4 * Write file data over NFS.
6 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
9 #include <linux/types.h>
10 #include <linux/slab.h>
12 #include <linux/pagemap.h>
13 #include <linux/file.h>
14 #include <linux/writeback.h>
15 #include <linux/swap.h>
16 #include <linux/migrate.h>
18 #include <linux/sunrpc/clnt.h>
19 #include <linux/nfs_fs.h>
20 #include <linux/nfs_mount.h>
21 #include <linux/nfs_page.h>
22 #include <linux/backing-dev.h>
24 #include <asm/uaccess.h>
26 #include "delegation.h"
32 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
34 #define MIN_POOL_WRITE (32)
35 #define MIN_POOL_COMMIT (4)
38 * Local function declarations
40 static void nfs_pageio_init_write(struct nfs_pageio_descriptor
*desc
,
41 struct inode
*inode
, int ioflags
);
42 static void nfs_redirty_request(struct nfs_page
*req
);
43 static const struct rpc_call_ops nfs_write_partial_ops
;
44 static const struct rpc_call_ops nfs_write_full_ops
;
45 static const struct rpc_call_ops nfs_commit_ops
;
47 static struct kmem_cache
*nfs_wdata_cachep
;
48 static mempool_t
*nfs_wdata_mempool
;
49 static mempool_t
*nfs_commit_mempool
;
51 struct nfs_write_data
*nfs_commitdata_alloc(void)
53 struct nfs_write_data
*p
= mempool_alloc(nfs_commit_mempool
, GFP_NOFS
);
56 memset(p
, 0, sizeof(*p
));
57 INIT_LIST_HEAD(&p
->pages
);
58 p
->res
.seq_res
.sr_slotid
= NFS4_MAX_SLOT_TABLE
;
63 void nfs_commit_free(struct nfs_write_data
*p
)
65 if (p
&& (p
->pagevec
!= &p
->page_array
[0]))
67 mempool_free(p
, nfs_commit_mempool
);
70 struct nfs_write_data
*nfs_writedata_alloc(unsigned int pagecount
)
72 struct nfs_write_data
*p
= mempool_alloc(nfs_wdata_mempool
, GFP_NOFS
);
75 memset(p
, 0, sizeof(*p
));
76 INIT_LIST_HEAD(&p
->pages
);
77 p
->npages
= pagecount
;
78 p
->res
.seq_res
.sr_slotid
= NFS4_MAX_SLOT_TABLE
;
79 if (pagecount
<= ARRAY_SIZE(p
->page_array
))
80 p
->pagevec
= p
->page_array
;
82 p
->pagevec
= kcalloc(pagecount
, sizeof(struct page
*), GFP_NOFS
);
84 mempool_free(p
, nfs_wdata_mempool
);
92 void nfs_writedata_free(struct nfs_write_data
*p
)
94 if (p
&& (p
->pagevec
!= &p
->page_array
[0]))
96 mempool_free(p
, nfs_wdata_mempool
);
99 static void nfs_writedata_release(struct nfs_write_data
*wdata
)
101 put_nfs_open_context(wdata
->args
.context
);
102 nfs_writedata_free(wdata
);
105 static void nfs_context_set_write_error(struct nfs_open_context
*ctx
, int error
)
109 set_bit(NFS_CONTEXT_ERROR_WRITE
, &ctx
->flags
);
112 static struct nfs_page
*nfs_page_find_request_locked(struct page
*page
)
114 struct nfs_page
*req
= NULL
;
116 if (PagePrivate(page
)) {
117 req
= (struct nfs_page
*)page_private(page
);
119 kref_get(&req
->wb_kref
);
124 static struct nfs_page
*nfs_page_find_request(struct page
*page
)
126 struct inode
*inode
= page
->mapping
->host
;
127 struct nfs_page
*req
= NULL
;
129 spin_lock(&inode
->i_lock
);
130 req
= nfs_page_find_request_locked(page
);
131 spin_unlock(&inode
->i_lock
);
135 /* Adjust the file length if we're writing beyond the end */
136 static void nfs_grow_file(struct page
*page
, unsigned int offset
, unsigned int count
)
138 struct inode
*inode
= page
->mapping
->host
;
142 spin_lock(&inode
->i_lock
);
143 i_size
= i_size_read(inode
);
144 end_index
= (i_size
- 1) >> PAGE_CACHE_SHIFT
;
145 if (i_size
> 0 && page
->index
< end_index
)
147 end
= ((loff_t
)page
->index
<< PAGE_CACHE_SHIFT
) + ((loff_t
)offset
+count
);
150 i_size_write(inode
, end
);
151 nfs_inc_stats(inode
, NFSIOS_EXTENDWRITE
);
153 spin_unlock(&inode
->i_lock
);
156 /* A writeback failed: mark the page as bad, and invalidate the page cache */
157 static void nfs_set_pageerror(struct page
*page
)
160 nfs_zap_mapping(page
->mapping
->host
, page
->mapping
);
163 /* We can set the PG_uptodate flag if we see that a write request
164 * covers the full page.
