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 void nfs_pageio_init_write(struct nfs_pageio_descriptor
*desc
,
38 struct inode
*inode
, int ioflags
);
39 static void nfs_redirty_request(struct nfs_page
*req
);
40 static const struct rpc_call_ops nfs_write_partial_ops
;
41 static const struct rpc_call_ops nfs_write_full_ops
;
42 static const struct rpc_call_ops nfs_commit_ops
;
44 static struct kmem_cache
*nfs_wdata_cachep
;
45 static mempool_t
*nfs_wdata_mempool
;
46 static mempool_t
*nfs_commit_mempool
;
48 struct nfs_write_data
*nfs_commitdata_alloc(void)
50 struct nfs_write_data
*p
= mempool_alloc(nfs_commit_mempool
, GFP_NOFS
);
53 memset(p
, 0, sizeof(*p
));
54 INIT_LIST_HEAD(&p
->pages
);
59 void nfs_commit_free(struct nfs_write_data
*p
)
61 if (p
&& (p
->pagevec
!= &p
->page_array
[0]))
63 mempool_free(p
, nfs_commit_mempool
);
66 struct nfs_write_data
*nfs_writedata_alloc(unsigned int pagecount
)
68 struct nfs_write_data
*p
= mempool_alloc(nfs_wdata_mempool
, GFP_NOFS
);
71 memset(p
, 0, sizeof(*p
));
72 INIT_LIST_HEAD(&p
->pages
);
73 p
->npages
= pagecount
;
74 if (pagecount
<= ARRAY_SIZE(p
->page_array
))
75 p
->pagevec
= p
->page_array
;
77 p
->pagevec
= kcalloc(pagecount
, sizeof(struct page
*), GFP_NOFS
);
79 mempool_free(p
, nfs_wdata_mempool
);
87 static void nfs_writedata_free(struct nfs_write_data
*p
)
89 if (p
&& (p
->pagevec
!= &p
->page_array
[0]))
91 mempool_free(p
, nfs_wdata_mempool
);
94 void nfs_writedata_release(void *data
)
96 struct nfs_write_data
*wdata
= data
;
98 put_nfs_open_context(wdata
->args
.context
);
99 nfs_writedata_free(wdata
);
102 static void nfs_context_set_write_error(struct nfs_open_context
*ctx
, int error
)
106 set_bit(NFS_CONTEXT_ERROR_WRITE
, &ctx
->flags
);
109 static struct nfs_page
*nfs_page_find_request_locked(struct page
*page
)
111 struct nfs_page
*req
= NULL
;
113 if (PagePrivate(page
)) {
114 req
= (struct nfs_page
*)page_private(page
);
116 kref_get(&req
->wb_kref
);
121 static struct nfs_page
*nfs_page_find_request(struct page
*page
)
123 struct inode
*inode
= page
->mapping
->host
;
124 struct nfs_page
*req
= NULL
;
126 spin_lock(&inode
->i_lock
);
127 req
= nfs_page_find_request_locked(page
);
128 spin_unlock(&inode
->i_lock
);
132 /* Adjust the file length if we're writing beyond the end */
133 static void nfs_grow_file(struct page
*page
, unsigned int offset
, unsigned int count
)
135 struct inode
*inode
= page
->mapping
->host
;
139 spin_lock(&inode
->i_lock
);
140 i_size
= i_size_read(inode
);
141 end_index
= (i_size
- 1) >> PAGE_CACHE_SHIFT
;
142 if (i_size
> 0 && page
->index
< end_index
)
144 end
= ((loff_t
)page
->index
<< PAGE_CACHE_SHIFT
) + ((loff_t
)offset
+count
);
147 i_size_write(inode
, end
);
148 nfs_inc_stats(inode
, NFSIOS_EXTENDWRITE
);
150 spin_unlock(&inode
->i_lock
);
153 /* A writeback failed: mark the page as bad, and invalidate the page cache */
154 static void nfs_set_pageerror(struct page
*page
)
157 nfs_zap_mapping(page
->mapping
->host
, page
->mapping
);
160 /* We can set the PG_uptodate flag if we see that a write request
161 * covers the full page.
163 static void nfs_mark_uptodate(struct page
*page
, unsigned int base
, unsigned int count
)
165 if (PageUptodate(page
))
169 if (count
!= nfs_page_length(page
))
171 SetPageUptodate(page
);
174 static int wb_priority(struct writeback_control
*wbc
)
176 if (wbc
->for_reclaim
)
177 return FLUSH_HIGHPRI
| FLUSH_STABLE
;
178 if (wbc
->for_kupdate
)
184 * NFS congestion control
187 int nfs_congestion_kb
;
189 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
190 #define NFS_CONGESTION_OFF_THRESH \
191 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
193 static int nfs_set_page_writeback(struct page
*page
)
195 int ret
= test_set_page_writeback(page
);
198 struct inode
*inode
= page
->mapping
->host
;
199 struct nfs_server
*nfss
= NFS_SERVER(inode
);
201 if (atomic_long_inc_return(&nfss
->writeback
) >
202 NFS_CONGESTION_ON_THRESH
)
203 set_bdi_congested(&nfss
->backing_dev_info
, WRITE
);
208 static void nfs_end_page_writeback(struct page
*page
)
210 struct inode
*inode
= page
->mapping
->host
;
211 struct nfs_server
*nfss
= NFS_SERVER(inode
);
213 end_page_writeback(page
);
214 if (atomic_long_dec_return(&nfss
->writeback
) < NFS_CONGESTION_OFF_THRESH
)
215 clear_bdi_congested(&nfss
->backing_dev_info
, WRITE
);
219 * Find an associated nfs write request, and prepare to flush it out
220 * May return an error if the user signalled nfs_wait_on_request().
