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
< NFS_PAGEVEC_SIZE
)
102 p
->pagevec
= &p
->page_array
[0];
104 size_t size
= ++pagecount
* sizeof(struct page
*);
105 p
->pagevec
= kzalloc(size
, GFP_NOFS
);
107 mempool_free(p
, nfs_commit_mempool
);
115 void nfs_commit_free(struct nfs_write_data
*p
)
117 if (p
&& (p
->pagevec
!= &p
->page_array
[0]))
119 mempool_free(p
, nfs_commit_mempool
);
122 struct nfs_write_data
*nfs_writedata_alloc(unsigned int pagecount
)
124 struct nfs_write_data
*p
= mempool_alloc(nfs_wdata_mempool
, SLAB_NOFS
);
127 memset(p
, 0, sizeof(*p
));
128 INIT_LIST_HEAD(&p
->pages
);
129 if (pagecount
< NFS_PAGEVEC_SIZE
)
130 p
->pagevec
= &p
->page_array
[0];
132 size_t size
= ++pagecount
* sizeof(struct page
*);
133 p
->pagevec
= kmalloc(size
, GFP_NOFS
);
135 memset(p
->pagevec
, 0, size
);
137 mempool_free(p
, nfs_wdata_mempool
);
145 void nfs_writedata_free(struct nfs_write_data
*p
)
147 if (p
&& (p
->pagevec
!= &p
->page_array
[0]))
149 mempool_free(p
, nfs_wdata_mempool
);
152 void nfs_writedata_release(void *wdata
)
154 nfs_writedata_free(wdata
);
157 /* Adjust the file length if we're writing beyond the end */
158 static void nfs_grow_file(struct page
*page
, unsigned int offset
, unsigned int count
)
160 struct inode
*inode
= page
->mapping
->host
;
161 loff_t end
, i_size
= i_size_read(inode
);
162 unsigned long end_index
= (i_size
- 1) >> PAGE_CACHE_SHIFT
;
164 if (i_size
> 0 && page
->index
< end_index
)
166 end
= ((loff_t
)page
->index
<< PAGE_CACHE_SHIFT
) + ((loff_t
)offset
+count
);
169 nfs_inc_stats(inode
, NFSIOS_EXTENDWRITE
);
170 i_size_write(inode
, end
);
173 /* We can set the PG_uptodate flag if we see that a write request
174 * covers the full page.
176 static void nfs_mark_uptodate(struct page
*page
, unsigned int base
, unsigned int count
)
180 if (PageUptodate(page
))
184 if (count
== PAGE_CACHE_SIZE
) {
185 SetPageUptodate(page
);
189 end_offs
= i_size_read(page
->mapping
->host
) - 1;
192 /* Is this the last page? */
193 if (page
->index
!= (unsigned long)(end_offs
>> PAGE_CACHE_SHIFT
))
195 /* This is the last page: set PG_uptodate if we cover the entire
196 * extent of the data, then zero the rest of the page.
198 if (count
== (unsigned int)(end_offs
& (PAGE_CACHE_SIZE
- 1)) + 1) {
199 memclear_highpage_flush(page
, count
, PAGE_CACHE_SIZE
- count
);
200 SetPageUptodate(page
);
205 * Write a page synchronously.
206 * Offset is the data offset within the page.
208 static int nfs_writepage_sync(struct nfs_open_context
*ctx
, struct inode
*inode
,
209 struct page
*page
, unsigned int offset
, unsigned int count
,
212 unsigned int wsize
= NFS_SERVER(inode
)->wsize
;
213 int result
, written
= 0;
214 struct nfs_write_data
*wdata
;
216 wdata
= nfs_writedata_alloc(1);
221 wdata
->cred
= ctx
->cred
;
222 wdata
->inode
= inode
;
223 wdata
->args
.fh
= NFS_FH(inode
);
224 wdata
->args
.context
= ctx
;
225 wdata
->args
.pages
= &page
;
226 wdata
->args
.stable
= NFS_FILE_SYNC
;
227 wdata
->args
.pgbase
= offset
;
228 wdata
->args
.count
= wsize
;
229 wdata
->res
.fattr
= &wdata
->fattr
;
230 wdata
->res
.verf
= &wdata
->verf
;
232 dprintk("NFS: nfs_writepage_sync(%s/%Ld %d@%Ld)\n",
234 (long long)NFS_FILEID(inode
),
235 count
, (long long)(page_offset(page
) + offset
));
237 set_page_writeback(page
);
238 nfs_begin_data_update(inode
);
241 wdata
->args
.count
= count
;
242 wdata
->args
.offset
= page_offset(page
) + wdata
->args
.pgbase
;
244 result
= NFS_PROTO(inode
)->write(wdata
);
247 /* Must mark the page invalid after I/O error */
248 ClearPageUptodate(page
);
251 if (result
< wdata
->args
.count
)
252 printk(KERN_WARNING
"NFS: short write, count=%u, result=%d\n",
253 wdata
->args
.count
, result
);
255 wdata
->args
.offset
+= result
;
256 wdata
->args
.pgbase
+= result
;
259 nfs_add_stats(inode
, NFSIOS_SERVERWRITTENBYTES
, result
);
261 /* Update file length */
262 nfs_grow_file(page
, offset
, written
);
263 /* Set the PG_uptodate flag? */
264 nfs_mark_uptodate(page
, offset
, written
);
267 ClearPageError(page
);
270 nfs_end_data_update(inode
);
271 end_page_writeback(page
);
272 nfs_writedata_free(wdata
);
273 return written
? written
: result
;
276 static int nfs_writepage_async(struct nfs_open_context
*ctx
,
277 struct inode
*inode
, struct page
*page
,
278 unsigned int offset
, unsigned int count
)
280 struct nfs_page
*req
;
282 req
= nfs_update_request(ctx
, inode
, page
, offset
, count
);
285 /* Update file length */
286 nfs_grow_file(page
, offset
, count
);
287 /* Set the PG_uptodate flag? */
288 nfs_mark_uptodate(page
, offset
, count
);
289 nfs_unlock_request(req
);
293 static int wb_priority(struct writeback_control
*wbc
)
295 if (wbc
->for_reclaim
)
296 return FLUSH_HIGHPRI
;
297 if (wbc
->for_kupdate
)
303 * Write an mmapped page to the server.
