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>
17 #include <linux/sunrpc/clnt.h>
18 #include <linux/nfs_fs.h>
19 #include <linux/nfs_mount.h>
20 #include <linux/nfs_page.h>
21 #include <linux/backing-dev.h>
23 #include <asm/uaccess.h>
25 #include "delegation.h"
29 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
31 #define MIN_POOL_WRITE (32)
32 #define MIN_POOL_COMMIT (4)
35 * Local function declarations
37 static struct nfs_page
* nfs_update_request(struct nfs_open_context
*,
39 unsigned int, unsigned int);
40 static void nfs_pageio_init_write(struct nfs_pageio_descriptor
*desc
,
41 struct inode
*inode
, int ioflags
);
42 static const struct rpc_call_ops nfs_write_partial_ops
;
43 static const struct rpc_call_ops nfs_write_full_ops
;
44 static const struct rpc_call_ops nfs_commit_ops
;
46 static struct kmem_cache
*nfs_wdata_cachep
;
47 static mempool_t
*nfs_wdata_mempool
;
48 static mempool_t
*nfs_commit_mempool
;
50 struct nfs_write_data
*nfs_commit_alloc(void)
52 struct nfs_write_data
*p
= mempool_alloc(nfs_commit_mempool
, GFP_NOFS
);
55 memset(p
, 0, sizeof(*p
));
56 INIT_LIST_HEAD(&p
->pages
);
61 static void nfs_commit_rcu_free(struct rcu_head
*head
)
63 struct nfs_write_data
*p
= container_of(head
, struct nfs_write_data
, task
.u
.tk_rcu
);
64 if (p
&& (p
->pagevec
!= &p
->page_array
[0]))
66 mempool_free(p
, nfs_commit_mempool
);
69 void nfs_commit_free(struct nfs_write_data
*wdata
)
71 call_rcu_bh(&wdata
->task
.u
.tk_rcu
, nfs_commit_rcu_free
);
74 struct nfs_write_data
*nfs_writedata_alloc(unsigned int pagecount
)
76 struct nfs_write_data
*p
= mempool_alloc(nfs_wdata_mempool
, GFP_NOFS
);
79 memset(p
, 0, sizeof(*p
));
80 INIT_LIST_HEAD(&p
->pages
);
81 p
->npages
= pagecount
;
82 if (pagecount
<= ARRAY_SIZE(p
->page_array
))
83 p
->pagevec
= p
->page_array
;
85 p
->pagevec
= kcalloc(pagecount
, sizeof(struct page
*), GFP_NOFS
);
87 mempool_free(p
, nfs_wdata_mempool
);
95 static void nfs_writedata_rcu_free(struct rcu_head
*head
)
97 struct nfs_write_data
*p
= container_of(head
, struct nfs_write_data
, task
.u
.tk_rcu
);
98 if (p
&& (p
->pagevec
!= &p
->page_array
[0]))
100 mempool_free(p
, nfs_wdata_mempool
);
103 static void nfs_writedata_free(struct nfs_write_data
*wdata
)
105 call_rcu_bh(&wdata
->task
.u
.tk_rcu
, nfs_writedata_rcu_free
);
108 void nfs_writedata_release(void *wdata
)
110 nfs_writedata_free(wdata
);
113 static struct nfs_page
*nfs_page_find_request_locked(struct page
*page
)
115 struct nfs_page
*req
= NULL
;
117 if (PagePrivate(page
)) {
118 req
= (struct nfs_page
*)page_private(page
);
120 kref_get(&req
->wb_kref
);
125 static struct nfs_page
*nfs_page_find_request(struct page
*page
)
127 struct inode
*inode
= page
->mapping
->host
;
128 struct nfs_page
*req
= NULL
;
130 spin_lock(&inode
->i_lock
);
131 req
= nfs_page_find_request_locked(page
);
132 spin_unlock(&inode
->i_lock
);
136 /* Adjust the file length if we're writing beyond the end */
137 static void nfs_grow_file(struct page
*page
, unsigned int offset
, unsigned int count
)
139 struct inode
*inode
= page
->mapping
->host
;
140 loff_t end
, i_size
= i_size_read(inode
);
141 pgoff_t end_index
= (i_size
- 1) >> PAGE_CACHE_SHIFT
;
143 if (i_size
> 0 && page
->index
< end_index
)
145 end
= ((loff_t
)page
->index
<< PAGE_CACHE_SHIFT
) + ((loff_t
)offset
+count
);
148 nfs_inc_stats(inode
, NFSIOS_EXTENDWRITE
);
149 i_size_write(inode
, end
);
152 /* A writeback failed: mark the page as bad, and invalidate the page cache */
153 static void nfs_set_pageerror(struct page
*page
)
156 nfs_zap_mapping(page
->mapping
->host
, page
->mapping
);
159 /* We can set the PG_uptodate flag if we see that a write request
160 * covers the full page.
