NFSv4.1: layoutcommit
[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 #include "pnfs.h"
32
33 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
34
35 #define MIN_POOL_WRITE          (32)
36 #define MIN_POOL_COMMIT         (4)
37
38 /*
39  * Local function declarations
40  */
41 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
42                                   struct inode *inode, int ioflags);
43 static void nfs_redirty_request(struct nfs_page *req);
44 static const struct rpc_call_ops nfs_write_partial_ops;
45 static const struct rpc_call_ops nfs_write_full_ops;
46 static const struct rpc_call_ops nfs_commit_ops;
47
48 static struct kmem_cache *nfs_wdata_cachep;
49 static mempool_t *nfs_wdata_mempool;
50 static mempool_t *nfs_commit_mempool;
51
52 struct nfs_write_data *nfs_commitdata_alloc(void)
53 {
54         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
55
56         if (p) {
57                 memset(p, 0, sizeof(*p));
58                 INIT_LIST_HEAD(&p->pages);
59         }
60         return p;
61 }
62 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
63
64 void nfs_commit_free(struct nfs_write_data *p)
65 {
66         if (p && (p->pagevec != &p->page_array[0]))
67                 kfree(p->pagevec);
68         mempool_free(p, nfs_commit_mempool);
69 }
70 EXPORT_SYMBOL_GPL(nfs_commit_free);
71
72 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
73 {
74         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
75
76         if (p) {
77                 memset(p, 0, sizeof(*p));
78                 INIT_LIST_HEAD(&p->pages);
79                 p->npages = pagecount;
80                 if (pagecount <= ARRAY_SIZE(p->page_array))
81                         p->pagevec = p->page_array;
82                 else {
83                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
84                         if (!p->pagevec) {
85                                 mempool_free(p, nfs_wdata_mempool);
86                                 p = NULL;
87                         }
88                 }
89         }
90         return p;
91 }
92
93 void nfs_writedata_free(struct nfs_write_data *p)
94 {
95         if (p && (p->pagevec != &p->page_array[0]))
96                 kfree(p->pagevec);
97         mempool_free(p, nfs_wdata_mempool);
98 }
99
100 static void nfs_writedata_release(struct nfs_write_data *wdata)
101 {
102         put_lseg(wdata->lseg);
103         put_nfs_open_context(wdata->args.context);
104         nfs_writedata_free(wdata);
105 }
106
107 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
108 {
109         ctx->error = error;
110         smp_wmb();
111         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
112 }
113
114 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
115 {
116         struct nfs_page *req = NULL;
117
118         if (PagePrivate(page)) {
119                 req = (struct nfs_page *)page_private(page);
120                 if (req != NULL)
121                         kref_get(&req->wb_kref);
122         }
123         return req;
124 }
125
126 static struct nfs_page *nfs_page_find_request(struct page *page)
127 {
128         struct inode *inode = page->mapping->host;
129         struct nfs_page *req = NULL;
130
131         spin_lock(&inode->i_lock);
132         req = nfs_page_find_request_locked(page);
133         spin_unlock(&inode->i_lock);
134         return req;
135 }
136
137 /* Adjust the file length if we're writing beyond the end */
138 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
139 {
140         struct inode *inode = page->mapping->host;
141         loff_t end, i_size;
142         pgoff_t end_index;
143
144         spin_lock(&inode->i_lock);
145         i_size = i_size_read(inode);
146         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
147         if (i_size > 0 && page->index < end_index)
148                 goto out;
149         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
150         if (i_size >= end)
151                 goto out;
152         i_size_write(inode, end);
153         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
154 out:
155         spin_unlock(&inode->i_lock);
156 }
157
158 /* A writeback failed: mark the page as bad, and invalidate the page cache */
159 static void nfs_set_pageerror(struct page *page)
160 {
161         SetPageError(page);
162         nfs_zap_mapping(page->mapping->host, page->mapping);
163 }
164
165 /* We can set the PG_uptodate flag if we see that a write request
166  * covers the full page.
167  */
168 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
169 {
170         if (PageUptodate(page))
171                 return;
172         if (base != 0)
173                 return;
174         if (count != nfs_page_length(page))
175                 return;
176         SetPageUptodate(page);
177 }
178
179 static int wb_priority(struct writeback_control *wbc)
180 {
181         if (wbc->for_reclaim)
182                 return FLUSH_HIGHPRI | FLUSH_STABLE;
183         if (wbc->for_kupdate || wbc->for_background)
184                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
185         return FLUSH_COND_STABLE;
186 }
187
188 /*
189  * NFS congestion control
190  */
191
192 int nfs_congestion_kb;
193
194 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
195 #define NFS_CONGESTION_OFF_THRESH       \
196         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
197
198 static int nfs_set_page_writeback(struct page *page)
199 {
200         int ret = test_set_page_writeback(page);
201
202         if (!ret) {
203                 struct inode *inode = page->mapping->host;
204                 struct nfs_server *nfss = NFS_SERVER(inode);
205
206                 page_cache_get(page);
207                 if (atomic_long_inc_return(&nfss->writeback) >
208                                 NFS_CONGESTION_ON_THRESH) {
209                         set_bdi_congested(&nfss->backing_dev_info,
210                                                 BLK_RW_ASYNC);
211                 }
212         }
213         return ret;
214 }
215
216 static void nfs_end_page_writeback(struct page *page)
217 {
218         struct inode *inode = page->mapping->host;
219         struct nfs_server *nfss = NFS_SERVER(inode);
220
221         end_page_writeback(page);
222         page_cache_release(page);
223         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
224                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
225 }
226
227 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
228 {
229         struct inode *inode = page->mapping->host;
230         struct nfs_page *req;
231         int ret;
232
233         spin_lock(&inode->i_lock);
234         for (;;) {
235                 req = nfs_page_find_request_locked(page);
236                 if (req == NULL)
237                         break;
238                 if (nfs_set_page_tag_locked(req))
239                         break;
240                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
241                  *       then the call to nfs_set_page_tag_locked() will always
242                  *       succeed provided that someone hasn't already marked the
243                  *       request as dirty (in which case we don't care).
244                  */
245                 spin_unlock(&inode->i_lock);
246                 if (!nonblock)
247                         ret = nfs_wait_on_request(req);
248                 else
249                         ret = -EAGAIN;
250                 nfs_release_request(req);
251                 if (ret != 0)
252                         return ERR_PTR(ret);
253                 spin_lock(&inode->i_lock);
254         }
255         spin_unlock(&inode->i_lock);
256         return req;
257 }
258
259 /*
260  * Find an associated nfs write request, and prepare to flush it out
261  * May return an error if the user signalled nfs_wait_on_request().