305 int nfs_writepage(struct page
*page
, struct writeback_control
*wbc
)
307 struct nfs_open_context
*ctx
;
308 struct inode
*inode
= page
->mapping
->host
;
309 unsigned long end_index
;
310 unsigned offset
= PAGE_CACHE_SIZE
;
311 loff_t i_size
= i_size_read(inode
);
312 int inode_referenced
= 0;
313 int priority
= wb_priority(wbc
);
316 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGE
);
317 nfs_add_stats(inode
, NFSIOS_WRITEPAGES
, 1);
320 * Note: We need to ensure that we have a reference to the inode
321 * if we are to do asynchronous writes. If not, waiting
322 * in nfs_wait_on_request() may deadlock with clear_inode().
324 * If igrab() fails here, then it is in any case safe to
325 * call nfs_wb_page(), since there will be no pending writes.
327 if (igrab(inode
) != 0)
328 inode_referenced
= 1;
329 end_index
= i_size
>> PAGE_CACHE_SHIFT
;
331 /* Ensure we've flushed out any previous writes */
332 nfs_wb_page_priority(inode
, page
, priority
);
335 if (page
->index
< end_index
)
337 /* things got complicated... */
338 offset
= i_size
& (PAGE_CACHE_SIZE
-1);
340 /* OK, are we completely out? */
341 err
= 0; /* potential race with truncate - ignore */
342 if (page
->index
>= end_index
+1 || !offset
)
345 ctx
= nfs_find_open_context(inode
, NULL
, FMODE_WRITE
);
351 if (!IS_SYNC(inode
) && inode_referenced
) {
352 err
= nfs_writepage_async(ctx
, inode
, page
, 0, offset
);
353 if (!wbc
->for_writepages
)
354 nfs_flush_inode(inode
, 0, 0, wb_priority(wbc
));
356 err
= nfs_writepage_sync(ctx
, inode
, page
, 0,
360 redirty_page_for_writepage(wbc
, page
);
365 put_nfs_open_context(ctx
);
368 if (inode_referenced
)
374 * Note: causes nfs_update_request() to block on the assumption
375 * that the writeback is generated due to memory pressure.
377 int nfs_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
379 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
380 struct inode
*inode
= mapping
->host
;
383 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGES
);
385 err
= generic_writepages(mapping
, wbc
);
388 while (test_and_set_bit(BDI_write_congested
, &bdi
->state
) != 0) {
389 if (wbc
->nonblocking
)
391 nfs_wait_on_write_congestion(mapping
, 0);
393 err
= nfs_flush_inode(inode
, 0, 0, wb_priority(wbc
));
396 nfs_add_stats(inode
, NFSIOS_WRITEPAGES
, err
);
397 wbc
->nr_to_write
-= err
;
398 if (!wbc
->nonblocking
&& wbc
->sync_mode
== WB_SYNC_ALL
) {
399 err
= nfs_wait_on_requests(inode
, 0, 0);
403 err
= nfs_commit_inode(inode
, wb_priority(wbc
));
405 wbc
->nr_to_write
-= err
;
409 clear_bit(BDI_write_congested
, &bdi
->state
);
410 wake_up_all(&nfs_write_congestion
);
415 * Insert a write request into an inode
417 static int nfs_inode_add_request(struct inode
*inode
, struct nfs_page
*req
)
419 struct nfs_inode
*nfsi
= NFS_I(inode
);
422 error
= radix_tree_insert(&nfsi
->nfs_page_tree
, req
->wb_index
, req
);
423 BUG_ON(error
== -EEXIST
);
428 nfs_begin_data_update(inode
);
429 if (nfs_have_delegation(inode
, FMODE_WRITE
))
432 SetPagePrivate(req
->wb_page
);
434 atomic_inc(&req
->wb_count
);
439 * Insert a write request into an inode
441 static void nfs_inode_remove_request(struct nfs_page
*req
)
443 struct inode
*inode
= req
->wb_context
->dentry
->d_inode
;
444 struct nfs_inode
*nfsi
= NFS_I(inode
);
446 BUG_ON (!NFS_WBACK_BUSY(req
));
448 spin_lock(&nfsi
->req_lock
);
449 ClearPagePrivate(req
->wb_page
);
450 radix_tree_delete(&nfsi
->nfs_page_tree
, req
->wb_index
);
453 spin_unlock(&nfsi
->req_lock
);
454 nfs_end_data_update(inode
);
457 spin_unlock(&nfsi
->req_lock
);
458 nfs_clear_request(req
);
459 nfs_release_request(req
);
465 static inline struct nfs_page
*
466 _nfs_find_request(struct inode
*inode
, unsigned long index
)
468 struct nfs_inode
*nfsi
= NFS_I(inode
);
469 struct nfs_page
*req
;
471 req
= (struct nfs_page
*)radix_tree_lookup(&nfsi
->nfs_page_tree
, index
);
473 atomic_inc(&req
->wb_count
);
477 static struct nfs_page
*
478 nfs_find_request(struct inode
*inode
, unsigned long index
)
480 struct nfs_page
*req
;
481 struct nfs_inode
*nfsi
= NFS_I(inode
);
483 spin_lock(&nfsi
->req_lock
);
484 req
= _nfs_find_request(inode
, index
);
485 spin_unlock(&nfsi
->req_lock
);
490 * Add a request to the inode's dirty list.
