4 * Writing file data over NFS.
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.
13 * Just like readahead, no async I/O is performed if wsize < PAGE_SIZE.
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.
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.
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
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
33 * Also note that because of the way pages are invalidated in
34 * nfs_revalidate_inode, the following assertions hold:
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.
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.
46 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
49 #include <linux/config.h>
50 #include <linux/types.h>
51 #include <linux/slab.h>
53 #include <linux/pagemap.h>
54 #include <linux/file.h>
55 #include <linux/mpage.h>
56 #include <linux/writeback.h>
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>
65 #include "delegation.h"
67 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
69 #define MIN_POOL_WRITE (32)
70 #define MIN_POOL_COMMIT (4)
73 * Local function declarations
75 static struct nfs_page
* nfs_update_request(struct nfs_open_context
*,
78 unsigned int, unsigned int);
79 static void nfs_writeback_done_partial(struct nfs_write_data
*, int);
80 static void nfs_writeback_done_full(struct nfs_write_data
*, int);
81 static int nfs_wait_on_write_congestion(struct address_space
*, int);
82 static int nfs_wait_on_requests(struct inode
*, unsigned long, unsigned int);
83 static int nfs_flush_inode(struct inode
*inode
, unsigned long idx_start
,
84 unsigned int npages
, int how
);
86 static kmem_cache_t
*nfs_wdata_cachep
;
87 mempool_t
*nfs_wdata_mempool
;
88 static mempool_t
*nfs_commit_mempool
;
90 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion
);
92 static inline struct nfs_write_data
*nfs_commit_alloc(void)
94 struct nfs_write_data
*p
= mempool_alloc(nfs_commit_mempool
, SLAB_NOFS
);
96 memset(p
, 0, sizeof(*p
));
97 INIT_LIST_HEAD(&p
->pages
);
102 static inline void nfs_commit_free(struct nfs_write_data
*p
)
104 mempool_free(p
, nfs_commit_mempool
);
107 static void nfs_writedata_release(struct rpc_task
*task
)
109 struct nfs_write_data
*wdata
= (struct nfs_write_data
*)task
->tk_calldata
;
110 nfs_writedata_free(wdata
);
113 /* Adjust the file length if we're writing beyond the end */
114 static void nfs_grow_file(struct page
*page
, unsigned int offset
, unsigned int count
)
116 struct inode
*inode
= page
->mapping
->host
;
117 loff_t end
, i_size
= i_size_read(inode
);
118 unsigned long end_index
= (i_size
- 1) >> PAGE_CACHE_SHIFT
;
120 if (i_size
> 0 && page
->index
< end_index
)
122 end
= ((loff_t
)page
->index
<< PAGE_CACHE_SHIFT
) + ((loff_t
)offset
+count
);
125 i_size_write(inode
, end
);
128 /* We can set the PG_uptodate flag if we see that a write request
129 * covers the full page.
131 static void nfs_mark_uptodate(struct page
*page
, unsigned int base
, unsigned int count
)
135 if (PageUptodate(page
))
139 if (count
== PAGE_CACHE_SIZE
) {
140 SetPageUptodate(page
);
144 end_offs
= i_size_read(page
->mapping
->host
) - 1;
147 /* Is this the last page? */
148 if (page
->index
!= (unsigned long)(end_offs
>> PAGE_CACHE_SHIFT
))
150 /* This is the last page: set PG_uptodate if we cover the entire
151 * extent of the data, then zero the rest of the page.
153 if (count
== (unsigned int)(end_offs
& (PAGE_CACHE_SIZE
- 1)) + 1) {
154 memclear_highpage_flush(page
, count
, PAGE_CACHE_SIZE
- count
);
155 SetPageUptodate(page
);
160 * Write a page synchronously.
161 * Offset is the data offset within the page.
