Linux 2.6.31-rc6
[linux-2.6] / 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 #include "nfs4_fs.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 void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
39                                   struct inode *inode, int ioflags);
40 static void nfs_redirty_request(struct nfs_page *req);
41 static const struct rpc_call_ops nfs_write_partial_ops;
42 static const struct rpc_call_ops nfs_write_full_ops;
43 static const struct rpc_call_ops nfs_commit_ops;
44
45 static struct kmem_cache *nfs_wdata_cachep;
46 static mempool_t *nfs_wdata_mempool;
47 static mempool_t *nfs_commit_mempool;
48
49 struct nfs_write_data *nfs_commitdata_alloc(void)
50 {
51         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
52
53         if (p) {
54                 memset(p, 0, sizeof(*p));
55                 INIT_LIST_HEAD(&p->pages);
56                 p->res.seq_res.sr_slotid = NFS4_MAX_SLOT_TABLE;
57         }
58         return p;
59 }
60
61 void nfs_commit_free(struct nfs_write_data *p)
62 {
63         if (p && (p->pagevec != &p->page_array[0]))
64                 kfree(p->pagevec);
65         mempool_free(p, nfs_commit_mempool);
66 }
67
68 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
69 {
70         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
71
72         if (p) {
73                 memset(p, 0, sizeof(*p));
74                 INIT_LIST_HEAD(&p->pages);
75                 p->npages = pagecount;
76                 p->res.seq_res.sr_slotid = NFS4_MAX_SLOT_TABLE;
77                 if (pagecount <= ARRAY_SIZE(p->page_array))
78                         p->pagevec = p->page_array;
79                 else {
80                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
81                         if (!p->pagevec) {
82                                 mempool_free(p, nfs_wdata_mempool);
83                                 p = NULL;
84                         }
85                 }
86         }
87         return p;
88 }
89
90 void nfs_writedata_free(struct nfs_write_data *p)
91 {
92         if (p && (p->pagevec != &p->page_array[0]))
93                 kfree(p->pagevec);
94         mempool_free(p, nfs_wdata_mempool);
95 }
96
97 static void nfs_writedata_release(struct nfs_write_data *wdata)
98 {
99         put_nfs_open_context(wdata->args.context);
100         nfs_writedata_free(wdata);
101 }
102
103 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
104 {
105         ctx->error = error;
106         smp_wmb();
107         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
108 }
109
110 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
111 {
112         struct nfs_page *req = NULL;
113
114         if (PagePrivate(page)) {
115                 req = (struct nfs_page *)page_private(page);
116                 if (req != NULL)
117                         kref_get(&req->wb_kref);
118         }
119         return req;
120 }
121
122 static struct nfs_page *nfs_page_find_request(struct page *page)
123 {
124         struct inode *inode = page->mapping->host;
125         struct nfs_page *req = NULL;
126
127         spin_lock(&inode->i_lock);
128         req = nfs_page_find_request_locked(page);
129         spin_unlock(&inode->i_lock);
130         return req;
131 }
132
133 /* Adjust the file length if we're writing beyond the end */
134 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
135 {
136         struct inode *inode = page->mapping->host;
137         loff_t end, i_size;
138         pgoff_t end_index;
139
140         spin_lock(&inode->i_lock);
141         i_size = i_size_read(inode);
142         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
143         if (i_size > 0 && page->index < end_index)
144                 goto out;
145         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
146         if (i_size >= end)
147                 goto out;
148         i_size_write(inode, end);
149         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
150 out:
151         spin_unlock(&inode->i_lock);
152 }
153
154 /* A writeback failed: mark the page as bad, and invalidate the page cache */
155 static void nfs_set_pageerror(struct page *page)
156 {
157         SetPageError(page);
158         nfs_zap_mapping(page->mapping->host, page->mapping);
159 }
160
161 /* We can set the PG_uptodate flag if we see that a write request
162  * covers the full page.
163  */
164 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
165 {
166         if (PageUptodate(page))
167                 return;
168         if (base != 0)
169                 return;
170         if (count != nfs_page_length(page))
171                 return;
172         SetPageUptodate(page);
173 }
174
175 static int wb_priority(struct writeback_control *wbc)
176 {
177         if (wbc->for_reclaim)
178                 return FLUSH_HIGHPRI | FLUSH_STABLE;
179         if (wbc->for_kupdate)
180                 return FLUSH_LOWPRI;
181         return 0;
182 }
183
184 /*
185  * NFS congestion control
186  */
187
188 int nfs_congestion_kb;
189
190 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
191 #define NFS_CONGESTION_OFF_THRESH       \
192         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
193
194 static int nfs_set_page_writeback(struct page *page)
195 {
196         int ret = test_set_page_writeback(page);
197
198         if (!ret) {
199                 struct inode *inode = page->mapping->host;
200                 struct nfs_server *nfss = NFS_SERVER(inode);
201
202                 if (atomic_long_inc_return(&nfss->writeback) >
203                                 NFS_CONGESTION_ON_THRESH) {
204                         set_bdi_congested(&nfss->backing_dev_info,
205                                                 BLK_RW_ASYNC);
206                 }
207         }
208         return ret;
209 }
210
211 static void nfs_end_page_writeback(struct page *page)
212 {
213         struct inode *inode = page->mapping->host;
214         struct nfs_server *nfss = NFS_SERVER(inode);
215
216         end_page_writeback(page);
217         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
218                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
219 }
220
221 /*
222  * Find an associated nfs write request, and prepare to flush it out
223  * May return an error if the user signalled nfs_wait_on_request().
224  */
225 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
226                                 struct page *page)
227 {
228         struct inode *inode = page->mapping->host;
229         struct nfs_page *req;
230         int ret;
231
232         spin_lock(&inode->i_lock);
233         for(;;) {
234                 req = nfs_page_find_request_locked(page);
235                 if (req == NULL) {
236                         spin_unlock(&inode->i_lock);
237                         return 0;
238                 }
239                 if (nfs_set_page_tag_locked(req))
240                         break;
241                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
242                  *       then the call to nfs_set_page_tag_locked() will always
243                  *       succeed provided that someone hasn't already marked the
244                  *       request as dirty (in which case we don't care).