162 static void nfs_mark_uptodate(struct page
*page
, unsigned int base
, unsigned int count
)
164 if (PageUptodate(page
))
168 if (count
!= nfs_page_length(page
))
170 if (count
!= PAGE_CACHE_SIZE
)
171 zero_user_page(page
, count
, PAGE_CACHE_SIZE
- count
, KM_USER0
);
172 SetPageUptodate(page
);
175 static int nfs_writepage_setup(struct nfs_open_context
*ctx
, struct page
*page
,
176 unsigned int offset
, unsigned int count
)
178 struct nfs_page
*req
;
182 req
= nfs_update_request(ctx
, page
, offset
, count
);
188 ret
= nfs_wb_page(page
->mapping
->host
, page
);
192 /* Update file length */
193 nfs_grow_file(page
, offset
, count
);
194 nfs_unlock_request(req
);
198 static int wb_priority(struct writeback_control
*wbc
)
200 if (wbc
->for_reclaim
)
201 return FLUSH_HIGHPRI
| FLUSH_STABLE
;
202 if (wbc
->for_kupdate
)
208 * NFS congestion control
211 int nfs_congestion_kb
;
213 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
214 #define NFS_CONGESTION_OFF_THRESH \
215 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
217 static int nfs_set_page_writeback(struct page
*page
)
219 int ret
= test_set_page_writeback(page
);
222 struct inode
*inode
= page
->mapping
->host
;
223 struct nfs_server
*nfss
= NFS_SERVER(inode
);
225 if (atomic_long_inc_return(&nfss
->writeback
) >
226 NFS_CONGESTION_ON_THRESH
)
227 set_bdi_congested(&nfss
->backing_dev_info
, WRITE
);
232 static void nfs_end_page_writeback(struct page
*page
)
234 struct inode
*inode
= page
->mapping
->host
;
235 struct nfs_server
*nfss
= NFS_SERVER(inode
);
237 end_page_writeback(page
);
238 if (atomic_long_dec_return(&nfss
->writeback
) < NFS_CONGESTION_OFF_THRESH
) {
239 clear_bdi_congested(&nfss
->backing_dev_info
, WRITE
);
240 congestion_end(WRITE
);
245 * Find an associated nfs write request, and prepare to flush it out
246 * May return an error if the user signalled nfs_wait_on_request().
248 static int nfs_page_async_flush(struct nfs_pageio_descriptor
*pgio
,
251 struct inode
*inode
= page
->mapping
->host
;
252 struct nfs_inode
*nfsi
= NFS_I(inode
);
253 struct nfs_page
*req
;
256 spin_lock(&inode
->i_lock
);
258 req
= nfs_page_find_request_locked(page
);
260 spin_unlock(&inode
->i_lock
);
263 if (nfs_lock_request_dontget(req
))
265 /* Note: If we hold the page lock, as is the case in nfs_writepage,
266 * then the call to nfs_lock_request_dontget() will always
267 * succeed provided that someone hasn't already marked the
268 * request as dirty (in which case we don't care).
270 spin_unlock(&inode
->i_lock
);
271 ret
= nfs_wait_on_request(req
);
272 nfs_release_request(req
);
275 spin_lock(&inode
->i_lock
);
277 if (test_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
278 /* This request is marked for commit */
279 spin_unlock(&inode
->i_lock
);
280 nfs_unlock_request(req
);
281 nfs_pageio_complete(pgio
);
284 if (nfs_set_page_writeback(page
) != 0) {
285 spin_unlock(&inode
->i_lock
);
288 radix_tree_tag_set(&nfsi
->nfs_page_tree
, req
->wb_index
,
289 NFS_PAGE_TAG_LOCKED
);
290 spin_unlock(&inode
->i_lock
);
291 nfs_pageio_add_request(pgio
, req
);
295 static int nfs_do_writepage(struct page
*page
, struct writeback_control
*wbc
, struct nfs_pageio_descriptor
*pgio
)
297 struct inode
*inode
= page
->mapping
->host
;
299 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGE
);
300 nfs_add_stats(inode
, NFSIOS_WRITEPAGES
, 1);
302 nfs_pageio_cond_complete(pgio
, page
->index
);
303 return nfs_page_async_flush(pgio
, page
);
307 * Write an mmapped page to the server.
309 static int nfs_writepage_locked(struct page
*page
, struct writeback_control
*wbc
)
311 struct nfs_pageio_descriptor pgio
;
314 nfs_pageio_init_write(&pgio
, page
->mapping
->host
, wb_priority(wbc
));
315 err
= nfs_do_writepage(page
, wbc
, &pgio
);
316 nfs_pageio_complete(&pgio
);
319 if (pgio
.pg_error
< 0)
320 return pgio
.pg_error
;
324 int nfs_writepage(struct page
*page
, struct writeback_control
*wbc
)
328 ret
= nfs_writepage_locked(page
, wbc
);
333 static int nfs_writepages_callback(struct page
*page
, struct writeback_control
*wbc
, void *data
)
337 ret
= nfs_do_writepage(page
, wbc
, data
);
342 int nfs_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
344 struct inode
*inode
= mapping
->host
;
345 struct nfs_pageio_descriptor pgio
;
348 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGES
);
350 nfs_pageio_init_write(&pgio
, inode
, wb_priority(wbc
));
351 err
= write_cache_pages(mapping
, wbc
, nfs_writepages_callback
, &pgio
);
352 nfs_pageio_complete(&pgio
);
355 if (pgio
.pg_error
< 0)
