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
, bool nonblock
)
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
);
245 ret
= nfs_wait_on_request(req
);
248 nfs_release_request(req
);
251 spin_lock(&inode
->i_lock
);
253 spin_unlock(&inode
->i_lock
);
258 * Find an associated nfs write request, and prepare to flush it out
259 * May return an error if the user signalled nfs_wait_on_request().
261 static int nfs_page_async_flush(struct nfs_pageio_descriptor
*pgio
,
262 struct page
*page
, bool nonblock
)
264 struct nfs_page
*req
;
267 req
= nfs_find_and_lock_request(page
, nonblock
);
274 ret
= nfs_set_page_writeback(page
);
276 BUG_ON(test_bit(PG_CLEAN
, &req
->wb_flags
));
278 if (!nfs_pageio_add_request(pgio
, req
)) {
279 nfs_redirty_request(req
);
280 ret
= pgio
->pg_error
;
286 static int nfs_do_writepage(struct page
*page
, struct writeback_control
*wbc
, struct nfs_pageio_descriptor
*pgio
)
288 struct inode
*inode
= page
->mapping
->host
;
291 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGE
);
292 nfs_add_stats(inode
, NFSIOS_WRITEPAGES
, 1);
294 nfs_pageio_cond_complete(pgio
, page
->index
);
295 ret
= nfs_page_async_flush(pgio
, page
,
296 wbc
->sync_mode
== WB_SYNC_NONE
||
297 wbc
->nonblocking
!= 0);
298 if (ret
== -EAGAIN
) {
299 redirty_page_for_writepage(wbc
, page
);
306 * Write an mmapped page to the server.
308 static int nfs_writepage_locked(struct page
*page
, struct writeback_control
*wbc
)
310 struct nfs_pageio_descriptor pgio
;
313 nfs_pageio_init_write(&pgio
, page
->mapping
->host
, wb_priority(wbc
));
314 err
= nfs_do_writepage(page
, wbc
, &pgio
);
315 nfs_pageio_complete(&pgio
);
318 if (pgio
.pg_error
< 0)
319 return pgio
.pg_error
;
323 int nfs_writepage(struct page
*page
, struct writeback_control
*wbc
)
327 ret
= nfs_writepage_locked(page
, wbc
);
332 static int nfs_writepages_callback(struct page
*page
, struct writeback_control
*wbc
, void *data
)
336 ret
= nfs_do_writepage(page
, wbc
, data
);
341 int nfs_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
343 struct inode
*inode
= mapping
->host
;
344 unsigned long *bitlock
= &NFS_I(inode
)->flags
;
345 struct nfs_pageio_descriptor pgio
;
348 /* Stop dirtying of new pages while we sync */
349 err
= wait_on_bit_lock(bitlock
, NFS_INO_FLUSHING
,
350 nfs_wait_bit_killable
, TASK_KILLABLE
);
354 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGES
);
356 nfs_pageio_init_write(&pgio
, inode
, wb_priority(wbc
));
357 err
= write_cache_pages(mapping
, wbc
, nfs_writepages_callback
, &pgio
);
358 nfs_pageio_complete(&pgio
);
360 clear_bit_unlock(NFS_INO_FLUSHING
, bitlock
);
361 smp_mb__after_clear_bit();
362 wake_up_bit(bitlock
, NFS_INO_FLUSHING
);
375 * Insert a write request into an inode
377 static int nfs_inode_add_request(struct inode
*inode
, struct nfs_page
*req
)
379 struct nfs_inode
*nfsi
= NFS_I(inode
);
382 error
= radix_tree_preload(GFP_NOFS
);
386 /* Lock the request! */
387 nfs_lock_request_dontget(req
);
389 spin_lock(&inode
->i_lock
);
390 error
= radix_tree_insert(&nfsi
->nfs_page_tree
, req
->wb_index
, req
);
394 if (nfs_have_delegation(inode
, FMODE_WRITE
))
397 SetPagePrivate(req
->wb_page
);
398 set_page_private(req
->wb_page
, (unsigned long)req
);
400 kref_get(&req
->wb_kref
);
401 radix_tree_tag_set(&nfsi
->nfs_page_tree
, req
->wb_index
,
402 NFS_PAGE_TAG_LOCKED
);
403 spin_unlock(&inode
->i_lock
);
404 radix_tree_preload_end();
410 * Remove a write request from an inode
412 static void nfs_inode_remove_request(struct nfs_page
*req
)
414 struct inode
*inode
= req
->wb_context
->path
.dentry
->d_inode
;
415 struct nfs_inode
*nfsi
= NFS_I(inode
);
417 BUG_ON (!NFS_WBACK_BUSY(req
));
419 spin_lock(&inode
->i_lock
);
420 set_page_private(req
->wb_page
, 0);
421 ClearPagePrivate(req
->wb_page
);
422 radix_tree_delete(&nfsi
->nfs_page_tree
, req
->wb_index
);
425 spin_unlock(&inode
->i_lock
);
428 spin_unlock(&inode
->i_lock
);
429 nfs_clear_request(req
);
430 nfs_release_request(req
);
434 nfs_mark_request_dirty(struct nfs_page
*req
)
436 __set_page_dirty_nobuffers(req
->wb_page
);
437 __mark_inode_dirty(req
->wb_page
->mapping
->host
, I_DIRTY_DATASYNC
);
440 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
442 * Add a request to the inode's commit list.