245                  */
246                 spin_unlock(&inode->i_lock);
247                 ret = nfs_wait_on_request(req);
248                 nfs_release_request(req);
249                 if (ret != 0)
250                         return ret;
251                 spin_lock(&inode->i_lock);
252         }
253         if (test_bit(PG_CLEAN, &req->wb_flags)) {
254                 spin_unlock(&inode->i_lock);
255                 BUG();
256         }
257         if (nfs_set_page_writeback(page) != 0) {
258                 spin_unlock(&inode->i_lock);
259                 BUG();
260         }
261         spin_unlock(&inode->i_lock);
262         if (!nfs_pageio_add_request(pgio, req)) {
263                 nfs_redirty_request(req);
264                 return pgio->pg_error;
265         }
266         return 0;
267 }
268
269 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
270 {
271         struct inode *inode = page->mapping->host;
272
273         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
274         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
275
276         nfs_pageio_cond_complete(pgio, page->index);
277         return nfs_page_async_flush(pgio, page);
278 }
279
280 /*
281  * Write an mmapped page to the server.
282  */
283 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
284 {
285         struct nfs_pageio_descriptor pgio;
286         int err;
287
288         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
289         err = nfs_do_writepage(page, wbc, &pgio);
290         nfs_pageio_complete(&pgio);
291         if (err < 0)
292                 return err;
293         if (pgio.pg_error < 0)
294                 return pgio.pg_error;
295         return 0;
296 }
297
298 int nfs_writepage(struct page *page, struct writeback_control *wbc)
299 {
300         int ret;
301
302         ret = nfs_writepage_locked(page, wbc);
303         unlock_page(page);
304         return ret;
305 }
306
307 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
308 {
309         int ret;
310
311         ret = nfs_do_writepage(page, wbc, data);
312         unlock_page(page);
313         return ret;
314 }
315
316 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
317 {
318         struct inode *inode = mapping->host;
319         unsigned long *bitlock = &NFS_I(inode)->flags;
320         struct nfs_pageio_descriptor pgio;
321         int err;
322
323         /* Stop dirtying of new pages while we sync */
324         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
325                         nfs_wait_bit_killable, TASK_KILLABLE);
326         if (err)
327                 goto out_err;
328
329         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
330
331         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
332         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
333         nfs_pageio_complete(&pgio);
334
335         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
336         smp_mb__after_clear_bit();
337         wake_up_bit(bitlock, NFS_INO_FLUSHING);
338
339         if (err < 0)
340                 goto out_err;
341         err = pgio.pg_error;
342         if (err < 0)
343                 goto out_err;
344         return 0;
345 out_err:
346         return err;
347 }
348
349 /*
350  * Insert a write request into an inode
351  */
352 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
353 {
354         struct nfs_inode *nfsi = NFS_I(inode);
355         int error;
356
357         error = radix_tree_preload(GFP_NOFS);
358         if (error != 0)
359                 goto out;
360
361         /* Lock the request! */
362         nfs_lock_request_dontget(req);
363
364         spin_lock(&inode->i_lock);
365         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
366         BUG_ON(error);
367         if (!nfsi->npages) {
368                 igrab(inode);
369                 if (nfs_have_delegation(inode, FMODE_WRITE))
370                         nfsi->change_attr++;
371         }
372         SetPagePrivate(req->wb_page);
373         set_page_private(req->wb_page, (unsigned long)req);
374         nfsi->npages++;
375         kref_get(&req->wb_kref);
376         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
377                                 NFS_PAGE_TAG_LOCKED);
378         spin_unlock(&inode->i_lock);
379         radix_tree_preload_end();
380 out:
381         return error;
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->path.dentry->d_inode;
390         struct nfs_inode *nfsi = NFS_I(inode);
391
392         BUG_ON (!NFS_WBACK_BUSY(req));
393
394         spin_lock(&inode->i_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(&inode->i_lock);
401                 iput(inode);
402         } else
403                 spin_unlock(&inode->i_lock);
404         nfs_clear_request(req);
405         nfs_release_request(req);
406 }
407
408 static void
409 nfs_mark_request_dirty(struct nfs_page *req)
410 {
411         __set_page_dirty_nobuffers(req->wb_page);
412 }
413
414 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
415 /*
416  * Add a request to the inode's commit list.
417  */
418 static void
419 nfs_mark_request_commit(struct nfs_page *req)
420 {
421         struct inode *inode = req->wb_context->path.dentry->d_inode;
422         struct nfs_inode *nfsi = NFS_I(inode);
423
424         spin_lock(&inode->i_lock);
425         set_bit(PG_CLEAN, &(req)->wb_flags);
426         radix_tree_tag_set(&nfsi->nfs_page_tree,
427                         req->wb_index,
428                         NFS_PAGE_TAG_COMMIT);
429         spin_unlock(&inode->i_lock);
430         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
431         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
432         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
433 }
434
435 static int
436 nfs_clear_request_commit(struct nfs_page *req)
437 {
438         struct page *page = req->wb_page;
439
440         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
441                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
442                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
443                 return 1;
444         }
445         return 0;
446 }
447
448 static inline
449 int nfs_write_need_commit(struct nfs_write_data *data)
450 {
451         return data->verf.committed != NFS_FILE_SYNC;
452 }
453
454 static inline
455 int nfs_reschedule_unstable_write(struct nfs_page *req)
456 {
457         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
458                 nfs_mark_request_commit(req);
459                 return 1;
460         }
461         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
462                 nfs_mark_request_dirty(req);
463                 return 1;
464         }
465         return 0;
466 }
467 #else
468 static inline void
469 nfs_mark_request_commit(struct nfs_page *req)
470 {
471 }
472
473 static inline int
474 nfs_clear_request_commit(struct nfs_page *req)
475 {
476         return 0;
477 }
478
479 static inline
480 int nfs_write_need_commit(struct nfs_write_data *data)
481 {
482         return 0;
483 }
484
485 static inline
486 int nfs_reschedule_unstable_write(struct nfs_page *req)
487 {
488         return 0;
489 }
490 #endif
491
492 /*
493  * Wait for a request to complete.
