NFS: add I/O performance counters
[linux-2.6.git] / fs / nfs / write.c
1 /*
2  * linux/fs/nfs/write.c
3  *
4  * Writing file data over NFS.
5  *
6  * We do it like this: When a (user) process wishes to write data to an
7  * NFS file, a write request is allocated that contains the RPC task data
8  * plus some info on the page to be written, and added to the inode's
9  * write chain. If the process writes past the end of the page, an async
10  * RPC call to write the page is scheduled immediately; otherwise, the call
11  * is delayed for a few seconds.
12  *
13  * Just like readahead, no async I/O is performed if wsize < PAGE_SIZE.
14  *
15  * Write requests are kept on the inode's writeback list. Each entry in
16  * that list references the page (portion) to be written. When the
17  * cache timeout has expired, the RPC task is woken up, and tries to
18  * lock the page. As soon as it manages to do so, the request is moved
19  * from the writeback list to the writelock list.
20  *
21  * Note: we must make sure never to confuse the inode passed in the
22  * write_page request with the one in page->inode. As far as I understand
23  * it, these are different when doing a swap-out.
24  *
25  * To understand everything that goes on here and in the NFS read code,
26  * one should be aware that a page is locked in exactly one of the following
27  * cases:
28  *
29  *  -   A write request is in progress.
30  *  -   A user process is in generic_file_write/nfs_update_page
31  *  -   A user process is in generic_file_read
32  *
33  * Also note that because of the way pages are invalidated in
34  * nfs_revalidate_inode, the following assertions hold:
35  *
36  *  -   If a page is dirty, there will be no read requests (a page will
37  *      not be re-read unless invalidated by nfs_revalidate_inode).
38  *  -   If the page is not uptodate, there will be no pending write
39  *      requests, and no process will be in nfs_update_page.
40  *
41  * FIXME: Interaction with the vmscan routines is not optimal yet.
42  * Either vmscan must be made nfs-savvy, or we need a different page
43  * reclaim concept that supports something like FS-independent
44  * buffer_heads with a b_ops-> field.
45  *
46  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
47  */
48
49 #include <linux/config.h>
50 #include <linux/types.h>
51 #include <linux/slab.h>
52 #include <linux/mm.h>
53 #include <linux/pagemap.h>
54 #include <linux/file.h>
55 #include <linux/mpage.h>
56 #include <linux/writeback.h>
57
58 #include <linux/sunrpc/clnt.h>
59 #include <linux/nfs_fs.h>
60 #include <linux/nfs_mount.h>
61 #include <linux/nfs_page.h>
62 #include <asm/uaccess.h>
63 #include <linux/smp_lock.h>
64
65 #include "delegation.h"
66 #include "iostat.h"
67
68 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
69
70 #define MIN_POOL_WRITE          (32)
71 #define MIN_POOL_COMMIT         (4)
72
73 /*
74  * Local function declarations
75  */
76 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
77                                             struct inode *,
78                                             struct page *,
79                                             unsigned int, unsigned int);
80 static void nfs_writeback_done_partial(struct nfs_write_data *, int);
81 static void nfs_writeback_done_full(struct nfs_write_data *, int);
82 static int nfs_wait_on_write_congestion(struct address_space *, int);
83 static int nfs_wait_on_requests(struct inode *, unsigned long, unsigned int);
84 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
85                            unsigned int npages, int how);
86
87 static kmem_cache_t *nfs_wdata_cachep;
88 mempool_t *nfs_wdata_mempool;
89 static mempool_t *nfs_commit_mempool;
90
91 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
92
93 static inline struct nfs_write_data *nfs_commit_alloc(unsigned int pagecount)
94 {
95         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS);
96
97         if (p) {
98                 memset(p, 0, sizeof(*p));
99                 INIT_LIST_HEAD(&p->pages);
100                 if (pagecount < NFS_PAGEVEC_SIZE)
101                         p->pagevec = &p->page_array[0];
102                 else {
103                         size_t size = ++pagecount * sizeof(struct page *);
104                         p->pagevec = kzalloc(size, GFP_NOFS);
105                         if (!p->pagevec) {
106                                 mempool_free(p, nfs_commit_mempool);
107                                 p = NULL;
108                         }
109                 }
110         }
111         return p;
112 }
113
114 static inline void nfs_commit_free(struct nfs_write_data *p)
115 {
116         if (p && (p->pagevec != &p->page_array[0]))
117                 kfree(p->pagevec);
118         mempool_free(p, nfs_commit_mempool);
119 }
120
121 void nfs_writedata_release(void *wdata)
122 {
123         nfs_writedata_free(wdata);
124 }
125
126 /* Adjust the file length if we're writing beyond the end */
127 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
128 {
129         struct inode *inode = page->mapping->host;
130         loff_t end, i_size = i_size_read(inode);
131         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
132
133         if (i_size > 0 && page->index < end_index)
134                 return;
135         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
136         if (i_size >= end)
137                 return;
138         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
139         i_size_write(inode, end);
140 }
141
142 /* We can set the PG_uptodate flag if we see that a write request
143  * covers the full page.
144  */
145 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
146 {
147         loff_t end_offs;
148
149         if (PageUptodate(page))
150                 return;
151         if (base != 0)
152                 return;
153         if (count == PAGE_CACHE_SIZE) {
154                 SetPageUptodate(page);
155                 return;
156         }
157
158         end_offs = i_size_read(page->mapping->host) - 1;
159         if (end_offs < 0)
160                 return;
161         /* Is this the last page? */
162         if (page->index != (unsigned long)(end_offs >> PAGE_CACHE_SHIFT))
163                 return;
164         /* This is the last page: set PG_uptodate if we cover the entire
165          * extent of the data, then zero the rest of the page.
166          */
167         if (count == (unsigned int)(end_offs & (PAGE_CACHE_SIZE - 1)) + 1) {
168                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
169                 SetPageUptodate(page);
170         }
171 }
172
173 /*
174  * Write a page synchronously.
