NFS: Fix a list corruption problem
[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
17 #include <linux/sunrpc/clnt.h>
18 #include <linux/nfs_fs.h>
19 #include <linux/nfs_mount.h>
20 #include <linux/nfs_page.h>
21 #include <linux/backing-dev.h>
22
23 #include <asm/uaccess.h>
24 #include <linux/smp_lock.h>
25
26 #include "delegation.h"
27 #include "internal.h"
28 #include "iostat.h"
29
30 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
31
32 #define MIN_POOL_WRITE          (32)
33 #define MIN_POOL_COMMIT         (4)
34
35 /*
36  * Local function declarations
37  */
38 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
39                                             struct page *,
40                                             unsigned int, unsigned int);
41 static void nfs_mark_request_dirty(struct nfs_page *req);
42 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how);
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_commit_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_rcu_free(struct rcu_head *head)
63 {
64         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
65         if (p && (p->pagevec != &p->page_array[0]))
66                 kfree(p->pagevec);
67         mempool_free(p, nfs_commit_mempool);
68 }
69
70 void nfs_commit_free(struct nfs_write_data *wdata)
71 {
72         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
73 }
74
75 struct nfs_write_data *nfs_writedata_alloc(size_t len)
76 {
77         unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
78         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
79
80         if (p) {
81                 memset(p, 0, sizeof(*p));
82                 INIT_LIST_HEAD(&p->pages);
83                 p->npages = pagecount;
84                 if (pagecount <= ARRAY_SIZE(p->page_array))
85                         p->pagevec = p->page_array;
86                 else {
87                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
88                         if (!p->pagevec) {
89                                 mempool_free(p, nfs_wdata_mempool);
90                                 p = NULL;
91                         }
92                 }
93         }
94         return p;
95 }
96
97 static void nfs_writedata_rcu_free(struct rcu_head *head)
98 {
99         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
100         if (p && (p->pagevec != &p->page_array[0]))
101                 kfree(p->pagevec);
102         mempool_free(p, nfs_wdata_mempool);
103 }
104
105 static void nfs_writedata_free(struct nfs_write_data *wdata)
106 {
107         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
108 }
109
110 void nfs_writedata_release(void *wdata)
111 {
112         nfs_writedata_free(wdata);
113 }
114
115 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
116 {
117         struct nfs_page *req = NULL;
118
119         if (PagePrivate(page)) {
120                 req = (struct nfs_page *)page_private(page);
121                 if (req != NULL)
122                         atomic_inc(&req->wb_count);
123         }
124         return req;
125 }
126
127 static struct nfs_page *nfs_page_find_request(struct page *page)
128 {
129         struct nfs_page *req = NULL;
130         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
131
132         spin_lock(req_lock);
133         req = nfs_page_find_request_locked(page);
134         spin_unlock(req_lock);
135         return req;
136 }
137
138 /* Adjust the file length if we're writing beyond the end */
139 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
140 {
141         struct inode *inode = page->mapping->host;
142         loff_t end, i_size = i_size_read(inode);
143         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
144
145         if (i_size > 0 && page->index < end_index)
146                 return;
147         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
148         if (i_size >= end)
149                 return;
150         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
151         i_size_write(inode, end);
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         if (count != PAGE_CACHE_SIZE)
173                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
174         SetPageUptodate(page);
175 }
176
177 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
178                 unsigned int offset, unsigned int count)
179 {
180         struct nfs_page *req;
181         int ret;
182
183         for (;;) {
184                 req = nfs_update_request(ctx, page, offset, count);
185                 if (!IS_ERR(req))
186                         break;
187                 ret = PTR_ERR(req);
188                 if (ret != -EBUSY)
189                         return ret;
190                 ret = nfs_wb_page(page->mapping->host, page);
191                 if (ret != 0)
192                         return ret;
193         }
194         /* Update file length */
195         nfs_grow_file(page, offset, count);
196         /* Set the PG_uptodate flag? */
197         nfs_mark_uptodate(page, offset, count);
198         nfs_unlock_request(req);
199         return 0;
200 }
201
202 static int wb_priority(struct writeback_control *wbc)
203 {
204         if (wbc->for_reclaim)
205                 return FLUSH_HIGHPRI;
206         if (wbc->for_kupdate)
207                 return FLUSH_LOWPRI;
208         return 0;
209 }
210
211 /*
212  * NFS congestion control
213  */
214
215 int nfs_congestion_kb;
216
217 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
218 #define NFS_CONGESTION_OFF_THRESH       \
219         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
220
221 static int nfs_set_page_writeback(struct page *page)
222 {
223         int ret = test_set_page_writeback(page);
224
225         if (!ret) {
226                 struct inode *inode = page->mapping->host;
227                 struct nfs_server *nfss = NFS_SERVER(inode);
228
229                 if (atomic_inc_return(&nfss->writeback) >
230                                 NFS_CONGESTION_ON_THRESH)
231                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
232         }
233         return ret;
234 }
235
236 static void nfs_end_page_writeback(struct page *page)
237 {
238         struct inode *inode = page->mapping->host;
239         struct nfs_server *nfss = NFS_SERVER(inode);
240
241         end_page_writeback(page);
242         if (atomic_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) {
243                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
244                 congestion_end(WRITE);
245         }
246 }
247
248 /*
249  * Find an associated nfs write request, and prepare to flush it out
250  * Returns 1 if there was no write request, or if the request was
251  * already tagged by nfs_set_page_dirty.Returns 0 if the request
252  * was not tagged.
253  * May also return an error if the user signalled nfs_wait_on_request().