166 static void nfs_mark_uptodate(struct page
*page
, unsigned int base
, unsigned int count
)
168 if (PageUptodate(page
))
172 if (count
!= nfs_page_length(page
))
174 SetPageUptodate(page
);
177 static int wb_priority(struct writeback_control
*wbc
)
179 if (wbc
->for_reclaim
)
180 return FLUSH_HIGHPRI
| FLUSH_STABLE
;
181 if (wbc
->for_kupdate
|| wbc
->for_background
)
187 * NFS congestion control
190 int nfs_congestion_kb
;
192 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
193 #define NFS_CONGESTION_OFF_THRESH \
194 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
196 static int nfs_set_page_writeback(struct page
*page
)
198 int ret
= test_set_page_writeback(page
);
201 struct inode
*inode
= page
->mapping
->host
;
202 struct nfs_server
*nfss
= NFS_SERVER(inode
);
204 page_cache_get(page
);
205 if (atomic_long_inc_return(&nfss
->writeback
) >
206 NFS_CONGESTION_ON_THRESH
) {
207 set_bdi_congested(&nfss
->backing_dev_info
,
214 static void nfs_end_page_writeback(struct page
*page
)
216 struct inode
*inode
= page
->mapping
->host
;
217 struct nfs_server
*nfss
= NFS_SERVER(inode
);
219 end_page_writeback(page
);
220 page_cache_release(page
);
221 if (atomic_long_dec_return(&nfss
->writeback
) < NFS_CONGESTION_OFF_THRESH
)
222 clear_bdi_congested(&nfss
->backing_dev_info
, BLK_RW_ASYNC
);
225 static struct nfs_page
*nfs_find_and_lock_request(struct page
*page
)
227 struct inode
*inode
= page
->mapping
->host
;
228 struct nfs_page
*req
;
231 spin_lock(&inode
->i_lock
);
233 req
= nfs_page_find_request_locked(page
);
236 if (nfs_set_page_tag_locked(req
))
238 /* Note: If we hold the page lock, as is the case in nfs_writepage,
239 * then the call to nfs_set_page_tag_locked() will always
240 * succeed provided that someone hasn't already marked the
241 * request as dirty (in which case we don't care).
243 spin_unlock(&inode
->i_lock
);
244 ret
= nfs_wait_on_request(req
);
245 nfs_release_request(req
);
248 spin_lock(&inode
->i_lock
);
250 spin_unlock(&inode
->i_lock
);
255 * Find an associated nfs write request, and prepare to flush it out
256 * May return an error if the user signalled nfs_wait_on_request().
258 static int nfs_page_async_flush(struct nfs_pageio_descriptor
*pgio
,
261 struct nfs_page
*req
;
264 req
= nfs_find_and_lock_request(page
);
271 ret
= nfs_set_page_writeback(page
);
273 BUG_ON(test_bit(PG_CLEAN
, &req
->wb_flags
));
275 if (!nfs_pageio_add_request(pgio
, req
)) {
276 nfs_redirty_request(req
);
277 ret
= pgio
->pg_error
;
283 static int nfs_do_writepage(struct page
*page
, struct writeback_control
*wbc
, struct nfs_pageio_descriptor
*pgio
)
285 struct inode
*inode
= page
->mapping
->host
;
287 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGE
);
288 nfs_add_stats(inode
, NFSIOS_WRITEPAGES
, 1);
290 nfs_pageio_cond_complete(pgio
, page
->index
);
291 return nfs_page_async_flush(pgio
, page
);
295 * Write an mmapped page to the server.
297 static int nfs_writepage_locked(struct page
*page
, struct writeback_control
*wbc
)
299 struct nfs_pageio_descriptor pgio
;
302 nfs_pageio_init_write(&pgio
, page
->mapping
->host
, wb_priority(wbc
));
303 err
= nfs_do_writepage(page
, wbc
, &pgio
);
304 nfs_pageio_complete(&pgio
);
307 if (pgio
.pg_error
< 0)
308 return pgio
.pg_error
;
312 int nfs_writepage(struct page
*page
, struct writeback_control
*wbc
)
316 ret
= nfs_writepage_locked(page
, wbc
);
321 static int nfs_writepages_callback(struct page
*page
, struct writeback_control
*wbc
, void *data
)
325 ret
= nfs_do_writepage(page
, wbc
, data
);
330 int nfs_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
332 struct inode
*inode
= mapping
->host
;
333 unsigned long *bitlock
= &NFS_I(inode
)->flags
;
334 struct nfs_pageio_descriptor pgio
;
337 /* Stop dirtying of new pages while we sync */
338 err
= wait_on_bit_lock(bitlock
, NFS_INO_FLUSHING
,
339 nfs_wait_bit_killable
, TASK_KILLABLE
);
343 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGES
);
345 nfs_pageio_init_write(&pgio
, inode
, wb_priority(wbc
));
346 err
= write_cache_pages(mapping
, wbc
, nfs_writepages_callback
, &pgio
);
347 nfs_pageio_complete(&pgio
);
349 clear_bit_unlock(NFS_INO_FLUSHING
, bitlock
);
350 smp_mb__after_clear_bit();
351 wake_up_bit(bitlock
, NFS_INO_FLUSHING
);
364 * Insert a write request into an inode
366 static int nfs_inode_add_request(struct inode
*inode
, struct nfs_page
*req
)
368 struct nfs_inode
*nfsi
= NFS_I(inode
);
371 error
= radix_tree_preload(GFP_NOFS
);
375 /* Lock the request! */
376 nfs_lock_request_dontget(req
);
378 spin_lock(&inode
->i_lock
);
379 error
= radix_tree_insert(&nfsi
->nfs_page_tree
, req
->wb_index
, req
);
383 if (nfs_have_delegation(inode
, FMODE_WRITE
))
386 SetPagePrivate(req
->wb_page
);
387 set_page_private(req
->wb_page
, (unsigned long)req
);
389 kref_get(&req
->wb_kref
);
390 radix_tree_tag_set(&nfsi
->nfs_page_tree
, req
->wb_index
,
391 NFS_PAGE_TAG_LOCKED
);
392 spin_unlock(&inode
->i_lock
);
393 radix_tree_preload_end();
399 * Remove a write request from an inode
401 static void nfs_inode_remove_request(struct nfs_page
*req
)
403 struct inode
*inode
= req
->wb_context
->path
.dentry
->d_inode
;
404 struct nfs_inode
*nfsi
= NFS_I(inode
);
406 BUG_ON (!NFS_WBACK_BUSY(req
));
408 spin_lock(&inode
->i_lock
);
409 set_page_private(req
->wb_page
, 0);
410 ClearPagePrivate(req
->wb_page
);
411 radix_tree_delete(&nfsi
->nfs_page_tree
, req
->wb_index
);
414 spin_unlock(&inode
->i_lock
);
417 spin_unlock(&inode
->i_lock
);
418 nfs_clear_request(req
);
419 nfs_release_request(req
);
423 nfs_mark_request_dirty(struct nfs_page
*req
)
425 __set_page_dirty_nobuffers(req
->wb_page
);
426 __mark_inode_dirty(req
->wb_page
->mapping
->host
, I_DIRTY_DATASYNC
);
429 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
431 * Add a request to the inode's commit list.