494  *
495  * Interruptible by fatal signals only.
496  */
497 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
498 {
499         struct nfs_inode *nfsi = NFS_I(inode);
500         struct nfs_page *req;
501         pgoff_t idx_end, next;
502         unsigned int            res = 0;
503         int                     error;
504
505         if (npages == 0)
506                 idx_end = ~0;
507         else
508                 idx_end = idx_start + npages - 1;
509
510         next = idx_start;
511         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
512                 if (req->wb_index > idx_end)
513                         break;
514
515                 next = req->wb_index + 1;
516                 BUG_ON(!NFS_WBACK_BUSY(req));
517
518                 kref_get(&req->wb_kref);
519                 spin_unlock(&inode->i_lock);
520                 error = nfs_wait_on_request(req);
521                 nfs_release_request(req);
522                 spin_lock(&inode->i_lock);
523                 if (error < 0)
524                         return error;
525                 res++;
526         }
527         return res;
528 }
529
530 static void nfs_cancel_commit_list(struct list_head *head)
531 {
532         struct nfs_page *req;
533
534         while(!list_empty(head)) {
535                 req = nfs_list_entry(head->next);
536                 nfs_list_remove_request(req);
537                 nfs_clear_request_commit(req);
538                 nfs_inode_remove_request(req);
539                 nfs_unlock_request(req);
540         }
541 }
542
543 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
544 static int
545 nfs_need_commit(struct nfs_inode *nfsi)
546 {
547         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
548 }
549
550 /*
551  * nfs_scan_commit - Scan an inode for commit requests
552  * @inode: NFS inode to scan
553  * @dst: destination list
554  * @idx_start: lower bound of page->index to scan.
555  * @npages: idx_start + npages sets the upper bound to scan.
556  *
557  * Moves requests from the inode's 'commit' request list.
558  * The requests are *not* checked to ensure that they form a contiguous set.
559  */
560 static int
561 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
562 {
563         struct nfs_inode *nfsi = NFS_I(inode);
564
565         if (!nfs_need_commit(nfsi))
566                 return 0;
567
568         return nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
569 }
570 #else
571 static inline int nfs_need_commit(struct nfs_inode *nfsi)
572 {
573         return 0;
574 }
575
576 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
577 {
578         return 0;
579 }
580 #endif
581
582 /*
583  * Search for an existing write request, and attempt to update
584  * it to reflect a new dirty region on a given page.
585  *
586  * If the attempt fails, then the existing request is flushed out
587  * to disk.
588  */
589 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
590                 struct page *page,
591                 unsigned int offset,
592                 unsigned int bytes)
593 {
594         struct nfs_page *req;
595         unsigned int rqend;
596         unsigned int end;
597         int error;
598
599         if (!PagePrivate(page))
600                 return NULL;
601
602         end = offset + bytes;
603         spin_lock(&inode->i_lock);
604
605         for (;;) {
606                 req = nfs_page_find_request_locked(page);
607                 if (req == NULL)
608                         goto out_unlock;
609
610                 rqend = req->wb_offset + req->wb_bytes;
611                 /*
612                  * Tell the caller to flush out the request if
613                  * the offsets are non-contiguous.
614                  * Note: nfs_flush_incompatible() will already
615                  * have flushed out requests having wrong owners.
616                  */
617                 if (offset > rqend
618                     || end < req->wb_offset)
619                         goto out_flushme;
620
621                 if (nfs_set_page_tag_locked(req))
622                         break;
623
624                 /* The request is locked, so wait and then retry */
625                 spin_unlock(&inode->i_lock);
626                 error = nfs_wait_on_request(req);
627                 nfs_release_request(req);
628                 if (error != 0)
629                         goto out_err;
630                 spin_lock(&inode->i_lock);
631         }
632
633         if (nfs_clear_request_commit(req))
634                 radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
635                                 req->wb_index, NFS_PAGE_TAG_COMMIT);
636
637         /* Okay, the request matches. Update the region */
638         if (offset < req->wb_offset) {
639                 req->wb_offset = offset;
640                 req->wb_pgbase = offset;
641         }
642         if (end > rqend)
643                 req->wb_bytes = end - req->wb_offset;
644         else
645                 req->wb_bytes = rqend - req->wb_offset;
646 out_unlock:
647         spin_unlock(&inode->i_lock);
648         return req;
649 out_flushme:
650         spin_unlock(&inode->i_lock);
651         nfs_release_request(req);
652         error = nfs_wb_page(inode, page);
653 out_err:
654         return ERR_PTR(error);
655 }
656
657 /*
658  * Try to update an existing write request, or create one if there is none.
659  *
660  * Note: Should always be called with the Page Lock held to prevent races
661  * if we have to add a new request. Also assumes that the caller has
662  * already called nfs_flush_incompatible() if necessary.
663  */
664 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
665                 struct page *page, unsigned int offset, unsigned int bytes)
666 {
667         struct inode *inode = page->mapping->host;
668         struct nfs_page *req;
669         int error;
670
671         req = nfs_try_to_update_request(inode, page, offset, bytes);
672         if (req != NULL)
673                 goto out;
674         req = nfs_create_request(ctx, inode, page, offset, bytes);
675         if (IS_ERR(req))
676                 goto out;
677         error = nfs_inode_add_request(inode, req);
678         if (error != 0) {
679                 nfs_release_request(req);
680                 req = ERR_PTR(error);
681         }
682 out:
683         return req;
684 }
685
686 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
687                 unsigned int offset, unsigned int count)
688 {
689         struct nfs_page *req;
690
691         req = nfs_setup_write_request(ctx, page, offset, count);
692         if (IS_ERR(req))
693                 return PTR_ERR(req);
694         /* Update file length */
695         nfs_grow_file(page, offset, count);
696         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
697         nfs_clear_page_tag_locked(req);
698         return 0;
699 }
700
701 int nfs_flush_incompatible(struct file *file, struct page *page)
702 {
703         struct nfs_open_context *ctx = nfs_file_open_context(file);
704         struct nfs_page *req;
705         int do_flush, status;
706         /*
707          * Look for a request corresponding to this page. If there
708          * is one, and it belongs to another file, we flush it out
709          * before we try to copy anything into the page. Do this
710          * due to the lack of an ACCESS-type call in NFSv2.
