4 * Write file data over NFS.
6 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
9 #include <linux/types.h>
10 #include <linux/slab.h>
12 #include <linux/pagemap.h>
13 #include <linux/file.h>
14 #include <linux/writeback.h>
15 #include <linux/swap.h>
16 #include <linux/migrate.h>
18 #include <linux/sunrpc/clnt.h>
19 #include <linux/nfs_fs.h>
20 #include <linux/nfs_mount.h>
21 #include <linux/nfs_page.h>
22 #include <linux/backing-dev.h>
23 #include <linux/export.h>
24 #include <linux/freezer.h>
25 #include <linux/wait.h>
27 #include <asm/uaccess.h>
29 #include "delegation.h"
38 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
40 #define MIN_POOL_WRITE (32)
41 #define MIN_POOL_COMMIT (4)
44 * Local function declarations
46 static void nfs_redirty_request(struct nfs_page
*req
);
47 static const struct rpc_call_ops nfs_commit_ops
;
48 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops
;
49 static const struct nfs_commit_completion_ops nfs_commit_completion_ops
;
50 static const struct nfs_rw_ops nfs_rw_write_ops
;
51 static void nfs_clear_request_commit(struct nfs_page
*req
);
52 static void nfs_init_cinfo_from_inode(struct nfs_commit_info
*cinfo
,
54 static struct nfs_page
*
55 nfs_page_search_commits_for_head_request_locked(struct nfs_inode
*nfsi
,
58 static struct kmem_cache
*nfs_wdata_cachep
;
59 static mempool_t
*nfs_wdata_mempool
;
60 static struct kmem_cache
*nfs_cdata_cachep
;
61 static mempool_t
*nfs_commit_mempool
;
63 struct nfs_commit_data
*nfs_commitdata_alloc(void)
65 struct nfs_commit_data
*p
= mempool_alloc(nfs_commit_mempool
, GFP_NOIO
);
68 memset(p
, 0, sizeof(*p
));
69 INIT_LIST_HEAD(&p
->pages
);
73 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc
);
75 void nfs_commit_free(struct nfs_commit_data
*p
)
77 mempool_free(p
, nfs_commit_mempool
);
79 EXPORT_SYMBOL_GPL(nfs_commit_free
);
81 static struct nfs_pgio_header
*nfs_writehdr_alloc(void)
83 struct nfs_pgio_header
*p
= mempool_alloc(nfs_wdata_mempool
, GFP_NOIO
);
86 memset(p
, 0, sizeof(*p
));
90 static void nfs_writehdr_free(struct nfs_pgio_header
*hdr
)
92 mempool_free(hdr
, nfs_wdata_mempool
);
95 static void nfs_context_set_write_error(struct nfs_open_context
*ctx
, int error
)
99 set_bit(NFS_CONTEXT_ERROR_WRITE
, &ctx
->flags
);
103 * nfs_page_find_head_request_locked - find head request associated with @page
105 * must be called while holding the inode lock.
107 * returns matching head request with reference held, or NULL if not found.
109 static struct nfs_page
*
110 nfs_page_find_head_request_locked(struct nfs_inode
*nfsi
, struct page
*page
)
112 struct nfs_page
*req
= NULL
;
114 if (PagePrivate(page
))
115 req
= (struct nfs_page
*)page_private(page
);
116 else if (unlikely(PageSwapCache(page
)))
117 req
= nfs_page_search_commits_for_head_request_locked(nfsi
,
121 WARN_ON_ONCE(req
->wb_head
!= req
);
122 kref_get(&req
->wb_kref
);
129 * nfs_page_find_head_request - find head request associated with @page
131 * returns matching head request with reference held, or NULL if not found.
133 static struct nfs_page
*nfs_page_find_head_request(struct page
*page
)
135 struct inode
*inode
= page_file_mapping(page
)->host
;
136 struct nfs_page
*req
= NULL
;
138 spin_lock(&inode
->i_lock
);
139 req
= nfs_page_find_head_request_locked(NFS_I(inode
), page
);
140 spin_unlock(&inode
->i_lock
);
144 /* Adjust the file length if we're writing beyond the end */
145 static void nfs_grow_file(struct page
*page
, unsigned int offset
, unsigned int count
)
147 struct inode
*inode
= page_file_mapping(page
)->host
;
151 spin_lock(&inode
->i_lock
);
152 i_size
= i_size_read(inode
);
153 end_index
= (i_size
- 1) >> PAGE_SHIFT
;
154 if (i_size
> 0 && page_file_index(page
) < end_index
)
156 end
= page_file_offset(page
) + ((loff_t
)offset
+count
);
159 i_size_write(inode
, end
);
160 nfs_inc_stats(inode
, NFSIOS_EXTENDWRITE
);
162 spin_unlock(&inode
->i_lock
);
165 /* A writeback failed: mark the page as bad, and invalidate the page cache */
166 static void nfs_set_pageerror(struct page
*page
)
168 nfs_zap_mapping(page_file_mapping(page
)->host
, page_file_mapping(page
));
172 * nfs_page_group_search_locked
173 * @head - head request of page group
174 * @page_offset - offset into page
176 * Search page group with head @head to find a request that contains the
177 * page offset @page_offset.
179 * Returns a pointer to the first matching nfs request, or NULL if no
182 * Must be called with the page group lock held
184 static struct nfs_page
*
185 nfs_page_group_search_locked(struct nfs_page
*head
, unsigned int page_offset
)
187 struct nfs_page
*req
;
189 WARN_ON_ONCE(head
!= head
->wb_head
);
190 WARN_ON_ONCE(!test_bit(PG_HEADLOCK
, &head
->wb_head
->wb_flags
));
194 if (page_offset
>= req
->wb_pgbase
&&
195 page_offset
< (req
->wb_pgbase
+ req
->wb_bytes
))
198 req
= req
->wb_this_page
;
199 } while (req
!= head
);
205 * nfs_page_group_covers_page
206 * @head - head request of page group
208 * Return true if the page group with head @head covers the whole page,
209 * returns false otherwise
211 static bool nfs_page_group_covers_page(struct nfs_page
*req
)
213 struct nfs_page
*tmp
;
214 unsigned int pos
= 0;
215 unsigned int len
= nfs_page_length(req
->wb_page
);
217 nfs_page_group_lock(req
, false);
220 tmp
= nfs_page_group_search_locked(req
->wb_head
, pos
);
222 /* no way this should happen */
223 WARN_ON_ONCE(tmp
->wb_pgbase
!= pos
);
224 pos
+= tmp
->wb_bytes
- (pos
- tmp
->wb_pgbase
);
226 } while (tmp
&& pos
< len
);
228 nfs_page_group_unlock(req
);
229 WARN_ON_ONCE(pos
> len
);
233 /* We can set the PG_uptodate flag if we see that a write request
234 * covers the full page.