163 static int nfs_writepage_sync(struct nfs_open_context
*ctx
, struct inode
*inode
,
164 struct page
*page
, unsigned int offset
, unsigned int count
,
167 unsigned int wsize
= NFS_SERVER(inode
)->wsize
;
168 int result
, written
= 0;
169 struct nfs_write_data
*wdata
;
171 wdata
= nfs_writedata_alloc();
176 wdata
->cred
= ctx
->cred
;
177 wdata
->inode
= inode
;
178 wdata
->args
.fh
= NFS_FH(inode
);
179 wdata
->args
.context
= ctx
;
180 wdata
->args
.pages
= &page
;
181 wdata
->args
.stable
= NFS_FILE_SYNC
;
182 wdata
->args
.pgbase
= offset
;
183 wdata
->args
.count
= wsize
;
184 wdata
->res
.fattr
= &wdata
->fattr
;
185 wdata
->res
.verf
= &wdata
->verf
;
187 dprintk("NFS: nfs_writepage_sync(%s/%Ld %d@%Ld)\n",
189 (long long)NFS_FILEID(inode
),
190 count
, (long long)(page_offset(page
) + offset
));
192 set_page_writeback(page
);
193 nfs_begin_data_update(inode
);
196 wdata
->args
.count
= count
;
197 wdata
->args
.offset
= page_offset(page
) + wdata
->args
.pgbase
;
199 result
= NFS_PROTO(inode
)->write(wdata
);
202 /* Must mark the page invalid after I/O error */
203 ClearPageUptodate(page
);
206 if (result
< wdata
->args
.count
)
207 printk(KERN_WARNING
"NFS: short write, count=%u, result=%d\n",
208 wdata
->args
.count
, result
);
210 wdata
->args
.offset
+= result
;
211 wdata
->args
.pgbase
+= result
;
215 /* Update file length */
216 nfs_grow_file(page
, offset
, written
);
217 /* Set the PG_uptodate flag? */
218 nfs_mark_uptodate(page
, offset
, written
);
221 ClearPageError(page
);
224 nfs_end_data_update(inode
);
225 end_page_writeback(page
);
226 nfs_writedata_free(wdata
);
227 return written
? written
: result
;
230 static int nfs_writepage_async(struct nfs_open_context
*ctx
,
231 struct inode
*inode
, struct page
*page
,
232 unsigned int offset
, unsigned int count
)
234 struct nfs_page
*req
;
236 req
= nfs_update_request(ctx
, inode
, page
, offset
, count
);
239 /* Update file length */
240 nfs_grow_file(page
, offset
, count
);
241 /* Set the PG_uptodate flag? */
242 nfs_mark_uptodate(page
, offset
, count
);
243 nfs_unlock_request(req
);
247 static int wb_priority(struct writeback_control
*wbc
)
249 if (wbc
->for_reclaim
)
250 return FLUSH_HIGHPRI
;
251 if (wbc
->for_kupdate
)
257 * Write an mmapped page to the server.
259 int nfs_writepage(struct page
*page
, struct writeback_control
*wbc
)
261 struct nfs_open_context
*ctx
;
262 struct inode
*inode
= page
->mapping
->host
;
263 unsigned long end_index
;
264 unsigned offset
= PAGE_CACHE_SIZE
;
265 loff_t i_size
= i_size_read(inode
);
266 int inode_referenced
= 0;
267 int priority
= wb_priority(wbc
);
271 * Note: We need to ensure that we have a reference to the inode
272 * if we are to do asynchronous writes. If not, waiting
273 * in nfs_wait_on_request() may deadlock with clear_inode().
275 * If igrab() fails here, then it is in any case safe to
276 * call nfs_wb_page(), since there will be no pending writes.
278 if (igrab(inode
) != 0)
279 inode_referenced
= 1;
280 end_index
= i_size
>> PAGE_CACHE_SHIFT
;
282 /* Ensure we've flushed out any previous writes */
283 nfs_wb_page_priority(inode
, page
, priority
);
286 if (page
->index
< end_index
)
288 /* things got complicated... */
289 offset
= i_size
& (PAGE_CACHE_SIZE
-1);
291 /* OK, are we completely out? */
292 err
= 0; /* potential race with truncate - ignore */
293 if (page
->index
>= end_index
+1 || !offset
)
296 ctx
= nfs_find_open_context(inode
, NULL
, FMODE_WRITE
);
302 if (!IS_SYNC(inode
) && inode_referenced
) {
303 err
= nfs_writepage_async(ctx
, inode
, page
, 0, offset
);
304 if (!wbc
->for_writepages
)
305 nfs_flush_inode(inode
, 0, 0, wb_priority(wbc
));
307 err
= nfs_writepage_sync(ctx
, inode
, page
, 0,
311 redirty_page_for_writepage(wbc
, page
);
316 put_nfs_open_context(ctx
);
319 if (inode_referenced
)
325 * Note: causes nfs_update_request() to block on the assumption
326 * that the writeback is generated due to memory pressure.
328 int nfs_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
330 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
331 struct inode
*inode
= mapping
->host
;
334 err
= generic_writepages(mapping
, wbc
);
337 while (test_and_set_bit(BDI_write_congested
, &bdi
->state
) != 0) {
338 if (wbc
->nonblocking
)
340 nfs_wait_on_write_congestion(mapping
, 0);
342 err
= nfs_flush_inode(inode
, 0, 0, wb_priority(wbc
));
345 wbc
->nr_to_write
-= err
;
346 if (!wbc
->nonblocking
&& wbc
->sync_mode
== WB_SYNC_ALL
) {
347 err
= nfs_wait_on_requests(inode
, 0, 0);
351 err
= nfs_commit_inode(inode
, wb_priority(wbc
));
353 wbc
->nr_to_write
-= err
;
357 clear_bit(BDI_write_congested
, &bdi
->state
);
358 wake_up_all(&nfs_write_congestion
);
363 * Insert a write request into an inode
365 static int nfs_inode_add_request(struct inode
*inode
, struct nfs_page
*req
)
367 struct nfs_inode
*nfsi
= NFS_I(inode
);
370 error
= radix_tree_insert(&nfsi
->nfs_page_tree
, req
->wb_index
, req
);
371 BUG_ON(error
== -EEXIST
);
376 nfs_begin_data_update(inode
);
377 if (nfs_have_delegation(inode
, FMODE_WRITE
))
381 atomic_inc(&req
->wb_count
);
386 * Insert a write request into an inode
388 static void nfs_inode_remove_request(struct nfs_page
*req
)
390 struct inode
*inode
= req
->wb_context
->dentry
->d_inode
;
391 struct nfs_inode
*nfsi
= NFS_I(inode
);
393 BUG_ON (!NFS_WBACK_BUSY(req
));
395 spin_lock(&nfsi
->req_lock
);
396 radix_tree_delete(&nfsi
->nfs_page_tree
, req
->wb_index
);
399 spin_unlock(&nfsi
->req_lock
);
400 nfs_end_data_update(inode
);
403 spin_unlock(&nfsi
->req_lock
);
404 nfs_clear_request(req
);
405 nfs_release_request(req
);
411 static inline struct nfs_page
*
412 _nfs_find_request(struct inode
*inode
, unsigned long index
)
414 struct nfs_inode
*nfsi
= NFS_I(inode
);
415 struct nfs_page
*req
;
417 req
= (struct nfs_page
*)radix_tree_lookup(&nfsi
->nfs_page_tree
, index
);
419 atomic_inc(&req
->wb_count
);
423 static struct nfs_page
*
424 nfs_find_request(struct inode
*inode
, unsigned long index
)
426 struct nfs_page
*req
;
427 struct nfs_inode
*nfsi
= NFS_I(inode
);
429 spin_lock(&nfsi
->req_lock
);
430 req
= _nfs_find_request(inode
, index
);
431 spin_unlock(&nfsi
->req_lock
);
436 * Add a request to the inode's dirty list.