711          * Also do the same if we find a request from an existing
712          * dropped page.
713          */
714         do {
715                 req = nfs_page_find_request(page);
716                 if (req == NULL)
717                         return 0;
718                 do_flush = req->wb_page != page || req->wb_context != ctx;
719                 nfs_release_request(req);
720                 if (!do_flush)
721                         return 0;
722                 status = nfs_wb_page(page->mapping->host, page);
723         } while (status == 0);
724         return status;
725 }
726
727 /*
728  * If the page cache is marked as unsafe or invalid, then we can't rely on
729  * the PageUptodate() flag. In this case, we will need to turn off
730  * write optimisations that depend on the page contents being correct.
731  */
732 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
733 {
734         return PageUptodate(page) &&
735                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
736 }
737
738 /*
739  * Update and possibly write a cached page of an NFS file.
740  *
741  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
742  * things with a page scheduled for an RPC call (e.g. invalidate it).
743  */
744 int nfs_updatepage(struct file *file, struct page *page,
745                 unsigned int offset, unsigned int count)
746 {
747         struct nfs_open_context *ctx = nfs_file_open_context(file);
748         struct inode    *inode = page->mapping->host;
749         int             status = 0;
750
751         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
752
753         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
754                 file->f_path.dentry->d_parent->d_name.name,
755                 file->f_path.dentry->d_name.name, count,
756                 (long long)(page_offset(page) + offset));
757
758         /* If we're not using byte range locks, and we know the page
759          * is up to date, it may be more efficient to extend the write
760          * to cover the entire page in order to avoid fragmentation
761          * inefficiencies.
762          */
763         if (nfs_write_pageuptodate(page, inode) &&
764                         inode->i_flock == NULL &&
765                         !(file->f_flags & O_SYNC)) {
766                 count = max(count + offset, nfs_page_length(page));
767                 offset = 0;
768         }
769
770         status = nfs_writepage_setup(ctx, page, offset, count);
771         if (status < 0)
772                 nfs_set_pageerror(page);
773         else
774                 __set_page_dirty_nobuffers(page);
775
776         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
777                         status, (long long)i_size_read(inode));
778         return status;
779 }
780
781 static void nfs_writepage_release(struct nfs_page *req)
782 {
783
784         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
785                 nfs_end_page_writeback(req->wb_page);
786                 nfs_inode_remove_request(req);
787         } else
788                 nfs_end_page_writeback(req->wb_page);
789         nfs_clear_page_tag_locked(req);
790 }
791
792 static int flush_task_priority(int how)
793 {
794         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
795                 case FLUSH_HIGHPRI:
796                         return RPC_PRIORITY_HIGH;
797                 case FLUSH_LOWPRI:
798                         return RPC_PRIORITY_LOW;
799         }
800         return RPC_PRIORITY_NORMAL;
801 }
802
803 /*
804  * Set up the argument/result storage required for the RPC call.
805  */
806 static int nfs_write_rpcsetup(struct nfs_page *req,
807                 struct nfs_write_data *data,
808                 const struct rpc_call_ops *call_ops,
809                 unsigned int count, unsigned int offset,
810                 int how)
811 {
812         struct inode *inode = req->wb_context->path.dentry->d_inode;
813         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
814         int priority = flush_task_priority(how);
815         struct rpc_task *task;
816         struct rpc_message msg = {
817                 .rpc_argp = &data->args,
818                 .rpc_resp = &data->res,
819                 .rpc_cred = req->wb_context->cred,
820         };
821         struct rpc_task_setup task_setup_data = {
822                 .rpc_client = NFS_CLIENT(inode),
823                 .task = &data->task,
824                 .rpc_message = &msg,
825                 .callback_ops = call_ops,
826                 .callback_data = data,
827                 .workqueue = nfsiod_workqueue,
828                 .flags = flags,
829                 .priority = priority,
830         };
831
832         /* Set up the RPC argument and reply structs
833          * NB: take care not to mess about with data->commit et al. */
834
835         data->req = req;
836         data->inode = inode = req->wb_context->path.dentry->d_inode;
837         data->cred = msg.rpc_cred;
838
839         data->args.fh     = NFS_FH(inode);
840         data->args.offset = req_offset(req) + offset;
841         data->args.pgbase = req->wb_pgbase + offset;
842         data->args.pages  = data->pagevec;
843         data->args.count  = count;
844         data->args.context = get_nfs_open_context(req->wb_context);
845         data->args.stable  = NFS_UNSTABLE;
846         if (how & FLUSH_STABLE) {
847                 data->args.stable = NFS_DATA_SYNC;
848                 if (!nfs_need_commit(NFS_I(inode)))
849                         data->args.stable = NFS_FILE_SYNC;
850         }
851
852         data->res.fattr   = &data->fattr;
853         data->res.count   = count;
854         data->res.verf    = &data->verf;
855         nfs_fattr_init(&data->fattr);
856
857         /* Set up the initial task struct.  */
858         NFS_PROTO(inode)->write_setup(data, &msg);
859
860         dprintk("NFS: %5u initiated write call "
861                 "(req %s/%lld, %u bytes @ offset %llu)\n",
862                 data->task.tk_pid,
863                 inode->i_sb->s_id,
864                 (long long)NFS_FILEID(inode),
865                 count,
866                 (unsigned long long)data->args.offset);
867
868         task = rpc_run_task(&task_setup_data);
869         if (IS_ERR(task))
870                 return PTR_ERR(task);
871         rpc_put_task(task);
872         return 0;
873 }
874
875 /* If a nfs_flush_* function fails, it should remove reqs from @head and
876  * call this on each, which will prepare them to be retried on next
877  * writeback using standard nfs.