445 nfs_mark_request_commit(struct nfs_page
*req
)
447 struct inode
*inode
= req
->wb_context
->path
.dentry
->d_inode
;
448 struct nfs_inode
*nfsi
= NFS_I(inode
);
450 spin_lock(&inode
->i_lock
);
451 set_bit(PG_CLEAN
, &(req
)->wb_flags
);
452 radix_tree_tag_set(&nfsi
->nfs_page_tree
,
454 NFS_PAGE_TAG_COMMIT
);
456 spin_unlock(&inode
->i_lock
);
457 inc_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
458 inc_bdi_stat(req
->wb_page
->mapping
->backing_dev_info
, BDI_RECLAIMABLE
);
459 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
463 nfs_clear_request_commit(struct nfs_page
*req
)
465 struct page
*page
= req
->wb_page
;
467 if (test_and_clear_bit(PG_CLEAN
, &(req
)->wb_flags
)) {
468 dec_zone_page_state(page
, NR_UNSTABLE_NFS
);
469 dec_bdi_stat(page
->mapping
->backing_dev_info
, BDI_RECLAIMABLE
);
476 int nfs_write_need_commit(struct nfs_write_data
*data
)
478 return data
->verf
.committed
!= NFS_FILE_SYNC
;
482 int nfs_reschedule_unstable_write(struct nfs_page
*req
)
484 if (test_and_clear_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
485 nfs_mark_request_commit(req
);
488 if (test_and_clear_bit(PG_NEED_RESCHED
, &req
->wb_flags
)) {
489 nfs_mark_request_dirty(req
);
496 nfs_mark_request_commit(struct nfs_page
*req
)
501 nfs_clear_request_commit(struct nfs_page
*req
)
507 int nfs_write_need_commit(struct nfs_write_data
*data
)
513 int nfs_reschedule_unstable_write(struct nfs_page
*req
)
519 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
521 nfs_need_commit(struct nfs_inode
*nfsi
)
523 return radix_tree_tagged(&nfsi
->nfs_page_tree
, NFS_PAGE_TAG_COMMIT
);
527 * nfs_scan_commit - Scan an inode for commit requests
528 * @inode: NFS inode to scan
529 * @dst: destination list
530 * @idx_start: lower bound of page->index to scan.
531 * @npages: idx_start + npages sets the upper bound to scan.
533 * Moves requests from the inode's 'commit' request list.
534 * The requests are *not* checked to ensure that they form a contiguous set.
537 nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, pgoff_t idx_start
, unsigned int npages
)
539 struct nfs_inode
*nfsi
= NFS_I(inode
);
542 if (!nfs_need_commit(nfsi
))
545 ret
= nfs_scan_list(nfsi
, dst
, idx_start
, npages
, NFS_PAGE_TAG_COMMIT
);
547 nfsi
->ncommit
-= ret
;
548 if (nfs_need_commit(NFS_I(inode
)))
549 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
553 static inline int nfs_need_commit(struct nfs_inode
*nfsi
)
558 static inline int nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, pgoff_t idx_start
, unsigned int npages
)
565 * Search for an existing write request, and attempt to update
566 * it to reflect a new dirty region on a given page.
568 * If the attempt fails, then the existing request is flushed out
571 static struct nfs_page
*nfs_try_to_update_request(struct inode
*inode
,
576 struct nfs_page
*req
;
581 if (!PagePrivate(page
))
584 end
= offset
+ bytes
;
585 spin_lock(&inode
->i_lock
);
588 req
= nfs_page_find_request_locked(page
);
592 rqend
= req
->wb_offset
+ req
->wb_bytes
;
594 * Tell the caller to flush out the request if
595 * the offsets are non-contiguous.
596 * Note: nfs_flush_incompatible() will already
597 * have flushed out requests having wrong owners.
600 || end
< req
->wb_offset
)
603 if (nfs_set_page_tag_locked(req
))
606 /* The request is locked, so wait and then retry */
607 spin_unlock(&inode
->i_lock
);
608 error
= nfs_wait_on_request(req
);
609 nfs_release_request(req
);
612 spin_lock(&inode
->i_lock
);
615 if (nfs_clear_request_commit(req
) &&
616 radix_tree_tag_clear(&NFS_I(inode
)->nfs_page_tree
,
617 req
->wb_index
, NFS_PAGE_TAG_COMMIT
) != NULL
)
618 NFS_I(inode
)->ncommit
--;
620 /* Okay, the request matches. Update the region */
621 if (offset
< req
->wb_offset
) {
622 req
->wb_offset
= offset
;
623 req
->wb_pgbase
= offset
;
626 req
->wb_bytes
= end
- req
->wb_offset
;
628 req
->wb_bytes
= rqend
- req
->wb_offset
;
630 spin_unlock(&inode
->i_lock
);
633 spin_unlock(&inode
->i_lock
);
634 nfs_release_request(req
);
635 error
= nfs_wb_page(inode
, page
);
637 return ERR_PTR(error
);
641 * Try to update an existing write request, or create one if there is none.