222 static int nfs_page_async_flush(struct nfs_pageio_descriptor
*pgio
,
225 struct inode
*inode
= page
->mapping
->host
;
226 struct nfs_page
*req
;
229 spin_lock(&inode
->i_lock
);
231 req
= nfs_page_find_request_locked(page
);
233 spin_unlock(&inode
->i_lock
);
236 if (nfs_set_page_tag_locked(req
))
238 /* Note: If we hold the page lock, as is the case in nfs_writepage,
239 * then the call to nfs_set_page_tag_locked() will always
240 * succeed provided that someone hasn't already marked the
241 * request as dirty (in which case we don't care).
243 spin_unlock(&inode
->i_lock
);
244 ret
= nfs_wait_on_request(req
);
245 nfs_release_request(req
);
248 spin_lock(&inode
->i_lock
);
250 if (test_bit(PG_CLEAN
, &req
->wb_flags
)) {
251 spin_unlock(&inode
->i_lock
);
254 if (nfs_set_page_writeback(page
) != 0) {
255 spin_unlock(&inode
->i_lock
);
258 spin_unlock(&inode
->i_lock
);
259 if (!nfs_pageio_add_request(pgio
, req
)) {
260 nfs_redirty_request(req
);
261 return pgio
->pg_error
;
266 static int nfs_do_writepage(struct page
*page
, struct writeback_control
*wbc
, struct nfs_pageio_descriptor
*pgio
)
268 struct inode
*inode
= page
->mapping
->host
;
270 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGE
);
271 nfs_add_stats(inode
, NFSIOS_WRITEPAGES
, 1);
273 nfs_pageio_cond_complete(pgio
, page
->index
);
274 return nfs_page_async_flush(pgio
, page
);
278 * Write an mmapped page to the server.
280 static int nfs_writepage_locked(struct page
*page
, struct writeback_control
*wbc
)
282 struct nfs_pageio_descriptor pgio
;
285 nfs_pageio_init_write(&pgio
, page
->mapping
->host
, wb_priority(wbc
));
286 err
= nfs_do_writepage(page
, wbc
, &pgio
);
287 nfs_pageio_complete(&pgio
);
290 if (pgio
.pg_error
< 0)
291 return pgio
.pg_error
;
295 int nfs_writepage(struct page
*page
, struct writeback_control
*wbc
)
299 ret
= nfs_writepage_locked(page
, wbc
);
304 static int nfs_writepages_callback(struct page
*page
, struct writeback_control
*wbc
, void *data
)
308 ret
= nfs_do_writepage(page
, wbc
, data
);
313 int nfs_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
315 struct inode
*inode
= mapping
->host
;
316 unsigned long *bitlock
= &NFS_I(inode
)->flags
;
317 struct nfs_pageio_descriptor pgio
;
320 /* Stop dirtying of new pages while we sync */
321 err
= wait_on_bit_lock(bitlock
, NFS_INO_FLUSHING
,
322 nfs_wait_bit_killable
, TASK_KILLABLE
);
326 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGES
);
328 nfs_pageio_init_write(&pgio
, inode
, wb_priority(wbc
));
329 err
= write_cache_pages(mapping
, wbc
, nfs_writepages_callback
, &pgio
);
330 nfs_pageio_complete(&pgio
);
332 clear_bit_unlock(NFS_INO_FLUSHING
, bitlock
);
333 smp_mb__after_clear_bit();
334 wake_up_bit(bitlock
, NFS_INO_FLUSHING
);
347 * Insert a write request into an inode
349 static int nfs_inode_add_request(struct inode
*inode
, struct nfs_page
*req
)
351 struct nfs_inode
*nfsi
= NFS_I(inode
);
354 error
= radix_tree_preload(GFP_NOFS
);
358 /* Lock the request! */
359 nfs_lock_request_dontget(req
);
361 spin_lock(&inode
->i_lock
);
362 error
= radix_tree_insert(&nfsi
->nfs_page_tree
, req
->wb_index
, req
);
366 if (nfs_have_delegation(inode
, FMODE_WRITE
))
369 SetPagePrivate(req
->wb_page
);
370 set_page_private(req
->wb_page
, (unsigned long)req
);
372 kref_get(&req
->wb_kref
);
373 radix_tree_tag_set(&nfsi
->nfs_page_tree
, req
->wb_index
,
374 NFS_PAGE_TAG_LOCKED
);
375 spin_unlock(&inode
->i_lock
);
376 radix_tree_preload_end();
382 * Remove a write request from an inode
384 static void nfs_inode_remove_request(struct nfs_page
*req
)
386 struct inode
*inode
= req
->wb_context
->path
.dentry
->d_inode
;
387 struct nfs_inode
*nfsi
= NFS_I(inode
);
389 BUG_ON (!NFS_WBACK_BUSY(req
));
391 spin_lock(&inode
->i_lock
);
392 set_page_private(req
->wb_page
, 0);
393 ClearPagePrivate(req
->wb_page
);
394 radix_tree_delete(&nfsi
->nfs_page_tree
, req
->wb_index
);
397 spin_unlock(&inode
->i_lock
);
400 spin_unlock(&inode
->i_lock
);
401 nfs_clear_request(req
);
402 nfs_release_request(req
);
406 nfs_mark_request_dirty(struct nfs_page
*req
)
408 __set_page_dirty_nobuffers(req
->wb_page
);
411 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
413 * Add a request to the inode's commit list.