878  */
879 static void nfs_redirty_request(struct nfs_page *req)
880 {
881         nfs_mark_request_dirty(req);
882         nfs_end_page_writeback(req->wb_page);
883         nfs_clear_page_tag_locked(req);
884 }
885
886 /*
887  * Generate multiple small requests to write out a single
888  * contiguous dirty area on one page.
889  */
890 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
891 {
892         struct nfs_page *req = nfs_list_entry(head->next);
893         struct page *page = req->wb_page;
894         struct nfs_write_data *data;
895         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
896         unsigned int offset;
897         int requests = 0;
898         int ret = 0;
899         LIST_HEAD(list);
900
901         nfs_list_remove_request(req);
902
903         nbytes = count;
904         do {
905                 size_t len = min(nbytes, wsize);
906
907                 data = nfs_writedata_alloc(1);
908                 if (!data)
909                         goto out_bad;
910                 list_add(&data->pages, &list);
911                 requests++;
912                 nbytes -= len;
913         } while (nbytes != 0);
914         atomic_set(&req->wb_complete, requests);
915
916         ClearPageError(page);
917         offset = 0;
918         nbytes = count;
919         do {
920                 int ret2;
921
922                 data = list_entry(list.next, struct nfs_write_data, pages);
923                 list_del_init(&data->pages);
924
925                 data->pagevec[0] = page;
926
927                 if (nbytes < wsize)
928                         wsize = nbytes;
929                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
930                                    wsize, offset, how);
931                 if (ret == 0)
932                         ret = ret2;
933                 offset += wsize;
934                 nbytes -= wsize;
935         } while (nbytes != 0);
936
937         return ret;
938
939 out_bad:
940         while (!list_empty(&list)) {
941                 data = list_entry(list.next, struct nfs_write_data, pages);
942                 list_del(&data->pages);
943                 nfs_writedata_release(data);
944         }
945         nfs_redirty_request(req);
946         return -ENOMEM;
947 }
948
949 /*
950  * Create an RPC task for the given write request and kick it.
951  * The page must have been locked by the caller.
952  *
953  * It may happen that the page we're passed is not marked dirty.
954  * This is the case if nfs_updatepage detects a conflicting request
955  * that has been written but not committed.
956  */
957 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
958 {
959         struct nfs_page         *req;
960         struct page             **pages;
961         struct nfs_write_data   *data;
962
963         data = nfs_writedata_alloc(npages);
964         if (!data)
965                 goto out_bad;
966
967         pages = data->pagevec;
968         while (!list_empty(head)) {
969                 req = nfs_list_entry(head->next);
970                 nfs_list_remove_request(req);
971                 nfs_list_add_request(req, &data->pages);
972                 ClearPageError(req->wb_page);
973                 *pages++ = req->wb_page;
974         }
975         req = nfs_list_entry(data->pages.next);
976
977         /* Set up the argument struct */
978         return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
979  out_bad:
980         while (!list_empty(head)) {
981                 req = nfs_list_entry(head->next);
982                 nfs_list_remove_request(req);
983                 nfs_redirty_request(req);
984         }
985         return -ENOMEM;
986 }
987
988 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
989                                   struct inode *inode, int ioflags)
990 {
991         size_t wsize = NFS_SERVER(inode)->wsize;
992
993         if (wsize < PAGE_CACHE_SIZE)
994                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
995         else
996                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
997 }
998
999 /*
1000  * Handle a write reply that flushed part of a page.
1001  */
1002 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1003 {
1004         struct nfs_write_data   *data = calldata;
1005
1006         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1007                 task->tk_pid,
1008                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1009                 (long long)
1010                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1011                 data->req->wb_bytes, (long long)req_offset(data->req));
1012
1013         nfs_writeback_done(task, data);
1014 }
1015
1016 static void nfs_writeback_release_partial(void *calldata)
1017 {
1018         struct nfs_write_data   *data = calldata;
1019         struct nfs_page         *req = data->req;
1020         struct page             *page = req->wb_page;
1021         int status = data->task.tk_status;
1022
1023         if (status < 0) {
1024                 nfs_set_pageerror(page);
1025                 nfs_context_set_write_error(req->wb_context, status);
1026                 dprintk(", error = %d\n", status);
1027                 goto out;
1028         }
1029
1030         if (nfs_write_need_commit(data)) {
1031                 struct inode *inode = page->mapping->host;
1032
1033                 spin_lock(&inode->i_lock);
1034                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1035                         /* Do nothing we need to resend the writes */
1036                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1037                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1038                         dprintk(" defer commit\n");
1039                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1040                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1041                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1042                         dprintk(" server reboot detected\n");
1043                 }
1044                 spin_unlock(&inode->i_lock);
1045         } else
1046                 dprintk(" OK\n");
1047
1048 out:
1049         if (atomic_dec_and_test(&req->wb_complete))
1050                 nfs_writepage_release(req);
1051         nfs_writedata_release(calldata);
1052 }
1053
1054 #if defined(CONFIG_NFS_V4_1)
1055 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1056 {
1057         struct nfs_write_data *data = calldata;
1058         struct nfs_client *clp = (NFS_SERVER(data->inode))->nfs_client;
1059
1060         if (nfs4_setup_sequence(clp, &data->args.seq_args,
1061                                 &data->res.seq_res, 1, task))
1062                 return;
1063         rpc_call_start(task);
1064 }
1065 #endif /* CONFIG_NFS_V4_1 */
1066
1067 static const struct rpc_call_ops nfs_write_partial_ops = {
1068 #if defined(CONFIG_NFS_V4_1)
1069         .rpc_call_prepare = nfs_write_prepare,
1070 #endif /* CONFIG_NFS_V4_1 */
1071         .rpc_call_done = nfs_writeback_done_partial,
1072         .rpc_release = nfs_writeback_release_partial,
1073 };
1074
1075 /*
1076  * Handle a write reply that flushes a whole page.