175  * Offset is the data offset within the page.
176  */
177 static int nfs_writepage_sync(struct nfs_open_context *ctx, struct inode *inode,
178                 struct page *page, unsigned int offset, unsigned int count,
179                 int how)
180 {
181         unsigned int    wsize = NFS_SERVER(inode)->wsize;
182         int             result, written = 0;
183         struct nfs_write_data *wdata;
184
185         wdata = nfs_writedata_alloc(1);
186         if (!wdata)
187                 return -ENOMEM;
188
189         wdata->flags = how;
190         wdata->cred = ctx->cred;
191         wdata->inode = inode;
192         wdata->args.fh = NFS_FH(inode);
193         wdata->args.context = ctx;
194         wdata->args.pages = &page;
195         wdata->args.stable = NFS_FILE_SYNC;
196         wdata->args.pgbase = offset;
197         wdata->args.count = wsize;
198         wdata->res.fattr = &wdata->fattr;
199         wdata->res.verf = &wdata->verf;
200
201         dprintk("NFS:      nfs_writepage_sync(%s/%Ld %d@%Ld)\n",
202                 inode->i_sb->s_id,
203                 (long long)NFS_FILEID(inode),
204                 count, (long long)(page_offset(page) + offset));
205
206         set_page_writeback(page);
207         nfs_begin_data_update(inode);
208         do {
209                 if (count < wsize)
210                         wdata->args.count = count;
211                 wdata->args.offset = page_offset(page) + wdata->args.pgbase;
212
213                 result = NFS_PROTO(inode)->write(wdata);
214
215                 if (result < 0) {
216                         /* Must mark the page invalid after I/O error */
217                         ClearPageUptodate(page);
218                         goto io_error;
219                 }
220                 if (result < wdata->args.count)
221                         printk(KERN_WARNING "NFS: short write, count=%u, result=%d\n",
222                                         wdata->args.count, result);
223
224                 wdata->args.offset += result;
225                 wdata->args.pgbase += result;
226                 written += result;
227                 count -= result;
228                 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, result);
229         } while (count);
230         /* Update file length */
231         nfs_grow_file(page, offset, written);
232         /* Set the PG_uptodate flag? */
233         nfs_mark_uptodate(page, offset, written);
234
235         if (PageError(page))
236                 ClearPageError(page);
237
238 io_error:
239         nfs_end_data_update(inode);
240         end_page_writeback(page);
241         nfs_writedata_free(wdata);
242         return written ? written : result;
243 }
244
245 static int nfs_writepage_async(struct nfs_open_context *ctx,
246                 struct inode *inode, struct page *page,
247                 unsigned int offset, unsigned int count)
248 {
249         struct nfs_page *req;
250
251         req = nfs_update_request(ctx, inode, page, offset, count);
252         if (IS_ERR(req))
253                 return PTR_ERR(req);
254         /* Update file length */
255         nfs_grow_file(page, offset, count);
256         /* Set the PG_uptodate flag? */
257         nfs_mark_uptodate(page, offset, count);
258         nfs_unlock_request(req);
259         return 0;
260 }
261
262 static int wb_priority(struct writeback_control *wbc)
263 {
264         if (wbc->for_reclaim)
265                 return FLUSH_HIGHPRI;
266         if (wbc->for_kupdate)
267                 return FLUSH_LOWPRI;
268         return 0;
269 }
270
271 /*
272  * Write an mmapped page to the server.
273  */
274 int nfs_writepage(struct page *page, struct writeback_control *wbc)
275 {
276         struct nfs_open_context *ctx;
277         struct inode *inode = page->mapping->host;
278         unsigned long end_index;
279         unsigned offset = PAGE_CACHE_SIZE;
280         loff_t i_size = i_size_read(inode);
281         int inode_referenced = 0;
282         int priority = wb_priority(wbc);
283         int err;
284
285         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
286         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
287
288         /*
289          * Note: We need to ensure that we have a reference to the inode
290          *       if we are to do asynchronous writes. If not, waiting
291          *       in nfs_wait_on_request() may deadlock with clear_inode().
292          *
293          *       If igrab() fails here, then it is in any case safe to
294          *       call nfs_wb_page(), since there will be no pending writes.
295          */
296         if (igrab(inode) != 0)
297                 inode_referenced = 1;
298         end_index = i_size >> PAGE_CACHE_SHIFT;
299
300         /* Ensure we've flushed out any previous writes */
301         nfs_wb_page_priority(inode, page, priority);
302
303         /* easy case */
304         if (page->index < end_index)
305                 goto do_it;
306         /* things got complicated... */
307         offset = i_size & (PAGE_CACHE_SIZE-1);
308
309         /* OK, are we completely out? */
310         err = 0; /* potential race with truncate - ignore */
311         if (page->index >= end_index+1 || !offset)
312                 goto out;
313 do_it:
314         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
315         if (ctx == NULL) {
316                 err = -EBADF;
317                 goto out;
318         }
319         lock_kernel();
320         if (!IS_SYNC(inode) && inode_referenced) {
321                 err = nfs_writepage_async(ctx, inode, page, 0, offset);
322                 if (!wbc->for_writepages)
323                         nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
324         } else {
325                 err = nfs_writepage_sync(ctx, inode, page, 0,
326                                                 offset, priority);
327                 if (err >= 0) {
328                         if (err != offset)
329                                 redirty_page_for_writepage(wbc, page);
330                         err = 0;
331                 }
332         }
333         unlock_kernel();
334         put_nfs_open_context(ctx);
335 out:
336         unlock_page(page);
337         if (inode_referenced)
338                 iput(inode);
339         return err; 
340 }
341
342 /*
343  * Note: causes nfs_update_request() to block on the assumption
344  *       that the writeback is generated due to memory pressure.