643 * Note: Should always be called with the Page Lock held to prevent races
644 * if we have to add a new request. Also assumes that the caller has
645 * already called nfs_flush_incompatible() if necessary.
647 static struct nfs_page
* nfs_setup_write_request(struct nfs_open_context
* ctx
,
648 struct page
*page
, unsigned int offset
, unsigned int bytes
)
650 struct inode
*inode
= page
->mapping
->host
;
651 struct nfs_page
*req
;
654 req
= nfs_try_to_update_request(inode
, page
, offset
, bytes
);
657 req
= nfs_create_request(ctx
, inode
, page
, offset
, bytes
);
660 error
= nfs_inode_add_request(inode
, req
);
662 nfs_release_request(req
);
663 req
= ERR_PTR(error
);
669 static int nfs_writepage_setup(struct nfs_open_context
*ctx
, struct page
*page
,
670 unsigned int offset
, unsigned int count
)
672 struct nfs_page
*req
;
674 req
= nfs_setup_write_request(ctx
, page
, offset
, count
);
677 nfs_mark_request_dirty(req
);
678 /* Update file length */
679 nfs_grow_file(page
, offset
, count
);
680 nfs_mark_uptodate(page
, req
->wb_pgbase
, req
->wb_bytes
);
681 nfs_mark_request_dirty(req
);
682 nfs_clear_page_tag_locked(req
);
686 int nfs_flush_incompatible(struct file
*file
, struct page
*page
)
688 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
689 struct nfs_page
*req
;
690 int do_flush
, status
;
692 * Look for a request corresponding to this page. If there
693 * is one, and it belongs to another file, we flush it out
694 * before we try to copy anything into the page. Do this
695 * due to the lack of an ACCESS-type call in NFSv2.
696 * Also do the same if we find a request from an existing
700 req
= nfs_page_find_request(page
);
703 do_flush
= req
->wb_page
!= page
|| req
->wb_context
!= ctx
||
704 req
->wb_lock_context
->lockowner
!= current
->files
||
705 req
->wb_lock_context
->pid
!= current
->tgid
;
706 nfs_release_request(req
);
709 status
= nfs_wb_page(page
->mapping
->host
, page
);
710 } while (status
== 0);
715 * If the page cache is marked as unsafe or invalid, then we can't rely on
716 * the PageUptodate() flag. In this case, we will need to turn off
717 * write optimisations that depend on the page contents being correct.
719 static int nfs_write_pageuptodate(struct page
*page
, struct inode
*inode
)
721 return PageUptodate(page
) &&
722 !(NFS_I(inode
)->cache_validity
& (NFS_INO_REVAL_PAGECACHE
|NFS_INO_INVALID_DATA
));
725 int nfs_updatepage(struct file
*file
, struct page
*page
,
726 unsigned int offset
, unsigned int count
)
728 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
729 struct inode
*inode
= page
->mapping
->host
;
732 nfs_inc_stats(inode
, NFSIOS_VFSUPDATEPAGE
);
734 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
735 file
->f_path
.dentry
->d_parent
->d_name
.name
,
736 file
->f_path
.dentry
->d_name
.name
, count
,
737 (long long)(page_offset(page
) + offset
));
739 /* If we're not using byte range locks, and we know the page
740 * is up to date, it may be more efficient to extend the write
741 * to cover the entire page in order to avoid fragmentation
744 if (nfs_write_pageuptodate(page
, inode
) &&
745 inode
->i_flock
== NULL
&&
746 !(file
->f_flags
& O_DSYNC
)) {
747 count
= max(count
+ offset
, nfs_page_length(page
));
751 status
= nfs_writepage_setup(ctx
, page
, offset
, count
);
753 nfs_set_pageerror(page
);
755 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
756 status
, (long long)i_size_read(inode
));
760 static void nfs_writepage_release(struct nfs_page
*req
)
762 struct page
*page
= req
->wb_page
;
764 if (PageError(req
->wb_page
) || !nfs_reschedule_unstable_write(req
))
765 nfs_inode_remove_request(req
);
766 nfs_clear_page_tag_locked(req
);
767 nfs_end_page_writeback(page
);
770 static int flush_task_priority(int how
)
772 switch (how
& (FLUSH_HIGHPRI
|FLUSH_LOWPRI
)) {
774 return RPC_PRIORITY_HIGH
;
776 return RPC_PRIORITY_LOW
;
778 return RPC_PRIORITY_NORMAL
;
782 * Set up the argument/result storage required for the RPC call.
