605e292501f493cbf921c8ee0a4b22caf863572c
[linux-2.6.git] / fs / nfs / write.c
1 /*
2  * linux/fs/nfs/write.c
3  *
4  * Write file data over NFS.
5  *
6  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
7  */
8
9 #include <linux/types.h>
10 #include <linux/slab.h>
11 #include <linux/mm.h>
12 #include <linux/pagemap.h>
13 #include <linux/file.h>
14 #include <linux/writeback.h>
15 #include <linux/swap.h>
16 #include <linux/migrate.h>
17
18 #include <linux/sunrpc/clnt.h>
19 #include <linux/nfs_fs.h>
20 #include <linux/nfs_mount.h>
21 #include <linux/nfs_page.h>
22 #include <linux/backing-dev.h>
23
24 #include <asm/uaccess.h>
25
26 #include "delegation.h"
27 #include "internal.h"
28 #include "iostat.h"
29 #include "nfs4_fs.h"
30 #include "fscache.h"
31
32 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
33
34 #define MIN_POOL_WRITE          (32)
35 #define MIN_POOL_COMMIT         (4)
36
37 /*
38  * Local function declarations
39  */
40 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
41                                   struct inode *inode, int ioflags);
42 static void nfs_redirty_request(struct nfs_page *req);
43 static const struct rpc_call_ops nfs_write_partial_ops;
44 static const struct rpc_call_ops nfs_write_full_ops;
45 static const struct rpc_call_ops nfs_commit_ops;
46
47 static struct kmem_cache *nfs_wdata_cachep;
48 static mempool_t *nfs_wdata_mempool;
49 static mempool_t *nfs_commit_mempool;
50
51 struct nfs_write_data *nfs_commitdata_alloc(void)
52 {
53         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
54
55         if (p) {
56                 memset(p, 0, sizeof(*p));
57                 INIT_LIST_HEAD(&p->pages);
58         }
59         return p;
60 }
61
62 void nfs_commit_free(struct nfs_write_data *p)
63 {
64         if (p && (p->pagevec != &p->page_array[0]))
65                 kfree(p->pagevec);
66         mempool_free(p, nfs_commit_mempool);
67 }
68
69 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
70 {
71         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
72
73         if (p) {
74                 memset(p, 0, sizeof(*p));
75                 INIT_LIST_HEAD(&p->pages);
76                 p->npages = pagecount;
77                 if (pagecount <= ARRAY_SIZE(p->page_array))
78                         p->pagevec = p->page_array;
79                 else {
80                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
81                         if (!p->pagevec) {
82                                 mempool_free(p, nfs_wdata_mempool);
83                                 p = NULL;
84                         }
85                 }
86         }
87         return p;
88 }
89
90 void nfs_writedata_free(struct nfs_write_data *p)
91 {
92         if (p && (p->pagevec != &p->page_array[0]))
93                 kfree(p->pagevec);
94         mempool_free(p, nfs_wdata_mempool);
95 }
96
97 static void nfs_writedata_release(struct nfs_write_data *wdata)
98 {
99         put_nfs_open_context(wdata->args.context);
100         nfs_writedata_free(wdata);
101 }
102
103 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
104 {
105         ctx->error = error;
106         smp_wmb();
107         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
108 }
109
110 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
111 {
112         struct nfs_page *req = NULL;
113
114         if (PagePrivate(page)) {
115                 req = (struct nfs_page *)page_private(page);
116                 if (req != NULL)
117                         kref_get(&req->wb_kref);
118         }
119         return req;
120 }
121
122 static struct nfs_page *nfs_page_find_request(struct page *page)
123 {
124         struct inode *inode = page->mapping->host;
125         struct nfs_page *req = NULL;
126
127         spin_lock(&inode->i_lock);
128         req = nfs_page_find_request_locked(page);
129         spin_unlock(&inode->i_lock);
130         return req;
131 }
132
133 /* Adjust the file length if we're writing beyond the end */
134 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
135 {
136         struct inode *inode = page->mapping->host;
137         loff_t end, i_size;
138         pgoff_t end_index;
139
140         spin_lock(&inode->i_lock);
141         i_size = i_size_read(inode);
142         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
143         if (i_size > 0 && page->index < end_index)
144                 goto out;
145         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
146         if (i_size >= end)
147                 goto out;
148         i_size_write(inode, end);
149         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
150 out:
151         spin_unlock(&inode->i_lock);
152 }
153
154 /* A writeback failed: mark the page as bad, and invalidate the page cache */
155 static void nfs_set_pageerror(struct page *page)
156 {
157         SetPageError(page);
158         nfs_zap_mapping(page->mapping->host, page->mapping);
159 }
160
161 /* We can set the PG_uptodate flag if we see that a write request
162  * covers the full page.
163  */
164 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
165 {
166         if (PageUptodate(page))
167                 return;
168         if (base != 0)
169                 return;
170         if (count != nfs_page_length(page))
171                 return;
172         SetPageUptodate(page);
173 }
174
175 static int wb_priority(struct writeback_control *wbc)
176 {
177         if (wbc->for_reclaim)
178                 return FLUSH_HIGHPRI | FLUSH_STABLE;
179         if (wbc->for_kupdate || wbc->for_background)
180                 return FLUSH_LOWPRI;
181         return 0;
182 }
183
184 /*
185  * NFS congestion control
186  */
187
188 int nfs_congestion_kb;
189
190 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
191 #define NFS_CONGESTION_OFF_THRESH       \
192         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
193
194 static int nfs_set_page_writeback(struct page *page)
195 {
196         int ret = test_set_page_writeback(page);
197
198         if (!ret) {
199                 struct inode *inode = page->mapping->host;
200                 struct nfs_server *nfss = NFS_SERVER(inode);
201
202                 page_cache_get(page);
203                 if (atomic_long_inc_return(&nfss->writeback) >
204                                 NFS_CONGESTION_ON_THRESH) {
205                         set_bdi_congested(&nfss->backing_dev_info,
206                                                 BLK_RW_ASYNC);
207                 }
208         }
209         return ret;
210 }
211
212 static void nfs_end_page_writeback(struct page *page)
213 {
214         struct inode *inode = page->mapping->host;
215         struct nfs_server *nfss = NFS_SERVER(inode);
216
217         end_page_writeback(page);
218         page_cache_release(page);
219         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
220                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
221 }
222
223 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
224 {
225         struct inode *inode = page->mapping->host;
226         struct nfs_page *req;
227         int ret;
228
229         spin_lock(&inode->i_lock);
230         for (;;) {
231                 req = nfs_page_find_request_locked(page);
232                 if (req == NULL)
233                         break;
234                 if (nfs_set_page_tag_locked(req))
235                         break;
236                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
237                  *       then the call to nfs_set_page_tag_locked() will always
238                  *       succeed provided that someone hasn't already marked the
239                  *       request as dirty (in which case we don't care).