345  */
346 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
347 {
348         struct backing_dev_info *bdi = mapping->backing_dev_info;
349         struct inode *inode = mapping->host;
350         int err;
351
352         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
353
354         err = generic_writepages(mapping, wbc);
355         if (err)
356                 return err;
357         while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
358                 if (wbc->nonblocking)
359                         return 0;
360                 nfs_wait_on_write_congestion(mapping, 0);
361         }
362         err = nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
363         if (err < 0)
364                 goto out;
365         nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
366         wbc->nr_to_write -= err;
367         if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) {
368                 err = nfs_wait_on_requests(inode, 0, 0);
369                 if (err < 0)
370                         goto out;
371         }
372         err = nfs_commit_inode(inode, wb_priority(wbc));
373         if (err > 0) {
374                 wbc->nr_to_write -= err;
375                 err = 0;
376         }
377 out:
378         clear_bit(BDI_write_congested, &bdi->state);
379         wake_up_all(&nfs_write_congestion);
380         return err;
381 }
382
383 /*
384  * Insert a write request into an inode
385  */
386 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
387 {
388         struct nfs_inode *nfsi = NFS_I(inode);
389         int error;
390
391         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
392         BUG_ON(error == -EEXIST);
393         if (error)
394                 return error;
395         if (!nfsi->npages) {
396                 igrab(inode);
397                 nfs_begin_data_update(inode);
398                 if (nfs_have_delegation(inode, FMODE_WRITE))
399                         nfsi->change_attr++;
400         }
401         nfsi->npages++;
402         atomic_inc(&req->wb_count);
403         return 0;
404 }
405
406 /*
407  * Insert a write request into an inode
408  */
409 static void nfs_inode_remove_request(struct nfs_page *req)
410 {
411         struct inode *inode = req->wb_context->dentry->d_inode;
412         struct nfs_inode *nfsi = NFS_I(inode);
413
414         BUG_ON (!NFS_WBACK_BUSY(req));
415
416         spin_lock(&nfsi->req_lock);
417         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
418         nfsi->npages--;
419         if (!nfsi->npages) {
420                 spin_unlock(&nfsi->req_lock);
421                 nfs_end_data_update(inode);
422                 iput(inode);
423         } else
424                 spin_unlock(&nfsi->req_lock);
425         nfs_clear_request(req);
426         nfs_release_request(req);
427 }
428
429 /*
430  * Find a request
431  */
432 static inline struct nfs_page *
433 _nfs_find_request(struct inode *inode, unsigned long index)
434 {
435         struct nfs_inode *nfsi = NFS_I(inode);
436         struct nfs_page *req;
437
438         req = (struct nfs_page*)radix_tree_lookup(&nfsi->nfs_page_tree, index);
439         if (req)
440                 atomic_inc(&req->wb_count);
441         return req;
442 }
443
444 static struct nfs_page *
445 nfs_find_request(struct inode *inode, unsigned long index)
446 {
447         struct nfs_page         *req;
448         struct nfs_inode        *nfsi = NFS_I(inode);
449
450         spin_lock(&nfsi->req_lock);
451         req = _nfs_find_request(inode, index);
452         spin_unlock(&nfsi->req_lock);
453         return req;
454 }
455
456 /*
457  * Add a request to the inode's dirty list.
458  */
459 static void
460 nfs_mark_request_dirty(struct nfs_page *req)
461 {
462         struct inode *inode = req->wb_context->dentry->d_inode;
463         struct nfs_inode *nfsi = NFS_I(inode);
464
465         spin_lock(&nfsi->req_lock);
466         radix_tree_tag_set(&nfsi->nfs_page_tree,
467                         req->wb_index, NFS_PAGE_TAG_DIRTY);
468         nfs_list_add_request(req, &nfsi->dirty);
469         nfsi->ndirty++;
470         spin_unlock(&nfsi->req_lock);
471         inc_page_state(nr_dirty);
472         mark_inode_dirty(inode);
473 }
474
475 /*
476  * Check if a request is dirty
477  */
478 static inline int
479 nfs_dirty_request(struct nfs_page *req)
480 {
481         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
482         return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty;
483 }
484
485 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
486 /*
487  * Add a request to the inode's commit list.
488  */
489 static void
490 nfs_mark_request_commit(struct nfs_page *req)
491 {
492         struct inode *inode = req->wb_context->dentry->d_inode;
493         struct nfs_inode *nfsi = NFS_I(inode);
494
495         spin_lock(&nfsi->req_lock);
496         nfs_list_add_request(req, &nfsi->commit);
497         nfsi->ncommit++;
498         spin_unlock(&nfsi->req_lock);
499         inc_page_state(nr_unstable);
500         mark_inode_dirty(inode);
501 }
502 #endif
503
504 /*
505  * Wait for a request to complete.
506  *
507  * Interruptible by signals only if mounted with intr flag.
508  */
509 static int
510 nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
511 {
512         struct nfs_inode *nfsi = NFS_I(inode);
513         struct nfs_page *req;
514         unsigned long           idx_end, next;
515         unsigned int            res = 0;
516         int                     error;
517
518         if (npages == 0)
519                 idx_end = ~0;
520         else
521                 idx_end = idx_start + npages - 1;
522
523         spin_lock(&nfsi->req_lock);
524         next = idx_start;
525         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
526                 if (req->wb_index > idx_end)
527                         break;
528
529                 next = req->wb_index + 1;
530                 BUG_ON(!NFS_WBACK_BUSY(req));
531
532                 atomic_inc(&req->wb_count);
533                 spin_unlock(&nfsi->req_lock);
534                 error = nfs_wait_on_request(req);
535                 nfs_release_request(req);
536                 if (error < 0)
537                         return error;
538                 spin_lock(&nfsi->req_lock);
539                 res++;
540         }
541         spin_unlock(&nfsi->req_lock);
542         return res;
543 }
544
545 /*
546  * nfs_scan_dirty - Scan an inode for dirty requests
547  * @inode: NFS inode to scan
548  * @dst: destination list
549  * @idx_start: lower bound of page->index to scan.
550  * @npages: idx_start + npages sets the upper bound to scan.