784 static int nfs_write_rpcsetup(struct nfs_page
*req
,
785 struct nfs_write_data
*data
,
786 const struct rpc_call_ops
*call_ops
,
787 unsigned int count
, unsigned int offset
,
790 struct inode
*inode
= req
->wb_context
->path
.dentry
->d_inode
;
791 int priority
= flush_task_priority(how
);
792 struct rpc_task
*task
;
793 struct rpc_message msg
= {
794 .rpc_argp
= &data
->args
,
795 .rpc_resp
= &data
->res
,
796 .rpc_cred
= req
->wb_context
->cred
,
798 struct rpc_task_setup task_setup_data
= {
799 .rpc_client
= NFS_CLIENT(inode
),
802 .callback_ops
= call_ops
,
803 .callback_data
= data
,
804 .workqueue
= nfsiod_workqueue
,
805 .flags
= RPC_TASK_ASYNC
,
806 .priority
= priority
,
810 /* Set up the RPC argument and reply structs
811 * NB: take care not to mess about with data->commit et al. */
814 data
->inode
= inode
= req
->wb_context
->path
.dentry
->d_inode
;
815 data
->cred
= msg
.rpc_cred
;
817 data
->args
.fh
= NFS_FH(inode
);
818 data
->args
.offset
= req_offset(req
) + offset
;
819 data
->args
.pgbase
= req
->wb_pgbase
+ offset
;
820 data
->args
.pages
= data
->pagevec
;
821 data
->args
.count
= count
;
822 data
->args
.context
= get_nfs_open_context(req
->wb_context
);
823 data
->args
.lock_context
= req
->wb_lock_context
;
824 data
->args
.stable
= NFS_UNSTABLE
;
825 if (how
& FLUSH_STABLE
) {
826 data
->args
.stable
= NFS_DATA_SYNC
;
827 if (!nfs_need_commit(NFS_I(inode
)))
828 data
->args
.stable
= NFS_FILE_SYNC
;
831 data
->res
.fattr
= &data
->fattr
;
832 data
->res
.count
= count
;
833 data
->res
.verf
= &data
->verf
;
834 nfs_fattr_init(&data
->fattr
);
836 /* Set up the initial task struct. */
837 NFS_PROTO(inode
)->write_setup(data
, &msg
);
839 dprintk("NFS: %5u initiated write call "
840 "(req %s/%lld, %u bytes @ offset %llu)\n",
843 (long long)NFS_FILEID(inode
),
845 (unsigned long long)data
->args
.offset
);
847 task
= rpc_run_task(&task_setup_data
);
852 if (how
& FLUSH_SYNC
) {
853 ret
= rpc_wait_for_completion_task(task
);
855 ret
= task
->tk_status
;
862 /* If a nfs_flush_* function fails, it should remove reqs from @head and
863 * call this on each, which will prepare them to be retried on next
864 * writeback using standard nfs.
866 static void nfs_redirty_request(struct nfs_page
*req
)
868 struct page
*page
= req
->wb_page
;
870 nfs_mark_request_dirty(req
);
871 nfs_clear_page_tag_locked(req
);
872 nfs_end_page_writeback(page
);
876 * Generate multiple small requests to write out a single
877 * contiguous dirty area on one page.
879 static int nfs_flush_multi(struct inode
*inode
, struct list_head
*head
, unsigned int npages
, size_t count
, int how
)
881 struct nfs_page
*req
= nfs_list_entry(head
->next
);
882 struct page
*page
= req
->wb_page
;
883 struct nfs_write_data
*data
;
884 size_t wsize
= NFS_SERVER(inode
)->wsize
, nbytes
;
890 nfs_list_remove_request(req
);
894 size_t len
= min(nbytes
, wsize
);
896 data
= nfs_writedata_alloc(1);
899 list_add(&data
->pages
, &list
);
902 } while (nbytes
!= 0);
903 atomic_set(&req
->wb_complete
, requests
);
905 ClearPageError(page
);
911 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
912 list_del_init(&data
->pages
);
914 data
->pagevec
[0] = page
;
918 ret2
= nfs_write_rpcsetup(req
, data
, &nfs_write_partial_ops
,
924 } while (nbytes
!= 0);
929 while (!list_empty(&list
)) {
930 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
931 list_del(&data
->pages
);
932 nfs_writedata_release(data
);
934 nfs_redirty_request(req
);
939 * Create an RPC task for the given write request and kick it.
940 * The page must have been locked by the caller.
942 * It may happen that the page we're passed is not marked dirty.
943 * This is the case if nfs_updatepage detects a conflicting request
944 * that has been written but not committed.
946 static int nfs_flush_one(struct inode
*inode
, struct list_head
*head
, unsigned int npages
, size_t count
, int how
)
948 struct nfs_page
*req
;
950 struct nfs_write_data
*data
;
952 data
= nfs_writedata_alloc(npages
);
956 pages
= data
->pagevec
;
957 while (!list_empty(head
)) {
958 req
= nfs_list_entry(head
->next
);
959 nfs_list_remove_request(req
);
960 nfs_list_add_request(req
, &data
->pages
);
961 ClearPageError(req
->wb_page
);
962 *pages
++ = req
->wb_page
;
964 req
= nfs_list_entry(data
->pages
.next
);
966 /* Set up the argument struct */
967 return nfs_write_rpcsetup(req
, data
, &nfs_write_full_ops
, count
, 0, how
);
969 while (!list_empty(head
)) {
970 req
= nfs_list_entry(head
->next
);
971 nfs_list_remove_request(req
);
972 nfs_redirty_request(req
);
977 static void nfs_pageio_init_write(struct nfs_pageio_descriptor
*pgio
,
978 struct inode
*inode
, int ioflags
)
980 size_t wsize
= NFS_SERVER(inode
)->wsize
;
982 if (wsize
< PAGE_CACHE_SIZE
)
983 nfs_pageio_init(pgio
, inode
, nfs_flush_multi
, wsize
, ioflags
);
985 nfs_pageio_init(pgio
, inode
, nfs_flush_one
, wsize
, ioflags
);
989 * Handle a write reply that flushed part of a page.