356 return pgio
.pg_error
;
361 * Insert a write request into an inode
363 static int nfs_inode_add_request(struct inode
*inode
, struct nfs_page
*req
)
365 struct nfs_inode
*nfsi
= NFS_I(inode
);
368 error
= radix_tree_insert(&nfsi
->nfs_page_tree
, req
->wb_index
, req
);
369 BUG_ON(error
== -EEXIST
);
374 nfs_begin_data_update(inode
);
375 if (nfs_have_delegation(inode
, FMODE_WRITE
))
378 SetPagePrivate(req
->wb_page
);
379 set_page_private(req
->wb_page
, (unsigned long)req
);
381 kref_get(&req
->wb_kref
);
386 * Remove a write request from an inode
388 static void nfs_inode_remove_request(struct nfs_page
*req
)
390 struct inode
*inode
= req
->wb_context
->path
.dentry
->d_inode
;
391 struct nfs_inode
*nfsi
= NFS_I(inode
);
393 BUG_ON (!NFS_WBACK_BUSY(req
));
395 spin_lock(&inode
->i_lock
);
396 set_page_private(req
->wb_page
, 0);
397 ClearPagePrivate(req
->wb_page
);
398 radix_tree_delete(&nfsi
->nfs_page_tree
, req
->wb_index
);
401 spin_unlock(&inode
->i_lock
);
402 nfs_end_data_update(inode
);
405 spin_unlock(&inode
->i_lock
);
406 nfs_clear_request(req
);
407 nfs_release_request(req
);
411 nfs_redirty_request(struct nfs_page
*req
)
413 __set_page_dirty_nobuffers(req
->wb_page
);
417 * Check if a request is dirty
420 nfs_dirty_request(struct nfs_page
*req
)
422 struct page
*page
= req
->wb_page
;
424 if (page
== NULL
|| test_bit(PG_NEED_COMMIT
, &req
->wb_flags
))
426 return !PageWriteback(req
->wb_page
);
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
);
441 set_bit(PG_NEED_COMMIT
, &(req
)->wb_flags
);
442 radix_tree_tag_set(&nfsi
->nfs_page_tree
,
444 NFS_PAGE_TAG_COMMIT
);
445 spin_unlock(&inode
->i_lock
);
446 inc_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
447 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
451 int nfs_write_need_commit(struct nfs_write_data
*data
)
453 return data
->verf
.committed
!= NFS_FILE_SYNC
;
457 int nfs_reschedule_unstable_write(struct nfs_page
*req
)
459 if (test_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
460 nfs_mark_request_commit(req
);
463 if (test_and_clear_bit(PG_NEED_RESCHED
, &req
->wb_flags
)) {
464 nfs_redirty_request(req
);
471 nfs_mark_request_commit(struct nfs_page
*req
)
476 int nfs_write_need_commit(struct nfs_write_data
*data
)
482 int nfs_reschedule_unstable_write(struct nfs_page
*req
)
489 * Wait for a request to complete.
491 * Interruptible by signals only if mounted with intr flag.
493 static int nfs_wait_on_requests_locked(struct inode
*inode
, pgoff_t idx_start
, unsigned int npages
)
495 struct nfs_inode
*nfsi
= NFS_I(inode
);
496 struct nfs_page
*req
;
497 pgoff_t idx_end
, next
;
498 unsigned int res
= 0;
504 idx_end
= idx_start
+ npages
- 1;
507 while (radix_tree_gang_lookup_tag(&nfsi
->nfs_page_tree
, (void **)&req
, next
, 1, NFS_PAGE_TAG_LOCKED
)) {
508 if (req
->wb_index
> idx_end
)
511 next
= req
->wb_index
+ 1;
512 BUG_ON(!NFS_WBACK_BUSY(req
));
514 kref_get(&req
->wb_kref
);
515 spin_unlock(&inode
->i_lock
);
516 error
= nfs_wait_on_request(req
);
517 nfs_release_request(req
);
518 spin_lock(&inode
->i_lock
);
526 static void nfs_cancel_commit_list(struct list_head
*head
)
528 struct nfs_page
*req
;
530 while(!list_empty(head
)) {
531 req
= nfs_list_entry(head
->next
);
532 dec_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
533 nfs_list_remove_request(req
);
534 clear_bit(PG_NEED_COMMIT
, &(req
)->wb_flags
);
535 nfs_inode_remove_request(req
);
536 nfs_unlock_request(req
);
540 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
542 * nfs_scan_commit - Scan an inode for commit requests
543 * @inode: NFS inode to scan
544 * @dst: destination list
545 * @idx_start: lower bound of page->index to scan.
546 * @npages: idx_start + npages sets the upper bound to scan.
548 * Moves requests from the inode's 'commit' request list.
549 * The requests are *not* checked to ensure that they form a contiguous set.
552 nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, pgoff_t idx_start
, unsigned int npages
)
554 struct nfs_inode
*nfsi
= NFS_I(inode
);
557 if (nfsi
->ncommit
!= 0) {
558 res
= nfs_scan_list(nfsi
, dst
, idx_start
, npages
,
559 NFS_PAGE_TAG_COMMIT
);
560 nfsi
->ncommit
-= res
;
565 static inline int nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, pgoff_t idx_start
, unsigned int npages
)
572 * Try to update any existing write request, or create one if there is none.
573 * In order to match, the request's credentials must match those of
574 * the calling process.
576 * Note: Should always be called with the Page Lock held!
