[PATCH] zoned vm counters: conversion of nr_unstable to per zone counter
[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 int nfs_wait_on_write_congestion(struct address_space *, int);
81 static int nfs_wait_on_requests(struct inode *, unsigned long, unsigned int);
82 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
83                            unsigned int npages, int how);
84 static const struct rpc_call_ops nfs_write_partial_ops;
85 static const struct rpc_call_ops nfs_write_full_ops;
86 static const struct rpc_call_ops nfs_commit_ops;
87
88 static kmem_cache_t *nfs_wdata_cachep;
89 static mempool_t *nfs_wdata_mempool;
90 static mempool_t *nfs_commit_mempool;
91
92 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
93
94 struct nfs_write_data *nfs_commit_alloc(unsigned int pagecount)
95 {
96         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS);
97
98         if (p) {
99                 memset(p, 0, sizeof(*p));
100                 INIT_LIST_HEAD(&p->pages);
101                 if (pagecount <= ARRAY_SIZE(p->page_array))
102                         p->pagevec = p->page_array;
103                 else {
104                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), 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 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 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
122 {
123         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, SLAB_NOFS);
124
125         if (p) {
126                 memset(p, 0, sizeof(*p));
127                 INIT_LIST_HEAD(&p->pages);
128                 if (pagecount <= ARRAY_SIZE(p->page_array))
129                         p->pagevec = p->page_array;
130                 else {
131                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
132                         if (!p->pagevec) {
133                                 mempool_free(p, nfs_wdata_mempool);
134                                 p = NULL;
135                         }
136                 }
137         }
138         return p;
139 }
140
141 void nfs_writedata_free(struct nfs_write_data *p)
142 {
143         if (p && (p->pagevec != &p->page_array[0]))
144                 kfree(p->pagevec);
145         mempool_free(p, nfs_wdata_mempool);
146 }
147
148 void nfs_writedata_release(void *wdata)
149 {
150         nfs_writedata_free(wdata);
151 }
152
153 /* Adjust the file length if we're writing beyond the end */
154 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
155 {
156         struct inode *inode = page->mapping->host;
157         loff_t end, i_size = i_size_read(inode);
158         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
159
160         if (i_size > 0 && page->index < end_index)
161                 return;
162         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
163         if (i_size >= end)
164                 return;
165         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
166         i_size_write(inode, end);
167 }
168
169 /* We can set the PG_uptodate flag if we see that a write request
170  * covers the full page.
171  */
172 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
173 {
174         loff_t end_offs;
175
176         if (PageUptodate(page))
177                 return;
178         if (base != 0)
179                 return;
180         if (count == PAGE_CACHE_SIZE) {
181                 SetPageUptodate(page);
182                 return;
183         }
184
185         end_offs = i_size_read(page->mapping->host) - 1;
186         if (end_offs < 0)
187                 return;
188         /* Is this the last page? */
189         if (page->index != (unsigned long)(end_offs >> PAGE_CACHE_SHIFT))
190                 return;
191         /* This is the last page: set PG_uptodate if we cover the entire
192          * extent of the data, then zero the rest of the page.
193          */
194         if (count == (unsigned int)(end_offs & (PAGE_CACHE_SIZE - 1)) + 1) {
195                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
196                 SetPageUptodate(page);
197         }
198 }
199
200 /*
201  * Write a page synchronously.
202  * Offset is the data offset within the page.
203  */
204 static int nfs_writepage_sync(struct nfs_open_context *ctx, struct inode *inode,
205                 struct page *page, unsigned int offset, unsigned int count,
206                 int how)
207 {
208         unsigned int    wsize = NFS_SERVER(inode)->wsize;
209         int             result, written = 0;
210         struct nfs_write_data *wdata;
211
212         wdata = nfs_writedata_alloc(1);
213         if (!wdata)
214                 return -ENOMEM;
215
216         wdata->flags = how;
217         wdata->cred = ctx->cred;
218         wdata->inode = inode;
219         wdata->args.fh = NFS_FH(inode);
220         wdata->args.context = ctx;
221         wdata->args.pages = &page;
222         wdata->args.stable = NFS_FILE_SYNC;
223         wdata->args.pgbase = offset;
224         wdata->args.count = wsize;
225         wdata->res.fattr = &wdata->fattr;
226         wdata->res.verf = &wdata->verf;
227
228         dprintk("NFS:      nfs_writepage_sync(%s/%Ld %d@%Ld)\n",
229                 inode->i_sb->s_id,
230                 (long long)NFS_FILEID(inode),
231                 count, (long long)(page_offset(page) + offset));
232
233         set_page_writeback(page);
234         nfs_begin_data_update(inode);
235         do {
236                 if (count < wsize)
237                         wdata->args.count = count;
238                 wdata->args.offset = page_offset(page) + wdata->args.pgbase;
239
240                 result = NFS_PROTO(inode)->write(wdata);
241
242                 if (result < 0) {
243                         /* Must mark the page invalid after I/O error */
244                         ClearPageUptodate(page);
245                         goto io_error;
246                 }
247                 if (result < wdata->args.count)
248                         printk(KERN_WARNING "NFS: short write, count=%u, result=%d\n",
249                                         wdata->args.count, result);
250
251                 wdata->args.offset += result;
252                 wdata->args.pgbase += result;
253                 written += result;
254                 count -= result;
255                 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, result);
256         } while (count);
257         /* Update file length */
258         nfs_grow_file(page, offset, written);
259         /* Set the PG_uptodate flag? */
260         nfs_mark_uptodate(page, offset, written);
261
262         if (PageError(page))
263                 ClearPageError(page);
264
265 io_error:
266         nfs_end_data_update(inode);
267         end_page_writeback(page);
268         nfs_writedata_free(wdata);
269         return written ? written : result;
270 }
271
272 static int nfs_writepage_async(struct nfs_open_context *ctx,
273                 struct inode *inode, struct page *page,
274                 unsigned int offset, unsigned int count)
275 {
276         struct nfs_page *req;
277
278         req = nfs_update_request(ctx, inode, page, offset, count);
279         if (IS_ERR(req))
280                 return PTR_ERR(req);
281         /* Update file length */
282         nfs_grow_file(page, offset, count);
283         /* Set the PG_uptodate flag? */
284         nfs_mark_uptodate(page, offset, count);
285         nfs_unlock_request(req);
286         return 0;
287 }
288
289 static int wb_priority(struct writeback_control *wbc)
290 {
291         if (wbc->for_reclaim)
292                 return FLUSH_HIGHPRI;
293         if (wbc->for_kupdate)
294                 return FLUSH_LOWPRI;
295         return 0;
296 }
297
298 /*
299  * Write an mmapped page to the server.