493 nfs_mark_request_dirty(struct nfs_page
*req
)
495 struct inode
*inode
= req
->wb_context
->dentry
->d_inode
;
496 struct nfs_inode
*nfsi
= NFS_I(inode
);
498 spin_lock(&nfsi
->req_lock
);
499 radix_tree_tag_set(&nfsi
->nfs_page_tree
,
500 req
->wb_index
, NFS_PAGE_TAG_DIRTY
);
501 nfs_list_add_request(req
, &nfsi
->dirty
);
503 spin_unlock(&nfsi
->req_lock
);
504 inc_page_state(nr_dirty
);
505 mark_inode_dirty(inode
);
509 * Check if a request is dirty
512 nfs_dirty_request(struct nfs_page
*req
)
514 struct nfs_inode
*nfsi
= NFS_I(req
->wb_context
->dentry
->d_inode
);
515 return !list_empty(&req
->wb_list
) && req
->wb_list_head
== &nfsi
->dirty
;
518 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
520 * Add a request to the inode's commit list.
523 nfs_mark_request_commit(struct nfs_page
*req
)
525 struct inode
*inode
= req
->wb_context
->dentry
->d_inode
;
526 struct nfs_inode
*nfsi
= NFS_I(inode
);
528 spin_lock(&nfsi
->req_lock
);
529 nfs_list_add_request(req
, &nfsi
->commit
);
531 spin_unlock(&nfsi
->req_lock
);
532 inc_page_state(nr_unstable
);
533 mark_inode_dirty(inode
);
538 * Wait for a request to complete.
540 * Interruptible by signals only if mounted with intr flag.
542 static int nfs_wait_on_requests_locked(struct inode
*inode
, unsigned long idx_start
, unsigned int npages
)
544 struct nfs_inode
*nfsi
= NFS_I(inode
);
545 struct nfs_page
*req
;
546 unsigned long idx_end
, next
;
547 unsigned int res
= 0;
553 idx_end
= idx_start
+ npages
- 1;
556 while (radix_tree_gang_lookup_tag(&nfsi
->nfs_page_tree
, (void **)&req
, next
, 1, NFS_PAGE_TAG_WRITEBACK
)) {
557 if (req
->wb_index
> idx_end
)
560 next
= req
->wb_index
+ 1;
561 BUG_ON(!NFS_WBACK_BUSY(req
));
563 atomic_inc(&req
->wb_count
);
564 spin_unlock(&nfsi
->req_lock
);
565 error
= nfs_wait_on_request(req
);
566 nfs_release_request(req
);
567 spin_lock(&nfsi
->req_lock
);
575 static int nfs_wait_on_requests(struct inode
*inode
, unsigned long idx_start
, unsigned int npages
)
577 struct nfs_inode
*nfsi
= NFS_I(inode
);
580 spin_lock(&nfsi
->req_lock
);
581 ret
= nfs_wait_on_requests_locked(inode
, idx_start
, npages
);
582 spin_unlock(&nfsi
->req_lock
);
587 * nfs_scan_dirty - Scan an inode for dirty requests
588 * @inode: NFS inode to scan
589 * @dst: destination list
590 * @idx_start: lower bound of page->index to scan.
591 * @npages: idx_start + npages sets the upper bound to scan.
593 * Moves requests from the inode's dirty page list.
594 * The requests are *not* checked to ensure that they form a contiguous set.
597 nfs_scan_dirty(struct inode
*inode
, struct list_head
*dst
, unsigned long idx_start
, unsigned int npages
)
599 struct nfs_inode
*nfsi
= NFS_I(inode
);
602 if (nfsi
->ndirty
!= 0) {
603 res
= nfs_scan_lock_dirty(nfsi
, dst
, idx_start
, npages
);
605 sub_page_state(nr_dirty
,res
);
606 if ((nfsi
->ndirty
== 0) != list_empty(&nfsi
->dirty
))
607 printk(KERN_ERR
"NFS: desynchronized value of nfs_i.ndirty.\n");
612 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
614 * nfs_scan_commit - Scan an inode for commit requests
615 * @inode: NFS inode to scan
616 * @dst: destination list
617 * @idx_start: lower bound of page->index to scan.
618 * @npages: idx_start + npages sets the upper bound to scan.
620 * Moves requests from the inode's 'commit' request list.
621 * The requests are *not* checked to ensure that they form a contiguous set.
