NFS: Throttle page dirtying while we're flushing to disk
[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
25 #include "delegation.h"
26 #include "internal.h"
27 #include "iostat.h"
28
29 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
30
31 #define MIN_POOL_WRITE          (32)
32 #define MIN_POOL_COMMIT         (4)
33
34 /*
35  * Local function declarations
36  */
37 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
38                                   struct inode *inode, int ioflags);
39 static void nfs_redirty_request(struct nfs_page *req);
40 static const struct rpc_call_ops nfs_write_partial_ops;
41 static const struct rpc_call_ops nfs_write_full_ops;
42 static const struct rpc_call_ops nfs_commit_ops;
43
44 static struct kmem_cache *nfs_wdata_cachep;
45 static mempool_t *nfs_wdata_mempool;
46 static mempool_t *nfs_commit_mempool;
47
48 struct nfs_write_data *nfs_commitdata_alloc(void)
49 {
50         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
51
52         if (p) {
53                 memset(p, 0, sizeof(*p));
54                 INIT_LIST_HEAD(&p->pages);
55         }
56         return p;
57 }
58
59 void nfs_commit_free(struct nfs_write_data *p)
60 {
61         if (p && (p->pagevec != &p->page_array[0]))
62                 kfree(p->pagevec);
63         mempool_free(p, nfs_commit_mempool);
64 }
65
66 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
67 {
68         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
69
70         if (p) {
71                 memset(p, 0, sizeof(*p));
72                 INIT_LIST_HEAD(&p->pages);
73                 p->npages = pagecount;
74                 if (pagecount <= ARRAY_SIZE(p->page_array))
75                         p->pagevec = p->page_array;
76                 else {
77                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
78                         if (!p->pagevec) {
79                                 mempool_free(p, nfs_wdata_mempool);
80                                 p = NULL;
81                         }
82                 }
83         }
84         return p;
85 }
86
87 static void nfs_writedata_free(struct nfs_write_data *p)
88 {
89         if (p && (p->pagevec != &p->page_array[0]))
90                 kfree(p->pagevec);
91         mempool_free(p, nfs_wdata_mempool);
92 }
93
94 void nfs_writedata_release(void *data)
95 {
96         struct nfs_write_data *wdata = data;
97
98         put_nfs_open_context(wdata->args.context);
99         nfs_writedata_free(wdata);
100 }
101
102 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
103 {
104         ctx->error = error;
105         smp_wmb();
106         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
107 }
108
109 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
110 {
111         struct nfs_page *req = NULL;
112
113         if (PagePrivate(page)) {
114                 req = (struct nfs_page *)page_private(page);
115                 if (req != NULL)
116                         kref_get(&req->wb_kref);
117         }
118         return req;
119 }
120
121 static struct nfs_page *nfs_page_find_request(struct page *page)
122 {
123         struct inode *inode = page->mapping->host;
124         struct nfs_page *req = NULL;
125
126         spin_lock(&inode->i_lock);
127         req = nfs_page_find_request_locked(page);
128         spin_unlock(&inode->i_lock);
129         return req;
130 }
131
132 /* Adjust the file length if we're writing beyond the end */
133 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
134 {
135         struct inode *inode = page->mapping->host;
136         loff_t end, i_size;
137         pgoff_t end_index;
138
139         spin_lock(&inode->i_lock);
140         i_size = i_size_read(inode);
141         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
142         if (i_size > 0 && page->index < end_index)
143                 goto out;
144         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
145         if (i_size >= end)
146                 goto out;
147         i_size_write(inode, end);
148         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
149 out:
150         spin_unlock(&inode->i_lock);
151 }
152
153 /* A writeback failed: mark the page as bad, and invalidate the page cache */
154 static void nfs_set_pageerror(struct page *page)
155 {
156         SetPageError(page);
157         nfs_zap_mapping(page->mapping->host, page->mapping);
158 }
159
160 /* We can set the PG_uptodate flag if we see that a write request
161  * covers the full page.
162  */
163 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
164 {
165         if (PageUptodate(page))
166                 return;
167         if (base != 0)
168                 return;
169         if (count != nfs_page_length(page))
170                 return;
171         SetPageUptodate(page);
172 }
173
174 static int wb_priority(struct writeback_control *wbc)
175 {
176         if (wbc->for_reclaim)
177                 return FLUSH_HIGHPRI | FLUSH_STABLE;
178         if (wbc->for_kupdate)
179                 return FLUSH_LOWPRI;
180         return 0;
181 }
182
183 /*
184  * NFS congestion control
185  */
186
187 int nfs_congestion_kb;
188
189 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
190 #define NFS_CONGESTION_OFF_THRESH       \
191         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
192
193 static int nfs_set_page_writeback(struct page *page)
194 {
195         int ret = test_set_page_writeback(page);
196
197         if (!ret) {
198                 struct inode *inode = page->mapping->host;
199                 struct nfs_server *nfss = NFS_SERVER(inode);
200
201                 if (atomic_long_inc_return(&nfss->writeback) >
202                                 NFS_CONGESTION_ON_THRESH)
203                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
204         }
205         return ret;
206 }
207
208 static void nfs_end_page_writeback(struct page *page)
209 {
210         struct inode *inode = page->mapping->host;
211         struct nfs_server *nfss = NFS_SERVER(inode);
212
213         end_page_writeback(page);
214         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
215                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
216 }
217
218 /*
219  * Find an associated nfs write request, and prepare to flush it out
220  * May return an error if the user signalled nfs_wait_on_request().
221  */
222 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
223                                 struct page *page)
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                         spin_unlock(&inode->i_lock);
234                         return 0;
235                 }
236                 if (nfs_set_page_tag_locked(req))
237                         break;
238                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
239                  *       then the call to nfs_set_page_tag_locked() will always
240                  *       succeed provided that someone hasn't already marked the
241                  *       request as dirty (in which case we don't care).