236 static void nfs_mark_uptodate(struct nfs_page
*req
)
238 if (PageUptodate(req
->wb_page
))
240 if (!nfs_page_group_covers_page(req
))
242 SetPageUptodate(req
->wb_page
);
245 static int wb_priority(struct writeback_control
*wbc
)
249 if (wbc
->sync_mode
== WB_SYNC_ALL
)
250 ret
= FLUSH_COND_STABLE
;
255 * NFS congestion control
258 int nfs_congestion_kb
;
260 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
261 #define NFS_CONGESTION_OFF_THRESH \
262 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
264 static void nfs_set_page_writeback(struct page
*page
)
266 struct nfs_server
*nfss
= NFS_SERVER(page_file_mapping(page
)->host
);
267 int ret
= test_set_page_writeback(page
);
269 WARN_ON_ONCE(ret
!= 0);
271 if (atomic_long_inc_return(&nfss
->writeback
) >
272 NFS_CONGESTION_ON_THRESH
) {
273 set_bdi_congested(&nfss
->backing_dev_info
,
278 static void nfs_end_page_writeback(struct nfs_page
*req
)
280 struct inode
*inode
= page_file_mapping(req
->wb_page
)->host
;
281 struct nfs_server
*nfss
= NFS_SERVER(inode
);
283 if (!nfs_page_group_sync_on_bit(req
, PG_WB_END
))
286 end_page_writeback(req
->wb_page
);
287 if (atomic_long_dec_return(&nfss
->writeback
) < NFS_CONGESTION_OFF_THRESH
)
288 clear_bdi_congested(&nfss
->backing_dev_info
, BLK_RW_ASYNC
);
292 /* nfs_page_group_clear_bits
293 * @req - an nfs request
294 * clears all page group related bits from @req
297 nfs_page_group_clear_bits(struct nfs_page
*req
)
299 clear_bit(PG_TEARDOWN
, &req
->wb_flags
);
300 clear_bit(PG_UNLOCKPAGE
, &req
->wb_flags
);
301 clear_bit(PG_UPTODATE
, &req
->wb_flags
);
302 clear_bit(PG_WB_END
, &req
->wb_flags
);
303 clear_bit(PG_REMOVE
, &req
->wb_flags
);
308 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req
310 * this is a helper function for nfs_lock_and_join_requests
312 * @inode - inode associated with request page group, must be holding inode lock
313 * @head - head request of page group, must be holding head lock
314 * @req - request that couldn't lock and needs to wait on the req bit lock
315 * @nonblock - if true, don't actually wait
317 * NOTE: this must be called holding page_group bit lock and inode spin lock
318 * and BOTH will be released before returning.
320 * returns 0 on success, < 0 on error.
323 nfs_unroll_locks_and_wait(struct inode
*inode
, struct nfs_page
*head
,
324 struct nfs_page
*req
, bool nonblock
)
325 __releases(&inode
->i_lock
)
327 struct nfs_page
*tmp
;
330 /* relinquish all the locks successfully grabbed this run */
331 for (tmp
= head
; tmp
!= req
; tmp
= tmp
->wb_this_page
)
332 nfs_unlock_request(tmp
);
334 WARN_ON_ONCE(test_bit(PG_TEARDOWN
, &req
->wb_flags
));
336 /* grab a ref on the request that will be waited on */
337 kref_get(&req
->wb_kref
);
339 nfs_page_group_unlock(head
);
340 spin_unlock(&inode
->i_lock
);
342 /* release ref from nfs_page_find_head_request_locked */
343 nfs_release_request(head
);
346 ret
= nfs_wait_on_request(req
);
349 nfs_release_request(req
);
355 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
357 * @destroy_list - request list (using wb_this_page) terminated by @old_head
358 * @old_head - the old head of the list
360 * All subrequests must be locked and removed from all lists, so at this point
361 * they are only "active" in this function, and possibly in nfs_wait_on_request
362 * with a reference held by some other context.
365 nfs_destroy_unlinked_subrequests(struct nfs_page
*destroy_list
,
366 struct nfs_page
*old_head
)
368 while (destroy_list
) {
369 struct nfs_page
*subreq
= destroy_list
;
371 destroy_list
= (subreq
->wb_this_page
== old_head
) ?
372 NULL
: subreq
->wb_this_page
;
374 WARN_ON_ONCE(old_head
!= subreq
->wb_head
);
376 /* make sure old group is not used */
377 subreq
->wb_head
= subreq
;
378 subreq
->wb_this_page
= subreq
;
380 /* subreq is now totally disconnected from page group or any
381 * write / commit lists. last chance to wake any waiters */
382 nfs_unlock_request(subreq
);
384 if (!test_bit(PG_TEARDOWN
, &subreq
->wb_flags
)) {
385 /* release ref on old head request */
386 nfs_release_request(old_head
);
388 nfs_page_group_clear_bits(subreq
);
390 /* release the PG_INODE_REF reference */
391 if (test_and_clear_bit(PG_INODE_REF
, &subreq
->wb_flags
))
392 nfs_release_request(subreq
);
396 WARN_ON_ONCE(test_bit(PG_CLEAN
, &subreq
->wb_flags
));
397 /* zombie requests have already released the last
398 * reference and were waiting on the rest of the
399 * group to complete. Since it's no longer part of a
400 * group, simply free the request */
401 nfs_page_group_clear_bits(subreq
);
402 nfs_free_request(subreq
);
408 * nfs_lock_and_join_requests - join all subreqs to the head req and return
409 * a locked reference, cancelling any pending
410 * operations for this page.
412 * @page - the page used to lookup the "page group" of nfs_page structures
413 * @nonblock - if true, don't block waiting for request locks
415 * This function joins all sub requests to the head request by first
416 * locking all requests in the group, cancelling any pending operations
417 * and finally updating the head request to cover the whole range covered by
418 * the (former) group. All subrequests are removed from any write or commit
419 * lists, unlinked from the group and destroyed.
421 * Returns a locked, referenced pointer to the head request - which after
422 * this call is guaranteed to be the only request associated with the page.
423 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
424 * error was encountered.
426 static struct nfs_page
*
427 nfs_lock_and_join_requests(struct page
*page
, bool nonblock
)
429 struct inode
*inode
= page_file_mapping(page
)->host
;
430 struct nfs_page
*head
, *subreq
;
431 struct nfs_page
*destroy_list
= NULL
;
432 unsigned int total_bytes
;
438 WARN_ON_ONCE(destroy_list
);
440 spin_lock(&inode
->i_lock
);
443 * A reference is taken only on the head request which acts as a
444 * reference to the whole page group - the group will not be destroyed
445 * until the head reference is released.