439 nfs_mark_request_dirty(struct nfs_page
*req
)
441 struct inode
*inode
= req
->wb_context
->dentry
->d_inode
;
442 struct nfs_inode
*nfsi
= NFS_I(inode
);
444 spin_lock(&nfsi
->req_lock
);
445 radix_tree_tag_set(&nfsi
->nfs_page_tree
,
446 req
->wb_index
, NFS_PAGE_TAG_DIRTY
);
447 nfs_list_add_request(req
, &nfsi
->dirty
);
449 spin_unlock(&nfsi
->req_lock
);
450 inc_page_state(nr_dirty
);
451 mark_inode_dirty(inode
);
455 * Check if a request is dirty
458 nfs_dirty_request(struct nfs_page
*req
)
460 struct nfs_inode
*nfsi
= NFS_I(req
->wb_context
->dentry
->d_inode
);
461 return !list_empty(&req
->wb_list
) && req
->wb_list_head
== &nfsi
->dirty
;
464 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
466 * Add a request to the inode's commit list.
469 nfs_mark_request_commit(struct nfs_page
*req
)
471 struct inode
*inode
= req
->wb_context
->dentry
->d_inode
;
472 struct nfs_inode
*nfsi
= NFS_I(inode
);
474 spin_lock(&nfsi
->req_lock
);
475 nfs_list_add_request(req
, &nfsi
->commit
);
477 spin_unlock(&nfsi
->req_lock
);
478 inc_page_state(nr_unstable
);
479 mark_inode_dirty(inode
);
484 * Wait for a request to complete.
486 * Interruptible by signals only if mounted with intr flag.
489 nfs_wait_on_requests(struct inode
*inode
, unsigned long idx_start
, unsigned int npages
)
491 struct nfs_inode
*nfsi
= NFS_I(inode
);
492 struct nfs_page
*req
;
493 unsigned long idx_end
, next
;
494 unsigned int res
= 0;
500 idx_end
= idx_start
+ npages
- 1;
502 spin_lock(&nfsi
->req_lock
);
504 while (radix_tree_gang_lookup_tag(&nfsi
->nfs_page_tree
, (void **)&req
, next
, 1, NFS_PAGE_TAG_WRITEBACK
)) {
505 if (req
->wb_index
> idx_end
)
508 next
= req
->wb_index
+ 1;
509 BUG_ON(!NFS_WBACK_BUSY(req
));
511 atomic_inc(&req
->wb_count
);
512 spin_unlock(&nfsi
->req_lock
);
513 error
= nfs_wait_on_request(req
);
514 nfs_release_request(req
);
517 spin_lock(&nfsi
->req_lock
);
520 spin_unlock(&nfsi
->req_lock
);
525 * nfs_scan_dirty - Scan an inode for dirty requests
526 * @inode: NFS inode to scan
527 * @dst: destination list
528 * @idx_start: lower bound of page->index to scan.
529 * @npages: idx_start + npages sets the upper bound to scan.
531 * Moves requests from the inode's dirty page list.
532 * The requests are *not* checked to ensure that they form a contiguous set.
535 nfs_scan_dirty(struct inode
*inode
, struct list_head
*dst
, unsigned long idx_start
, unsigned int npages
)
537 struct nfs_inode
*nfsi
= NFS_I(inode
);
540 if (nfsi
->ndirty
!= 0) {
541 res
= nfs_scan_lock_dirty(nfsi
, dst
, idx_start
, npages
);
543 sub_page_state(nr_dirty
,res
);
544 if ((nfsi
->ndirty
== 0) != list_empty(&nfsi
->dirty
))
545 printk(KERN_ERR
"NFS: desynchronized value of nfs_i.ndirty.\n");
550 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
552 * nfs_scan_commit - Scan an inode for commit requests
553 * @inode: NFS inode to scan
554 * @dst: destination list
555 * @idx_start: lower bound of page->index to scan.