242                  */
243                 spin_unlock(&inode->i_lock);
244                 ret = nfs_wait_on_request(req);
245                 nfs_release_request(req);
246                 if (ret != 0)
247                         return ret;
248                 spin_lock(&inode->i_lock);
249         }
250         if (test_bit(PG_CLEAN, &req->wb_flags)) {
251                 spin_unlock(&inode->i_lock);
252                 BUG();
253         }
254         if (nfs_set_page_writeback(page) != 0) {
255                 spin_unlock(&inode->i_lock);
256                 BUG();
257         }
258         spin_unlock(&inode->i_lock);
259         if (!nfs_pageio_add_request(pgio, req)) {
260                 nfs_redirty_request(req);
261                 return pgio->pg_error;
262         }
263         return 0;
264 }
265
266 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
267 {
268         struct inode *inode = page->mapping->host;
269
270         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
271         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
272
273         nfs_pageio_cond_complete(pgio, page->index);
274         return nfs_page_async_flush(pgio, page);
275 }
276
277 /*
278  * Write an mmapped page to the server.
279  */
280 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
281 {
282         struct nfs_pageio_descriptor pgio;
283         int err;
284
285         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
286         err = nfs_do_writepage(page, wbc, &pgio);
287         nfs_pageio_complete(&pgio);
288         if (err < 0)
289                 return err;
290         if (pgio.pg_error < 0)
291                 return pgio.pg_error;
292         return 0;
293 }
294
295 int nfs_writepage(struct page *page, struct writeback_control *wbc)
296 {
297         int ret;
298
299         ret = nfs_writepage_locked(page, wbc);
300         unlock_page(page);
301         return ret;
302 }
303
304 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
305 {
306         int ret;
307
308         ret = nfs_do_writepage(page, wbc, data);
309         unlock_page(page);
310         return ret;
311 }
312
313 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
314 {
315         struct inode *inode = mapping->host;
316         unsigned long *bitlock = &NFS_I(inode)->flags;
317         struct nfs_pageio_descriptor pgio;
318         int err;
319
320         /* Stop dirtying of new pages while we sync */
321         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
322                         nfs_wait_bit_killable, TASK_KILLABLE);
323         if (err)
324                 goto out_err;
325
326         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
327
328         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
329         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
330         nfs_pageio_complete(&pgio);
331
332         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
333         smp_mb__after_clear_bit();
334         wake_up_bit(bitlock, NFS_INO_FLUSHING);
335
336         if (err < 0)
337                 goto out_err;
338         err = pgio.pg_error;
339         if (err < 0)
340                 goto out_err;
341         return 0;
342 out_err:
343         return err;
344 }
345
346 /*
347  * Insert a write request into an inode
348  */
349 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
350 {
351         struct nfs_inode *nfsi = NFS_I(inode);
352         int error;
353
354         error = radix_tree_preload(GFP_NOFS);
355         if (error != 0)
356                 goto out;
357
358         /* Lock the request! */
359         nfs_lock_request_dontget(req);
360
361         spin_lock(&inode->i_lock);
362         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
363         BUG_ON(error);
364         if (!nfsi->npages) {
365                 igrab(inode);
366                 if (nfs_have_delegation(inode, FMODE_WRITE))
367                         nfsi->change_attr++;
368         }
369         SetPagePrivate(req->wb_page);
370         set_page_private(req->wb_page, (unsigned long)req);
371         nfsi->npages++;
372         kref_get(&req->wb_kref);
373         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
374                                 NFS_PAGE_TAG_LOCKED);
375         spin_unlock(&inode->i_lock);
376         radix_tree_preload_end();
377 out:
378         return error;
379 }
380
381 /*
382  * Remove a write request from an inode
383  */
384 static void nfs_inode_remove_request(struct nfs_page *req)
385 {
386         struct inode *inode = req->wb_context->path.dentry->d_inode;
387         struct nfs_inode *nfsi = NFS_I(inode);
388
389         BUG_ON (!NFS_WBACK_BUSY(req));
390
391         spin_lock(&inode->i_lock);
392         set_page_private(req->wb_page, 0);
393         ClearPagePrivate(req->wb_page);
394         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
395         nfsi->npages--;
396         if (!nfsi->npages) {
397                 spin_unlock(&inode->i_lock);
398                 iput(inode);
399         } else
400                 spin_unlock(&inode->i_lock);
401         nfs_clear_request(req);
402         nfs_release_request(req);
403 }
404
405 static void
406 nfs_mark_request_dirty(struct nfs_page *req)
407 {
408         __set_page_dirty_nobuffers(req->wb_page);
409 }
410
411 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
412 /*
413  * Add a request to the inode's commit list.
414  */
415 static void
416 nfs_mark_request_commit(struct nfs_page *req)
417 {
418         struct inode *inode = req->wb_context->path.dentry->d_inode;
419         struct nfs_inode *nfsi = NFS_I(inode);
420
421         spin_lock(&inode->i_lock);
422         set_bit(PG_CLEAN, &(req)->wb_flags);
423         radix_tree_tag_set(&nfsi->nfs_page_tree,
424                         req->wb_index,
425                         NFS_PAGE_TAG_COMMIT);
426         spin_unlock(&inode->i_lock);
427         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
428         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
429         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
430 }
431
432 static int
433 nfs_clear_request_commit(struct nfs_page *req)
434 {
435         struct page *page = req->wb_page;
436
437         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
438                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
439                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
440                 return 1;
441         }
442         return 0;
443 }
444
445 static inline
446 int nfs_write_need_commit(struct nfs_write_data *data)
447 {
448         return data->verf.committed != NFS_FILE_SYNC;
449 }
450
451 static inline
452 int nfs_reschedule_unstable_write(struct nfs_page *req)
453 {
454         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
455                 nfs_mark_request_commit(req);
456                 return 1;
457         }
458         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
459                 nfs_mark_request_dirty(req);
460                 return 1;
461         }
462         return 0;
463 }
464 #else
465 static inline void
466 nfs_mark_request_commit(struct nfs_page *req)
467 {
468 }
469
470 static inline int
471 nfs_clear_request_commit(struct nfs_page *req)
472 {
473         return 0;
474 }
475
476 static inline
477 int nfs_write_need_commit(struct nfs_write_data *data)
478 {
479         return 0;
480 }
481
482 static inline
483 int nfs_reschedule_unstable_write(struct nfs_page *req)
484 {
485         return 0;
486 }
487 #endif
488
489 /*
490  * Wait for a request to complete.