578 static struct nfs_page
* nfs_update_request(struct nfs_open_context
* ctx
,
579 struct page
*page
, unsigned int offset
, unsigned int bytes
)
581 struct address_space
*mapping
= page
->mapping
;
582 struct inode
*inode
= mapping
->host
;
583 struct nfs_page
*req
, *new = NULL
;
586 end
= offset
+ bytes
;
589 /* Loop over all inode entries and see if we find
590 * A request for the page we wish to update
592 spin_lock(&inode
->i_lock
);
593 req
= nfs_page_find_request_locked(page
);
595 if (!nfs_lock_request_dontget(req
)) {
598 spin_unlock(&inode
->i_lock
);
599 error
= nfs_wait_on_request(req
);
600 nfs_release_request(req
);
603 nfs_release_request(new);
604 return ERR_PTR(error
);
608 spin_unlock(&inode
->i_lock
);
610 nfs_release_request(new);
616 nfs_lock_request_dontget(new);
617 error
= nfs_inode_add_request(inode
, new);
619 spin_unlock(&inode
->i_lock
);
620 nfs_unlock_request(new);
621 return ERR_PTR(error
);
623 spin_unlock(&inode
->i_lock
);
626 spin_unlock(&inode
->i_lock
);
628 new = nfs_create_request(ctx
, inode
, page
, offset
, bytes
);
633 /* We have a request for our page.
634 * If the creds don't match, or the
635 * page addresses don't match,
636 * tell the caller to wait on the conflicting
639 rqend
= req
->wb_offset
+ req
->wb_bytes
;
640 if (req
->wb_context
!= ctx
641 || req
->wb_page
!= page
642 || !nfs_dirty_request(req
)
643 || offset
> rqend
|| end
< req
->wb_offset
) {
644 nfs_unlock_request(req
);
645 return ERR_PTR(-EBUSY
);
648 /* Okay, the request matches. Update the region */
649 if (offset
< req
->wb_offset
) {
650 req
->wb_offset
= offset
;
651 req
->wb_pgbase
= offset
;
652 req
->wb_bytes
= rqend
- req
->wb_offset
;
656 req
->wb_bytes
= end
- req
->wb_offset
;
661 int nfs_flush_incompatible(struct file
*file
, struct page
*page
)
663 struct nfs_open_context
*ctx
= (struct nfs_open_context
*)file
->private_data
;
664 struct nfs_page
*req
;
665 int do_flush
, status
;
667 * Look for a request corresponding to this page. If there
668 * is one, and it belongs to another file, we flush it out
669 * before we try to copy anything into the page. Do this
670 * due to the lack of an ACCESS-type call in NFSv2.
671 * Also do the same if we find a request from an existing
675 req
= nfs_page_find_request(page
);
678 do_flush
= req
->wb_page
!= page
|| req
->wb_context
!= ctx
679 || !nfs_dirty_request(req
);
680 nfs_release_request(req
);
683 status
= nfs_wb_page(page
->mapping
->host
, page
);
684 } while (status
== 0);
689 * Update and possibly write a cached page of an NFS file.
691 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
692 * things with a page scheduled for an RPC call (e.g. invalidate it).
694 int nfs_updatepage(struct file
*file
, struct page
*page
,
695 unsigned int offset
, unsigned int count
)
697 struct nfs_open_context
*ctx
= (struct nfs_open_context
*)file
->private_data
;
698 struct inode
*inode
= page
->mapping
->host
;
701 nfs_inc_stats(inode
, NFSIOS_VFSUPDATEPAGE
);
703 dprintk("NFS: nfs_updatepage(%s/%s %d@%Ld)\n",
704 file
->f_path
.dentry
->d_parent
->d_name
.name
,
705 file
->f_path
.dentry
->d_name
.name
, count
,
706 (long long)(page_offset(page
) +offset
));
708 /* If we're not using byte range locks, and we know the page
709 * is entirely in cache, it may be more efficient to avoid
710 * fragmenting write requests.
712 if (PageUptodate(page
) && inode
->i_flock
== NULL
&& !(file
->f_mode
& O_SYNC
)) {
713 count
= max(count
+ offset
, nfs_page_length(page
));
717 status
= nfs_writepage_setup(ctx
, page
, offset
, count
);
718 __set_page_dirty_nobuffers(page
);
720 dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n",
721 status
, (long long)i_size_read(inode
));
723 nfs_set_pageerror(page
);
727 static void nfs_writepage_release(struct nfs_page
*req
)
730 if (PageError(req
->wb_page
)) {
731 nfs_end_page_writeback(req
->wb_page
);
732 nfs_inode_remove_request(req
);
733 } else if (!nfs_reschedule_unstable_write(req
)) {
734 /* Set the PG_uptodate flag */
735 nfs_mark_uptodate(req
->wb_page
, req
->wb_pgbase
, req
->wb_bytes
);
736 nfs_end_page_writeback(req
->wb_page
);
737 nfs_inode_remove_request(req
);
739 nfs_end_page_writeback(req
->wb_page
);
740 nfs_clear_page_tag_locked(req
);
743 static inline int flush_task_priority(int how
)
745 switch (how
& (FLUSH_HIGHPRI
|FLUSH_LOWPRI
)) {
747 return RPC_PRIORITY_HIGH
;
749 return RPC_PRIORITY_LOW
;
751 return RPC_PRIORITY_NORMAL
;
755 * Set up the argument/result storage required for the RPC call.