434 nfs_mark_request_commit(struct nfs_page
*req
)
436 struct inode
*inode
= req
->wb_context
->path
.dentry
->d_inode
;
437 struct nfs_inode
*nfsi
= NFS_I(inode
);
439 spin_lock(&inode
->i_lock
);
440 set_bit(PG_CLEAN
, &(req
)->wb_flags
);
441 radix_tree_tag_set(&nfsi
->nfs_page_tree
,
443 NFS_PAGE_TAG_COMMIT
);
445 spin_unlock(&inode
->i_lock
);
446 inc_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
447 inc_bdi_stat(req
->wb_page
->mapping
->backing_dev_info
, BDI_RECLAIMABLE
);
448 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
452 nfs_clear_request_commit(struct nfs_page
*req
)
454 struct page
*page
= req
->wb_page
;
456 if (test_and_clear_bit(PG_CLEAN
, &(req
)->wb_flags
)) {
457 dec_zone_page_state(page
, NR_UNSTABLE_NFS
);
458 dec_bdi_stat(page
->mapping
->backing_dev_info
, BDI_RECLAIMABLE
);
465 int nfs_write_need_commit(struct nfs_write_data
*data
)
467 return data
->verf
.committed
!= NFS_FILE_SYNC
;
471 int nfs_reschedule_unstable_write(struct nfs_page
*req
)
473 if (test_and_clear_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
474 nfs_mark_request_commit(req
);
477 if (test_and_clear_bit(PG_NEED_RESCHED
, &req
->wb_flags
)) {
478 nfs_mark_request_dirty(req
);
485 nfs_mark_request_commit(struct nfs_page
*req
)
490 nfs_clear_request_commit(struct nfs_page
*req
)
496 int nfs_write_need_commit(struct nfs_write_data
*data
)
502 int nfs_reschedule_unstable_write(struct nfs_page
*req
)
508 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
510 nfs_need_commit(struct nfs_inode
*nfsi
)
512 return radix_tree_tagged(&nfsi
->nfs_page_tree
, NFS_PAGE_TAG_COMMIT
);
516 * nfs_scan_commit - Scan an inode for commit requests
517 * @inode: NFS inode to scan
518 * @dst: destination list
519 * @idx_start: lower bound of page->index to scan.
520 * @npages: idx_start + npages sets the upper bound to scan.
522 * Moves requests from the inode's 'commit' request list.
523 * The requests are *not* checked to ensure that they form a contiguous set.
526 nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, pgoff_t idx_start
, unsigned int npages
)
528 struct nfs_inode
*nfsi
= NFS_I(inode
);
531 if (!nfs_need_commit(nfsi
))
534 ret
= nfs_scan_list(nfsi
, dst
, idx_start
, npages
, NFS_PAGE_TAG_COMMIT
);
536 nfsi
->ncommit
-= ret
;
537 if (nfs_need_commit(NFS_I(inode
)))
538 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
542 static inline int nfs_need_commit(struct nfs_inode
*nfsi
)
547 static inline int nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, pgoff_t idx_start
, unsigned int npages
)
554 * Search for an existing write request, and attempt to update
555 * it to reflect a new dirty region on a given page.
557 * If the attempt fails, then the existing request is flushed out
560 static struct nfs_page
*nfs_try_to_update_request(struct inode
*inode
,
565 struct nfs_page
*req
;
570 if (!PagePrivate(page
))
573 end
= offset
+ bytes
;
574 spin_lock(&inode
->i_lock
);
577 req
= nfs_page_find_request_locked(page
);
581 rqend
= req
->wb_offset
+ req
->wb_bytes
;
583 * Tell the caller to flush out the request if
584 * the offsets are non-contiguous.
585 * Note: nfs_flush_incompatible() will already
586 * have flushed out requests having wrong owners.
589 || end
< req
->wb_offset
)
592 if (nfs_set_page_tag_locked(req
))
595 /* The request is locked, so wait and then retry */
596 spin_unlock(&inode
->i_lock
);
597 error
= nfs_wait_on_request(req
);
598 nfs_release_request(req
);
601 spin_lock(&inode
->i_lock
);
604 if (nfs_clear_request_commit(req
) &&
605 radix_tree_tag_clear(&NFS_I(inode
)->nfs_page_tree
,
606 req
->wb_index
, NFS_PAGE_TAG_COMMIT
) != NULL
)
607 NFS_I(inode
)->ncommit
--;
609 /* Okay, the request matches. Update the region */
610 if (offset
< req
->wb_offset
) {
611 req
->wb_offset
= offset
;
612 req
->wb_pgbase
= offset
;
615 req
->wb_bytes
= end
- req
->wb_offset
;
617 req
->wb_bytes
= rqend
- req
->wb_offset
;
619 spin_unlock(&inode
->i_lock
);
622 spin_unlock(&inode
->i_lock
);
623 nfs_release_request(req
);
624 error
= nfs_wb_page(inode
, page
);
626 return ERR_PTR(error
);
630 * Try to update an existing write request, or create one if there is none.
632 * Note: Should always be called with the Page Lock held to prevent races
633 * if we have to add a new request. Also assumes that the caller has
634 * already called nfs_flush_incompatible() if necessary.