491  *
492  * Interruptible by fatal signals only.
493  */
494 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
495 {
496         struct nfs_inode *nfsi = NFS_I(inode);
497         struct nfs_page *req;
498         pgoff_t idx_end, next;
499         unsigned int            res = 0;
500         int                     error;
501
502         if (npages == 0)
503                 idx_end = ~0;
504         else
505                 idx_end = idx_start + npages - 1;
506
507         next = idx_start;
508         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
509                 if (req->wb_index > idx_end)
510                         break;
511
512                 next = req->wb_index + 1;
513                 BUG_ON(!NFS_WBACK_BUSY(req));
514
515                 kref_get(&req->wb_kref);
516                 spin_unlock(&inode->i_lock);
517                 error = nfs_wait_on_request(req);
518                 nfs_release_request(req);
519                 spin_lock(&inode->i_lock);
520                 if (error < 0)
521                         return error;
522                 res++;
523         }
524         return res;
525 }
526
527 static void nfs_cancel_commit_list(struct list_head *head)
528 {
529         struct nfs_page *req;
530
531         while(!list_empty(head)) {
532                 req = nfs_list_entry(head->next);
533                 nfs_list_remove_request(req);
534                 nfs_clear_request_commit(req);
535                 nfs_inode_remove_request(req);
536                 nfs_unlock_request(req);
537         }
538 }
539
540 static int
541 nfs_need_commit(struct nfs_inode *nfsi)
542 {
543         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
544 }
545
546 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
547 /*
548  * nfs_scan_commit - Scan an inode for commit requests
549  * @inode: NFS inode to scan
550  * @dst: destination list
551  * @idx_start: lower bound of page->index to scan.
552  * @npages: idx_start + npages sets the upper bound to scan.
553  *
554  * Moves requests from the inode's 'commit' request list.
555  * The requests are *not* checked to ensure that they form a contiguous set.
556  */
557 static int
558 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
559 {
560         struct nfs_inode *nfsi = NFS_I(inode);
561
562         if (!nfs_need_commit(nfsi))
563                 return 0;
564
565         return nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
566 }
567 #else
568 static inline int nfs_need_commit(struct nfs_inode *nfsi)
569 {
570         return 0;
571 }
572
573 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
574 {
575         return 0;
576 }
577 #endif
578
579 /*
580  * Search for an existing write request, and attempt to update
581  * it to reflect a new dirty region on a given page.
582  *
583  * If the attempt fails, then the existing request is flushed out
584  * to disk.
585  */
586 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
587                 struct page *page,
588                 unsigned int offset,
589                 unsigned int bytes)
590 {
591         struct nfs_page *req;
592         unsigned int rqend;
593         unsigned int end;
594         int error;
595
596         if (!PagePrivate(page))
597                 return NULL;
598
599         end = offset + bytes;
600         spin_lock(&inode->i_lock);
601
602         for (;;) {
603                 req = nfs_page_find_request_locked(page);
604                 if (req == NULL)
605                         goto out_unlock;
606
607                 rqend = req->wb_offset + req->wb_bytes;
608                 /*
609                  * Tell the caller to flush out the request if
610                  * the offsets are non-contiguous.
611                  * Note: nfs_flush_incompatible() will already
612                  * have flushed out requests having wrong owners.
613                  */
614                 if (offset > rqend
615                     || end < req->wb_offset)
616                         goto out_flushme;
617
618                 if (nfs_set_page_tag_locked(req))
619                         break;
620
621                 /* The request is locked, so wait and then retry */
622                 spin_unlock(&inode->i_lock);
623                 error = nfs_wait_on_request(req);
624                 nfs_release_request(req);
625                 if (error != 0)
626                         goto out_err;
627                 spin_lock(&inode->i_lock);
628         }
629
630         if (nfs_clear_request_commit(req))
631                 radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
632                                 req->wb_index, NFS_PAGE_TAG_COMMIT);
633
634         /* Okay, the request matches. Update the region */
635         if (offset < req->wb_offset) {
636                 req->wb_offset = offset;
637                 req->wb_pgbase = offset;
638         }
639         if (end > rqend)
640                 req->wb_bytes = end - req->wb_offset;
641         else
642                 req->wb_bytes = rqend - req->wb_offset;
643 out_unlock:
644         spin_unlock(&inode->i_lock);
645         return req;
646 out_flushme:
647         spin_unlock(&inode->i_lock);
648         nfs_release_request(req);
649         error = nfs_wb_page(inode, page);
650 out_err:
651         return ERR_PTR(error);
652 }
653
654 /*
655  * Try to update an existing write request, or create one if there is none.
656  *
657  * Note: Should always be called with the Page Lock held to prevent races
658  * if we have to add a new request. Also assumes that the caller has
659  * already called nfs_flush_incompatible() if necessary.