300  */
301 int nfs_writepage(struct page *page, struct writeback_control *wbc)
302 {
303         struct nfs_open_context *ctx;
304         struct inode *inode = page->mapping->host;
305         unsigned long end_index;
306         unsigned offset = PAGE_CACHE_SIZE;
307         loff_t i_size = i_size_read(inode);
308         int inode_referenced = 0;
309         int priority = wb_priority(wbc);
310         int err;
311
312         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
313         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
314
315         /*
316          * Note: We need to ensure that we have a reference to the inode
317          *       if we are to do asynchronous writes. If not, waiting
318          *       in nfs_wait_on_request() may deadlock with clear_inode().
319          *
320          *       If igrab() fails here, then it is in any case safe to
321          *       call nfs_wb_page(), since there will be no pending writes.
322          */
323         if (igrab(inode) != 0)
324                 inode_referenced = 1;
325         end_index = i_size >> PAGE_CACHE_SHIFT;
326
327         /* Ensure we've flushed out any previous writes */
328         nfs_wb_page_priority(inode, page, priority);
329
330         /* easy case */
331         if (page->index < end_index)
332                 goto do_it;
333         /* things got complicated... */
334         offset = i_size & (PAGE_CACHE_SIZE-1);
335
336         /* OK, are we completely out? */
337         err = 0; /* potential race with truncate - ignore */
338         if (page->index >= end_index+1 || !offset)
339                 goto out;
340 do_it:
341         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
342         if (ctx == NULL) {
343                 err = -EBADF;
344                 goto out;
345         }
346         lock_kernel();
347         if (!IS_SYNC(inode) && inode_referenced) {
348                 err = nfs_writepage_async(ctx, inode, page, 0, offset);
349                 if (!wbc->for_writepages)
350                         nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
351         } else {
352                 err = nfs_writepage_sync(ctx, inode, page, 0,
353                                                 offset, priority);
354                 if (err >= 0) {
355                         if (err != offset)
356                                 redirty_page_for_writepage(wbc, page);
357                         err = 0;
358                 }
359         }
360         unlock_kernel();
361         put_nfs_open_context(ctx);
362 out:
363         unlock_page(page);
364         if (inode_referenced)
365                 iput(inode);
366         return err; 
367 }
368
369 /*
370  * Note: causes nfs_update_request() to block on the assumption
371  *       that the writeback is generated due to memory pressure.
372  */
373 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
374 {
375         struct backing_dev_info *bdi = mapping->backing_dev_info;
376         struct inode *inode = mapping->host;
377         int err;
378
379         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
380
381         err = generic_writepages(mapping, wbc);
382         if (err)
383                 return err;
384         while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
385                 if (wbc->nonblocking)
386                         return 0;
387                 nfs_wait_on_write_congestion(mapping, 0);
388         }
389         err = nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
390         if (err < 0)
391                 goto out;
392         nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
393         wbc->nr_to_write -= err;
394         if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) {
395                 err = nfs_wait_on_requests(inode, 0, 0);
396                 if (err < 0)
397                         goto out;
398         }
399         err = nfs_commit_inode(inode, wb_priority(wbc));
400         if (err > 0) {
401                 wbc->nr_to_write -= err;
402                 err = 0;
403         }
404 out:
405         clear_bit(BDI_write_congested, &bdi->state);
406         wake_up_all(&nfs_write_congestion);
407         return err;
408 }
409
410 /*
411  * Insert a write request into an inode
412  */
413 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
414 {
415         struct nfs_inode *nfsi = NFS_I(inode);
416         int error;
417
418         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
419         BUG_ON(error == -EEXIST);
420         if (error)
421                 return error;
422         if (!nfsi->npages) {
423                 igrab(inode);
424                 nfs_begin_data_update(inode);
425                 if (nfs_have_delegation(inode, FMODE_WRITE))
426                         nfsi->change_attr++;
427         }
428         SetPagePrivate(req->wb_page);
429         nfsi->npages++;
430         atomic_inc(&req->wb_count);
431         return 0;
432 }
433
434 /*
435  * Insert a write request into an inode
436  */
437 static void nfs_inode_remove_request(struct nfs_page *req)
438 {
439         struct inode *inode = req->wb_context->dentry->d_inode;
440         struct nfs_inode *nfsi = NFS_I(inode);
441
442         BUG_ON (!NFS_WBACK_BUSY(req));
443
444         spin_lock(&nfsi->req_lock);
445         ClearPagePrivate(req->wb_page);
446         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
447         nfsi->npages--;
448         if (!nfsi->npages) {
449                 spin_unlock(&nfsi->req_lock);
450                 nfs_end_data_update(inode);
451                 iput(inode);
452         } else
453                 spin_unlock(&nfsi->req_lock);
454         nfs_clear_request(req);
455         nfs_release_request(req);
456 }
457
458 /*
459  * Find a request
460  */
461 static inline struct nfs_page *
462 _nfs_find_request(struct inode *inode, unsigned long index)
463 {
464         struct nfs_inode *nfsi = NFS_I(inode);
465         struct nfs_page *req;
466
467         req = (struct nfs_page*)radix_tree_lookup(&nfsi->nfs_page_tree, index);
468         if (req)
469                 atomic_inc(&req->wb_count);
470         return req;
471 }
472
473 static struct nfs_page *
474 nfs_find_request(struct inode *inode, unsigned long index)
475 {
476         struct nfs_page         *req;
477         struct nfs_inode        *nfsi = NFS_I(inode);
478
479         spin_lock(&nfsi->req_lock);
480         req = _nfs_find_request(inode, index);
481         spin_unlock(&nfsi->req_lock);
482         return req;
483 }
484
485 /*
486  * Add a request to the inode's dirty list.