624 nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, unsigned long idx_start
, unsigned int npages
)
626 struct nfs_inode
*nfsi
= NFS_I(inode
);
629 if (nfsi
->ncommit
!= 0) {
630 res
= nfs_scan_list(&nfsi
->commit
, dst
, idx_start
, npages
);
631 nfsi
->ncommit
-= res
;
632 if ((nfsi
->ncommit
== 0) != list_empty(&nfsi
->commit
))
633 printk(KERN_ERR
"NFS: desynchronized value of nfs_i.ncommit.\n");
638 static inline int nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, unsigned long idx_start
, unsigned int npages
)
644 static int nfs_wait_on_write_congestion(struct address_space
*mapping
, int intr
)
646 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
652 if (!bdi_write_congested(bdi
))
655 nfs_inc_stats(mapping
->host
, NFSIOS_CONGESTIONWAIT
);
658 struct rpc_clnt
*clnt
= NFS_CLIENT(mapping
->host
);
661 rpc_clnt_sigmask(clnt
, &oldset
);
662 prepare_to_wait(&nfs_write_congestion
, &wait
, TASK_INTERRUPTIBLE
);
663 if (bdi_write_congested(bdi
)) {
669 rpc_clnt_sigunmask(clnt
, &oldset
);
671 prepare_to_wait(&nfs_write_congestion
, &wait
, TASK_UNINTERRUPTIBLE
);
672 if (bdi_write_congested(bdi
))
675 finish_wait(&nfs_write_congestion
, &wait
);
681 * Try to update any existing write request, or create one if there is none.
682 * In order to match, the request's credentials must match those of
683 * the calling process.
685 * Note: Should always be called with the Page Lock held!
687 static struct nfs_page
* nfs_update_request(struct nfs_open_context
* ctx
,
688 struct inode
*inode
, struct page
*page
,
689 unsigned int offset
, unsigned int bytes
)
691 struct nfs_server
*server
= NFS_SERVER(inode
);
692 struct nfs_inode
*nfsi
= NFS_I(inode
);
693 struct nfs_page
*req
, *new = NULL
;
694 unsigned long rqend
, end
;
696 end
= offset
+ bytes
;
698 if (nfs_wait_on_write_congestion(page
->mapping
, server
->flags
& NFS_MOUNT_INTR
))
699 return ERR_PTR(-ERESTARTSYS
);
701 /* Loop over all inode entries and see if we find
702 * A request for the page we wish to update
704 spin_lock(&nfsi
->req_lock
);
705 req
= _nfs_find_request(inode
, page
->index
);
707 if (!nfs_lock_request_dontget(req
)) {
709 spin_unlock(&nfsi
->req_lock
);
710 error
= nfs_wait_on_request(req
);
711 nfs_release_request(req
);
714 nfs_release_request(new);
715 return ERR_PTR(error
);
719 spin_unlock(&nfsi
->req_lock
);
721 nfs_release_request(new);
727 nfs_lock_request_dontget(new);
728 error
= nfs_inode_add_request(inode
, new);
730 spin_unlock(&nfsi
->req_lock
);
731 nfs_unlock_request(new);
732 return ERR_PTR(error
);
734 spin_unlock(&nfsi
->req_lock
);
735 nfs_mark_request_dirty(new);
738 spin_unlock(&nfsi
->req_lock
);
740 new = nfs_create_request(ctx
, inode
, page
, offset
, bytes
);
745 /* We have a request for our page.
746 * If the creds don't match, or the
747 * page addresses don't match,
748 * tell the caller to wait on the conflicting
751 rqend
= req
->wb_offset
+ req
->wb_bytes
;
752 if (req
->wb_context
!= ctx
753 || req
->wb_page
!= page
754 || !nfs_dirty_request(req
)
755 || offset
> rqend
|| end
< req
->wb_offset
) {
756 nfs_unlock_request(req
);
757 return ERR_PTR(-EBUSY
);
760 /* Okay, the request matches. Update the region */
761 if (offset
< req
->wb_offset
) {
762 req
->wb_offset
= offset
;
763 req
->wb_pgbase
= offset
;
764 req
->wb_bytes
= rqend
- req
->wb_offset
;
768 req
->wb_bytes
= end
- req
->wb_offset
;
773 int nfs_flush_incompatible(struct file
*file
, struct page
*page
)
775 struct nfs_open_context
*ctx
= (struct nfs_open_context
*)file
->private_data
;
776 struct inode
*inode
= page
->mapping
->host
;
777 struct nfs_page
*req
;
780 * Look for a request corresponding to this page. If there
781 * is one, and it belongs to another file, we flush it out
782 * before we try to copy anything into the page. Do this
783 * due to the lack of an ACCESS-type call in NFSv2.
784 * Also do the same if we find a request from an existing
787 req
= nfs_find_request(inode
, page
->index
);
789 if (req
->wb_page
!= page
|| ctx
!= req
->wb_context
)
790 status
= nfs_wb_page(inode
, page
);
791 nfs_release_request(req
);
793 return (status
< 0) ? status
: 0;
797 * Update and possibly write a cached page of an NFS file.
799 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
800 * things with a page scheduled for an RPC call (e.g. invalidate it).
802 int nfs_updatepage(struct file
*file
, struct page
*page
,
803 unsigned int offset
, unsigned int count
)
805 struct nfs_open_context
*ctx
= (struct nfs_open_context
*)file
->private_data
;
806 struct inode
*inode
= page
->mapping
->host
;
807 struct nfs_page
*req
;
810 nfs_inc_stats(inode
, NFSIOS_VFSUPDATEPAGE
);
812 dprintk("NFS: nfs_updatepage(%s/%s %d@%Ld)\n",
813 file
->f_dentry
->d_parent
->d_name
.name
,
814 file
->f_dentry
->d_name
.name
, count
,
815 (long long)(page_offset(page
) +offset
));
817 if (IS_SYNC(inode
)) {
818 status
= nfs_writepage_sync(ctx
, inode
, page
, offset
, count
, 0);
820 if (offset
== 0 && status
== PAGE_CACHE_SIZE
)
821 SetPageUptodate(page
);
827 /* If we're not using byte range locks, and we know the page
828 * is entirely in cache, it may be more efficient to avoid
829 * fragmenting write requests.