991 static void nfs_writeback_done_partial(struct rpc_task
*task
, void *calldata
)
993 struct nfs_write_data
*data
= calldata
;
995 dprintk("NFS: %5u write(%s/%lld %d@%lld)",
997 data
->req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
999 NFS_FILEID(data
->req
->wb_context
->path
.dentry
->d_inode
),
1000 data
->req
->wb_bytes
, (long long)req_offset(data
->req
));
1002 nfs_writeback_done(task
, data
);
1005 static void nfs_writeback_release_partial(void *calldata
)
1007 struct nfs_write_data
*data
= calldata
;
1008 struct nfs_page
*req
= data
->req
;
1009 struct page
*page
= req
->wb_page
;
1010 int status
= data
->task
.tk_status
;
1013 nfs_set_pageerror(page
);
1014 nfs_context_set_write_error(req
->wb_context
, status
);
1015 dprintk(", error = %d\n", status
);
1019 if (nfs_write_need_commit(data
)) {
1020 struct inode
*inode
= page
->mapping
->host
;
1022 spin_lock(&inode
->i_lock
);
1023 if (test_bit(PG_NEED_RESCHED
, &req
->wb_flags
)) {
1024 /* Do nothing we need to resend the writes */
1025 } else if (!test_and_set_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
1026 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1027 dprintk(" defer commit\n");
1028 } else if (memcmp(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
))) {
1029 set_bit(PG_NEED_RESCHED
, &req
->wb_flags
);
1030 clear_bit(PG_NEED_COMMIT
, &req
->wb_flags
);
1031 dprintk(" server reboot detected\n");
1033 spin_unlock(&inode
->i_lock
);
1038 if (atomic_dec_and_test(&req
->wb_complete
))
1039 nfs_writepage_release(req
);
1040 nfs_writedata_release(calldata
);
1043 #if defined(CONFIG_NFS_V4_1)
1044 void nfs_write_prepare(struct rpc_task
*task
, void *calldata
)
1046 struct nfs_write_data
*data
= calldata
;
1048 if (nfs4_setup_sequence(NFS_SERVER(data
->inode
),
1049 &data
->args
.seq_args
,
1050 &data
->res
.seq_res
, 1, task
))
1052 rpc_call_start(task
);
1054 #endif /* CONFIG_NFS_V4_1 */
1056 static const struct rpc_call_ops nfs_write_partial_ops
= {
1057 #if defined(CONFIG_NFS_V4_1)
1058 .rpc_call_prepare
= nfs_write_prepare
,
1059 #endif /* CONFIG_NFS_V4_1 */
1060 .rpc_call_done
= nfs_writeback_done_partial
,
1061 .rpc_release
= nfs_writeback_release_partial
,
1064 static void nfs_writeback_done_full(struct rpc_task
*task
, void *calldata
)
1066 struct nfs_write_data
*data
= calldata
;
1068 nfs_writeback_done(task
, data
);
1071 static void nfs_writeback_release_full(void *calldata
)
1073 struct nfs_write_data
*data
= calldata
;
1074 int status
= data
->task
.tk_status
;
1076 /* Update attributes as result of writeback. */
1077 while (!list_empty(&data
->pages
)) {
1078 struct nfs_page
*req
= nfs_list_entry(data
->pages
.next
);
1079 struct page
*page
= req
->wb_page
;
1081 nfs_list_remove_request(req
);
1083 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1085 req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
1086 (long long)NFS_FILEID(req
->wb_context
->path
.dentry
->d_inode
),
1088 (long long)req_offset(req
));
1091 nfs_set_pageerror(page
);
1092 nfs_context_set_write_error(req
->wb_context
, status
);
1093 dprintk(", error = %d\n", status
);
1094 goto remove_request
;
1097 if (nfs_write_need_commit(data
)) {
1098 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1099 nfs_mark_request_commit(req
);
1100 dprintk(" marked for commit\n");
1105 nfs_inode_remove_request(req
);
1107 nfs_clear_page_tag_locked(req
);
1108 nfs_end_page_writeback(page
);
1110 nfs_writedata_release(calldata
);
1113 static const struct rpc_call_ops nfs_write_full_ops
= {
1114 #if defined(CONFIG_NFS_V4_1)
1115 .rpc_call_prepare
= nfs_write_prepare
,
1116 #endif /* CONFIG_NFS_V4_1 */
1117 .rpc_call_done
= nfs_writeback_done_full
,
1118 .rpc_release
= nfs_writeback_release_full
,
1123 * This function is called when the WRITE call is complete.