556 * @npages: idx_start + npages sets the upper bound to scan.
558 * Moves requests from the inode's 'commit' request list.
559 * The requests are *not* checked to ensure that they form a contiguous set.
562 nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, unsigned long idx_start
, unsigned int npages
)
564 struct nfs_inode
*nfsi
= NFS_I(inode
);
567 if (nfsi
->ncommit
!= 0) {
568 res
= nfs_scan_list(&nfsi
->commit
, dst
, idx_start
, npages
);
569 nfsi
->ncommit
-= res
;
570 if ((nfsi
->ncommit
== 0) != list_empty(&nfsi
->commit
))
571 printk(KERN_ERR
"NFS: desynchronized value of nfs_i.ncommit.\n");
577 static int nfs_wait_on_write_congestion(struct address_space
*mapping
, int intr
)
579 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
585 if (!bdi_write_congested(bdi
))
588 struct rpc_clnt
*clnt
= NFS_CLIENT(mapping
->host
);
591 rpc_clnt_sigmask(clnt
, &oldset
);
592 prepare_to_wait(&nfs_write_congestion
, &wait
, TASK_INTERRUPTIBLE
);
593 if (bdi_write_congested(bdi
)) {
599 rpc_clnt_sigunmask(clnt
, &oldset
);
601 prepare_to_wait(&nfs_write_congestion
, &wait
, TASK_UNINTERRUPTIBLE
);
602 if (bdi_write_congested(bdi
))
605 finish_wait(&nfs_write_congestion
, &wait
);
611 * Try to update any existing write request, or create one if there is none.
612 * In order to match, the request's credentials must match those of
613 * the calling process.
615 * Note: Should always be called with the Page Lock held!
617 static struct nfs_page
* nfs_update_request(struct nfs_open_context
* ctx
,
618 struct inode
*inode
, struct page
*page
,
619 unsigned int offset
, unsigned int bytes
)
621 struct nfs_server
*server
= NFS_SERVER(inode
);
622 struct nfs_inode
*nfsi
= NFS_I(inode
);
623 struct nfs_page
*req
, *new = NULL
;
624 unsigned long rqend
, end
;
626 end
= offset
+ bytes
;
628 if (nfs_wait_on_write_congestion(page
->mapping
, server
->flags
& NFS_MOUNT_INTR
))
629 return ERR_PTR(-ERESTARTSYS
);
631 /* Loop over all inode entries and see if we find
632 * A request for the page we wish to update
634 spin_lock(&nfsi
->req_lock
);
635 req
= _nfs_find_request(inode
, page
->index
);
637 if (!nfs_lock_request_dontget(req
)) {
639 spin_unlock(&nfsi
->req_lock
);
640 error
= nfs_wait_on_request(req
);
641 nfs_release_request(req
);
643 return ERR_PTR(error
);
646 spin_unlock(&nfsi
->req_lock
);
648 nfs_release_request(new);
654 nfs_lock_request_dontget(new);
655 error
= nfs_inode_add_request(inode
, new);
657 spin_unlock(&nfsi
->req_lock
);
658 nfs_unlock_request(new);
659 return ERR_PTR(error
);
661 spin_unlock(&nfsi
->req_lock
);
662 nfs_mark_request_dirty(new);
665 spin_unlock(&nfsi
->req_lock
);
667 new = nfs_create_request(ctx
, inode
, page
, offset
, bytes
);
672 /* We have a request for our page.
673 * If the creds don't match, or the
674 * page addresses don't match,
675 * tell the caller to wait on the conflicting
678 rqend
= req
->wb_offset
+ req
->wb_bytes
;
679 if (req
->wb_context
!= ctx
680 || req
->wb_page
!= page
681 || !nfs_dirty_request(req
)
682 || offset
> rqend
|| end
< req
->wb_offset
) {
683 nfs_unlock_request(req
);
684 return ERR_PTR(-EBUSY
);
687 /* Okay, the request matches. Update the region */
688 if (offset
< req
->wb_offset
) {
689 req
->wb_offset
= offset
;
690 req
->wb_pgbase
= offset
;
691 req
->wb_bytes
= rqend
- req
->wb_offset
;
695 req
->wb_bytes
= end
- req
->wb_offset
;
700 int nfs_flush_incompatible(struct file
*file
, struct page
*page
)
702 struct nfs_open_context
*ctx
= (struct nfs_open_context
*)file
->private_data
;
703 struct inode
*inode
= page
->mapping
->host
;
704 struct nfs_page
*req
;
707 * Look for a request corresponding to this page. If there
708 * is one, and it belongs to another file, we flush it out
709 * before we try to copy anything into the page. Do this
710 * due to the lack of an ACCESS-type call in NFSv2.
711 * Also do the same if we find a request from an existing
714 req
= nfs_find_request(inode
, page
->index
);
716 if (req
->wb_page
!= page
|| ctx
!= req
->wb_context
)
717 status
= nfs_wb_page(inode
, page
);
718 nfs_release_request(req
);
720 return (status
< 0) ? status
: 0;
724 * Update and possibly write a cached page of an NFS file.
726 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
727 * things with a page scheduled for an RPC call (e.g. invalidate it).