831 if (PageUptodate(page
) && inode
->i_flock
== NULL
&& !(file
->f_mode
& O_SYNC
)) {
832 loff_t end_offs
= i_size_read(inode
) - 1;
833 unsigned long end_index
= end_offs
>> PAGE_CACHE_SHIFT
;
837 if (unlikely(end_offs
< 0)) {
839 } else if (page
->index
== end_index
) {
841 pglen
= (unsigned int)(end_offs
& (PAGE_CACHE_SIZE
-1)) + 1;
844 } else if (page
->index
< end_index
)
845 count
= PAGE_CACHE_SIZE
;
849 * Try to find an NFS request corresponding to this page
851 * If the existing request cannot be updated, we must flush
855 req
= nfs_update_request(ctx
, inode
, page
, offset
, count
);
856 status
= (IS_ERR(req
)) ? PTR_ERR(req
) : 0;
857 if (status
!= -EBUSY
)
859 /* Request could not be updated. Flush it out and try again */
860 status
= nfs_wb_page(inode
, page
);
861 } while (status
>= 0);
867 /* Update file length */
868 nfs_grow_file(page
, offset
, count
);
869 /* Set the PG_uptodate flag? */
870 nfs_mark_uptodate(page
, req
->wb_pgbase
, req
->wb_bytes
);
871 nfs_unlock_request(req
);
873 dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n",
874 status
, (long long)i_size_read(inode
));
876 ClearPageUptodate(page
);
880 static void nfs_writepage_release(struct nfs_page
*req
)
882 end_page_writeback(req
->wb_page
);
884 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
885 if (!PageError(req
->wb_page
)) {
886 if (NFS_NEED_RESCHED(req
)) {
887 nfs_mark_request_dirty(req
);
889 } else if (NFS_NEED_COMMIT(req
)) {
890 nfs_mark_request_commit(req
);
894 nfs_inode_remove_request(req
);
897 nfs_clear_commit(req
);
898 nfs_clear_reschedule(req
);
900 nfs_inode_remove_request(req
);
902 nfs_clear_page_writeback(req
);
905 static inline int flush_task_priority(int how
)
907 switch (how
& (FLUSH_HIGHPRI
|FLUSH_LOWPRI
)) {
909 return RPC_PRIORITY_HIGH
;
911 return RPC_PRIORITY_LOW
;
913 return RPC_PRIORITY_NORMAL
;
917 * Set up the argument/result storage required for the RPC call.
919 static void nfs_write_rpcsetup(struct nfs_page
*req
,
920 struct nfs_write_data
*data
,
921 const struct rpc_call_ops
*call_ops
,
922 unsigned int count
, unsigned int offset
,
928 /* Set up the RPC argument and reply structs
929 * NB: take care not to mess about with data->commit et al. */
932 data
->inode
= inode
= req
->wb_context
->dentry
->d_inode
;
933 data
->cred
= req
->wb_context
->cred
;
935 data
->args
.fh
= NFS_FH(inode
);
936 data
->args
.offset
= req_offset(req
) + offset
;
937 data
->args
.pgbase
= req
->wb_pgbase
+ offset
;
938 data
->args
.pages
= data
->pagevec
;
939 data
->args
.count
= count
;
940 data
->args
.context
= req
->wb_context
;
942 data
->res
.fattr
= &data
->fattr
;
943 data
->res
.count
= count
;
944 data
->res
.verf
= &data
->verf
;
945 nfs_fattr_init(&data
->fattr
);
947 /* Set up the initial task struct. */
948 flags
= (how
& FLUSH_SYNC
) ? 0 : RPC_TASK_ASYNC
;
949 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), flags
, call_ops
, data
);
950 NFS_PROTO(inode
)->write_setup(data
, how
);
952 data
->task
.tk_priority
= flush_task_priority(how
);
953 data
->task
.tk_cookie
= (unsigned long)inode
;
955 dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
958 (long long)NFS_FILEID(inode
),
960 (unsigned long long)data
->args
.offset
);
963 static void nfs_execute_write(struct nfs_write_data
*data
)
965 struct rpc_clnt
*clnt
= NFS_CLIENT(data
->inode
);
968 rpc_clnt_sigmask(clnt
, &oldset
);
970 rpc_execute(&data
->task
);
972 rpc_clnt_sigunmask(clnt
, &oldset
);
976 * Generate multiple small requests to write out a single
977 * contiguous dirty area on one page.
979 static int nfs_flush_multi(struct inode
*inode
, struct list_head
*head
, int how
)
981 struct nfs_page
*req
= nfs_list_entry(head
->next
);
982 struct page
*page
= req
->wb_page
;
983 struct nfs_write_data
*data
;
984 unsigned int wsize
= NFS_SERVER(inode
)->wsize
;
985 unsigned int nbytes
, offset
;
989 nfs_list_remove_request(req
);
991 nbytes
= req
->wb_bytes
;
993 data
= nfs_writedata_alloc(1);
996 list_add(&data
->pages
, &list
);
1002 atomic_set(&req
->wb_complete
, requests
);
1004 ClearPageError(page
);
1005 set_page_writeback(page
);
1007 nbytes
= req
->wb_bytes
;
1009 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
1010 list_del_init(&data
->pages
);
1012 data
->pagevec
[0] = page
;
1014 if (nbytes
> wsize
) {
1015 nfs_write_rpcsetup(req
, data
, &nfs_write_partial_ops
,
1016 wsize
, offset
, how
);
1020 nfs_write_rpcsetup(req
, data
, &nfs_write_partial_ops
,
1021 nbytes
, offset
, how
);
1024 nfs_execute_write(data
);
1025 } while (nbytes
!= 0);
1030 while (!list_empty(&list
)) {
1031 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
1032 list_del(&data
->pages
);
1033 nfs_writedata_free(data
);
1035 nfs_mark_request_dirty(req
);
1036 nfs_clear_page_writeback(req
);
1041 * Create an RPC task for the given write request and kick it.
