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 void nfs_context_set_write_error(struct nfs_open_context
*ctx
, int error
)
117 set_bit(NFS_CONTEXT_ERROR_WRITE
, &ctx
->flags
);
120 static struct nfs_page
*nfs_page_find_request_locked(struct page
*page
)
122 struct nfs_page
*req
= NULL
;
124 if (PagePrivate(page
)) {
125 req
= (struct nfs_page
*)page_private(page
);
127 kref_get(&req
->wb_kref
);
132 static struct nfs_page
*nfs_page_find_request(struct page
*page
)
134 struct inode
*inode
= page
->mapping
->host
;
135 struct nfs_page
*req
= NULL
;
137 spin_lock(&inode
->i_lock
);
138 req
= nfs_page_find_request_locked(page
);
139 spin_unlock(&inode
->i_lock
);
143 /* Adjust the file length if we're writing beyond the end */
144 static void nfs_grow_file(struct page
*page
, unsigned int offset
, unsigned int count
)
146 struct inode
*inode
= page
->mapping
->host
;
147 loff_t end
, i_size
= i_size_read(inode
);
148 pgoff_t end_index
= (i_size
- 1) >> PAGE_CACHE_SHIFT
;
150 if (i_size
> 0 && page
->index
< end_index
)
152 end
= ((loff_t
)page
->index
<< PAGE_CACHE_SHIFT
) + ((loff_t
)offset
+count
);
155 nfs_inc_stats(inode
, NFSIOS_EXTENDWRITE
);
156 i_size_write(inode
, end
);
159 /* A writeback failed: mark the page as bad, and invalidate the page cache */
160 static void nfs_set_pageerror(struct page
*page
)
163 nfs_zap_mapping(page
->mapping
->host
, page
->mapping
);
166 /* We can set the PG_uptodate flag if we see that a write request
167 * covers the full page.
169 static void nfs_mark_uptodate(struct page
*page
, unsigned int base
, unsigned int count
)
171 if (PageUptodate(page
))
175 if (count
!= nfs_page_length(page
))
177 SetPageUptodate(page
);
180 static int nfs_writepage_setup(struct nfs_open_context
*ctx
, struct page
*page
,
181 unsigned int offset
, unsigned int count
)
183 struct nfs_page
*req
;
187 req
= nfs_update_request(ctx
, page
, offset
, count
);
193 ret
= nfs_wb_page(page
->mapping
->host
, page
);
197 /* Update file length */
198 nfs_grow_file(page
, offset
, count
);
199 nfs_unlock_request(req
);
203 static int wb_priority(struct writeback_control
*wbc
)
205 if (wbc
->for_reclaim
)
206 return FLUSH_HIGHPRI
| FLUSH_STABLE
;
207 if (wbc
->for_kupdate
)
213 * NFS congestion control
216 int nfs_congestion_kb
;
218 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
219 #define NFS_CONGESTION_OFF_THRESH \
220 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
222 static int nfs_set_page_writeback(struct page
*page
)
224 int ret
= test_set_page_writeback(page
);
227 struct inode
*inode
= page
->mapping
->host
;
228 struct nfs_server
*nfss
= NFS_SERVER(inode
);
230 if (atomic_long_inc_return(&nfss
->writeback
) >
231 NFS_CONGESTION_ON_THRESH
)
232 set_bdi_congested(&nfss
->backing_dev_info
, WRITE
);
237 static void nfs_end_page_writeback(struct page
*page
)
239 struct inode
*inode
= page
->mapping
->host
;
240 struct nfs_server
*nfss
= NFS_SERVER(inode
);
242 end_page_writeback(page
);
243 if (atomic_long_dec_return(&nfss
->writeback
) < NFS_CONGESTION_OFF_THRESH
)
244 clear_bdi_congested(&nfss
->backing_dev_info
, WRITE
);
248 * Find an associated nfs write request, and prepare to flush it out
249 * May return an error if the user signalled nfs_wait_on_request().
251 static int nfs_page_async_flush(struct nfs_pageio_descriptor
*pgio
,
254 struct inode
*inode
= page
->mapping
->host
;
255 struct nfs_inode
*nfsi
= NFS_I(inode
);
256 struct nfs_page
*req
;
259 spin_lock(&inode
->i_lock
);
261 req
= nfs_page_find_request_locked(page
);
263 spin_unlock(&inode
->i_lock
);
266 if (nfs_lock_request_dontget(req
))
268 /* Note: If we hold the page lock, as is the case in nfs_writepage,
269 * then the call to nfs_lock_request_dontget() will always
270 * succeed provided that someone hasn't already marked the
271 * request as dirty (in which case we don't care).