757 static void nfs_write_rpcsetup(struct nfs_page
*req
,
758 struct nfs_write_data
*data
,
759 const struct rpc_call_ops
*call_ops
,
760 unsigned int count
, unsigned int offset
,
766 /* Set up the RPC argument and reply structs
767 * NB: take care not to mess about with data->commit et al. */
770 data
->inode
= inode
= req
->wb_context
->path
.dentry
->d_inode
;
771 data
->cred
= req
->wb_context
->cred
;
773 data
->args
.fh
= NFS_FH(inode
);
774 data
->args
.offset
= req_offset(req
) + offset
;
775 data
->args
.pgbase
= req
->wb_pgbase
+ offset
;
776 data
->args
.pages
= data
->pagevec
;
777 data
->args
.count
= count
;
778 data
->args
.context
= req
->wb_context
;
780 data
->res
.fattr
= &data
->fattr
;
781 data
->res
.count
= count
;
782 data
->res
.verf
= &data
->verf
;
783 nfs_fattr_init(&data
->fattr
);
785 /* Set up the initial task struct. */
786 flags
= (how
& FLUSH_SYNC
) ? 0 : RPC_TASK_ASYNC
;
787 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), flags
, call_ops
, data
);
788 NFS_PROTO(inode
)->write_setup(data
, how
);
790 data
->task
.tk_priority
= flush_task_priority(how
);
791 data
->task
.tk_cookie
= (unsigned long)inode
;
793 dprintk("NFS: %5u initiated write call "
794 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
797 (long long)NFS_FILEID(inode
),
799 (unsigned long long)data
->args
.offset
);
802 static void nfs_execute_write(struct nfs_write_data
*data
)
804 struct rpc_clnt
*clnt
= NFS_CLIENT(data
->inode
);
807 rpc_clnt_sigmask(clnt
, &oldset
);
808 rpc_execute(&data
->task
);
809 rpc_clnt_sigunmask(clnt
, &oldset
);
813 * Generate multiple small requests to write out a single
814 * contiguous dirty area on one page.
816 static int nfs_flush_multi(struct inode
*inode
, struct list_head
*head
, unsigned int npages
, size_t count
, int how
)
818 struct nfs_page
*req
= nfs_list_entry(head
->next
);
819 struct page
*page
= req
->wb_page
;
820 struct nfs_write_data
*data
;
821 size_t wsize
= NFS_SERVER(inode
)->wsize
, nbytes
;
826 nfs_list_remove_request(req
);
830 size_t len
= min(nbytes
, wsize
);
832 data
= nfs_writedata_alloc(1);
835 list_add(&data
->pages
, &list
);
838 } while (nbytes
!= 0);
839 atomic_set(&req
->wb_complete
, requests
);
841 ClearPageError(page
);
845 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
846 list_del_init(&data
->pages
);
848 data
->pagevec
[0] = page
;
852 nfs_write_rpcsetup(req
, data
, &nfs_write_partial_ops
,
856 nfs_execute_write(data
);
857 } while (nbytes
!= 0);
862 while (!list_empty(&list
)) {
863 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
864 list_del(&data
->pages
);
865 nfs_writedata_release(data
);
867 nfs_redirty_request(req
);
868 nfs_end_page_writeback(req
->wb_page
);
869 nfs_clear_page_tag_locked(req
);
874 * Create an RPC task for the given write request and kick it.
875 * The page must have been locked by the caller.
877 * It may happen that the page we're passed is not marked dirty.
878 * This is the case if nfs_updatepage detects a conflicting request
879 * that has been written but not committed.
881 static int nfs_flush_one(struct inode
*inode
, struct list_head
*head
, unsigned int npages
, size_t count
, int how
)
883 struct nfs_page
*req
;
885 struct nfs_write_data
*data
;
887 data
= nfs_writedata_alloc(npages
);
891 pages
= data
->pagevec
;
892 while (!list_empty(head
)) {
893 req
= nfs_list_entry(head
->next
);
894 nfs_list_remove_request(req
);
895 nfs_list_add_request(req
, &data
->pages
);
896 ClearPageError(req
->wb_page
);
897 *pages
++ = req
->wb_page
;
899 req
= nfs_list_entry(data
->pages
.next
);
901 /* Set up the argument struct */
902 nfs_write_rpcsetup(req
, data
, &nfs_write_full_ops
, count
, 0, how
);
904 nfs_execute_write(data
);
907 while (!list_empty(head
)) {
908 req
= nfs_list_entry(head
->next
);
909 nfs_list_remove_request(req
);
910 nfs_redirty_request(req
);
911 nfs_end_page_writeback(req
->wb_page
);
912 nfs_clear_page_tag_locked(req
);
917 static void nfs_pageio_init_write(struct nfs_pageio_descriptor
*pgio
,
918 struct inode
*inode
, int ioflags
)
920 int wsize
= NFS_SERVER(inode
)->wsize
;
922 if (wsize
< PAGE_CACHE_SIZE
)
923 nfs_pageio_init(pgio
, inode
, nfs_flush_multi
, wsize
, ioflags
);
925 nfs_pageio_init(pgio
, inode
, nfs_flush_one
, wsize
, ioflags
);
929 * Handle a write reply that flushed part of a page.
931 static void nfs_writeback_done_partial(struct rpc_task
*task
, void *calldata
)
933 struct nfs_write_data
*data
= calldata
;
934 struct nfs_page
*req
= data
->req
;
935 struct page
*page
= req
->wb_page
;
937 dprintk("NFS: write (%s/%Ld %d@%Ld)",
938 req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
939 (long long)NFS_FILEID(req
->wb_context
->path
.dentry
->d_inode
),
941 (long long)req_offset(req
));
943 if (nfs_writeback_done(task
, data
) != 0)
946 if (task
->tk_status
< 0) {
947 nfs_set_pageerror(page
);
948 req
->wb_context
->error
= task
->tk_status
;
949 dprintk(", error = %d\n", task
->tk_status
);
953 if (nfs_write_need_commit(data
)) {
954 struct inode
*inode
= page
->mapping
->host
;
956 spin_lock(&inode
->i_lock
);
957 if (test_bit(PG_NEED_RESCHED
, &req
->wb_flags
)) {
958 /* Do nothing we need to resend the writes */
959 } else if (!test_and_set_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
960 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
961 dprintk(" defer commit\n");
962 } else if (memcmp(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
))) {
963 set_bit(PG_NEED_RESCHED
, &req
->wb_flags
);
964 clear_bit(PG_NEED_COMMIT
, &req
->wb_flags
);
965 dprintk(" server reboot detected\n");
967 spin_unlock(&inode
->i_lock
);
972 if (atomic_dec_and_test(&req
->wb_complete
))
973 nfs_writepage_release(req
);
976 static const struct rpc_call_ops nfs_write_partial_ops
= {
977 .rpc_call_done
= nfs_writeback_done_partial
,
978 .rpc_release
= nfs_writedata_release
,
982 * Handle a write reply that flushes a whole page.