551  *
552  * Moves requests from the inode's dirty page list.
553  * The requests are *not* checked to ensure that they form a contiguous set.
554  */
555 static int
556 nfs_scan_dirty(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
557 {
558         struct nfs_inode *nfsi = NFS_I(inode);
559         int res = 0;
560
561         if (nfsi->ndirty != 0) {
562                 res = nfs_scan_lock_dirty(nfsi, dst, idx_start, npages);
563                 nfsi->ndirty -= res;
564                 sub_page_state(nr_dirty,res);
565                 if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty))
566                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n");
567         }
568         return res;
569 }
570
571 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
572 /*
573  * nfs_scan_commit - Scan an inode for commit requests
574  * @inode: NFS inode to scan
575  * @dst: destination list
576  * @idx_start: lower bound of page->index to scan.
577  * @npages: idx_start + npages sets the upper bound to scan.
578  *
579  * Moves requests from the inode's 'commit' request list.
580  * The requests are *not* checked to ensure that they form a contiguous set.
581  */
582 static int
583 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
584 {
585         struct nfs_inode *nfsi = NFS_I(inode);
586         int res = 0;
587
588         if (nfsi->ncommit != 0) {
589                 res = nfs_scan_list(&nfsi->commit, dst, idx_start, npages);
590                 nfsi->ncommit -= res;
591                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
592                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
593         }
594         return res;
595 }
596 #endif
597
598 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
599 {
600         struct backing_dev_info *bdi = mapping->backing_dev_info;
601         DEFINE_WAIT(wait);
602         int ret = 0;
603
604         might_sleep();
605
606         if (!bdi_write_congested(bdi))
607                 return 0;
608
609         nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
610
611         if (intr) {
612                 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
613                 sigset_t oldset;
614
615                 rpc_clnt_sigmask(clnt, &oldset);
616                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
617                 if (bdi_write_congested(bdi)) {
618                         if (signalled())
619                                 ret = -ERESTARTSYS;
620                         else
621                                 schedule();
622                 }
623                 rpc_clnt_sigunmask(clnt, &oldset);
624         } else {
625                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
626                 if (bdi_write_congested(bdi))
627                         schedule();
628         }
629         finish_wait(&nfs_write_congestion, &wait);
630         return ret;
631 }
632
633
634 /*
635  * Try to update any existing write request, or create one if there is none.
636  * In order to match, the request's credentials must match those of
637  * the calling process.
638  *
639  * Note: Should always be called with the Page Lock held!
640  */
641 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
642                 struct inode *inode, struct page *page,
643                 unsigned int offset, unsigned int bytes)
644 {
645         struct nfs_server *server = NFS_SERVER(inode);
646         struct nfs_inode *nfsi = NFS_I(inode);
647         struct nfs_page         *req, *new = NULL;
648         unsigned long           rqend, end;
649
650         end = offset + bytes;
651
652         if (nfs_wait_on_write_congestion(page->mapping, server->flags & NFS_MOUNT_INTR))
653                 return ERR_PTR(-ERESTARTSYS);
654         for (;;) {
655                 /* Loop over all inode entries and see if we find
656                  * A request for the page we wish to update
657                  */
658                 spin_lock(&nfsi->req_lock);
659                 req = _nfs_find_request(inode, page->index);
660                 if (req) {
661                         if (!nfs_lock_request_dontget(req)) {
662                                 int error;
663                                 spin_unlock(&nfsi->req_lock);
664                                 error = nfs_wait_on_request(req);
665                                 nfs_release_request(req);
666                                 if (error < 0) {
667                                         if (new)
668                                                 nfs_release_request(new);
669                                         return ERR_PTR(error);
670                                 }
671                                 continue;
672                         }
673                         spin_unlock(&nfsi->req_lock);
674                         if (new)
675                                 nfs_release_request(new);
676                         break;
677                 }
678
679                 if (new) {
680                         int error;
681                         nfs_lock_request_dontget(new);
682                         error = nfs_inode_add_request(inode, new);
683                         if (error) {
684                                 spin_unlock(&nfsi->req_lock);
685                                 nfs_unlock_request(new);
686                                 return ERR_PTR(error);
687                         }
688                         spin_unlock(&nfsi->req_lock);
689                         nfs_mark_request_dirty(new);
690                         return new;
691                 }
692                 spin_unlock(&nfsi->req_lock);
693
694                 new = nfs_create_request(ctx, inode, page, offset, bytes);
695                 if (IS_ERR(new))
696                         return new;
697         }
698
699         /* We have a request for our page.
700          * If the creds don't match, or the
701          * page addresses don't match,
702          * tell the caller to wait on the conflicting
703          * request.
704          */
705         rqend = req->wb_offset + req->wb_bytes;
706         if (req->wb_context != ctx
707             || req->wb_page != page
708             || !nfs_dirty_request(req)
709             || offset > rqend || end < req->wb_offset) {
710                 nfs_unlock_request(req);
711                 return ERR_PTR(-EBUSY);
712         }
713
714         /* Okay, the request matches. Update the region */
715         if (offset < req->wb_offset) {
716                 req->wb_offset = offset;
717                 req->wb_pgbase = offset;
718                 req->wb_bytes = rqend - req->wb_offset;
719         }
720
721         if (end > rqend)
722                 req->wb_bytes = end - req->wb_offset;
723
724         return req;
725 }
726
727 int nfs_flush_incompatible(struct file *file, struct page *page)
728 {
729         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
730         struct inode    *inode = page->mapping->host;
731         struct nfs_page *req;
732         int             status = 0;
733         /*
734          * Look for a request corresponding to this page. If there
735          * is one, and it belongs to another file, we flush it out
736          * before we try to copy anything into the page. Do this
737          * due to the lack of an ACCESS-type call in NFSv2.
738          * Also do the same if we find a request from an existing
739          * dropped page.