487  */
488 static void
489 nfs_mark_request_dirty(struct nfs_page *req)
490 {
491         struct inode *inode = req->wb_context->dentry->d_inode;
492         struct nfs_inode *nfsi = NFS_I(inode);
493
494         spin_lock(&nfsi->req_lock);
495         radix_tree_tag_set(&nfsi->nfs_page_tree,
496                         req->wb_index, NFS_PAGE_TAG_DIRTY);
497         nfs_list_add_request(req, &nfsi->dirty);
498         nfsi->ndirty++;
499         spin_unlock(&nfsi->req_lock);
500         inc_zone_page_state(req->wb_page, NR_FILE_DIRTY);
501         mark_inode_dirty(inode);
502 }
503
504 /*
505  * Check if a request is dirty
506  */
507 static inline int
508 nfs_dirty_request(struct nfs_page *req)
509 {
510         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
511         return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty;
512 }
513
514 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
515 /*
516  * Add a request to the inode's commit list.
517  */
518 static void
519 nfs_mark_request_commit(struct nfs_page *req)
520 {
521         struct inode *inode = req->wb_context->dentry->d_inode;
522         struct nfs_inode *nfsi = NFS_I(inode);
523
524         spin_lock(&nfsi->req_lock);
525         nfs_list_add_request(req, &nfsi->commit);
526         nfsi->ncommit++;
527         spin_unlock(&nfsi->req_lock);
528         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
529         mark_inode_dirty(inode);
530 }
531 #endif
532
533 /*
534  * Wait for a request to complete.
535  *
536  * Interruptible by signals only if mounted with intr flag.
537  */
538 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
539 {
540         struct nfs_inode *nfsi = NFS_I(inode);
541         struct nfs_page *req;
542         unsigned long           idx_end, next;
543         unsigned int            res = 0;
544         int                     error;
545
546         if (npages == 0)
547                 idx_end = ~0;
548         else
549                 idx_end = idx_start + npages - 1;
550
551         next = idx_start;
552         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
553                 if (req->wb_index > idx_end)
554                         break;
555
556                 next = req->wb_index + 1;
557                 BUG_ON(!NFS_WBACK_BUSY(req));
558
559                 atomic_inc(&req->wb_count);
560                 spin_unlock(&nfsi->req_lock);
561                 error = nfs_wait_on_request(req);
562                 nfs_release_request(req);
563                 spin_lock(&nfsi->req_lock);
564                 if (error < 0)
565                         return error;
566                 res++;
567         }
568         return res;
569 }
570
571 static int nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
572 {
573         struct nfs_inode *nfsi = NFS_I(inode);
574         int ret;
575
576         spin_lock(&nfsi->req_lock);
577         ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
578         spin_unlock(&nfsi->req_lock);
579         return ret;
580 }
581
582 static void nfs_cancel_requests(struct list_head *head)
583 {
584         struct nfs_page *req;
585         while(!list_empty(head)) {
586                 req = nfs_list_entry(head->next);
587                 nfs_list_remove_request(req);
588                 nfs_inode_remove_request(req);
589                 nfs_clear_page_writeback(req);
590         }
591 }
592
593 /*
594  * nfs_scan_dirty - Scan an inode for dirty requests
595  * @inode: NFS inode to scan
596  * @dst: destination list
597  * @idx_start: lower bound of page->index to scan.
598  * @npages: idx_start + npages sets the upper bound to scan.
599  *
600  * Moves requests from the inode's dirty page list.
601  * The requests are *not* checked to ensure that they form a contiguous set.
602  */
603 static int
604 nfs_scan_dirty(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
605 {
606         struct nfs_inode *nfsi = NFS_I(inode);
607         int res = 0;
608
609         if (nfsi->ndirty != 0) {
610                 res = nfs_scan_lock_dirty(nfsi, dst, idx_start, npages);
611                 nfsi->ndirty -= res;
612                 if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty))
613                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n");
614         }
615         return res;
616 }
617
618 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
619 /*
620  * nfs_scan_commit - Scan an inode for commit requests
621  * @inode: NFS inode to scan
622  * @dst: destination list
623  * @idx_start: lower bound of page->index to scan.
624  * @npages: idx_start + npages sets the upper bound to scan.
625  *
626  * Moves requests from the inode's 'commit' request list.
627  * The requests are *not* checked to ensure that they form a contiguous set.
628  */
629 static int
630 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
631 {
632         struct nfs_inode *nfsi = NFS_I(inode);
633         int res = 0;
634
635         if (nfsi->ncommit != 0) {
636                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
637                 nfsi->ncommit -= res;
638                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
639                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
640         }
641         return res;
642 }
643 #else
644 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
645 {
646         return 0;
647 }
648 #endif
649
650 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
651 {
652         struct backing_dev_info *bdi = mapping->backing_dev_info;
653         DEFINE_WAIT(wait);
654         int ret = 0;
655
656         might_sleep();
657
658         if (!bdi_write_congested(bdi))
659                 return 0;
660
661         nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
662
663         if (intr) {
664                 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
665                 sigset_t oldset;
666
667                 rpc_clnt_sigmask(clnt, &oldset);
668                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
669                 if (bdi_write_congested(bdi)) {
670                         if (signalled())
671                                 ret = -ERESTARTSYS;
672                         else
673                                 schedule();
674                 }
675                 rpc_clnt_sigunmask(clnt, &oldset);
676         } else {
677                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
678                 if (bdi_write_congested(bdi))
679                         schedule();
680         }
681         finish_wait(&nfs_write_congestion, &wait);
682         return ret;
683 }
684
685
686 /*
687  * Try to update any existing write request, or create one if there is none.