273 spin_unlock(&inode
->i_lock
);
274 ret
= nfs_wait_on_request(req
);
275 nfs_release_request(req
);
278 spin_lock(&inode
->i_lock
);
280 if (test_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
281 /* This request is marked for commit */
282 spin_unlock(&inode
->i_lock
);
283 nfs_unlock_request(req
);
284 nfs_pageio_complete(pgio
);
287 if (nfs_set_page_writeback(page
) != 0) {
288 spin_unlock(&inode
->i_lock
);
291 radix_tree_tag_set(&nfsi
->nfs_page_tree
, req
->wb_index
,
292 NFS_PAGE_TAG_LOCKED
);
293 spin_unlock(&inode
->i_lock
);
294 nfs_pageio_add_request(pgio
, req
);
298 static int nfs_do_writepage(struct page
*page
, struct writeback_control
*wbc
, struct nfs_pageio_descriptor
*pgio
)
300 struct inode
*inode
= page
->mapping
->host
;
302 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGE
);
303 nfs_add_stats(inode
, NFSIOS_WRITEPAGES
, 1);
305 nfs_pageio_cond_complete(pgio
, page
->index
);
306 return nfs_page_async_flush(pgio
, page
);
310 * Write an mmapped page to the server.
312 static int nfs_writepage_locked(struct page
*page
, struct writeback_control
*wbc
)
314 struct nfs_pageio_descriptor pgio
;
317 nfs_pageio_init_write(&pgio
, page
->mapping
->host
, wb_priority(wbc
));
318 err
= nfs_do_writepage(page
, wbc
, &pgio
);
319 nfs_pageio_complete(&pgio
);
322 if (pgio
.pg_error
< 0)
323 return pgio
.pg_error
;
327 int nfs_writepage(struct page
*page
, struct writeback_control
*wbc
)
331 ret
= nfs_writepage_locked(page
, wbc
);
336 static int nfs_writepages_callback(struct page
*page
, struct writeback_control
*wbc
, void *data
)
340 ret
= nfs_do_writepage(page
, wbc
, data
);
345 int nfs_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
347 struct inode
*inode
= mapping
->host
;
348 struct nfs_pageio_descriptor pgio
;
351 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGES
);
353 nfs_pageio_init_write(&pgio
, inode
, wb_priority(wbc
));
354 err
= write_cache_pages(mapping
, wbc
, nfs_writepages_callback
, &pgio
);
355 nfs_pageio_complete(&pgio
);
358 if (pgio
.pg_error
< 0)
359 return pgio
.pg_error
;
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_insert(&nfsi
->nfs_page_tree
, req
->wb_index
, req
);
372 BUG_ON(error
== -EEXIST
);
377 if (nfs_have_delegation(inode
, FMODE_WRITE
))
380 SetPagePrivate(req
->wb_page
);
381 set_page_private(req
->wb_page
, (unsigned long)req
);
383 kref_get(&req
->wb_kref
);
388 * Remove a write request from an inode
390 static void nfs_inode_remove_request(struct nfs_page
*req
)
392 struct inode
*inode
= req
->wb_context
->path
.dentry
->d_inode
;
393 struct nfs_inode
*nfsi
= NFS_I(inode
);
395 BUG_ON (!NFS_WBACK_BUSY(req
));
397 spin_lock(&inode
->i_lock
);
398 set_page_private(req
->wb_page
, 0);
399 ClearPagePrivate(req
->wb_page
);
400 radix_tree_delete(&nfsi
->nfs_page_tree
, req
->wb_index
);
403 spin_unlock(&inode
->i_lock
);
406 spin_unlock(&inode
->i_lock
);
407 nfs_clear_request(req
);
408 nfs_release_request(req
);
412 nfs_redirty_request(struct nfs_page
*req
)
414 __set_page_dirty_nobuffers(req
->wb_page
);
418 * Check if a request is dirty
421 nfs_dirty_request(struct nfs_page
*req
)
423 struct page
*page
= req
->wb_page
;
425 if (page
== NULL
|| test_bit(PG_NEED_COMMIT
, &req
->wb_flags
))
427 return !PageWriteback(req
->wb_page
);
430 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
432 * Add a request to the inode's commit list.
435 nfs_mark_request_commit(struct nfs_page
*req
)
437 struct inode
*inode
= req
->wb_context
->path
.dentry
->d_inode
;
438 struct nfs_inode
*nfsi
= NFS_I(inode
);
440 spin_lock(&inode
->i_lock
);
442 set_bit(PG_NEED_COMMIT
, &(req
)->wb_flags
);
443 radix_tree_tag_set(&nfsi
->nfs_page_tree
,
445 NFS_PAGE_TAG_COMMIT
);
446 spin_unlock(&inode
->i_lock
);
447 inc_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
448 inc_bdi_stat(req
->wb_page
->mapping
->backing_dev_info
, BDI_RECLAIMABLE
);
449 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
453 int nfs_write_need_commit(struct nfs_write_data
*data
)
455 return data
->verf
.committed
!= NFS_FILE_SYNC
;
459 int nfs_reschedule_unstable_write(struct nfs_page
*req
)
461 if (test_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
462 nfs_mark_request_commit(req
);
465 if (test_and_clear_bit(PG_NEED_RESCHED
, &req
->wb_flags
)) {
466 nfs_redirty_request(req
);
473 nfs_mark_request_commit(struct nfs_page
*req
)
478 int nfs_write_need_commit(struct nfs_write_data
*data
)
484 int nfs_reschedule_unstable_write(struct nfs_page
*req
)
491 * Wait for a request to complete.
493 * Interruptible by fatal signals only.
