = nfs_commitdata_alloc();
1318
1319         if (!data)
1320                 goto out_bad;
1321
1322         /* Set up the argument struct */
1323         return nfs_commit_rpcsetup(head, data, how);
1324  out_bad:
1325         while (!list_empty(head)) {
1326                 req = nfs_list_entry(head->next);
1327                 nfs_list_remove_request(req);
1328                 nfs_mark_request_commit(req);
1329                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1330                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1331                                 BDI_RECLAIMABLE);
1332                 nfs_clear_page_tag_locked(req);
1333         }
1334         return -ENOMEM;
1335 }
1336
1337 /*
1338  * COMMIT call returned
1339  */
1340 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1341 {
1342         struct nfs_write_data   *data = calldata;
1343
1344         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1345                                 task->tk_pid, task->tk_status);
1346
1347         /* Call the NFS version-specific code */
1348         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1349                 return;
1350 }
1351
1352 static void nfs_commit_release(void *calldata)
1353 {
1354         struct nfs_write_data   *data = calldata;
1355         struct nfs_page         *req;
1356         int status = data->task.tk_status;
1357
1358         while (!list_empty(&data->pages)) {
1359                 req = nfs_list_entry(data->pages.next);
1360                 nfs_list_remove_request(req);
1361                 nfs_clear_request_commit(req);
1362
1363                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1364                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1365                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1366                         req->wb_bytes,
1367                         (long long)req_offset(req));
1368                 if (status < 0) {
1369                         nfs_context_set_write_error(req->wb_context, status);
1370                         nfs_inode_remove_request(req);
1371                         dprintk(", error = %d\n", status);
1372                         goto next;
1373                 }
1374
1375                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1376                  * returned by the server against all stored verfs. */
1377                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1378                         /* We have a match */
1379                         nfs_inode_remove_request(req);
1380                         dprintk(" OK\n");
1381                         goto next;
1382                 }
1383                 /* We have a mismatch. Write the page again */
1384                 dprintk(" mismatch\n");
1385                 nfs_mark_request_dirty(req);
1386         next:
1387                 nfs_clear_page_tag_locked(req);
1388         }
1389         nfs_commitdata_release(calldata);
1390 }
1391
1392 static const struct rpc_call_ops nfs_commit_ops = {
1393 #if defined(CONFIG_NFS_V4_1)
1394         .rpc_call_prepare = nfs_write_prepare,
1395 #endif /* CONFIG_NFS_V4_1 */
1396         .rpc_call_done = nfs_commit_done,
1397         .rpc_release = nfs_commit_release,
1398 };
1399
1400 static int nfs_commit_inode(struct inode *inode, int how)
1401 {
1402         LIST_HEAD(head);
1403         int res;
1404
1405         spin_lock(&inode->i_lock);
1406         res = nfs_scan_commit(inode, &head, 0, 0);
1407         spin_unlock(&inode->i_lock);
1408         if (res) {
1409                 int error = nfs_commit_list(inode, &head, how);
1410                 if (error < 0)
1411                         return error;
1412         }
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 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, 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 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1463 {
1464         struct inode *inode = mapping->host;
1465         pgoff_t idx_start, idx_end;
1466         unsigned int npages = 0;
1467         LIST_HEAD(head);
1468         int nocommit = how & FLUSH_NOCOMMIT;
1469         long pages, ret;
1470
1471         /* FIXME */
1472         if (wbc->range_cyclic)
1473                 idx_start = 0;
1474         else {
1475                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1476                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1477                 if (idx_end > idx_start) {
1478                         pgoff_t l_npages = 1 + idx_end - idx_start;
1479                         npages = l_npages;
1480                         if (sizeof(npages) != sizeof(l_npages) &&
1481                                         (pgoff_t)npages != l_npages)
1482                                 npages = 0;
1483                 }
1484         }
1485         how &= ~FLUSH_NOCOMMIT;
1486         spin_lock(&inode->i_lock);
1487         do {
1488                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1489                 if (ret != 0)
1490                         continue;
1491                 if (nocommit)
1492                         break;
1493                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1494                 if (pages == 0)
1495                         break;
1496                 if (how & FLUSH_INVALIDATE) {
1497                         spin_unlock(&inode->i_lock);
1498                         nfs_cancel_commit_list(&head);
1499                         ret = pages;
1500                         spin_lock(&inode->i_lock);
1501                         continue;
1502                 }
1503                 pages += nfs_scan_commit(inode, &head, 0, 0);
1504                 spin_unlock(&inode->i_lock);
1505                 ret = nfs_commit_list(inode, &head, how);
1506                 spin_lock(&inode->i_lock);
1507
1508         } while (ret >= 0);
1509         spin_unlock(&inode->i_lock);
1510         return ret;
1511 }
1512
1513 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1514 {
1515         int ret;
1516
1517         ret = nfs_writepages(mapping, wbc);
1518         if (ret < 0)
1519                 goto out;
1520         ret = nfs_sync_mapping_wait(mapping, wbc, how);
1521         if (ret < 0)
1522                 goto out;
1523         return 0;
1524 out:
1525         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1526         return ret;
1527 }
1528
1529 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1530 static int nfs_write_mapping(struct address_space *mapping, int how)
1531 {
1532         struct writeback_control wbc = {
1533                 .bdi = mapping->backing_dev_info,
1534                 .sync_mode = WB_SYNC_ALL,
1535                 .nr_to_write = LONG_MAX,
1536                 .range_start = 0,
1537                 .range_end = LLONG_MAX,
1538         };
1539
1540         return __nfs_write_mapping(mapping, &wbc, how);
1541 }
1542
1543 /*
1544  * flush the inode to disk.