254  */
255 static int nfs_page_mark_flush(struct page *page)
256 {
257         struct nfs_page *req;
258         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
259         int ret;
260
261         spin_lock(req_lock);
262         for(;;) {
263                 req = nfs_page_find_request_locked(page);
264                 if (req == NULL) {
265                         spin_unlock(req_lock);
266                         return 1;
267                 }
268                 if (nfs_lock_request_dontget(req))
269                         break;
270                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
271                  *       then the call to nfs_lock_request_dontget() will always
272                  *       succeed provided that someone hasn't already marked the
273                  *       request as dirty (in which case we don't care).
274                  */
275                 spin_unlock(req_lock);
276                 ret = nfs_wait_on_request(req);
277                 nfs_release_request(req);
278                 if (ret != 0)
279                         return ret;
280                 spin_lock(req_lock);
281         }
282         spin_unlock(req_lock);
283         if (nfs_set_page_writeback(page) == 0) {
284                 nfs_list_remove_request(req);
285                 nfs_mark_request_dirty(req);
286         }
287         ret = test_bit(PG_NEED_FLUSH, &req->wb_flags);
288         nfs_unlock_request(req);
289         return ret;
290 }
291
292 /*
293  * Write an mmapped page to the server.
294  */
295 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
296 {
297         struct nfs_open_context *ctx;
298         struct inode *inode = page->mapping->host;
299         unsigned offset;
300         int err;
301
302         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
303         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
304
305         err = nfs_page_mark_flush(page);
306         if (err <= 0)
307                 goto out;
308         err = 0;
309         offset = nfs_page_length(page);
310         if (!offset)
311                 goto out;
312
313         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
314         if (ctx == NULL) {
315                 err = -EBADF;
316                 goto out;
317         }
318         err = nfs_writepage_setup(ctx, page, 0, offset);
319         put_nfs_open_context(ctx);
320         if (err != 0)
321                 goto out;
322         err = nfs_page_mark_flush(page);
323         if (err > 0)
324                 err = 0;
325 out:
326         if (!wbc->for_writepages)
327                 nfs_flush_mapping(page->mapping, wbc, FLUSH_STABLE|wb_priority(wbc));
328         return err;
329 }
330
331 int nfs_writepage(struct page *page, struct writeback_control *wbc)
332 {
333         int err;
334
335         err = nfs_writepage_locked(page, wbc);
336         unlock_page(page);
337         return err; 
338 }
339
340 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
341 {
342         struct inode *inode = mapping->host;
343         int err;
344
345         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
346
347         err = generic_writepages(mapping, wbc);
348         if (err)
349                 return err;
350         err = nfs_flush_mapping(mapping, wbc, wb_priority(wbc));
351         if (err < 0)
352                 goto out;
353         nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
354         err = 0;
355 out:
356         return err;
357 }
358
359 /*
360  * Insert a write request into an inode
361  */
362 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
363 {
364         struct nfs_inode *nfsi = NFS_I(inode);
365         int error;
366
367         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
368         BUG_ON(error == -EEXIST);
369         if (error)
370                 return error;
371         if (!nfsi->npages) {
372                 igrab(inode);
373                 nfs_begin_data_update(inode);
374                 if (nfs_have_delegation(inode, FMODE_WRITE))
375                         nfsi->change_attr++;
376         }
377         SetPagePrivate(req->wb_page);
378         set_page_private(req->wb_page, (unsigned long)req);
379         nfsi->npages++;
380         atomic_inc(&req->wb_count);
381         return 0;
382 }
383
384 /*
385  * Remove a write request from an inode
386  */
387 static void nfs_inode_remove_request(struct nfs_page *req)
388 {
389         struct inode *inode = req->wb_context->dentry->d_inode;
390         struct nfs_inode *nfsi = NFS_I(inode);
391
392         BUG_ON (!NFS_WBACK_BUSY(req));
393
394         spin_lock(&nfsi->req_lock);
395         set_page_private(req->wb_page, 0);
396         ClearPagePrivate(req->wb_page);
397         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
398         nfsi->npages--;
399         if (!nfsi->npages) {
400                 spin_unlock(&nfsi->req_lock);
401                 nfs_end_data_update(inode);
402                 iput(inode);
403         } else
404                 spin_unlock(&nfsi->req_lock);
405         nfs_clear_request(req);
406         nfs_release_request(req);
407 }
408
409 /*
410  * Add a request to the inode's dirty list.
411  */
412 static void
413 nfs_mark_request_dirty(struct nfs_page *req)
414 {
415         struct inode *inode = req->wb_context->dentry->d_inode;
416         struct nfs_inode *nfsi = NFS_I(inode);
417
418         spin_lock(&nfsi->req_lock);
419         radix_tree_tag_set(&nfsi->nfs_page_tree,
420                         req->wb_index, NFS_PAGE_TAG_DIRTY);
421         nfs_list_add_request(req, &nfsi->dirty);
422         nfsi->ndirty++;
423         spin_unlock(&nfsi->req_lock);
424         __mark_inode_dirty(inode, I_DIRTY_PAGES);
425 }
426
427 static void
428 nfs_redirty_request(struct nfs_page *req)
429 {
430         __set_page_dirty_nobuffers(req->wb_page);
431 }
432
433 /*
434  * Check if a request is dirty
435  */
436 static inline int
437 nfs_dirty_request(struct nfs_page *req)
438 {
439         struct page *page = req->wb_page;
440
441         if (page == NULL)
442                 return 0;
443         return !PageWriteback(req->wb_page);
444 }
445
446 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
447 /*
448  * Add a request to the inode's commit list.
449  */
450 static void
451 nfs_mark_request_commit(struct nfs_page *req)
452 {
453         struct inode *inode = req->wb_context->dentry->d_inode;
454         struct nfs_inode *nfsi = NFS_I(inode);
455
456         spin_lock(&nfsi->req_lock);
457         nfs_list_add_request(req, &nfsi->commit);
458         nfsi->ncommit++;
459         spin_unlock(&nfsi->req_lock);
460         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
461         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
462 }
463 #endif
464
465 /*
466  * Wait for a request to complete.