984 * FIXME: There is an inherent race with invalidate_inode_pages and
985 * writebacks since the page->count is kept > 1 for as long
986 * as the page has a write request pending.
988 static void nfs_writeback_done_full(struct rpc_task
*task
, void *calldata
)
990 struct nfs_write_data
*data
= calldata
;
991 struct nfs_page
*req
;
994 if (nfs_writeback_done(task
, data
) != 0)
997 /* Update attributes as result of writeback. */
998 while (!list_empty(&data
->pages
)) {
999 req
= nfs_list_entry(data
->pages
.next
);
1000 nfs_list_remove_request(req
);
1001 page
= req
->wb_page
;
1003 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1004 req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
1005 (long long)NFS_FILEID(req
->wb_context
->path
.dentry
->d_inode
),
1007 (long long)req_offset(req
));
1009 if (task
->tk_status
< 0) {
1010 nfs_set_pageerror(page
);
1011 req
->wb_context
->error
= task
->tk_status
;
1012 dprintk(", error = %d\n", task
->tk_status
);
1013 goto remove_request
;
1016 if (nfs_write_need_commit(data
)) {
1017 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1018 nfs_mark_request_commit(req
);
1019 nfs_end_page_writeback(page
);
1020 dprintk(" marked for commit\n");
1023 /* Set the PG_uptodate flag? */
1024 nfs_mark_uptodate(page
, req
->wb_pgbase
, req
->wb_bytes
);
1027 nfs_end_page_writeback(page
);
1028 nfs_inode_remove_request(req
);
1030 nfs_clear_page_tag_locked(req
);
1034 static const struct rpc_call_ops nfs_write_full_ops
= {
1035 .rpc_call_done
= nfs_writeback_done_full
,
1036 .rpc_release
= nfs_writedata_release
,
1041 * This function is called when the WRITE call is complete.
1043 int nfs_writeback_done(struct rpc_task
*task
, struct nfs_write_data
*data
)
1045 struct nfs_writeargs
*argp
= &data
->args
;
1046 struct nfs_writeres
*resp
= &data
->res
;
1049 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1050 task
->tk_pid
, task
->tk_status
);
1053 * ->write_done will attempt to use post-op attributes to detect
1054 * conflicting writes by other clients. A strict interpretation
1055 * of close-to-open would allow us to continue caching even if
1056 * another writer had changed the file, but some applications
1057 * depend on tighter cache coherency when writing.
1059 status
= NFS_PROTO(data
->inode
)->write_done(task
, data
);
1062 nfs_add_stats(data
->inode
, NFSIOS_SERVERWRITTENBYTES
, resp
->count
);
1064 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1065 if (resp
->verf
->committed
< argp
->stable
&& task
->tk_status
>= 0) {
1066 /* We tried a write call, but the server did not
1067 * commit data to stable storage even though we
1069 * Note: There is a known bug in Tru64 < 5.0 in which
1070 * the server reports NFS_DATA_SYNC, but performs
1071 * NFS_FILE_SYNC. We therefore implement this checking
1072 * as a dprintk() in order to avoid filling syslog.
1074 static unsigned long complain
;
1076 if (time_before(complain
, jiffies
)) {
1077 dprintk("NFS: faulty NFS server %s:"
1078 " (committed = %d) != (stable = %d)\n",
1079 NFS_SERVER(data
->inode
)->nfs_client
->cl_hostname
,
1080 resp
->verf
->committed
, argp
->stable
);
1081 complain
= jiffies
+ 300 * HZ
;
1085 /* Is this a short write? */
1086 if (task
->tk_status
>= 0 && resp
->count
< argp
->count
) {
1087 static unsigned long complain
;
1089 nfs_inc_stats(data
->inode
, NFSIOS_SHORTWRITE
);
1091 /* Has the server at least made some progress? */
1092 if (resp
->count
!= 0) {
1093 /* Was this an NFSv2 write or an NFSv3 stable write? */
1094 if (resp
->verf
->committed
!= NFS_UNSTABLE
) {
1095 /* Resend from where the server left off */
1096 argp
->offset
+= resp
->count
;
1097 argp
->pgbase
+= resp
->count
;
1098 argp
->count
-= resp
->count
;
1100 /* Resend as a stable write in order to avoid
1101 * headaches in the case of a server crash.
1103 argp
->stable
= NFS_FILE_SYNC
;
1105 rpc_restart_call(task
);
1108 if (time_before(complain
, jiffies
)) {
1110 "NFS: Server wrote zero bytes, expected %u.\n",
1112 complain
= jiffies
+ 300 * HZ
;
1114 /* Can't do anything about it except throw an error. */
1115 task
->tk_status
= -EIO
;
1121 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1122 void nfs_commit_release(void *wdata
)
1124 nfs_commit_free(wdata
);
1128 * Set up the argument/result storage required for the RPC call.