740          */
741         req = nfs_find_request(inode, page->index);
742         if (req) {
743                 if (req->wb_page != page || ctx != req->wb_context)
744                         status = nfs_wb_page(inode, page);
745                 nfs_release_request(req);
746         }
747         return (status < 0) ? status : 0;
748 }
749
750 /*
751  * Update and possibly write a cached page of an NFS file.
752  *
753  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
754  * things with a page scheduled for an RPC call (e.g. invalidate it).
755  */
756 int nfs_updatepage(struct file *file, struct page *page,
757                 unsigned int offset, unsigned int count)
758 {
759         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
760         struct inode    *inode = page->mapping->host;
761         struct nfs_page *req;
762         int             status = 0;
763
764         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
765
766         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
767                 file->f_dentry->d_parent->d_name.name,
768                 file->f_dentry->d_name.name, count,
769                 (long long)(page_offset(page) +offset));
770
771         if (IS_SYNC(inode)) {
772                 status = nfs_writepage_sync(ctx, inode, page, offset, count, 0);
773                 if (status > 0) {
774                         if (offset == 0 && status == PAGE_CACHE_SIZE)
775                                 SetPageUptodate(page);
776                         return 0;
777                 }
778                 return status;
779         }
780
781         /* If we're not using byte range locks, and we know the page
782          * is entirely in cache, it may be more efficient to avoid
783          * fragmenting write requests.
784          */
785         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
786                 loff_t end_offs = i_size_read(inode) - 1;
787                 unsigned long end_index = end_offs >> PAGE_CACHE_SHIFT;
788
789                 count += offset;
790                 offset = 0;
791                 if (unlikely(end_offs < 0)) {
792                         /* Do nothing */
793                 } else if (page->index == end_index) {
794                         unsigned int pglen;
795                         pglen = (unsigned int)(end_offs & (PAGE_CACHE_SIZE-1)) + 1;
796                         if (count < pglen)
797                                 count = pglen;
798                 } else if (page->index < end_index)
799                         count = PAGE_CACHE_SIZE;
800         }
801
802         /*
803          * Try to find an NFS request corresponding to this page
804          * and update it.
805          * If the existing request cannot be updated, we must flush
806          * it out now.
807          */
808         do {
809                 req = nfs_update_request(ctx, inode, page, offset, count);
810                 status = (IS_ERR(req)) ? PTR_ERR(req) : 0;
811                 if (status != -EBUSY)
812                         break;
813                 /* Request could not be updated. Flush it out and try again */
814                 status = nfs_wb_page(inode, page);
815         } while (status >= 0);
816         if (status < 0)
817                 goto done;
818
819         status = 0;
820
821         /* Update file length */
822         nfs_grow_file(page, offset, count);
823         /* Set the PG_uptodate flag? */
824         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
825         nfs_unlock_request(req);
826 done:
827         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
828                         status, (long long)i_size_read(inode));
829         if (status < 0)
830                 ClearPageUptodate(page);
831         return status;
832 }
833
834 static void nfs_writepage_release(struct nfs_page *req)
835 {
836         end_page_writeback(req->wb_page);
837
838 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
839         if (!PageError(req->wb_page)) {
840                 if (NFS_NEED_RESCHED(req)) {
841                         nfs_mark_request_dirty(req);
842                         goto out;
843                 } else if (NFS_NEED_COMMIT(req)) {
844                         nfs_mark_request_commit(req);
845                         goto out;
846                 }
847         }
848         nfs_inode_remove_request(req);
849
850 out:
851         nfs_clear_commit(req);
852         nfs_clear_reschedule(req);
853 #else
854         nfs_inode_remove_request(req);
855 #endif
856         nfs_clear_page_writeback(req);
857 }
858
859 static inline int flush_task_priority(int how)
860 {
861         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
862                 case FLUSH_HIGHPRI:
863                         return RPC_PRIORITY_HIGH;
864                 case FLUSH_LOWPRI:
865                         return RPC_PRIORITY_LOW;
866         }
867         return RPC_PRIORITY_NORMAL;
868 }
869
870 /*
871  * Set up the argument/result storage required for the RPC call.
872  */
873 static void nfs_write_rpcsetup(struct nfs_page *req,
874                 struct nfs_write_data *data,
875                 unsigned int count, unsigned int offset,
876                 int how)
877 {
878         struct inode            *inode;
879
880         /* Set up the RPC argument and reply structs
881          * NB: take care not to mess about with data->commit et al. */
882
883         data->req = req;
884         data->inode = inode = req->wb_context->dentry->d_inode;
885         data->cred = req->wb_context->cred;
886
887         data->args.fh     = NFS_FH(inode);
888         data->args.offset = req_offset(req) + offset;
889         data->args.pgbase = req->wb_pgbase + offset;
890         data->args.pages  = data->pagevec;
891         data->args.count  = count;
892         data->args.context = req->wb_context;
893
894         data->res.fattr   = &data->fattr;
895         data->res.count   = count;
896         data->res.verf    = &data->verf;
897         nfs_fattr_init(&data->fattr);
898
899         NFS_PROTO(inode)->write_setup(data, how);
900
901         data->task.tk_priority = flush_task_priority(how);
902         data->task.tk_cookie = (unsigned long)inode;
903
904         dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
905                 data->task.tk_pid,
906                 inode->i_sb->s_id,
907                 (long long)NFS_FILEID(inode),
908                 count,
909                 (unsigned long long)data->args.offset);
910 }
911
912 static void nfs_execute_write(struct nfs_write_data *data)
913 {
914         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
915         sigset_t oldset;
916
917         rpc_clnt_sigmask(clnt, &oldset);
918         lock_kernel();
919         rpc_execute(&data->task);
920         unlock_kernel();
921         rpc_clnt_sigunmask(clnt, &oldset);
922 }
923
924 /*
925  * Generate multiple small requests to write out a single
926  * contiguous dirty area on one page.