467  *
468  * Interruptible by signals only if mounted with intr flag.
469  */
470 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
471 {
472         struct nfs_inode *nfsi = NFS_I(inode);
473         struct nfs_page *req;
474         unsigned long           idx_end, next;
475         unsigned int            res = 0;
476         int                     error;
477
478         if (npages == 0)
479                 idx_end = ~0;
480         else
481                 idx_end = idx_start + npages - 1;
482
483         next = idx_start;
484         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
485                 if (req->wb_index > idx_end)
486                         break;
487
488                 next = req->wb_index + 1;
489                 BUG_ON(!NFS_WBACK_BUSY(req));
490
491                 atomic_inc(&req->wb_count);
492                 spin_unlock(&nfsi->req_lock);
493                 error = nfs_wait_on_request(req);
494                 nfs_release_request(req);
495                 spin_lock(&nfsi->req_lock);
496                 if (error < 0)
497                         return error;
498                 res++;
499         }
500         return res;
501 }
502
503 static void nfs_cancel_dirty_list(struct list_head *head)
504 {
505         struct nfs_page *req;
506         while(!list_empty(head)) {
507                 req = nfs_list_entry(head->next);
508                 nfs_list_remove_request(req);
509                 nfs_end_page_writeback(req->wb_page);
510                 nfs_inode_remove_request(req);
511                 nfs_clear_page_writeback(req);
512         }
513 }
514
515 static void nfs_cancel_commit_list(struct list_head *head)
516 {
517         struct nfs_page *req;
518
519         while(!list_empty(head)) {
520                 req = nfs_list_entry(head->next);
521                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
522                 nfs_list_remove_request(req);
523                 nfs_inode_remove_request(req);
524                 nfs_unlock_request(req);
525         }
526 }
527
528 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
529 /*
530  * nfs_scan_commit - Scan an inode for commit requests
531  * @inode: NFS inode to scan
532  * @dst: destination list
533  * @idx_start: lower bound of page->index to scan.
534  * @npages: idx_start + npages sets the upper bound to scan.
535  *
536  * Moves requests from the inode's 'commit' request list.
537  * The requests are *not* checked to ensure that they form a contiguous set.
538  */
539 static int
540 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
541 {
542         struct nfs_inode *nfsi = NFS_I(inode);
543         int res = 0;
544
545         if (nfsi->ncommit != 0) {
546                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
547                 nfsi->ncommit -= res;
548                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
549                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
550         }
551         return res;
552 }
553 #else
554 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
555 {
556         return 0;
557 }
558 #endif
559
560 static int nfs_wait_on_write_congestion(struct address_space *mapping)
561 {
562         struct inode *inode = mapping->host;
563         struct backing_dev_info *bdi = mapping->backing_dev_info;
564         int ret = 0;
565
566         might_sleep();
567
568         if (!bdi_write_congested(bdi))
569                 return 0;
570
571         nfs_inc_stats(inode, NFSIOS_CONGESTIONWAIT);
572
573         do {
574                 struct rpc_clnt *clnt = NFS_CLIENT(inode);
575                 sigset_t oldset;
576
577                 rpc_clnt_sigmask(clnt, &oldset);
578                 ret = congestion_wait_interruptible(WRITE, HZ/10);
579                 rpc_clnt_sigunmask(clnt, &oldset);
580                 if (ret == -ERESTARTSYS)
581                         break;
582                 ret = 0;
583         } while (bdi_write_congested(bdi));
584
585         return ret;
586 }
587
588 /*
589  * Try to update any existing write request, or create one if there is none.
590  * In order to match, the request's credentials must match those of
591  * the calling process.
592  *
593  * Note: Should always be called with the Page Lock held!
594  */
595 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
596                 struct page *page, unsigned int offset, unsigned int bytes)
597 {
598         struct address_space *mapping = page->mapping;
599         struct inode *inode = mapping->host;
600         struct nfs_inode *nfsi = NFS_I(inode);
601         struct nfs_page         *req, *new = NULL;
602         unsigned long           rqend, end;
603
604         end = offset + bytes;
605
606         if (nfs_wait_on_write_congestion(mapping))
607                 return ERR_PTR(-ERESTARTSYS);
608         for (;;) {
609                 /* Loop over all inode entries and see if we find
610                  * A request for the page we wish to update
611                  */
612                 spin_lock(&nfsi->req_lock);
613                 req = nfs_page_find_request_locked(page);
614                 if (req) {
615                         if (!nfs_lock_request_dontget(req)) {
616                                 int error;
617
618                                 spin_unlock(&nfsi->req_lock);
619                                 error = nfs_wait_on_request(req);
620                                 nfs_release_request(req);
621                                 if (error < 0) {
622                                         if (new)
623                                                 nfs_release_request(new);
624                                         return ERR_PTR(error);
625                                 }
626                                 continue;
627                         }
628                         spin_unlock(&nfsi->req_lock);
629                         if (new)
630                                 nfs_release_request(new);
631                         break;
632                 }
633
634                 if (new) {
635                         int error;
636                         nfs_lock_request_dontget(new);
637                         error = nfs_inode_add_request(inode, new);
638                         if (error) {
639                                 spin_unlock(&nfsi->req_lock);
640                                 nfs_unlock_request(new);
641                                 return ERR_PTR(error);
642                         }
643                         spin_unlock(&nfsi->req_lock);
644                         return new;
645                 }
646                 spin_unlock(&nfsi->req_lock);
647
648                 new = nfs_create_request(ctx, inode, page, offset, bytes);
649                 if (IS_ERR(new))
650                         return new;
651         }
652
653         /* We have a request for our page.
