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"
68 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
70 #define MIN_POOL_WRITE (32)
71 #define MIN_POOL_COMMIT (4)
74 * Local function declarations
76 static struct nfs_page
* nfs_update_request(struct nfs_open_context
*,
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
;
88 static kmem_cache_t
*nfs_wdata_cachep
;
89 static mempool_t
*nfs_wdata_mempool
;
90 static mempool_t
*nfs_commit_mempool
;
92 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion
);
94 struct nfs_write_data
*nfs_commit_alloc(unsigned int pagecount
)
96 struct nfs_write_data
*p
= mempool_alloc(nfs_commit_mempool
, SLAB_NOFS
);
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
;
104 p
->pagevec
= kcalloc(pagecount
, sizeof(struct page
*), GFP_NOFS
);
106 mempool_free(p
, nfs_commit_mempool
);
114 void nfs_commit_free(struct nfs_write_data
*p
)
116 if (p
&& (p
->pagevec
!= &p
->page_array
[0]))
118 mempool_free(p
, nfs_commit_mempool
);
121 struct nfs_write_data
*nfs_writedata_alloc(unsigned int pagecount
)
123 struct nfs_write_data
*p
= mempool_alloc(nfs_wdata_mempool
, SLAB_NOFS
);
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
;
131 p
->pagevec
= kcalloc(pagecount
, sizeof(struct page
*), GFP_NOFS
);
133 mempool_free(p
, nfs_wdata_mempool
);
141 void nfs_writedata_free(struct nfs_write_data
*p
)
143 if (p
&& (p
->pagevec
!= &p
->page_array
[0]))
145 mempool_free(p
, nfs_wdata_mempool
);
148 void nfs_writedata_release(void *wdata
)
150 nfs_writedata_free(wdata
);
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
)
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
;
160 if (i_size
> 0 && page
->index
< end_index
)
162 end
= ((loff_t
)page
->index
<< PAGE_CACHE_SHIFT
) + ((loff_t
)offset
+count
);
165 nfs_inc_stats(inode
, NFSIOS_EXTENDWRITE
);
166 i_size_write(inode
, end
);
169 /* We can set the PG_uptodate flag if we see that a write request
170 * covers the full page.
172 static void nfs_mark_uptodate(struct page
*page
, unsigned int base
, unsigned int count
)
176 if (PageUptodate(page
))
180 if (count
== PAGE_CACHE_SIZE
) {
181 SetPageUptodate(page
);
185 end_offs
= i_size_read(page
->mapping
->host
) - 1;
188 /* Is this the last page? */
189 if (page
->index
!= (unsigned long)(end_offs
>> PAGE_CACHE_SHIFT
))
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.
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
);
201 * Write a page synchronously.
202 * Offset is the data offset within the page.
204 static int nfs_writepage_sync(struct nfs_open_context
*ctx
, struct inode
*inode
,
205 struct page
*page
, unsigned int offset
, unsigned int count
,
208 unsigned int wsize
= NFS_SERVER(inode
)->wsize
;
209 int result
, written
= 0;
210 struct nfs_write_data
*wdata
;
212 wdata
= nfs_writedata_alloc(1);
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
;
228 dprintk("NFS: nfs_writepage_sync(%s/%Ld %d@%Ld)\n",
230 (long long)NFS_FILEID(inode
),
231 count
, (long long)(page_offset(page
) + offset
));
233 set_page_writeback(page
);
234 nfs_begin_data_update(inode
);
237 wdata
->args
.count
= count
;
238 wdata
->args
.offset
= page_offset(page
) + wdata
->args
.pgbase
;
240 result
= NFS_PROTO(inode
)->write(wdata
);
243 /* Must mark the page invalid after I/O error */
244 ClearPageUptodate(page
);
247 if (result
< wdata
->args
.count
)
248 printk(KERN_WARNING
"NFS: short write, count=%u, result=%d\n",
249 wdata
->args
.count
, result
);
251 wdata
->args
.offset
+= result
;
252 wdata
->args
.pgbase
+= result
;
255 nfs_add_stats(inode
, NFSIOS_SERVERWRITTENBYTES
, result
);
257 /* Update file length */
258 nfs_grow_file(page
, offset
, written
);
259 /* Set the PG_uptodate flag? */
260 nfs_mark_uptodate(page
, offset
, written
);
263 ClearPageError(page
);
266 nfs_end_data_update(inode
);
267 end_page_writeback(page
);
268 nfs_writedata_free(wdata
);
269 return written
? written
: result
;
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
)
276 struct nfs_page
*req
;
278 req
= nfs_update_request(ctx
, inode
, page
, offset
, count
);
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
);
289 static int wb_priority(struct writeback_control
*wbc
)
291 if (wbc
->for_reclaim
)
292 return FLUSH_HIGHPRI
;
293 if (wbc
->for_kupdate
)
299 * Write an mmapped page to the server.
301 int nfs_writepage(struct page
*page
, struct writeback_control
*wbc
)
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
);
312 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGE
);
313 nfs_add_stats(inode
, NFSIOS_WRITEPAGES
, 1);
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().
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.