1545  */
1546 int nfs_wb_all(struct inode *inode)
1547 {
1548         return nfs_write_mapping(inode->i_mapping, 0);
1549 }
1550
1551 int nfs_wb_nocommit(struct inode *inode)
1552 {
1553         return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1554 }
1555
1556 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1557 {
1558         struct nfs_page *req;
1559         loff_t range_start = page_offset(page);
1560         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1561         struct writeback_control wbc = {
1562                 .bdi = page->mapping->backing_dev_info,
1563                 .sync_mode = WB_SYNC_ALL,
1564                 .nr_to_write = LONG_MAX,
1565                 .range_start = range_start,
1566                 .range_end = range_end,
1567         };
1568         int ret = 0;
1569
1570         BUG_ON(!PageLocked(page));
1571         for (;;) {
1572                 req = nfs_page_find_request(page);
1573                 if (req == NULL)
1574                         goto out;
1575                 if (test_bit(PG_CLEAN, &req->wb_flags)) {
1576                         nfs_release_request(req);
1577                         break;
1578                 }
1579                 if (nfs_lock_request_dontget(req)) {
1580                         nfs_inode_remove_request(req);
1581                         /*
1582                          * In case nfs_inode_remove_request has marked the
1583                          * page as being dirty
1584                          */
1585                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1586                         nfs_unlock_request(req);
1587                         break;
1588                 }
1589                 ret = nfs_wait_on_request(req);
1590                 nfs_release_request(req);
1591                 if (ret < 0)
1592                         goto out;
1593         }
1594         if (!PagePrivate(page))
1595                 return 0;
1596         ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1597 out:
1598         return ret;
1599 }
1600
1601 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1602                                 int how)
1603 {
1604         loff_t range_start = page_offset(page);
1605         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1606         struct writeback_control wbc = {
1607                 .bdi = page->mapping->backing_dev_info,
1608                 .sync_mode = WB_SYNC_ALL,
1609                 .nr_to_write = LONG_MAX,
1610                 .range_start = range_start,
1611                 .range_end = range_end,
1612         };
1613         int ret;
1614
1615         do {
1616                 if (clear_page_dirty_for_io(page)) {
1617                         ret = nfs_writepage_locked(page, &wbc);
1618                         if (ret < 0)
1619                                 goto out_error;
1620                 } else if (!PagePrivate(page))
1621                         break;
1622                 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1623                 if (ret < 0)
1624                         goto out_error;
1625         } while (PagePrivate(page));
1626         return 0;
1627 out_error:
1628         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1629         return ret;
1630 }
1631
1632 /*
1633  * Write back all requests on one page - we do this before reading it.
1634  */
1635 int nfs_wb_page(struct inode *inode, struct page* page)
1636 {
1637         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1638 }
1639
1640 #ifdef CONFIG_MIGRATION
1641 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1642                 struct page *page)
1643 {
1644         struct nfs_page *req;
1645         int ret;
1646
1647         nfs_fscache_release_page(page, GFP_KERNEL);
1648
1649         req = nfs_find_and_lock_request(page);
1650         ret = PTR_ERR(req);
1651         if (IS_ERR(req))
1652                 goto out;
1653
1654         ret = migrate_page(mapping, newpage, page);
1655         if (!req)
1656                 goto out;
1657         if (ret)
1658                 goto out_unlock;
1659         page_cache_get(newpage);
1660         spin_lock(&mapping->host->i_lock);
1661         req->wb_page = newpage;
1662         SetPagePrivate(newpage);
1663         set_page_private(newpage, (unsigned long)req);
1664         ClearPagePrivate(page);
1665         set_page_private(page, 0);
1666         spin_unlock(&mapping->host->i_lock);
1667         page_cache_release(page);
1668 out_unlock:
1669         nfs_clear_page_tag_locked(req);
1670 out:
1671         return ret;
1672 }
1673 #endif
1674
1675 int __init nfs_init_writepagecache(void)
1676 {
1677         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1678                                              sizeof(struct nfs_write_data),
1679                                              0, SLAB_HWCACHE_ALIGN,
1680                                              NULL);
1681         if (nfs_wdata_cachep == NULL)
1682                 return -ENOMEM;
1683
1684         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1685                                                      nfs_wdata_cachep);
1686         if (nfs_wdata_mempool == NULL)
1687                 return -ENOMEM;
1688
1689         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1690                                                       nfs_wdata_cachep);
1691         if (nfs_commit_mempool == NULL)
1692                 return -ENOMEM;
1693
1694         /*
1695          * NFS congestion size, scale with available memory.
1696          *
1697          *  64MB:    8192k
1698          * 128MB:   11585k
1699          * 256MB:   16384k
1700          * 512MB:   23170k
1701          *   1GB:   32768k
1702          *   2GB:   46340k
1703          *   4GB:   65536k
1704          *   8GB:   92681k
1705          *  16GB:  131072k
1706          *
1707          * This allows larger machines to have larger/more transfers.
1708          * Limit the default to 256M
1709          */
1710         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1711         if (nfs_congestion_kb > 256*1024)
1712                 nfs_congestion_kb = 256*1024;
1713
1714         return 0;
1715 }
1716
1717 void nfs_destroy_writepagecache(void)
1718 {
1719         mempool_destroy(nfs_commit_mempool);
1720         mempool_destroy(nfs_wdata_mempool);
1721         kmem_cache_destroy(nfs_wdata_cachep);
1722 }
1723