654          * If the creds don't match, or the
655          * page addresses don't match,
656          * tell the caller to wait on the conflicting
657          * request.
658          */
659         rqend = req->wb_offset + req->wb_bytes;
660         if (req->wb_context != ctx
661             || req->wb_page != page
662             || !nfs_dirty_request(req)
663             || offset > rqend || end < req->wb_offset) {
664                 nfs_unlock_request(req);
665                 return ERR_PTR(-EBUSY);
666         }
667
668         /* Okay, the request matches. Update the region */
669         if (offset < req->wb_offset) {
670                 req->wb_offset = offset;
671                 req->wb_pgbase = offset;
672                 req->wb_bytes = rqend - req->wb_offset;
673         }
674
675         if (end > rqend)
676                 req->wb_bytes = end - req->wb_offset;
677
678         return req;
679 }
680
681 int nfs_flush_incompatible(struct file *file, struct page *page)
682 {
683         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
684         struct nfs_page *req;
685         int do_flush, status;
686         /*
687          * Look for a request corresponding to this page. If there
688          * is one, and it belongs to another file, we flush it out
689          * before we try to copy anything into the page. Do this
690          * due to the lack of an ACCESS-type call in NFSv2.
691          * Also do the same if we find a request from an existing
692          * dropped page.
693          */
694         do {
695                 req = nfs_page_find_request(page);
696                 if (req == NULL)
697                         return 0;
698                 do_flush = req->wb_page != page || req->wb_context != ctx
699                         || !nfs_dirty_request(req);
700                 nfs_release_request(req);
701                 if (!do_flush)
702                         return 0;
703                 status = nfs_wb_page(page->mapping->host, page);
704         } while (status == 0);
705         return status;
706 }
707
708 /*
709  * Update and possibly write a cached page of an NFS file.
710  *
711  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
712  * things with a page scheduled for an RPC call (e.g. invalidate it).
713  */
714 int nfs_updatepage(struct file *file, struct page *page,
715                 unsigned int offset, unsigned int count)
716 {
717         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
718         struct inode    *inode = page->mapping->host;
719         int             status = 0;
720
721         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
722
723         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
724                 file->f_path.dentry->d_parent->d_name.name,
725                 file->f_path.dentry->d_name.name, count,
726                 (long long)(page_offset(page) +offset));
727
728         /* If we're not using byte range locks, and we know the page
729          * is entirely in cache, it may be more efficient to avoid
730          * fragmenting write requests.
731          */
732         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
733                 count = max(count + offset, nfs_page_length(page));
734                 offset = 0;
735         }
736
737         status = nfs_writepage_setup(ctx, page, offset, count);
738         __set_page_dirty_nobuffers(page);
739
740         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
741                         status, (long long)i_size_read(inode));
742         if (status < 0)
743                 nfs_set_pageerror(page);
744         return status;
745 }
746
747 static void nfs_writepage_release(struct nfs_page *req)
748 {
749         nfs_end_page_writeback(req->wb_page);
750
751 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
752         if (!PageError(req->wb_page)) {
753                 if (NFS_NEED_RESCHED(req)) {
754                         nfs_redirty_request(req);
755                         goto out;
756                 } else if (NFS_NEED_COMMIT(req)) {
757                         nfs_mark_request_commit(req);
758                         goto out;
759                 }
760         }
761         nfs_inode_remove_request(req);
762
763 out:
764         nfs_clear_commit(req);
765         nfs_clear_reschedule(req);
766 #else
767         nfs_inode_remove_request(req);
768 #endif
769         nfs_clear_page_writeback(req);
770 }
771
772 static inline int flush_task_priority(int how)
773 {
774         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
775                 case FLUSH_HIGHPRI:
776                         return RPC_PRIORITY_HIGH;
777                 case FLUSH_LOWPRI:
778                         return RPC_PRIORITY_LOW;
779         }
780         return RPC_PRIORITY_NORMAL;
781 }
782
783 /*
784  * Set up the argument/result storage required for the RPC call.
785  */
786 static void nfs_write_rpcsetup(struct nfs_page *req,
787                 struct nfs_write_data *data,
788                 const struct rpc_call_ops *call_ops,
789                 unsigned int count, unsigned int offset,
790                 int how)
791 {
792         struct inode            *inode;
793         int flags;
794
795         /* Set up the RPC argument and reply structs
796          * NB: take care not to mess about with data->commit et al. */
797
798         data->req = req;
799         data->inode = inode = req->wb_context->dentry->d_inode;
800         data->cred = req->wb_context->cred;
801
802         data->args.fh     = NFS_FH(inode);
803         data->args.offset = req_offset(req) + offset;
804         data->args.pgbase = req->wb_pgbase + offset;
805         data->args.pages  = data->pagevec;
806         data->args.count  = count;
807         data->args.context = req->wb_context;
808
809         data->res.fattr   = &data->fattr;
810         data->res.count   = count;
811         data->res.verf    = &data->verf;
812         nfs_fattr_init(&data->fattr);
813
814         /* Set up the initial task struct.  */
815         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
816         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
817         NFS_PROTO(inode)->write_setup(data, how);
818
819         data->task.tk_priority = flush_task_priority(how);
820         data->task.tk_cookie = (unsigned long)inode;
821
822         dprintk("NFS: %5u initiated write call "
823                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
824                 data->task.tk_pid,
825                 inode->i_sb->s_id,
826                 (long long)NFS_FILEID(inode),
827                 count,
828                 (unsigned long long)data->args.offset);
829 }
830
831 static void nfs_execute_write(struct nfs_write_data *data)
832 {
833         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
834         sigset_t oldset;
835
836         rpc_clnt_sigmask(clnt, &oldset);
837         rpc_execute(&data->task);
838         rpc_clnt_sigunmask(clnt, &oldset);
839 }
840
841 /*
842  * Generate multiple small requests to write out a single
843  * contiguous dirty area on one page.
