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