927  */
928 static int nfs_flush_multi(struct list_head *head, struct inode *inode, int how)
929 {
930         struct nfs_page *req = nfs_list_entry(head->next);
931         struct page *page = req->wb_page;
932         struct nfs_write_data *data;
933         unsigned int wsize = NFS_SERVER(inode)->wsize;
934         unsigned int nbytes, offset;
935         int requests = 0;
936         LIST_HEAD(list);
937
938         nfs_list_remove_request(req);
939
940         nbytes = req->wb_bytes;
941         for (;;) {
942                 data = nfs_writedata_alloc(1);
943                 if (!data)
944                         goto out_bad;
945                 list_add(&data->pages, &list);
946                 requests++;
947                 if (nbytes <= wsize)
948                         break;
949                 nbytes -= wsize;
950         }
951         atomic_set(&req->wb_complete, requests);
952
953         ClearPageError(page);
954         set_page_writeback(page);
955         offset = 0;
956         nbytes = req->wb_bytes;
957         do {
958                 data = list_entry(list.next, struct nfs_write_data, pages);
959                 list_del_init(&data->pages);
960
961                 data->pagevec[0] = page;
962                 data->complete = nfs_writeback_done_partial;
963
964                 if (nbytes > wsize) {
965                         nfs_write_rpcsetup(req, data, wsize, offset, how);
966                         offset += wsize;
967                         nbytes -= wsize;
968                 } else {
969                         nfs_write_rpcsetup(req, data, nbytes, offset, how);
970                         nbytes = 0;
971                 }
972                 nfs_execute_write(data);
973         } while (nbytes != 0);
974
975         return 0;
976
977 out_bad:
978         while (!list_empty(&list)) {
979                 data = list_entry(list.next, struct nfs_write_data, pages);
980                 list_del(&data->pages);
981                 nfs_writedata_free(data);
982         }
983         nfs_mark_request_dirty(req);
984         nfs_clear_page_writeback(req);
985         return -ENOMEM;
986 }
987
988 /*
989  * Create an RPC task for the given write request and kick it.
990  * The page must have been locked by the caller.
991  *
992  * It may happen that the page we're passed is not marked dirty.
993  * This is the case if nfs_updatepage detects a conflicting request
994  * that has been written but not committed.
995  */
996 static int nfs_flush_one(struct list_head *head, struct inode *inode, int how)
997 {
998         struct nfs_page         *req;
999         struct page             **pages;
1000         struct nfs_write_data   *data;
1001         unsigned int            count;
1002
1003         if (NFS_SERVER(inode)->wsize < PAGE_CACHE_SIZE)
1004                 return nfs_flush_multi(head, inode, how);
1005
1006         data = nfs_writedata_alloc(NFS_SERVER(inode)->wpages);
1007         if (!data)
1008                 goto out_bad;
1009
1010         pages = data->pagevec;
1011         count = 0;
1012         while (!list_empty(head)) {
1013                 req = nfs_list_entry(head->next);
1014                 nfs_list_remove_request(req);
1015                 nfs_list_add_request(req, &data->pages);
1016                 ClearPageError(req->wb_page);
1017                 set_page_writeback(req->wb_page);
1018                 *pages++ = req->wb_page;
1019                 count += req->wb_bytes;
1020         }
1021         req = nfs_list_entry(data->pages.next);
1022
1023         data->complete = nfs_writeback_done_full;
1024         /* Set up the argument struct */
1025         nfs_write_rpcsetup(req, data, count, 0, how);
1026
1027         nfs_execute_write(data);
1028         return 0;
1029  out_bad:
1030         while (!list_empty(head)) {
1031                 struct nfs_page *req = nfs_list_entry(head->next);
1032                 nfs_list_remove_request(req);
1033                 nfs_mark_request_dirty(req);
1034                 nfs_clear_page_writeback(req);
1035         }
1036         return -ENOMEM;
1037 }
1038
1039 static int
1040 nfs_flush_list(struct list_head *head, int wpages, int how)
1041 {
1042         LIST_HEAD(one_request);
1043         struct nfs_page         *req;
1044         int                     error = 0;
1045         unsigned int            pages = 0;
1046
1047         while (!list_empty(head)) {
1048                 pages += nfs_coalesce_requests(head, &one_request, wpages);
1049                 req = nfs_list_entry(one_request.next);
1050                 error = nfs_flush_one(&one_request, req->wb_context->dentry->d_inode, how);
1051                 if (error < 0)
1052                         break;
1053         }
1054         if (error >= 0)
1055                 return pages;
1056
1057         while (!list_empty(head)) {
1058                 req = nfs_list_entry(head->next);
1059                 nfs_list_remove_request(req);
1060                 nfs_mark_request_dirty(req);
1061                 nfs_clear_page_writeback(req);
1062         }
1063         return error;
1064 }
1065
1066 /*
1067  * Handle a write reply that flushed part of a page.
1068  */
1069 static void nfs_writeback_done_partial(struct nfs_write_data *data, int status)
1070 {
1071         struct nfs_page         *req = data->req;
1072         struct page             *page = req->wb_page;
1073
1074         dprintk("NFS: write (%s/%Ld %d@%Ld)",
1075                 req->wb_context->dentry->d_inode->i_sb->s_id,
1076                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1077                 req->wb_bytes,
1078                 (long long)req_offset(req));
1079
1080         if (status < 0) {
1081                 ClearPageUptodate(page);
1082                 SetPageError(page);
1083                 req->wb_context->error = status;
1084                 dprintk(", error = %d\n", status);
1085         } else {
1086 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1087                 if (data->verf.committed < NFS_FILE_SYNC) {
1088                         if (!NFS_NEED_COMMIT(req)) {
1089                                 nfs_defer_commit(req);
1090                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1091                                 dprintk(" defer commit\n");
1092                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1093                                 nfs_defer_reschedule(req);
1094                                 dprintk(" server reboot detected\n");
1095                         }
1096                 } else
1097 #endif
1098                         dprintk(" OK\n");
1099         }
1100
1101         if (atomic_dec_and_test(&req->wb_complete))
1102                 nfs_writepage_release(req);
1103 }
1104
1105 /*
1106  * Handle a write reply that flushes a whole page.