495 static int nfs_wait_on_requests_locked(struct inode
*inode
, pgoff_t idx_start
, unsigned int npages
)
497 struct nfs_inode
*nfsi
= NFS_I(inode
);
498 struct nfs_page
*req
;
499 pgoff_t idx_end
, next
;
500 unsigned int res
= 0;
506 idx_end
= idx_start
+ npages
- 1;
509 while (radix_tree_gang_lookup_tag(&nfsi
->nfs_page_tree
, (void **)&req
, next
, 1, NFS_PAGE_TAG_LOCKED
)) {
510 if (req
->wb_index
> idx_end
)
513 next
= req
->wb_index
+ 1;
514 BUG_ON(!NFS_WBACK_BUSY(req
));
516 kref_get(&req
->wb_kref
);
517 spin_unlock(&inode
->i_lock
);
518 error
= nfs_wait_on_request(req
);
519 nfs_release_request(req
);
520 spin_lock(&inode
->i_lock
);
528 static void nfs_cancel_commit_list(struct list_head
*head
)
530 struct nfs_page
*req
;
532 while(!list_empty(head
)) {
533 req
= nfs_list_entry(head
->next
);
534 dec_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
535 dec_bdi_stat(req
->wb_page
->mapping
->backing_dev_info
,
537 nfs_list_remove_request(req
);
538 clear_bit(PG_NEED_COMMIT
, &(req
)->wb_flags
);
539 nfs_inode_remove_request(req
);
540 nfs_unlock_request(req
);
544 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
546 * nfs_scan_commit - Scan an inode for commit requests
547 * @inode: NFS inode to scan
548 * @dst: destination list
549 * @idx_start: lower bound of page->index to scan.
550 * @npages: idx_start + npages sets the upper bound to scan.
552 * Moves requests from the inode's 'commit' request list.
553 * The requests are *not* checked to ensure that they form a contiguous set.
556 nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, pgoff_t idx_start
, unsigned int npages
)
558 struct nfs_inode
*nfsi
= NFS_I(inode
);
561 if (nfsi
->ncommit
!= 0) {
562 res
= nfs_scan_list(nfsi
, dst
, idx_start
, npages
,
563 NFS_PAGE_TAG_COMMIT
);
564 nfsi
->ncommit
-= res
;
569 static inline int nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, pgoff_t idx_start
, unsigned int npages
)
576 * Try to update any existing write request, or create one if there is none.
577 * In order to match, the request's credentials must match those of
578 * the calling process.
580 * Note: Should always be called with the Page Lock held!
582 static struct nfs_page
* nfs_update_request(struct nfs_open_context
* ctx
,
583 struct page
*page
, unsigned int offset
, unsigned int bytes
)
585 struct address_space
*mapping
= page
->mapping
;
586 struct inode
*inode
= mapping
->host
;
587 struct nfs_page
*req
, *new = NULL
;
590 end
= offset
+ bytes
;
593 /* Loop over all inode entries and see if we find
594 * A request for the page we wish to update
596 spin_lock(&inode
->i_lock
);
597 req
= nfs_page_find_request_locked(page
);
599 if (!nfs_lock_request_dontget(req
)) {
602 spin_unlock(&inode
->i_lock
);
603 error
= nfs_wait_on_request(req
);
604 nfs_release_request(req
);
607 nfs_release_request(new);
608 return ERR_PTR(error
);
612 spin_unlock(&inode
->i_lock
);
614 nfs_release_request(new);
620 nfs_lock_request_dontget(new);
621 error
= nfs_inode_add_request(inode
, new);
623 spin_unlock(&inode
->i_lock
);
624 nfs_unlock_request(new);
625 return ERR_PTR(error
);
627 spin_unlock(&inode
->i_lock
);
631 spin_unlock(&inode
->i_lock
);
633 new = nfs_create_request(ctx
, inode
, page
, offset
, bytes
);
638 /* We have a request for our page.
639 * If the creds don't match, or the
640 * page addresses don't match,
641 * tell the caller to wait on the conflicting
644 rqend
= req
->wb_offset
+ req
->wb_bytes
;
645 if (req
->wb_context
!= ctx
646 || req
->wb_page
!= page
647 || !nfs_dirty_request(req
)
648 || offset
> rqend
|| end
< req
->wb_offset
) {
649 nfs_unlock_request(req
);
650 return ERR_PTR(-EBUSY
);
653 /* Okay, the request matches. Update the region */
654 if (offset
< req
->wb_offset
) {
655 req
->wb_offset
= offset
;
656 req
->wb_pgbase
= offset
;
657 req
->wb_bytes
= max(end
, rqend
) - req
->wb_offset
;
662 req
->wb_bytes
= end
- req
->wb_offset
;
666 /* If this page might potentially be marked as up to date,
667 * then we need to zero any uninitalised data. */
668 if (req
->wb_pgbase
== 0 && req
->wb_bytes
!= PAGE_CACHE_SIZE
669 && !PageUptodate(req
->wb_page
))
670 zero_user_page(req
->wb_page
, req
->wb_bytes
,
671 PAGE_CACHE_SIZE
- req
->wb_bytes
,
676 int nfs_flush_incompatible(struct file
*file
, struct page
*page
)
678 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
679 struct nfs_page
*req
;
680 int do_flush
, status
;
682 * Look for a request corresponding to this page. If there
683 * is one, and it belongs to another file, we flush it out
684 * before we try to copy anything into the page. Do this
685 * due to the lack of an ACCESS-type call in NFSv2.