1130 static void nfs_commit_rpcsetup(struct list_head
*head
,
1131 struct nfs_write_data
*data
,
1134 struct nfs_page
*first
;
1135 struct inode
*inode
;
1138 /* Set up the RPC argument and reply structs
1139 * NB: take care not to mess about with data->commit et al. */
1141 list_splice_init(head
, &data
->pages
);
1142 first
= nfs_list_entry(data
->pages
.next
);
1143 inode
= first
->wb_context
->path
.dentry
->d_inode
;
1145 data
->inode
= inode
;
1146 data
->cred
= first
->wb_context
->cred
;
1148 data
->args
.fh
= NFS_FH(data
->inode
);
1149 /* Note: we always request a commit of the entire inode */
1150 data
->args
.offset
= 0;
1151 data
->args
.count
= 0;
1152 data
->res
.count
= 0;
1153 data
->res
.fattr
= &data
->fattr
;
1154 data
->res
.verf
= &data
->verf
;
1155 nfs_fattr_init(&data
->fattr
);
1157 /* Set up the initial task struct. */
1158 flags
= (how
& FLUSH_SYNC
) ? 0 : RPC_TASK_ASYNC
;
1159 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), flags
, &nfs_commit_ops
, data
);
1160 NFS_PROTO(inode
)->commit_setup(data
, how
);
1162 data
->task
.tk_priority
= flush_task_priority(how
);
1163 data
->task
.tk_cookie
= (unsigned long)inode
;
1165 dprintk("NFS: %5u initiated commit call\n", data
->task
.tk_pid
);
1169 * Commit dirty pages
1172 nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
)
1174 struct nfs_write_data
*data
;
1175 struct nfs_page
*req
;
1177 data
= nfs_commit_alloc();
1182 /* Set up the argument struct */
1183 nfs_commit_rpcsetup(head
, data
, how
);
1185 nfs_execute_write(data
);
1188 while (!list_empty(head
)) {
1189 req
= nfs_list_entry(head
->next
);
1190 nfs_list_remove_request(req
);
1191 nfs_mark_request_commit(req
);
1192 dec_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
1193 nfs_clear_page_tag_locked(req
);
1199 * COMMIT call returned
1201 static void nfs_commit_done(struct rpc_task
*task
, void *calldata
)
1203 struct nfs_write_data
*data
= calldata
;
1204 struct nfs_page
*req
;
1206 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1207 task
->tk_pid
, task
->tk_status
);
1209 /* Call the NFS version-specific code */
1210 if (NFS_PROTO(data
->inode
)->commit_done(task
, data
) != 0)
1213 while (!list_empty(&data
->pages
)) {
1214 req
= nfs_list_entry(data
->pages
.next
);
1215 nfs_list_remove_request(req
);
1216 clear_bit(PG_NEED_COMMIT
, &(req
)->wb_flags
);
1217 dec_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
1219 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1220 req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
1221 (long long)NFS_FILEID(req
->wb_context
->path
.dentry
->d_inode
),
1223 (long long)req_offset(req
));
1224 if (task
->tk_status
< 0) {
1225 req
->wb_context
->error
= task
->tk_status
;
1226 nfs_inode_remove_request(req
);
1227 dprintk(", error = %d\n", task
->tk_status
);
1231 /* Okay, COMMIT succeeded, apparently. Check the verifier
1232 * returned by the server against all stored verfs. */
1233 if (!memcmp(req
->wb_verf
.verifier
, data
->verf
.verifier
, sizeof(data
->verf
.verifier
))) {
1234 /* We have a match */
1235 /* Set the PG_uptodate flag */
1236 nfs_mark_uptodate(req
->wb_page
, req
->wb_pgbase
,
1238 nfs_inode_remove_request(req
);
1242 /* We have a mismatch. Write the page again */
1243 dprintk(" mismatch\n");
1244 nfs_redirty_request(req
);
1246 nfs_clear_page_tag_locked(req
);
1250 static const struct rpc_call_ops nfs_commit_ops
= {
1251 .rpc_call_done
= nfs_commit_done
,
1252 .rpc_release
= nfs_commit_release
,
1255 int nfs_commit_inode(struct inode
*inode
, int how
)
1260 spin_lock(&inode
->i_lock
);
1261 res
= nfs_scan_commit(inode
, &head
, 0, 0);
1262 spin_unlock(&inode
->i_lock
);
1264 int error
= nfs_commit_list(inode
, &head
, how
);
1271 static inline int nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
)
1277 long nfs_sync_mapping_wait(struct address_space
*mapping
, struct writeback_control
*wbc
, int how
)
1279 struct inode
*inode
= mapping
->host
;
1280 pgoff_t idx_start
, idx_end
;
1281 unsigned int npages
= 0;
1283 int nocommit
= how
& FLUSH_NOCOMMIT
;
1287 if (wbc
->range_cyclic
)
1290 idx_start
= wbc
->range_start
>> PAGE_CACHE_SHIFT
;
1291 idx_end
= wbc
->range_end
>> PAGE_CACHE_SHIFT
;
1292 if (idx_end
> idx_start
) {
1293 pgoff_t l_npages
= 1 + idx_end
- idx_start
;
1295 if (sizeof(npages
) != sizeof(l_npages
) &&
1296 (pgoff_t
)npages
!= l_npages
)
1300 how
&= ~FLUSH_NOCOMMIT
;
1301 spin_lock(&inode
->i_lock
);
1303 ret
= nfs_wait_on_requests_locked(inode
, idx_start
, npages
);
1308 pages
= nfs_scan_commit(inode
, &head
, idx_start
, npages
);
1311 if (how
& FLUSH_INVALIDATE
) {
1312 spin_unlock(&inode
->i_lock
);
1313 nfs_cancel_commit_list(&head
);
1315 spin_lock(&inode
->i_lock
);
1318 pages
+= nfs_scan_commit(inode
, &head
, 0, 0);
1319 spin_unlock(&inode
->i_lock
);
1320 ret
= nfs_commit_list(inode
, &head
, how
);
1321 spin_lock(&inode
->i_lock
);
1324 spin_unlock(&inode
->i_lock
);
1328 static int nfs_write_mapping(struct address_space
*mapping
, int how
)
1330 struct writeback_control wbc
= {
1331 .bdi
= mapping
->backing_dev_info
,
1332 .sync_mode
= WB_SYNC_ALL
,
1333 .nr_to_write
= LONG_MAX
,
1334 .for_writepages
= 1,
1339 ret
= nfs_writepages(mapping
, &wbc
);
1342 ret
= nfs_sync_mapping_wait(mapping
, &wbc
, how
);
1347 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
1352 * flush the inode to disk.