323 if (igrab(inode
) != 0)
324 inode_referenced
= 1;
325 end_index
= i_size
>> PAGE_CACHE_SHIFT
;
327 /* Ensure we've flushed out any previous writes */
328 nfs_wb_page_priority(inode
, page
, priority
);
331 if (page
->index
< end_index
)
333 /* things got complicated... */
334 offset
= i_size
& (PAGE_CACHE_SIZE
-1);
336 /* OK, are we completely out? */
337 err
= 0; /* potential race with truncate - ignore */
338 if (page
->index
>= end_index
+1 || !offset
)
341 ctx
= nfs_find_open_context(inode
, NULL
, FMODE_WRITE
);
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
));
352 err
= nfs_writepage_sync(ctx
, inode
, page
, 0,
356 redirty_page_for_writepage(wbc
, page
);
361 put_nfs_open_context(ctx
);
364 if (inode_referenced
)
370 * Note: causes nfs_update_request() to block on the assumption
371 * that the writeback is generated due to memory pressure.
373 int nfs_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
375 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
376 struct inode
*inode
= mapping
->host
;
379 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGES
);
381 err
= generic_writepages(mapping
, wbc
);
384 while (test_and_set_bit(BDI_write_congested
, &bdi
->state
) != 0) {
385 if (wbc
->nonblocking
)
387 nfs_wait_on_write_congestion(mapping
, 0);
389 err
= nfs_flush_inode(inode
, 0, 0, wb_priority(wbc
));
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);
399 err
= nfs_commit_inode(inode
, wb_priority(wbc
));
401 wbc
->nr_to_write
-= err
;
405 clear_bit(BDI_write_congested
, &bdi
->state
);
406 wake_up_all(&nfs_write_congestion
);
411 * Insert a write request into an inode
413 static int nfs_inode_add_request(struct inode
*inode
, struct nfs_page
*req
)
415 struct nfs_inode
*nfsi
= NFS_I(inode
);
418 error
= radix_tree_insert(&nfsi
->nfs_page_tree
, req
->wb_index
, req
);
419 BUG_ON(error
== -EEXIST
);
424 nfs_begin_data_update(inode
);
425 if (nfs_have_delegation(inode
, FMODE_WRITE
))
428 SetPagePrivate(req
->wb_page
);
430 atomic_inc(&req
->wb_count
);
435 * Insert a write request into an inode
437 static void nfs_inode_remove_request(struct nfs_page
*req
)
439 struct inode
*inode
= req
->wb_context
->dentry
->d_inode
;
440 struct nfs_inode
*nfsi
= NFS_I(inode
);
442 BUG_ON (!NFS_WBACK_BUSY(req
));
444 spin_lock(&nfsi
->req_lock
);
445 ClearPagePrivate(req
->wb_page
);
446 radix_tree_delete(&nfsi
->nfs_page_tree
, req
->wb_index
);
449 spin_unlock(&nfsi
->req_lock
);
450 nfs_end_data_update(inode
);
453 spin_unlock(&nfsi
->req_lock
);
454 nfs_clear_request(req
);
455 nfs_release_request(req
);
461 static inline struct nfs_page
*
462 _nfs_find_request(struct inode
*inode
, unsigned long index
)
464 struct nfs_inode
*nfsi
= NFS_I(inode
);
465 struct nfs_page
*req
;
467 req
= (struct nfs_page
*)radix_tree_lookup(&nfsi
->nfs_page_tree
, index
);
469 atomic_inc(&req
->wb_count
);
473 static struct nfs_page
*
474 nfs_find_request(struct inode
*inode
, unsigned long index
)
476 struct nfs_page
*req
;
477 struct nfs_inode
*nfsi
= NFS_I(inode
);
479 spin_lock(&nfsi
->req_lock
);
480 req
= _nfs_find_request(inode
, index
);
481 spin_unlock(&nfsi
->req_lock
);
486 * Add a request to the inode's dirty list.
489 nfs_mark_request_dirty(struct nfs_page
*req
)
491 struct inode
*inode
= req
->wb_context
->dentry
->d_inode
;
492 struct nfs_inode
*nfsi
= NFS_I(inode
);
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
);
499 spin_unlock(&nfsi
->req_lock
);
500 inc_page_state(nr_dirty
);
501 mark_inode_dirty(inode
);
505 * Check if a request is dirty
508 nfs_dirty_request(struct nfs_page
*req
)
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
;
514 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
516 * Add a request to the inode's commit list.
519 nfs_mark_request_commit(struct nfs_page
*req
)
521 struct inode
*inode
= req
->wb_context
->dentry
->d_inode
;
522 struct nfs_inode
*nfsi
= NFS_I(inode
);
524 spin_lock(&nfsi
->req_lock
);
525 nfs_list_add_request(req
, &nfsi
->commit
);
527 spin_unlock(&nfsi
->req_lock
);
528 inc_page_state(nr_unstable
);
529 mark_inode_dirty(inode
);
534 * Wait for a request to complete.
536 * Interruptible by signals only if mounted with intr flag.