1125 int nfs_writeback_done(struct rpc_task
*task
, struct nfs_write_data
*data
)
1127 struct nfs_writeargs
*argp
= &data
->args
;
1128 struct nfs_writeres
*resp
= &data
->res
;
1129 struct nfs_server
*server
= NFS_SERVER(data
->inode
);
1132 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1133 task
->tk_pid
, task
->tk_status
);
1136 * ->write_done will attempt to use post-op attributes to detect
1137 * conflicting writes by other clients. A strict interpretation
1138 * of close-to-open would allow us to continue caching even if
1139 * another writer had changed the file, but some applications
1140 * depend on tighter cache coherency when writing.
1142 status
= NFS_PROTO(data
->inode
)->write_done(task
, data
);
1145 nfs_add_stats(data
->inode
, NFSIOS_SERVERWRITTENBYTES
, resp
->count
);
1147 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1148 if (resp
->verf
->committed
< argp
->stable
&& task
->tk_status
>= 0) {
1149 /* We tried a write call, but the server did not
1150 * commit data to stable storage even though we
1152 * Note: There is a known bug in Tru64 < 5.0 in which
1153 * the server reports NFS_DATA_SYNC, but performs
1154 * NFS_FILE_SYNC. We therefore implement this checking
1155 * as a dprintk() in order to avoid filling syslog.
1157 static unsigned long complain
;
1159 if (time_before(complain
, jiffies
)) {
1160 dprintk("NFS: faulty NFS server %s:"
1161 " (committed = %d) != (stable = %d)\n",
1162 server
->nfs_client
->cl_hostname
,
1163 resp
->verf
->committed
, argp
->stable
);
1164 complain
= jiffies
+ 300 * HZ
;
1168 /* Is this a short write? */
1169 if (task
->tk_status
>= 0 && resp
->count
< argp
->count
) {
1170 static unsigned long complain
;
1172 nfs_inc_stats(data
->inode
, NFSIOS_SHORTWRITE
);
1174 /* Has the server at least made some progress? */
1175 if (resp
->count
!= 0) {
1176 /* Was this an NFSv2 write or an NFSv3 stable write? */
1177 if (resp
->verf
->committed
!= NFS_UNSTABLE
) {
1178 /* Resend from where the server left off */
1179 argp
->offset
+= resp
->count
;
1180 argp
->pgbase
+= resp
->count
;
1181 argp
->count
-= resp
->count
;
1183 /* Resend as a stable write in order to avoid
1184 * headaches in the case of a server crash.
1186 argp
->stable
= NFS_FILE_SYNC
;
1188 nfs_restart_rpc(task
, server
->nfs_client
);
1191 if (time_before(complain
, jiffies
)) {
1193 "NFS: Server wrote zero bytes, expected %u.\n",
1195 complain
= jiffies
+ 300 * HZ
;
1197 /* Can't do anything about it except throw an error. */
1198 task
->tk_status
= -EIO
;
1204 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1205 static int nfs_commit_set_lock(struct nfs_inode
*nfsi
, int may_wait
)
1207 if (!test_and_set_bit(NFS_INO_COMMIT
, &nfsi
->flags
))
1209 if (may_wait
&& !out_of_line_wait_on_bit_lock(&nfsi
->flags
,
1210 NFS_INO_COMMIT
, nfs_wait_bit_killable
,
1216 static void nfs_commit_clear_lock(struct nfs_inode
*nfsi
)
1218 clear_bit(NFS_INO_COMMIT
, &nfsi
->flags
);
1219 smp_mb__after_clear_bit();
1220 wake_up_bit(&nfsi
->flags
, NFS_INO_COMMIT
);
1224 static void nfs_commitdata_release(void *data
)
1226 struct nfs_write_data
*wdata
= data
;
1228 put_nfs_open_context(wdata
->args
.context
);
1229 nfs_commit_free(wdata
);
1233 * Set up the argument/result storage required for the RPC call.