660  */
661 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
662                 struct page *page, unsigned int offset, unsigned int bytes)
663 {
664         struct inode *inode = page->mapping->host;
665         struct nfs_page *req;
666         int error;
667
668         req = nfs_try_to_update_request(inode, page, offset, bytes);
669         if (req != NULL)
670                 goto out;
671         req = nfs_create_request(ctx, inode, page, offset, bytes);
672         if (IS_ERR(req))
673                 goto out;
674         error = nfs_inode_add_request(inode, req);
675         if (error != 0) {
676                 nfs_release_request(req);
677                 req = ERR_PTR(error);
678         }
679 out:
680         return req;
681 }
682
683 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
684                 unsigned int offset, unsigned int count)
685 {
686         struct nfs_page *req;
687
688         req = nfs_setup_write_request(ctx, page, offset, count);
689         if (IS_ERR(req))
690                 return PTR_ERR(req);
691         /* Update file length */
692         nfs_grow_file(page, offset, count);
693         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
694         nfs_clear_page_tag_locked(req);
695         return 0;
696 }
697
698 int nfs_flush_incompatible(struct file *file, struct page *page)
699 {
700         struct nfs_open_context *ctx = nfs_file_open_context(file);
701         struct nfs_page *req;
702         int do_flush, status;
703         /*
704          * Look for a request corresponding to this page. If there
705          * is one, and it belongs to another file, we flush it out
706          * before we try to copy anything into the page. Do this
707          * due to the lack of an ACCESS-type call in NFSv2.
708          * Also do the same if we find a request from an existing
709          * dropped page.
710          */
711         do {
712                 req = nfs_page_find_request(page);
713                 if (req == NULL)
714                         return 0;
715                 do_flush = req->wb_page != page || req->wb_context != ctx;
716                 nfs_release_request(req);
717                 if (!do_flush)
718                         return 0;
719                 status = nfs_wb_page(page->mapping->host, page);
720         } while (status == 0);
721         return status;
722 }
723
724 /*
725  * If the page cache is marked as unsafe or invalid, then we can't rely on
726  * the PageUptodate() flag. In this case, we will need to turn off
727  * write optimisations that depend on the page contents being correct.
728  */
729 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
730 {
731         return PageUptodate(page) &&
732                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
733 }
734
735 /*
736  * Update and possibly write a cached page of an NFS file.
737  *
738  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
739  * things with a page scheduled for an RPC call (e.g. invalidate it).
740  */
741 int nfs_updatepage(struct file *file, struct page *page,
742                 unsigned int offset, unsigned int count)
743 {
744         struct nfs_open_context *ctx = nfs_file_open_context(file);
745         struct inode    *inode = page->mapping->host;
746         int             status = 0;
747
748         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
749
750         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
751                 file->f_path.dentry->d_parent->d_name.name,
752                 file->f_path.dentry->d_name.name, count,
753                 (long long)(page_offset(page) + offset));
754
755         /* If we're not using byte range locks, and we know the page
756          * is up to date, it may be more efficient to extend the write
757          * to cover the entire page in order to avoid fragmentation
758          * inefficiencies.
759          */
760         if (nfs_write_pageuptodate(page, inode) &&
761                         inode->i_flock == NULL &&
762                         !(file->f_flags & O_SYNC)) {
763                 count = max(count + offset, nfs_page_length(page));
764                 offset = 0;
765         }
766
767         status = nfs_writepage_setup(ctx, page, offset, count);
768         if (status < 0)
769                 nfs_set_pageerror(page);
770         else
771                 __set_page_dirty_nobuffers(page);
772
773         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
774                         status, (long long)i_size_read(inode));
775         return status;
776 }
777
778 static void nfs_writepage_release(struct nfs_page *req)
779 {
780
781         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
782                 nfs_end_page_writeback(req->wb_page);
783                 nfs_inode_remove_request(req);
784         } else
785                 nfs_end_page_writeback(req->wb_page);
786         nfs_clear_page_tag_locked(req);
787 }
788
789 static int flush_task_priority(int how)
790 {
791         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
792                 case FLUSH_HIGHPRI:
793                         return RPC_PRIORITY_HIGH;
794                 case FLUSH_LOWPRI:
795                         return RPC_PRIORITY_LOW;
796         }
797         return RPC_PRIORITY_NORMAL;
798 }
799
800 /*
801  * Set up the argument/result storage required for the RPC call.
802  */
803 static int nfs_write_rpcsetup(struct nfs_page *req,
804                 struct nfs_write_data *data,
805                 const struct rpc_call_ops *call_ops,
806                 unsigned int count, unsigned int offset,
807                 int how)
808 {
809         struct inode *inode = req->wb_context->path.dentry->d_inode;
810         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
811         int priority = flush_task_priority(how);
812         struct rpc_task *task;
813         struct rpc_message msg = {
814                 .rpc_argp = &data->args,
815                 .rpc_resp = &data->res,
816                 .rpc_cred = req->wb_context->cred,
817         };
818         struct rpc_task_setup task_setup_data = {
819                 .rpc_client = NFS_CLIENT(inode),
820                 .task = &data->task,
821                 .rpc_message = &msg,
822                 .callback_ops = call_ops,
823                 .callback_data = data,
824                 .workqueue = nfsiod_workqueue,
825                 .flags = flags,
826                 .priority = priority,
827         };
828
829         /* Set up the RPC argument and reply structs
830          * NB: take care not to mess about with data->commit et al. */
831
832         data->req = req;
833         data->inode = inode = req->wb_context->path.dentry->d_inode;
834         data->cred = msg.rpc_cred;
835
836         data->args.fh     = NFS_FH(inode);
837         data->args.offset = req_offset(req) + offset;
838         data->args.pgbase = req->wb_pgbase + offset;
839         data->args.pages  = data->pagevec;
840         data->args.count  = count;
841         data->args.context = get_nfs_open_context(req->wb_context);
842         data->args.stable  = NFS_UNSTABLE;
843         if (how & FLUSH_STABLE) {
844                 data->args.stable = NFS_DATA_SYNC;
845                 if (!nfs_need_commit(NFS_I(inode)))
846                         data->args.stable = NFS_FILE_SYNC;
847         }
848
849         data->res.fattr   = &data->fattr;
850         data->res.count   = count;
851         data->res.verf    = &data->verf;
852         nfs_fattr_init(&data->fattr);
853
854         /* Set up the initial task struct.  */
855         NFS_PROTO(inode)->write_setup(data, &msg);
856
857         dprintk("NFS: %5u initiated write call "
858                 "(req %s/%lld, %u bytes @ offset %llu)\n",
859                 data->task.tk_pid,
860                 inode->i_sb->s_id,
861                 (long long)NFS_FILEID(inode),
862                 count,
863                 (unsigned long long)data->args.offset);
864
865         task = rpc_run_task(&task_setup_data);
866         if (IS_ERR(task))
867                 return PTR_ERR(task);
868         rpc_put_task(task);
869         return 0;
870 }
871
872 /* If a nfs_flush_* function fails, it should remove reqs from @head and
873  * call this on each, which will prepare them to be retried on next
874  * writeback using standard nfs.