844  */
845 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
846 {
847         struct nfs_page *req = nfs_list_entry(head->next);
848         struct page *page = req->wb_page;
849         struct nfs_write_data *data;
850         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
851         unsigned int offset;
852         int requests = 0;
853         LIST_HEAD(list);
854
855         nfs_list_remove_request(req);
856
857         nbytes = req->wb_bytes;
858         do {
859                 size_t len = min(nbytes, wsize);
860
861                 data = nfs_writedata_alloc(len);
862                 if (!data)
863                         goto out_bad;
864                 list_add(&data->pages, &list);
865                 requests++;
866                 nbytes -= len;
867         } while (nbytes != 0);
868         atomic_set(&req->wb_complete, requests);
869
870         ClearPageError(page);
871         offset = 0;
872         nbytes = req->wb_bytes;
873         do {
874                 data = list_entry(list.next, struct nfs_write_data, pages);
875                 list_del_init(&data->pages);
876
877                 data->pagevec[0] = page;
878
879                 if (nbytes > wsize) {
880                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
881                                         wsize, offset, how);
882                         offset += wsize;
883                         nbytes -= wsize;
884                 } else {
885                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
886                                         nbytes, offset, how);
887                         nbytes = 0;
888                 }
889                 nfs_execute_write(data);
890         } while (nbytes != 0);
891
892         return 0;
893
894 out_bad:
895         while (!list_empty(&list)) {
896                 data = list_entry(list.next, struct nfs_write_data, pages);
897                 list_del(&data->pages);
898                 nfs_writedata_release(data);
899         }
900         nfs_end_page_writeback(req->wb_page);
901         nfs_redirty_request(req);
902         nfs_clear_page_writeback(req);
903         return -ENOMEM;
904 }
905
906 /*
907  * Create an RPC task for the given write request and kick it.
908  * The page must have been locked by the caller.
909  *
910  * It may happen that the page we're passed is not marked dirty.
911  * This is the case if nfs_updatepage detects a conflicting request
912  * that has been written but not committed.
913  */
914 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
915 {
916         struct nfs_page         *req;
917         struct page             **pages;
918         struct nfs_write_data   *data;
919         unsigned int            count;
920
921         data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
922         if (!data)
923                 goto out_bad;
924
925         pages = data->pagevec;
926         count = 0;
927         while (!list_empty(head)) {
928                 req = nfs_list_entry(head->next);
929                 nfs_list_remove_request(req);
930                 nfs_list_add_request(req, &data->pages);
931                 ClearPageError(req->wb_page);
932                 *pages++ = req->wb_page;
933                 count += req->wb_bytes;
934         }
935         req = nfs_list_entry(data->pages.next);
936
937         /* Set up the argument struct */
938         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
939
940         nfs_execute_write(data);
941         return 0;
942  out_bad:
943         while (!list_empty(head)) {
944                 struct nfs_page *req = nfs_list_entry(head->next);
945                 nfs_list_remove_request(req);
946                 nfs_end_page_writeback(req->wb_page);
947                 nfs_redirty_request(req);
948                 nfs_clear_page_writeback(req);
949         }
950         return -ENOMEM;
951 }
952
953 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
954 {
955         LIST_HEAD(one_request);
956         int (*flush_one)(struct inode *, struct list_head *, int);
957         struct nfs_page *req;
958         int wpages = NFS_SERVER(inode)->wpages;
959         int wsize = NFS_SERVER(inode)->wsize;
960         int error;
961
962         flush_one = nfs_flush_one;
963         if (wsize < PAGE_CACHE_SIZE)
964                 flush_one = nfs_flush_multi;
965         /* For single writes, FLUSH_STABLE is more efficient */
966         if (npages <= wpages && npages == NFS_I(inode)->npages
967                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
968                 how |= FLUSH_STABLE;
969
970         do {
971                 nfs_coalesce_requests(head, &one_request, wpages);
972                 req = nfs_list_entry(one_request.next);
973                 error = flush_one(inode, &one_request, how);
974                 if (error < 0)
975                         goto out_err;
976         } while (!list_empty(head));
977         return 0;
978 out_err:
979         while (!list_empty(head)) {
980                 req = nfs_list_entry(head->next);
981                 nfs_list_remove_request(req);
982                 nfs_end_page_writeback(req->wb_page);
983                 nfs_redirty_request(req);
984                 nfs_clear_page_writeback(req);
985         }
986         return error;
987 }
988
989 /*
990  * Handle a write reply that flushed part of a page.
991  */
992 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
993 {
994         struct nfs_write_data   *data = calldata;
995         struct nfs_page         *req = data->req;
996         struct page             *page = req->wb_page;
997
998         dprintk("NFS: write (%s/%Ld %d@%Ld)",
999                 req->wb_context->dentry->d_inode->i_sb->s_id,
1000                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1001                 req->wb_bytes,
1002                 (long long)req_offset(req));
1003
1004         if (nfs_writeback_done(task, data) != 0)
1005                 return;
1006
1007         if (task->tk_status < 0) {
1008                 nfs_set_pageerror(page);
1009                 req->wb_context->error = task->tk_status;
1010                 dprintk(", error = %d\n", task->tk_status);
1011         } else {
1012 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1013                 if (data->verf.committed < NFS_FILE_SYNC) {
1014                         if (!NFS_NEED_COMMIT(req)) {
1015                                 nfs_defer_commit(req);
1016                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1017                                 dprintk(" defer commit\n");
1018                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1019                                 nfs_defer_reschedule(req);
1020                                 dprintk(" server reboot detected\n");
1021                         }
1022                 } else
1023 #endif
1024                         dprintk(" OK\n");
1025         }
1026
1027         if (atomic_dec_and_test(&req->wb_complete))
1028                 nfs_writepage_release(req);
1029 }
1030
1031 static const struct rpc_call_ops nfs_write_partial_ops = {
1032         .rpc_call_done = nfs_writeback_done_partial,
1033         .rpc_release = nfs_writedata_release,
1034 };
1035
1036 /*
1037  * Handle a write reply that flushes a whole page.