686 * Also do the same if we find a request from an existing
690 req
= nfs_page_find_request(page
);
693 do_flush
= req
->wb_page
!= page
|| req
->wb_context
!= ctx
694 || !nfs_dirty_request(req
);
695 nfs_release_request(req
);
698 status
= nfs_wb_page(page
->mapping
->host
, page
);
699 } while (status
== 0);
704 * Update and possibly write a cached page of an NFS file.
706 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
707 * things with a page scheduled for an RPC call (e.g. invalidate it).
709 int nfs_updatepage(struct file
*file
, struct page
*page
,
710 unsigned int offset
, unsigned int count
)
712 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
713 struct inode
*inode
= page
->mapping
->host
;
716 nfs_inc_stats(inode
, NFSIOS_VFSUPDATEPAGE
);
718 dprintk("NFS: nfs_updatepage(%s/%s %d@%Ld)\n",
719 file
->f_path
.dentry
->d_parent
->d_name
.name
,
720 file
->f_path
.dentry
->d_name
.name
, count
,
721 (long long)(page_offset(page
) +offset
));
723 /* If we're not using byte range locks, and we know the page
724 * is entirely in cache, it may be more efficient to avoid
725 * fragmenting write requests.
727 if (PageUptodate(page
) && inode
->i_flock
== NULL
&& !(file
->f_mode
& O_SYNC
)) {
728 count
= max(count
+ offset
, nfs_page_length(page
));
732 status
= nfs_writepage_setup(ctx
, page
, offset
, count
);
733 __set_page_dirty_nobuffers(page
);
735 dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n",
736 status
, (long long)i_size_read(inode
));
738 nfs_set_pageerror(page
);
742 static void nfs_writepage_release(struct nfs_page
*req
)
745 if (PageError(req
->wb_page
)) {
746 nfs_end_page_writeback(req
->wb_page
);
747 nfs_inode_remove_request(req
);
748 } else if (!nfs_reschedule_unstable_write(req
)) {
749 /* Set the PG_uptodate flag */
750 nfs_mark_uptodate(req
->wb_page
, req
->wb_pgbase
, req
->wb_bytes
);
751 nfs_end_page_writeback(req
->wb_page
);
752 nfs_inode_remove_request(req
);
754 nfs_end_page_writeback(req
->wb_page
);
755 nfs_clear_page_tag_locked(req
);
758 static inline int flush_task_priority(int how
)
760 switch (how
& (FLUSH_HIGHPRI
|FLUSH_LOWPRI
)) {
762 return RPC_PRIORITY_HIGH
;
764 return RPC_PRIORITY_LOW
;
766 return RPC_PRIORITY_NORMAL
;
770 * Set up the argument/result storage required for the RPC call.
772 static void nfs_write_rpcsetup(struct nfs_page
*req
,
773 struct nfs_write_data
*data
,
774 const struct rpc_call_ops
*call_ops
,
775 unsigned int count
, unsigned int offset
,
781 /* Set up the RPC argument and reply structs
782 * NB: take care not to mess about with data->commit et al. */
785 data
->inode
= inode
= req
->wb_context
->path
.dentry
->d_inode
;
786 data
->cred
= req
->wb_context
->cred
;
788 data
->args
.fh
= NFS_FH(inode
);
789 data
->args
.offset
= req_offset(req
) + offset
;
790 data
->args
.pgbase
= req
->wb_pgbase
+ offset
;
791 data
->args
.pages
= data
->pagevec
;
792 data
->args
.count
= count
;
793 data
->args
.context
= req
->wb_context
;
795 data
->res
.fattr
= &data
->fattr
;
796 data
->res
.count
= count
;
797 data
->res
.verf
= &data
->verf
;
798 nfs_fattr_init(&data
->fattr
);
800 /* Set up the initial task struct. */
801 flags
= (how
& FLUSH_SYNC
) ? 0 : RPC_TASK_ASYNC
;
802 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), flags
, call_ops
, data
);
803 NFS_PROTO(inode
)->write_setup(data
, how
);
805 data
->task
.tk_priority
= flush_task_priority(how
);
806 data
->task
.tk_cookie
= (unsigned long)inode
;
808 dprintk("NFS: %5u initiated write call "
809 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
812 (long long)NFS_FILEID(inode
),
814 (unsigned long long)data
->args
.offset
);
817 static void nfs_execute_write(struct nfs_write_data
*data
)
819 rpc_execute(&data
->task
);
823 * Generate multiple small requests to write out a single
824 * contiguous dirty area on one page.