688  * In order to match, the request's credentials must match those of
689  * the calling process.
690  *
691  * Note: Should always be called with the Page Lock held!
692  */
693 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
694                 struct inode *inode, struct page *page,
695                 unsigned int offset, unsigned int bytes)
696 {
697         struct nfs_server *server = NFS_SERVER(inode);
698         struct nfs_inode *nfsi = NFS_I(inode);
699         struct nfs_page         *req, *new = NULL;
700         unsigned long           rqend, end;
701
702         end = offset + bytes;
703
704         if (nfs_wait_on_write_congestion(page->mapping, server->flags & NFS_MOUNT_INTR))
705                 return ERR_PTR(-ERESTARTSYS);
706         for (;;) {
707                 /* Loop over all inode entries and see if we find
708                  * A request for the page we wish to update
709                  */
710                 spin_lock(&nfsi->req_lock);
711                 req = _nfs_find_request(inode, page->index);
712                 if (req) {
713                         if (!nfs_lock_request_dontget(req)) {
714                                 int error;
715                                 spin_unlock(&nfsi->req_lock);
716                                 error = nfs_wait_on_request(req);
717                                 nfs_release_request(req);
718                                 if (error < 0) {
719                                         if (new)
720                                                 nfs_release_request(new);
721                                         return ERR_PTR(error);
722                                 }
723                                 continue;
724                         }
725                         spin_unlock(&nfsi->req_lock);
726                         if (new)
727                                 nfs_release_request(new);
728                         break;
729                 }
730
731                 if (new) {
732                         int error;
733                         nfs_lock_request_dontget(new);
734                         error = nfs_inode_add_request(inode, new);
735                         if (error) {
736                                 spin_unlock(&nfsi->req_lock);
737                                 nfs_unlock_request(new);
738                                 return ERR_PTR(error);
739                         }
740                         spin_unlock(&nfsi->req_lock);
741                         nfs_mark_request_dirty(new);
742                         return new;
743                 }
744                 spin_unlock(&nfsi->req_lock);
745
746                 new = nfs_create_request(ctx, inode, page, offset, bytes);
747                 if (IS_ERR(new))
748                         return new;
749         }
750
751         /* We have a request for our page.
752          * If the creds don't match, or the
753          * page addresses don't match,
754          * tell the caller to wait on the conflicting
755          * request.
756          */
757         rqend = req->wb_offset + req->wb_bytes;
758         if (req->wb_context != ctx
759             || req->wb_page != page
760             || !nfs_dirty_request(req)
761             || offset > rqend || end < req->wb_offset) {
762                 nfs_unlock_request(req);
763                 return ERR_PTR(-EBUSY);
764         }
765
766         /* Okay, the request matches. Update the region */
767         if (offset < req->wb_offset) {
768                 req->wb_offset = offset;
769                 req->wb_pgbase = offset;
770                 req->wb_bytes = rqend - req->wb_offset;
771         }
772
773         if (end > rqend)
774                 req->wb_bytes = end - req->wb_offset;
775
776         return req;
777 }
778
779 int nfs_flush_incompatible(struct file *file, struct page *page)
780 {
781         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
782         struct inode    *inode = page->mapping->host;
783         struct nfs_page *req;
784         int             status = 0;
785         /*
786          * Look for a request corresponding to this page. If there
787          * is one, and it belongs to another file, we flush it out
788          * before we try to copy anything into the page. Do this
789          * due to the lack of an ACCESS-type call in NFSv2.
790          * Also do the same if we find a request from an existing
791          * dropped page.
792          */
793         req = nfs_find_request(inode, page->index);
794         if (req) {
795                 if (req->wb_page != page || ctx != req->wb_context)
796                         status = nfs_wb_page(inode, page);
797                 nfs_release_request(req);
798         }
799         return (status < 0) ? status : 0;
800 }
801
802 /*
803  * Update and possibly write a cached page of an NFS file.
804  *
805  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
806  * things with a page scheduled for an RPC call (e.g. invalidate it).
807  */
808 int nfs_updatepage(struct file *file, struct page *page,
809                 unsigned int offset, unsigned int count)
810 {
811         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
812         struct inode    *inode = page->mapping->host;
813         struct nfs_page *req;
814         int             status = 0;
815
816         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
817
818         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
819                 file->f_dentry->d_parent->d_name.name,
820                 file->f_dentry->d_name.name, count,
821                 (long long)(page_offset(page) +offset));
822
823         if (IS_SYNC(inode)) {
824                 status = nfs_writepage_sync(ctx, inode, page, offset, count, 0);
825                 if (status > 0) {
826                         if (offset == 0 && status == PAGE_CACHE_SIZE)
827                                 SetPageUptodate(page);
828                         return 0;
829                 }
830                 return status;
831         }
832
833         /* If we're not using byte range locks, and we know the page
834          * is entirely in cache, it may be more efficient to avoid
835          * fragmenting write requests.