416 nfs_mark_request_commit(struct nfs_page
*req
)
418 struct inode
*inode
= req
->wb_context
->path
.dentry
->d_inode
;
419 struct nfs_inode
*nfsi
= NFS_I(inode
);
421 spin_lock(&inode
->i_lock
);
422 set_bit(PG_CLEAN
, &(req
)->wb_flags
);
423 radix_tree_tag_set(&nfsi
->nfs_page_tree
,
425 NFS_PAGE_TAG_COMMIT
);
426 spin_unlock(&inode
->i_lock
);
427 inc_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
428 inc_bdi_stat(req
->wb_page
->mapping
->backing_dev_info
, BDI_RECLAIMABLE
);
429 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
433 nfs_clear_request_commit(struct nfs_page
*req
)
435 struct page
*page
= req
->wb_page
;
437 if (test_and_clear_bit(PG_CLEAN
, &(req
)->wb_flags
)) {
438 dec_zone_page_state(page
, NR_UNSTABLE_NFS
);
439 dec_bdi_stat(page
->mapping
->backing_dev_info
, BDI_RECLAIMABLE
);
446 int nfs_write_need_commit(struct nfs_write_data
*data
)
448 return data
->verf
.committed
!= NFS_FILE_SYNC
;
452 int nfs_reschedule_unstable_write(struct nfs_page
*req
)
454 if (test_and_clear_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
455 nfs_mark_request_commit(req
);
458 if (test_and_clear_bit(PG_NEED_RESCHED
, &req
->wb_flags
)) {
459 nfs_mark_request_dirty(req
);
466 nfs_mark_request_commit(struct nfs_page
*req
)
471 nfs_clear_request_commit(struct nfs_page
*req
)
477 int nfs_write_need_commit(struct nfs_write_data
*data
)
483 int nfs_reschedule_unstable_write(struct nfs_page
*req
)
490 * Wait for a request to complete.
492 * Interruptible by fatal signals only.
494 static int nfs_wait_on_requests_locked(struct inode
*inode
, pgoff_t idx_start
, unsigned int npages
)
496 struct nfs_inode
*nfsi
= NFS_I(inode
);
497 struct nfs_page
*req
;
498 pgoff_t idx_end
, next
;
499 unsigned int res
= 0;
505 idx_end
= idx_start
+ npages
- 1;
508 while (radix_tree_gang_lookup_tag(&nfsi
->nfs_page_tree
, (void **)&req
, next
, 1, NFS_PAGE_TAG_LOCKED
)) {
509 if (req
->wb_index
> idx_end
)
512 next
= req
->wb_index
+ 1;
513 BUG_ON(!NFS_WBACK_BUSY(req
));
515 kref_get(&req
->wb_kref
);
516 spin_unlock(&inode
->i_lock
);
517 error
= nfs_wait_on_request(req
);
518 nfs_release_request(req
);
519 spin_lock(&inode
->i_lock
);
527 static void nfs_cancel_commit_list(struct list_head
*head
)
529 struct nfs_page
*req
;
531 while(!list_empty(head
)) {
532 req
= nfs_list_entry(head
->next
);
533 nfs_list_remove_request(req
);
534 nfs_clear_request_commit(req
);
535 nfs_inode_remove_request(req
);
536 nfs_unlock_request(req
);
540 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
542 nfs_need_commit(struct nfs_inode
*nfsi
)
544 return radix_tree_tagged(&nfsi
->nfs_page_tree
, NFS_PAGE_TAG_COMMIT
);
548 * nfs_scan_commit - Scan an inode for commit requests
549 * @inode: NFS inode to scan
550 * @dst: destination list
551 * @idx_start: lower bound of page->index to scan.
552 * @npages: idx_start + npages sets the upper bound to scan.
554 * Moves requests from the inode's 'commit' request list.
555 * The requests are *not* checked to ensure that they form a contiguous set.
558 nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, pgoff_t idx_start
, unsigned int npages
)
560 struct nfs_inode
*nfsi
= NFS_I(inode
);
562 if (!nfs_need_commit(nfsi
))
565 return nfs_scan_list(nfsi
, dst
, idx_start
, npages
, NFS_PAGE_TAG_COMMIT
);
568 static inline int nfs_need_commit(struct nfs_inode
*nfsi
)
573 static inline int nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
, pgoff_t idx_start
, unsigned int npages
)
580 * Search for an existing write request, and attempt to update
581 * it to reflect a new dirty region on a given page.
583 * If the attempt fails, then the existing request is flushed out
586 static struct nfs_page
*nfs_try_to_update_request(struct inode
*inode
,
591 struct nfs_page
*req
;
596 if (!PagePrivate(page
))
599 end
= offset
+ bytes
;
600 spin_lock(&inode
->i_lock
);
603 req
= nfs_page_find_request_locked(page
);
607 rqend
= req
->wb_offset
+ req
->wb_bytes
;
609 * Tell the caller to flush out the request if
610 * the offsets are non-contiguous.
611 * Note: nfs_flush_incompatible() will already
612 * have flushed out requests having wrong owners.
615 || end
< req
->wb_offset
)
618 if (nfs_set_page_tag_locked(req
))
621 /* The request is locked, so wait and then retry */
622 spin_unlock(&inode
->i_lock
);
623 error
= nfs_wait_on_request(req
);
624 nfs_release_request(req
);
627 spin_lock(&inode
->i_lock
);
630 if (nfs_clear_request_commit(req
))
631 radix_tree_tag_clear(&NFS_I(inode
)->nfs_page_tree
,
632 req
->wb_index
, NFS_PAGE_TAG_COMMIT
);
634 /* Okay, the request matches. Update the region */
635 if (offset
< req
->wb_offset
) {
636 req
->wb_offset
= offset
;
637 req
->wb_pgbase
= offset
;
640 req
->wb_bytes
= end
- req
->wb_offset
;
642 req
->wb_bytes
= rqend
- req
->wb_offset
;
644 spin_unlock(&inode
->i_lock
);
647 spin_unlock(&inode
->i_lock
);
648 nfs_release_request(req
);
649 error
= nfs_wb_page(inode
, page
);
651 return ERR_PTR(error
);
655 * Try to update an existing write request, or create one if there is none.
657 * Note: Should always be called with the Page Lock held to prevent races
658 * if we have to add a new request. Also assumes that the caller has
659 * already called nfs_flush_incompatible() if necessary.