826 static int nfs_flush_multi(struct inode
*inode
, struct list_head
*head
, unsigned int npages
, size_t count
, int how
)
828 struct nfs_page
*req
= nfs_list_entry(head
->next
);
829 struct page
*page
= req
->wb_page
;
830 struct nfs_write_data
*data
;
831 size_t wsize
= NFS_SERVER(inode
)->wsize
, nbytes
;
836 nfs_list_remove_request(req
);
840 size_t len
= min(nbytes
, wsize
);
842 data
= nfs_writedata_alloc(1);
845 list_add(&data
->pages
, &list
);
848 } while (nbytes
!= 0);
849 atomic_set(&req
->wb_complete
, requests
);
851 ClearPageError(page
);
855 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
856 list_del_init(&data
->pages
);
858 data
->pagevec
[0] = page
;
862 nfs_write_rpcsetup(req
, data
, &nfs_write_partial_ops
,
866 nfs_execute_write(data
);
867 } while (nbytes
!= 0);
872 while (!list_empty(&list
)) {
873 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
874 list_del(&data
->pages
);
875 nfs_writedata_release(data
);
877 nfs_redirty_request(req
);
878 nfs_end_page_writeback(req
->wb_page
);
879 nfs_clear_page_tag_locked(req
);
884 * Create an RPC task for the given write request and kick it.
885 * The page must have been locked by the caller.
887 * It may happen that the page we're passed is not marked dirty.
888 * This is the case if nfs_updatepage detects a conflicting request
889 * that has been written but not committed.
891 static int nfs_flush_one(struct inode
*inode
, struct list_head
*head
, unsigned int npages
, size_t count
, int how
)
893 struct nfs_page
*req
;
895 struct nfs_write_data
*data
;
897 data
= nfs_writedata_alloc(npages
);
901 pages
= data
->pagevec
;
902 while (!list_empty(head
)) {
903 req
= nfs_list_entry(head
->next
);
904 nfs_list_remove_request(req
);
905 nfs_list_add_request(req
, &data
->pages
);
906 ClearPageError(req
->wb_page
);
907 *pages
++ = req
->wb_page
;
909 req
= nfs_list_entry(data
->pages
.next
);
911 /* Set up the argument struct */
912 nfs_write_rpcsetup(req
, data
, &nfs_write_full_ops
, count
, 0, how
);
914 nfs_execute_write(data
);
917 while (!list_empty(head
)) {
918 req
= nfs_list_entry(head
->next
);
919 nfs_list_remove_request(req
);
920 nfs_redirty_request(req
);
921 nfs_end_page_writeback(req
->wb_page
);
922 nfs_clear_page_tag_locked(req
);
927 static void nfs_pageio_init_write(struct nfs_pageio_descriptor
*pgio
,
928 struct inode
*inode
, int ioflags
)
930 int wsize
= NFS_SERVER(inode
)->wsize
;
932 if (wsize
< PAGE_CACHE_SIZE
)
933 nfs_pageio_init(pgio
, inode
, nfs_flush_multi
, wsize
, ioflags
);
935 nfs_pageio_init(pgio
, inode
, nfs_flush_one
, wsize
, ioflags
);
939 * Handle a write reply that flushed part of a page.
941 static void nfs_writeback_done_partial(struct rpc_task
*task
, void *calldata
)
943 struct nfs_write_data
*data
= calldata
;
944 struct nfs_page
*req
= data
->req
;
945 struct page
*page
= req
->wb_page
;
947 dprintk("NFS: write (%s/%Ld %d@%Ld)",
948 req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
949 (long long)NFS_FILEID(req
->wb_context
->path
.dentry
->d_inode
),
951 (long long)req_offset(req
));
953 if (nfs_writeback_done(task
, data
) != 0)
956 if (task
->tk_status
< 0) {
957 nfs_set_pageerror(page
);
958 nfs_context_set_write_error(req
->wb_context
, task
->tk_status
);
959 dprintk(", error = %d\n", task
->tk_status
);
963 if (nfs_write_need_commit(data
)) {
964 struct inode
*inode
= page
->mapping
->host
;
966 spin_lock(&inode
->i_lock
);
967 if (test_bit(PG_NEED_RESCHED
, &req
->wb_flags
)) {
968 /* Do nothing we need to resend the writes */
969 } else if (!test_and_set_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
970 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
971 dprintk(" defer commit\n");
972 } else if (memcmp(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
))) {
973 set_bit(PG_NEED_RESCHED
, &req
->wb_flags
);
974 clear_bit(PG_NEED_COMMIT
, &req
->wb_flags
);
975 dprintk(" server reboot detected\n");
977 spin_unlock(&inode
->i_lock
);
982 if (atomic_dec_and_test(&req
->wb_complete
))
983 nfs_writepage_release(req
);
986 static const struct rpc_call_ops nfs_write_partial_ops
= {
987 .rpc_call_done
= nfs_writeback_done_partial
,
988 .rpc_release
= nfs_writedata_release
,
992 * Handle a write reply that flushes a whole page.
994 * FIXME: There is an inherent race with invalidate_inode_pages and
995 * writebacks since the page->count is kept > 1 for as long
996 * as the page has a write request pending.