447 head
= nfs_page_find_head_request_locked(NFS_I(inode
), page
);
450 spin_unlock(&inode
->i_lock
);
454 /* holding inode lock, so always make a non-blocking call to try the
456 ret
= nfs_page_group_lock(head
, true);
458 spin_unlock(&inode
->i_lock
);
460 if (!nonblock
&& ret
== -EAGAIN
) {
461 nfs_page_group_lock_wait(head
);
462 nfs_release_request(head
);
466 nfs_release_request(head
);
470 /* lock each request in the page group */
474 * Subrequests are always contiguous, non overlapping
475 * and in order - but may be repeated (mirrored writes).
477 if (subreq
->wb_offset
== (head
->wb_offset
+ total_bytes
)) {
478 /* keep track of how many bytes this group covers */
479 total_bytes
+= subreq
->wb_bytes
;
480 } else if (WARN_ON_ONCE(subreq
->wb_offset
< head
->wb_offset
||
481 ((subreq
->wb_offset
+ subreq
->wb_bytes
) >
482 (head
->wb_offset
+ total_bytes
)))) {
483 nfs_page_group_unlock(head
);
484 spin_unlock(&inode
->i_lock
);
485 return ERR_PTR(-EIO
);
488 if (!nfs_lock_request(subreq
)) {
489 /* releases page group bit lock and
490 * inode spin lock and all references */
491 ret
= nfs_unroll_locks_and_wait(inode
, head
,
500 subreq
= subreq
->wb_this_page
;
501 } while (subreq
!= head
);
503 /* Now that all requests are locked, make sure they aren't on any list.
504 * Commit list removal accounting is done after locks are dropped */
507 nfs_clear_request_commit(subreq
);
508 subreq
= subreq
->wb_this_page
;
509 } while (subreq
!= head
);
511 /* unlink subrequests from head, destroy them later */
512 if (head
->wb_this_page
!= head
) {
513 /* destroy list will be terminated by head */
514 destroy_list
= head
->wb_this_page
;
515 head
->wb_this_page
= head
;
517 /* change head request to cover whole range that
518 * the former page group covered */
519 head
->wb_bytes
= total_bytes
;
523 * prepare head request to be added to new pgio descriptor
525 nfs_page_group_clear_bits(head
);
528 * some part of the group was still on the inode list - otherwise
529 * the group wouldn't be involved in async write.
530 * grab a reference for the head request, iff it needs one.
532 if (!test_and_set_bit(PG_INODE_REF
, &head
->wb_flags
))
533 kref_get(&head
->wb_kref
);
535 nfs_page_group_unlock(head
);
537 /* drop lock to clean uprequests on destroy list */
538 spin_unlock(&inode
->i_lock
);
540 nfs_destroy_unlinked_subrequests(destroy_list
, head
);
542 /* still holds ref on head from nfs_page_find_head_request_locked
543 * and still has lock on head from lock loop */
547 static void nfs_write_error_remove_page(struct nfs_page
*req
)
549 nfs_unlock_request(req
);
550 nfs_end_page_writeback(req
);
551 nfs_release_request(req
);
552 generic_error_remove_page(page_file_mapping(req
->wb_page
),
557 * Find an associated nfs write request, and prepare to flush it out
558 * May return an error if the user signalled nfs_wait_on_request().
560 static int nfs_page_async_flush(struct nfs_pageio_descriptor
*pgio
,
561 struct page
*page
, bool nonblock
,
564 struct nfs_page
*req
;
567 req
= nfs_lock_and_join_requests(page
, nonblock
);
574 nfs_set_page_writeback(page
);
575 WARN_ON_ONCE(test_bit(PG_CLEAN
, &req
->wb_flags
));
578 if (!nfs_pageio_add_request(pgio
, req
)) {
579 ret
= pgio
->pg_error
;
581 * Remove the problematic req upon fatal errors
582 * in launder case, while other dirty pages can
583 * still be around until they get flushed.
585 if (nfs_error_is_fatal(ret
)) {
586 nfs_context_set_write_error(req
->wb_context
, ret
);
588 nfs_write_error_remove_page(req
);
592 nfs_redirty_request(req
);
595 nfs_add_stats(page_file_mapping(page
)->host
,
596 NFSIOS_WRITEPAGES
, 1);
601 static int nfs_do_writepage(struct page
*page
, struct writeback_control
*wbc
,
602 struct nfs_pageio_descriptor
*pgio
, bool launder
)
606 nfs_pageio_cond_complete(pgio
, page_file_index(page
));
607 ret
= nfs_page_async_flush(pgio
, page
, wbc
->sync_mode
== WB_SYNC_NONE
,
609 if (ret
== -EAGAIN
) {
610 redirty_page_for_writepage(wbc
, page
);
617 * Write an mmapped page to the server.
619 static int nfs_writepage_locked(struct page
*page
,
620 struct writeback_control
*wbc
,
623 struct nfs_pageio_descriptor pgio
;
624 struct inode
*inode
= page_file_mapping(page
)->host
;
627 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGE
);
628 nfs_pageio_init_write(&pgio
, inode
, wb_priority(wbc
),
629 false, &nfs_async_write_completion_ops
);
630 err
= nfs_do_writepage(page
, wbc
, &pgio
, launder
);
631 nfs_pageio_complete(&pgio
);
634 if (pgio
.pg_error
< 0)
635 return pgio
.pg_error
;
639 int nfs_writepage(struct page
*page
, struct writeback_control
*wbc
)
643 ret
= nfs_writepage_locked(page
, wbc
, false);
648 static int nfs_writepages_callback(struct page
*page
, struct writeback_control
*wbc
, void *data
)
652 ret
= nfs_do_writepage(page
, wbc
, data
, false);
657 int nfs_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
659 struct inode
*inode
= mapping
->host
;
660 unsigned long *bitlock
= &NFS_I(inode
)->flags
;
661 struct nfs_pageio_descriptor pgio
;
664 /* Stop dirtying of new pages while we sync */
665 err
= wait_on_bit_lock_action(bitlock
, NFS_INO_FLUSHING
,
666 nfs_wait_bit_killable
, TASK_KILLABLE
);
670 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGES
);
672 nfs_pageio_init_write(&pgio
, inode
, wb_priority(wbc
), false,
673 &nfs_async_write_completion_ops
);
674 err
= write_cache_pages(mapping
, wbc
, nfs_writepages_callback
, &pgio
);
675 nfs_pageio_complete(&pgio
);
677 clear_bit_unlock(NFS_INO_FLUSHING
, bitlock
);
678 smp_mb__after_atomic();
679 wake_up_bit(bitlock
, NFS_INO_FLUSHING
);
692 * Insert a write request into an inode
694 static void nfs_inode_add_request(struct inode
*inode
, struct nfs_page
*req
)
696 struct nfs_inode
*nfsi
= NFS_I(inode
);
698 WARN_ON_ONCE(req
->wb_this_page
!= req
);
700 /* Lock the request! */
701 nfs_lock_request(req
);
703 spin_lock(&inode
->i_lock
);
704 if (!nfsi
->nrequests
&&
705 NFS_PROTO(inode
)->have_delegation(inode
, FMODE_WRITE
))
708 * Swap-space should not get truncated. Hence no need to plug the race
709 * with invalidate/truncate.