1077  *
1078  * FIXME: There is an inherent race with invalidate_inode_pages and
1079  *        writebacks since the page->count is kept > 1 for as long
1080  *        as the page has a write request pending.
1081  */
1082 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1083 {
1084         struct nfs_write_data   *data = calldata;
1085
1086         nfs_writeback_done(task, data);
1087 }
1088
1089 static void nfs_writeback_release_full(void *calldata)
1090 {
1091         struct nfs_write_data   *data = calldata;
1092         int status = data->task.tk_status;
1093
1094         /* Update attributes as result of writeback. */
1095         while (!list_empty(&data->pages)) {
1096                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1097                 struct page *page = req->wb_page;
1098
1099                 nfs_list_remove_request(req);
1100
1101                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1102                         data->task.tk_pid,
1103                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1104                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1105                         req->wb_bytes,
1106                         (long long)req_offset(req));
1107
1108                 if (status < 0) {
1109                         nfs_set_pageerror(page);
1110                         nfs_context_set_write_error(req->wb_context, status);
1111                         dprintk(", error = %d\n", status);
1112                         goto remove_request;
1113                 }
1114
1115                 if (nfs_write_need_commit(data)) {
1116                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1117                         nfs_mark_request_commit(req);
1118                         nfs_end_page_writeback(page);
1119                         dprintk(" marked for commit\n");
1120                         goto next;
1121                 }
1122                 dprintk(" OK\n");
1123 remove_request:
1124                 nfs_end_page_writeback(page);
1125                 nfs_inode_remove_request(req);
1126         next:
1127                 nfs_clear_page_tag_locked(req);
1128         }
1129         nfs_writedata_release(calldata);
1130 }
1131
1132 static const struct rpc_call_ops nfs_write_full_ops = {
1133 #if defined(CONFIG_NFS_V4_1)
1134         .rpc_call_prepare = nfs_write_prepare,
1135 #endif /* CONFIG_NFS_V4_1 */
1136         .rpc_call_done = nfs_writeback_done_full,
1137         .rpc_release = nfs_writeback_release_full,
1138 };
1139
1140
1141 /*
1142  * This function is called when the WRITE call is complete.
1143  */
1144 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1145 {
1146         struct nfs_writeargs    *argp = &data->args;
1147         struct nfs_writeres     *resp = &data->res;
1148         struct nfs_server       *server = NFS_SERVER(data->inode);
1149         int status;
1150
1151         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1152                 task->tk_pid, task->tk_status);
1153
1154         /*
1155          * ->write_done will attempt to use post-op attributes to detect
1156          * conflicting writes by other clients.  A strict interpretation
1157          * of close-to-open would allow us to continue caching even if
1158          * another writer had changed the file, but some applications
1159          * depend on tighter cache coherency when writing.
1160          */
1161         status = NFS_PROTO(data->inode)->write_done(task, data);
1162         if (status != 0)
1163                 return status;
1164         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1165
1166 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1167         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1168                 /* We tried a write call, but the server did not
1169                  * commit data to stable storage even though we
1170                  * requested it.
1171                  * Note: There is a known bug in Tru64 < 5.0 in which
1172                  *       the server reports NFS_DATA_SYNC, but performs
1173                  *       NFS_FILE_SYNC. We therefore implement this checking
1174                  *       as a dprintk() in order to avoid filling syslog.
1175                  */
1176                 static unsigned long    complain;
1177
1178                 if (time_before(complain, jiffies)) {
1179                         dprintk("NFS:       faulty NFS server %s:"
1180                                 " (committed = %d) != (stable = %d)\n",
1181                                 server->nfs_client->cl_hostname,
1182                                 resp->verf->committed, argp->stable);
1183                         complain = jiffies + 300 * HZ;
1184                 }
1185         }
1186 #endif
1187         /* Is this a short write? */
1188         if (task->tk_status >= 0 && resp->count < argp->count) {
1189                 static unsigned long    complain;
1190
1191                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1192
1193                 /* Has the server at least made some progress? */
1194                 if (resp->count != 0) {
1195                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1196                         if (resp->verf->committed != NFS_UNSTABLE) {
1197                                 /* Resend from where the server left off */
1198                                 argp->offset += resp->count;
1199                                 argp->pgbase += resp->count;
1200                                 argp->count -= resp->count;
1201                         } else {
1202                                 /* Resend as a stable write in order to avoid
1203                                  * headaches in the case of a server crash.
1204                                  */
1205                                 argp->stable = NFS_FILE_SYNC;
1206                         }
1207                         nfs4_restart_rpc(task, server->nfs_client);
1208                         return -EAGAIN;
1209                 }
1210                 if (time_before(complain, jiffies)) {
1211                         printk(KERN_WARNING
1212                                "NFS: Server wrote zero bytes, expected %u.\n",
1213                                         argp->count);
1214                         complain = jiffies + 300 * HZ;
1215                 }
1216                 /* Can't do anything about it except throw an error. */
1217                 task->tk_status = -EIO;
1218         }
1219         nfs4_sequence_free_slot(server->nfs_client, &data->res.seq_res);
1220         return 0;
1221 }
1222
1223
1224 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1225 void nfs_commitdata_release(void *data)
1226 {
1227         struct nfs_write_data *wdata = data;
1228
1229         put_nfs_open_context(wdata->args.context);
1230         nfs_commit_free(wdata);
1231 }
1232
1233 /*
1234  * Set up the argument/result storage required for the RPC call.