240                  */
241                 spin_unlock(&inode->i_lock);
242                 if (!nonblock)
243                         ret = nfs_wait_on_request(req);
244                 else
245                         ret = -EAGAIN;
246                 nfs_release_request(req);
247                 if (ret != 0)
248                         return ERR_PTR(ret);
249                 spin_lock(&inode->i_lock);
250         }
251         spin_unlock(&inode->i_lock);
252         return req;
253 }
254
255 /*
256  * Find an associated nfs write request, and prepare to flush it out
257  * May return an error if the user signalled nfs_wait_on_request().
258  */
259 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
260                                 struct page *page, bool nonblock)
261 {
262         struct nfs_page *req;
263         int ret = 0;
264
265         req = nfs_find_and_lock_request(page, nonblock);
266         if (!req)
267                 goto out;
268         ret = PTR_ERR(req);
269         if (IS_ERR(req))
270                 goto out;
271
272         ret = nfs_set_page_writeback(page);
273         BUG_ON(ret != 0);
274         BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
275
276         if (!nfs_pageio_add_request(pgio, req)) {
277                 nfs_redirty_request(req);
278                 ret = pgio->pg_error;
279         }
280 out:
281         return ret;
282 }
283
284 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
285 {
286         struct inode *inode = page->mapping->host;
287         int ret;
288
289         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
290         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
291
292         nfs_pageio_cond_complete(pgio, page->index);
293         ret = nfs_page_async_flush(pgio, page,
294                         wbc->sync_mode == WB_SYNC_NONE ||
295                         wbc->nonblocking != 0);
296         if (ret == -EAGAIN) {
297                 redirty_page_for_writepage(wbc, page);
298                 ret = 0;
299         }
300         return ret;
301 }
302
303 /*
304  * Write an mmapped page to the server.
305  */
306 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
307 {
308         struct nfs_pageio_descriptor pgio;
309         int err;
310
311         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
312         err = nfs_do_writepage(page, wbc, &pgio);
313         nfs_pageio_complete(&pgio);
314         if (err < 0)
315                 return err;
316         if (pgio.pg_error < 0)
317                 return pgio.pg_error;
318         return 0;
319 }
320
321 int nfs_writepage(struct page *page, struct writeback_control *wbc)
322 {
323         int ret;
324
325         ret = nfs_writepage_locked(page, wbc);
326         unlock_page(page);
327         return ret;
328 }
329
330 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
331 {
332         int ret;
333
334         ret = nfs_do_writepage(page, wbc, data);
335         unlock_page(page);
336         return ret;
337 }
338
339 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
340 {
341         struct inode *inode = mapping->host;
342         unsigned long *bitlock = &NFS_I(inode)->flags;
343         struct nfs_pageio_descriptor pgio;
344         int err;
345
346         /* Stop dirtying of new pages while we sync */
347         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
348                         nfs_wait_bit_killable, TASK_KILLABLE);
349         if (err)
350                 goto out_err;
351
352         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
353
354         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
355         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
356         nfs_pageio_complete(&pgio);
357
358         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
359         smp_mb__after_clear_bit();
360         wake_up_bit(bitlock, NFS_INO_FLUSHING);
361
362         if (err < 0)
363                 goto out_err;
364         err = pgio.pg_error;
365         if (err < 0)
366                 goto out_err;
367         return 0;
368 out_err:
369         return err;
370 }
371
372 /*
373  * Insert a write request into an inode
374  */
375 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
376 {
377         struct nfs_inode *nfsi = NFS_I(inode);
378         int error;
379
380         error = radix_tree_preload(GFP_NOFS);
381         if (error != 0)
382                 goto out;
383
384         /* Lock the request! */
385         nfs_lock_request_dontget(req);
386
387         spin_lock(&inode->i_lock);
388         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
389         BUG_ON(error);
390         if (!nfsi->npages) {
391                 igrab(inode);
392                 if (nfs_have_delegation(inode, FMODE_WRITE))
393                         nfsi->change_attr++;
394         }
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         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
421         nfsi->npages--;
422         if (!nfsi->npages) {
423                 spin_unlock(&inode->i_lock);
424                 iput(inode);
425         } else
426                 spin_unlock(&inode->i_lock);
427         nfs_clear_request(req);
428         nfs_release_request(req);
429 }
430
431 static void
432 nfs_mark_request_dirty(struct nfs_page *req)
433 {
434         __set_page_dirty_nobuffers(req->wb_page);
435         __mark_inode_dirty(req->wb_page->mapping->host, I_DIRTY_DATASYNC);
436 }
437
438 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
439 /*
440  * Add a request to the inode's commit list.