1235 static int nfs_commit_rpcsetup(struct list_head
*head
,
1236 struct nfs_write_data
*data
,
1239 struct nfs_page
*first
= nfs_list_entry(head
->next
);
1240 struct inode
*inode
= first
->wb_context
->path
.dentry
->d_inode
;
1241 int priority
= flush_task_priority(how
);
1242 struct rpc_task
*task
;
1243 struct rpc_message msg
= {
1244 .rpc_argp
= &data
->args
,
1245 .rpc_resp
= &data
->res
,
1246 .rpc_cred
= first
->wb_context
->cred
,
1248 struct rpc_task_setup task_setup_data
= {
1249 .task
= &data
->task
,
1250 .rpc_client
= NFS_CLIENT(inode
),
1251 .rpc_message
= &msg
,
1252 .callback_ops
= &nfs_commit_ops
,
1253 .callback_data
= data
,
1254 .workqueue
= nfsiod_workqueue
,
1255 .flags
= RPC_TASK_ASYNC
,
1256 .priority
= priority
,
1259 /* Set up the RPC argument and reply structs
1260 * NB: take care not to mess about with data->commit et al. */
1262 list_splice_init(head
, &data
->pages
);
1264 data
->inode
= inode
;
1265 data
->cred
= msg
.rpc_cred
;
1267 data
->args
.fh
= NFS_FH(data
->inode
);
1268 /* Note: we always request a commit of the entire inode */
1269 data
->args
.offset
= 0;
1270 data
->args
.count
= 0;
1271 data
->args
.context
= get_nfs_open_context(first
->wb_context
);
1272 data
->res
.count
= 0;
1273 data
->res
.fattr
= &data
->fattr
;
1274 data
->res
.verf
= &data
->verf
;
1275 nfs_fattr_init(&data
->fattr
);
1277 /* Set up the initial task struct. */
1278 NFS_PROTO(inode
)->commit_setup(data
, &msg
);
1280 dprintk("NFS: %5u initiated commit call\n", data
->task
.tk_pid
);
1282 task
= rpc_run_task(&task_setup_data
);
1284 return PTR_ERR(task
);
1290 * Commit dirty pages
1293 nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
)
1295 struct nfs_write_data
*data
;
1296 struct nfs_page
*req
;
1298 data
= nfs_commitdata_alloc();
1303 /* Set up the argument struct */
1304 return nfs_commit_rpcsetup(head
, data
, how
);
1306 while (!list_empty(head
)) {
1307 req
= nfs_list_entry(head
->next
);
1308 nfs_list_remove_request(req
);
1309 nfs_mark_request_commit(req
);
1310 dec_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
1311 dec_bdi_stat(req
->wb_page
->mapping
->backing_dev_info
,
1313 nfs_clear_page_tag_locked(req
);
1315 nfs_commit_clear_lock(NFS_I(inode
));
1320 * COMMIT call returned
1322 static void nfs_commit_done(struct rpc_task
*task
, void *calldata
)
1324 struct nfs_write_data
*data
= calldata
;
1326 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1327 task
->tk_pid
, task
->tk_status
);
1329 /* Call the NFS version-specific code */
1330 if (NFS_PROTO(data
->inode
)->commit_done(task
, data
) != 0)
1334 static void nfs_commit_release(void *calldata
)
1336 struct nfs_write_data
*data
= calldata
;
1337 struct nfs_page
*req
;
1338 int status
= data
->task
.tk_status
;
1340 while (!list_empty(&data
->pages
)) {
1341 req
= nfs_list_entry(data
->pages
.next
);
1342 nfs_list_remove_request(req
);
1343 nfs_clear_request_commit(req
);
1345 dprintk("NFS: commit (%s/%lld %d@%lld)",
1346 req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
1347 (long long)NFS_FILEID(req
->wb_context
->path
.dentry
->d_inode
),
1349 (long long)req_offset(req
));
1351 nfs_context_set_write_error(req
->wb_context
, status
);
1352 nfs_inode_remove_request(req
);
1353 dprintk(", error = %d\n", status
);
1357 /* Okay, COMMIT succeeded, apparently. Check the verifier
1358 * returned by the server against all stored verfs. */
1359 if (!memcmp(req
->wb_verf
.verifier
, data
->verf
.verifier
, sizeof(data
->verf
.verifier
))) {
1360 /* We have a match */
1361 nfs_inode_remove_request(req
);
1365 /* We have a mismatch. Write the page again */
1366 dprintk(" mismatch\n");
1367 nfs_mark_request_dirty(req
);
1369 nfs_clear_page_tag_locked(req
);
1371 nfs_commit_clear_lock(NFS_I(data
->inode
));
1372 nfs_commitdata_release(calldata
);
1375 static const struct rpc_call_ops nfs_commit_ops
= {
1376 #if defined(CONFIG_NFS_V4_1)
1377 .rpc_call_prepare
= nfs_write_prepare
,
1378 #endif /* CONFIG_NFS_V4_1 */
1379 .rpc_call_done
= nfs_commit_done
,
1380 .rpc_release
= nfs_commit_release
,
1383 int nfs_commit_inode(struct inode
*inode
, int how
)
1386 int may_wait
= how
& FLUSH_SYNC
;
1389 if (!nfs_commit_set_lock(NFS_I(inode
), may_wait
))
1390 goto out_mark_dirty
;
1391 spin_lock(&inode
->i_lock
);
1392 res
= nfs_scan_commit(inode
, &head
, 0, 0);
1393 spin_unlock(&inode
->i_lock
);
1395 int error
= nfs_commit_list(inode
, &head
, how
);
1399 wait_on_bit(&NFS_I(inode
)->flags
, NFS_INO_COMMIT
,
1400 nfs_wait_bit_killable
,
1403 goto out_mark_dirty
;
1405 nfs_commit_clear_lock(NFS_I(inode
));
1407 /* Note: If we exit without ensuring that the commit is complete,
1408 * we must mark the inode as dirty. Otherwise, future calls to
1409 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1410 * that the data is on the disk.