711 if (likely(!PageSwapCache(req
->wb_page
))) {
712 set_bit(PG_MAPPED
, &req
->wb_flags
);
713 SetPagePrivate(req
->wb_page
);
714 set_page_private(req
->wb_page
, (unsigned long)req
);
717 /* this a head request for a page group - mark it as having an
718 * extra reference so sub groups can follow suit.
719 * This flag also informs pgio layer when to bump nrequests when
720 * adding subrequests. */
721 WARN_ON(test_and_set_bit(PG_INODE_REF
, &req
->wb_flags
));
722 kref_get(&req
->wb_kref
);
723 spin_unlock(&inode
->i_lock
);
727 * Remove a write request from an inode
729 static void nfs_inode_remove_request(struct nfs_page
*req
)
731 struct inode
*inode
= d_inode(req
->wb_context
->dentry
);
732 struct nfs_inode
*nfsi
= NFS_I(inode
);
733 struct nfs_page
*head
;
735 if (nfs_page_group_sync_on_bit(req
, PG_REMOVE
)) {
738 spin_lock(&inode
->i_lock
);
739 if (likely(head
->wb_page
&& !PageSwapCache(head
->wb_page
))) {
740 set_page_private(head
->wb_page
, 0);
741 ClearPagePrivate(head
->wb_page
);
742 smp_mb__after_atomic();
743 wake_up_page(head
->wb_page
, PG_private
);
744 clear_bit(PG_MAPPED
, &head
->wb_flags
);
747 spin_unlock(&inode
->i_lock
);
749 spin_lock(&inode
->i_lock
);
751 spin_unlock(&inode
->i_lock
);
754 if (test_and_clear_bit(PG_INODE_REF
, &req
->wb_flags
))
755 nfs_release_request(req
);
759 nfs_mark_request_dirty(struct nfs_page
*req
)
762 __set_page_dirty_nobuffers(req
->wb_page
);
766 * nfs_page_search_commits_for_head_request_locked
768 * Search through commit lists on @inode for the head request for @page.
769 * Must be called while holding the inode (which is cinfo) lock.
771 * Returns the head request if found, or NULL if not found.
773 static struct nfs_page
*
774 nfs_page_search_commits_for_head_request_locked(struct nfs_inode
*nfsi
,
777 struct nfs_page
*freq
, *t
;
778 struct nfs_commit_info cinfo
;
779 struct inode
*inode
= &nfsi
->vfs_inode
;
781 nfs_init_cinfo_from_inode(&cinfo
, inode
);
783 /* search through pnfs commit lists */
784 freq
= pnfs_search_commit_reqs(inode
, &cinfo
, page
);
786 return freq
->wb_head
;
788 /* Linearly search the commit list for the correct request */
789 list_for_each_entry_safe(freq
, t
, &cinfo
.mds
->list
, wb_list
) {
790 if (freq
->wb_page
== page
)
791 return freq
->wb_head
;
798 * nfs_request_add_commit_list_locked - add request to a commit list
799 * @req: pointer to a struct nfs_page
800 * @dst: commit list head
801 * @cinfo: holds list lock and accounting info
803 * This sets the PG_CLEAN bit, updates the cinfo count of
804 * number of outstanding requests requiring a commit as well as
807 * The caller must hold cinfo->inode->i_lock, and the nfs_page lock.
810 nfs_request_add_commit_list_locked(struct nfs_page
*req
, struct list_head
*dst
,
811 struct nfs_commit_info
*cinfo
)
813 set_bit(PG_CLEAN
, &req
->wb_flags
);
814 nfs_list_add_request(req
, dst
);
815 cinfo
->mds
->ncommit
++;
817 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked
);
820 * nfs_request_add_commit_list - add request to a commit list
821 * @req: pointer to a struct nfs_page
822 * @dst: commit list head
823 * @cinfo: holds list lock and accounting info
825 * This sets the PG_CLEAN bit, updates the cinfo count of
826 * number of outstanding requests requiring a commit as well as
829 * The caller must _not_ hold the cinfo->lock, but must be
830 * holding the nfs_page lock.
833 nfs_request_add_commit_list(struct nfs_page
*req
, struct nfs_commit_info
*cinfo
)
835 spin_lock(&cinfo
->inode
->i_lock
);
836 nfs_request_add_commit_list_locked(req
, &cinfo
->mds
->list
, cinfo
);
837 spin_unlock(&cinfo
->inode
->i_lock
);
839 nfs_mark_page_unstable(req
->wb_page
, cinfo
);
841 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list
);
844 * nfs_request_remove_commit_list - Remove request from a commit list
845 * @req: pointer to a nfs_page
846 * @cinfo: holds list lock and accounting info
848 * This clears the PG_CLEAN bit, and updates the cinfo's count of
849 * number of outstanding requests requiring a commit
850 * It does not update the MM page stats.
852 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
855 nfs_request_remove_commit_list(struct nfs_page
*req
,
856 struct nfs_commit_info
*cinfo
)
858 if (!test_and_clear_bit(PG_CLEAN
, &(req
)->wb_flags
))
860 nfs_list_remove_request(req
);
861 cinfo
->mds
->ncommit
--;
863 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list
);
865 static void nfs_init_cinfo_from_inode(struct nfs_commit_info
*cinfo
,
868 cinfo
->inode
= inode
;
869 cinfo
->mds
= &NFS_I(inode
)->commit_info
;
870 cinfo
->ds
= pnfs_get_ds_info(inode
);
872 cinfo
->completion_ops
= &nfs_commit_completion_ops
;
875 void nfs_init_cinfo(struct nfs_commit_info
*cinfo
,
877 struct nfs_direct_req
*dreq
)
880 nfs_init_cinfo_from_dreq(cinfo
, dreq
);
882 nfs_init_cinfo_from_inode(cinfo
, inode
);
884 EXPORT_SYMBOL_GPL(nfs_init_cinfo
);
887 * Add a request to the inode's commit list.