441  */
442 static void
443 nfs_mark_request_commit(struct nfs_page *req)
444 {
445         struct inode *inode = req->wb_context->path.dentry->d_inode;
446         struct nfs_inode *nfsi = NFS_I(inode);
447
448         spin_lock(&inode->i_lock);
449         set_bit(PG_CLEAN, &(req)->wb_flags);
450         radix_tree_tag_set(&nfsi->nfs_page_tree,
451                         req->wb_index,
452                         NFS_PAGE_TAG_COMMIT);
453         nfsi->ncommit++;
454         spin_unlock(&inode->i_lock);
455         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
456         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
457         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
458 }
459
460 static int
461 nfs_clear_request_commit(struct nfs_page *req)
462 {
463         struct page *page = req->wb_page;
464
465         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
466                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
467                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
468                 return 1;
469         }
470         return 0;
471 }
472
473 static inline
474 int nfs_write_need_commit(struct nfs_write_data *data)
475 {
476         return data->verf.committed != NFS_FILE_SYNC;
477 }
478
479 static inline
480 int nfs_reschedule_unstable_write(struct nfs_page *req)
481 {
482         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
483                 nfs_mark_request_commit(req);
484                 return 1;
485         }
486         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
487                 nfs_mark_request_dirty(req);
488                 return 1;
489         }
490         return 0;
491 }
492 #else
493 static inline void
494 nfs_mark_request_commit(struct nfs_page *req)
495 {
496 }
497
498 static inline int
499 nfs_clear_request_commit(struct nfs_page *req)
500 {
501         return 0;
502 }
503
504 static inline
505 int nfs_write_need_commit(struct nfs_write_data *data)
506 {
507         return 0;
508 }
509
510 static inline
511 int nfs_reschedule_unstable_write(struct nfs_page *req)
512 {
513         return 0;
514 }
515 #endif
516
517 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
518 static int
519 nfs_need_commit(struct nfs_inode *nfsi)
520 {
521         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
522 }
523
524 /*
525  * nfs_scan_commit - Scan an inode for commit requests
526  * @inode: NFS inode to scan
527  * @dst: destination list
528  * @idx_start: lower bound of page->index to scan.
529  * @npages: idx_start + npages sets the upper bound to scan.
530  *
531  * Moves requests from the inode's 'commit' request list.
532  * The requests are *not* checked to ensure that they form a contiguous set.
533  */
534 static int
535 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
536 {
537         struct nfs_inode *nfsi = NFS_I(inode);
538         int ret;
539
540         if (!nfs_need_commit(nfsi))
541                 return 0;
542
543         ret = nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
544         if (ret > 0)
545                 nfsi->ncommit -= ret;
546         if (nfs_need_commit(NFS_I(inode)))
547                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
548         return ret;
549 }
550 #else
551 static inline int nfs_need_commit(struct nfs_inode *nfsi)
552 {
553         return 0;
554 }
555
556 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
557 {
558         return 0;
559 }
560 #endif
561
562 /*
563  * Search for an existing write request, and attempt to update
564  * it to reflect a new dirty region on a given page.
565  *
566  * If the attempt fails, then the existing request is flushed out
567  * to disk.
568  */
569 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
570                 struct page *page,
571                 unsigned int offset,
572                 unsigned int bytes)
573 {
574         struct nfs_page *req;
575         unsigned int rqend;
576         unsigned int end;
577         int error;
578
579         if (!PagePrivate(page))
580                 return NULL;
581
582         end = offset + bytes;
583         spin_lock(&inode->i_lock);
584
585         for (;;) {
586                 req = nfs_page_find_request_locked(page);
587                 if (req == NULL)
588                         goto out_unlock;
589
590                 rqend = req->wb_offset + req->wb_bytes;
591                 /*
592                  * Tell the caller to flush out the request if
593                  * the offsets are non-contiguous.
594                  * Note: nfs_flush_incompatible() will already
595                  * have flushed out requests having wrong owners.
596                  */
597                 if (offset > rqend
598                     || end < req->wb_offset)
599                         goto out_flushme;
600
601                 if (nfs_set_page_tag_locked(req))
602                         break;
603
604                 /* The request is locked, so wait and then retry */
605                 spin_unlock(&inode->i_lock);
606                 error = nfs_wait_on_request(req);
607                 nfs_release_request(req);
608                 if (error != 0)
609                         goto out_err;
610                 spin_lock(&inode->i_lock);
611         }
612
613         if (nfs_clear_request_commit(req) &&
614                         radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
615                                 req->wb_index, NFS_PAGE_TAG_COMMIT) != NULL)
616                 NFS_I(inode)->ncommit--;
617
618         /* Okay, the request matches. Update the region */
619         if (offset < req->wb_offset) {
620                 req->wb_offset = offset;
621                 req->wb_pgbase = offset;
622         }
623         if (end > rqend)
624                 req->wb_bytes = end - req->wb_offset;
625         else
626                 req->wb_bytes = rqend - req->wb_offset;
627 out_unlock:
628         spin_unlock(&inode->i_lock);
629         return req;
630 out_flushme:
631         spin_unlock(&inode->i_lock);
632         nfs_release_request(req);
633         error = nfs_wb_page(inode, page);
634 out_err:
635         return ERR_PTR(error);
636 }
637
638 /*
639  * Try to update an existing write request, or create one if there is none.
640  *
641  * Note: Should always be called with the Page Lock held to prevent races
642  * if we have to add a new request. Also assumes that the caller has
643  * already called nfs_flush_incompatible() if necessary.
644  */
645 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
646                 struct page *page, unsigned int offset, unsigned int bytes)
647 {
648         struct inode *inode = page->mapping->host;
649         struct nfs_page *req;
650         int error;
651
652         req = nfs_try_to_update_request(inode, page, offset, bytes);
653         if (req != NULL)
654                 goto out;
655         req = nfs_create_request(ctx, inode, page, offset, bytes);
656         if (IS_ERR(req))
657                 goto out;
658         error = nfs_inode_add_request(inode, req);
659         if (error != 0) {
660                 nfs_release_request(req);
661                 req = ERR_PTR(error);
662         }
663 out:
664         return req;
665 }
666
667 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
668                 unsigned int offset, unsigned int count)
669 {
670         struct nfs_page *req;
671
672         req = nfs_setup_write_request(ctx, page, offset, count);
673         if (IS_ERR(req))
674                 return PTR_ERR(req);
675         nfs_mark_request_dirty(req);
676         /* Update file length */
677         nfs_grow_file(page, offset, count);
678         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
679         nfs_mark_request_dirty(req);
680         nfs_clear_page_tag_locked(req);
681         return 0;
682 }
683
684 int nfs_flush_incompatible(struct file *file, struct page *page)
685 {
686         struct nfs_open_context *ctx = nfs_file_open_context(file);
687         struct nfs_page *req;
688         int do_flush, status;
689         /*
690          * Look for a request corresponding to this page. If there
691          * is one, and it belongs to another file, we flush it out
692          * before we try to copy anything into the page. Do this
693          * due to the lack of an ACCESS-type call in NFSv2.