661 static struct nfs_page
* nfs_setup_write_request(struct nfs_open_context
* ctx
,
662 struct page
*page
, unsigned int offset
, unsigned int bytes
)
664 struct inode
*inode
= page
->mapping
->host
;
665 struct nfs_page
*req
;
668 req
= nfs_try_to_update_request(inode
, page
, offset
, bytes
);
671 req
= nfs_create_request(ctx
, inode
, page
, offset
, bytes
);
674 error
= nfs_inode_add_request(inode
, req
);
676 nfs_release_request(req
);
677 req
= ERR_PTR(error
);
683 static int nfs_writepage_setup(struct nfs_open_context
*ctx
, struct page
*page
,
684 unsigned int offset
, unsigned int count
)
686 struct nfs_page
*req
;
688 req
= nfs_setup_write_request(ctx
, page
, offset
, count
);
691 /* Update file length */
692 nfs_grow_file(page
, offset
, count
);
693 nfs_mark_uptodate(page
, req
->wb_pgbase
, req
->wb_bytes
);
694 nfs_clear_page_tag_locked(req
);
698 int nfs_flush_incompatible(struct file
*file
, struct page
*page
)
700 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
701 struct nfs_page
*req
;
702 int do_flush
, status
;
704 * Look for a request corresponding to this page. If there
705 * is one, and it belongs to another file, we flush it out
706 * before we try to copy anything into the page. Do this
707 * due to the lack of an ACCESS-type call in NFSv2.
708 * Also do the same if we find a request from an existing
712 req
= nfs_page_find_request(page
);
715 do_flush
= req
->wb_page
!= page
|| req
->wb_context
!= ctx
;
716 nfs_release_request(req
);
719 status
= nfs_wb_page(page
->mapping
->host
, page
);
720 } while (status
== 0);
725 * If the page cache is marked as unsafe or invalid, then we can't rely on
726 * the PageUptodate() flag. In this case, we will need to turn off
727 * write optimisations that depend on the page contents being correct.
729 static int nfs_write_pageuptodate(struct page
*page
, struct inode
*inode
)
731 return PageUptodate(page
) &&
732 !(NFS_I(inode
)->cache_validity
& (NFS_INO_REVAL_PAGECACHE
|NFS_INO_INVALID_DATA
));
736 * Update and possibly write a cached page of an NFS file.
738 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
739 * things with a page scheduled for an RPC call (e.g. invalidate it).
741 int nfs_updatepage(struct file
*file
, struct page
*page
,
742 unsigned int offset
, unsigned int count
)
744 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
745 struct inode
*inode
= page
->mapping
->host
;
748 nfs_inc_stats(inode
, NFSIOS_VFSUPDATEPAGE
);
750 dprintk("NFS: nfs_updatepage(%s/%s %d@%lld)\n",
751 file
->f_path
.dentry
->d_parent
->d_name
.name
,
752 file
->f_path
.dentry
->d_name
.name
, count
,
753 (long long)(page_offset(page
) + offset
));
755 /* If we're not using byte range locks, and we know the page
756 * is up to date, it may be more efficient to extend the write
757 * to cover the entire page in order to avoid fragmentation
760 if (nfs_write_pageuptodate(page
, inode
) &&
761 inode
->i_flock
== NULL
&&
762 !(file
->f_flags
& O_SYNC
)) {
763 count
= max(count
+ offset
, nfs_page_length(page
));
767 status
= nfs_writepage_setup(ctx
, page
, offset
, count
);
769 nfs_set_pageerror(page
);
771 __set_page_dirty_nobuffers(page
);
773 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
774 status
, (long long)i_size_read(inode
));
778 static void nfs_writepage_release(struct nfs_page
*req
)
781 if (PageError(req
->wb_page
) || !nfs_reschedule_unstable_write(req
)) {
782 nfs_end_page_writeback(req
->wb_page
);
783 nfs_inode_remove_request(req
);
785 nfs_end_page_writeback(req
->wb_page
);
786 nfs_clear_page_tag_locked(req
);
789 static int flush_task_priority(int how
)
791 switch (how
& (FLUSH_HIGHPRI
|FLUSH_LOWPRI
)) {
793 return RPC_PRIORITY_HIGH
;
795 return RPC_PRIORITY_LOW
;
797 return RPC_PRIORITY_NORMAL
;
801 * Set up the argument/result storage required for the RPC call.
803 static int nfs_write_rpcsetup(struct nfs_page
*req
,
804 struct nfs_write_data
*data
,
805 const struct rpc_call_ops
*call_ops
,
806 unsigned int count
, unsigned int offset
,
809 struct inode
*inode
= req
->wb_context
->path
.dentry
->d_inode
;
810 int flags
= (how
& FLUSH_SYNC
) ? 0 : RPC_TASK_ASYNC
;
811 int priority
= flush_task_priority(how
);
812 struct rpc_task
*task
;
813 struct rpc_message msg
= {
814 .rpc_argp
= &data
->args
,
815 .rpc_resp
= &data
->res
,
816 .rpc_cred
= req
->wb_context
->cred
,
818 struct rpc_task_setup task_setup_data
= {
819 .rpc_client
= NFS_CLIENT(inode
),
822 .callback_ops
= call_ops
,
823 .callback_data
= data
,
824 .workqueue
= nfsiod_workqueue
,
826 .priority
= priority
,
829 /* Set up the RPC argument and reply structs
830 * NB: take care not to mess about with data->commit et al. */
833 data
->inode
= inode
= req
->wb_context
->path
.dentry
->d_inode
;
834 data
->cred
= msg
.rpc_cred
;
836 data
->args
.fh
= NFS_FH(inode
);
837 data
->args
.offset
= req_offset(req
) + offset
;
838 data
->args
.pgbase
= req
->wb_pgbase
+ offset
;
839 data
->args
.pages
= data
->pagevec
;
840 data
->args
.count
= count
;
841 data
->args
.context
= get_nfs_open_context(req
->wb_context
);
842 data
->args
.stable
= NFS_UNSTABLE
;
843 if (how
& FLUSH_STABLE
) {
844 data
->args
.stable
= NFS_DATA_SYNC
;
845 if (!nfs_need_commit(NFS_I(inode
)))
846 data
->args
.stable
= NFS_FILE_SYNC
;
849 data
->res
.fattr
= &data
->fattr
;
850 data
->res
.count
= count
;
851 data
->res
.verf
= &data
->verf
;
852 nfs_fattr_init(&data
->fattr
);
854 /* Set up the initial task struct. */
855 NFS_PROTO(inode
)->write_setup(data
, &msg
);
857 dprintk("NFS: %5u initiated write call "
858 "(req %s/%lld, %u bytes @ offset %llu)\n",
861 (long long)NFS_FILEID(inode
),
863 (unsigned long long)data
->args
.offset
);
865 task
= rpc_run_task(&task_setup_data
);
867 return PTR_ERR(task
);
872 /* If a nfs_flush_* function fails, it should remove reqs from @head and
873 * call this on each, which will prepare them to be retried on next
874 * writeback using standard nfs.