1235  */
1236 static int nfs_commit_rpcsetup(struct list_head *head,
1237                 struct nfs_write_data *data,
1238                 int how)
1239 {
1240         struct nfs_page *first = nfs_list_entry(head->next);
1241         struct inode *inode = first->wb_context->path.dentry->d_inode;
1242         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1243         int priority = flush_task_priority(how);
1244         struct rpc_task *task;
1245         struct rpc_message msg = {
1246                 .rpc_argp = &data->args,
1247                 .rpc_resp = &data->res,
1248                 .rpc_cred = first->wb_context->cred,
1249         };
1250         struct rpc_task_setup task_setup_data = {
1251                 .task = &data->task,
1252                 .rpc_client = NFS_CLIENT(inode),
1253                 .rpc_message = &msg,
1254                 .callback_ops = &nfs_commit_ops,
1255                 .callback_data = data,
1256                 .workqueue = nfsiod_workqueue,
1257                 .flags = flags,
1258                 .priority = priority,
1259         };
1260
1261         /* Set up the RPC argument and reply structs
1262          * NB: take care not to mess about with data->commit et al. */
1263
1264         list_splice_init(head, &data->pages);
1265
1266         data->inode       = inode;
1267         data->cred        = msg.rpc_cred;
1268
1269         data->args.fh     = NFS_FH(data->inode);
1270         /* Note: we always request a commit of the entire inode */
1271         data->args.offset = 0;
1272         data->args.count  = 0;
1273         data->args.context = get_nfs_open_context(first->wb_context);
1274         data->res.count   = 0;
1275         data->res.fattr   = &data->fattr;
1276         data->res.verf    = &data->verf;
1277         nfs_fattr_init(&data->fattr);
1278
1279         /* Set up the initial task struct.  */
1280         NFS_PROTO(inode)->commit_setup(data, &msg);
1281
1282         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1283
1284         task = rpc_run_task(&task_setup_data);
1285         if (IS_ERR(task))
1286                 return PTR_ERR(task);
1287         rpc_put_task(task);
1288         return 0;
1289 }
1290
1291 /*
1292  * Commit dirty pages
1293  */
1294 static int
1295 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1296 {
1297         struct nfs_write_data   *data;
1298         struct nfs_page         *req;
1299
1300         data = nfs_commitdata_alloc();
1301
1302         if (!data)
1303                 goto out_bad;
1304
1305         /* Set up the argument struct */
1306         return nfs_commit_rpcsetup(head, data, how);
1307  out_bad:
1308         while (!list_empty(head)) {
1309                 req = nfs_list_entry(head->next);
1310                 nfs_list_remove_request(req);
1311                 nfs_mark_request_commit(req);
1312                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1313                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1314                                 BDI_RECLAIMABLE);
1315                 nfs_clear_page_tag_locked(req);
1316         }
1317         return -ENOMEM;
1318 }
1319
1320 /*
1321  * COMMIT call returned
1322  */
1323 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1324 {
1325         struct nfs_write_data   *data = calldata;
1326
1327         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1328                                 task->tk_pid, task->tk_status);
1329
1330         /* Call the NFS version-specific code */
1331         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1332                 return;
1333 }
1334
1335 static void nfs_commit_release(void *calldata)
1336 {
1337         struct nfs_write_data   *data = calldata;
1338         struct nfs_page         *req;
1339         int status = data->task.tk_status;
1340
1341         while (!list_empty(&data->pages)) {
1342                 req = nfs_list_entry(data->pages.next);
1343                 nfs_list_remove_request(req);
1344                 nfs_clear_request_commit(req);
1345
1346                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1347                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1348                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1349                         req->wb_bytes,
1350                         (long long)req_offset(req));
1351                 if (status < 0) {
1352                         nfs_context_set_write_error(req->wb_context, status);
1353                         nfs_inode_remove_request(req);
1354                         dprintk(", error = %d\n", status);
1355                         goto next;
1356                 }
1357
1358                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1359                  * returned by the server against all stored verfs. */
1360                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1361                         /* We have a match */
1362                         nfs_inode_remove_request(req);
1363                         dprintk(" OK\n");
1364                         goto next;
1365                 }
1366                 /* We have a mismatch. Write the page again */
1367                 dprintk(" mismatch\n");
1368                 nfs_mark_request_dirty(req);
1369         next:
1370                 nfs_clear_page_tag_locked(req);
1371         }
1372         nfs_commitdata_release(calldata);
1373 }
1374
1375 static const struct rpc_call_ops nfs_commit_ops = {
1376 #if defined(CONFIG_NFS_V4_1)
1377         .rpc_call_prepare = nfs_write_prepare,
1378 #endif /* CONFIG_NFS_V4_1 */
1379         .rpc_call_done = nfs_commit_done,
1380         .rpc_release = nfs_commit_release,
1381 };
1382
1383 int nfs_commit_inode(struct inode *inode, int how)
1384 {
1385         LIST_HEAD(head);
1386         int res;
1387
1388         spin_lock(&inode->i_lock);
1389         res = nfs_scan_commit(inode, &head, 0, 0);
1390         spin_unlock(&inode->i_lock);
1391         if (res) {
1392                 int error = nfs_commit_list(inode, &head, how);
1393                 if (error < 0)
1394                         return error;
1395         }
1396         return res;
1397 }
1398 #else
1399 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1400 {
1401         return 0;
1402 }
1403 #endif
1404
1405 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1406 {
1407         struct inode *inode = mapping->host;
1408         pgoff_t idx_start, idx_end;
1409         unsigned int npages = 0;
1410         LIST_HEAD(head);
1411         int nocommit = how & FLUSH_NOCOMMIT;
1412         long pages, ret;
1413
1414         /* FIXME */
1415         if (wbc->range_cyclic)
1416                 idx_start = 0;
1417         else {
1418                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1419                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1420                 if (idx_end > idx_start) {
1421                         pgoff_t l_npages = 1 + idx_end - idx_start;
1422                         npages = l_npages;
1423                         if (sizeof(npages) != sizeof(l_npages) &&
1424                                         (pgoff_t)npages != l_npages)
1425                                 npages = 0;
1426                 }
1427         }
1428         how &= ~FLUSH_NOCOMMIT;
1429         spin_lock(&inode->i_lock);
1430         do {
1431                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1432                 if (ret != 0)
1433                         continue;
1434                 if (nocommit)
1435                         break;
1436                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1437                 if (pages == 0)
1438                         break;
1439                 if (how & FLUSH_INVALIDATE) {
1440                         spin_unlock(&inode->i_lock);
1441                         nfs_cancel_commit_list(&head);
1442                         ret = pages;
1443                         spin_lock(&inode->i_lock);
1444                         continue;
1445                 }
1446                 pages += nfs_scan_commit(inode, &head, 0, 0);
1447                 spin_unlock(&inode->i_lock);
1448                 ret = nfs_commit_list(inode, &head, how);
1449                 spin_lock(&inode->i_lock);
1450
1451         } while (ret >= 0);
1452         spin_unlock(&inode->i_lock);
1453         return ret;
1454 }
1455
1456 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1457 {
1458         int ret;
1459
1460         ret = nfs_writepages(mapping, wbc);
1461         if (ret < 0)
1462                 goto out;
1463         ret = nfs_sync_mapping_wait(mapping, wbc, how);
1464         if (ret < 0)
1465                 goto out;
1466         return 0;
1467 out:
1468         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1469         return ret;
1470 }
1471
1472 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1473 static int nfs_write_mapping(struct address_space *mapping, int how)
1474 {
1475         struct writeback_control wbc = {
1476                 .bdi = mapping->backing_dev_info,
1477                 .sync_mode = WB_SYNC_ALL,
1478                 .nr_to_write = LONG_MAX,
1479                 .range_start = 0,
1480                 .range_end = LLONG_MAX,
1481                 .for_writepages = 1,
1482         };
1483
1484         return __nfs_write_mapping(mapping, &wbc, how);
1485 }
1486
1487 /*
1488  * flush the inode to disk.