694          * Also do the same if we find a request from an existing
695          * dropped page.
696          */
697         do {
698                 req = nfs_page_find_request(page);
699                 if (req == NULL)
700                         return 0;
701                 do_flush = req->wb_page != page || req->wb_context != ctx ||
702                         req->wb_lock_context->lockowner != current->files ||
703                         req->wb_lock_context->pid != current->tgid;
704                 nfs_release_request(req);
705                 if (!do_flush)
706                         return 0;
707                 status = nfs_wb_page(page->mapping->host, page);
708         } while (status == 0);
709         return status;
710 }
711
712 /*
713  * If the page cache is marked as unsafe or invalid, then we can't rely on
714  * the PageUptodate() flag. In this case, we will need to turn off
715  * write optimisations that depend on the page contents being correct.
716  */
717 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
718 {
719         return PageUptodate(page) &&
720                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
721 }
722
723 /*
724  * Update and possibly write a cached page of an NFS file.
725  *
726  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
727  * things with a page scheduled for an RPC call (e.g. invalidate it).
728  */
729 int nfs_updatepage(struct file *file, struct page *page,
730                 unsigned int offset, unsigned int count)
731 {
732         struct nfs_open_context *ctx = nfs_file_open_context(file);
733         struct inode    *inode = page->mapping->host;
734         int             status = 0;
735
736         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
737
738         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
739                 file->f_path.dentry->d_parent->d_name.name,
740                 file->f_path.dentry->d_name.name, count,
741                 (long long)(page_offset(page) + offset));
742
743         /* If we're not using byte range locks, and we know the page
744          * is up to date, it may be more efficient to extend the write
745          * to cover the entire page in order to avoid fragmentation
746          * inefficiencies.
747          */
748         if (nfs_write_pageuptodate(page, inode) &&
749                         inode->i_flock == NULL &&
750                         !(file->f_flags & O_DSYNC)) {
751                 count = max(count + offset, nfs_page_length(page));
752                 offset = 0;
753         }
754
755         status = nfs_writepage_setup(ctx, page, offset, count);
756         if (status < 0)
757                 nfs_set_pageerror(page);
758
759         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
760                         status, (long long)i_size_read(inode));
761         return status;
762 }
763
764 static void nfs_writepage_release(struct nfs_page *req)
765 {
766         struct page *page = req->wb_page;
767
768         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req))
769                 nfs_inode_remove_request(req);
770         nfs_clear_page_tag_locked(req);
771         nfs_end_page_writeback(page);
772 }
773
774 static int flush_task_priority(int how)
775 {
776         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
777                 case FLUSH_HIGHPRI:
778                         return RPC_PRIORITY_HIGH;
779                 case FLUSH_LOWPRI:
780                         return RPC_PRIORITY_LOW;
781         }
782         return RPC_PRIORITY_NORMAL;
783 }
784
785 /*
786  * Set up the argument/result storage required for the RPC call.
787  */
788 static int nfs_write_rpcsetup(struct nfs_page *req,
789                 struct nfs_write_data *data,
790                 const struct rpc_call_ops *call_ops,
791                 unsigned int count, unsigned int offset,
792                 int how)
793 {
794         struct inode *inode = req->wb_context->path.dentry->d_inode;
795         int priority = flush_task_priority(how);
796         struct rpc_task *task;
797         struct rpc_message msg = {
798                 .rpc_argp = &data->args,
799                 .rpc_resp = &data->res,
800                 .rpc_cred = req->wb_context->cred,
801         };
802         struct rpc_task_setup task_setup_data = {
803                 .rpc_client = NFS_CLIENT(inode),
804                 .task = &data->task,
805                 .rpc_message = &msg,
806                 .callback_ops = call_ops,
807                 .callback_data = data,
808                 .workqueue = nfsiod_workqueue,
809                 .flags = RPC_TASK_ASYNC,
810                 .priority = priority,
811         };
812         int ret = 0;
813
814         /* Set up the RPC argument and reply structs
815          * NB: take care not to mess about with data->commit et al. */
816
817         data->req = req;
818         data->inode = inode = req->wb_context->path.dentry->d_inode;
819         data->cred = msg.rpc_cred;
820
821         data->args.fh     = NFS_FH(inode);
822         data->args.offset = req_offset(req) + offset;
823         data->args.pgbase = req->wb_pgbase + offset;
824         data->args.pages  = data->pagevec;
825         data->args.count  = count;
826         data->args.context = get_nfs_open_context(req->wb_context);
827         data->args.lock_context = req->wb_lock_context;
828         data->args.stable  = NFS_UNSTABLE;
829         if (how & FLUSH_STABLE) {
830                 data->args.stable = NFS_DATA_SYNC;
831                 if (!nfs_need_commit(NFS_I(inode)))
832                         data->args.stable = NFS_FILE_SYNC;
833         }
834
835         data->res.fattr   = &data->fattr;
836         data->res.count   = count;
837         data->res.verf    = &data->verf;
838         nfs_fattr_init(&data->fattr);
839
840         /* Set up the initial task struct.  */
841         NFS_PROTO(inode)->write_setup(data, &msg);
842
843         dprintk("NFS: %5u initiated write call "
844                 "(req %s/%lld, %u bytes @ offset %llu)\n",
845                 data->task.tk_pid,
846                 inode->i_sb->s_id,
847                 (long long)NFS_FILEID(inode),
848                 count,
849                 (unsigned long long)data->args.offset);
850
851         task = rpc_run_task(&task_setup_data);
852         if (IS_ERR(task)) {
853                 ret = PTR_ERR(task);
854                 goto out;
855         }
856         if (how & FLUSH_SYNC) {
857                 ret = rpc_wait_for_completion_task(task);
858                 if (ret == 0)
859                         ret = task->tk_status;
860         }
861         rpc_put_task(task);
862 out:
863         return ret;
864 }
865
866 /* If a nfs_flush_* function fails, it should remove reqs from @head and
867  * call this on each, which will prepare them to be retried on next
868  * writeback using standard nfs.
