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