262  */
263 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
264                                 struct page *page, bool nonblock)
265 {
266         struct nfs_page *req;
267         int ret = 0;
268
269         req = nfs_find_and_lock_request(page, nonblock);
270         if (!req)
271                 goto out;
272         ret = PTR_ERR(req);
273         if (IS_ERR(req))
274                 goto out;
275
276         ret = nfs_set_page_writeback(page);
277         BUG_ON(ret != 0);
278         BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
279
280         if (!nfs_pageio_add_request(pgio, req)) {
281                 nfs_redirty_request(req);
282                 ret = pgio->pg_error;
283         }
284 out:
285         return ret;
286 }
287
288 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
289 {
290         struct inode *inode = page->mapping->host;
291         int ret;
292
293         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
294         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
295
296         nfs_pageio_cond_complete(pgio, page->index);
297         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
298         if (ret == -EAGAIN) {
299                 redirty_page_for_writepage(wbc, page);
300                 ret = 0;
301         }
302         return ret;
303 }
304
305 /*
306  * Write an mmapped page to the server.
307  */
308 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
309 {
310         struct nfs_pageio_descriptor pgio;
311         int err;
312
313         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
314         err = nfs_do_writepage(page, wbc, &pgio);
315         nfs_pageio_complete(&pgio);
316         if (err < 0)
317                 return err;
318         if (pgio.pg_error < 0)
319                 return pgio.pg_error;
320         return 0;
321 }
322
323 int nfs_writepage(struct page *page, struct writeback_control *wbc)
324 {
325         int ret;
326
327         ret = nfs_writepage_locked(page, wbc);
328         unlock_page(page);
329         return ret;
330 }
331
332 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
333 {
334         int ret;
335
336         ret = nfs_do_writepage(page, wbc, data);
337         unlock_page(page);
338         return ret;
339 }
340
341 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
342 {
343         struct inode *inode = mapping->host;
344         unsigned long *bitlock = &NFS_I(inode)->flags;
345         struct nfs_pageio_descriptor pgio;
346         int err;
347
348         /* Stop dirtying of new pages while we sync */
349         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
350                         nfs_wait_bit_killable, TASK_KILLABLE);
351         if (err)
352                 goto out_err;
353
354         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
355
356         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
357         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
358         nfs_pageio_complete(&pgio);
359
360         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
361         smp_mb__after_clear_bit();
362         wake_up_bit(bitlock, NFS_INO_FLUSHING);
363
364         if (err < 0)
365                 goto out_err;
366         err = pgio.pg_error;
367         if (err < 0)
368                 goto out_err;
369         return 0;
370 out_err:
371         return err;
372 }
373
374 /*
375  * Insert a write request into an inode
376  */
377 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
378 {
379         struct nfs_inode *nfsi = NFS_I(inode);
380         int error;
381
382         error = radix_tree_preload(GFP_NOFS);
383         if (error != 0)
384                 goto out;
385
386         /* Lock the request! */
387         nfs_lock_request_dontget(req);
388
389         spin_lock(&inode->i_lock);
390         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
391         BUG_ON(error);
392         if (!nfsi->npages) {
393                 igrab(inode);
394                 if (nfs_have_delegation(inode, FMODE_WRITE))
395                         nfsi->change_attr++;
396         }
397         set_bit(PG_MAPPED, &req->wb_flags);
398         SetPagePrivate(req->wb_page);
399         set_page_private(req->wb_page, (unsigned long)req);
400         nfsi->npages++;
401         kref_get(&req->wb_kref);
402         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
403                                 NFS_PAGE_TAG_LOCKED);
404         spin_unlock(&inode->i_lock);
405         radix_tree_preload_end();
406 out:
407         return error;
408 }
409
410 /*
411  * Remove a write request from an inode
412  */
413 static void nfs_inode_remove_request(struct nfs_page *req)
414 {
415         struct inode *inode = req->wb_context->path.dentry->d_inode;
416         struct nfs_inode *nfsi = NFS_I(inode);
417
418         BUG_ON (!NFS_WBACK_BUSY(req));
419
420         spin_lock(&inode->i_lock);
421         set_page_private(req->wb_page, 0);
422         ClearPagePrivate(req->wb_page);
423         clear_bit(PG_MAPPED, &req->wb_flags);
424         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
425         nfsi->npages--;
426         if (!nfsi->npages) {
427                 spin_unlock(&inode->i_lock);
428                 iput(inode);
429         } else
430                 spin_unlock(&inode->i_lock);
431         nfs_release_request(req);
432 }
433
434 static void
435 nfs_mark_request_dirty(struct nfs_page *req)
436 {
437         __set_page_dirty_nobuffers(req->wb_page);
438         __mark_inode_dirty(req->wb_page->mapping->host, I_DIRTY_DATASYNC);
439 }
440
441 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
442 /*
443  * Add a request to the inode's commit list.
444  */
445 static void
446 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg)
447 {
448         struct inode *inode = req->wb_context->path.dentry->d_inode;
449         struct nfs_inode *nfsi = NFS_I(inode);
450
451         spin_lock(&inode->i_lock);
452         set_bit(PG_CLEAN, &(req)->wb_flags);
453         radix_tree_tag_set(&nfsi->nfs_page_tree,
454                         req->wb_index,
455                         NFS_PAGE_TAG_COMMIT);
456         nfsi->ncommit++;
457         spin_unlock(&inode->i_lock);
458         pnfs_mark_request_commit(req, lseg);
459         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
460         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
461         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
462 }
463
464 static int
465 nfs_clear_request_commit(struct nfs_page *req)
466 {
467         struct page *page = req->wb_page;
468
469         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
470                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
471                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
472                 return 1;
473         }
474         return 0;
475 }
476
477 static inline
478 int nfs_write_need_commit(struct nfs_write_data *data)
479 {
480         if (data->verf.committed == NFS_DATA_SYNC)
481                 return data->lseg == NULL;
482         else
483                 return data->verf.committed != NFS_FILE_SYNC;
484 }
485
486 static inline
487 int nfs_reschedule_unstable_write(struct nfs_page *req,
488                                   struct nfs_write_data *data)
489 {
490         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
491                 nfs_mark_request_commit(req, data->lseg);
492                 return 1;
493         }
494         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
495                 nfs_mark_request_dirty(req);
496                 return 1;
497         }
498         return 0;
499 }
500 #else
501 static inline void
502 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg)
503 {
504 }
505
506 static inline int
507 nfs_clear_request_commit(struct nfs_page *req)
508 {
509         return 0;
510 }
511
512 static inline
513 int nfs_write_need_commit(struct nfs_write_data *data)
514 {
515         return 0;
516 }
517
518 static inline
519 int nfs_reschedule_unstable_write(struct nfs_page *req,
520                                   struct nfs_write_data *data)
521 {
522         return 0;
523 }
524 #endif
525
526 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
527 static int
528 nfs_need_commit(struct nfs_inode *nfsi)
529 {
530         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
531 }
532
533 /*
534  * nfs_scan_commit - Scan an inode for commit requests
535  * @inode: NFS inode to scan
536  * @dst: destination list
537  * @idx_start: lower bound of page->index to scan.