1413 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
1417 static int nfs_commit_unstable_pages(struct inode
*inode
, struct writeback_control
*wbc
)
1419 struct nfs_inode
*nfsi
= NFS_I(inode
);
1420 int flags
= FLUSH_SYNC
;
1423 /* Don't commit yet if this is a non-blocking flush and there are
1424 * lots of outstanding writes for this mapping.
1426 if (wbc
->sync_mode
== WB_SYNC_NONE
&&
1427 nfsi
->ncommit
<= (nfsi
->npages
>> 1))
1428 goto out_mark_dirty
;
1430 if (wbc
->nonblocking
|| wbc
->for_background
)
1432 ret
= nfs_commit_inode(inode
, flags
);
1434 if (wbc
->sync_mode
== WB_SYNC_NONE
) {
1435 if (ret
< wbc
->nr_to_write
)
1436 wbc
->nr_to_write
-= ret
;
1438 wbc
->nr_to_write
= 0;
1443 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
1447 static int nfs_commit_unstable_pages(struct inode
*inode
, struct writeback_control
*wbc
)
1453 int nfs_write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1455 return nfs_commit_unstable_pages(inode
, wbc
);
1459 * flush the inode to disk.
1461 int nfs_wb_all(struct inode
*inode
)
1463 struct writeback_control wbc
= {
1464 .sync_mode
= WB_SYNC_ALL
,
1465 .nr_to_write
= LONG_MAX
,
1467 .range_end
= LLONG_MAX
,
1470 return sync_inode(inode
, &wbc
);
1473 int nfs_wb_page_cancel(struct inode
*inode
, struct page
*page
)
1475 struct nfs_page
*req
;
1478 BUG_ON(!PageLocked(page
));
1480 wait_on_page_writeback(page
);
1481 req
= nfs_page_find_request(page
);
1484 if (nfs_lock_request_dontget(req
)) {
1485 nfs_inode_remove_request(req
);
1487 * In case nfs_inode_remove_request has marked the
1488 * page as being dirty
1490 cancel_dirty_page(page
, PAGE_CACHE_SIZE
);
1491 nfs_unlock_request(req
);
1494 ret
= nfs_wait_on_request(req
);
1495 nfs_release_request(req
);
1503 * Write back all requests on one page - we do this before reading it.
1505 int nfs_wb_page(struct inode
*inode
, struct page
*page
)
1507 loff_t range_start
= page_offset(page
);
1508 loff_t range_end
= range_start
+ (loff_t
)(PAGE_CACHE_SIZE
- 1);
1509 struct writeback_control wbc
= {
1510 .sync_mode
= WB_SYNC_ALL
,
1512 .range_start
= range_start
,
1513 .range_end
= range_end
,
1518 wait_on_page_writeback(page
);
1519 if (clear_page_dirty_for_io(page
)) {
1520 ret
= nfs_writepage_locked(page
, &wbc
);
1525 if (!PagePrivate(page
))
1527 ret
= nfs_commit_inode(inode
, FLUSH_SYNC
);
1536 #ifdef CONFIG_MIGRATION
1537 int nfs_migrate_page(struct address_space
*mapping
, struct page
*newpage
,
1540 struct nfs_page
*req
;
1543 nfs_fscache_release_page(page
, GFP_KERNEL
);
1545 req
= nfs_find_and_lock_request(page
, false);
1550 ret
= migrate_page(mapping
, newpage
, page
);
1555 page_cache_get(newpage
);
1556 spin_lock(&mapping
->host
->i_lock
);
1557 req
->wb_page
= newpage
;
1558 SetPagePrivate(newpage
);
1559 set_page_private(newpage
, (unsigned long)req
);
1560 ClearPagePrivate(page
);
1561 set_page_private(page
, 0);
1562 spin_unlock(&mapping
->host
->i_lock
);
1563 page_cache_release(page
);
1565 nfs_clear_page_tag_locked(req
);
1571 int __init
nfs_init_writepagecache(void)
1573 nfs_wdata_cachep
= kmem_cache_create("nfs_write_data",
1574 sizeof(struct nfs_write_data
),
1575 0, SLAB_HWCACHE_ALIGN
,
1577 if (nfs_wdata_cachep
== NULL
)
1580 nfs_wdata_mempool
= mempool_create_slab_pool(MIN_POOL_WRITE
,
1582 if (nfs_wdata_mempool
== NULL
)
1585 nfs_commit_mempool
= mempool_create_slab_pool(MIN_POOL_COMMIT
,
1587 if (nfs_commit_mempool
== NULL
)
1591 * NFS congestion size, scale with available memory.
1603 * This allows larger machines to have larger/more transfers.
1604 * Limit the default to 256M
1606 nfs_congestion_kb
= (16*int_sqrt(totalram_pages
)) << (PAGE_SHIFT
-10);
1607 if (nfs_congestion_kb
> 256*1024)
1608 nfs_congestion_kb
= 256*1024;
1613 void nfs_destroy_writepagecache(void)
1615 mempool_destroy(nfs_commit_mempool
);
1616 mempool_destroy(nfs_wdata_mempool
);
1617 kmem_cache_destroy(nfs_wdata_cachep
);