836          */
837         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
838                 loff_t end_offs = i_size_read(inode) - 1;
839                 unsigned long end_index = end_offs >> PAGE_CACHE_SHIFT;
840
841                 count += offset;
842                 offset = 0;
843                 if (unlikely(end_offs < 0)) {
844                         /* Do nothing */
845                 } else if (page->index == end_index) {
846                         unsigned int pglen;
847                         pglen = (unsigned int)(end_offs & (PAGE_CACHE_SIZE-1)) + 1;
848                         if (count < pglen)
849                                 count = pglen;
850                 } else if (page->index < end_index)
851                         count = PAGE_CACHE_SIZE;
852         }
853
854         /*
855          * Try to find an NFS request corresponding to this page
856          * and update it.
857          * If the existing request cannot be updated, we must flush
858          * it out now.
859          */
860         do {
861                 req = nfs_update_request(ctx, inode, page, offset, count);
862                 status = (IS_ERR(req)) ? PTR_ERR(req) : 0;
863                 if (status != -EBUSY)
864                         break;
865                 /* Request could not be updated. Flush it out and try again */
866                 status = nfs_wb_page(inode, page);
867         } while (status >= 0);
868         if (status < 0)
869                 goto done;
870
871         status = 0;
872
873         /* Update file length */
874         nfs_grow_file(page, offset, count);
875         /* Set the PG_uptodate flag? */
876         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
877         nfs_unlock_request(req);
878 done:
879         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
880                         status, (long long)i_size_read(inode));
881         if (status < 0)
882                 ClearPageUptodate(page);
883         return status;
884 }
885
886 static void nfs_writepage_release(struct nfs_page *req)
887 {
888         end_page_writeback(req->wb_page);
889
890 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
891         if (!PageError(req->wb_page)) {
892                 if (NFS_NEED_RESCHED(req)) {
893                         nfs_mark_request_dirty(req);
894                         goto out;
895                 } else if (NFS_NEED_COMMIT(req)) {
896                         nfs_mark_request_commit(req);
897                         goto out;
898                 }
899         }
900         nfs_inode_remove_request(req);
901
902 out:
903         nfs_clear_commit(req);
904         nfs_clear_reschedule(req);
905 #else
906         nfs_inode_remove_request(req);
907 #endif
908         nfs_clear_page_writeback(req);
909 }
910
911 static inline int flush_task_priority(int how)
912 {
913         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
914                 case FLUSH_HIGHPRI:
915                         return RPC_PRIORITY_HIGH;
916                 case FLUSH_LOWPRI:
917                         return RPC_PRIORITY_LOW;
918         }
919         return RPC_PRIORITY_NORMAL;
920 }
921
922 /*
923  * Set up the argument/result storage required for the RPC call.
924  */
925 static void nfs_write_rpcsetup(struct nfs_page *req,
926                 struct nfs_write_data *data,
927                 const struct rpc_call_ops *call_ops,
928                 unsigned int count, unsigned int offset,
929                 int how)
930 {
931         struct inode            *inode;
932         int flags;
933
934         /* Set up the RPC argument and reply structs
935          * NB: take care not to mess about with data->commit et al. */
936
937         data->req = req;
938         data->inode = inode = req->wb_context->dentry->d_inode;
939         data->cred = req->wb_context->cred;
940
941         data->args.fh     = NFS_FH(inode);
942         data->args.offset = req_offset(req) + offset;
943         data->args.pgbase = req->wb_pgbase + offset;
944         data->args.pages  = data->pagevec;
945         data->args.count  = count;
946         data->args.context = req->wb_context;
947
948         data->res.fattr   = &data->fattr;
949         data->res.count   = count;
950         data->res.verf    = &data->verf;
951         nfs_fattr_init(&data->fattr);
952
953         /* Set up the initial task struct.  */
954         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
955         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
956         NFS_PROTO(inode)->write_setup(data, how);
957
958         data->task.tk_priority = flush_task_priority(how);
959         data->task.tk_cookie = (unsigned long)inode;
960
961         dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
962                 data->task.tk_pid,
963                 inode->i_sb->s_id,
964                 (long long)NFS_FILEID(inode),
965                 count,
966                 (unsigned long long)data->args.offset);
967 }
968
969 static void nfs_execute_write(struct nfs_write_data *data)
970 {
971         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
972         sigset_t oldset;
973
974         rpc_clnt_sigmask(clnt, &oldset);
975         lock_kernel();
976         rpc_execute(&data->task);
977         unlock_kernel();
978         rpc_clnt_sigunmask(clnt, &oldset);
979 }
980
981 /*
982  * Generate multiple small requests to write out a single
983  * contiguous dirty area on one page.
984  */
985 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
986 {
987         struct nfs_page *req = nfs_list_entry(head->next);
988         struct page *page = req->wb_page;
989         struct nfs_write_data *data;
990         unsigned int wsize = NFS_SERVER(inode)->wsize;
991         unsigned int nbytes, offset;
992         int requests = 0;
993         LIST_HEAD(list);
994
995         nfs_list_remove_request(req);
996
997         nbytes = req->wb_bytes;
998         for (;;) {
999                 data = nfs_writedata_alloc(1);
1000                 if (!data)
1001                         goto out_bad;
1002                 list_add(&data->pages, &list);
1003                 requests++;
1004                 if (nbytes <= wsize)
1005                         break;
1006                 nbytes -= wsize;
1007         }
1008         atomic_set(&req->wb_complete, requests);
1009
1010         ClearPageError(page);
1011         set_page_writeback(page);
1012         offset = 0;
1013         nbytes = req->wb_bytes;
1014         do {
1015                 data = list_entry(list.next, struct nfs_write_data, pages);
1016                 list_del_init(&data->pages);
1017
1018                 data->pagevec[0] = page;
1019
1020                 if (nbytes > wsize) {
1021                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
1022                                         wsize, offset, how);
1023                         offset += wsize;
1024                         nbytes -= wsize;
1025                 } else {
1026                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
1027                                         nbytes, offset, how);
1028                         nbytes = 0;
1029                 }
1030                 nfs_execute_write(data);
1031         } while (nbytes != 0);
1032
1033         return 0;
1034
1035 out_bad:
1036         while (!list_empty(&list)) {
1037                 data = list_entry(list.next, struct nfs_write_data, pages);
1038                 list_del(&data->pages);
1039                 nfs_writedata_free(data);
1040         }
1041         nfs_mark_request_dirty(req);
1042         nfs_clear_page_writeback(req);
1043         return -ENOMEM;
1044 }
1045
1046 /*
1047  * Create an RPC task for the given write request and kick it.