636 static struct nfs_page
* nfs_setup_write_request(struct nfs_open_context
* ctx
,
637 struct page
*page
, unsigned int offset
, unsigned int bytes
)
639 struct inode
*inode
= page
->mapping
->host
;
640 struct nfs_page
*req
;
643 req
= nfs_try_to_update_request(inode
, page
, offset
, bytes
);
646 req
= nfs_create_request(ctx
, inode
, page
, offset
, bytes
);
649 error
= nfs_inode_add_request(inode
, req
);
651 nfs_release_request(req
);
652 req
= ERR_PTR(error
);
658 static int nfs_writepage_setup(struct nfs_open_context
*ctx
, struct page
*page
,
659 unsigned int offset
, unsigned int count
)
661 struct nfs_page
*req
;
663 req
= nfs_setup_write_request(ctx
, page
, offset
, count
);
666 nfs_mark_request_dirty(req
);
667 /* Update file length */
668 nfs_grow_file(page
, offset
, count
);
669 nfs_mark_uptodate(page
, req
->wb_pgbase
, req
->wb_bytes
);
670 nfs_mark_request_dirty(req
);
671 nfs_clear_page_tag_locked(req
);
675 int nfs_flush_incompatible(struct file
*file
, struct page
*page
)
677 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
678 struct nfs_page
*req
;
679 int do_flush
, status
;
681 * Look for a request corresponding to this page. If there
682 * is one, and it belongs to another file, we flush it out
683 * before we try to copy anything into the page. Do this
684 * due to the lack of an ACCESS-type call in NFSv2.
685 * Also do the same if we find a request from an existing
689 req
= nfs_page_find_request(page
);
692 do_flush
= req
->wb_page
!= page
|| req
->wb_context
!= ctx
;
693 nfs_release_request(req
);
696 status
= nfs_wb_page(page
->mapping
->host
, page
);
697 } while (status
== 0);
702 * If the page cache is marked as unsafe or invalid, then we can't rely on
703 * the PageUptodate() flag. In this case, we will need to turn off
704 * write optimisations that depend on the page contents being correct.
706 static int nfs_write_pageuptodate(struct page
*page
, struct inode
*inode
)
708 return PageUptodate(page
) &&
709 !(NFS_I(inode
)->cache_validity
& (NFS_INO_REVAL_PAGECACHE
|NFS_INO_INVALID_DATA
));
713 * Update and possibly write a cached page of an NFS file.
715 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
716 * things with a page scheduled for an RPC call (e.g. invalidate it).
718 int nfs_updatepage(struct file
*file
, struct page
*page
,
719 unsigned int offset
, unsigned int count
)
721 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
722 struct inode
*inode
= page
->mapping
->host
;
725 nfs_inc_stats(inode
, NFSIOS_VFSUPDATEPAGE
);
727 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
728 file
->f_path
.dentry
->d_parent
->d_name
.name
,
729 file
->f_path
.dentry
->d_name
.name
, count
,
730 (long long)(page_offset(page
) + offset
));
732 /* If we're not using byte range locks, and we know the page
733 * is up to date, it may be more efficient to extend the write
734 * to cover the entire page in order to avoid fragmentation
737 if (nfs_write_pageuptodate(page
, inode
) &&
738 inode
->i_flock
== NULL
&&
739 !(file
->f_flags
& O_DSYNC
)) {
740 count
= max(count
+ offset
, nfs_page_length(page
));
744 status
= nfs_writepage_setup(ctx
, page
, offset
, count
);
746 nfs_set_pageerror(page
);
748 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
749 status
, (long long)i_size_read(inode
));
753 static void nfs_writepage_release(struct nfs_page
*req
)
755 struct page
*page
= req
->wb_page
;
757 if (PageError(req
->wb_page
) || !nfs_reschedule_unstable_write(req
))
758 nfs_inode_remove_request(req
);
759 nfs_clear_page_tag_locked(req
);
760 nfs_end_page_writeback(page
);
763 static int flush_task_priority(int how
)
765 switch (how
& (FLUSH_HIGHPRI
|FLUSH_LOWPRI
)) {
767 return RPC_PRIORITY_HIGH
;
769 return RPC_PRIORITY_LOW
;
771 return RPC_PRIORITY_NORMAL
;
775 * Set up the argument/result storage required for the RPC call.
777 static int nfs_write_rpcsetup(struct nfs_page
*req
,
778 struct nfs_write_data
*data
,
779 const struct rpc_call_ops
*call_ops
,
780 unsigned int count
, unsigned int offset
,
783 struct inode
*inode
= req
->wb_context
->path
.dentry
->d_inode
;
784 int flags
= (how
& FLUSH_SYNC
) ? 0 : RPC_TASK_ASYNC
;
785 int priority
= flush_task_priority(how
);
786 struct rpc_task
*task
;
787 struct rpc_message msg
= {
788 .rpc_argp
= &data
->args
,
789 .rpc_resp
= &data
->res
,
790 .rpc_cred
= req
->wb_context
->cred
,
792 struct rpc_task_setup task_setup_data
= {
793 .rpc_client
= NFS_CLIENT(inode
),
796 .callback_ops
= call_ops
,
797 .callback_data
= data
,
798 .workqueue
= nfsiod_workqueue
,
800 .priority
= priority
,
803 /* Set up the RPC argument and reply structs
804 * NB: take care not to mess about with data->commit et al. */
807 data
->inode
= inode
= req
->wb_context
->path
.dentry
->d_inode
;
808 data
->cred
= msg
.rpc_cred
;
810 data
->args
.fh
= NFS_FH(inode
);
811 data
->args
.offset
= req_offset(req
) + offset
;
812 data
->args
.pgbase
= req
->wb_pgbase
+ offset
;
813 data
->args
.pages
= data
->pagevec
;
814 data
->args
.count
= count
;
815 data
->args
.context
= get_nfs_open_context(req
->wb_context
);
816 data
->args
.stable
= NFS_UNSTABLE
;
817 if (how
& FLUSH_STABLE
) {
818 data
->args
.stable
= NFS_DATA_SYNC
;
819 if (!nfs_need_commit(NFS_I(inode
)))
820 data
->args
.stable
= NFS_FILE_SYNC
;
823 data
->res
.fattr
= &data
->fattr
;
824 data
->res
.count
= count
;
825 data
->res
.verf
= &data
->verf
;
826 nfs_fattr_init(&data
->fattr
);
828 /* Set up the initial task struct. */
829 NFS_PROTO(inode
)->write_setup(data
, &msg
);
831 dprintk("NFS: %5u initiated write call "
832 "(req %s/%lld, %u bytes @ offset %llu)\n",
835 (long long)NFS_FILEID(inode
),
837 (unsigned long long)data
->args
.offset
);
839 task
= rpc_run_task(&task_setup_data
);
841 return PTR_ERR(task
);
846 /* If a nfs_flush_* function fails, it should remove reqs from @head and
847 * call this on each, which will prepare them to be retried on next
848 * writeback using standard nfs.