1042 * The page must have been locked by the caller.
1044 * It may happen that the page we're passed is not marked dirty.
1045 * This is the case if nfs_updatepage detects a conflicting request
1046 * that has been written but not committed.
1048 static int nfs_flush_one(struct inode
*inode
, struct list_head
*head
, int how
)
1050 struct nfs_page
*req
;
1051 struct page
**pages
;
1052 struct nfs_write_data
*data
;
1055 data
= nfs_writedata_alloc(NFS_SERVER(inode
)->wpages
);
1059 pages
= data
->pagevec
;
1061 while (!list_empty(head
)) {
1062 req
= nfs_list_entry(head
->next
);
1063 nfs_list_remove_request(req
);
1064 nfs_list_add_request(req
, &data
->pages
);
1065 ClearPageError(req
->wb_page
);
1066 set_page_writeback(req
->wb_page
);
1067 *pages
++ = req
->wb_page
;
1068 count
+= req
->wb_bytes
;
1070 req
= nfs_list_entry(data
->pages
.next
);
1072 /* Set up the argument struct */
1073 nfs_write_rpcsetup(req
, data
, &nfs_write_full_ops
, count
, 0, how
);
1075 nfs_execute_write(data
);
1078 while (!list_empty(head
)) {
1079 struct nfs_page
*req
= nfs_list_entry(head
->next
);
1080 nfs_list_remove_request(req
);
1081 nfs_mark_request_dirty(req
);
1082 nfs_clear_page_writeback(req
);
1087 static int nfs_flush_list(struct inode
*inode
, struct list_head
*head
, int npages
, int how
)
1089 LIST_HEAD(one_request
);
1090 int (*flush_one
)(struct inode
*, struct list_head
*, int);
1091 struct nfs_page
*req
;
1092 int wpages
= NFS_SERVER(inode
)->wpages
;
1093 int wsize
= NFS_SERVER(inode
)->wsize
;
1096 flush_one
= nfs_flush_one
;
1097 if (wsize
< PAGE_CACHE_SIZE
)
1098 flush_one
= nfs_flush_multi
;
1099 /* For single writes, FLUSH_STABLE is more efficient */
1100 if (npages
<= wpages
&& npages
== NFS_I(inode
)->npages
1101 && nfs_list_entry(head
->next
)->wb_bytes
<= wsize
)
1102 how
|= FLUSH_STABLE
;
1105 nfs_coalesce_requests(head
, &one_request
, wpages
);
1106 req
= nfs_list_entry(one_request
.next
);
1107 error
= flush_one(inode
, &one_request
, how
);
1110 } while (!list_empty(head
));
1113 while (!list_empty(head
)) {
1114 req
= nfs_list_entry(head
->next
);
1115 nfs_list_remove_request(req
);
1116 nfs_mark_request_dirty(req
);
1117 nfs_clear_page_writeback(req
);
1123 * Handle a write reply that flushed part of a page.
1125 static void nfs_writeback_done_partial(struct rpc_task
*task
, void *calldata
)
1127 struct nfs_write_data
*data
= calldata
;
1128 struct nfs_page
*req
= data
->req
;
1129 struct page
*page
= req
->wb_page
;
1131 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1132 req
->wb_context
->dentry
->d_inode
->i_sb
->s_id
,
1133 (long long)NFS_FILEID(req
->wb_context
->dentry
->d_inode
),
1135 (long long)req_offset(req
));
1137 if (nfs_writeback_done(task
, data
) != 0)
1140 if (task
->tk_status
< 0) {
1141 ClearPageUptodate(page
);
1143 req
->wb_context
->error
= task
->tk_status
;
1144 dprintk(", error = %d\n", task
->tk_status
);
1146 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1147 if (data
->verf
.committed
< NFS_FILE_SYNC
) {
1148 if (!NFS_NEED_COMMIT(req
)) {
1149 nfs_defer_commit(req
);
1150 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1151 dprintk(" defer commit\n");
1152 } else if (memcmp(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
))) {
1153 nfs_defer_reschedule(req
);
1154 dprintk(" server reboot detected\n");
1161 if (atomic_dec_and_test(&req
->wb_complete
))
1162 nfs_writepage_release(req
);
1165 static const struct rpc_call_ops nfs_write_partial_ops
= {
1166 .rpc_call_done
= nfs_writeback_done_partial
,
1167 .rpc_release
= nfs_writedata_release
,
1171 * Handle a write reply that flushes a whole page.
1173 * FIXME: There is an inherent race with invalidate_inode_pages and
1174 * writebacks since the page->count is kept > 1 for as long
1175 * as the page has a write request pending.