1038  *
1039  * FIXME: There is an inherent race with invalidate_inode_pages and
1040  *        writebacks since the page->count is kept > 1 for as long
1041  *        as the page has a write request pending.
1042  */
1043 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1044 {
1045         struct nfs_write_data   *data = calldata;
1046         struct nfs_page         *req;
1047         struct page             *page;
1048
1049         if (nfs_writeback_done(task, data) != 0)
1050                 return;
1051
1052         /* Update attributes as result of writeback. */
1053         while (!list_empty(&data->pages)) {
1054                 req = nfs_list_entry(data->pages.next);
1055                 nfs_list_remove_request(req);
1056                 page = req->wb_page;
1057
1058                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1059                         req->wb_context->dentry->d_inode->i_sb->s_id,
1060                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1061                         req->wb_bytes,
1062                         (long long)req_offset(req));
1063
1064                 if (task->tk_status < 0) {
1065                         nfs_set_pageerror(page);
1066                         req->wb_context->error = task->tk_status;
1067                         nfs_end_page_writeback(page);
1068                         nfs_inode_remove_request(req);
1069                         dprintk(", error = %d\n", task->tk_status);
1070                         goto next;
1071                 }
1072                 nfs_end_page_writeback(page);
1073
1074 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1075                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1076                         nfs_inode_remove_request(req);
1077                         dprintk(" OK\n");
1078                         goto next;
1079                 }
1080                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1081                 nfs_mark_request_commit(req);
1082                 dprintk(" marked for commit\n");
1083 #else
1084                 nfs_inode_remove_request(req);
1085 #endif
1086         next:
1087                 nfs_clear_page_writeback(req);
1088         }
1089 }
1090
1091 static const struct rpc_call_ops nfs_write_full_ops = {
1092         .rpc_call_done = nfs_writeback_done_full,
1093         .rpc_release = nfs_writedata_release,
1094 };
1095
1096
1097 /*
1098  * This function is called when the WRITE call is complete.
1099  */
1100 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1101 {
1102         struct nfs_writeargs    *argp = &data->args;
1103         struct nfs_writeres     *resp = &data->res;
1104         int status;
1105
1106         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1107                 task->tk_pid, task->tk_status);
1108
1109         /*
1110          * ->write_done will attempt to use post-op attributes to detect
1111          * conflicting writes by other clients.  A strict interpretation
1112          * of close-to-open would allow us to continue caching even if
1113          * another writer had changed the file, but some applications
1114          * depend on tighter cache coherency when writing.
1115          */
1116         status = NFS_PROTO(data->inode)->write_done(task, data);
1117         if (status != 0)
1118                 return status;
1119         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1120
1121 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1122         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1123                 /* We tried a write call, but the server did not
1124                  * commit data to stable storage even though we
1125                  * requested it.
1126                  * Note: There is a known bug in Tru64 < 5.0 in which
1127                  *       the server reports NFS_DATA_SYNC, but performs
1128                  *       NFS_FILE_SYNC. We therefore implement this checking
1129                  *       as a dprintk() in order to avoid filling syslog.
1130                  */
1131                 static unsigned long    complain;
1132
1133                 if (time_before(complain, jiffies)) {
1134                         dprintk("NFS: faulty NFS server %s:"
1135                                 " (committed = %d) != (stable = %d)\n",
1136                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1137                                 resp->verf->committed, argp->stable);
1138                         complain = jiffies + 300 * HZ;
1139                 }
1140         }
1141 #endif
1142         /* Is this a short write? */
1143         if (task->tk_status >= 0 && resp->count < argp->count) {
1144                 static unsigned long    complain;
1145
1146                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1147
1148                 /* Has the server at least made some progress? */
1149                 if (resp->count != 0) {
1150                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1151                         if (resp->verf->committed != NFS_UNSTABLE) {
1152                                 /* Resend from where the server left off */
1153                                 argp->offset += resp->count;
1154                                 argp->pgbase += resp->count;
1155                                 argp->count -= resp->count;
1156                         } else {
1157                                 /* Resend as a stable write in order to avoid
1158                                  * headaches in the case of a server crash.
1159                                  */
1160                                 argp->stable = NFS_FILE_SYNC;
1161                         }
1162                         rpc_restart_call(task);
1163                         return -EAGAIN;
1164                 }
1165                 if (time_before(complain, jiffies)) {
1166                         printk(KERN_WARNING
1167                                "NFS: Server wrote zero bytes, expected %u.\n",
1168                                         argp->count);
1169                         complain = jiffies + 300 * HZ;
1170                 }
1171                 /* Can't do anything about it except throw an error. */
1172                 task->tk_status = -EIO;
1173         }
1174         return 0;
1175 }
1176
1177
1178 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1179 void nfs_commit_release(void *wdata)
1180 {
1181         nfs_commit_free(wdata);
1182 }
1183
1184 /*
1185  * Set up the argument/result storage required for the RPC call.