869  */
870 static void nfs_redirty_request(struct nfs_page *req)
871 {
872         struct page *page = req->wb_page;
873
874         nfs_mark_request_dirty(req);
875         nfs_clear_page_tag_locked(req);
876         nfs_end_page_writeback(page);
877 }
878
879 /*
880  * Generate multiple small requests to write out a single
881  * contiguous dirty area on one page.
882  */
883 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
884 {
885         struct nfs_page *req = nfs_list_entry(head->next);
886         struct page *page = req->wb_page;
887         struct nfs_write_data *data;
888         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
889         unsigned int offset;
890         int requests = 0;
891         int ret = 0;
892         LIST_HEAD(list);
893
894         nfs_list_remove_request(req);
895
896         nbytes = count;
897         do {
898                 size_t len = min(nbytes, wsize);
899
900                 data = nfs_writedata_alloc(1);
901                 if (!data)
902                         goto out_bad;
903                 list_add(&data->pages, &list);
904                 requests++;
905                 nbytes -= len;
906         } while (nbytes != 0);
907         atomic_set(&req->wb_complete, requests);
908
909         ClearPageError(page);
910         offset = 0;
911         nbytes = count;
912         do {
913                 int ret2;
914
915                 data = list_entry(list.next, struct nfs_write_data, pages);
916                 list_del_init(&data->pages);
917
918                 data->pagevec[0] = page;
919
920                 if (nbytes < wsize)
921                         wsize = nbytes;
922                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
923                                    wsize, offset, how);
924                 if (ret == 0)
925                         ret = ret2;
926                 offset += wsize;
927                 nbytes -= wsize;
928         } while (nbytes != 0);
929
930         return ret;
931
932 out_bad:
933         while (!list_empty(&list)) {
934                 data = list_entry(list.next, struct nfs_write_data, pages);
935                 list_del(&data->pages);
936                 nfs_writedata_release(data);
937         }
938         nfs_redirty_request(req);
939         return -ENOMEM;
940 }
941
942 /*
943  * Create an RPC task for the given write request and kick it.
944  * The page must have been locked by the caller.
945  *
946  * It may happen that the page we're passed is not marked dirty.
947  * This is the case if nfs_updatepage detects a conflicting request
948  * that has been written but not committed.
949  */
950 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
951 {
952         struct nfs_page         *req;
953         struct page             **pages;
954         struct nfs_write_data   *data;
955
956         data = nfs_writedata_alloc(npages);
957         if (!data)
958                 goto out_bad;
959
960         pages = data->pagevec;
961         while (!list_empty(head)) {
962                 req = nfs_list_entry(head->next);
963                 nfs_list_remove_request(req);
964                 nfs_list_add_request(req, &data->pages);
965                 ClearPageError(req->wb_page);
966                 *pages++ = req->wb_page;
967         }
968         req = nfs_list_entry(data->pages.next);
969
970         /* Set up the argument struct */
971         return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
972  out_bad:
973         while (!list_empty(head)) {
974                 req = nfs_list_entry(head->next);
975                 nfs_list_remove_request(req);
976                 nfs_redirty_request(req);
977         }
978         return -ENOMEM;
979 }
980
981 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
982                                   struct inode *inode, int ioflags)
983 {
984         size_t wsize = NFS_SERVER(inode)->wsize;
985
986         if (wsize < PAGE_CACHE_SIZE)
987                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
988         else
989                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
990 }
991
992 /*
993  * Handle a write reply that flushed part of a page.
994  */
995 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
996 {
997         struct nfs_write_data   *data = calldata;
998
999         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1000                 task->tk_pid,
1001                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1002                 (long long)
1003                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1004                 data->req->wb_bytes, (long long)req_offset(data->req));
1005
1006         nfs_writeback_done(task, data);
1007 }
1008
1009 static void nfs_writeback_release_partial(void *calldata)
1010 {
1011         struct nfs_write_data   *data = calldata;
1012         struct nfs_page         *req = data->req;
1013         struct page             *page = req->wb_page;
1014         int status = data->task.tk_status;
1015
1016         if (status < 0) {
1017                 nfs_set_pageerror(page);
1018                 nfs_context_set_write_error(req->wb_context, status);
1019                 dprintk(", error = %d\n", status);
1020                 goto out;
1021         }
1022
1023         if (nfs_write_need_commit(data)) {
1024                 struct inode *inode = page->mapping->host;
1025
1026                 spin_lock(&inode->i_lock);
1027                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1028                         /* Do nothing we need to resend the writes */
1029                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1030                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1031                         dprintk(" defer commit\n");
1032                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1033                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1034                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1035                         dprintk(" server reboot detected\n");
1036                 }
1037                 spin_unlock(&inode->i_lock);
1038         } else
1039                 dprintk(" OK\n");
1040
1041 out:
1042         if (atomic_dec_and_test(&req->wb_complete))
1043                 nfs_writepage_release(req);
1044         nfs_writedata_release(calldata);
1045 }
1046
1047 #if defined(CONFIG_NFS_V4_1)
1048 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1049 {
1050         struct nfs_write_data *data = calldata;
1051
1052         if (nfs4_setup_sequence(NFS_SERVER(data->inode),
1053                                 &data->args.seq_args,
1054                                 &data->res.seq_res, 1, task))
1055                 return;
1056         rpc_call_start(task);
1057 }
1058 #endif /* CONFIG_NFS_V4_1 */
1059
1060 static const struct rpc_call_ops nfs_write_partial_ops = {
1061 #if defined(CONFIG_NFS_V4_1)
1062         .rpc_call_prepare = nfs_write_prepare,
1063 #endif /* CONFIG_NFS_V4_1 */
1064         .rpc_call_done = nfs_writeback_done_partial,
1065         .rpc_release = nfs_writeback_release_partial,
1066 };
1067
1068 /*
1069  * Handle a write reply that flushes a whole page.