850 static void nfs_redirty_request(struct nfs_page
*req
)
852 struct page
*page
= req
->wb_page
;
854 nfs_mark_request_dirty(req
);
855 nfs_clear_page_tag_locked(req
);
856 nfs_end_page_writeback(page
);
860 * Generate multiple small requests to write out a single
861 * contiguous dirty area on one page.
863 static int nfs_flush_multi(struct inode
*inode
, struct list_head
*head
, unsigned int npages
, size_t count
, int how
)
865 struct nfs_page
*req
= nfs_list_entry(head
->next
);
866 struct page
*page
= req
->wb_page
;
867 struct nfs_write_data
*data
;
868 size_t wsize
= NFS_SERVER(inode
)->wsize
, nbytes
;
874 nfs_list_remove_request(req
);
878 size_t len
= min(nbytes
, wsize
);
880 data
= nfs_writedata_alloc(1);
883 list_add(&data
->pages
, &list
);
886 } while (nbytes
!= 0);
887 atomic_set(&req
->wb_complete
, requests
);
889 ClearPageError(page
);
895 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
896 list_del_init(&data
->pages
);
898 data
->pagevec
[0] = page
;
902 ret2
= nfs_write_rpcsetup(req
, data
, &nfs_write_partial_ops
,
908 } while (nbytes
!= 0);
913 while (!list_empty(&list
)) {
914 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
915 list_del(&data
->pages
);
916 nfs_writedata_release(data
);
918 nfs_redirty_request(req
);
923 * Create an RPC task for the given write request and kick it.
924 * The page must have been locked by the caller.
926 * It may happen that the page we're passed is not marked dirty.
927 * This is the case if nfs_updatepage detects a conflicting request
928 * that has been written but not committed.
930 static int nfs_flush_one(struct inode
*inode
, struct list_head
*head
, unsigned int npages
, size_t count
, int how
)
932 struct nfs_page
*req
;
934 struct nfs_write_data
*data
;
936 data
= nfs_writedata_alloc(npages
);
940 pages
= data
->pagevec
;
941 while (!list_empty(head
)) {
942 req
= nfs_list_entry(head
->next
);
943 nfs_list_remove_request(req
);
944 nfs_list_add_request(req
, &data
->pages
);
945 ClearPageError(req
->wb_page
);
946 *pages
++ = req
->wb_page
;
948 req
= nfs_list_entry(data
->pages
.next
);
950 /* Set up the argument struct */
951 return nfs_write_rpcsetup(req
, data
, &nfs_write_full_ops
, count
, 0, how
);
953 while (!list_empty(head
)) {
954 req
= nfs_list_entry(head
->next
);
955 nfs_list_remove_request(req
);
956 nfs_redirty_request(req
);
961 static void nfs_pageio_init_write(struct nfs_pageio_descriptor
*pgio
,
962 struct inode
*inode
, int ioflags
)
964 size_t wsize
= NFS_SERVER(inode
)->wsize
;
966 if (wsize
< PAGE_CACHE_SIZE
)
967 nfs_pageio_init(pgio
, inode
, nfs_flush_multi
, wsize
, ioflags
);
969 nfs_pageio_init(pgio
, inode
, nfs_flush_one
, wsize
, ioflags
);
973 * Handle a write reply that flushed part of a page.