1048  * The page must have been locked by the caller.
1049  *
1050  * It may happen that the page we're passed is not marked dirty.
1051  * This is the case if nfs_updatepage detects a conflicting request
1052  * that has been written but not committed.
1053  */
1054 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
1055 {
1056         struct nfs_page         *req;
1057         struct page             **pages;
1058         struct nfs_write_data   *data;
1059         unsigned int            count;
1060
1061         data = nfs_writedata_alloc(NFS_SERVER(inode)->wpages);
1062         if (!data)
1063                 goto out_bad;
1064
1065         pages = data->pagevec;
1066         count = 0;
1067         while (!list_empty(head)) {
1068                 req = nfs_list_entry(head->next);
1069                 nfs_list_remove_request(req);
1070                 nfs_list_add_request(req, &data->pages);
1071                 ClearPageError(req->wb_page);
1072                 set_page_writeback(req->wb_page);
1073                 *pages++ = req->wb_page;
1074                 count += req->wb_bytes;
1075         }
1076         req = nfs_list_entry(data->pages.next);
1077
1078         /* Set up the argument struct */
1079         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
1080
1081         nfs_execute_write(data);
1082         return 0;
1083  out_bad:
1084         while (!list_empty(head)) {
1085                 struct nfs_page *req = nfs_list_entry(head->next);
1086                 nfs_list_remove_request(req);
1087                 nfs_mark_request_dirty(req);
1088                 nfs_clear_page_writeback(req);
1089         }
1090         return -ENOMEM;
1091 }
1092
1093 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
1094 {
1095         LIST_HEAD(one_request);
1096         int (*flush_one)(struct inode *, struct list_head *, int);
1097         struct nfs_page *req;
1098         int wpages = NFS_SERVER(inode)->wpages;
1099         int wsize = NFS_SERVER(inode)->wsize;
1100         int error;
1101
1102         flush_one = nfs_flush_one;
1103         if (wsize < PAGE_CACHE_SIZE)
1104                 flush_one = nfs_flush_multi;
1105         /* For single writes, FLUSH_STABLE is more efficient */
1106         if (npages <= wpages && npages == NFS_I(inode)->npages
1107                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
1108                 how |= FLUSH_STABLE;
1109
1110         do {
1111                 nfs_coalesce_requests(head, &one_request, wpages);
1112                 req = nfs_list_entry(one_request.next);
1113                 error = flush_one(inode, &one_request, how);
1114                 if (error < 0)
1115                         goto out_err;
1116         } while (!list_empty(head));
1117         return 0;
1118 out_err:
1119         while (!list_empty(head)) {
1120                 req = nfs_list_entry(head->next);
1121                 nfs_list_remove_request(req);
1122                 nfs_mark_request_dirty(req);
1123                 nfs_clear_page_writeback(req);
1124         }
1125         return error;
1126 }
1127
1128 /*
1129  * Handle a write reply that flushed part of a page.
1130  */
1131 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1132 {
1133         struct nfs_write_data   *data = calldata;
1134         struct nfs_page         *req = data->req;
1135         struct page             *page = req->wb_page;
1136
1137         dprintk("NFS: write (%s/%Ld %d@%Ld)",
1138                 req->wb_context->dentry->d_inode->i_sb->s_id,
1139                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1140                 req->wb_bytes,
1141                 (long long)req_offset(req));
1142
1143         if (nfs_writeback_done(task, data) != 0)
1144                 return;
1145
1146         if (task->tk_status < 0) {
1147                 ClearPageUptodate(page);
1148                 SetPageError(page);
1149                 req->wb_context->error = task->tk_status;
1150                 dprintk(", error = %d\n", task->tk_status);
1151         } else {
1152 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1153                 if (data->verf.committed < NFS_FILE_SYNC) {
1154                         if (!NFS_NEED_COMMIT(req)) {
1155                                 nfs_defer_commit(req);
1156                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1157                                 dprintk(" defer commit\n");
1158                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1159                                 nfs_defer_reschedule(req);
1160                                 dprintk(" server reboot detected\n");
1161                         }
1162                 } else
1163 #endif
1164                         dprintk(" OK\n");
1165         }
1166
1167         if (atomic_dec_and_test(&req->wb_complete))
1168                 nfs_writepage_release(req);
1169 }
1170
1171 static const struct rpc_call_ops nfs_write_partial_ops = {
1172         .rpc_call_done = nfs_writeback_done_partial,
1173         .rpc_release = nfs_writedata_release,
1174 };
1175
1176 /*
1177  * Handle a write reply that flushes a whole page.
1178  *
1179  * FIXME: There is an inherent race with invalidate_inode_pages and
1180  *        writebacks since the page->count is kept > 1 for as long
1181  *        as the page has a write request pending.