1177 static void nfs_writeback_done_full(struct rpc_task
*task
, void *calldata
)
1179 struct nfs_write_data
*data
= calldata
;
1180 struct nfs_page
*req
;
1183 if (nfs_writeback_done(task
, data
) != 0)
1186 /* Update attributes as result of writeback. */
1187 while (!list_empty(&data
->pages
)) {
1188 req
= nfs_list_entry(data
->pages
.next
);
1189 nfs_list_remove_request(req
);
1190 page
= req
->wb_page
;
1192 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1193 req
->wb_context
->dentry
->d_inode
->i_sb
->s_id
,
1194 (long long)NFS_FILEID(req
->wb_context
->dentry
->d_inode
),
1196 (long long)req_offset(req
));
1198 if (task
->tk_status
< 0) {
1199 ClearPageUptodate(page
);
1201 req
->wb_context
->error
= task
->tk_status
;
1202 end_page_writeback(page
);
1203 nfs_inode_remove_request(req
);
1204 dprintk(", error = %d\n", task
->tk_status
);
1207 end_page_writeback(page
);
1209 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1210 if (data
->args
.stable
!= NFS_UNSTABLE
|| data
->verf
.committed
== NFS_FILE_SYNC
) {
1211 nfs_inode_remove_request(req
);
1215 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1216 nfs_mark_request_commit(req
);
1217 dprintk(" marked for commit\n");
1219 nfs_inode_remove_request(req
);
1222 nfs_clear_page_writeback(req
);
1226 static const struct rpc_call_ops nfs_write_full_ops
= {
1227 .rpc_call_done
= nfs_writeback_done_full
,
1228 .rpc_release
= nfs_writedata_release
,
1233 * This function is called when the WRITE call is complete.
1235 int nfs_writeback_done(struct rpc_task
*task
, struct nfs_write_data
*data
)
1237 struct nfs_writeargs
*argp
= &data
->args
;
1238 struct nfs_writeres
*resp
= &data
->res
;
1241 dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1242 task
->tk_pid
, task
->tk_status
);
1244 /* Call the NFS version-specific code */
1245 status
= NFS_PROTO(data
->inode
)->write_done(task
, data
);
1248 nfs_add_stats(data
->inode
, NFSIOS_SERVERWRITTENBYTES
, resp
->count
);
1250 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1251 if (resp
->verf
->committed
< argp
->stable
&& task
->tk_status
>= 0) {
1252 /* We tried a write call, but the server did not
1253 * commit data to stable storage even though we
1255 * Note: There is a known bug in Tru64 < 5.0 in which
1256 * the server reports NFS_DATA_SYNC, but performs
1257 * NFS_FILE_SYNC. We therefore implement this checking
1258 * as a dprintk() in order to avoid filling syslog.
1260 static unsigned long complain
;
1262 if (time_before(complain
, jiffies
)) {
1263 dprintk("NFS: faulty NFS server %s:"
1264 " (committed = %d) != (stable = %d)\n",
1265 NFS_SERVER(data
->inode
)->hostname
,
1266 resp
->verf
->committed
, argp
->stable
);
1267 complain
= jiffies
+ 300 * HZ
;
1271 /* Is this a short write? */
1272 if (task
->tk_status
>= 0 && resp
->count
< argp
->count
) {
1273 static unsigned long complain
;
1275 nfs_inc_stats(data
->inode
, NFSIOS_SHORTWRITE
);
1277 /* Has the server at least made some progress? */
1278 if (resp
->count
!= 0) {
1279 /* Was this an NFSv2 write or an NFSv3 stable write? */
1280 if (resp
->verf
->committed
!= NFS_UNSTABLE
) {
1281 /* Resend from where the server left off */
1282 argp
->offset
+= resp
->count
;
1283 argp
->pgbase
+= resp
->count
;
1284 argp
->count
-= resp
->count
;
1286 /* Resend as a stable write in order to avoid
1287 * headaches in the case of a server crash.
1289 argp
->stable
= NFS_FILE_SYNC
;
1291 rpc_restart_call(task
);
1294 if (time_before(complain
, jiffies
)) {
1296 "NFS: Server wrote zero bytes, expected %u.\n",
1298 complain
= jiffies
+ 300 * HZ
;
1300 /* Can't do anything about it except throw an error. */
1301 task
->tk_status
= -EIO
;
1307 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1308 void nfs_commit_release(void *wdata
)
1310 nfs_commit_free(wdata
);
1314 * Set up the argument/result storage required for the RPC call.