1070  *
1071  * FIXME: There is an inherent race with invalidate_inode_pages and
1072  *        writebacks since the page->count is kept > 1 for as long
1073  *        as the page has a write request pending.
1074  */
1075 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1076 {
1077         struct nfs_write_data   *data = calldata;
1078
1079         nfs_writeback_done(task, data);
1080 }
1081
1082 static void nfs_writeback_release_full(void *calldata)
1083 {
1084         struct nfs_write_data   *data = calldata;
1085         int status = data->task.tk_status;
1086
1087         /* Update attributes as result of writeback. */
1088         while (!list_empty(&data->pages)) {
1089                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1090                 struct page *page = req->wb_page;
1091
1092                 nfs_list_remove_request(req);
1093
1094                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1095                         data->task.tk_pid,
1096                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1097                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1098                         req->wb_bytes,
1099                         (long long)req_offset(req));
1100
1101                 if (status < 0) {
1102                         nfs_set_pageerror(page);
1103                         nfs_context_set_write_error(req->wb_context, status);
1104                         dprintk(", error = %d\n", status);
1105                         goto remove_request;
1106                 }
1107
1108                 if (nfs_write_need_commit(data)) {
1109                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1110                         nfs_mark_request_commit(req);
1111                         dprintk(" marked for commit\n");
1112                         goto next;
1113                 }
1114                 dprintk(" OK\n");
1115 remove_request:
1116                 nfs_inode_remove_request(req);
1117         next:
1118                 nfs_clear_page_tag_locked(req);
1119                 nfs_end_page_writeback(page);
1120         }
1121         nfs_writedata_release(calldata);
1122 }
1123
1124 static const struct rpc_call_ops nfs_write_full_ops = {
1125 #if defined(CONFIG_NFS_V4_1)
1126         .rpc_call_prepare = nfs_write_prepare,
1127 #endif /* CONFIG_NFS_V4_1 */
1128         .rpc_call_done = nfs_writeback_done_full,
1129         .rpc_release = nfs_writeback_release_full,
1130 };
1131
1132
1133 /*
1134  * This function is called when the WRITE call is complete.
1135  */
1136 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1137 {
1138         struct nfs_writeargs    *argp = &data->args;
1139         struct nfs_writeres     *resp = &data->res;
1140         struct nfs_server       *server = NFS_SERVER(data->inode);
1141         int status;
1142
1143         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1144                 task->tk_pid, task->tk_status);
1145
1146         /*
1147          * ->write_done will attempt to use post-op attributes to detect
1148          * conflicting writes by other clients.  A strict interpretation
1149          * of close-to-open would allow us to continue caching even if
1150          * another writer had changed the file, but some applications
1151          * depend on tighter cache coherency when writing.
1152          */
1153         status = NFS_PROTO(data->inode)->write_done(task, data);
1154         if (status != 0)
1155                 return status;
1156         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1157
1158 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1159         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1160                 /* We tried a write call, but the server did not
1161                  * commit data to stable storage even though we
1162                  * requested it.
1163                  * Note: There is a known bug in Tru64 < 5.0 in which
1164                  *       the server reports NFS_DATA_SYNC, but performs
1165                  *       NFS_FILE_SYNC. We therefore implement this checking
1166                  *       as a dprintk() in order to avoid filling syslog.
1167                  */
1168                 static unsigned long    complain;
1169
1170                 if (time_before(complain, jiffies)) {
1171                         dprintk("NFS:       faulty NFS server %s:"
1172                                 " (committed = %d) != (stable = %d)\n",
1173                                 server->nfs_client->cl_hostname,
1174                                 resp->verf->committed, argp->stable);
1175                         complain = jiffies + 300 * HZ;
1176                 }
1177         }
1178 #endif
1179         /* Is this a short write? */
1180         if (task->tk_status >= 0 && resp->count < argp->count) {
1181                 static unsigned long    complain;
1182
1183                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1184
1185                 /* Has the server at least made some progress? */
1186                 if (resp->count != 0) {
1187                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1188                         if (resp->verf->committed != NFS_UNSTABLE) {
1189                                 /* Resend from where the server left off */
1190                                 argp->offset += resp->count;
1191                                 argp->pgbase += resp->count;
1192                                 argp->count -= resp->count;
1193                         } else {
1194                                 /* Resend as a stable write in order to avoid
1195                                  * headaches in the case of a server crash.
1196                                  */
1197                                 argp->stable = NFS_FILE_SYNC;
1198                         }
1199                         nfs_restart_rpc(task, server->nfs_client);
1200                         return -EAGAIN;
1201                 }
1202                 if (time_before(complain, jiffies)) {
1203                         printk(KERN_WARNING
1204                                "NFS: Server wrote zero bytes, expected %u.\n",
1205                                         argp->count);
1206                         complain = jiffies + 300 * HZ;
1207                 }
1208                 /* Can't do anything about it except throw an error. */
1209                 task->tk_status = -EIO;
1210         }
1211         return 0;
1212 }
1213
1214
1215 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1216 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1217 {
1218         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1219                 return 1;
1220         if (may_wait && !out_of_line_wait_on_bit_lock(&nfsi->flags,
1221                                 NFS_INO_COMMIT, nfs_wait_bit_killable,
1222                                 TASK_KILLABLE))
1223                 return 1;
1224         return 0;
1225 }
1226
1227 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1228 {
1229         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1230         smp_mb__after_clear_bit();
1231         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1232 }
1233
1234
1235 static void nfs_commitdata_release(void *data)
1236 {
1237         struct nfs_write_data *wdata = data;
1238
1239         put_nfs_open_context(wdata->args.context);
1240         nfs_commit_free(wdata);
1241 }
1242
1243 /*
1244  * Set up the argument/result storage required for the RPC call.