890 nfs_mark_request_commit(struct nfs_page
*req
, struct pnfs_layout_segment
*lseg
,
891 struct nfs_commit_info
*cinfo
, u32 ds_commit_idx
)
893 if (pnfs_mark_request_commit(req
, lseg
, cinfo
, ds_commit_idx
))
895 nfs_request_add_commit_list(req
, cinfo
);
899 nfs_clear_page_commit(struct page
*page
)
901 dec_zone_page_state(page
, NR_UNSTABLE_NFS
);
902 dec_wb_stat(&inode_to_bdi(page_file_mapping(page
)->host
)->wb
,
906 /* Called holding inode (/cinfo) lock */
908 nfs_clear_request_commit(struct nfs_page
*req
)
910 if (test_bit(PG_CLEAN
, &req
->wb_flags
)) {
911 struct inode
*inode
= d_inode(req
->wb_context
->dentry
);
912 struct nfs_commit_info cinfo
;
914 nfs_init_cinfo_from_inode(&cinfo
, inode
);
915 if (!pnfs_clear_request_commit(req
, &cinfo
)) {
916 nfs_request_remove_commit_list(req
, &cinfo
);
918 nfs_clear_page_commit(req
->wb_page
);
922 int nfs_write_need_commit(struct nfs_pgio_header
*hdr
)
924 if (hdr
->verf
.committed
== NFS_DATA_SYNC
)
925 return hdr
->lseg
== NULL
;
926 return hdr
->verf
.committed
!= NFS_FILE_SYNC
;
929 static void nfs_write_completion(struct nfs_pgio_header
*hdr
)
931 struct nfs_commit_info cinfo
;
932 unsigned long bytes
= 0;
934 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
936 nfs_init_cinfo_from_inode(&cinfo
, hdr
->inode
);
937 while (!list_empty(&hdr
->pages
)) {
938 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
940 bytes
+= req
->wb_bytes
;
941 nfs_list_remove_request(req
);
942 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
) &&
943 (hdr
->good_bytes
< bytes
)) {
944 nfs_set_pageerror(req
->wb_page
);
945 nfs_context_set_write_error(req
->wb_context
, hdr
->error
);
948 if (nfs_write_need_commit(hdr
)) {
949 memcpy(&req
->wb_verf
, &hdr
->verf
.verifier
, sizeof(req
->wb_verf
));
950 nfs_mark_request_commit(req
, hdr
->lseg
, &cinfo
,
951 hdr
->pgio_mirror_idx
);
955 nfs_inode_remove_request(req
);
957 nfs_unlock_request(req
);
958 nfs_end_page_writeback(req
);
959 nfs_release_request(req
);
966 nfs_reqs_to_commit(struct nfs_commit_info
*cinfo
)
968 return cinfo
->mds
->ncommit
;
971 /* cinfo->inode->i_lock held by caller */
973 nfs_scan_commit_list(struct list_head
*src
, struct list_head
*dst
,
974 struct nfs_commit_info
*cinfo
, int max
)
976 struct nfs_page
*req
, *tmp
;
979 list_for_each_entry_safe(req
, tmp
, src
, wb_list
) {
980 if (!nfs_lock_request(req
))
982 kref_get(&req
->wb_kref
);
983 if (cond_resched_lock(&cinfo
->inode
->i_lock
))
984 list_safe_reset_next(req
, tmp
, wb_list
);
985 nfs_request_remove_commit_list(req
, cinfo
);
986 nfs_list_add_request(req
, dst
);
988 if ((ret
== max
) && !cinfo
->dreq
)
995 * nfs_scan_commit - Scan an inode for commit requests
996 * @inode: NFS inode to scan
997 * @dst: mds destination list
998 * @cinfo: mds and ds lists of reqs ready to commit
1000 * Moves requests from the inode's 'commit' request list.
1001 * The requests are *not* checked to ensure that they form a contiguous set.
1004 nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
,
1005 struct nfs_commit_info
*cinfo
)
1009 spin_lock(&cinfo
->inode
->i_lock
);
1010 if (cinfo
->mds
->ncommit
> 0) {
1011 const int max
= INT_MAX
;
1013 ret
= nfs_scan_commit_list(&cinfo
->mds
->list
, dst
,
1015 ret
+= pnfs_scan_commit_lists(inode
, cinfo
, max
- ret
);
1017 spin_unlock(&cinfo
->inode
->i_lock
);
1022 * Search for an existing write request, and attempt to update
1023 * it to reflect a new dirty region on a given page.
1025 * If the attempt fails, then the existing request is flushed out
1028 static struct nfs_page
*nfs_try_to_update_request(struct inode
*inode
,
1030 unsigned int offset
,
1033 struct nfs_page
*req
;
1038 if (!PagePrivate(page
))
1041 end
= offset
+ bytes
;
1042 spin_lock(&inode
->i_lock
);
1045 req
= nfs_page_find_head_request_locked(NFS_I(inode
), page
);
1049 /* should be handled by nfs_flush_incompatible */
1050 WARN_ON_ONCE(req
->wb_head
!= req
);
1051 WARN_ON_ONCE(req
->wb_this_page
!= req
);
1053 rqend
= req
->wb_offset
+ req
->wb_bytes
;
1055 * Tell the caller to flush out the request if
1056 * the offsets are non-contiguous.
1057 * Note: nfs_flush_incompatible() will already
1058 * have flushed out requests having wrong owners.
1061 || end
< req
->wb_offset
)
1064 if (nfs_lock_request(req
))
1067 /* The request is locked, so wait and then retry */
1068 spin_unlock(&inode
->i_lock
);
1069 error
= nfs_wait_on_request(req
);
1070 nfs_release_request(req
);
1073 spin_lock(&inode
->i_lock
);
1076 /* Okay, the request matches. Update the region */
1077 if (offset
< req
->wb_offset
) {
1078 req
->wb_offset
= offset
;
1079 req
->wb_pgbase
= offset
;
1082 req
->wb_bytes
= end
- req
->wb_offset
;
1084 req
->wb_bytes
= rqend
- req
->wb_offset
;
1087 nfs_clear_request_commit(req
);
1088 spin_unlock(&inode
->i_lock
);
1091 spin_unlock(&inode
->i_lock
);
1092 nfs_release_request(req
);
1093 error
= nfs_wb_page(inode
, page
);
1095 return ERR_PTR(error
);
1099 * Try to update an existing write request, or create one if there is none.
1101 * Note: Should always be called with the Page Lock held to prevent races
1102 * if we have to add a new request. Also assumes that the caller has
1103 * already called nfs_flush_incompatible() if necessary.
1105 static struct nfs_page
* nfs_setup_write_request(struct nfs_open_context
* ctx
,
1106 struct page
*page
, unsigned int offset
, unsigned int bytes
)
1108 struct inode
*inode
= page_file_mapping(page
)->host
;
1109 struct nfs_page
*req
;
1111 req
= nfs_try_to_update_request(inode
, page
, offset
, bytes
);
1114 req
= nfs_create_request(ctx
, page
, NULL
, offset
, bytes
);
1117 nfs_inode_add_request(inode
, req
);
1122 static int nfs_writepage_setup(struct nfs_open_context
*ctx
, struct page
*page
,
1123 unsigned int offset
, unsigned int count
)
1125 struct nfs_page
*req
;
1127 req
= nfs_setup_write_request(ctx
, page
, offset
, count
);
1129 return PTR_ERR(req
);
1130 /* Update file length */
1131 nfs_grow_file(page
, offset
, count
);
1132 nfs_mark_uptodate(req
);
1133 nfs_mark_request_dirty(req
);
1134 nfs_unlock_and_release_request(req
);
1138 int nfs_flush_incompatible(struct file
*file
, struct page
*page
)
1140 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
1141 struct nfs_lock_context
*l_ctx
;
1142 struct file_lock_context
*flctx
= file_inode(file
)->i_flctx
;
1143 struct nfs_page
*req
;
1144 int do_flush
, status
;
1146 * Look for a request corresponding to this page. If there
1147 * is one, and it belongs to another file, we flush it out
1148 * before we try to copy anything into the page. Do this
1149 * due to the lack of an ACCESS-type call in NFSv2.