538 static int nfs_wait_on_requests_locked(struct inode
*inode
, unsigned long idx_start
, unsigned int npages
)
540 struct nfs_inode
*nfsi
= NFS_I(inode
);
541 struct nfs_page
*req
;
542 unsigned long idx_end
, next
;
543 unsigned int res
= 0;
549 idx_end
= idx_start
+ npages
- 1;
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
)
556 next
= req
->wb_index
+ 1;
557 BUG_ON(!NFS_WBACK_BUSY(req
));
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
);
571 static int nfs_wait_on_requests(struct inode
*inode
, unsigned long idx_start
, unsigned int npages
)
573 struct nfs_inode
*nfsi
= NFS_I(inode
);
576 spin_lock(&nfsi
->req_lock
);
577 ret
= nfs_wait_on_requests_locked(inode
, idx_start
, npages
);
578 spin_unlock(&nfsi
->req_lock
);
583 * nfs_scan_dirty - Scan an inode for dirty requests
584 * @inode: NFS inode to scan
585 * @dst: destination list
586 * @idx_start: lower bound of page->index to scan.
587 * @npages: idx_start + npages sets the upper bound to scan.
589 * Moves requests from the inode's dirty page list.
590 * The requests are *not* checked to ensure that they form a contiguous set.
593 nfs_scan_dirty(struct inode
*inode
, struct list_head
*dst
, unsigned long idx_start
, unsigned int npages
)
595 struct nfs_inode
*nfsi
= NFS_I(inode
);
598 if (nfsi
->ndirty
!= 0) {
599 res
= nfs_scan_lock_dirty(nfsi
, dst
, idx_start
, npages
);
601 sub_page_state(nr_dirty
,res
);
602 if ((nfsi
->ndirty
== 0) != list_empty(&nfsi
->dirty
))
603 printk(KERN_ERR
"NFS: desynchronized value of nfs_i.ndirty.\n");
608 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
610 * nfs_scan_commit - Scan an inode for commit requests
611 * @inode: NFS inode to scan
612 * @dst: destination list
613 * @idx_start: lower bound of page->index to scan.
614 * @npages: idx_start + npages sets the upper bound to scan.
616 * Moves requests from the inode's 'commit' request list.
617 * The requests are *not* checked to ensure that they form a contiguous set.
620 nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, unsigned long idx_start
, unsigned int npages
)
622 struct nfs_inode
*nfsi
= NFS_I(inode
);
625 if (nfsi
->ncommit
!= 0) {
626 res
= nfs_scan_list(&nfsi
->commit
, dst
, idx_start
, npages
);
627 nfsi
->ncommit
-= res
;
628 if ((nfsi
->ncommit
== 0) != list_empty(&nfsi
->commit
))
629 printk(KERN_ERR
"NFS: desynchronized value of nfs_i.ncommit.\n");
634 static inline int nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, unsigned long idx_start
, unsigned int npages
)
640 static int nfs_wait_on_write_congestion(struct address_space
*mapping
, int intr
)
642 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
648 if (!bdi_write_congested(bdi
))
651 nfs_inc_stats(mapping
->host
, NFSIOS_CONGESTIONWAIT
);
654 struct rpc_clnt
*clnt
= NFS_CLIENT(mapping
->host
);
657 rpc_clnt_sigmask(clnt
, &oldset
);
658 prepare_to_wait(&nfs_write_congestion
, &wait
, TASK_INTERRUPTIBLE
);
659 if (bdi_write_congested(bdi
)) {
665 rpc_clnt_sigunmask(clnt
, &oldset
);
667 prepare_to_wait(&nfs_write_congestion
, &wait
, TASK_UNINTERRUPTIBLE
);
668 if (bdi_write_congested(bdi
))
671 finish_wait(&nfs_write_congestion
, &wait
);
677 * Try to update any existing write request, or create one if there is none.
678 * In order to match, the request's credentials must match those of
679 * the calling process.
681 * Note: Should always be called with the Page Lock held!
683 static struct nfs_page
* nfs_update_request(struct nfs_open_context
* ctx
,
684 struct inode
*inode
, struct page
*page
,
685 unsigned int offset
, unsigned int bytes
)
687 struct nfs_server
*server
= NFS_SERVER(inode
);
688 struct nfs_inode
*nfsi
= NFS_I(inode
);
689 struct nfs_page
*req
, *new = NULL
;
690 unsigned long rqend
, end
;
692 end
= offset
+ bytes
;
694 if (nfs_wait_on_write_congestion(page
->mapping
, server
->flags
& NFS_MOUNT_INTR
))
695 return ERR_PTR(-ERESTARTSYS
);
697 /* Loop over all inode entries and see if we find
698 * A request for the page we wish to update
700 spin_lock(&nfsi
->req_lock
);
701 req
= _nfs_find_request(inode
, page
->index
);
703 if (!nfs_lock_request_dontget(req
)) {
705 spin_unlock(&nfsi
->req_lock
);
706 error
= nfs_wait_on_request(req
);
707 nfs_release_request(req
);
710 nfs_release_request(new);
711 return ERR_PTR(error
);
715 spin_unlock(&nfsi
->req_lock
);
717 nfs_release_request(new);
723 nfs_lock_request_dontget(new);
724 error
= nfs_inode_add_request(inode
, new);
726 spin_unlock(&nfsi
->req_lock
);
727 nfs_unlock_request(new);
728 return ERR_PTR(error
);
730 spin_unlock(&nfsi
->req_lock
);
731 nfs_mark_request_dirty(new);
734 spin_unlock(&nfsi
->req_lock
);
736 new = nfs_create_request(ctx
, inode
, page
, offset
, bytes
);
741 /* We have a request for our page.