876 static void nfs_redirty_request(struct nfs_page
*req
)
878 nfs_mark_request_dirty(req
);
879 nfs_end_page_writeback(req
->wb_page
);
880 nfs_clear_page_tag_locked(req
);
884 * Generate multiple small requests to write out a single
885 * contiguous dirty area on one page.
887 static int nfs_flush_multi(struct inode
*inode
, struct list_head
*head
, unsigned int npages
, size_t count
, int how
)
889 struct nfs_page
*req
= nfs_list_entry(head
->next
);
890 struct page
*page
= req
->wb_page
;
891 struct nfs_write_data
*data
;
892 size_t wsize
= NFS_SERVER(inode
)->wsize
, nbytes
;
898 nfs_list_remove_request(req
);
902 size_t len
= min(nbytes
, wsize
);
904 data
= nfs_writedata_alloc(1);
907 list_add(&data
->pages
, &list
);
910 } while (nbytes
!= 0);
911 atomic_set(&req
->wb_complete
, requests
);
913 ClearPageError(page
);
919 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
920 list_del_init(&data
->pages
);
922 data
->pagevec
[0] = page
;
926 ret2
= nfs_write_rpcsetup(req
, data
, &nfs_write_partial_ops
,
932 } while (nbytes
!= 0);
937 while (!list_empty(&list
)) {
938 data
= list_entry(list
.next
, struct nfs_write_data
, pages
);
939 list_del(&data
->pages
);
940 nfs_writedata_release(data
);
942 nfs_redirty_request(req
);
947 * Create an RPC task for the given write request and kick it.
948 * The page must have been locked by the caller.
950 * It may happen that the page we're passed is not marked dirty.
951 * This is the case if nfs_updatepage detects a conflicting request
952 * that has been written but not committed.
954 static int nfs_flush_one(struct inode
*inode
, struct list_head
*head
, unsigned int npages
, size_t count
, int how
)
956 struct nfs_page
*req
;
958 struct nfs_write_data
*data
;
960 data
= nfs_writedata_alloc(npages
);
964 pages
= data
->pagevec
;
965 while (!list_empty(head
)) {
966 req
= nfs_list_entry(head
->next
);
967 nfs_list_remove_request(req
);
968 nfs_list_add_request(req
, &data
->pages
);
969 ClearPageError(req
->wb_page
);
970 *pages
++ = req
->wb_page
;
972 req
= nfs_list_entry(data
->pages
.next
);
974 /* Set up the argument struct */
975 return nfs_write_rpcsetup(req
, data
, &nfs_write_full_ops
, count
, 0, how
);
977 while (!list_empty(head
)) {
978 req
= nfs_list_entry(head
->next
);
979 nfs_list_remove_request(req
);
980 nfs_redirty_request(req
);
985 static void nfs_pageio_init_write(struct nfs_pageio_descriptor
*pgio
,
986 struct inode
*inode
, int ioflags
)
988 size_t wsize
= NFS_SERVER(inode
)->wsize
;
990 if (wsize
< PAGE_CACHE_SIZE
)
991 nfs_pageio_init(pgio
, inode
, nfs_flush_multi
, wsize
, ioflags
);
993 nfs_pageio_init(pgio
, inode
, nfs_flush_one
, wsize
, ioflags
);
997 * Handle a write reply that flushed part of a page.
999 static void nfs_writeback_done_partial(struct rpc_task
*task
, void *calldata
)
1001 struct nfs_write_data
*data
= calldata
;
1003 dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1005 data
->req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
1007 NFS_FILEID(data
->req
->wb_context
->path
.dentry
->d_inode
),
1008 data
->req
->wb_bytes
, (long long)req_offset(data
->req
));
1010 nfs_writeback_done(task
, data
);
1013 static void nfs_writeback_release_partial(void *calldata
)
1015 struct nfs_write_data
*data
= calldata
;
1016 struct nfs_page
*req
= data
->req
;
1017 struct page
*page
= req
->wb_page
;
1018 int status
= data
->task
.tk_status
;
1021 nfs_set_pageerror(page
);
1022 nfs_context_set_write_error(req
->wb_context
, status
);
1023 dprintk(", error = %d\n", status
);
1027 if (nfs_write_need_commit(data
)) {
1028 struct inode
*inode
= page
->mapping
->host
;
1030 spin_lock(&inode
->i_lock
);
1031 if (test_bit(PG_NEED_RESCHED
, &req
->wb_flags
)) {
1032 /* Do nothing we need to resend the writes */
1033 } else if (!test_and_set_bit(PG_NEED_COMMIT
, &req
->wb_flags
)) {
1034 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1035 dprintk(" defer commit\n");
1036 } else if (memcmp(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
))) {
1037 set_bit(PG_NEED_RESCHED
, &req
->wb_flags
);
1038 clear_bit(PG_NEED_COMMIT
, &req
->wb_flags
);
1039 dprintk(" server reboot detected\n");
1041 spin_unlock(&inode
->i_lock
);
1046 if (atomic_dec_and_test(&req
->wb_complete
))
1047 nfs_writepage_release(req
);
1048 nfs_writedata_release(calldata
);
1051 static const struct rpc_call_ops nfs_write_partial_ops
= {
1052 .rpc_call_done
= nfs_writeback_done_partial
,
1053 .rpc_release
= nfs_writeback_release_partial
,
1057 * Handle a write reply that flushes a whole page.