1150 * Also do the same if we find a request from an existing
1154 req
= nfs_page_find_head_request(page
);
1157 l_ctx
= req
->wb_lock_context
;
1158 do_flush
= req
->wb_page
!= page
||
1159 !nfs_match_open_context(req
->wb_context
, ctx
);
1160 /* for now, flush if more than 1 request in page_group */
1161 do_flush
|= req
->wb_this_page
!= req
;
1162 if (l_ctx
&& flctx
&&
1163 !(list_empty_careful(&flctx
->flc_posix
) &&
1164 list_empty_careful(&flctx
->flc_flock
))) {
1165 do_flush
|= l_ctx
->lockowner
.l_owner
!= current
->files
1166 || l_ctx
->lockowner
.l_pid
!= current
->tgid
;
1168 nfs_release_request(req
);
1171 status
= nfs_wb_page(page_file_mapping(page
)->host
, page
);
1172 } while (status
== 0);
1177 * Avoid buffered writes when a open context credential's key would
1180 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1182 * Return 0 and set a credential flag which triggers the inode to flush
1183 * and performs NFS_FILE_SYNC writes if the key will expired within
1184 * RPC_KEY_EXPIRE_TIMEO.
1187 nfs_key_timeout_notify(struct file
*filp
, struct inode
*inode
)
1189 struct nfs_open_context
*ctx
= nfs_file_open_context(filp
);
1190 struct rpc_auth
*auth
= NFS_SERVER(inode
)->client
->cl_auth
;
1192 return rpcauth_key_timeout_notify(auth
, ctx
->cred
);
1196 * Test if the open context credential key is marked to expire soon.
1198 bool nfs_ctx_key_to_expire(struct nfs_open_context
*ctx
, struct inode
*inode
)
1200 struct rpc_auth
*auth
= NFS_SERVER(inode
)->client
->cl_auth
;
1202 return rpcauth_cred_key_to_expire(auth
, ctx
->cred
);
1206 * If the page cache is marked as unsafe or invalid, then we can't rely on
1207 * the PageUptodate() flag. In this case, we will need to turn off
1208 * write optimisations that depend on the page contents being correct.
1210 static bool nfs_write_pageuptodate(struct page
*page
, struct inode
*inode
)
1212 struct nfs_inode
*nfsi
= NFS_I(inode
);
1214 if (nfs_have_delegated_attributes(inode
))
1216 if (nfsi
->cache_validity
& NFS_INO_REVAL_PAGECACHE
)
1219 if (test_bit(NFS_INO_INVALIDATING
, &nfsi
->flags
))
1222 if (nfsi
->cache_validity
& NFS_INO_INVALID_DATA
)
1224 return PageUptodate(page
) != 0;
1228 is_whole_file_wrlock(struct file_lock
*fl
)
1230 return fl
->fl_start
== 0 && fl
->fl_end
== OFFSET_MAX
&&
1231 fl
->fl_type
== F_WRLCK
;
1234 /* If we know the page is up to date, and we're not using byte range locks (or
1235 * if we have the whole file locked for writing), it may be more efficient to
1236 * extend the write to cover the entire page in order to avoid fragmentation
1239 * If the file is opened for synchronous writes then we can just skip the rest
1242 static int nfs_can_extend_write(struct file
*file
, struct page
*page
, struct inode
*inode
)
1245 struct file_lock_context
*flctx
= inode
->i_flctx
;
1246 struct file_lock
*fl
;
1248 if (file
->f_flags
& O_DSYNC
)
1250 if (!nfs_write_pageuptodate(page
, inode
))
1252 if (NFS_PROTO(inode
)->have_delegation(inode
, FMODE_WRITE
))
1254 if (!flctx
|| (list_empty_careful(&flctx
->flc_flock
) &&
1255 list_empty_careful(&flctx
->flc_posix
)))
1258 /* Check to see if there are whole file write locks */
1260 spin_lock(&flctx
->flc_lock
);
1261 if (!list_empty(&flctx
->flc_posix
)) {
1262 fl
= list_first_entry(&flctx
->flc_posix
, struct file_lock
,
1264 if (is_whole_file_wrlock(fl
))
1266 } else if (!list_empty(&flctx
->flc_flock
)) {
1267 fl
= list_first_entry(&flctx
->flc_flock
, struct file_lock
,
1269 if (fl
->fl_type
== F_WRLCK
)
1272 spin_unlock(&flctx
->flc_lock
);
1277 * Update and possibly write a cached page of an NFS file.
1279 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1280 * things with a page scheduled for an RPC call (e.g. invalidate it).
1282 int nfs_updatepage(struct file
*file
, struct page
*page
,
1283 unsigned int offset
, unsigned int count
)
1285 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
1286 struct inode
*inode
= page_file_mapping(page
)->host
;
1289 nfs_inc_stats(inode
, NFSIOS_VFSUPDATEPAGE
);
1291 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1292 file
, count
, (long long)(page_file_offset(page
) + offset
));
1294 if (nfs_can_extend_write(file
, page
, inode
)) {
1295 count
= max(count
+ offset
, nfs_page_length(page
));
1299 status
= nfs_writepage_setup(ctx
, page
, offset
, count
);
1301 nfs_set_pageerror(page
);
1303 __set_page_dirty_nobuffers(page
);
1305 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1306 status
, (long long)i_size_read(inode
));
1310 static int flush_task_priority(int how
)
1312 switch (how
& (FLUSH_HIGHPRI
|FLUSH_LOWPRI
)) {
1314 return RPC_PRIORITY_HIGH
;
1316 return RPC_PRIORITY_LOW
;
1318 return RPC_PRIORITY_NORMAL
;
1321 static void nfs_initiate_write(struct nfs_pgio_header
*hdr
,
1322 struct rpc_message
*msg
,
1323 const struct nfs_rpc_ops
*rpc_ops
,
1324 struct rpc_task_setup
*task_setup_data
, int how
)
1326 int priority
= flush_task_priority(how
);
1328 task_setup_data
->priority
= priority
;
1329 rpc_ops
->write_setup(hdr
, msg
);
1331 nfs4_state_protect_write(NFS_SERVER(hdr
->inode
)->nfs_client
,
1332 &task_setup_data
->rpc_client
, msg
, hdr
);
1335 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1336 * call this on each, which will prepare them to be retried on next
1337 * writeback using standard nfs.