1186  */
1187 static void nfs_commit_rpcsetup(struct list_head *head,
1188                 struct nfs_write_data *data,
1189                 int how)
1190 {
1191         struct nfs_page         *first;
1192         struct inode            *inode;
1193         int flags;
1194
1195         /* Set up the RPC argument and reply structs
1196          * NB: take care not to mess about with data->commit et al. */
1197
1198         list_splice_init(head, &data->pages);
1199         first = nfs_list_entry(data->pages.next);
1200         inode = first->wb_context->dentry->d_inode;
1201
1202         data->inode       = inode;
1203         data->cred        = first->wb_context->cred;
1204
1205         data->args.fh     = NFS_FH(data->inode);
1206         /* Note: we always request a commit of the entire inode */
1207         data->args.offset = 0;
1208         data->args.count  = 0;
1209         data->res.count   = 0;
1210         data->res.fattr   = &data->fattr;
1211         data->res.verf    = &data->verf;
1212         nfs_fattr_init(&data->fattr);
1213
1214         /* Set up the initial task struct.  */
1215         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1216         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1217         NFS_PROTO(inode)->commit_setup(data, how);
1218
1219         data->task.tk_priority = flush_task_priority(how);
1220         data->task.tk_cookie = (unsigned long)inode;
1221         
1222         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1223 }
1224
1225 /*
1226  * Commit dirty pages
1227  */
1228 static int
1229 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1230 {
1231         struct nfs_write_data   *data;
1232         struct nfs_page         *req;
1233
1234         data = nfs_commit_alloc();
1235
1236         if (!data)
1237                 goto out_bad;
1238
1239         /* Set up the argument struct */
1240         nfs_commit_rpcsetup(head, data, how);
1241
1242         nfs_execute_write(data);
1243         return 0;
1244  out_bad:
1245         while (!list_empty(head)) {
1246                 req = nfs_list_entry(head->next);
1247                 nfs_list_remove_request(req);
1248                 nfs_mark_request_commit(req);
1249                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1250                 nfs_clear_page_writeback(req);
1251         }
1252         return -ENOMEM;
1253 }
1254
1255 /*
1256  * COMMIT call returned
1257  */
1258 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1259 {
1260         struct nfs_write_data   *data = calldata;
1261         struct nfs_page         *req;
1262
1263         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1264                                 task->tk_pid, task->tk_status);
1265
1266         /* Call the NFS version-specific code */
1267         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1268                 return;
1269
1270         while (!list_empty(&data->pages)) {
1271                 req = nfs_list_entry(data->pages.next);
1272                 nfs_list_remove_request(req);
1273                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1274
1275                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1276                         req->wb_context->dentry->d_inode->i_sb->s_id,
1277                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1278                         req->wb_bytes,
1279                         (long long)req_offset(req));
1280                 if (task->tk_status < 0) {
1281                         req->wb_context->error = task->tk_status;
1282                         nfs_inode_remove_request(req);
1283                         dprintk(", error = %d\n", task->tk_status);
1284                         goto next;
1285                 }
1286
1287                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1288                  * returned by the server against all stored verfs. */
1289                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1290                         /* We have a match */
1291                         nfs_inode_remove_request(req);
1292                         dprintk(" OK\n");
1293                         goto next;
1294                 }
1295                 /* We have a mismatch. Write the page again */
1296                 dprintk(" mismatch\n");
1297                 nfs_redirty_request(req);
1298         next:
1299                 nfs_clear_page_writeback(req);
1300         }
1301 }
1302
1303 static const struct rpc_call_ops nfs_commit_ops = {
1304         .rpc_call_done = nfs_commit_done,
1305         .rpc_release = nfs_commit_release,
1306 };
1307 #else
1308 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1309 {
1310         return 0;
1311 }
1312 #endif
1313
1314 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1315 {
1316         struct nfs_inode *nfsi = NFS_I(mapping->host);
1317         LIST_HEAD(head);
1318         long res;
1319
1320         spin_lock(&nfsi->req_lock);
1321         res = nfs_scan_dirty(mapping, wbc, &head);
1322         spin_unlock(&nfsi->req_lock);
1323         if (res) {
1324                 int error = nfs_flush_list(mapping->host, &head, res, how);
1325                 if (error < 0)
1326                         return error;
1327         }
1328         return res;
1329 }
1330
1331 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1332 int nfs_commit_inode(struct inode *inode, int how)
1333 {
1334         struct nfs_inode *nfsi = NFS_I(inode);
1335         LIST_HEAD(head);
1336         int res;
1337
1338         spin_lock(&nfsi->req_lock);
1339         res = nfs_scan_commit(inode, &head, 0, 0);
1340         spin_unlock(&nfsi->req_lock);
1341         if (res) {
1342                 int error = nfs_commit_list(inode, &head, how);
1343                 if (error < 0)
1344                         return error;
1345         }
1346         return res;
1347 }
1348 #endif
1349
1350 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1351 {
1352         struct inode *inode = mapping->host;
1353         struct nfs_inode *nfsi = NFS_I(inode);
1354         unsigned long idx_start, idx_end;
1355         unsigned int npages = 0;
1356         LIST_HEAD(head);
1357         int nocommit = how & FLUSH_NOCOMMIT;
1358         long pages, ret;
1359
1360         /* FIXME */
1361         if (wbc->range_cyclic)
1362                 idx_start = 0;
1363         else {
1364                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1365                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1366                 if (idx_end > idx_start) {
1367                         unsigned long l_npages = 1 + idx_end - idx_start;
1368                         npages = l_npages;
1369                         if (sizeof(npages) != sizeof(l_npages) &&
1370                                         (unsigned long)npages != l_npages)
1371                                 npages = 0;
1372                 }
1373         }
1374         how &= ~FLUSH_NOCOMMIT;
1375         spin_lock(&nfsi->req_lock);
1376         do {
1377                 wbc->pages_skipped = 0;
1378                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1379                 if (ret != 0)
1380                         continue;
1381                 pages = nfs_scan_dirty(mapping, wbc, &head);
1382                 if (pages != 0) {
1383                         spin_unlock(&nfsi->req_lock);
1384                         if (how & FLUSH_INVALIDATE) {
1385                                 nfs_cancel_dirty_list(&head);
1386                                 ret = pages;
1387                         } else
1388                                 ret = nfs_flush_list(inode, &head, pages, how);
1389                         spin_lock(&nfsi->req_lock);
1390                         continue;
1391                 }
1392                 if (wbc->pages_skipped != 0)
1393                         continue;
1394                 if (nocommit)
1395                         break;
1396                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1397                 if (pages == 0) {
1398                         if (wbc->pages_skipped != 0)
1399                                 continue;
1400                         break;
1401                 }
1402                 if (how & FLUSH_INVALIDATE) {
1403                         spin_unlock(&nfsi->req_lock);
1404                         nfs_cancel_commit_list(&head);
1405                         ret = pages;
1406                         spin_lock(&nfsi->req_lock);
1407                         continue;
1408                 }
1409                 pages += nfs_scan_commit(inode, &head, 0, 0);
1410                 spin_unlock(&nfsi->req_lock);
1411                 ret = nfs_commit_list(inode, &head, how);
1412                 spin_lock(&nfsi->req_lock);
1413         } while (ret >= 0);
1414         spin_unlock(&nfsi->req_lock);
1415         return ret;
1416 }
1417
1418 /*
1419  * flush the inode to disk.