729 int nfs_updatepage(struct file
*file
, struct page
*page
,
730 unsigned int offset
, unsigned int count
)
732 struct nfs_open_context
*ctx
= (struct nfs_open_context
*)file
->private_data
;
733 struct inode
*inode
= page
->mapping
->host
;
734 struct nfs_page
*req
;
737 dprintk("NFS: nfs_updatepage(%s/%s %d@%Ld)\n",
738 file
->f_dentry
->d_parent
->d_name
.name
,
739 file
->f_dentry
->d_name
.name
, count
,
740 (long long)(page_offset(page
) +offset
));
742 if (IS_SYNC(inode
)) {
743 status
= nfs_writepage_sync(ctx
, inode
, page
, offset
, count
, 0);
745 if (offset
== 0 && status
== PAGE_CACHE_SIZE
)
746 SetPageUptodate(page
);
752 /* If we're not using byte range locks, and we know the page
753 * is entirely in cache, it may be more efficient to avoid
754 * fragmenting write requests.
756 if (PageUptodate(page
) && inode
->i_flock
== NULL
&& !(file
->f_mode
& O_SYNC
)) {
757 loff_t end_offs
= i_size_read(inode
) - 1;
758 unsigned long end_index
= end_offs
>> PAGE_CACHE_SHIFT
;
762 if (unlikely(end_offs
< 0)) {
764 } else if (page
->index
== end_index
) {
766 pglen
= (unsigned int)(end_offs
& (PAGE_CACHE_SIZE
-1)) + 1;
769 } else if (page
->index
< end_index
)
770 count
= PAGE_CACHE_SIZE
;
774 * Try to find an NFS request corresponding to this page
776 * If the existing request cannot be updated, we must flush
780 req
= nfs_update_request(ctx
, inode
, page
, offset
, count
);
781 status
= (IS_ERR(req
)) ? PTR_ERR(req
) : 0;
782 if (status
!= -EBUSY
)
784 /* Request could not be updated. Flush it out and try again */
785 status
= nfs_wb_page(inode
, page
);
786 } while (status
>= 0);
792 /* Update file length */
793 nfs_grow_file(page
, offset
, count
);
794 /* Set the PG_uptodate flag? */
795 nfs_mark_uptodate(page
, req
->wb_pgbase
, req
->wb_bytes
);
796 nfs_unlock_request(req
);
798 dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n",
799 status
, (long long)i_size_read(inode
));
801 ClearPageUptodate(page
);
805 static void nfs_writepage_release(struct nfs_page
*req
)
807 end_page_writeback(req
->wb_page
);
809 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
810 if (!PageError(req
->wb_page
)) {
811 if (NFS_NEED_RESCHED(req
)) {
812 nfs_mark_request_dirty(req
);
814 } else if (NFS_NEED_COMMIT(req
)) {
815 nfs_mark_request_commit(req
);
819 nfs_inode_remove_request(req
);
822 nfs_clear_commit(req
);
823 nfs_clear_reschedule(req
);
825 nfs_inode_remove_request(req
);
827 nfs_clear_page_writeback(req
);
830 static inline int flush_task_priority(int how
)
832 switch (how
& (FLUSH_HIGHPRI
|FLUSH_LOWPRI
)) {
834 return RPC_PRIORITY_HIGH
;
836 return RPC_PRIORITY_LOW
;
838 return RPC_PRIORITY_NORMAL
;
842 * Set up the argument/result storage required for the RPC call.
844 static void nfs_write_rpcsetup(struct nfs_page
*req
,
845 struct nfs_write_data
*data
,
846 unsigned int count
, unsigned int offset
,
851 /* Set up the RPC argument and reply structs
852 * NB: take care not to mess about with data->commit et al. */
855 data
->inode
= inode
= req
->wb_context
->dentry
->d_inode
;
856 data
->cred
= req
->wb_context
->cred
;
858 data
->args
.fh
= NFS_FH(inode
);
859 data
->args
.offset
= req_offset(req
) + offset
;
860 data
->args
.pgbase
= req
->wb_pgbase
+ offset
;
861 data
->args
.pages
= data
->pagevec
;
862 data
->args
.count
= count
;
863 data
->args
.context
= req
->wb_context
;
865 data
->res
.fattr
= &data
->fattr
;
866 data
->res
.count
= count
;
867 data
->res
.verf
= &data
->verf
;
868 nfs_fattr_init(&data
->fattr
);
870 NFS_PROTO(inode
)->write_setup(data
, how
);
872 data
->task
.tk_priority
= flush_task_priority(how
);
873 data
->task
.tk_cookie
= (unsigned long)inode
;
874 data
->task
.tk_calldata
= data
;
875 /* Release requests */
876 data
->task
.tk_release
= nfs_writedata_release
;
878 dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
881 (long long)NFS_FILEID(inode
),
883 (unsigned long long)data
->args
.offset
);
886 static void nfs_execute_write(struct nfs_write_data
*data
)
888 struct rpc_clnt
*clnt
= NFS_CLIENT(data
->inode
);
891 rpc_clnt_sigmask(clnt
, &oldset
);
893 rpc_execute(&data
->task
);
895 rpc_clnt_sigunmask(clnt
, &oldset
);
899 * Generate multiple small requests to write out a single
900 * contiguous dirty area on one page.