1339 static void nfs_redirty_request(struct nfs_page
*req
)
1341 nfs_mark_request_dirty(req
);
1342 set_bit(NFS_CONTEXT_RESEND_WRITES
, &req
->wb_context
->flags
);
1343 nfs_unlock_request(req
);
1344 nfs_end_page_writeback(req
);
1345 nfs_release_request(req
);
1348 static void nfs_async_write_error(struct list_head
*head
)
1350 struct nfs_page
*req
;
1352 while (!list_empty(head
)) {
1353 req
= nfs_list_entry(head
->next
);
1354 nfs_list_remove_request(req
);
1355 nfs_redirty_request(req
);
1359 static void nfs_async_write_reschedule_io(struct nfs_pgio_header
*hdr
)
1361 nfs_async_write_error(&hdr
->pages
);
1364 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops
= {
1365 .error_cleanup
= nfs_async_write_error
,
1366 .completion
= nfs_write_completion
,
1367 .reschedule_io
= nfs_async_write_reschedule_io
,
1370 void nfs_pageio_init_write(struct nfs_pageio_descriptor
*pgio
,
1371 struct inode
*inode
, int ioflags
, bool force_mds
,
1372 const struct nfs_pgio_completion_ops
*compl_ops
)
1374 struct nfs_server
*server
= NFS_SERVER(inode
);
1375 const struct nfs_pageio_ops
*pg_ops
= &nfs_pgio_rw_ops
;
1377 #ifdef CONFIG_NFS_V4_1
1378 if (server
->pnfs_curr_ld
&& !force_mds
)
1379 pg_ops
= server
->pnfs_curr_ld
->pg_write_ops
;
1381 nfs_pageio_init(pgio
, inode
, pg_ops
, compl_ops
, &nfs_rw_write_ops
,
1382 server
->wsize
, ioflags
);
1384 EXPORT_SYMBOL_GPL(nfs_pageio_init_write
);
1386 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor
*pgio
)
1388 struct nfs_pgio_mirror
*mirror
;
1390 if (pgio
->pg_ops
&& pgio
->pg_ops
->pg_cleanup
)
1391 pgio
->pg_ops
->pg_cleanup(pgio
);
1393 pgio
->pg_ops
= &nfs_pgio_rw_ops
;
1395 nfs_pageio_stop_mirroring(pgio
);
1397 mirror
= &pgio
->pg_mirrors
[0];
1398 mirror
->pg_bsize
= NFS_SERVER(pgio
->pg_inode
)->wsize
;
1400 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds
);
1403 void nfs_commit_prepare(struct rpc_task
*task
, void *calldata
)
1405 struct nfs_commit_data
*data
= calldata
;
1407 NFS_PROTO(data
->inode
)->commit_rpc_prepare(task
, data
);
1411 * Special version of should_remove_suid() that ignores capabilities.
1413 static int nfs_should_remove_suid(const struct inode
*inode
)
1415 umode_t mode
= inode
->i_mode
;
1418 /* suid always must be killed */
1419 if (unlikely(mode
& S_ISUID
))
1420 kill
= ATTR_KILL_SUID
;
1423 * sgid without any exec bits is just a mandatory locking mark; leave
1424 * it alone. If some exec bits are set, it's a real sgid; kill it.
1426 if (unlikely((mode
& S_ISGID
) && (mode
& S_IXGRP
)))
1427 kill
|= ATTR_KILL_SGID
;
1429 if (unlikely(kill
&& S_ISREG(mode
)))
1435 static void nfs_writeback_check_extend(struct nfs_pgio_header
*hdr
,
1436 struct nfs_fattr
*fattr
)
1438 struct nfs_pgio_args
*argp
= &hdr
->args
;
1439 struct nfs_pgio_res
*resp
= &hdr
->res
;
1440 u64 size
= argp
->offset
+ resp
->count
;
1442 if (!(fattr
->valid
& NFS_ATTR_FATTR_SIZE
))
1444 if (nfs_size_to_loff_t(fattr
->size
) < i_size_read(hdr
->inode
)) {
1445 fattr
->valid
&= ~NFS_ATTR_FATTR_SIZE
;
1448 if (size
!= fattr
->size
)
1450 /* Set attribute barrier */
1451 nfs_fattr_set_barrier(fattr
);
1452 /* ...and update size */
1453 fattr
->valid
|= NFS_ATTR_FATTR_SIZE
;
1456 void nfs_writeback_update_inode(struct nfs_pgio_header
*hdr
)
1458 struct nfs_fattr
*fattr
= &hdr
->fattr
;
1459 struct inode
*inode
= hdr
->inode
;
1461 spin_lock(&inode
->i_lock
);
1462 nfs_writeback_check_extend(hdr
, fattr
);
1463 nfs_post_op_update_inode_force_wcc_locked(inode
, fattr
);
1464 spin_unlock(&inode
->i_lock
);
1466 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode
);
1469 * This function is called when the WRITE call is complete.
1471 static int nfs_writeback_done(struct rpc_task
*task
,
1472 struct nfs_pgio_header
*hdr
,
1473 struct inode
*inode
)
1478 * ->write_done will attempt to use post-op attributes to detect
1479 * conflicting writes by other clients. A strict interpretation
1480 * of close-to-open would allow us to continue caching even if
1481 * another writer had changed the file, but some applications
1482 * depend on tighter cache coherency when writing.
1484 status
= NFS_PROTO(inode
)->write_done(task
, hdr
);
1487 nfs_add_stats(inode
, NFSIOS_SERVERWRITTENBYTES
, hdr
->res
.count
);
1489 if (hdr
->res
.verf
->committed
< hdr
->args
.stable
&&
1490 task
->tk_status
>= 0) {
1491 /* We tried a write call, but the server did not
1492 * commit data to stable storage even though we
1494 * Note: There is a known bug in Tru64 < 5.0 in which
1495 * the server reports NFS_DATA_SYNC, but performs
1496 * NFS_FILE_SYNC. We therefore implement this checking
1497 * as a dprintk() in order to avoid filling syslog.
1499 static unsigned long complain
;
1501 /* Note this will print the MDS for a DS write */
1502 if (time_before(complain
, jiffies
)) {
1503 dprintk("NFS: faulty NFS server %s:"
1504 " (committed = %d) != (stable = %d)\n",
1505 NFS_SERVER(inode
)->nfs_client
->cl_hostname
,
1506 hdr
->res
.verf
->committed
, hdr
->args
.stable
);
1507 complain
= jiffies
+ 300 * HZ
;
1511 /* Deal with the suid/sgid bit corner case */
1512 if (nfs_should_remove_suid(inode
))
1513 nfs_mark_for_revalidate(inode
);
1518 * This function is called when the WRITE call is complete.