538  * @npages: idx_start + npages sets the upper bound to scan.
539  *
540  * Moves requests from the inode's 'commit' request list.
541  * The requests are *not* checked to ensure that they form a contiguous set.
542  */
543 static int
544 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
545 {
546         struct nfs_inode *nfsi = NFS_I(inode);
547         int ret;
548
549         if (!nfs_need_commit(nfsi))
550                 return 0;
551
552         ret = nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
553         if (ret > 0)
554                 nfsi->ncommit -= ret;
555         if (nfs_need_commit(NFS_I(inode)))
556                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
557         return ret;
558 }
559 #else
560 static inline int nfs_need_commit(struct nfs_inode *nfsi)
561 {
562         return 0;
563 }
564
565 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
566 {
567         return 0;
568 }
569 #endif
570
571 /*
572  * Search for an existing write request, and attempt to update
573  * it to reflect a new dirty region on a given page.
574  *
575  * If the attempt fails, then the existing request is flushed out
576  * to disk.
577  */
578 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
579                 struct page *page,
580                 unsigned int offset,
581                 unsigned int bytes)
582 {
583         struct nfs_page *req;
584         unsigned int rqend;
585         unsigned int end;
586         int error;
587
588         if (!PagePrivate(page))
589                 return NULL;
590
591         end = offset + bytes;
592         spin_lock(&inode->i_lock);
593
594         for (;;) {
595                 req = nfs_page_find_request_locked(page);
596                 if (req == NULL)
597                         goto out_unlock;
598
599                 rqend = req->wb_offset + req->wb_bytes;
600                 /*
601                  * Tell the caller to flush out the request if
602                  * the offsets are non-contiguous.
603                  * Note: nfs_flush_incompatible() will already
604                  * have flushed out requests having wrong owners.
605                  */
606                 if (offset > rqend
607                     || end < req->wb_offset)
608                         goto out_flushme;
609
610                 if (nfs_set_page_tag_locked(req))
611                         break;
612
613                 /* The request is locked, so wait and then retry */
614                 spin_unlock(&inode->i_lock);
615                 error = nfs_wait_on_request(req);
616                 nfs_release_request(req);
617                 if (error != 0)
618                         goto out_err;
619                 spin_lock(&inode->i_lock);
620         }
621
622         if (nfs_clear_request_commit(req) &&
623             radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
624                                  req->wb_index, NFS_PAGE_TAG_COMMIT) != NULL) {
625                 NFS_I(inode)->ncommit--;
626                 pnfs_clear_request_commit(req);
627         }
628
629         /* Okay, the request matches. Update the region */
630         if (offset < req->wb_offset) {
631                 req->wb_offset = offset;
632                 req->wb_pgbase = offset;
633         }
634         if (end > rqend)
635                 req->wb_bytes = end - req->wb_offset;
636         else
637                 req->wb_bytes = rqend - req->wb_offset;
638 out_unlock:
639         spin_unlock(&inode->i_lock);
640         return req;
641 out_flushme:
642         spin_unlock(&inode->i_lock);
643         nfs_release_request(req);
644         error = nfs_wb_page(inode, page);
645 out_err:
646         return ERR_PTR(error);
647 }
648
649 /*
650  * Try to update an existing write request, or create one if there is none.
651  *
652  * Note: Should always be called with the Page Lock held to prevent races
653  * if we have to add a new request. Also assumes that the caller has
654  * already called nfs_flush_incompatible() if necessary.
655  */
656 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
657                 struct page *page, unsigned int offset, unsigned int bytes)
658 {
659         struct inode *inode = page->mapping->host;
660         struct nfs_page *req;
661         int error;
662
663         req = nfs_try_to_update_request(inode, page, offset, bytes);
664         if (req != NULL)
665                 goto out;
666         req = nfs_create_request(ctx, inode, page, offset, bytes);
667         if (IS_ERR(req))
668                 goto out;
669         error = nfs_inode_add_request(inode, req);
670         if (error != 0) {
671                 nfs_release_request(req);
672                 req = ERR_PTR(error);
673         }
674 out:
675         return req;
676 }
677
678 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
679                 unsigned int offset, unsigned int count)
680 {
681         struct nfs_page *req;
682
683         req = nfs_setup_write_request(ctx, page, offset, count);
684         if (IS_ERR(req))
685                 return PTR_ERR(req);
686         nfs_mark_request_dirty(req);
687         /* Update file length */
688         nfs_grow_file(page, offset, count);
689         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
690         nfs_mark_request_dirty(req);
691         nfs_clear_page_tag_locked(req);
692         return 0;
693 }
694
695 int nfs_flush_incompatible(struct file *file, struct page *page)
696 {
697         struct nfs_open_context *ctx = nfs_file_open_context(file);
698         struct nfs_page *req;
699         int do_flush, status;
700         /*
701          * Look for a request corresponding to this page. If there
702          * is one, and it belongs to another file, we flush it out
703          * before we try to copy anything into the page. Do this
704          * due to the lack of an ACCESS-type call in NFSv2.
705          * Also do the same if we find a request from an existing
706          * dropped page.
707          */
708         do {
709                 req = nfs_page_find_request(page);
710                 if (req == NULL)
711                         return 0;
712                 do_flush = req->wb_page != page || req->wb_context != ctx ||
713                         req->wb_lock_context->lockowner != current->files ||
714                         req->wb_lock_context->pid != current->tgid;
715                 nfs_release_request(req);
716                 if (!do_flush)
717                         return 0;
718                 status = nfs_wb_page(page->mapping->host, page);
719         } while (status == 0);
720         return status;
721 }
722
723 /*
724  * If the page cache is marked as unsafe or invalid, then we can't rely on
725  * the PageUptodate() flag. In this case, we will need to turn off
726  * write optimisations that depend on the page contents being correct.
727  */
728 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
729 {
730         return PageUptodate(page) &&
731                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
732 }
733
734 /*
735  * Update and possibly write a cached page of an NFS file.