742 * If the creds don't match, or the
743 * page addresses don't match,
744 * tell the caller to wait on the conflicting
747 rqend
= req
->wb_offset
+ req
->wb_bytes
;
748 if (req
->wb_context
!= ctx
749 || req
->wb_page
!= page
750 || !nfs_dirty_request(req
)
751 || offset
> rqend
|| end
< req
->wb_offset
) {
752 nfs_unlock_request(req
);
753 return ERR_PTR(-EBUSY
);
756 /* Okay, the request matches. Update the region */
757 if (offset
< req
->wb_offset
) {
758 req
->wb_offset
= offset
;
759 req
->wb_pgbase
= offset
;
760 req
->wb_bytes
= rqend
- req
->wb_offset
;
764 req
->wb_bytes
= end
- req
->wb_offset
;
769 int nfs_flush_incompatible(struct file
*file
, struct page
*page
)
771 struct nfs_open_context
*ctx
= (struct nfs_open_context
*)file
->private_data
;
772 struct inode
*inode
= page
->mapping
->host
;
773 struct nfs_page
*req
;
776 * Look for a request corresponding to this page. If there
777 * is one, and it belongs to another file, we flush it out
778 * before we try to copy anything into the page. Do this
779 * due to the lack of an ACCESS-type call in NFSv2.
780 * Also do the same if we find a request from an existing
783 req
= nfs_find_request(inode
, page
->index
);
785 if (req
->wb_page
!= page
|| ctx
!= req
->wb_context
)
786 status
= nfs_wb_page(inode
, page
);
787 nfs_release_request(req
);
789 return (status
< 0) ? status
: 0;
793 * Update and possibly write a cached page of an NFS file.
795 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
796 * things with a page scheduled for an RPC call (e.g. invalidate it).
798 int nfs_updatepage(struct file
*file
, struct page
*page
,
799 unsigned int offset
, unsigned int count
)
801 struct nfs_open_context
*ctx
= (struct nfs_open_context
*)file
->private_data
;
802 struct inode
*inode
= page
->mapping
->host
;
803 struct nfs_page
*req
;
806 nfs_inc_stats(inode
, NFSIOS_VFSUPDATEPAGE
);
808 dprintk("NFS: nfs_updatepage(%s/%s %d@%Ld)\n",
809 file
->f_dentry
->d_parent
->d_name
.name
,
810 file
->f_dentry
->d_name
.name
, count
,
811 (long long)(page_offset(page
) +offset
));
813 if (IS_SYNC(inode
)) {
814 status
= nfs_writepage_sync(ctx
, inode
, page
, offset
, count
, 0);
816 if (offset
== 0 && status
== PAGE_CACHE_SIZE
)
817 SetPageUptodate(page
);
823 /* If we're not using byte range locks, and we know the page
824 * is entirely in cache, it may be more efficient to avoid
825 * fragmenting write requests.
827 if (PageUptodate(page
) && inode
->i_flock
== NULL
&& !(file
->f_mode
& O_SYNC
)) {
828 loff_t end_offs
= i_size_read(inode
) - 1;
829 unsigned long end_index
= end_offs
>> PAGE_CACHE_SHIFT
;
833 if (unlikely(end_offs
< 0)) {
835 } else if (page
->index
== end_index
) {
837 pglen
= (unsigned int)(end_offs
& (PAGE_CACHE_SIZE
-1)) + 1;
840 } else if (page
->index
< end_index
)
841 count
= PAGE_CACHE_SIZE
;
845 * Try to find an NFS request corresponding to this page
847 * If the existing request cannot be updated, we must flush
851 req
= nfs_update_request(ctx
, inode
, page
, offset
, count
);
852 status
= (IS_ERR(req
)) ? PTR_ERR(req
) : 0;
853 if (status
!= -EBUSY
)
855 /* Request could not be updated. Flush it out and try again */
856 status
= nfs_wb_page(inode
, page
);
857 } while (status
>= 0);
863 /* Update file length */
864 nfs_grow_file(page
, offset
, count
);
865 /* Set the PG_uptodate flag? */
866 nfs_mark_uptodate(page
, req
->wb_pgbase
, req
->wb_bytes
);
867 nfs_unlock_request(req
);
869 dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n",
870 status
, (long long)i_size_read(inode
));
872 ClearPageUptodate(page
);
876 static void nfs_writepage_release(struct nfs_page
*req
)
878 end_page_writeback(req
->wb_page
);
880 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
881 if (!PageError(req
->wb_page
)) {
882 if (NFS_NEED_RESCHED(req
)) {
883 nfs_mark_request_dirty(req
);
885 } else if (NFS_NEED_COMMIT(req
)) {
886 nfs_mark_request_commit(req
);
890 nfs_inode_remove_request(req
);
893 nfs_clear_commit(req
);
894 nfs_clear_reschedule(req
);
896 nfs_inode_remove_request(req
);
898 nfs_clear_page_writeback(req
);
901 static inline int flush_task_priority(int how
)
903 switch (how
& (FLUSH_HIGHPRI
|FLUSH_LOWPRI
)) {
905 return RPC_PRIORITY_HIGH
;
907 return RPC_PRIORITY_LOW
;
909 return RPC_PRIORITY_NORMAL
;
913 * Set up the argument/result storage required for the RPC call.