1107  *
1108  * FIXME: There is an inherent race with invalidate_inode_pages and
1109  *        writebacks since the page->count is kept > 1 for as long
1110  *        as the page has a write request pending.
1111  */
1112 static void nfs_writeback_done_full(struct nfs_write_data *data, int status)
1113 {
1114         struct nfs_page         *req;
1115         struct page             *page;
1116
1117         /* Update attributes as result of writeback. */
1118         while (!list_empty(&data->pages)) {
1119                 req = nfs_list_entry(data->pages.next);
1120                 nfs_list_remove_request(req);
1121                 page = req->wb_page;
1122
1123                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1124                         req->wb_context->dentry->d_inode->i_sb->s_id,
1125                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1126                         req->wb_bytes,
1127                         (long long)req_offset(req));
1128
1129                 if (status < 0) {
1130                         ClearPageUptodate(page);
1131                         SetPageError(page);
1132                         req->wb_context->error = status;
1133                         end_page_writeback(page);
1134                         nfs_inode_remove_request(req);
1135                         dprintk(", error = %d\n", status);
1136                         goto next;
1137                 }
1138                 end_page_writeback(page);
1139
1140 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1141                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1142                         nfs_inode_remove_request(req);
1143                         dprintk(" OK\n");
1144                         goto next;
1145                 }
1146                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1147                 nfs_mark_request_commit(req);
1148                 dprintk(" marked for commit\n");
1149 #else
1150                 nfs_inode_remove_request(req);
1151 #endif
1152         next:
1153                 nfs_clear_page_writeback(req);
1154         }
1155 }
1156
1157 /*
1158  * This function is called when the WRITE call is complete.
1159  */
1160 void nfs_writeback_done(struct rpc_task *task, void *calldata)
1161 {
1162         struct nfs_write_data   *data = calldata;
1163         struct nfs_writeargs    *argp = &data->args;
1164         struct nfs_writeres     *resp = &data->res;
1165
1166         dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1167                 task->tk_pid, task->tk_status);
1168
1169         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1170
1171 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1172         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1173                 /* We tried a write call, but the server did not
1174                  * commit data to stable storage even though we
1175                  * requested it.
1176                  * Note: There is a known bug in Tru64 < 5.0 in which
1177                  *       the server reports NFS_DATA_SYNC, but performs
1178                  *       NFS_FILE_SYNC. We therefore implement this checking
1179                  *       as a dprintk() in order to avoid filling syslog.
1180                  */
1181                 static unsigned long    complain;
1182
1183                 if (time_before(complain, jiffies)) {
1184                         dprintk("NFS: faulty NFS server %s:"
1185                                 " (committed = %d) != (stable = %d)\n",
1186                                 NFS_SERVER(data->inode)->hostname,
1187                                 resp->verf->committed, argp->stable);
1188                         complain = jiffies + 300 * HZ;
1189                 }
1190         }
1191 #endif
1192         /* Is this a short write? */
1193         if (task->tk_status >= 0 && resp->count < argp->count) {
1194                 static unsigned long    complain;
1195
1196                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1197
1198                 /* Has the server at least made some progress? */
1199                 if (resp->count != 0) {
1200                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1201                         if (resp->verf->committed != NFS_UNSTABLE) {
1202                                 /* Resend from where the server left off */
1203                                 argp->offset += resp->count;
1204                                 argp->pgbase += resp->count;
1205                                 argp->count -= resp->count;
1206                         } else {
1207                                 /* Resend as a stable write in order to avoid
1208                                  * headaches in the case of a server crash.
1209                                  */
1210                                 argp->stable = NFS_FILE_SYNC;
1211                         }
1212                         rpc_restart_call(task);
1213                         return;
1214                 }
1215                 if (time_before(complain, jiffies)) {
1216                         printk(KERN_WARNING
1217                                "NFS: Server wrote zero bytes, expected %u.\n",
1218                                         argp->count);
1219                         complain = jiffies + 300 * HZ;
1220                 }
1221                 /* Can't do anything about it except throw an error. */
1222                 task->tk_status = -EIO;
1223         }
1224
1225         /*
1226          * Process the nfs_page list
1227          */
1228         data->complete(data, task->tk_status);
1229 }
1230
1231
1232 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1233 void nfs_commit_release(void *wdata)
1234 {
1235         nfs_commit_free(wdata);
1236 }
1237
1238 /*
1239  * Set up the argument/result storage required for the RPC call.