1489  */
1490 int nfs_wb_all(struct inode *inode)
1491 {
1492         return nfs_write_mapping(inode->i_mapping, 0);
1493 }
1494
1495 int nfs_wb_nocommit(struct inode *inode)
1496 {
1497         return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1498 }
1499
1500 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1501 {
1502         struct nfs_page *req;
1503         loff_t range_start = page_offset(page);
1504         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1505         struct writeback_control wbc = {
1506                 .bdi = page->mapping->backing_dev_info,
1507                 .sync_mode = WB_SYNC_ALL,
1508                 .nr_to_write = LONG_MAX,
1509                 .range_start = range_start,
1510                 .range_end = range_end,
1511         };
1512         int ret = 0;
1513
1514         BUG_ON(!PageLocked(page));
1515         for (;;) {
1516                 req = nfs_page_find_request(page);
1517                 if (req == NULL)
1518                         goto out;
1519                 if (test_bit(PG_CLEAN, &req->wb_flags)) {
1520                         nfs_release_request(req);
1521                         break;
1522                 }
1523                 if (nfs_lock_request_dontget(req)) {
1524                         nfs_inode_remove_request(req);
1525                         /*
1526                          * In case nfs_inode_remove_request has marked the
1527                          * page as being dirty
1528                          */
1529                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1530                         nfs_unlock_request(req);
1531                         break;
1532                 }
1533                 ret = nfs_wait_on_request(req);
1534                 if (ret < 0)
1535                         goto out;
1536         }
1537         if (!PagePrivate(page))
1538                 return 0;
1539         ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1540 out:
1541         return ret;
1542 }
1543
1544 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1545                                 int how)
1546 {
1547         loff_t range_start = page_offset(page);
1548         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1549         struct writeback_control wbc = {
1550                 .bdi = page->mapping->backing_dev_info,
1551                 .sync_mode = WB_SYNC_ALL,
1552                 .nr_to_write = LONG_MAX,
1553                 .range_start = range_start,
1554                 .range_end = range_end,
1555         };
1556         int ret;
1557
1558         do {
1559                 if (clear_page_dirty_for_io(page)) {
1560                         ret = nfs_writepage_locked(page, &wbc);
1561                         if (ret < 0)
1562                                 goto out_error;
1563                 } else if (!PagePrivate(page))
1564                         break;
1565                 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1566                 if (ret < 0)
1567                         goto out_error;
1568         } while (PagePrivate(page));
1569         return 0;
1570 out_error:
1571         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1572         return ret;
1573 }
1574
1575 /*
1576  * Write back all requests on one page - we do this before reading it.
1577  */
1578 int nfs_wb_page(struct inode *inode, struct page* page)
1579 {
1580         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1581 }
1582
1583 int __init nfs_init_writepagecache(void)
1584 {
1585         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1586                                              sizeof(struct nfs_write_data),
1587                                              0, SLAB_HWCACHE_ALIGN,
1588                                              NULL);
1589         if (nfs_wdata_cachep == NULL)
1590                 return -ENOMEM;
1591
1592         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1593                                                      nfs_wdata_cachep);
1594         if (nfs_wdata_mempool == NULL)
1595                 return -ENOMEM;
1596
1597         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1598                                                       nfs_wdata_cachep);
1599         if (nfs_commit_mempool == NULL)
1600                 return -ENOMEM;
1601
1602         /*
1603          * NFS congestion size, scale with available memory.
1604          *
1605          *  64MB:    8192k
1606          * 128MB:   11585k
1607          * 256MB:   16384k
1608          * 512MB:   23170k
1609          *   1GB:   32768k
1610          *   2GB:   46340k
1611          *   4GB:   65536k
1612          *   8GB:   92681k
1613          *  16GB:  131072k
1614          *
1615          * This allows larger machines to have larger/more transfers.
1616          * Limit the default to 256M
1617          */
1618         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1619         if (nfs_congestion_kb > 256*1024)
1620                 nfs_congestion_kb = 256*1024;
1621
1622         return 0;
1623 }
1624
1625 void nfs_destroy_writepagecache(void)
1626 {
1627         mempool_destroy(nfs_commit_mempool);
1628         mempool_destroy(nfs_wdata_mempool);
1629         kmem_cache_destroy(nfs_wdata_cachep);
1630 }
1631