1182  */
1183 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1184 {
1185         struct nfs_write_data   *data = calldata;
1186         struct nfs_page         *req;
1187         struct page             *page;
1188
1189         if (nfs_writeback_done(task, data) != 0)
1190                 return;
1191
1192         /* Update attributes as result of writeback. */
1193         while (!list_empty(&data->pages)) {
1194                 req = nfs_list_entry(data->pages.next);
1195                 nfs_list_remove_request(req);
1196                 page = req->wb_page;
1197
1198                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1199                         req->wb_context->dentry->d_inode->i_sb->s_id,
1200                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1201                         req->wb_bytes,
1202                         (long long)req_offset(req));
1203
1204                 if (task->tk_status < 0) {
1205                         ClearPageUptodate(page);
1206                         SetPageError(page);
1207                         req->wb_context->error = task->tk_status;
1208                         end_page_writeback(page);
1209                         nfs_inode_remove_request(req);
1210                         dprintk(", error = %d\n", task->tk_status);
1211                         goto next;
1212                 }
1213                 end_page_writeback(page);
1214
1215 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1216                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1217                         nfs_inode_remove_request(req);
1218                         dprintk(" OK\n");
1219                         goto next;
1220                 }
1221                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1222                 nfs_mark_request_commit(req);
1223                 dprintk(" marked for commit\n");
1224 #else
1225                 nfs_inode_remove_request(req);
1226 #endif
1227         next:
1228                 nfs_clear_page_writeback(req);
1229         }
1230 }
1231
1232 static const struct rpc_call_ops nfs_write_full_ops = {
1233         .rpc_call_done = nfs_writeback_done_full,
1234         .rpc_release = nfs_writedata_release,
1235 };
1236
1237
1238 /*
1239  * This function is called when the WRITE call is complete.
1240  */
1241 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1242 {
1243         struct nfs_writeargs    *argp = &data->args;
1244         struct nfs_writeres     *resp = &data->res;
1245         int status;
1246
1247         dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1248                 task->tk_pid, task->tk_status);
1249
1250         /* Call the NFS version-specific code */
1251         status = NFS_PROTO(data->inode)->write_done(task, data);
1252         if (status != 0)
1253                 return status;
1254         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1255
1256 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1257         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1258                 /* We tried a write call, but the server did not
1259                  * commit data to stable storage even though we
1260                  * requested it.
1261                  * Note: There is a known bug in Tru64 < 5.0 in which
1262                  *       the server reports NFS_DATA_SYNC, but performs
1263                  *       NFS_FILE_SYNC. We therefore implement this checking
1264                  *       as a dprintk() in order to avoid filling syslog.
1265                  */
1266                 static unsigned long    complain;
1267
1268                 if (time_before(complain, jiffies)) {
1269                         dprintk("NFS: faulty NFS server %s:"
1270                                 " (committed = %d) != (stable = %d)\n",
1271                                 NFS_SERVER(data->inode)->hostname,
1272                                 resp->verf->committed, argp->stable);
1273                         complain = jiffies + 300 * HZ;
1274                 }
1275         }
1276 #endif
1277         /* Is this a short write? */
1278         if (task->tk_status >= 0 && resp->count < argp->count) {
1279                 static unsigned long    complain;
1280
1281                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1282
1283                 /* Has the server at least made some progress? */
1284                 if (resp->count != 0) {
1285                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1286                         if (resp->verf->committed != NFS_UNSTABLE) {
1287                                 /* Resend from where the server left off */
1288                                 argp->offset += resp->count;
1289                                 argp->pgbase += resp->count;
1290                                 argp->count -= resp->count;
1291                         } else {
1292                                 /* Resend as a stable write in order to avoid
1293                                  * headaches in the case of a server crash.
1294                                  */
1295                                 argp->stable = NFS_FILE_SYNC;
1296                         }
1297                         rpc_restart_call(task);
1298                         return -EAGAIN;
1299                 }
1300                 if (time_before(complain, jiffies)) {
1301                         printk(KERN_WARNING
1302                                "NFS: Server wrote zero bytes, expected %u.\n",
1303                                         argp->count);
1304                         complain = jiffies + 300 * HZ;
1305                 }
1306                 /* Can't do anything about it except throw an error. */
1307                 task->tk_status = -EIO;
1308         }
1309         return 0;
1310 }
1311
1312
1313 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1314 void nfs_commit_release(void *wdata)
1315 {
1316         nfs_commit_free(wdata);
1317 }
1318
1319 /*
1320  * Set up the argument/result storage required for the RPC call.
1321  */
1322 static void nfs_commit_rpcsetup(struct list_head *head,
1323                 struct nfs_write_data *data,
1324                 int how)
1325 {
1326         struct nfs_page         *first;
1327         struct inode            *inode;
1328         int flags;
1329
1330         /* Set up the RPC argument and reply structs
1331          * NB: take care not to mess about with data->commit et al. */
1332
1333         list_splice_init(head, &data->pages);
1334         first = nfs_list_entry(data->pages.next);
1335         inode = first->wb_context->dentry->d_inode;
1336
1337         data->inode       = inode;
1338         data->cred        = first->wb_context->cred;
1339
1340         data->args.fh     = NFS_FH(data->inode);
1341         /* Note: we always request a commit of the entire inode */
1342         data->args.offset = 0;
1343         data->args.count  = 0;
1344         data->res.count   = 0;
1345         data->res.fattr   = &data->fattr;
1346         data->res.verf    = &data->verf;
1347         nfs_fattr_init(&data->fattr);
1348
1349         /* Set up the initial task struct.  */
1350         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1351         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1352         NFS_PROTO(inode)->commit_setup(data, how);
1353
1354         data->task.tk_priority = flush_task_priority(how);
1355         data->task.tk_cookie = (unsigned long)inode;
1356         
1357         dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1358 }