1245  */
1246 static int nfs_commit_rpcsetup(struct list_head *head,
1247                 struct nfs_write_data *data,
1248                 int how)
1249 {
1250         struct nfs_page *first = nfs_list_entry(head->next);
1251         struct inode *inode = first->wb_context->path.dentry->d_inode;
1252         int priority = flush_task_priority(how);
1253         struct rpc_task *task;
1254         struct rpc_message msg = {
1255                 .rpc_argp = &data->args,
1256                 .rpc_resp = &data->res,
1257                 .rpc_cred = first->wb_context->cred,
1258         };
1259         struct rpc_task_setup task_setup_data = {
1260                 .task = &data->task,
1261                 .rpc_client = NFS_CLIENT(inode),
1262                 .rpc_message = &msg,
1263                 .callback_ops = &nfs_commit_ops,
1264                 .callback_data = data,
1265                 .workqueue = nfsiod_workqueue,
1266                 .flags = RPC_TASK_ASYNC,
1267                 .priority = priority,
1268         };
1269
1270         /* Set up the RPC argument and reply structs
1271          * NB: take care not to mess about with data->commit et al. */
1272
1273         list_splice_init(head, &data->pages);
1274
1275         data->inode       = inode;
1276         data->cred        = msg.rpc_cred;
1277
1278         data->args.fh     = NFS_FH(data->inode);
1279         /* Note: we always request a commit of the entire inode */
1280         data->args.offset = 0;
1281         data->args.count  = 0;
1282         data->args.context = get_nfs_open_context(first->wb_context);
1283         data->res.count   = 0;
1284         data->res.fattr   = &data->fattr;
1285         data->res.verf    = &data->verf;
1286         nfs_fattr_init(&data->fattr);
1287
1288         /* Set up the initial task struct.  */
1289         NFS_PROTO(inode)->commit_setup(data, &msg);
1290
1291         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1292
1293         task = rpc_run_task(&task_setup_data);
1294         if (IS_ERR(task))
1295                 return PTR_ERR(task);
1296         rpc_put_task(task);
1297         return 0;
1298 }
1299
1300 /*
1301  * Commit dirty pages
1302  */
1303 static int
1304 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1305 {
1306         struct nfs_write_data   *data;
1307         struct nfs_page         *req;
1308
1309         data = nfs_commitdata_alloc();
1310
1311         if (!data)
1312                 goto out_bad;
1313
1314         /* Set up the argument struct */
1315         return nfs_commit_rpcsetup(head, data, how);
1316  out_bad:
1317         while (!list_empty(head)) {
1318                 req = nfs_list_entry(head->next);
1319                 nfs_list_remove_request(req);
1320                 nfs_mark_request_commit(req);
1321                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1322                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1323                                 BDI_RECLAIMABLE);
1324                 nfs_clear_page_tag_locked(req);
1325         }
1326         nfs_commit_clear_lock(NFS_I(inode));
1327         return -ENOMEM;
1328 }
1329
1330 /*
1331  * COMMIT call returned
1332  */
1333 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1334 {
1335         struct nfs_write_data   *data = calldata;
1336
1337         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1338                                 task->tk_pid, task->tk_status);
1339
1340         /* Call the NFS version-specific code */
1341         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1342                 return;
1343 }
1344
1345 static void nfs_commit_release(void *calldata)
1346 {
1347         struct nfs_write_data   *data = calldata;
1348         struct nfs_page         *req;
1349         int status = data->task.tk_status;
1350
1351         while (!list_empty(&data->pages)) {
1352                 req = nfs_list_entry(data->pages.next);
1353                 nfs_list_remove_request(req);
1354                 nfs_clear_request_commit(req);
1355
1356                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1357                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1358                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1359                         req->wb_bytes,
1360                         (long long)req_offset(req));
1361                 if (status < 0) {
1362                         nfs_context_set_write_error(req->wb_context, status);
1363                         nfs_inode_remove_request(req);
1364                         dprintk(", error = %d\n", status);
1365                         goto next;
1366                 }
1367
1368                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1369                  * returned by the server against all stored verfs. */
1370                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1371                         /* We have a match */
1372                         nfs_inode_remove_request(req);
1373                         dprintk(" OK\n");
1374                         goto next;
1375                 }
1376                 /* We have a mismatch. Write the page again */
1377                 dprintk(" mismatch\n");
1378                 nfs_mark_request_dirty(req);
1379         next:
1380                 nfs_clear_page_tag_locked(req);
1381         }
1382         nfs_commit_clear_lock(NFS_I(data->inode));
1383         nfs_commitdata_release(calldata);
1384 }
1385
1386 static const struct rpc_call_ops nfs_commit_ops = {
1387 #if defined(CONFIG_NFS_V4_1)
1388         .rpc_call_prepare = nfs_write_prepare,
1389 #endif /* CONFIG_NFS_V4_1 */
1390         .rpc_call_done = nfs_commit_done,
1391         .rpc_release = nfs_commit_release,
1392 };
1393
1394 int nfs_commit_inode(struct inode *inode, int how)
1395 {
1396         LIST_HEAD(head);
1397         int may_wait = how & FLUSH_SYNC;
1398         int res = 0;
1399
1400         if (!nfs_commit_set_lock(NFS_I(inode), may_wait))
1401                 goto out_mark_dirty;
1402         spin_lock(&inode->i_lock);
1403         res = nfs_scan_commit(inode, &head, 0, 0);
1404         spin_unlock(&inode->i_lock);
1405         if (res) {
1406                 int error = nfs_commit_list(inode, &head, how);
1407                 if (error < 0)
1408                         return error;
1409                 if (may_wait)
1410                         wait_on_bit(&NFS_I(inode)->flags, NFS_INO_COMMIT,
1411                                         nfs_wait_bit_killable,
1412                                         TASK_KILLABLE);
1413                 else
1414                         goto out_mark_dirty;
1415         } else
1416                 nfs_commit_clear_lock(NFS_I(inode));
1417         return res;
1418         /* Note: If we exit without ensuring that the commit is complete,
1419          * we must mark the inode as dirty. Otherwise, future calls to
1420          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1421          * that the data is on the disk.