1316 static void nfs_commit_rpcsetup(struct list_head
*head
,
1317 struct nfs_write_data
*data
,
1320 struct nfs_page
*first
;
1321 struct inode
*inode
;
1324 /* Set up the RPC argument and reply structs
1325 * NB: take care not to mess about with data->commit et al. */
1327 list_splice_init(head
, &data
->pages
);
1328 first
= nfs_list_entry(data
->pages
.next
);
1329 inode
= first
->wb_context
->dentry
->d_inode
;
1331 data
->inode
= inode
;
1332 data
->cred
= first
->wb_context
->cred
;
1334 data
->args
.fh
= NFS_FH(data
->inode
);
1335 /* Note: we always request a commit of the entire inode */
1336 data
->args
.offset
= 0;
1337 data
->args
.count
= 0;
1338 data
->res
.count
= 0;
1339 data
->res
.fattr
= &data
->fattr
;
1340 data
->res
.verf
= &data
->verf
;
1341 nfs_fattr_init(&data
->fattr
);
1343 /* Set up the initial task struct. */
1344 flags
= (how
& FLUSH_SYNC
) ? 0 : RPC_TASK_ASYNC
;
1345 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), flags
, &nfs_commit_ops
, data
);
1346 NFS_PROTO(inode
)->commit_setup(data
, how
);
1348 data
->task
.tk_priority
= flush_task_priority(how
);
1349 data
->task
.tk_cookie
= (unsigned long)inode
;
1351 dprintk("NFS: %4d initiated commit call\n", data
->task
.tk_pid
);
1355 * Commit dirty pages
1358 nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
)
1360 struct nfs_write_data
*data
;
1361 struct nfs_page
*req
;
1363 data
= nfs_commit_alloc(NFS_SERVER(inode
)->wpages
);
1368 /* Set up the argument struct */
1369 nfs_commit_rpcsetup(head
, data
, how
);
1371 nfs_execute_write(data
);
1374 while (!list_empty(head
)) {
1375 req
= nfs_list_entry(head
->next
);
1376 nfs_list_remove_request(req
);
1377 nfs_mark_request_commit(req
);
1378 nfs_clear_page_writeback(req
);
1384 * COMMIT call returned
1386 static void nfs_commit_done(struct rpc_task
*task
, void *calldata
)
1388 struct nfs_write_data
*data
= calldata
;
1389 struct nfs_page
*req
;
1392 dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1393 task
->tk_pid
, task
->tk_status
);
1395 /* Call the NFS version-specific code */
1396 if (NFS_PROTO(data
->inode
)->commit_done(task
, data
) != 0)
1399 while (!list_empty(&data
->pages
)) {
1400 req
= nfs_list_entry(data
->pages
.next
);
1401 nfs_list_remove_request(req
);
1403 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1404 req
->wb_context
->dentry
->d_inode
->i_sb
->s_id
,
1405 (long long)NFS_FILEID(req
->wb_context
->dentry
->d_inode
),
1407 (long long)req_offset(req
));
1408 if (task
->tk_status
< 0) {
1409 req
->wb_context
->error
= task
->tk_status
;
1410 nfs_inode_remove_request(req
);
1411 dprintk(", error = %d\n", task
->tk_status
);
1415 /* Okay, COMMIT succeeded, apparently. Check the verifier
1416 * returned by the server against all stored verfs. */
1417 if (!memcmp(req
->wb_verf
.verifier
, data
->verf
.verifier
, sizeof(data
->verf
.verifier
))) {
1418 /* We have a match */
1419 nfs_inode_remove_request(req
);
1423 /* We have a mismatch. Write the page again */
1424 dprintk(" mismatch\n");
1425 nfs_mark_request_dirty(req
);
1427 nfs_clear_page_writeback(req
);
1430 sub_page_state(nr_unstable
,res
);
1433 static const struct rpc_call_ops nfs_commit_ops
= {
1434 .rpc_call_done
= nfs_commit_done
,
1435 .rpc_release
= nfs_commit_release
,
1438 static inline int nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
)
1444 static int nfs_flush_inode(struct inode
*inode
, unsigned long idx_start
,
1445 unsigned int npages
, int how
)
1447 struct nfs_inode
*nfsi
= NFS_I(inode
);
1451 spin_lock(&nfsi
->req_lock
);
1452 res
= nfs_scan_dirty(inode
, &head
, idx_start
, npages
);
1453 spin_unlock(&nfsi
->req_lock
);
1455 int error
= nfs_flush_list(inode
, &head
, res
, how
);
1462 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1463 int nfs_commit_inode(struct inode
*inode
, int how
)
1465 struct nfs_inode
*nfsi
= NFS_I(inode
);
1469 spin_lock(&nfsi
->req_lock
);
1470 res
= nfs_scan_commit(inode
, &head
, 0, 0);
1471 spin_unlock(&nfsi
->req_lock
);
1473 int error
= nfs_commit_list(inode
, &head
, how
);
1481 int nfs_sync_inode_wait(struct inode
*inode
, unsigned long idx_start
,
1482 unsigned int npages
, int how
)
1484 struct nfs_inode
*nfsi
= NFS_I(inode
);
1486 int nocommit
= how
& FLUSH_NOCOMMIT
;
1489 how
&= ~FLUSH_NOCOMMIT
;
1490 spin_lock(&nfsi
->req_lock
);
1492 ret
= nfs_wait_on_requests_locked(inode
, idx_start
, npages
);
1495 pages
= nfs_scan_dirty(inode
, &head
, idx_start
, npages
);
1497 spin_unlock(&nfsi
->req_lock
);
1498 ret
= nfs_flush_list(inode
, &head
, pages
, how
);
1499 spin_lock(&nfsi
->req_lock
);
1504 pages
= nfs_scan_commit(inode
, &head
, 0, 0);
1507 spin_unlock(&nfsi
->req_lock
);
1508 ret
= nfs_commit_list(inode
, &head
, how
);
1509 spin_lock(&nfsi
->req_lock
);
1511 spin_unlock(&nfsi
->req_lock
);
1515 int nfs_init_writepagecache(void)
1517 nfs_wdata_cachep
= kmem_cache_create("nfs_write_data",
1518 sizeof(struct nfs_write_data
),
1519 0, SLAB_HWCACHE_ALIGN
,
1521 if (nfs_wdata_cachep
== NULL
)
1524 nfs_wdata_mempool
= mempool_create_slab_pool(MIN_POOL_WRITE
,
1526 if (nfs_wdata_mempool
== NULL
)
1529 nfs_commit_mempool
= mempool_create_slab_pool(MIN_POOL_COMMIT
,
1531 if (nfs_commit_mempool
== NULL
)
1537 void nfs_destroy_writepagecache(void)
1539 mempool_destroy(nfs_commit_mempool
);
1540 mempool_destroy(nfs_wdata_mempool
);
1541 if (kmem_cache_destroy(nfs_wdata_cachep
))
1542 printk(KERN_INFO
"nfs_write_data: not all structures were freed\n");