875  */
876 static void nfs_redirty_request(struct nfs_page *req)
877 {
878         nfs_mark_request_dirty(req);
879         nfs_end_page_writeback(req->wb_page);
880         nfs_clear_page_tag_locked(req);
881 }
882
883 /*
884  * Generate multiple small requests to write out a single
885  * contiguous dirty area on one page.
886  */
887 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
888 {
889         struct nfs_page *req = nfs_list_entry(head->next);
890         struct page *page = req->wb_page;
891         struct nfs_write_data *data;
892         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
893         unsigned int offset;
894         int requests = 0;
895         int ret = 0;
896         LIST_HEAD(list);
897
898         nfs_list_remove_request(req);
899
900         nbytes = count;
901         do {
902                 size_t len = min(nbytes, wsize);
903
904                 data = nfs_writedata_alloc(1);
905                 if (!data)
906                         goto out_bad;
907                 list_add(&data->pages, &list);
908                 requests++;
909                 nbytes -= len;
910         } while (nbytes != 0);
911         atomic_set(&req->wb_complete, requests);
912
913         ClearPageError(page);
914         offset = 0;
915         nbytes = count;
916         do {
917                 int ret2;
918
919                 data = list_entry(list.next, struct nfs_write_data, pages);
920                 list_del_init(&data->pages);
921
922                 data->pagevec[0] = page;
923
924                 if (nbytes < wsize)
925                         wsize = nbytes;
926                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
927                                    wsize, offset, how);
928                 if (ret == 0)
929                         ret = ret2;
930                 offset += wsize;
931                 nbytes -= wsize;
932         } while (nbytes != 0);
933
934         return ret;
935
936 out_bad:
937         while (!list_empty(&list)) {
938                 data = list_entry(list.next, struct nfs_write_data, pages);
939                 list_del(&data->pages);
940                 nfs_writedata_release(data);
941         }
942         nfs_redirty_request(req);
943         return -ENOMEM;
944 }
945
946 /*
947  * Create an RPC task for the given write request and kick it.
948  * The page must have been locked by the caller.
949  *
950  * It may happen that the page we're passed is not marked dirty.
951  * This is the case if nfs_updatepage detects a conflicting request
952  * that has been written but not committed.
953  */
954 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
955 {
956         struct nfs_page         *req;
957         struct page             **pages;
958         struct nfs_write_data   *data;
959
960         data = nfs_writedata_alloc(npages);
961         if (!data)
962                 goto out_bad;
963
964         pages = data->pagevec;
965         while (!list_empty(head)) {
966                 req = nfs_list_entry(head->next);
967                 nfs_list_remove_request(req);
968                 nfs_list_add_request(req, &data->pages);
969                 ClearPageError(req->wb_page);
970                 *pages++ = req->wb_page;
971         }
972         req = nfs_list_entry(data->pages.next);
973
974         /* Set up the argument struct */
975         return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
976  out_bad:
977         while (!list_empty(head)) {
978                 req = nfs_list_entry(head->next);
979                 nfs_list_remove_request(req);
980                 nfs_redirty_request(req);
981         }
982         return -ENOMEM;
983 }
984
985 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
986                                   struct inode *inode, int ioflags)
987 {
988         size_t wsize = NFS_SERVER(inode)->wsize;
989
990         if (wsize < PAGE_CACHE_SIZE)
991                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
992         else
993                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
994 }
995
996 /*
997  * Handle a write reply that flushed part of a page.
998  */
999 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1000 {
1001         struct nfs_write_data   *data = calldata;
1002
1003         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1004                 task->tk_pid,
1005                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1006                 (long long)
1007                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1008                 data->req->wb_bytes, (long long)req_offset(data->req));
1009
1010         nfs_writeback_done(task, data);
1011 }
1012
1013 static void nfs_writeback_release_partial(void *calldata)
1014 {
1015         struct nfs_write_data   *data = calldata;
1016         struct nfs_page         *req = data->req;
1017         struct page             *page = req->wb_page;
1018         int status = data->task.tk_status;
1019
1020         if (status < 0) {
1021                 nfs_set_pageerror(page);
1022                 nfs_context_set_write_error(req->wb_context, status);
1023                 dprintk(", error = %d\n", status);
1024                 goto out;
1025         }
1026
1027         if (nfs_write_need_commit(data)) {
1028                 struct inode *inode = page->mapping->host;
1029
1030                 spin_lock(&inode->i_lock);
1031                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1032                         /* Do nothing we need to resend the writes */
1033                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1034                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1035                         dprintk(" defer commit\n");
1036                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1037                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1038                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1039                         dprintk(" server reboot detected\n");
1040                 }
1041                 spin_unlock(&inode->i_lock);
1042         } else
1043                 dprintk(" OK\n");
1044
1045 out:
1046         if (atomic_dec_and_test(&req->wb_complete))
1047                 nfs_writepage_release(req);
1048         nfs_writedata_release(calldata);
1049 }
1050
1051 static const struct rpc_call_ops nfs_write_partial_ops = {
1052         .rpc_call_done = nfs_writeback_done_partial,
1053         .rpc_release = nfs_writeback_release_partial,
1054 };
1055
1056 /*
1057  * Handle a write reply that flushes a whole page.