915 static void nfs_write_rpcsetup(struct nfs_page
*req
,
916 struct nfs_write_data
*data
,
917 const struct rpc_call_ops
*call_ops
,
918 unsigned int count
, unsigned int offset
,
924 /* Set up the RPC argument and reply structs
925 * NB: take care not to mess about with data->commit et al. */
928 data
->inode
= inode
= req
->wb_context
->dentry
->d_inode
;
929 data
->cred
= req
->wb_context
->cred
;
931 data
->args
.fh
= NFS_FH(inode
);
932 data
->args
.offset
= req_offset(req
) + offset
;
933 data
->args
.pgbase
= req
->wb_pgbase
+ offset
;
934 data
->args
.pages
= data
->pagevec
;
935 data
->args
.count
= count
;
936 data
->args
.context
= req
->wb_context
;
938 data
->res
.fattr
= &data
->fattr
;
939 data
->res
.count
= count
;
940 data
->res
.verf
= &data
->verf
;
941 nfs_fattr_init(&data
->fattr
);
943 /* Set up the initial task struct. */
944 flags
= (how
& FLUSH_SYNC
) ? 0 : RPC_TASK_ASYNC
;
945 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), flags
, call_ops
, data
);
946 NFS_PROTO(inode
)->write_setup(data
, how
);
948 data
->task
.tk_priority
= flush_task_priority(how
);
949 data
->task
.tk_cookie
= (unsigned long)inode
;
951 dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
954 (long long)NFS_FILEID(inode
),
956 (unsigned long long)data
->args
.offset
);
959 static void nfs_execute_write(struct nfs_write_data
*data
)
961 struct rpc_clnt
*clnt
= NFS_CLIENT(data
->inode
);
964 rpc_clnt_sigmask(clnt
, &oldset
);
966 rpc_execute(&data
->task
);
968 rpc_clnt_sigunmask(clnt
, &oldset
);
972 * Generate multiple small requests to write out a single
973 * contiguous dirty area on one page.
975 static int nfs_flush_multi(struct inode
*inode
, struct list_head
*head
, int how
)
977 struct nfs_page
*req
= nfs_list_entry(head
->next
);
978 struct page
*page
= req
->wb_page
;
979 struct nfs_write_data
*data
;
980 unsigned int wsize
= NFS_SERVER(inode
)->wsize
;
981 unsigned int nbytes
, offset
;
985 nfs_list_remove_request(req
);
987 nbytes
= req
->wb_bytes
;
989 data
= nfs_writedata_alloc(1);
992 list_add(&data
->pages
, &list
);
998 atomic_set(&req
->wb_complete
, requests
);
1000 ClearPageError(page
);
1001 set_page_writeback(page
);
1003 nbytes
= req
->wb_bytes
;
1005 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
1006 list_del_init(&data
->pages
);
1008 data
->pagevec
[0] = page
;
1010 if (nbytes
> wsize
) {
1011 nfs_write_rpcsetup(req
, data
, &nfs_write_partial_ops
,
1012 wsize
, offset
, how
);
1016 nfs_write_rpcsetup(req
, data
, &nfs_write_partial_ops
,
1017 nbytes
, offset
, how
);
1020 nfs_execute_write(data
);
1021 } while (nbytes
!= 0);
1026 while (!list_empty(&list
)) {
1027 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
1028 list_del(&data
->pages
);
1029 nfs_writedata_free(data
);
1031 nfs_mark_request_dirty(req
);
1032 nfs_clear_page_writeback(req
);
1037 * Create an RPC task for the given write request and kick it.
1038 * The page must have been locked by the caller.
1040 * It may happen that the page we're passed is not marked dirty.
1041 * This is the case if nfs_updatepage detects a conflicting request
1042 * that has been written but not committed.