736  *
737  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
738  * things with a page scheduled for an RPC call (e.g. invalidate it).
739  */
740 int nfs_updatepage(struct file *file, struct page *page,
741                 unsigned int offset, unsigned int count)
742 {
743         struct nfs_open_context *ctx = nfs_file_open_context(file);
744         struct inode    *inode = page->mapping->host;
745         int             status = 0;
746
747         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
748
749         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
750                 file->f_path.dentry->d_parent->d_name.name,
751                 file->f_path.dentry->d_name.name, count,
752                 (long long)(page_offset(page) + offset));
753
754         /* If we're not using byte range locks, and we know the page
755          * is up to date, it may be more efficient to extend the write
756          * to cover the entire page in order to avoid fragmentation
757          * inefficiencies.
758          */
759         if (nfs_write_pageuptodate(page, inode) &&
760                         inode->i_flock == NULL &&
761                         !(file->f_flags & O_DSYNC)) {
762                 count = max(count + offset, nfs_page_length(page));
763                 offset = 0;
764         }
765
766         status = nfs_writepage_setup(ctx, page, offset, count);
767         if (status < 0)
768                 nfs_set_pageerror(page);
769
770         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
771                         status, (long long)i_size_read(inode));
772         return status;
773 }
774
775 static void nfs_writepage_release(struct nfs_page *req,
776                                   struct nfs_write_data *data)
777 {
778         struct page *page = req->wb_page;
779
780         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req, data))
781                 nfs_inode_remove_request(req);
782         nfs_clear_page_tag_locked(req);
783         nfs_end_page_writeback(page);
784 }
785
786 static int flush_task_priority(int how)
787 {
788         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
789                 case FLUSH_HIGHPRI:
790                         return RPC_PRIORITY_HIGH;
791                 case FLUSH_LOWPRI:
792                         return RPC_PRIORITY_LOW;
793         }
794         return RPC_PRIORITY_NORMAL;
795 }
796
797 int nfs_initiate_write(struct nfs_write_data *data,
798                        struct rpc_clnt *clnt,
799                        const struct rpc_call_ops *call_ops,
800                        int how)
801 {
802         struct inode *inode = data->inode;
803         int priority = flush_task_priority(how);
804         struct rpc_task *task;
805         struct rpc_message msg = {
806                 .rpc_argp = &data->args,
807                 .rpc_resp = &data->res,
808                 .rpc_cred = data->cred,
809         };
810         struct rpc_task_setup task_setup_data = {
811                 .rpc_client = clnt,
812                 .task = &data->task,
813                 .rpc_message = &msg,
814                 .callback_ops = call_ops,
815                 .callback_data = data,
816                 .workqueue = nfsiod_workqueue,
817                 .flags = RPC_TASK_ASYNC,
818                 .priority = priority,
819         };
820         int ret = 0;
821
822         /* Set up the initial task struct.  */
823         NFS_PROTO(inode)->write_setup(data, &msg);
824
825         dprintk("NFS: %5u initiated write call "
826                 "(req %s/%lld, %u bytes @ offset %llu)\n",
827                 data->task.tk_pid,
828                 inode->i_sb->s_id,
829                 (long long)NFS_FILEID(inode),
830                 data->args.count,
831                 (unsigned long long)data->args.offset);
832
833         task = rpc_run_task(&task_setup_data);
834         if (IS_ERR(task)) {
835                 ret = PTR_ERR(task);
836                 goto out;
837         }
838         if (how & FLUSH_SYNC) {
839                 ret = rpc_wait_for_completion_task(task);
840                 if (ret == 0)
841                         ret = task->tk_status;
842         }
843         rpc_put_task(task);
844 out:
845         return ret;
846 }
847 EXPORT_SYMBOL_GPL(nfs_initiate_write);
848
849 /*
850  * Set up the argument/result storage required for the RPC call.
851  */
852 static int nfs_write_rpcsetup(struct nfs_page *req,
853                 struct nfs_write_data *data,
854                 const struct rpc_call_ops *call_ops,
855                 unsigned int count, unsigned int offset,
856                 struct pnfs_layout_segment *lseg,
857                 int how)
858 {
859         struct inode *inode = req->wb_context->path.dentry->d_inode;
860
861         /* Set up the RPC argument and reply structs
862          * NB: take care not to mess about with data->commit et al. */
863
864         data->req = req;
865         data->inode = inode = req->wb_context->path.dentry->d_inode;
866         data->cred = req->wb_context->cred;
867         data->lseg = get_lseg(lseg);
868
869         data->args.fh     = NFS_FH(inode);
870         data->args.offset = req_offset(req) + offset;
871         data->args.pgbase = req->wb_pgbase + offset;
872         data->args.pages  = data->pagevec;
873         data->args.count  = count;
874         data->args.context = get_nfs_open_context(req->wb_context);
875         data->args.lock_context = req->wb_lock_context;
876         data->args.stable  = NFS_UNSTABLE;
877         if (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
878                 data->args.stable = NFS_DATA_SYNC;
879                 if (!nfs_need_commit(NFS_I(inode)))
880                         data->args.stable = NFS_FILE_SYNC;
881         }
882
883         data->res.fattr   = &data->fattr;
884         data->res.count   = count;
885         data->res.verf    = &data->verf;
886         nfs_fattr_init(&data->fattr);
887
888         if (data->lseg &&
889             (pnfs_try_to_write_data(data, call_ops, how) == PNFS_ATTEMPTED))
890                 return 0;
891
892         return nfs_initiate_write(data, NFS_CLIENT(inode), call_ops, how);
893 }
894
895 /* If a nfs_flush_* function fails, it should remove reqs from @head and
896  * call this on each, which will prepare them to be retried on next
897  * writeback using standard nfs.
898  */
899 static void nfs_redirty_request(struct nfs_page *req)
900 {
901         struct page *page = req->wb_page;
902
903         nfs_mark_request_dirty(req);
904         nfs_clear_page_tag_locked(req);
905         nfs_end_page_writeback(page);
906 }
907
908 /*
909  * Generate multiple small requests to write out a single
910  * contiguous dirty area on one page.