1058  *
1059  * FIXME: There is an inherent race with invalidate_inode_pages and
1060  *        writebacks since the page->count is kept > 1 for as long
1061  *        as the page has a write request pending.
1062  */
1063 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1064 {
1065         struct nfs_write_data   *data = calldata;
1066
1067         nfs_writeback_done(task, data);
1068 }
1069
1070 static void nfs_writeback_release_full(void *calldata)
1071 {
1072         struct nfs_write_data   *data = calldata;
1073         int status = data->task.tk_status;
1074
1075         /* Update attributes as result of writeback. */
1076         while (!list_empty(&data->pages)) {
1077                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1078                 struct page *page = req->wb_page;
1079
1080                 nfs_list_remove_request(req);
1081
1082                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1083                         data->task.tk_pid,
1084                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1085                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1086                         req->wb_bytes,
1087                         (long long)req_offset(req));
1088
1089                 if (status < 0) {
1090                         nfs_set_pageerror(page);
1091                         nfs_context_set_write_error(req->wb_context, status);
1092                         dprintk(", error = %d\n", status);
1093                         goto remove_request;
1094                 }
1095
1096                 if (nfs_write_need_commit(data)) {
1097                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1098                         nfs_mark_request_commit(req);
1099                         nfs_end_page_writeback(page);
1100                         dprintk(" marked for commit\n");
1101                         goto next;
1102                 }
1103                 dprintk(" OK\n");
1104 remove_request:
1105                 nfs_end_page_writeback(page);
1106                 nfs_inode_remove_request(req);
1107         next:
1108                 nfs_clear_page_tag_locked(req);
1109         }
1110         nfs_writedata_release(calldata);
1111 }
1112
1113 static const struct rpc_call_ops nfs_write_full_ops = {
1114         .rpc_call_done = nfs_writeback_done_full,
1115         .rpc_release = nfs_writeback_release_full,
1116 };
1117
1118
1119 /*
1120  * This function is called when the WRITE call is complete.
1121  */
1122 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1123 {
1124         struct nfs_writeargs    *argp = &data->args;
1125         struct nfs_writeres     *resp = &data->res;
1126         int status;
1127
1128         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1129                 task->tk_pid, task->tk_status);
1130
1131         /*
1132          * ->write_done will attempt to use post-op attributes to detect
1133          * conflicting writes by other clients.  A strict interpretation
1134          * of close-to-open would allow us to continue caching even if
1135          * another writer had changed the file, but some applications
1136          * depend on tighter cache coherency when writing.
1137          */
1138         status = NFS_PROTO(data->inode)->write_done(task, data);
1139         if (status != 0)
1140                 return status;
1141         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1142
1143 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1144         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1145                 /* We tried a write call, but the server did not
1146                  * commit data to stable storage even though we
1147                  * requested it.
1148                  * Note: There is a known bug in Tru64 < 5.0 in which
1149                  *       the server reports NFS_DATA_SYNC, but performs
1150                  *       NFS_FILE_SYNC. We therefore implement this checking
1151                  *       as a dprintk() in order to avoid filling syslog.
1152                  */
1153                 static unsigned long    complain;
1154
1155                 if (time_before(complain, jiffies)) {
1156                         dprintk("NFS:       faulty NFS server %s:"
1157                                 " (committed = %d) != (stable = %d)\n",
1158                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1159                                 resp->verf->committed, argp->stable);
1160                         complain = jiffies + 300 * HZ;
1161                 }
1162         }
1163 #endif
1164         /* Is this a short write? */
1165         if (task->tk_status >= 0 && resp->count < argp->count) {
1166                 static unsigned long    complain;
1167
1168                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1169
1170                 /* Has the server at least made some progress? */
1171                 if (resp->count != 0) {
1172                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1173                         if (resp->verf->committed != NFS_UNSTABLE) {
1174                                 /* Resend from where the server left off */
1175                                 argp->offset += resp->count;
1176                                 argp->pgbase += resp->count;
1177                                 argp->count -= resp->count;
1178                         } else {
1179                                 /* Resend as a stable write in order to avoid
1180                                  * headaches in the case of a server crash.
1181                                  */
1182                                 argp->stable = NFS_FILE_SYNC;
1183                         }
1184                         rpc_restart_call(task);
1185                         return -EAGAIN;
1186                 }
1187                 if (time_before(complain, jiffies)) {
1188                         printk(KERN_WARNING
1189                                "NFS: Server wrote zero bytes, expected %u.\n",
1190                                         argp->count);
1191                         complain = jiffies + 300 * HZ;
1192                 }
1193                 /* Can't do anything about it except throw an error. */
1194                 task->tk_status = -EIO;
1195         }
1196         return 0;
1197 }
1198
1199
1200 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1201 void nfs_commitdata_release(void *data)
1202 {
1203         struct nfs_write_data *wdata = data;
1204
1205         put_nfs_open_context(wdata->args.context);
1206         nfs_commit_free(wdata);
1207 }
1208
1209 /*
1210  * Set up the argument/result storage required for the RPC call.