902 static int nfs_flush_multi(struct list_head
*head
, struct inode
*inode
, int how
)
904 struct nfs_page
*req
= nfs_list_entry(head
->next
);
905 struct page
*page
= req
->wb_page
;
906 struct nfs_write_data
*data
;
907 unsigned int wsize
= NFS_SERVER(inode
)->wsize
;
908 unsigned int nbytes
, offset
;
912 nfs_list_remove_request(req
);
914 nbytes
= req
->wb_bytes
;
916 data
= nfs_writedata_alloc();
919 list_add(&data
->pages
, &list
);
925 atomic_set(&req
->wb_complete
, requests
);
927 ClearPageError(page
);
928 set_page_writeback(page
);
930 nbytes
= req
->wb_bytes
;
932 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
933 list_del_init(&data
->pages
);
935 data
->pagevec
[0] = page
;
936 data
->complete
= nfs_writeback_done_partial
;
938 if (nbytes
> wsize
) {
939 nfs_write_rpcsetup(req
, data
, wsize
, offset
, how
);
943 nfs_write_rpcsetup(req
, data
, nbytes
, offset
, how
);
946 nfs_execute_write(data
);
947 } while (nbytes
!= 0);
952 while (!list_empty(&list
)) {
953 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
954 list_del(&data
->pages
);
955 nfs_writedata_free(data
);
957 nfs_mark_request_dirty(req
);
958 nfs_clear_page_writeback(req
);
963 * Create an RPC task for the given write request and kick it.
964 * The page must have been locked by the caller.
966 * It may happen that the page we're passed is not marked dirty.
967 * This is the case if nfs_updatepage detects a conflicting request
968 * that has been written but not committed.
970 static int nfs_flush_one(struct list_head
*head
, struct inode
*inode
, int how
)
972 struct nfs_page
*req
;
974 struct nfs_write_data
*data
;
977 if (NFS_SERVER(inode
)->wsize
< PAGE_CACHE_SIZE
)
978 return nfs_flush_multi(head
, inode
, how
);
980 data
= nfs_writedata_alloc();
984 pages
= data
->pagevec
;
986 while (!list_empty(head
)) {
987 req
= nfs_list_entry(head
->next
);
988 nfs_list_remove_request(req
);
989 nfs_list_add_request(req
, &data
->pages
);
990 ClearPageError(req
->wb_page
);
991 set_page_writeback(req
->wb_page
);
992 *pages
++ = req
->wb_page
;
993 count
+= req
->wb_bytes
;
995 req
= nfs_list_entry(data
->pages
.next
);
997 data
->complete
= nfs_writeback_done_full
;
998 /* Set up the argument struct */
999 nfs_write_rpcsetup(req
, data
, count
, 0, how
);
1001 nfs_execute_write(data
);
1004 while (!list_empty(head
)) {
1005 struct nfs_page
*req
= nfs_list_entry(head
->next
);
1006 nfs_list_remove_request(req
);
1007 nfs_mark_request_dirty(req
);
1008 nfs_clear_page_writeback(req
);
1014 nfs_flush_list(struct list_head
*head
, int wpages
, int how
)
1016 LIST_HEAD(one_request
);
1017 struct nfs_page
*req
;
1019 unsigned int pages
= 0;
1021 while (!list_empty(head
)) {
1022 pages
+= nfs_coalesce_requests(head
, &one_request
, wpages
);
1023 req
= nfs_list_entry(one_request
.next
);
1024 error
= nfs_flush_one(&one_request
, req
->wb_context
->dentry
->d_inode
, how
);
1031 while (!list_empty(head
)) {
1032 req
= nfs_list_entry(head
->next
);
1033 nfs_list_remove_request(req
);
1034 nfs_mark_request_dirty(req
);
1035 nfs_clear_page_writeback(req
);
1041 * Handle a write reply that flushed part of a page.
1043 static void nfs_writeback_done_partial(struct nfs_write_data
*data
, int status
)
1045 struct nfs_page
*req
= data
->req
;
1046 struct page
*page
= req
->wb_page
;
1048 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1049 req
->wb_context
->dentry
->d_inode
->i_sb
->s_id
,
1050 (long long)NFS_FILEID(req
->wb_context
->dentry
->d_inode
),
1052 (long long)req_offset(req
));
1055 ClearPageUptodate(page
);
1057 req
->wb_context
->error
= status
;
1058 dprintk(", error = %d\n", status
);
1060 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1061 if (data
->verf
.committed
< NFS_FILE_SYNC
) {
1062 if (!NFS_NEED_COMMIT(req
)) {
1063 nfs_defer_commit(req
);
1064 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1065 dprintk(" defer commit\n");
1066 } else if (memcmp(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
))) {
1067 nfs_defer_reschedule(req
);
1068 dprintk(" server reboot detected\n");
1075 if (atomic_dec_and_test(&req
->wb_complete
))
1076 nfs_writepage_release(req
);
1080 * Handle a write reply that flushes a whole page.
1082 * FIXME: There is an inherent race with invalidate_inode_pages and
1083 * writebacks since the page->count is kept > 1 for as long
1084 * as the page has a write request pending.