1354 int nfs_wb_all(struct inode
*inode
)
1356 return nfs_write_mapping(inode
->i_mapping
, 0);
1359 int nfs_wb_nocommit(struct inode
*inode
)
1361 return nfs_write_mapping(inode
->i_mapping
, FLUSH_NOCOMMIT
);
1364 int nfs_wb_page_cancel(struct inode
*inode
, struct page
*page
)
1366 struct nfs_page
*req
;
1367 loff_t range_start
= page_offset(page
);
1368 loff_t range_end
= range_start
+ (loff_t
)(PAGE_CACHE_SIZE
- 1);
1369 struct writeback_control wbc
= {
1370 .bdi
= page
->mapping
->backing_dev_info
,
1371 .sync_mode
= WB_SYNC_ALL
,
1372 .nr_to_write
= LONG_MAX
,
1373 .range_start
= range_start
,
1374 .range_end
= range_end
,
1378 BUG_ON(!PageLocked(page
));
1380 req
= nfs_page_find_request(page
);
1383 if (test_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
1384 nfs_release_request(req
);
1387 if (nfs_lock_request_dontget(req
)) {
1388 nfs_inode_remove_request(req
);
1390 * In case nfs_inode_remove_request has marked the
1391 * page as being dirty
1393 cancel_dirty_page(page
, PAGE_CACHE_SIZE
);
1394 nfs_unlock_request(req
);
1397 ret
= nfs_wait_on_request(req
);
1401 if (!PagePrivate(page
))
1403 ret
= nfs_sync_mapping_wait(page
->mapping
, &wbc
, FLUSH_INVALIDATE
);
1408 int nfs_wb_page_priority(struct inode
*inode
, struct page
*page
, int how
)
1410 loff_t range_start
= page_offset(page
);
1411 loff_t range_end
= range_start
+ (loff_t
)(PAGE_CACHE_SIZE
- 1);
1412 struct writeback_control wbc
= {
1413 .bdi
= page
->mapping
->backing_dev_info
,
1414 .sync_mode
= WB_SYNC_ALL
,
1415 .nr_to_write
= LONG_MAX
,
1416 .range_start
= range_start
,
1417 .range_end
= range_end
,
1421 BUG_ON(!PageLocked(page
));
1422 if (clear_page_dirty_for_io(page
)) {
1423 ret
= nfs_writepage_locked(page
, &wbc
);
1427 if (!PagePrivate(page
))
1429 ret
= nfs_sync_mapping_wait(page
->mapping
, &wbc
, how
);
1433 __mark_inode_dirty(inode
, I_DIRTY_PAGES
);
1438 * Write back all requests on one page - we do this before reading it.
1440 int nfs_wb_page(struct inode
*inode
, struct page
* page
)
1442 return nfs_wb_page_priority(inode
, page
, FLUSH_STABLE
);
1445 int __init
nfs_init_writepagecache(void)
1447 nfs_wdata_cachep
= kmem_cache_create("nfs_write_data",
1448 sizeof(struct nfs_write_data
),
1449 0, SLAB_HWCACHE_ALIGN
,
1451 if (nfs_wdata_cachep
== NULL
)
1454 nfs_wdata_mempool
= mempool_create_slab_pool(MIN_POOL_WRITE
,
1456 if (nfs_wdata_mempool
== NULL
)
1459 nfs_commit_mempool
= mempool_create_slab_pool(MIN_POOL_COMMIT
,
1461 if (nfs_commit_mempool
== NULL
)
1465 * NFS congestion size, scale with available memory.
1477 * This allows larger machines to have larger/more transfers.
1478 * Limit the default to 256M
1480 nfs_congestion_kb
= (16*int_sqrt(totalram_pages
)) << (PAGE_SHIFT
-10);
1481 if (nfs_congestion_kb
> 256*1024)
1482 nfs_congestion_kb
= 256*1024;
1487 void nfs_destroy_writepagecache(void)
1489 mempool_destroy(nfs_commit_mempool
);
1490 mempool_destroy(nfs_wdata_mempool
);
1491 kmem_cache_destroy(nfs_wdata_cachep
);