1044 static int nfs_flush_one(struct inode
*inode
, struct list_head
*head
, int how
)
1046 struct nfs_page
*req
;
1047 struct page
**pages
;
1048 struct nfs_write_data
*data
;
1051 data
= nfs_writedata_alloc(NFS_SERVER(inode
)->wpages
);
1055 pages
= data
->pagevec
;
1057 while (!list_empty(head
)) {
1058 req
= nfs_list_entry(head
->next
);
1059 nfs_list_remove_request(req
);
1060 nfs_list_add_request(req
, &data
->pages
);
1061 ClearPageError(req
->wb_page
);
1062 set_page_writeback(req
->wb_page
);
1063 *pages
++ = req
->wb_page
;
1064 count
+= req
->wb_bytes
;
1066 req
= nfs_list_entry(data
->pages
.next
);
1068 /* Set up the argument struct */
1069 nfs_write_rpcsetup(req
, data
, &nfs_write_full_ops
, count
, 0, how
);
1071 nfs_execute_write(data
);
1074 while (!list_empty(head
)) {
1075 struct nfs_page
*req
= nfs_list_entry(head
->next
);
1076 nfs_list_remove_request(req
);
1077 nfs_mark_request_dirty(req
);
1078 nfs_clear_page_writeback(req
);
1083 static int nfs_flush_list(struct inode
*inode
, struct list_head
*head
, int npages
, int how
)
1085 LIST_HEAD(one_request
);
1086 int (*flush_one
)(struct inode
*, struct list_head
*, int);
1087 struct nfs_page
*req
;
1088 int wpages
= NFS_SERVER(inode
)->wpages
;
1089 int wsize
= NFS_SERVER(inode
)->wsize
;
1092 flush_one
= nfs_flush_one
;
1093 if (wsize
< PAGE_CACHE_SIZE
)
1094 flush_one
= nfs_flush_multi
;
1095 /* For single writes, FLUSH_STABLE is more efficient */
1096 if (npages
<= wpages
&& npages
== NFS_I(inode
)->npages
1097 && nfs_list_entry(head
->next
)->wb_bytes
<= wsize
)
1098 how
|= FLUSH_STABLE
;
1101 nfs_coalesce_requests(head
, &one_request
, wpages
);
1102 req
= nfs_list_entry(one_request
.next
);
1103 error
= flush_one(inode
, &one_request
, how
);
1106 } while (!list_empty(head
));
1109 while (!list_empty(head
)) {
1110 req
= nfs_list_entry(head
->next
);
1111 nfs_list_remove_request(req
);
1112 nfs_mark_request_dirty(req
);
1113 nfs_clear_page_writeback(req
);
1119 * Handle a write reply that flushed part of a page.
1121 static void nfs_writeback_done_partial(struct rpc_task
*task
, void *calldata
)
1123 struct nfs_write_data
*data
= calldata
;
1124 struct nfs_page
*req
= data
->req
;
1125 struct page
*page
= req
->wb_page
;
1127 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1128 req
->wb_context
->dentry
->d_inode
->i_sb
->s_id
,
1129 (long long)NFS_FILEID(req
->wb_context
->dentry
->d_inode
),
1131 (long long)req_offset(req
));
1133 if (nfs_writeback_done(task
, data
) != 0)
1136 if (task
->tk_status
< 0) {
1137 ClearPageUptodate(page
);
1139 req
->wb_context
->error
= task
->tk_status
;
1140 dprintk(", error = %d\n", task
->tk_status
);
1142 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1143 if (data
->verf
.committed
< NFS_FILE_SYNC
) {
1144 if (!NFS_NEED_COMMIT(req
)) {
1145 nfs_defer_commit(req
);
1146 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1147 dprintk(" defer commit\n");
1148 } else if (memcmp(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
))) {
1149 nfs_defer_reschedule(req
);
1150 dprintk(" server reboot detected\n");
1157 if (atomic_dec_and_test(&req
->wb_complete
))
1158 nfs_writepage_release(req
);
1161 static const struct rpc_call_ops nfs_write_partial_ops
= {
1162 .rpc_call_done
= nfs_writeback_done_partial
,
1163 .rpc_release
= nfs_writedata_release
,
1167 * Handle a write reply that flushes a whole page.
1169 * FIXME: There is an inherent race with invalidate_inode_pages and
1170 * writebacks since the page->count is kept > 1 for as long
1171 * as the page has a write request pending.
1173 static void nfs_writeback_done_full(struct rpc_task
*task
, void *calldata
)
1175 struct nfs_write_data
*data
= calldata
;
1176 struct nfs_page
*req
;
1179 if (nfs_writeback_done(task
, data
) != 0)
1182 /* Update attributes as result of writeback. */
1183 while (!list_empty(&data
->pages
)) {
1184 req
= nfs_list_entry(data
->pages
.next
);
1185 nfs_list_remove_request(req
);
1186 page
= req
->wb_page
;
1188 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1189 req
->wb_context
->dentry
->d_inode
->i_sb
->s_id
,
1190 (long long)NFS_FILEID(req
->wb_context
->dentry
->d_inode
),
1192 (long long)req_offset(req
));
1194 if (task
->tk_status
< 0) {
1195 ClearPageUptodate(page
);
1197 req
->wb_context
->error
= task
->tk_status
;
1198 end_page_writeback(page
);
1199 nfs_inode_remove_request(req
);
1200 dprintk(", error = %d\n", task
->tk_status
);
1203 end_page_writeback(page
);
1205 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1206 if (data
->args
.stable
!= NFS_UNSTABLE
|| data
->verf
.committed
== NFS_FILE_SYNC
) {
1207 nfs_inode_remove_request(req
);
1211 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1212 nfs_mark_request_commit(req
);
1213 dprintk(" marked for commit\n");
1215 nfs_inode_remove_request(req
);
1218 nfs_clear_page_writeback(req
);
1222 static const struct rpc_call_ops nfs_write_full_ops
= {
1223 .rpc_call_done
= nfs_writeback_done_full
,
1224 .rpc_release
= nfs_writedata_release
,
1229 * This function is called when the WRITE call is complete.