1086 static void nfs_writeback_done_full(struct nfs_write_data
*data
, int status
)
1088 struct nfs_page
*req
;
1091 /* Update attributes as result of writeback. */
1092 while (!list_empty(&data
->pages
)) {
1093 req
= nfs_list_entry(data
->pages
.next
);
1094 nfs_list_remove_request(req
);
1095 page
= req
->wb_page
;
1097 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1098 req
->wb_context
->dentry
->d_inode
->i_sb
->s_id
,
1099 (long long)NFS_FILEID(req
->wb_context
->dentry
->d_inode
),
1101 (long long)req_offset(req
));
1104 ClearPageUptodate(page
);
1106 req
->wb_context
->error
= status
;
1107 end_page_writeback(page
);
1108 nfs_inode_remove_request(req
);
1109 dprintk(", error = %d\n", status
);
1112 end_page_writeback(page
);
1114 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1115 if (data
->args
.stable
!= NFS_UNSTABLE
|| data
->verf
.committed
== NFS_FILE_SYNC
) {
1116 nfs_inode_remove_request(req
);
1120 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1121 nfs_mark_request_commit(req
);
1122 dprintk(" marked for commit\n");
1124 nfs_inode_remove_request(req
);
1127 nfs_clear_page_writeback(req
);
1132 * This function is called when the WRITE call is complete.
1134 void nfs_writeback_done(struct rpc_task
*task
)
1136 struct nfs_write_data
*data
= (struct nfs_write_data
*) task
->tk_calldata
;
1137 struct nfs_writeargs
*argp
= &data
->args
;
1138 struct nfs_writeres
*resp
= &data
->res
;
1140 dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1141 task
->tk_pid
, task
->tk_status
);
1143 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1144 if (resp
->verf
->committed
< argp
->stable
&& task
->tk_status
>= 0) {
1145 /* We tried a write call, but the server did not
1146 * commit data to stable storage even though we
1148 * Note: There is a known bug in Tru64 < 5.0 in which
1149 * the server reports NFS_DATA_SYNC, but performs
1150 * NFS_FILE_SYNC. We therefore implement this checking
1151 * as a dprintk() in order to avoid filling syslog.
1153 static unsigned long complain
;
1155 if (time_before(complain
, jiffies
)) {
1156 dprintk("NFS: faulty NFS server %s:"
1157 " (committed = %d) != (stable = %d)\n",
1158 NFS_SERVER(data
->inode
)->hostname
,
1159 resp
->verf
->committed
, argp
->stable
);
1160 complain
= jiffies
+ 300 * HZ
;
1164 /* Is this a short write? */
1165 if (task
->tk_status
>= 0 && resp
->count
< argp
->count
) {
1166 static unsigned long complain
;
1168 /* Has the server at least made some progress? */
1169 if (resp
->count
!= 0) {
1170 /* Was this an NFSv2 write or an NFSv3 stable write? */
1171 if (resp
->verf
->committed
!= NFS_UNSTABLE
) {
1172 /* Resend from where the server left off */
1173 argp
->offset
+= resp
->count
;
1174 argp
->pgbase
+= resp
->count
;
1175 argp
->count
-= resp
->count
;
1177 /* Resend as a stable write in order to avoid
1178 * headaches in the case of a server crash.
1180 argp
->stable
= NFS_FILE_SYNC
;
1182 rpc_restart_call(task
);
1185 if (time_before(complain
, jiffies
)) {
1187 "NFS: Server wrote zero bytes, expected %u.\n",
1189 complain
= jiffies
+ 300 * HZ
;
1191 /* Can't do anything about it except throw an error. */
1192 task
->tk_status
= -EIO
;
1196 * Process the nfs_page list
1198 data
->complete(data
, task
->tk_status
);
1202 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1203 static void nfs_commit_release(struct rpc_task
*task
)
1205 struct nfs_write_data
*wdata
= (struct nfs_write_data
*)task
->tk_calldata
;
1206 nfs_commit_free(wdata
);
1210 * Set up the argument/result storage required for the RPC call.