998 static void nfs_writeback_done_full(struct rpc_task
*task
, void *calldata
)
1000 struct nfs_write_data
*data
= calldata
;
1001 struct nfs_page
*req
;
1004 if (nfs_writeback_done(task
, data
) != 0)
1007 /* Update attributes as result of writeback. */
1008 while (!list_empty(&data
->pages
)) {
1009 req
= nfs_list_entry(data
->pages
.next
);
1010 nfs_list_remove_request(req
);
1011 page
= req
->wb_page
;
1013 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1014 req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
1015 (long long)NFS_FILEID(req
->wb_context
->path
.dentry
->d_inode
),
1017 (long long)req_offset(req
));
1019 if (task
->tk_status
< 0) {
1020 nfs_set_pageerror(page
);
1021 nfs_context_set_write_error(req
->wb_context
, task
->tk_status
);
1022 dprintk(", error = %d\n", task
->tk_status
);
1023 goto remove_request
;
1026 if (nfs_write_need_commit(data
)) {
1027 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1028 nfs_mark_request_commit(req
);
1029 nfs_end_page_writeback(page
);
1030 dprintk(" marked for commit\n");
1033 /* Set the PG_uptodate flag? */
1034 nfs_mark_uptodate(page
, req
->wb_pgbase
, req
->wb_bytes
);
1037 nfs_end_page_writeback(page
);
1038 nfs_inode_remove_request(req
);
1040 nfs_clear_page_tag_locked(req
);
1044 static const struct rpc_call_ops nfs_write_full_ops
= {
1045 .rpc_call_done
= nfs_writeback_done_full
,
1046 .rpc_release
= nfs_writedata_release
,
1051 * This function is called when the WRITE call is complete.
1053 int nfs_writeback_done(struct rpc_task
*task
, struct nfs_write_data
*data
)
1055 struct nfs_writeargs
*argp
= &data
->args
;
1056 struct nfs_writeres
*resp
= &data
->res
;
1059 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1060 task
->tk_pid
, task
->tk_status
);
1063 * ->write_done will attempt to use post-op attributes to detect
1064 * conflicting writes by other clients. A strict interpretation
1065 * of close-to-open would allow us to continue caching even if
1066 * another writer had changed the file, but some applications
1067 * depend on tighter cache coherency when writing.
1069 status
= NFS_PROTO(data
->inode
)->write_done(task
, data
);
1072 nfs_add_stats(data
->inode
, NFSIOS_SERVERWRITTENBYTES
, resp
->count
);
1074 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1075 if (resp
->verf
->committed
< argp
->stable
&& task
->tk_status
>= 0) {
1076 /* We tried a write call, but the server did not
1077 * commit data to stable storage even though we
1079 * Note: There is a known bug in Tru64 < 5.0 in which
1080 * the server reports NFS_DATA_SYNC, but performs
1081 * NFS_FILE_SYNC. We therefore implement this checking
1082 * as a dprintk() in order to avoid filling syslog.
1084 static unsigned long complain
;
1086 if (time_before(complain
, jiffies
)) {
1087 dprintk("NFS: faulty NFS server %s:"
1088 " (committed = %d) != (stable = %d)\n",
1089 NFS_SERVER(data
->inode
)->nfs_client
->cl_hostname
,
1090 resp
->verf
->committed
, argp
->stable
);
1091 complain
= jiffies
+ 300 * HZ
;
1095 /* Is this a short write? */
1096 if (task
->tk_status
>= 0 && resp
->count
< argp
->count
) {
1097 static unsigned long complain
;
1099 nfs_inc_stats(data
->inode
, NFSIOS_SHORTWRITE
);
1101 /* Has the server at least made some progress? */
1102 if (resp
->count
!= 0) {
1103 /* Was this an NFSv2 write or an NFSv3 stable write? */
1104 if (resp
->verf
->committed
!= NFS_UNSTABLE
) {
1105 /* Resend from where the server left off */
1106 argp
->offset
+= resp
->count
;
1107 argp
->pgbase
+= resp
->count
;
1108 argp
->count
-= resp
->count
;
1110 /* Resend as a stable write in order to avoid
1111 * headaches in the case of a server crash.
1113 argp
->stable
= NFS_FILE_SYNC
;
1115 rpc_restart_call(task
);
1118 if (time_before(complain
, jiffies
)) {
1120 "NFS: Server wrote zero bytes, expected %u.\n",
1122 complain
= jiffies
+ 300 * HZ
;
1124 /* Can't do anything about it except throw an error. */
1125 task
->tk_status
= -EIO
;
1131 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1132 void nfs_commit_release(void *wdata
)
1134 nfs_commit_free(wdata
);
1138 * Set up the argument/result storage required for the RPC call.