1059 * FIXME: There is an inherent race with invalidate_inode_pages and
1060 * writebacks since the page->count is kept > 1 for as long
1061 * as the page has a write request pending.
1063 static void nfs_writeback_done_full(struct rpc_task
*task
, void *calldata
)
1065 struct nfs_write_data
*data
= calldata
;
1067 nfs_writeback_done(task
, data
);
1070 static void nfs_writeback_release_full(void *calldata
)
1072 struct nfs_write_data
*data
= calldata
;
1073 int status
= data
->task
.tk_status
;
1075 /* Update attributes as result of writeback. */
1076 while (!list_empty(&data
->pages
)) {
1077 struct nfs_page
*req
= nfs_list_entry(data
->pages
.next
);
1078 struct page
*page
= req
->wb_page
;
1080 nfs_list_remove_request(req
);
1082 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1084 req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
1085 (long long)NFS_FILEID(req
->wb_context
->path
.dentry
->d_inode
),
1087 (long long)req_offset(req
));
1090 nfs_set_pageerror(page
);
1091 nfs_context_set_write_error(req
->wb_context
, status
);
1092 dprintk(", error = %d\n", status
);
1093 goto remove_request
;
1096 if (nfs_write_need_commit(data
)) {
1097 memcpy(&req
->wb_verf
, &data
->verf
, sizeof(req
->wb_verf
));
1098 nfs_mark_request_commit(req
);
1099 nfs_end_page_writeback(page
);
1100 dprintk(" marked for commit\n");
1105 nfs_end_page_writeback(page
);
1106 nfs_inode_remove_request(req
);
1108 nfs_clear_page_tag_locked(req
);
1110 nfs_writedata_release(calldata
);
1113 static const struct rpc_call_ops nfs_write_full_ops
= {
1114 .rpc_call_done
= nfs_writeback_done_full
,
1115 .rpc_release
= nfs_writeback_release_full
,
1120 * This function is called when the WRITE call is complete.
1122 int nfs_writeback_done(struct rpc_task
*task
, struct nfs_write_data
*data
)
1124 struct nfs_writeargs
*argp
= &data
->args
;
1125 struct nfs_writeres
*resp
= &data
->res
;
1128 dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1129 task
->tk_pid
, task
->tk_status
);
1132 * ->write_done will attempt to use post-op attributes to detect
1133 * conflicting writes by other clients. A strict interpretation
1134 * of close-to-open would allow us to continue caching even if
1135 * another writer had changed the file, but some applications
1136 * depend on tighter cache coherency when writing.
1138 status
= NFS_PROTO(data
->inode
)->write_done(task
, data
);
1141 nfs_add_stats(data
->inode
, NFSIOS_SERVERWRITTENBYTES
, resp
->count
);
1143 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1144 if (resp
->verf
->committed
< argp
->stable
&& task
->tk_status
>= 0) {
1145 /* We tried a write call, but the server did not
1146 * commit data to stable storage even though we
1148 * Note: There is a known bug in Tru64 < 5.0 in which
1149 * the server reports NFS_DATA_SYNC, but performs
1150 * NFS_FILE_SYNC. We therefore implement this checking
1151 * as a dprintk() in order to avoid filling syslog.
1153 static unsigned long complain
;
1155 if (time_before(complain
, jiffies
)) {
1156 dprintk("NFS: faulty NFS server %s:"
1157 " (committed = %d) != (stable = %d)\n",
1158 NFS_SERVER(data
->inode
)->nfs_client
->cl_hostname
,
1159 resp
->verf
->committed
, argp
->stable
);
1160 complain
= jiffies
+ 300 * HZ
;
1164 /* Is this a short write? */
1165 if (task
->tk_status
>= 0 && resp
->count
< argp
->count
) {
1166 static unsigned long complain
;
1168 nfs_inc_stats(data
->inode
, NFSIOS_SHORTWRITE
);
1170 /* Has the server at least made some progress? */
1171 if (resp
->count
!= 0) {
1172 /* Was this an NFSv2 write or an NFSv3 stable write? */
1173 if (resp
->verf
->committed
!= NFS_UNSTABLE
) {
1174 /* Resend from where the server left off */
1175 argp
->offset
+= resp
->count
;
1176 argp
->pgbase
+= resp
->count
;
1177 argp
->count
-= resp
->count
;
1179 /* Resend as a stable write in order to avoid
1180 * headaches in the case of a server crash.
1182 argp
->stable
= NFS_FILE_SYNC
;
1184 rpc_restart_call(task
);
1187 if (time_before(complain
, jiffies
)) {
1189 "NFS: Server wrote zero bytes, expected %u.\n",
1191 complain
= jiffies
+ 300 * HZ
;
1193 /* Can't do anything about it except throw an error. */
1194 task
->tk_status
= -EIO
;
1200 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1201 void nfs_commitdata_release(void *data
)
1203 struct nfs_write_data
*wdata
= data
;
1205 put_nfs_open_context(wdata
->args
.context
);
1206 nfs_commit_free(wdata
);
1210 * Set up the argument/result storage required for the RPC call.