1420  */
1421 int nfs_wb_all(struct inode *inode)
1422 {
1423         struct address_space *mapping = inode->i_mapping;
1424         struct writeback_control wbc = {
1425                 .bdi = mapping->backing_dev_info,
1426                 .sync_mode = WB_SYNC_ALL,
1427                 .nr_to_write = LONG_MAX,
1428                 .for_writepages = 1,
1429                 .range_cyclic = 1,
1430         };
1431         int ret;
1432
1433         ret = generic_writepages(mapping, &wbc);
1434         if (ret < 0)
1435                 goto out;
1436         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1437         if (ret >= 0)
1438                 return 0;
1439 out:
1440         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1441         return ret;
1442 }
1443
1444 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1445 {
1446         struct writeback_control wbc = {
1447                 .bdi = mapping->backing_dev_info,
1448                 .sync_mode = WB_SYNC_ALL,
1449                 .nr_to_write = LONG_MAX,
1450                 .range_start = range_start,
1451                 .range_end = range_end,
1452                 .for_writepages = 1,
1453         };
1454         int ret;
1455
1456         if (!(how & FLUSH_NOWRITEPAGE)) {
1457                 ret = generic_writepages(mapping, &wbc);
1458                 if (ret < 0)
1459                         goto out;
1460         }
1461         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1462         if (ret >= 0)
1463                 return 0;
1464 out:
1465         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1466         return ret;
1467 }
1468
1469 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1470 {
1471         loff_t range_start = page_offset(page);
1472         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1473         struct writeback_control wbc = {
1474                 .bdi = page->mapping->backing_dev_info,
1475                 .sync_mode = WB_SYNC_ALL,
1476                 .nr_to_write = LONG_MAX,
1477                 .range_start = range_start,
1478                 .range_end = range_end,
1479         };
1480         int ret;
1481
1482         BUG_ON(!PageLocked(page));
1483         if (!(how & FLUSH_NOWRITEPAGE) && clear_page_dirty_for_io(page)) {
1484                 ret = nfs_writepage_locked(page, &wbc);
1485                 if (ret < 0)
1486                         goto out;
1487         }
1488         if (!PagePrivate(page))
1489                 return 0;
1490         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1491         if (ret >= 0)
1492                 return 0;
1493 out:
1494         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1495         return ret;
1496 }
1497
1498 /*
1499  * Write back all requests on one page - we do this before reading it.
1500  */
1501 int nfs_wb_page(struct inode *inode, struct page* page)
1502 {
1503         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1504 }
1505
1506 int nfs_set_page_dirty(struct page *page)
1507 {
1508         struct nfs_page *req;
1509
1510         req = nfs_page_find_request(page);
1511         if (req != NULL) {
1512                 /* Mark any existing write requests for flushing */
1513                 set_bit(PG_NEED_FLUSH, &req->wb_flags);
1514                 nfs_release_request(req);
1515         }
1516         return __set_page_dirty_nobuffers(page);
1517 }
1518
1519
1520 int __init nfs_init_writepagecache(void)
1521 {
1522         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1523                                              sizeof(struct nfs_write_data),
1524                                              0, SLAB_HWCACHE_ALIGN,
1525                                              NULL, NULL);
1526         if (nfs_wdata_cachep == NULL)
1527                 return -ENOMEM;
1528
1529         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1530                                                      nfs_wdata_cachep);
1531         if (nfs_wdata_mempool == NULL)
1532                 return -ENOMEM;
1533
1534         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1535                                                       nfs_wdata_cachep);
1536         if (nfs_commit_mempool == NULL)
1537                 return -ENOMEM;
1538
1539         /*
1540          * NFS congestion size, scale with available memory.
1541          *
1542          *  64MB:    8192k
1543          * 128MB:   11585k
1544          * 256MB:   16384k
1545          * 512MB:   23170k
1546          *   1GB:   32768k
1547          *   2GB:   46340k
1548          *   4GB:   65536k
1549          *   8GB:   92681k
1550          *  16GB:  131072k
1551          *
1552          * This allows larger machines to have larger/more transfers.
1553          * Limit the default to 256M
1554          */
1555         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1556         if (nfs_congestion_kb > 256*1024)
1557                 nfs_congestion_kb = 256*1024;
1558
1559         return 0;
1560 }
1561
1562 void nfs_destroy_writepagecache(void)
1563 {
1564         mempool_destroy(nfs_commit_mempool);
1565         mempool_destroy(nfs_wdata_mempool);
1566         kmem_cache_destroy(nfs_wdata_cachep);
1567 }
1568