1359
1360 /*
1361  * Commit dirty pages
1362  */
1363 static int
1364 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1365 {
1366         struct nfs_write_data   *data;
1367         struct nfs_page         *req;
1368
1369         data = nfs_commit_alloc(NFS_SERVER(inode)->wpages);
1370
1371         if (!data)
1372                 goto out_bad;
1373
1374         /* Set up the argument struct */
1375         nfs_commit_rpcsetup(head, data, how);
1376
1377         nfs_execute_write(data);
1378         return 0;
1379  out_bad:
1380         while (!list_empty(head)) {
1381                 req = nfs_list_entry(head->next);
1382                 nfs_list_remove_request(req);
1383                 nfs_mark_request_commit(req);
1384                 nfs_clear_page_writeback(req);
1385         }
1386         return -ENOMEM;
1387 }
1388
1389 /*
1390  * COMMIT call returned
1391  */
1392 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1393 {
1394         struct nfs_write_data   *data = calldata;
1395         struct nfs_page         *req;
1396
1397         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1398                                 task->tk_pid, task->tk_status);
1399
1400         /* Call the NFS version-specific code */
1401         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1402                 return;
1403
1404         while (!list_empty(&data->pages)) {
1405                 req = nfs_list_entry(data->pages.next);
1406                 nfs_list_remove_request(req);
1407                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1408
1409                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1410                         req->wb_context->dentry->d_inode->i_sb->s_id,
1411                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1412                         req->wb_bytes,
1413                         (long long)req_offset(req));
1414                 if (task->tk_status < 0) {
1415                         req->wb_context->error = task->tk_status;
1416                         nfs_inode_remove_request(req);
1417                         dprintk(", error = %d\n", task->tk_status);
1418                         goto next;
1419                 }
1420
1421                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1422                  * returned by the server against all stored verfs. */
1423                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1424                         /* We have a match */
1425                         nfs_inode_remove_request(req);
1426                         dprintk(" OK\n");
1427                         goto next;
1428                 }
1429                 /* We have a mismatch. Write the page again */
1430                 dprintk(" mismatch\n");
1431                 nfs_mark_request_dirty(req);
1432         next:
1433                 nfs_clear_page_writeback(req);
1434         }
1435 }
1436
1437 static const struct rpc_call_ops nfs_commit_ops = {
1438         .rpc_call_done = nfs_commit_done,
1439         .rpc_release = nfs_commit_release,
1440 };
1441 #else
1442 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1443 {
1444         return 0;
1445 }
1446 #endif
1447
1448 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
1449                            unsigned int npages, int how)
1450 {
1451         struct nfs_inode *nfsi = NFS_I(inode);
1452         LIST_HEAD(head);
1453         int res;
1454
1455         spin_lock(&nfsi->req_lock);
1456         res = nfs_scan_dirty(inode, &head, idx_start, npages);
1457         spin_unlock(&nfsi->req_lock);
1458         if (res) {
1459                 int error = nfs_flush_list(inode, &head, res, how);
1460                 if (error < 0)
1461                         return error;
1462         }
1463         return res;
1464 }
1465
1466 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1467 int nfs_commit_inode(struct inode *inode, int how)
1468 {
1469         struct nfs_inode *nfsi = NFS_I(inode);
1470         LIST_HEAD(head);
1471         int res;
1472
1473         spin_lock(&nfsi->req_lock);
1474         res = nfs_scan_commit(inode, &head, 0, 0);
1475         spin_unlock(&nfsi->req_lock);
1476         if (res) {
1477                 int error = nfs_commit_list(inode, &head, how);
1478                 if (error < 0)
1479                         return error;
1480         }
1481         return res;
1482 }
1483 #endif
1484
1485 int nfs_sync_inode_wait(struct inode *inode, unsigned long idx_start,
1486                 unsigned int npages, int how)
1487 {
1488         struct nfs_inode *nfsi = NFS_I(inode);
1489         LIST_HEAD(head);
1490         int nocommit = how & FLUSH_NOCOMMIT;
1491         int pages, ret;
1492
1493         how &= ~FLUSH_NOCOMMIT;
1494         spin_lock(&nfsi->req_lock);
1495         do {
1496                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1497                 if (ret != 0)
1498                         continue;
1499                 pages = nfs_scan_dirty(inode, &head, idx_start, npages);
1500                 if (pages != 0) {
1501                         spin_unlock(&nfsi->req_lock);
1502                         if (how & FLUSH_INVALIDATE)
1503                                 nfs_cancel_requests(&head);
1504                         else
1505                                 ret = nfs_flush_list(inode, &head, pages, how);
1506                         spin_lock(&nfsi->req_lock);
1507                         continue;
1508                 }
1509                 if (nocommit)
1510                         break;
1511                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1512                 if (pages == 0)
1513                         break;
1514                 if (how & FLUSH_INVALIDATE) {
1515                         spin_unlock(&nfsi->req_lock);
1516                         nfs_cancel_requests(&head);
1517                         spin_lock(&nfsi->req_lock);
1518                         continue;
1519                 }
1520                 pages += nfs_scan_commit(inode, &head, 0, 0);
1521                 spin_unlock(&nfsi->req_lock);
1522                 ret = nfs_commit_list(inode, &head, how);
1523                 spin_lock(&nfsi->req_lock);
1524         } while (ret >= 0);
1525         spin_unlock(&nfsi->req_lock);
1526         return ret;
1527 }
1528
1529 int __init nfs_init_writepagecache(void)
1530 {
1531         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1532                                              sizeof(struct nfs_write_data),
1533                                              0, SLAB_HWCACHE_ALIGN,
1534                                              NULL, NULL);
1535         if (nfs_wdata_cachep == NULL)
1536                 return -ENOMEM;
1537
1538         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1539                                                      nfs_wdata_cachep);
1540         if (nfs_wdata_mempool == NULL)
1541                 return -ENOMEM;
1542
1543         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1544                                                       nfs_wdata_cachep);
1545         if (nfs_commit_mempool == NULL)
1546                 return -ENOMEM;
1547
1548         return 0;
1549 }
1550
1551 void nfs_destroy_writepagecache(void)
1552 {
1553         mempool_destroy(nfs_commit_mempool);
1554         mempool_destroy(nfs_wdata_mempool);
1555         if (kmem_cache_destroy(nfs_wdata_cachep))
1556                 printk(KERN_INFO "nfs_write_data: not all structures were freed\n");
1557 }
1558