1212 static void nfs_commit_rpcsetup(struct list_head
*head
,
1213 struct nfs_write_data
*data
, int how
)
1215 struct nfs_page
*first
;
1216 struct inode
*inode
;
1218 /* Set up the RPC argument and reply structs
1219 * NB: take care not to mess about with data->commit et al. */
1221 list_splice_init(head
, &data
->pages
);
1222 first
= nfs_list_entry(data
->pages
.next
);
1223 inode
= first
->wb_context
->dentry
->d_inode
;
1225 data
->inode
= inode
;
1226 data
->cred
= first
->wb_context
->cred
;
1228 data
->args
.fh
= NFS_FH(data
->inode
);
1229 /* Note: we always request a commit of the entire inode */
1230 data
->args
.offset
= 0;
1231 data
->args
.count
= 0;
1232 data
->res
.count
= 0;
1233 data
->res
.fattr
= &data
->fattr
;
1234 data
->res
.verf
= &data
->verf
;
1235 nfs_fattr_init(&data
->fattr
);
1237 NFS_PROTO(inode
)->commit_setup(data
, how
);
1239 data
->task
.tk_priority
= flush_task_priority(how
);
1240 data
->task
.tk_cookie
= (unsigned long)inode
;
1241 data
->task
.tk_calldata
= data
;
1242 /* Release requests */
1243 data
->task
.tk_release
= nfs_commit_release
;
1245 dprintk("NFS: %4d initiated commit call\n", data
->task
.tk_pid
);
1249 * Commit dirty pages
1252 nfs_commit_list(struct list_head
*head
, int how
)
1254 struct nfs_write_data
*data
;
1255 struct nfs_page
*req
;
1257 data
= nfs_commit_alloc();
1262 /* Set up the argument struct */
1263 nfs_commit_rpcsetup(head
, data
, how
);
1265 nfs_execute_write(data
);
1268 while (!list_empty(head
)) {
1269 req
= nfs_list_entry(head
->next
);
1270 nfs_list_remove_request(req
);
1271 nfs_mark_request_commit(req
);
1272 nfs_clear_page_writeback(req
);
1278 * COMMIT call returned
1281 nfs_commit_done(struct rpc_task
*task
)
1283 struct nfs_write_data
*data
= (struct nfs_write_data
*)task
->tk_calldata
;
1284 struct nfs_page
*req
;
1287 dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1288 task
->tk_pid
, task
->tk_status
);
1290 while (!list_empty(&data
->pages
)) {
1291 req
= nfs_list_entry(data
->pages
.next
);
1292 nfs_list_remove_request(req
);
1294 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1295 req
->wb_context
->dentry
->d_inode
->i_sb
->s_id
,
1296 (long long)NFS_FILEID(req
->wb_context
->dentry
->d_inode
),
1298 (long long)req_offset(req
));
1299 if (task
->tk_status
< 0) {
1300 req
->wb_context
->error
= task
->tk_status
;
1301 nfs_inode_remove_request(req
);
1302 dprintk(", error = %d\n", task
->tk_status
);
1306 /* Okay, COMMIT succeeded, apparently. Check the verifier
1307 * returned by the server against all stored verfs. */
1308 if (!memcmp(req
->wb_verf
.verifier
, data
->verf
.verifier
, sizeof(data
->verf
.verifier
))) {
1309 /* We have a match */
1310 nfs_inode_remove_request(req
);
1314 /* We have a mismatch. Write the page again */
1315 dprintk(" mismatch\n");
1316 nfs_mark_request_dirty(req
);
1318 nfs_clear_page_writeback(req
);
1321 sub_page_state(nr_unstable
,res
);
1325 static int nfs_flush_inode(struct inode
*inode
, unsigned long idx_start
,
1326 unsigned int npages
, int how
)
1328 struct nfs_inode
*nfsi
= NFS_I(inode
);
1333 spin_lock(&nfsi
->req_lock
);
1334 res
= nfs_scan_dirty(inode
, &head
, idx_start
, npages
);
1335 spin_unlock(&nfsi
->req_lock
);
1337 struct nfs_server
*server
= NFS_SERVER(inode
);
1339 /* For single writes, FLUSH_STABLE is more efficient */
1340 if (res
== nfsi
->npages
&& nfsi
->npages
<= server
->wpages
) {
1341 if (res
> 1 || nfs_list_entry(head
.next
)->wb_bytes
<= server
->wsize
)
1342 how
|= FLUSH_STABLE
;
1344 error
= nfs_flush_list(&head
, server
->wpages
, how
);
1351 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1352 int nfs_commit_inode(struct inode
*inode
, int how
)
1354 struct nfs_inode
*nfsi
= NFS_I(inode
);
1359 spin_lock(&nfsi
->req_lock
);
1360 res
= nfs_scan_commit(inode
, &head
, 0, 0);
1361 spin_unlock(&nfsi
->req_lock
);
1363 error
= nfs_commit_list(&head
, how
);
1371 int nfs_sync_inode(struct inode
*inode
, unsigned long idx_start
,
1372 unsigned int npages
, int how
)
1377 wait
= how
& FLUSH_WAIT
;
1383 error
= nfs_wait_on_requests(inode
, idx_start
, npages
);
1385 error
= nfs_flush_inode(inode
, idx_start
, npages
, how
);
1386 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1388 error
= nfs_commit_inode(inode
, how
);
1390 } while (error
> 0);
1394 int nfs_init_writepagecache(void)
1396 nfs_wdata_cachep
= kmem_cache_create("nfs_write_data",
1397 sizeof(struct nfs_write_data
),
1398 0, SLAB_HWCACHE_ALIGN
,
1400 if (nfs_wdata_cachep
== NULL
)
1403 nfs_wdata_mempool
= mempool_create(MIN_POOL_WRITE
,
1407 if (nfs_wdata_mempool
== NULL
)
1410 nfs_commit_mempool
= mempool_create(MIN_POOL_COMMIT
,
1414 if (nfs_commit_mempool
== NULL
)
1420 void nfs_destroy_writepagecache(void)
1422 mempool_destroy(nfs_commit_mempool
);
1423 mempool_destroy(nfs_wdata_mempool
);
1424 if (kmem_cache_destroy(nfs_wdata_cachep
))
1425 printk(KERN_INFO
"nfs_write_data: not all structures were freed\n");