911  */
912 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc)
913 {
914         struct nfs_page *req = nfs_list_entry(desc->pg_list.next);
915         struct page *page = req->wb_page;
916         struct nfs_write_data *data;
917         size_t wsize = NFS_SERVER(desc->pg_inode)->wsize, nbytes;
918         unsigned int offset;
919         int requests = 0;
920         int ret = 0;
921         struct pnfs_layout_segment *lseg;
922         LIST_HEAD(list);
923
924         nfs_list_remove_request(req);
925
926         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
927             (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit ||
928              desc->pg_count > wsize))
929                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
930
931
932         nbytes = desc->pg_count;
933         do {
934                 size_t len = min(nbytes, wsize);
935
936                 data = nfs_writedata_alloc(1);
937                 if (!data)
938                         goto out_bad;
939                 list_add(&data->pages, &list);
940                 requests++;
941                 nbytes -= len;
942         } while (nbytes != 0);
943         atomic_set(&req->wb_complete, requests);
944
945         BUG_ON(desc->pg_lseg);
946         lseg = pnfs_update_layout(desc->pg_inode, req->wb_context, IOMODE_RW);
947         ClearPageError(page);
948         offset = 0;
949         nbytes = desc->pg_count;
950         do {
951                 int ret2;
952
953                 data = list_entry(list.next, struct nfs_write_data, pages);
954                 list_del_init(&data->pages);
955
956                 data->pagevec[0] = page;
957
958                 if (nbytes < wsize)
959                         wsize = nbytes;
960                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
961                                           wsize, offset, lseg, desc->pg_ioflags);
962                 if (ret == 0)
963                         ret = ret2;
964                 offset += wsize;
965                 nbytes -= wsize;
966         } while (nbytes != 0);
967
968         put_lseg(lseg);
969         desc->pg_lseg = NULL;
970         return ret;
971
972 out_bad:
973         while (!list_empty(&list)) {
974                 data = list_entry(list.next, struct nfs_write_data, pages);
975                 list_del(&data->pages);
976                 nfs_writedata_free(data);
977         }
978         nfs_redirty_request(req);
979         return -ENOMEM;
980 }
981
982 /*
983  * Create an RPC task for the given write request and kick it.
984  * The page must have been locked by the caller.
985  *
986  * It may happen that the page we're passed is not marked dirty.
987  * This is the case if nfs_updatepage detects a conflicting request
988  * that has been written but not committed.
989  */
990 static int nfs_flush_one(struct nfs_pageio_descriptor *desc)
991 {
992         struct nfs_page         *req;
993         struct page             **pages;
994         struct nfs_write_data   *data;
995         struct list_head *head = &desc->pg_list;
996         struct pnfs_layout_segment *lseg = desc->pg_lseg;
997         int ret;
998
999         data = nfs_writedata_alloc(nfs_page_array_len(desc->pg_base,
1000                                                       desc->pg_count));
1001         if (!data) {
1002                 while (!list_empty(head)) {
1003                         req = nfs_list_entry(head->next);
1004                         nfs_list_remove_request(req);
1005                         nfs_redirty_request(req);
1006                 }
1007                 ret = -ENOMEM;
1008                 goto out;
1009         }
1010         pages = data->pagevec;
1011         while (!list_empty(head)) {
1012                 req = nfs_list_entry(head->next);
1013                 nfs_list_remove_request(req);
1014                 nfs_list_add_request(req, &data->pages);
1015                 ClearPageError(req->wb_page);
1016                 *pages++ = req->wb_page;
1017         }
1018         req = nfs_list_entry(data->pages.next);
1019         if ((!lseg) && list_is_singular(&data->pages))
1020                 lseg = pnfs_update_layout(desc->pg_inode, req->wb_context, IOMODE_RW);
1021
1022         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1023             (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit))
1024                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1025
1026         /* Set up the argument struct */
1027         ret = nfs_write_rpcsetup(req, data, &nfs_write_full_ops, desc->pg_count, 0, lseg, desc->pg_ioflags);
1028 out:
1029         put_lseg(lseg); /* Cleans any gotten in ->pg_test */
1030         desc->pg_lseg = NULL;
1031         return ret;
1032 }
1033
1034 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1035                                   struct inode *inode, int ioflags)
1036 {
1037         size_t wsize = NFS_SERVER(inode)->wsize;
1038
1039         pnfs_pageio_init_write(pgio, inode);
1040
1041         if (wsize < PAGE_CACHE_SIZE)
1042                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
1043         else
1044                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
1045 }
1046
1047 /*
1048  * Handle a write reply that flushed part of a page.
1049  */
1050 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1051 {
1052         struct nfs_write_data   *data = calldata;
1053
1054         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1055                 task->tk_pid,
1056                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1057                 (long long)
1058                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1059                 data->req->wb_bytes, (long long)req_offset(data->req));
1060
1061         nfs_writeback_done(task, data);
1062 }
1063
1064 static void nfs_writeback_release_partial(void *calldata)
1065 {
1066         struct nfs_write_data   *data = calldata;
1067         struct nfs_page         *req = data->req;
1068         struct page             *page = req->wb_page;
1069         int status = data->task.tk_status;
1070
1071         if (status < 0) {
1072                 nfs_set_pageerror(page);
1073                 nfs_context_set_write_error(req->wb_context, status);
1074                 dprintk(", error = %d\n", status);
1075                 goto out;
1076         }
1077
1078         if (nfs_write_need_commit(data)) {
1079                 struct inode *inode = page->mapping->host;
1080
1081                 spin_lock(&inode->i_lock);
1082                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1083                         /* Do nothing we need to resend the writes */
1084                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1085                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1086                         dprintk(" defer commit\n");
1087                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1088                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1089                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1090                         dprintk(" server reboot detected\n");
1091                 }
1092                 spin_unlock(&inode->i_lock);
1093         } else
1094                 dprintk(" OK\n");
1095
1096 out:
1097         if (atomic_dec_and_test(&req->wb_complete))
1098                 nfs_writepage_release(req, data);
1099         nfs_writedata_release(calldata);
1100 }
1101
1102 #if defined(CONFIG_NFS_V4_1)
1103 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1104 {
1105         struct nfs_write_data *data = calldata;
1106
1107         if (nfs4_setup_sequence(NFS_SERVER(data->inode),
1108                                 &data->args.seq_args,
1109                                 &data->res.seq_res, 1, task))
1110                 return;
1111         rpc_call_start(task);
1112 }
1113 #endif /* CONFIG_NFS_V4_1 */
1114
1115 static const struct rpc_call_ops nfs_write_partial_ops = {
1116 #if defined(CONFIG_NFS_V4_1)
1117         .rpc_call_prepare = nfs_write_prepare,
1118 #endif /* CONFIG_NFS_V4_1 */
1119         .rpc_call_done = nfs_writeback_done_partial,
1120         .rpc_release = nfs_writeback_release_partial,
1121 };
1122
1123 /*
1124  * Handle a write reply that flushes a whole page.
1125  *
1126  * FIXME: There is an inherent race with invalidate_inode_pages and
1127  *        writebacks since the page->count is kept > 1 for as long
1128  *        as the page has a write request pending.