1240  */
1241 static void nfs_commit_rpcsetup(struct list_head *head,
1242                 struct nfs_write_data *data, int how)
1243 {
1244         struct nfs_page         *first;
1245         struct inode            *inode;
1246
1247         /* Set up the RPC argument and reply structs
1248          * NB: take care not to mess about with data->commit et al. */
1249
1250         list_splice_init(head, &data->pages);
1251         first = nfs_list_entry(data->pages.next);
1252         inode = first->wb_context->dentry->d_inode;
1253
1254         data->inode       = inode;
1255         data->cred        = first->wb_context->cred;
1256
1257         data->args.fh     = NFS_FH(data->inode);
1258         /* Note: we always request a commit of the entire inode */
1259         data->args.offset = 0;
1260         data->args.count  = 0;
1261         data->res.count   = 0;
1262         data->res.fattr   = &data->fattr;
1263         data->res.verf    = &data->verf;
1264         nfs_fattr_init(&data->fattr);
1265         
1266         NFS_PROTO(inode)->commit_setup(data, how);
1267
1268         data->task.tk_priority = flush_task_priority(how);
1269         data->task.tk_cookie = (unsigned long)inode;
1270         
1271         dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1272 }
1273
1274 /*
1275  * Commit dirty pages
1276  */
1277 static int
1278 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1279 {
1280         struct nfs_write_data   *data;
1281         struct nfs_page         *req;
1282
1283         data = nfs_commit_alloc(NFS_SERVER(inode)->wpages);
1284
1285         if (!data)
1286                 goto out_bad;
1287
1288         /* Set up the argument struct */
1289         nfs_commit_rpcsetup(head, data, how);
1290
1291         nfs_execute_write(data);
1292         return 0;
1293  out_bad:
1294         while (!list_empty(head)) {
1295                 req = nfs_list_entry(head->next);
1296                 nfs_list_remove_request(req);
1297                 nfs_mark_request_commit(req);
1298                 nfs_clear_page_writeback(req);
1299         }
1300         return -ENOMEM;
1301 }
1302
1303 /*
1304  * COMMIT call returned
1305  */
1306 void nfs_commit_done(struct rpc_task *task, void *calldata)
1307 {
1308         struct nfs_write_data   *data = calldata;
1309         struct nfs_page         *req;
1310         int res = 0;
1311
1312         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1313                                 task->tk_pid, task->tk_status);
1314
1315         while (!list_empty(&data->pages)) {
1316                 req = nfs_list_entry(data->pages.next);
1317                 nfs_list_remove_request(req);
1318
1319                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1320                         req->wb_context->dentry->d_inode->i_sb->s_id,
1321                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1322                         req->wb_bytes,
1323                         (long long)req_offset(req));
1324                 if (task->tk_status < 0) {
1325                         req->wb_context->error = task->tk_status;
1326                         nfs_inode_remove_request(req);
1327                         dprintk(", error = %d\n", task->tk_status);
1328                         goto next;
1329                 }
1330
1331                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1332                  * returned by the server against all stored verfs. */
1333                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1334                         /* We have a match */
1335                         nfs_inode_remove_request(req);
1336                         dprintk(" OK\n");
1337                         goto next;
1338                 }
1339                 /* We have a mismatch. Write the page again */
1340                 dprintk(" mismatch\n");
1341                 nfs_mark_request_dirty(req);
1342         next:
1343                 nfs_clear_page_writeback(req);
1344                 res++;
1345         }
1346         sub_page_state(nr_unstable,res);
1347 }
1348 #endif
1349
1350 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
1351                            unsigned int npages, int how)
1352 {
1353         struct nfs_inode *nfsi = NFS_I(inode);
1354         LIST_HEAD(head);
1355         int                     res,
1356                                 error = 0;
1357
1358         spin_lock(&nfsi->req_lock);
1359         res = nfs_scan_dirty(inode, &head, idx_start, npages);
1360         spin_unlock(&nfsi->req_lock);
1361         if (res) {
1362                 struct nfs_server *server = NFS_SERVER(inode);
1363
1364                 /* For single writes, FLUSH_STABLE is more efficient */
1365                 if (res == nfsi->npages && nfsi->npages <= server->wpages) {
1366                         if (res > 1 || nfs_list_entry(head.next)->wb_bytes <= server->wsize)
1367                                 how |= FLUSH_STABLE;
1368                 }
1369                 error = nfs_flush_list(&head, server->wpages, how);
1370         }
1371         if (error < 0)
1372                 return error;
1373         return res;
1374 }
1375
1376 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1377 int nfs_commit_inode(struct inode *inode, int how)
1378 {
1379         struct nfs_inode *nfsi = NFS_I(inode);
1380         LIST_HEAD(head);
1381         int                     res,
1382                                 error = 0;
1383
1384         spin_lock(&nfsi->req_lock);
1385         res = nfs_scan_commit(inode, &head, 0, 0);
1386         spin_unlock(&nfsi->req_lock);
1387         if (res) {
1388                 error = nfs_commit_list(inode, &head, how);
1389                 if (error < 0)
1390                         return error;
1391         }
1392         return res;
1393 }
1394 #endif
1395
1396 int nfs_sync_inode(struct inode *inode, unsigned long idx_start,
1397                   unsigned int npages, int how)
1398 {
1399         int nocommit = how & FLUSH_NOCOMMIT;
1400         int wait = how & FLUSH_WAIT;
1401         int error;
1402
1403         how &= ~(FLUSH_WAIT|FLUSH_NOCOMMIT);
1404
1405         do {
1406                 if (wait) {
1407                         error = nfs_wait_on_requests(inode, idx_start, npages);
1408                         if (error != 0)
1409                                 continue;
1410                 }
1411                 error = nfs_flush_inode(inode, idx_start, npages, how);
1412                 if (error != 0)
1413                         continue;
1414                 if (!nocommit)
1415                         error = nfs_commit_inode(inode, how);
1416         } while (error > 0);
1417         return error;
1418 }
1419
1420 int nfs_init_writepagecache(void)
1421 {
1422         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1423                                              sizeof(struct nfs_write_data),
1424                                              0, SLAB_HWCACHE_ALIGN,
1425                                              NULL, NULL);
1426         if (nfs_wdata_cachep == NULL)
1427                 return -ENOMEM;
1428
1429         nfs_wdata_mempool = mempool_create(MIN_POOL_WRITE,
1430                                            mempool_alloc_slab,
1431                                            mempool_free_slab,
1432                                            nfs_wdata_cachep);
1433         if (nfs_wdata_mempool == NULL)
1434                 return -ENOMEM;
1435
1436         nfs_commit_mempool = mempool_create(MIN_POOL_COMMIT,
1437                                            mempool_alloc_slab,
1438                                            mempool_free_slab,
1439                                            nfs_wdata_cachep);
1440         if (nfs_commit_mempool == NULL)
1441                 return -ENOMEM;
1442
1443         return 0;
1444 }
1445
1446 void nfs_destroy_writepagecache(void)
1447 {
1448         mempool_destroy(nfs_commit_mempool);
1449         mempool_destroy(nfs_wdata_mempool);
1450         if (kmem_cache_destroy(nfs_wdata_cachep))
1451                 printk(KERN_INFO "nfs_write_data: not all structures were freed\n");
1452 }
1453