975 static void nfs_writeback_done_partial(struct rpc_task
*task
, void *calldata
)
977 struct nfs_write_data
*data
= calldata
;
979 dprintk("NFS: %5u write(%s/%lld %d@%lld)",
981 data
->req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
983 NFS_FILEID(data
->req
->wb_context
->path
.dentry
->d_inode
),
984 data
->req
->wb_bytes
, (long long)req_offset(data
->req
));
986 nfs_writeback_done(task
, data
);
989 static void nfs_writeback_release_partial(void *calldata
)
991 struct nfs_write_data
*data
= calldata
;
992 struct nfs_page
*req
= data
->req
;
993 struct page
*page
= req
->wb_page
;
994 int status
= data
->task
.tk_status
;
997 nfs_set_pageerror(page
);
998 nfs_context_set_write_error(req
->wb_context
, status
);
999 dprintk(", error = %d\n", status
);
1003 if (nfs_write_need_commit(data
)) {
1004 struct inode
*inode
= page
->mapping
->host
;
1006 spin_lock(&inode
->i_lock
);
1007 if (test_bit(PG_NEED_RESCHED
, &req
->wb_flags
)) {
1008 /* Do nothing we need to resend the writes */
1009 } else if (!test_and_set_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
1010 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1011 dprintk(" defer commit\n");
1012 } else if (memcmp(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
))) {
1013 set_bit(PG_NEED_RESCHED
, &req
->wb_flags
);
1014 clear_bit(PG_NEED_COMMIT
, &req
->wb_flags
);
1015 dprintk(" server reboot detected\n");
1017 spin_unlock(&inode
->i_lock
);
1022 if (atomic_dec_and_test(&req
->wb_complete
))
1023 nfs_writepage_release(req
);
1024 nfs_writedata_release(calldata
);
1027 #if defined(CONFIG_NFS_V4_1)
1028 void nfs_write_prepare(struct rpc_task
*task
, void *calldata
)
1030 struct nfs_write_data
*data
= calldata
;
1031 struct nfs_client
*clp
= (NFS_SERVER(data
->inode
))->nfs_client
;
1033 if (nfs4_setup_sequence(clp
, &data
->args
.seq_args
,
1034 &data
->res
.seq_res
, 1, task
))
1036 rpc_call_start(task
);
1038 #endif /* CONFIG_NFS_V4_1 */
1040 static const struct rpc_call_ops nfs_write_partial_ops
= {
1041 #if defined(CONFIG_NFS_V4_1)
1042 .rpc_call_prepare
= nfs_write_prepare
,
1043 #endif /* CONFIG_NFS_V4_1 */
1044 .rpc_call_done
= nfs_writeback_done_partial
,
1045 .rpc_release
= nfs_writeback_release_partial
,
1049 * Handle a write reply that flushes a whole page.
1051 * FIXME: There is an inherent race with invalidate_inode_pages and
1052 * writebacks since the page->count is kept > 1 for as long
1053 * as the page has a write request pending.
1055 static void nfs_writeback_done_full(struct rpc_task
*task
, void *calldata
)
1057 struct nfs_write_data
*data
= calldata
;
1059 nfs_writeback_done(task
, data
);
1062 static void nfs_writeback_release_full(void *calldata
)
1064 struct nfs_write_data
*data
= calldata
;
1065 int status
= data
->task
.tk_status
;
1067 /* Update attributes as result of writeback. */
1068 while (!list_empty(&data
->pages
)) {
1069 struct nfs_page
*req
= nfs_list_entry(data
->pages
.next
);
1070 struct page
*page
= req
->wb_page
;
1072 nfs_list_remove_request(req
);
1074 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1076 req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
1077 (long long)NFS_FILEID(req
->wb_context
->path
.dentry
->d_inode
),
1079 (long long)req_offset(req
));
1082 nfs_set_pageerror(page
);
1083 nfs_context_set_write_error(req
->wb_context
, status
);
1084 dprintk(", error = %d\n", status
);
1085 goto remove_request
;
1088 if (nfs_write_need_commit(data
)) {
1089 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1090 nfs_mark_request_commit(req
);
1091 dprintk(" marked for commit\n");
1096 nfs_inode_remove_request(req
);
1098 nfs_clear_page_tag_locked(req
);
1099 nfs_end_page_writeback(page
);
1101 nfs_writedata_release(calldata
);
1104 static const struct rpc_call_ops nfs_write_full_ops
= {
1105 #if defined(CONFIG_NFS_V4_1)
1106 .rpc_call_prepare
= nfs_write_prepare
,
1107 #endif /* CONFIG_NFS_V4_1 */
1108 .rpc_call_done
= nfs_writeback_done_full
,
1109 .rpc_release
= nfs_writeback_release_full
,
1114 * This function is called when the WRITE call is complete.
1116 int nfs_writeback_done(struct rpc_task
*task
, struct nfs_write_data
*data
)
1118 struct nfs_writeargs
*argp
= &data
->args
;
1119 struct nfs_writeres
*resp
= &data
->res
;
1120 struct nfs_server
*server
= NFS_SERVER(data
->inode
);
1123 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1124 task
->tk_pid
, task
->tk_status
);
1127 * ->write_done will attempt to use post-op attributes to detect
1128 * conflicting writes by other clients. A strict interpretation
1129 * of close-to-open would allow us to continue caching even if
1130 * another writer had changed the file, but some applications
1131 * depend on tighter cache coherency when writing.
1133 status
= NFS_PROTO(data
->inode
)->write_done(task
, data
);
1136 nfs_add_stats(data
->inode
, NFSIOS_SERVERWRITTENBYTES
, resp
->count
);
1138 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1139 if (resp
->verf
->committed
< argp
->stable
&& task
->tk_status
>= 0) {
1140 /* We tried a write call, but the server did not
1141 * commit data to stable storage even though we
1143 * Note: There is a known bug in Tru64 < 5.0 in which
1144 * the server reports NFS_DATA_SYNC, but performs
1145 * NFS_FILE_SYNC. We therefore implement this checking
1146 * as a dprintk() in order to avoid filling syslog.
1148 static unsigned long complain
;
1150 if (time_before(complain
, jiffies
)) {
1151 dprintk("NFS: faulty NFS server %s:"
1152 " (committed = %d) != (stable = %d)\n",
1153 server
->nfs_client
->cl_hostname
,
1154 resp
->verf
->committed
, argp
->stable
);
1155 complain
= jiffies
+ 300 * HZ
;
1159 /* Is this a short write? */
1160 if (task
->tk_status
>= 0 && resp
->count
< argp
->count
) {
1161 static unsigned long complain
;
1163 nfs_inc_stats(data
->inode
, NFSIOS_SHORTWRITE
);
1165 /* Has the server at least made some progress? */
1166 if (resp
->count
!= 0) {
1167 /* Was this an NFSv2 write or an NFSv3 stable write? */
1168 if (resp
->verf
->committed
!= NFS_UNSTABLE
) {
1169 /* Resend from where the server left off */
1170 argp
->offset
+= resp
->count
;
1171 argp
->pgbase
+= resp
->count
;
1172 argp
->count
-= resp
->count
;
1174 /* Resend as a stable write in order to avoid
1175 * headaches in the case of a server crash.