1129  */
1130 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1131 {
1132         struct nfs_write_data   *data = calldata;
1133
1134         nfs_writeback_done(task, data);
1135 }
1136
1137 static void nfs_writeback_release_full(void *calldata)
1138 {
1139         struct nfs_write_data   *data = calldata;
1140         int status = data->task.tk_status;
1141
1142         /* Update attributes as result of writeback. */
1143         while (!list_empty(&data->pages)) {
1144                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1145                 struct page *page = req->wb_page;
1146
1147                 nfs_list_remove_request(req);
1148
1149                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1150                         data->task.tk_pid,
1151                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1152                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1153                         req->wb_bytes,
1154                         (long long)req_offset(req));
1155
1156                 if (status < 0) {
1157                         nfs_set_pageerror(page);
1158                         nfs_context_set_write_error(req->wb_context, status);
1159                         dprintk(", error = %d\n", status);
1160                         goto remove_request;
1161                 }
1162
1163                 if (nfs_write_need_commit(data)) {
1164                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1165                         nfs_mark_request_commit(req, data->lseg);
1166                         dprintk(" marked for commit\n");
1167                         goto next;
1168                 }
1169                 dprintk(" OK\n");
1170 remove_request:
1171                 nfs_inode_remove_request(req);
1172         next:
1173                 nfs_clear_page_tag_locked(req);
1174                 nfs_end_page_writeback(page);
1175         }
1176         nfs_writedata_release(calldata);
1177 }
1178
1179 static const struct rpc_call_ops nfs_write_full_ops = {
1180 #if defined(CONFIG_NFS_V4_1)
1181         .rpc_call_prepare = nfs_write_prepare,
1182 #endif /* CONFIG_NFS_V4_1 */
1183         .rpc_call_done = nfs_writeback_done_full,
1184         .rpc_release = nfs_writeback_release_full,
1185 };
1186
1187
1188 /*
1189  * This function is called when the WRITE call is complete.
1190  */
1191 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1192 {
1193         struct nfs_writeargs    *argp = &data->args;
1194         struct nfs_writeres     *resp = &data->res;
1195         struct nfs_server       *server = NFS_SERVER(data->inode);
1196         int status;
1197
1198         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1199                 task->tk_pid, task->tk_status);
1200
1201         /*
1202          * ->write_done will attempt to use post-op attributes to detect
1203          * conflicting writes by other clients.  A strict interpretation
1204          * of close-to-open would allow us to continue caching even if
1205          * another writer had changed the file, but some applications
1206          * depend on tighter cache coherency when writing.
1207          */
1208         status = NFS_PROTO(data->inode)->write_done(task, data);
1209         if (status != 0)
1210                 return;
1211         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1212
1213 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1214         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1215                 /* We tried a write call, but the server did not
1216                  * commit data to stable storage even though we
1217                  * requested it.
1218                  * Note: There is a known bug in Tru64 < 5.0 in which
1219                  *       the server reports NFS_DATA_SYNC, but performs
1220                  *       NFS_FILE_SYNC. We therefore implement this checking
1221                  *       as a dprintk() in order to avoid filling syslog.
1222                  */
1223                 static unsigned long    complain;
1224
1225                 /* Note this will print the MDS for a DS write */
1226                 if (time_before(complain, jiffies)) {
1227                         dprintk("NFS:       faulty NFS server %s:"
1228                                 " (committed = %d) != (stable = %d)\n",
1229                                 server->nfs_client->cl_hostname,
1230                                 resp->verf->committed, argp->stable);
1231                         complain = jiffies + 300 * HZ;
1232                 }
1233         }
1234 #endif
1235         /* Is this a short write? */
1236         if (task->tk_status >= 0 && resp->count < argp->count) {
1237                 static unsigned long    complain;
1238
1239                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1240
1241                 /* Has the server at least made some progress? */
1242                 if (resp->count != 0) {
1243                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1244                         if (resp->verf->committed != NFS_UNSTABLE) {
1245                                 /* Resend from where the server left off */
1246                                 data->mds_offset += resp->count;
1247                                 argp->offset += resp->count;
1248                                 argp->pgbase += resp->count;
1249                                 argp->count -= resp->count;
1250                         } else {
1251                                 /* Resend as a stable write in order to avoid
1252                                  * headaches in the case of a server crash.
1253                                  */
1254                                 argp->stable = NFS_FILE_SYNC;
1255                         }
1256                         nfs_restart_rpc(task, server->nfs_client);
1257                         return;
1258                 }
1259                 if (time_before(complain, jiffies)) {
1260                         printk(KERN_WARNING
1261                                "NFS: Server wrote zero bytes, expected %u.\n",
1262                                         argp->count);
1263                         complain = jiffies + 300 * HZ;
1264                 }
1265                 /* Can't do anything about it except throw an error. */
1266                 task->tk_status = -EIO;
1267         }
1268         return;
1269 }
1270
1271
1272 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1273 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1274 {
1275         int ret;
1276
1277         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1278                 return 1;
1279         if (!may_wait)
1280                 return 0;
1281         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1282                                 NFS_INO_COMMIT,
1283                                 nfs_wait_bit_killable,
1284                                 TASK_KILLABLE);
1285         return (ret < 0) ? ret : 1;
1286 }
1287
1288 void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1289 {
1290         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1291         smp_mb__after_clear_bit();
1292         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1293 }
1294 EXPORT_SYMBOL_GPL(nfs_commit_clear_lock);
1295
1296 void nfs_commitdata_release(void *data)
1297 {
1298         struct nfs_write_data *wdata = data;
1299
1300         put_lseg(wdata->lseg);
1301         put_nfs_open_context(wdata->args.context);
1302         nfs_commit_free(wdata);
1303 }
1304 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1305
1306 int nfs_initiate_commit(struct nfs_write_data *data, struct rpc_clnt *clnt,
1307                         const struct rpc_call_ops *call_ops,
1308                         int how)
1309 {
1310         struct rpc_task *task;
1311         int priority = flush_task_priority(how);
1312         struct rpc_message msg = {
1313                 .rpc_argp = &data->args,
1314                 .rpc_resp = &data->res,
1315                 .rpc_cred = data->cred,
1316         };
1317         struct rpc_task_setup task_setup_data = {
1318                 .task = &data->task,
1319                 .rpc_client = clnt,
1320                 .rpc_message = &msg,
1321                 .callback_ops = call_ops,
1322                 .callback_data = data,
1323                 .workqueue = nfsiod_workqueue,
1324                 .flags = RPC_TASK_ASYNC,
1325                 .priority = priority,
1326         };
1327         /* Set up the initial task struct.  */
1328         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1329
1330         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1331
1332         task = rpc_run_task(&task_setup_data);
1333         if (IS_ERR(task))
1334                 return PTR_ERR(task);
1335         if (how & FLUSH_SYNC)
1336                 rpc_wait_for_completion_task(task);
1337         rpc_put_task(task);
1338         return 0;
1339 }
1340 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1341
1342 /*
1343  * Set up the argument/result storage required for the RPC call.