1212 static int nfs_commit_rpcsetup(struct list_head
*head
,
1213 struct nfs_write_data
*data
,
1216 struct nfs_page
*first
= nfs_list_entry(head
->next
);
1217 struct inode
*inode
= first
->wb_context
->path
.dentry
->d_inode
;
1218 int flags
= (how
& FLUSH_SYNC
) ? 0 : RPC_TASK_ASYNC
;
1219 int priority
= flush_task_priority(how
);
1220 struct rpc_task
*task
;
1221 struct rpc_message msg
= {
1222 .rpc_argp
= &data
->args
,
1223 .rpc_resp
= &data
->res
,
1224 .rpc_cred
= first
->wb_context
->cred
,
1226 struct rpc_task_setup task_setup_data
= {
1227 .task
= &data
->task
,
1228 .rpc_client
= NFS_CLIENT(inode
),
1229 .rpc_message
= &msg
,
1230 .callback_ops
= &nfs_commit_ops
,
1231 .callback_data
= data
,
1232 .workqueue
= nfsiod_workqueue
,
1234 .priority
= priority
,
1237 /* Set up the RPC argument and reply structs
1238 * NB: take care not to mess about with data->commit et al. */
1240 list_splice_init(head
, &data
->pages
);
1242 data
->inode
= inode
;
1243 data
->cred
= msg
.rpc_cred
;
1245 data
->args
.fh
= NFS_FH(data
->inode
);
1246 /* Note: we always request a commit of the entire inode */
1247 data
->args
.offset
= 0;
1248 data
->args
.count
= 0;
1249 data
->args
.context
= get_nfs_open_context(first
->wb_context
);
1250 data
->res
.count
= 0;
1251 data
->res
.fattr
= &data
->fattr
;
1252 data
->res
.verf
= &data
->verf
;
1253 nfs_fattr_init(&data
->fattr
);
1255 /* Set up the initial task struct. */
1256 NFS_PROTO(inode
)->commit_setup(data
, &msg
);
1258 dprintk("NFS: %5u initiated commit call\n", data
->task
.tk_pid
);
1260 task
= rpc_run_task(&task_setup_data
);
1262 return PTR_ERR(task
);
1268 * Commit dirty pages
1271 nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
)
1273 struct nfs_write_data
*data
;
1274 struct nfs_page
*req
;
1276 data
= nfs_commitdata_alloc();
1281 /* Set up the argument struct */
1282 return nfs_commit_rpcsetup(head
, data
, how
);
1284 while (!list_empty(head
)) {
1285 req
= nfs_list_entry(head
->next
);
1286 nfs_list_remove_request(req
);
1287 nfs_mark_request_commit(req
);
1288 dec_zone_page_state(req
->wb_page
, NR_UNSTABLE_NFS
);
1289 dec_bdi_stat(req
->wb_page
->mapping
->backing_dev_info
,
1291 nfs_clear_page_tag_locked(req
);
1297 * COMMIT call returned
1299 static void nfs_commit_done(struct rpc_task
*task
, void *calldata
)
1301 struct nfs_write_data
*data
= calldata
;
1303 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1304 task
->tk_pid
, task
->tk_status
);
1306 /* Call the NFS version-specific code */
1307 if (NFS_PROTO(data
->inode
)->commit_done(task
, data
) != 0)
1311 static void nfs_commit_release(void *calldata
)
1313 struct nfs_write_data
*data
= calldata
;
1314 struct nfs_page
*req
;
1315 int status
= data
->task
.tk_status
;
1317 while (!list_empty(&data
->pages
)) {
1318 req
= nfs_list_entry(data
->pages
.next
);
1319 nfs_list_remove_request(req
);
1320 nfs_clear_request_commit(req
);
1322 dprintk("NFS: commit (%s/%lld %d@%lld)",
1323 req
->wb_context
->path
.dentry
->d_inode
->i_sb
->s_id
,
1324 (long long)NFS_FILEID(req
->wb_context
->path
.dentry
->d_inode
),
1326 (long long)req_offset(req
));
1328 nfs_context_set_write_error(req
->wb_context
, status
);
1329 nfs_inode_remove_request(req
);
1330 dprintk(", error = %d\n", status
);
1334 /* Okay, COMMIT succeeded, apparently. Check the verifier
1335 * returned by the server against all stored verfs. */
1336 if (!memcmp(req
->wb_verf
.verifier
, data
->verf
.verifier
, sizeof(data
->verf
.verifier
))) {
1337 /* We have a match */
1338 nfs_inode_remove_request(req
);
1342 /* We have a mismatch. Write the page again */
1343 dprintk(" mismatch\n");
1344 nfs_mark_request_dirty(req
);
1346 nfs_clear_page_tag_locked(req
);
1348 nfs_commitdata_release(calldata
);
1351 static const struct rpc_call_ops nfs_commit_ops
= {
1352 .rpc_call_done
= nfs_commit_done
,
1353 .rpc_release
= nfs_commit_release
,
1356 int nfs_commit_inode(struct inode
*inode
, int how
)
1361 spin_lock(&inode
->i_lock
);
1362 res
= nfs_scan_commit(inode
, &head
, 0, 0);
1363 spin_unlock(&inode
->i_lock
);
1365 int error
= nfs_commit_list(inode
, &head
, how
);
1372 static inline int nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
)
1378 long nfs_sync_mapping_wait(struct address_space
*mapping
, struct writeback_control
*wbc
, int how
)
1380 struct inode
*inode
= mapping
->host
;
1381 pgoff_t idx_start
, idx_end
;
1382 unsigned int npages
= 0;
1384 int nocommit
= how
& FLUSH_NOCOMMIT
;
1388 if (wbc
->range_cyclic
)
1391 idx_start
= wbc
->range_start
>> PAGE_CACHE_SHIFT
;
1392 idx_end
= wbc
->range_end
>> PAGE_CACHE_SHIFT
;
1393 if (idx_end
> idx_start
) {
1394 pgoff_t l_npages
= 1 + idx_end
- idx_start
;
1396 if (sizeof(npages
) != sizeof(l_npages
) &&
1397 (pgoff_t
)npages
!= l_npages
)
1401 how
&= ~FLUSH_NOCOMMIT
;
1402 spin_lock(&inode
->i_lock
);
1404 ret
= nfs_wait_on_requests_locked(inode
, idx_start
, npages
);
1409 pages
= nfs_scan_commit(inode
, &head
, idx_start
, npages
);
1412 if (how
& FLUSH_INVALIDATE
) {
1413 spin_unlock(&inode
->i_lock
);
1414 nfs_cancel_commit_list(&head
);
1416 spin_lock(&inode
->i_lock
);
1419 pages
+= nfs_scan_commit(inode
, &head
, 0, 0);
1420 spin_unlock(&inode
->i_lock
);
1421 ret
= nfs_commit_list(inode
, &head
, how
);
1422 spin_lock(&inode
->i_lock
);
1425 spin_unlock(&inode
->i_lock
);
1429 static int __nfs_write_mapping(struct address_space
*mapping
, struct writeback_control
*wbc
, int how
)
1433 ret
= nfs_writepages(mapping
, wbc
);
1436 ret
= nfs_sync_mapping_wait(mapping
, wbc
, how
);
1441 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
1445 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1446 static int nfs_write_mapping(struct address_space
*mapping
, int how
)
1448 struct writeback_control wbc
= {
1449 .bdi
= mapping
->backing_dev_info
,
1450 .sync_mode
= WB_SYNC_ALL
,
1451 .nr_to_write
= LONG_MAX
,
1453 .range_end
= LLONG_MAX
,
1454 .for_writepages
= 1,
1457 return __nfs_write_mapping(mapping
, &wbc
, how
);
1461 * flush the inode to disk.