1211  */
1212 static int nfs_commit_rpcsetup(struct list_head *head,
1213                 struct nfs_write_data *data,
1214                 int how)
1215 {
1216         struct nfs_page *first = nfs_list_entry(head->next);
1217         struct inode *inode = first->wb_context->path.dentry->d_inode;
1218         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1219         int priority = flush_task_priority(how);
1220         struct rpc_task *task;
1221         struct rpc_message msg = {
1222                 .rpc_argp = &data->args,
1223                 .rpc_resp = &data->res,
1224                 .rpc_cred = first->wb_context->cred,
1225         };
1226         struct rpc_task_setup task_setup_data = {
1227                 .task = &data->task,
1228                 .rpc_client = NFS_CLIENT(inode),
1229                 .rpc_message = &msg,
1230                 .callback_ops = &nfs_commit_ops,
1231                 .callback_data = data,
1232                 .workqueue = nfsiod_workqueue,
1233                 .flags = flags,
1234                 .priority = priority,
1235         };
1236
1237         /* Set up the RPC argument and reply structs
1238          * NB: take care not to mess about with data->commit et al. */
1239
1240         list_splice_init(head, &data->pages);
1241
1242         data->inode       = inode;
1243         data->cred        = msg.rpc_cred;
1244
1245         data->args.fh     = NFS_FH(data->inode);
1246         /* Note: we always request a commit of the entire inode */
1247         data->args.offset = 0;
1248         data->args.count  = 0;
1249         data->args.context = get_nfs_open_context(first->wb_context);
1250         data->res.count   = 0;
1251         data->res.fattr   = &data->fattr;
1252         data->res.verf    = &data->verf;
1253         nfs_fattr_init(&data->fattr);
1254
1255         /* Set up the initial task struct.  */
1256         NFS_PROTO(inode)->commit_setup(data, &msg);
1257
1258         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1259
1260         task = rpc_run_task(&task_setup_data);
1261         if (IS_ERR(task))
1262                 return PTR_ERR(task);
1263         rpc_put_task(task);
1264         return 0;
1265 }
1266
1267 /*
1268  * Commit dirty pages
1269  */
1270 static int
1271 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1272 {
1273         struct nfs_write_data   *data;
1274         struct nfs_page         *req;
1275
1276         data = nfs_commitdata_alloc();
1277
1278         if (!data)
1279                 goto out_bad;
1280
1281         /* Set up the argument struct */
1282         return nfs_commit_rpcsetup(head, data, how);
1283  out_bad:
1284         while (!list_empty(head)) {
1285                 req = nfs_list_entry(head->next);
1286                 nfs_list_remove_request(req);
1287                 nfs_mark_request_commit(req);
1288                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1289                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1290                                 BDI_RECLAIMABLE);
1291                 nfs_clear_page_tag_locked(req);
1292         }
1293         return -ENOMEM;
1294 }
1295
1296 /*
1297  * COMMIT call returned
1298  */
1299 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1300 {
1301         struct nfs_write_data   *data = calldata;
1302
1303         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1304                                 task->tk_pid, task->tk_status);
1305
1306         /* Call the NFS version-specific code */
1307         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1308                 return;
1309 }
1310
1311 static void nfs_commit_release(void *calldata)
1312 {
1313         struct nfs_write_data   *data = calldata;
1314         struct nfs_page         *req;
1315         int status = data->task.tk_status;
1316
1317         while (!list_empty(&data->pages)) {
1318                 req = nfs_list_entry(data->pages.next);
1319                 nfs_list_remove_request(req);
1320                 nfs_clear_request_commit(req);
1321
1322                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1323                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1324                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1325                         req->wb_bytes,
1326                         (long long)req_offset(req));
1327                 if (status < 0) {
1328                         nfs_context_set_write_error(req->wb_context, status);
1329                         nfs_inode_remove_request(req);
1330                         dprintk(", error = %d\n", status);
1331                         goto next;
1332                 }
1333
1334                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1335                  * returned by the server against all stored verfs. */
1336                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1337                         /* We have a match */
1338                         nfs_inode_remove_request(req);
1339                         dprintk(" OK\n");
1340                         goto next;
1341                 }
1342                 /* We have a mismatch. Write the page again */
1343                 dprintk(" mismatch\n");
1344                 nfs_mark_request_dirty(req);
1345         next:
1346                 nfs_clear_page_tag_locked(req);
1347         }
1348         nfs_commitdata_release(calldata);
1349 }
1350
1351 static const struct rpc_call_ops nfs_commit_ops = {
1352         .rpc_call_done = nfs_commit_done,
1353         .rpc_release = nfs_commit_release,
1354 };
1355
1356 int nfs_commit_inode(struct inode *inode, int how)
1357 {
1358         LIST_HEAD(head);
1359         int res;
1360
1361         spin_lock(&inode->i_lock);
1362         res = nfs_scan_commit(inode, &head, 0, 0);
1363         spin_unlock(&inode->i_lock);
1364         if (res) {
1365                 int error = nfs_commit_list(inode, &head, how);
1366                 if (error < 0)
1367                         return error;
1368         }
1369         return res;
1370 }
1371 #else
1372 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1373 {
1374         return 0;
1375 }
1376 #endif
1377
1378 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1379 {
1380         struct inode *inode = mapping->host;
1381         pgoff_t idx_start, idx_end;
1382         unsigned int npages = 0;
1383         LIST_HEAD(head);
1384         int nocommit = how & FLUSH_NOCOMMIT;
1385         long pages, ret;
1386
1387         /* FIXME */
1388         if (wbc->range_cyclic)
1389                 idx_start = 0;
1390         else {
1391                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1392                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1393                 if (idx_end > idx_start) {
1394                         pgoff_t l_npages = 1 + idx_end - idx_start;
1395                         npages = l_npages;
1396                         if (sizeof(npages) != sizeof(l_npages) &&
1397                                         (pgoff_t)npages != l_npages)
1398                                 npages = 0;
1399                 }
1400         }
1401         how &= ~FLUSH_NOCOMMIT;
1402         spin_lock(&inode->i_lock);
1403         do {
1404                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1405                 if (ret != 0)
1406                         continue;
1407                 if (nocommit)
1408                         break;
1409                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1410                 if (pages == 0)
1411                         break;
1412                 if (how & FLUSH_INVALIDATE) {
1413                         spin_unlock(&inode->i_lock);
1414                         nfs_cancel_commit_list(&head);
1415                         ret = pages;
1416                         spin_lock(&inode->i_lock);
1417                         continue;
1418                 }
1419                 pages += nfs_scan_commit(inode, &head, 0, 0);
1420                 spin_unlock(&inode->i_lock);
1421                 ret = nfs_commit_list(inode, &head, how);
1422                 spin_lock(&inode->i_lock);
1423
1424         } while (ret >= 0);
1425         spin_unlock(&inode->i_lock);
1426         return ret;
1427 }
1428
1429 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1430 {
1431         int ret;
1432
1433         ret = nfs_writepages(mapping, wbc);
1434         if (ret < 0)
1435                 goto out;
1436         ret = nfs_sync_mapping_wait(mapping, wbc, how);
1437         if (ret < 0)
1438                 goto out;
1439         return 0;
1440 out:
1441         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1442         return ret;
1443 }
1444
1445 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1446 static int nfs_write_mapping(struct address_space *mapping, int how)
1447 {
1448         struct writeback_control wbc = {
1449                 .bdi = mapping->backing_dev_info,
1450                 .sync_mode = WB_SYNC_ALL,
1451                 .nr_to_write = LONG_MAX,
1452                 .range_start = 0,
1453                 .range_end = LLONG_MAX,
1454                 .for_writepages = 1,
1455         };
1456
1457         return __nfs_write_mapping(mapping, &wbc, how);
1458 }
1459
1460 /*
1461  * flush the inode to disk.