1344  */
1345 void nfs_init_commit(struct nfs_write_data *data,
1346                             struct list_head *head,
1347                             struct pnfs_layout_segment *lseg)
1348 {
1349         struct nfs_page *first = nfs_list_entry(head->next);
1350         struct inode *inode = first->wb_context->path.dentry->d_inode;
1351
1352         /* Set up the RPC argument and reply structs
1353          * NB: take care not to mess about with data->commit et al. */
1354
1355         list_splice_init(head, &data->pages);
1356
1357         data->inode       = inode;
1358         data->cred        = first->wb_context->cred;
1359         data->lseg        = lseg; /* reference transferred */
1360         data->mds_ops     = &nfs_commit_ops;
1361
1362         data->args.fh     = NFS_FH(data->inode);
1363         /* Note: we always request a commit of the entire inode */
1364         data->args.offset = 0;
1365         data->args.count  = 0;
1366         data->args.context = get_nfs_open_context(first->wb_context);
1367         data->res.count   = 0;
1368         data->res.fattr   = &data->fattr;
1369         data->res.verf    = &data->verf;
1370         nfs_fattr_init(&data->fattr);
1371 }
1372 EXPORT_SYMBOL_GPL(nfs_init_commit);
1373
1374 void nfs_retry_commit(struct list_head *page_list,
1375                       struct pnfs_layout_segment *lseg)
1376 {
1377         struct nfs_page *req;
1378
1379         while (!list_empty(page_list)) {
1380                 req = nfs_list_entry(page_list->next);
1381                 nfs_list_remove_request(req);
1382                 nfs_mark_request_commit(req, lseg);
1383                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1384                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1385                              BDI_RECLAIMABLE);
1386                 nfs_clear_page_tag_locked(req);
1387         }
1388 }
1389 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1390
1391 /*
1392  * Commit dirty pages
1393  */
1394 static int
1395 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1396 {
1397         struct nfs_write_data   *data;
1398
1399         data = nfs_commitdata_alloc();
1400
1401         if (!data)
1402                 goto out_bad;
1403
1404         /* Set up the argument struct */
1405         nfs_init_commit(data, head, NULL);
1406         return nfs_initiate_commit(data, NFS_CLIENT(inode), data->mds_ops, how);
1407  out_bad:
1408         nfs_retry_commit(head, NULL);
1409         nfs_commit_clear_lock(NFS_I(inode));
1410         return -ENOMEM;
1411 }
1412
1413 /*
1414  * COMMIT call returned
1415  */
1416 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1417 {
1418         struct nfs_write_data   *data = calldata;
1419
1420         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1421                                 task->tk_pid, task->tk_status);
1422
1423         /* Call the NFS version-specific code */
1424         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1425                 return;
1426 }
1427
1428 void nfs_commit_release_pages(struct nfs_write_data *data)
1429 {
1430         struct nfs_page *req;
1431         int status = data->task.tk_status;
1432
1433         while (!list_empty(&data->pages)) {
1434                 req = nfs_list_entry(data->pages.next);
1435                 nfs_list_remove_request(req);
1436                 nfs_clear_request_commit(req);
1437
1438                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1439                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1440                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1441                         req->wb_bytes,
1442                         (long long)req_offset(req));
1443                 if (status < 0) {
1444                         nfs_context_set_write_error(req->wb_context, status);
1445                         nfs_inode_remove_request(req);
1446                         dprintk(", error = %d\n", status);
1447                         goto next;
1448                 }
1449
1450                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1451                  * returned by the server against all stored verfs. */
1452                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1453                         /* We have a match */
1454                         nfs_inode_remove_request(req);
1455                         dprintk(" OK\n");
1456                         goto next;
1457                 }
1458                 /* We have a mismatch. Write the page again */
1459                 dprintk(" mismatch\n");
1460                 nfs_mark_request_dirty(req);
1461         next:
1462                 nfs_clear_page_tag_locked(req);
1463         }
1464 }
1465 EXPORT_SYMBOL_GPL(nfs_commit_release_pages);
1466
1467 static void nfs_commit_release(void *calldata)
1468 {
1469         struct nfs_write_data *data = calldata;
1470
1471         nfs_commit_release_pages(data);
1472         nfs_commit_clear_lock(NFS_I(data->inode));
1473         nfs_commitdata_release(calldata);
1474 }
1475
1476 static const struct rpc_call_ops nfs_commit_ops = {
1477 #if defined(CONFIG_NFS_V4_1)
1478         .rpc_call_prepare = nfs_write_prepare,
1479 #endif /* CONFIG_NFS_V4_1 */
1480         .rpc_call_done = nfs_commit_done,
1481         .rpc_release = nfs_commit_release,
1482 };
1483
1484 int nfs_commit_inode(struct inode *inode, int how)
1485 {
1486         LIST_HEAD(head);
1487         int may_wait = how & FLUSH_SYNC;
1488         int res;
1489
1490         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1491         if (res <= 0)
1492                 goto out_mark_dirty;
1493         spin_lock(&inode->i_lock);
1494         res = nfs_scan_commit(inode, &head, 0, 0);
1495         spin_unlock(&inode->i_lock);
1496         if (res) {
1497                 int error;
1498
1499                 error = pnfs_commit_list(inode, &head, how);
1500                 if (error == PNFS_NOT_ATTEMPTED)
1501                         error = nfs_commit_list(inode, &head, how);
1502                 if (error < 0)
1503                         return error;
1504                 if (!may_wait)
1505                         goto out_mark_dirty;
1506                 error = wait_on_bit(&NFS_I(inode)->flags,
1507                                 NFS_INO_COMMIT,
1508                                 nfs_wait_bit_killable,
1509                                 TASK_KILLABLE);
1510                 if (error < 0)
1511                         return error;
1512         } else
1513                 nfs_commit_clear_lock(NFS_I(inode));
1514         return res;
1515         /* Note: If we exit without ensuring that the commit is complete,
1516          * we must mark the inode as dirty. Otherwise, future calls to
1517          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1518          * that the data is on the disk.