1231 int nfs_writeback_done(struct rpc_task
*task
, struct nfs_write_data
*data
)
1233 struct nfs_writeargs
*argp
= &data
->args
;
1234 struct nfs_writeres
*resp
= &data
->res
;
1237 dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1238 task
->tk_pid
, task
->tk_status
);
1240 /* Call the NFS version-specific code */
1241 status
= NFS_PROTO(data
->inode
)->write_done(task
, data
);
1244 nfs_add_stats(data
->inode
, NFSIOS_SERVERWRITTENBYTES
, resp
->count
);
1246 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1247 if (resp
->verf
->committed
< argp
->stable
&& task
->tk_status
>= 0) {
1248 /* We tried a write call, but the server did not
1249 * commit data to stable storage even though we
1251 * Note: There is a known bug in Tru64 < 5.0 in which
1252 * the server reports NFS_DATA_SYNC, but performs
1253 * NFS_FILE_SYNC. We therefore implement this checking
1254 * as a dprintk() in order to avoid filling syslog.
1256 static unsigned long complain
;
1258 if (time_before(complain
, jiffies
)) {
1259 dprintk("NFS: faulty NFS server %s:"
1260 " (committed = %d) != (stable = %d)\n",
1261 NFS_SERVER(data
->inode
)->hostname
,
1262 resp
->verf
->committed
, argp
->stable
);
1263 complain
= jiffies
+ 300 * HZ
;
1267 /* Is this a short write? */
1268 if (task
->tk_status
>= 0 && resp
->count
< argp
->count
) {
1269 static unsigned long complain
;
1271 nfs_inc_stats(data
->inode
, NFSIOS_SHORTWRITE
);
1273 /* Has the server at least made some progress? */
1274 if (resp
->count
!= 0) {
1275 /* Was this an NFSv2 write or an NFSv3 stable write? */
1276 if (resp
->verf
->committed
!= NFS_UNSTABLE
) {
1277 /* Resend from where the server left off */
1278 argp
->offset
+= resp
->count
;
1279 argp
->pgbase
+= resp
->count
;
1280 argp
->count
-= resp
->count
;
1282 /* Resend as a stable write in order to avoid
1283 * headaches in the case of a server crash.
1285 argp
->stable
= NFS_FILE_SYNC
;
1287 rpc_restart_call(task
);
1290 if (time_before(complain
, jiffies
)) {
1292 "NFS: Server wrote zero bytes, expected %u.\n",
1294 complain
= jiffies
+ 300 * HZ
;
1296 /* Can't do anything about it except throw an error. */
1297 task
->tk_status
= -EIO
;
1303 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1304 void nfs_commit_release(void *wdata
)
1306 nfs_commit_free(wdata
);
1310 * Set up the argument/result storage required for the RPC call.
1312 static void nfs_commit_rpcsetup(struct list_head
*head
,
1313 struct nfs_write_data
*data
,
1316 struct nfs_page
*first
;
1317 struct inode
*inode
;
1320 /* Set up the RPC argument and reply structs
1321 * NB: take care not to mess about with data->commit et al. */
1323 list_splice_init(head
, &data
->pages
);
1324 first
= nfs_list_entry(data
->pages
.next
);
1325 inode
= first
->wb_context
->dentry
->d_inode
;
1327 data
->inode
= inode
;
1328 data
->cred
= first
->wb_context
->cred
;
1330 data
->args
.fh
= NFS_FH(data
->inode
);
1331 /* Note: we always request a commit of the entire inode */
1332 data
->args
.offset
= 0;
1333 data
->args
.count
= 0;
1334 data
->res
.count
= 0;
1335 data
->res
.fattr
= &data
->fattr
;
1336 data
->res
.verf
= &data
->verf
;
1337 nfs_fattr_init(&data
->fattr
);
1339 /* Set up the initial task struct. */
1340 flags
= (how
& FLUSH_SYNC
) ? 0 : RPC_TASK_ASYNC
;
1341 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), flags
, &nfs_commit_ops
, data
);
1342 NFS_PROTO(inode
)->commit_setup(data
, how
);
1344 data
->task
.tk_priority
= flush_task_priority(how
);
1345 data
->task
.tk_cookie
= (unsigned long)inode
;
1347 dprintk("NFS: %4d initiated commit call\n", data
->task
.tk_pid
);
1351 * Commit dirty pages
1354 nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
)
1356 struct nfs_write_data
*data
;
1357 struct nfs_page
*req
;
1359 data
= nfs_commit_alloc(NFS_SERVER(inode
)->wpages
);
1364 /* Set up the argument struct */
1365 nfs_commit_rpcsetup(head
, data
, how
);