1140 static void nfs_commit_rpcsetup(struct list_head
*head
,
1141 struct nfs_write_data
*data
,
1144 struct nfs_page
*first
;
1145 struct inode
*inode
;
1148 /* Set up the RPC argument and reply structs
1149 * NB: take care not to mess about with data->commit et al. */
1151 list_splice_init(head
, &data
->pages
);
1152 first
= nfs_list_entry(data
->pages
.next
);
1153 inode
= first
->wb_context
->path
.dentry
->d_inode
;
1155 data
->inode
= inode
;
1156 data
->cred
= first
->wb_context
->cred
;
1158 data
->args
.fh
= NFS_FH(data
->inode
);
1159 /* Note: we always request a commit of the entire inode */
1160 data
->args
.offset
= 0;
1161 data
->args
.count
= 0;
1162 data
->res
.count
= 0;
1163 data
->res
.fattr
= &data
->fattr
;
1164 data
->res
.verf
= &data
->verf
;
1165 nfs_fattr_init(&data
->fattr
);
1167 /* Set up the initial task struct. */
1168 flags
= (how
& FLUSH_SYNC
) ? 0 : RPC_TASK_ASYNC
;
1169 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), flags
, &nfs_commit_ops
, data
);
1170 NFS_PROTO(inode
)->commit_setup(data
, how
);
1172 data
->task
.tk_priority
= flush_task_priority(how
);
1173 data
->task
.tk_cookie
= (unsigned long)inode
;
1175 dprintk("NFS: %5u initiated commit call\n", data
->task
.tk_pid
);
1179 * Commit dirty pages
1182 nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
)
1184 struct nfs_write_data
*data
;
1185 struct nfs_page
*req
;
1187 data
= nfs_commit_alloc();
1192 /* Set up the argument struct */
1193 nfs_commit_rpcsetup(head
, data
, how
);
1195 nfs_execute_write(data
);
1198 while (!list_empty(head
)) {
1199 req
= nfs_list_entry(head
->next
);
1200 nfs_list_remove_request(req
);
1201 nfs_mark_request_commit(req
);
1202 dec_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
1203 dec_bdi_stat(req
->wb_page
->mapping
->backing_dev_info
,
1205 nfs_clear_page_tag_locked(req
);
1211 * COMMIT call returned
1213 static void nfs_commit_done(struct rpc_task
*task
, void *calldata
)
1215 struct nfs_write_data
*data
= calldata
;
1216 struct nfs_page
*req
;
1218 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1219 task
->tk_pid
, task
->tk_status
);
1221 /* Call the NFS version-specific code */
1222 if (NFS_PROTO(data
->inode
)->commit_done(task
, data
) != 0)
1225 while (!list_empty(&data
->pages
)) {
1226 req
= nfs_list_entry(data
->pages
.next
);
1227 nfs_list_remove_request(req
);
1228 clear_bit(PG_NEED_COMMIT
, &(req
)->wb_flags
);
1229 dec_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
1230 dec_bdi_stat(req
->wb_page
->mapping
->backing_dev_info
,
1233 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1234 req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
1235 (long long)NFS_FILEID(req
->wb_context
->path
.dentry
->d_inode
),
1237 (long long)req_offset(req
));
1238 if (task
->tk_status
< 0) {
1239 nfs_context_set_write_error(req
->wb_context
, task
->tk_status
);
1240 nfs_inode_remove_request(req
);
1241 dprintk(", error = %d\n", task
->tk_status
);
1245 /* Okay, COMMIT succeeded, apparently. Check the verifier
1246 * returned by the server against all stored verfs. */
1247 if (!memcmp(req
->wb_verf
.verifier
, data
->verf
.verifier
, sizeof(data
->verf
.verifier
))) {
1248 /* We have a match */
1249 /* Set the PG_uptodate flag */
1250 nfs_mark_uptodate(req
->wb_page
, req
->wb_pgbase
,
1252 nfs_inode_remove_request(req
);
1256 /* We have a mismatch. Write the page again */
1257 dprintk(" mismatch\n");
1258 nfs_redirty_request(req
);
1260 nfs_clear_page_tag_locked(req
);
1264 static const struct rpc_call_ops nfs_commit_ops
= {
1265 .rpc_call_done
= nfs_commit_done
,
1266 .rpc_release
= nfs_commit_release
,
1269 int nfs_commit_inode(struct inode
*inode
, int how
)
1274 spin_lock(&inode
->i_lock
);
1275 res
= nfs_scan_commit(inode
, &head
, 0, 0);
1276 spin_unlock(&inode
->i_lock
);
1278 int error
= nfs_commit_list(inode
, &head
, how
);
1285 static inline int nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
)
1291 long nfs_sync_mapping_wait(struct address_space
*mapping
, struct writeback_control
*wbc
, int how
)
1293 struct inode
*inode
= mapping
->host
;
1294 pgoff_t idx_start
, idx_end
;
1295 unsigned int npages
= 0;
1297 int nocommit
= how
& FLUSH_NOCOMMIT
;
1301 if (wbc
->range_cyclic
)
1304 idx_start
= wbc
->range_start
>> PAGE_CACHE_SHIFT
;
1305 idx_end
= wbc
->range_end
>> PAGE_CACHE_SHIFT
;
1306 if (idx_end
> idx_start
) {
1307 pgoff_t l_npages
= 1 + idx_end
- idx_start
;
1309 if (sizeof(npages
) != sizeof(l_npages
) &&
1310 (pgoff_t
)npages
!= l_npages
)
1314 how
&= ~FLUSH_NOCOMMIT
;
1315 spin_lock(&inode
->i_lock
);
1317 ret
= nfs_wait_on_requests_locked(inode
, idx_start
, npages
);
1322 pages
= nfs_scan_commit(inode
, &head
, idx_start
, npages
);
1325 if (how
& FLUSH_INVALIDATE
) {
1326 spin_unlock(&inode
->i_lock
);
1327 nfs_cancel_commit_list(&head
);
1329 spin_lock(&inode
->i_lock
);
1332 pages
+= nfs_scan_commit(inode
, &head
, 0, 0);
1333 spin_unlock(&inode
->i_lock
);
1334 ret
= nfs_commit_list(inode
, &head
, how
);
1335 spin_lock(&inode
->i_lock
);
1338 spin_unlock(&inode
->i_lock
);
1342 static int __nfs_write_mapping(struct address_space
*mapping
, struct writeback_control
*wbc
, int how
)
1346 ret
= nfs_writepages(mapping
, wbc
);
1349 ret
= nfs_sync_mapping_wait(mapping
, wbc
, how
);
1354 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
1358 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1359 static int nfs_write_mapping(struct address_space
*mapping
, int how
)
1361 struct writeback_control wbc
= {
1362 .bdi
= mapping
->backing_dev_info
,
1363 .sync_mode
= WB_SYNC_NONE
,
1364 .nr_to_write
= LONG_MAX
,
1365 .for_writepages
= 1,
1370 ret
= __nfs_write_mapping(mapping
, &wbc
, how
);
1373 wbc
.sync_mode
= WB_SYNC_ALL
;
1374 return __nfs_write_mapping(mapping
, &wbc
, how
);
1378 * flush the inode to disk.