1462  */
1463 int nfs_wb_all(struct inode *inode)
1464 {
1465         return nfs_write_mapping(inode->i_mapping, 0);
1466 }
1467
1468 int nfs_wb_nocommit(struct inode *inode)
1469 {
1470         return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1471 }
1472
1473 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1474 {
1475         struct nfs_page *req;
1476         loff_t range_start = page_offset(page);
1477         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1478         struct writeback_control wbc = {
1479                 .bdi = page->mapping->backing_dev_info,
1480                 .sync_mode = WB_SYNC_ALL,
1481                 .nr_to_write = LONG_MAX,
1482                 .range_start = range_start,
1483                 .range_end = range_end,
1484         };
1485         int ret = 0;
1486
1487         BUG_ON(!PageLocked(page));
1488         for (;;) {
1489                 req = nfs_page_find_request(page);
1490                 if (req == NULL)
1491                         goto out;
1492                 if (test_bit(PG_CLEAN, &req->wb_flags)) {
1493                         nfs_release_request(req);
1494                         break;
1495                 }
1496                 if (nfs_lock_request_dontget(req)) {
1497                         nfs_inode_remove_request(req);
1498                         /*
1499                          * In case nfs_inode_remove_request has marked the
1500                          * page as being dirty
1501                          */
1502                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1503                         nfs_unlock_request(req);
1504                         break;
1505                 }
1506                 ret = nfs_wait_on_request(req);
1507                 if (ret < 0)
1508                         goto out;
1509         }
1510         if (!PagePrivate(page))
1511                 return 0;
1512         ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1513 out:
1514         return ret;
1515 }
1516
1517 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1518                                 int how)
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                 .bdi = page->mapping->backing_dev_info,
1524                 .sync_mode = WB_SYNC_ALL,
1525                 .nr_to_write = LONG_MAX,
1526                 .range_start = range_start,
1527                 .range_end = range_end,
1528         };
1529         int ret;
1530
1531         do {
1532                 if (clear_page_dirty_for_io(page)) {
1533                         ret = nfs_writepage_locked(page, &wbc);
1534                         if (ret < 0)
1535                                 goto out_error;
1536                 } else if (!PagePrivate(page))
1537                         break;
1538                 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1539                 if (ret < 0)
1540                         goto out_error;
1541         } while (PagePrivate(page));
1542         return 0;
1543 out_error:
1544         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1545         return ret;
1546 }
1547
1548 /*
1549  * Write back all requests on one page - we do this before reading it.
1550  */
1551 int nfs_wb_page(struct inode *inode, struct page* page)
1552 {
1553         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1554 }
1555
1556 int __init nfs_init_writepagecache(void)
1557 {
1558         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1559                                              sizeof(struct nfs_write_data),
1560                                              0, SLAB_HWCACHE_ALIGN,
1561                                              NULL);
1562         if (nfs_wdata_cachep == NULL)
1563                 return -ENOMEM;
1564
1565         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1566                                                      nfs_wdata_cachep);
1567         if (nfs_wdata_mempool == NULL)
1568                 return -ENOMEM;
1569
1570         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1571                                                       nfs_wdata_cachep);
1572         if (nfs_commit_mempool == NULL)
1573                 return -ENOMEM;
1574
1575         /*
1576          * NFS congestion size, scale with available memory.
1577          *
1578          *  64MB:    8192k
1579          * 128MB:   11585k
1580          * 256MB:   16384k
1581          * 512MB:   23170k
1582          *   1GB:   32768k
1583          *   2GB:   46340k
1584          *   4GB:   65536k
1585          *   8GB:   92681k
1586          *  16GB:  131072k
1587          *
1588          * This allows larger machines to have larger/more transfers.
1589          * Limit the default to 256M
1590          */
1591         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1592         if (nfs_congestion_kb > 256*1024)
1593                 nfs_congestion_kb = 256*1024;
1594
1595         return 0;
1596 }
1597
1598 void nfs_destroy_writepagecache(void)
1599 {
1600         mempool_destroy(nfs_commit_mempool);
1601         mempool_destroy(nfs_wdata_mempool);
1602         kmem_cache_destroy(nfs_wdata_cachep);
1603 }
1604