1519          */
1520 out_mark_dirty:
1521         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1522         return res;
1523 }
1524
1525 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1526 {
1527         struct nfs_inode *nfsi = NFS_I(inode);
1528         int flags = FLUSH_SYNC;
1529         int ret = 0;
1530
1531         if (wbc->sync_mode == WB_SYNC_NONE) {
1532                 /* Don't commit yet if this is a non-blocking flush and there
1533                  * are a lot of outstanding writes for this mapping.
1534                  */
1535                 if (nfsi->ncommit <= (nfsi->npages >> 1))
1536                         goto out_mark_dirty;
1537
1538                 /* don't wait for the COMMIT response */
1539                 flags = 0;
1540         }
1541
1542         ret = nfs_commit_inode(inode, flags);
1543         if (ret >= 0) {
1544                 if (wbc->sync_mode == WB_SYNC_NONE) {
1545                         if (ret < wbc->nr_to_write)
1546                                 wbc->nr_to_write -= ret;
1547                         else
1548                                 wbc->nr_to_write = 0;
1549                 }
1550                 return 0;
1551         }
1552 out_mark_dirty:
1553         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1554         return ret;
1555 }
1556 #else
1557 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1558 {
1559         return 0;
1560 }
1561 #endif
1562
1563 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1564 {
1565         int ret;
1566
1567         ret = nfs_commit_unstable_pages(inode, wbc);
1568         if (ret >= 0 && test_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags)) {
1569                 int status, sync = wbc->sync_mode;
1570
1571                 if (wbc->nonblocking || wbc->for_background)
1572                                 sync = 0;
1573
1574                 status = pnfs_layoutcommit_inode(inode, sync);
1575                 if (status < 0)
1576                         return status;
1577         }
1578         return ret;
1579 }
1580
1581 /*
1582  * flush the inode to disk.
1583  */
1584 int nfs_wb_all(struct inode *inode)
1585 {
1586         struct writeback_control wbc = {
1587                 .sync_mode = WB_SYNC_ALL,
1588                 .nr_to_write = LONG_MAX,
1589                 .range_start = 0,
1590                 .range_end = LLONG_MAX,
1591         };
1592
1593         return sync_inode(inode, &wbc);
1594 }
1595
1596 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1597 {
1598         struct nfs_page *req;
1599         int ret = 0;
1600
1601         BUG_ON(!PageLocked(page));
1602         for (;;) {
1603                 wait_on_page_writeback(page);
1604                 req = nfs_page_find_request(page);
1605                 if (req == NULL)
1606                         break;
1607                 if (nfs_lock_request_dontget(req)) {
1608                         nfs_inode_remove_request(req);
1609                         /*
1610                          * In case nfs_inode_remove_request has marked the
1611                          * page as being dirty
1612                          */
1613                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1614                         nfs_unlock_request(req);
1615                         break;
1616                 }
1617                 ret = nfs_wait_on_request(req);
1618                 nfs_release_request(req);
1619                 if (ret < 0)
1620                         break;
1621         }
1622         return ret;
1623 }
1624
1625 /*
1626  * Write back all requests on one page - we do this before reading it.
1627  */
1628 int nfs_wb_page(struct inode *inode, struct page *page)
1629 {
1630         loff_t range_start = page_offset(page);
1631         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1632         struct writeback_control wbc = {
1633                 .sync_mode = WB_SYNC_ALL,
1634                 .nr_to_write = 0,
1635                 .range_start = range_start,
1636                 .range_end = range_end,
1637         };
1638         int ret;
1639
1640         for (;;) {
1641                 wait_on_page_writeback(page);
1642                 if (clear_page_dirty_for_io(page)) {
1643                         ret = nfs_writepage_locked(page, &wbc);
1644                         if (ret < 0)
1645                                 goto out_error;
1646                         continue;
1647                 }
1648                 if (!PagePrivate(page))
1649                         break;
1650                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1651                 if (ret < 0)
1652                         goto out_error;
1653         }
1654         return 0;
1655 out_error:
1656         return ret;
1657 }
1658
1659 #ifdef CONFIG_MIGRATION
1660 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1661                 struct page *page)
1662 {
1663         struct nfs_page *req;
1664         int ret;
1665
1666         nfs_fscache_release_page(page, GFP_KERNEL);
1667
1668         req = nfs_find_and_lock_request(page, false);
1669         ret = PTR_ERR(req);
1670         if (IS_ERR(req))
1671                 goto out;
1672
1673         ret = migrate_page(mapping, newpage, page);
1674         if (!req)
1675                 goto out;
1676         if (ret)
1677                 goto out_unlock;
1678         page_cache_get(newpage);
1679         spin_lock(&mapping->host->i_lock);
1680         req->wb_page = newpage;
1681         SetPagePrivate(newpage);
1682         set_page_private(newpage, (unsigned long)req);
1683         ClearPagePrivate(page);
1684         set_page_private(page, 0);
1685         spin_unlock(&mapping->host->i_lock);
1686         page_cache_release(page);
1687 out_unlock:
1688         nfs_clear_page_tag_locked(req);
1689 out:
1690         return ret;
1691 }
1692 #endif
1693
1694 int __init nfs_init_writepagecache(void)
1695 {
1696         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1697                                              sizeof(struct nfs_write_data),
1698                                              0, SLAB_HWCACHE_ALIGN,
1699                                              NULL);
1700         if (nfs_wdata_cachep == NULL)
1701                 return -ENOMEM;
1702
1703         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1704                                                      nfs_wdata_cachep);
1705         if (nfs_wdata_mempool == NULL)
1706                 return -ENOMEM;
1707
1708         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1709                                                       nfs_wdata_cachep);
1710         if (nfs_commit_mempool == NULL)
1711                 return -ENOMEM;
1712
1713         /*
1714          * NFS congestion size, scale with available memory.
1715          *
1716          *  64MB:    8192k
1717          * 128MB:   11585k
1718          * 256MB:   16384k
1719          * 512MB:   23170k
1720          *   1GB:   32768k
1721          *   2GB:   46340k
1722          *   4GB:   65536k
1723          *   8GB:   92681k
1724          *  16GB:  131072k
1725          *
1726          * This allows larger machines to have larger/more transfers.
1727          * Limit the default to 256M
1728          */
1729         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1730         if (nfs_congestion_kb > 256*1024)
1731                 nfs_congestion_kb = 256*1024;
1732
1733         return 0;
1734 }
1735
1736 void nfs_destroy_writepagecache(void)
1737 {
1738         mempool_destroy(nfs_commit_mempool);
1739         mempool_destroy(nfs_wdata_mempool);
1740         kmem_cache_destroy(nfs_wdata_cachep);
1741 }
1742