1177 argp
->stable
= NFS_FILE_SYNC
;
1179 nfs_restart_rpc(task
, server
->nfs_client
);
1182 if (time_before(complain
, jiffies
)) {
1184 "NFS: Server wrote zero bytes, expected %u.\n",
1186 complain
= jiffies
+ 300 * HZ
;
1188 /* Can't do anything about it except throw an error. */
1189 task
->tk_status
= -EIO
;
1195 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1196 static void nfs_commitdata_release(void *data
)
1198 struct nfs_write_data
*wdata
= data
;
1200 put_nfs_open_context(wdata
->args
.context
);
1201 nfs_commit_free(wdata
);
1205 * Set up the argument/result storage required for the RPC call.
1207 static int nfs_commit_rpcsetup(struct list_head
*head
,
1208 struct nfs_write_data
*data
,
1211 struct nfs_page
*first
= nfs_list_entry(head
->next
);
1212 struct inode
*inode
= first
->wb_context
->path
.dentry
->d_inode
;
1213 int flags
= (how
& FLUSH_SYNC
) ? 0 : RPC_TASK_ASYNC
;
1214 int priority
= flush_task_priority(how
);
1215 struct rpc_task
*task
;
1216 struct rpc_message msg
= {
1217 .rpc_argp
= &data
->args
,
1218 .rpc_resp
= &data
->res
,
1219 .rpc_cred
= first
->wb_context
->cred
,
1221 struct rpc_task_setup task_setup_data
= {
1222 .task
= &data
->task
,
1223 .rpc_client
= NFS_CLIENT(inode
),
1224 .rpc_message
= &msg
,
1225 .callback_ops
= &nfs_commit_ops
,
1226 .callback_data
= data
,
1227 .workqueue
= nfsiod_workqueue
,
1229 .priority
= priority
,
1232 /* Set up the RPC argument and reply structs
1233 * NB: take care not to mess about with data->commit et al. */
1235 list_splice_init(head
, &data
->pages
);
1237 data
->inode
= inode
;
1238 data
->cred
= msg
.rpc_cred
;
1240 data
->args
.fh
= NFS_FH(data
->inode
);
1241 /* Note: we always request a commit of the entire inode */
1242 data
->args
.offset
= 0;
1243 data
->args
.count
= 0;
1244 data
->args
.context
= get_nfs_open_context(first
->wb_context
);
1245 data
->res
.count
= 0;
1246 data
->res
.fattr
= &data
->fattr
;
1247 data
->res
.verf
= &data
->verf
;
1248 nfs_fattr_init(&data
->fattr
);
1250 /* Set up the initial task struct. */
1251 NFS_PROTO(inode
)->commit_setup(data
, &msg
);
1253 dprintk("NFS: %5u initiated commit call\n", data
->task
.tk_pid
);
1255 task
= rpc_run_task(&task_setup_data
);
1257 return PTR_ERR(task
);
1263 * Commit dirty pages
1266 nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
)
1268 struct nfs_write_data
*data
;
1269 struct nfs_page
*req
;
1271 data
= nfs_commitdata_alloc();
1276 /* Set up the argument struct */
1277 return nfs_commit_rpcsetup(head
, data
, how
);
1279 while (!list_empty(head
)) {
1280 req
= nfs_list_entry(head
->next
);
1281 nfs_list_remove_request(req
);
1282 nfs_mark_request_commit(req
);
1283 dec_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
1284 dec_bdi_stat(req
->wb_page
->mapping
->backing_dev_info
,
1286 nfs_clear_page_tag_locked(req
);
1292 * COMMIT call returned
1294 static void nfs_commit_done(struct rpc_task
*task
, void *calldata
)
1296 struct nfs_write_data
*data
= calldata
;
1298 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1299 task
->tk_pid
, task
->tk_status
);
1301 /* Call the NFS version-specific code */
1302 if (NFS_PROTO(data
->inode
)->commit_done(task
, data
) != 0)
1306 static void nfs_commit_release(void *calldata
)
1308 struct nfs_write_data
*data
= calldata
;
1309 struct nfs_page
*req
;
1310 int status
= data
->task
.tk_status
;
1312 while (!list_empty(&data
->pages
)) {
1313 req
= nfs_list_entry(data
->pages
.next
);
1314 nfs_list_remove_request(req
);
1315 nfs_clear_request_commit(req
);
1317 dprintk("NFS: commit (%s/%lld %d@%lld)",
1318 req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
1319 (long long)NFS_FILEID(req
->wb_context
->path
.dentry
->d_inode
),
1321 (long long)req_offset(req
));
1323 nfs_context_set_write_error(req
->wb_context
, status
);
1324 nfs_inode_remove_request(req
);
1325 dprintk(", error = %d\n", status
);
1329 /* Okay, COMMIT succeeded, apparently. Check the verifier
1330 * returned by the server against all stored verfs. */
1331 if (!memcmp(req
->wb_verf
.verifier
, data
->verf
.verifier
, sizeof(data
->verf
.verifier
))) {
1332 /* We have a match */
1333 nfs_inode_remove_request(req
);
1337 /* We have a mismatch. Write the page again */
1338 dprintk(" mismatch\n");
1339 nfs_mark_request_dirty(req
);
1341 nfs_clear_page_tag_locked(req
);
1343 nfs_commitdata_release(calldata
);
1346 static const struct rpc_call_ops nfs_commit_ops
= {
1347 #if defined(CONFIG_NFS_V4_1)
1348 .rpc_call_prepare
= nfs_write_prepare
,
1349 #endif /* CONFIG_NFS_V4_1 */
1350 .rpc_call_done
= nfs_commit_done
,
1351 .rpc_release
= nfs_commit_release
,
1354 static int nfs_commit_inode(struct inode
*inode
, int how
)
1359 spin_lock(&inode
->i_lock
);
1360 res
= nfs_scan_commit(inode
, &head
, 0, 0);
1361 spin_unlock(&inode
->i_lock
);
1363 int error
= nfs_commit_list(inode
, &head
, how
);
1370 static int nfs_commit_unstable_pages(struct inode
*inode
, struct writeback_control
*wbc
)
1372 struct nfs_inode
*nfsi
= NFS_I(inode
);
1373 int flags
= FLUSH_SYNC
;
1376 /* Don't commit yet if this is a non-blocking flush and there are
1377 * lots of outstanding writes for this mapping.