1367 nfs_execute_write(data
);
1370 while (!list_empty(head
)) {
1371 req
= nfs_list_entry(head
->next
);
1372 nfs_list_remove_request(req
);
1373 nfs_mark_request_commit(req
);
1374 nfs_clear_page_writeback(req
);
1380 * COMMIT call returned
1382 static void nfs_commit_done(struct rpc_task
*task
, void *calldata
)
1384 struct nfs_write_data
*data
= calldata
;
1385 struct nfs_page
*req
;
1388 dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1389 task
->tk_pid
, task
->tk_status
);
1391 /* Call the NFS version-specific code */
1392 if (NFS_PROTO(data
->inode
)->commit_done(task
, data
) != 0)
1395 while (!list_empty(&data
->pages
)) {
1396 req
= nfs_list_entry(data
->pages
.next
);
1397 nfs_list_remove_request(req
);
1399 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1400 req
->wb_context
->dentry
->d_inode
->i_sb
->s_id
,
1401 (long long)NFS_FILEID(req
->wb_context
->dentry
->d_inode
),
1403 (long long)req_offset(req
));
1404 if (task
->tk_status
< 0) {
1405 req
->wb_context
->error
= task
->tk_status
;
1406 nfs_inode_remove_request(req
);
1407 dprintk(", error = %d\n", task
->tk_status
);
1411 /* Okay, COMMIT succeeded, apparently. Check the verifier
1412 * returned by the server against all stored verfs. */
1413 if (!memcmp(req
->wb_verf
.verifier
, data
->verf
.verifier
, sizeof(data
->verf
.verifier
))) {
1414 /* We have a match */
1415 nfs_inode_remove_request(req
);
1419 /* We have a mismatch. Write the page again */
1420 dprintk(" mismatch\n");
1421 nfs_mark_request_dirty(req
);
1423 nfs_clear_page_writeback(req
);
1426 sub_page_state(nr_unstable
,res
);
1429 static const struct rpc_call_ops nfs_commit_ops
= {
1430 .rpc_call_done
= nfs_commit_done
,
1431 .rpc_release
= nfs_commit_release
,
1434 static inline int nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
)
1440 static int nfs_flush_inode(struct inode
*inode
, unsigned long idx_start
,
1441 unsigned int npages
, int how
)
1443 struct nfs_inode
*nfsi
= NFS_I(inode
);
1447 spin_lock(&nfsi
->req_lock
);
1448 res
= nfs_scan_dirty(inode
, &head
, idx_start
, npages
);
1449 spin_unlock(&nfsi
->req_lock
);
1451 int error
= nfs_flush_list(inode
, &head
, res
, how
);
1458 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1459 int nfs_commit_inode(struct inode
*inode
, int how
)
1461 struct nfs_inode
*nfsi
= NFS_I(inode
);
1465 spin_lock(&nfsi
->req_lock
);
1466 res
= nfs_scan_commit(inode
, &head
, 0, 0);
1467 spin_unlock(&nfsi
->req_lock
);
1469 int error
= nfs_commit_list(inode
, &head
, how
);
1477 int nfs_sync_inode_wait(struct inode
*inode
, unsigned long idx_start
,
1478 unsigned int npages
, int how
)
1480 struct nfs_inode
*nfsi
= NFS_I(inode
);
1482 int nocommit
= how
& FLUSH_NOCOMMIT
;
1485 how
&= ~FLUSH_NOCOMMIT
;
1486 spin_lock(&nfsi
->req_lock
);
1488 ret
= nfs_wait_on_requests_locked(inode
, idx_start
, npages
);
1491 pages
= nfs_scan_dirty(inode
, &head
, idx_start
, npages
);
1493 spin_unlock(&nfsi
->req_lock
);
1494 ret
= nfs_flush_list(inode
, &head
, pages
, how
);
1495 spin_lock(&nfsi
->req_lock
);
1500 pages
= nfs_scan_commit(inode
, &head
, 0, 0);
1503 spin_unlock(&nfsi
->req_lock
);
1504 ret
= nfs_commit_list(inode
, &head
, how
);
1505 spin_lock(&nfsi
->req_lock
);
1507 spin_unlock(&nfsi
->req_lock
);
1511 int nfs_init_writepagecache(void)
1513 nfs_wdata_cachep
= kmem_cache_create("nfs_write_data",
1514 sizeof(struct nfs_write_data
),
1515 0, SLAB_HWCACHE_ALIGN
,
1517 if (nfs_wdata_cachep
== NULL
)
1520 nfs_wdata_mempool
= mempool_create_slab_pool(MIN_POOL_WRITE
,
1522 if (nfs_wdata_mempool
== NULL
)
1525 nfs_commit_mempool
= mempool_create_slab_pool(MIN_POOL_COMMIT
,
1527 if (nfs_commit_mempool
== NULL
)
1533 void nfs_destroy_writepagecache(void)
1535 mempool_destroy(nfs_commit_mempool
);
1536 mempool_destroy(nfs_wdata_mempool
);
1537 if (kmem_cache_destroy(nfs_wdata_cachep
))
1538 printk(KERN_INFO
"nfs_write_data: not all structures were freed\n");