1380 int nfs_wb_all(struct inode
*inode
)
1382 return nfs_write_mapping(inode
->i_mapping
, 0);
1385 int nfs_wb_nocommit(struct inode
*inode
)
1387 return nfs_write_mapping(inode
->i_mapping
, FLUSH_NOCOMMIT
);
1390 int nfs_wb_page_cancel(struct inode
*inode
, struct page
*page
)
1392 struct nfs_page
*req
;
1393 loff_t range_start
= page_offset(page
);
1394 loff_t range_end
= range_start
+ (loff_t
)(PAGE_CACHE_SIZE
- 1);
1395 struct writeback_control wbc
= {
1396 .bdi
= page
->mapping
->backing_dev_info
,
1397 .sync_mode
= WB_SYNC_ALL
,
1398 .nr_to_write
= LONG_MAX
,
1399 .range_start
= range_start
,
1400 .range_end
= range_end
,
1404 BUG_ON(!PageLocked(page
));
1406 req
= nfs_page_find_request(page
);
1409 if (test_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
1410 nfs_release_request(req
);
1413 if (nfs_lock_request_dontget(req
)) {
1414 nfs_inode_remove_request(req
);
1416 * In case nfs_inode_remove_request has marked the
1417 * page as being dirty
1419 cancel_dirty_page(page
, PAGE_CACHE_SIZE
);
1420 nfs_unlock_request(req
);
1423 ret
= nfs_wait_on_request(req
);
1427 if (!PagePrivate(page
))
1429 ret
= nfs_sync_mapping_wait(page
->mapping
, &wbc
, FLUSH_INVALIDATE
);
1434 static int nfs_wb_page_priority(struct inode
*inode
, struct page
*page
,
1437 loff_t range_start
= page_offset(page
);
1438 loff_t range_end
= range_start
+ (loff_t
)(PAGE_CACHE_SIZE
- 1);
1439 struct writeback_control wbc
= {
1440 .bdi
= page
->mapping
->backing_dev_info
,
1441 .sync_mode
= WB_SYNC_ALL
,
1442 .nr_to_write
= LONG_MAX
,
1443 .range_start
= range_start
,
1444 .range_end
= range_end
,
1448 BUG_ON(!PageLocked(page
));
1449 if (clear_page_dirty_for_io(page
)) {
1450 ret
= nfs_writepage_locked(page
, &wbc
);
1454 if (!PagePrivate(page
))
1456 ret
= nfs_sync_mapping_wait(page
->mapping
, &wbc
, how
);
1460 __mark_inode_dirty(inode
, I_DIRTY_PAGES
);
1465 * Write back all requests on one page - we do this before reading it.
1467 int nfs_wb_page(struct inode
*inode
, struct page
* page
)
1469 return nfs_wb_page_priority(inode
, page
, FLUSH_STABLE
);
1472 int __init
nfs_init_writepagecache(void)
1474 nfs_wdata_cachep
= kmem_cache_create("nfs_write_data",
1475 sizeof(struct nfs_write_data
),
1476 0, SLAB_HWCACHE_ALIGN
,
1478 if (nfs_wdata_cachep
== NULL
)
1481 nfs_wdata_mempool
= mempool_create_slab_pool(MIN_POOL_WRITE
,
1483 if (nfs_wdata_mempool
== NULL
)
1486 nfs_commit_mempool
= mempool_create_slab_pool(MIN_POOL_COMMIT
,
1488 if (nfs_commit_mempool
== NULL
)
1492 * NFS congestion size, scale with available memory.
1504 * This allows larger machines to have larger/more transfers.
1505 * Limit the default to 256M
1507 nfs_congestion_kb
= (16*int_sqrt(totalram_pages
)) << (PAGE_SHIFT
-10);
1508 if (nfs_congestion_kb
> 256*1024)
1509 nfs_congestion_kb
= 256*1024;
1514 void nfs_destroy_writepagecache(void)
1516 mempool_destroy(nfs_commit_mempool
);
1517 mempool_destroy(nfs_wdata_mempool
);
1518 kmem_cache_destroy(nfs_wdata_cachep
);