1379 if (wbc
->sync_mode
== WB_SYNC_NONE
&&
1380 nfsi
->ncommit
<= (nfsi
->npages
>> 1))
1381 goto out_mark_dirty
;
1383 if (wbc
->nonblocking
|| wbc
->for_background
)
1385 ret
= nfs_commit_inode(inode
, flags
);
1387 if (wbc
->sync_mode
== WB_SYNC_NONE
) {
1388 if (ret
< wbc
->nr_to_write
)
1389 wbc
->nr_to_write
-= ret
;
1391 wbc
->nr_to_write
= 0;
1396 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
1400 static int nfs_commit_inode(struct inode
*inode
, int how
)
1405 static int nfs_commit_unstable_pages(struct inode
*inode
, struct writeback_control
*wbc
)
1411 int nfs_write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1413 return nfs_commit_unstable_pages(inode
, wbc
);
1417 * flush the inode to disk.
1419 int nfs_wb_all(struct inode
*inode
)
1421 struct writeback_control wbc
= {
1422 .sync_mode
= WB_SYNC_ALL
,
1423 .nr_to_write
= LONG_MAX
,
1425 .range_end
= LLONG_MAX
,
1428 return sync_inode(inode
, &wbc
);
1431 int nfs_wb_page_cancel(struct inode
*inode
, struct page
*page
)
1433 struct nfs_page
*req
;
1436 BUG_ON(!PageLocked(page
));
1438 req
= nfs_page_find_request(page
);
1441 if (nfs_lock_request_dontget(req
)) {
1442 nfs_inode_remove_request(req
);
1444 * In case nfs_inode_remove_request has marked the
1445 * page as being dirty
1447 cancel_dirty_page(page
, PAGE_CACHE_SIZE
);
1448 nfs_unlock_request(req
);
1451 ret
= nfs_wait_on_request(req
);
1452 nfs_release_request(req
);
1460 * Write back all requests on one page - we do this before reading it.
1462 int nfs_wb_page(struct inode
*inode
, struct page
*page
)
1464 loff_t range_start
= page_offset(page
);
1465 loff_t range_end
= range_start
+ (loff_t
)(PAGE_CACHE_SIZE
- 1);
1466 struct writeback_control wbc
= {
1467 .sync_mode
= WB_SYNC_ALL
,
1469 .range_start
= range_start
,
1470 .range_end
= range_end
,
1472 struct nfs_page
*req
;
1476 while(PagePrivate(page
)) {
1477 if (clear_page_dirty_for_io(page
)) {
1478 ret
= nfs_writepage_locked(page
, &wbc
);
1482 req
= nfs_find_and_lock_request(page
);
1489 need_commit
= test_bit(PG_CLEAN
, &req
->wb_flags
);
1490 nfs_clear_page_tag_locked(req
);
1492 ret
= nfs_commit_inode(inode
, FLUSH_SYNC
);
1502 #ifdef CONFIG_MIGRATION
1503 int nfs_migrate_page(struct address_space
*mapping
, struct page
*newpage
,
1506 struct nfs_page
*req
;
1509 nfs_fscache_release_page(page
, GFP_KERNEL
);
1511 req
= nfs_find_and_lock_request(page
);
1516 ret
= migrate_page(mapping
, newpage
, page
);
1521 page_cache_get(newpage
);
1522 spin_lock(&mapping
->host
->i_lock
);
1523 req
->wb_page
= newpage
;
1524 SetPagePrivate(newpage
);
1525 set_page_private(newpage
, (unsigned long)req
);
1526 ClearPagePrivate(page
);
1527 set_page_private(page
, 0);
1528 spin_unlock(&mapping
->host
->i_lock
);
1529 page_cache_release(page
);
1531 nfs_clear_page_tag_locked(req
);
1537 int __init
nfs_init_writepagecache(void)
1539 nfs_wdata_cachep
= kmem_cache_create("nfs_write_data",
1540 sizeof(struct nfs_write_data
),
1541 0, SLAB_HWCACHE_ALIGN
,
1543 if (nfs_wdata_cachep
== NULL
)
1546 nfs_wdata_mempool
= mempool_create_slab_pool(MIN_POOL_WRITE
,
1548 if (nfs_wdata_mempool
== NULL
)
1551 nfs_commit_mempool
= mempool_create_slab_pool(MIN_POOL_COMMIT
,
1553 if (nfs_commit_mempool
== NULL
)
1557 * NFS congestion size, scale with available memory.
1569 * This allows larger machines to have larger/more transfers.
1570 * Limit the default to 256M
1572 nfs_congestion_kb
= (16*int_sqrt(totalram_pages
)) << (PAGE_SHIFT
-10);
1573 if (nfs_congestion_kb
> 256*1024)
1574 nfs_congestion_kb
= 256*1024;
1579 void nfs_destroy_writepagecache(void)
1581 mempool_destroy(nfs_commit_mempool
);
1582 mempool_destroy(nfs_wdata_mempool
);
1583 kmem_cache_destroy(nfs_wdata_cachep
);