1422          */
1423 out_mark_dirty:
1424         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1425         return res;
1426 }
1427
1428 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1429 {
1430         struct nfs_inode *nfsi = NFS_I(inode);
1431         int flags = FLUSH_SYNC;
1432         int ret = 0;
1433
1434         if (wbc->sync_mode == WB_SYNC_NONE) {
1435                 /* Don't commit yet if this is a non-blocking flush and there
1436                  * are a lot of outstanding writes for this mapping.
1437                  */
1438                 if (nfsi->ncommit <= (nfsi->npages >> 1))
1439                         goto out_mark_dirty;
1440
1441                 /* don't wait for the COMMIT response */
1442                 flags = 0;
1443         }
1444
1445         ret = nfs_commit_inode(inode, flags);
1446         if (ret >= 0) {
1447                 if (wbc->sync_mode == WB_SYNC_NONE) {
1448                         if (ret < wbc->nr_to_write)
1449                                 wbc->nr_to_write -= ret;
1450                         else
1451                                 wbc->nr_to_write = 0;
1452                 }
1453                 return 0;
1454         }
1455 out_mark_dirty:
1456         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1457         return ret;
1458 }
1459 #else
1460 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1461 {
1462         return 0;
1463 }
1464 #endif
1465
1466 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1467 {
1468         return nfs_commit_unstable_pages(inode, wbc);
1469 }
1470
1471 /*
1472  * flush the inode to disk.
1473  */
1474 int nfs_wb_all(struct inode *inode)
1475 {
1476         struct writeback_control wbc = {
1477                 .sync_mode = WB_SYNC_ALL,
1478                 .nr_to_write = LONG_MAX,
1479                 .range_start = 0,
1480                 .range_end = LLONG_MAX,
1481         };
1482
1483         return sync_inode(inode, &wbc);
1484 }
1485
1486 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1487 {
1488         struct nfs_page *req;
1489         int ret = 0;
1490
1491         BUG_ON(!PageLocked(page));
1492         for (;;) {
1493                 wait_on_page_writeback(page);
1494                 req = nfs_page_find_request(page);
1495                 if (req == NULL)
1496                         break;
1497                 if (nfs_lock_request_dontget(req)) {
1498                         nfs_inode_remove_request(req);
1499                         /*
1500                          * In case nfs_inode_remove_request has marked the
1501                          * page as being dirty
1502                          */
1503                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1504                         nfs_unlock_request(req);
1505                         break;
1506                 }
1507                 ret = nfs_wait_on_request(req);
1508                 nfs_release_request(req);
1509                 if (ret < 0)
1510                         break;
1511         }
1512         return ret;
1513 }
1514
1515 /*
1516  * Write back all requests on one page - we do this before reading it.
1517  */
1518 int nfs_wb_page(struct inode *inode, struct page *page)
1519 {
1520         loff_t range_start = page_offset(page);
1521         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1522         struct writeback_control wbc = {
1523                 .sync_mode = WB_SYNC_ALL,
1524                 .nr_to_write = 0,
1525                 .range_start = range_start,
1526                 .range_end = range_end,
1527         };
1528         int ret;
1529
1530         for (;;) {
1531                 wait_on_page_writeback(page);
1532                 if (clear_page_dirty_for_io(page)) {
1533                         ret = nfs_writepage_locked(page, &wbc);
1534                         if (ret < 0)
1535                                 goto out_error;
1536                         continue;
1537                 }
1538                 if (!PagePrivate(page))
1539                         break;
1540                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1541                 if (ret < 0)
1542                         goto out_error;
1543         }
1544         return 0;
1545 out_error:
1546         return ret;
1547 }
1548
1549 #ifdef CONFIG_MIGRATION
1550 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1551                 struct page *page)
1552 {
1553         struct nfs_page *req;
1554         int ret;
1555
1556         nfs_fscache_release_page(page, GFP_KERNEL);
1557
1558         req = nfs_find_and_lock_request(page, false);
1559         ret = PTR_ERR(req);
1560         if (IS_ERR(req))
1561                 goto out;
1562
1563         ret = migrate_page(mapping, newpage, page);
1564         if (!req)
1565                 goto out;
1566         if (ret)
1567                 goto out_unlock;
1568         page_cache_get(newpage);
1569         spin_lock(&mapping->host->i_lock);
1570         req->wb_page = newpage;
1571         SetPagePrivate(newpage);
1572         set_page_private(newpage, (unsigned long)req);
1573         ClearPagePrivate(page);
1574         set_page_private(page, 0);
1575         spin_unlock(&mapping->host->i_lock);
1576         page_cache_release(page);
1577 out_unlock:
1578         nfs_clear_page_tag_locked(req);
1579 out:
1580         return ret;
1581 }
1582 #endif
1583
1584 int __init nfs_init_writepagecache(void)
1585 {
1586         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1587                                              sizeof(struct nfs_write_data),
1588                                              0, SLAB_HWCACHE_ALIGN,
1589                                              NULL);
1590         if (nfs_wdata_cachep == NULL)
1591                 return -ENOMEM;
1592
1593         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1594                                                      nfs_wdata_cachep);
1595         if (nfs_wdata_mempool == NULL)
1596                 return -ENOMEM;
1597
1598         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1599                                                       nfs_wdata_cachep);
1600         if (nfs_commit_mempool == NULL)
1601                 return -ENOMEM;
1602
1603         /*
1604          * NFS congestion size, scale with available memory.
1605          *
1606          *  64MB:    8192k
1607          * 128MB:   11585k
1608          * 256MB:   16384k
1609          * 512MB:   23170k
1610          *   1GB:   32768k
1611          *   2GB:   46340k
1612          *   4GB:   65536k
1613          *   8GB:   92681k
1614          *  16GB:  131072k
1615          *
1616          * This allows larger machines to have larger/more transfers.
1617          * Limit the default to 256M
1618          */
1619         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1620         if (nfs_congestion_kb > 256*1024)
1621                 nfs_congestion_kb = 256*1024;
1622
1623         return 0;
1624 }
1625
1626 void nfs_destroy_writepagecache(void)
1627 {
1628         mempool_destroy(nfs_commit_mempool);
1629         mempool_destroy(nfs_wdata_mempool);
1630         kmem_cache_destroy(nfs_wdata_cachep);
1631 }
1632