1463 int nfs_wb_all(struct inode
*inode
)
1465 return nfs_write_mapping(inode
->i_mapping
, 0);
1468 int nfs_wb_nocommit(struct inode
*inode
)
1470 return nfs_write_mapping(inode
->i_mapping
, FLUSH_NOCOMMIT
);
1473 int nfs_wb_page_cancel(struct inode
*inode
, struct page
*page
)
1475 struct nfs_page
*req
;
1476 loff_t range_start
= page_offset(page
);
1477 loff_t range_end
= range_start
+ (loff_t
)(PAGE_CACHE_SIZE
- 1);
1478 struct writeback_control wbc
= {
1479 .bdi
= page
->mapping
->backing_dev_info
,
1480 .sync_mode
= WB_SYNC_ALL
,
1481 .nr_to_write
= LONG_MAX
,
1482 .range_start
= range_start
,
1483 .range_end
= range_end
,
1487 BUG_ON(!PageLocked(page
));
1489 req
= nfs_page_find_request(page
);
1492 if (test_bit(PG_CLEAN
, &req
->wb_flags
)) {
1493 nfs_release_request(req
);
1496 if (nfs_lock_request_dontget(req
)) {
1497 nfs_inode_remove_request(req
);
1499 * In case nfs_inode_remove_request has marked the
1500 * page as being dirty
1502 cancel_dirty_page(page
, PAGE_CACHE_SIZE
);
1503 nfs_unlock_request(req
);
1506 ret
= nfs_wait_on_request(req
);
1510 if (!PagePrivate(page
))
1512 ret
= nfs_sync_mapping_wait(page
->mapping
, &wbc
, FLUSH_INVALIDATE
);
1517 static int nfs_wb_page_priority(struct inode
*inode
, struct page
*page
,
1520 loff_t range_start
= page_offset(page
);
1521 loff_t range_end
= range_start
+ (loff_t
)(PAGE_CACHE_SIZE
- 1);
1522 struct writeback_control wbc
= {
1523 .bdi
= page
->mapping
->backing_dev_info
,
1524 .sync_mode
= WB_SYNC_ALL
,
1525 .nr_to_write
= LONG_MAX
,
1526 .range_start
= range_start
,
1527 .range_end
= range_end
,
1532 if (clear_page_dirty_for_io(page
)) {
1533 ret
= nfs_writepage_locked(page
, &wbc
);
1536 } else if (!PagePrivate(page
))
1538 ret
= nfs_sync_mapping_wait(page
->mapping
, &wbc
, how
);
1541 } while (PagePrivate(page
));
1544 __mark_inode_dirty(inode
, I_DIRTY_PAGES
);
1549 * Write back all requests on one page - we do this before reading it.
1551 int nfs_wb_page(struct inode
*inode
, struct page
* page
)
1553 return nfs_wb_page_priority(inode
, page
, FLUSH_STABLE
);
1556 int __init
nfs_init_writepagecache(void)
1558 nfs_wdata_cachep
= kmem_cache_create("nfs_write_data",
1559 sizeof(struct nfs_write_data
),
1560 0, SLAB_HWCACHE_ALIGN
,
1562 if (nfs_wdata_cachep
== NULL
)
1565 nfs_wdata_mempool
= mempool_create_slab_pool(MIN_POOL_WRITE
,
1567 if (nfs_wdata_mempool
== NULL
)
1570 nfs_commit_mempool
= mempool_create_slab_pool(MIN_POOL_COMMIT
,
1572 if (nfs_commit_mempool
== NULL
)
1576 * NFS congestion size, scale with available memory.
1588 * This allows larger machines to have larger/more transfers.
1589 * Limit the default to 256M
1591 nfs_congestion_kb
= (16*int_sqrt(totalram_pages
)) << (PAGE_SHIFT
-10);
1592 if (nfs_congestion_kb
> 256*1024)
1593 nfs_congestion_kb
= 256*1024;
1598 void nfs_destroy_writepagecache(void)
1600 mempool_destroy(nfs_commit_mempool
);
1601 mempool_destroy(nfs_wdata_mempool
);
1602 kmem_cache_destroy(nfs_wdata_cachep
);