1520 static void nfs_writeback_result(struct rpc_task
*task
,
1521 struct nfs_pgio_header
*hdr
)
1523 struct nfs_pgio_args
*argp
= &hdr
->args
;
1524 struct nfs_pgio_res
*resp
= &hdr
->res
;
1526 if (resp
->count
< argp
->count
) {
1527 static unsigned long complain
;
1529 /* This a short write! */
1530 nfs_inc_stats(hdr
->inode
, NFSIOS_SHORTWRITE
);
1532 /* Has the server at least made some progress? */
1533 if (resp
->count
== 0) {
1534 if (time_before(complain
, jiffies
)) {
1536 "NFS: Server wrote zero bytes, expected %u.\n",
1538 complain
= jiffies
+ 300 * HZ
;
1540 nfs_set_pgio_error(hdr
, -EIO
, argp
->offset
);
1541 task
->tk_status
= -EIO
;
1545 /* For non rpc-based layout drivers, retry-through-MDS */
1546 if (!task
->tk_ops
) {
1547 hdr
->pnfs_error
= -EAGAIN
;
1551 /* Was this an NFSv2 write or an NFSv3 stable write? */
1552 if (resp
->verf
->committed
!= NFS_UNSTABLE
) {
1553 /* Resend from where the server left off */
1554 hdr
->mds_offset
+= resp
->count
;
1555 argp
->offset
+= resp
->count
;
1556 argp
->pgbase
+= resp
->count
;
1557 argp
->count
-= resp
->count
;
1559 /* Resend as a stable write in order to avoid
1560 * headaches in the case of a server crash.
1562 argp
->stable
= NFS_FILE_SYNC
;
1564 rpc_restart_call_prepare(task
);
1568 static int wait_on_commit(struct nfs_mds_commit_info
*cinfo
)
1570 return wait_on_atomic_t(&cinfo
->rpcs_out
,
1571 nfs_wait_atomic_killable
, TASK_KILLABLE
);
1574 static void nfs_commit_begin(struct nfs_mds_commit_info
*cinfo
)
1576 atomic_inc(&cinfo
->rpcs_out
);
1579 static void nfs_commit_end(struct nfs_mds_commit_info
*cinfo
)
1581 if (atomic_dec_and_test(&cinfo
->rpcs_out
))
1582 wake_up_atomic_t(&cinfo
->rpcs_out
);
1585 void nfs_commitdata_release(struct nfs_commit_data
*data
)
1587 put_nfs_open_context(data
->context
);
1588 nfs_commit_free(data
);
1590 EXPORT_SYMBOL_GPL(nfs_commitdata_release
);
1592 int nfs_initiate_commit(struct rpc_clnt
*clnt
, struct nfs_commit_data
*data
,
1593 const struct nfs_rpc_ops
*nfs_ops
,
1594 const struct rpc_call_ops
*call_ops
,
1597 struct rpc_task
*task
;
1598 int priority
= flush_task_priority(how
);
1599 struct rpc_message msg
= {
1600 .rpc_argp
= &data
->args
,
1601 .rpc_resp
= &data
->res
,
1602 .rpc_cred
= data
->cred
,
1604 struct rpc_task_setup task_setup_data
= {
1605 .task
= &data
->task
,
1607 .rpc_message
= &msg
,
1608 .callback_ops
= call_ops
,
1609 .callback_data
= data
,
1610 .workqueue
= nfsiod_workqueue
,
1611 .flags
= RPC_TASK_ASYNC
| flags
,
1612 .priority
= priority
,
1614 /* Set up the initial task struct. */
1615 nfs_ops
->commit_setup(data
, &msg
);
1617 dprintk("NFS: initiated commit call\n");
1619 nfs4_state_protect(NFS_SERVER(data
->inode
)->nfs_client
,
1620 NFS_SP4_MACH_CRED_COMMIT
, &task_setup_data
.rpc_client
, &msg
);
1622 task
= rpc_run_task(&task_setup_data
);
1624 return PTR_ERR(task
);
1625 if (how
& FLUSH_SYNC
)
1626 rpc_wait_for_completion_task(task
);
1630 EXPORT_SYMBOL_GPL(nfs_initiate_commit
);
1632 static loff_t
nfs_get_lwb(struct list_head
*head
)
1635 struct nfs_page
*req
;
1637 list_for_each_entry(req
, head
, wb_list
)
1638 if (lwb
< (req_offset(req
) + req
->wb_bytes
))
1639 lwb
= req_offset(req
) + req
->wb_bytes
;
1645 * Set up the argument/result storage required for the RPC call.
1647 void nfs_init_commit(struct nfs_commit_data
*data
,
1648 struct list_head
*head
,
1649 struct pnfs_layout_segment
*lseg
,
1650 struct nfs_commit_info
*cinfo
)
1652 struct nfs_page
*first
= nfs_list_entry(head
->next
);
1653 struct inode
*inode
= d_inode(first
->wb_context
->dentry
);
1655 /* Set up the RPC argument and reply structs
1656 * NB: take care not to mess about with data->commit et al. */
1658 list_splice_init(head
, &data
->pages
);
1660 data
->inode
= inode
;
1661 data
->cred
= first
->wb_context
->cred
;
1662 data
->lseg
= lseg
; /* reference transferred */
1663 /* only set lwb for pnfs commit */
1665 data
->lwb
= nfs_get_lwb(&data
->pages
);
1666 data
->mds_ops
= &nfs_commit_ops
;
1667 data
->completion_ops
= cinfo
->completion_ops
;
1668 data
->dreq
= cinfo
->dreq
;
1670 data
->args
.fh
= NFS_FH(data
->inode
);
1671 /* Note: we always request a commit of the entire inode */
1672 data
->args
.offset
= 0;
1673 data
->args
.count
= 0;
1674 data
->context
= get_nfs_open_context(first
->wb_context
);
1675 data
->res
.fattr
= &data
->fattr
;
1676 data
->res
.verf
= &data
->verf
;
1677 nfs_fattr_init(&data
->fattr
);
1679 EXPORT_SYMBOL_GPL(nfs_init_commit
);
1681 void nfs_retry_commit(struct list_head
*page_list
,
1682 struct pnfs_layout_segment
*lseg
,
1683 struct nfs_commit_info
*cinfo
,
1686 struct nfs_page
*req
;
1688 while (!list_empty(page_list
)) {
1689 req
= nfs_list_entry(page_list
->next
);
1690 nfs_list_remove_request(req
);
1691 nfs_mark_request_commit(req
, lseg
, cinfo
, ds_commit_idx
);
1693 nfs_clear_page_commit(req
->wb_page
);
1694 nfs_unlock_and_release_request(req
);
1697 EXPORT_SYMBOL_GPL(nfs_retry_commit
);
1700 nfs_commit_resched_write(struct nfs_commit_info
*cinfo
,
1701 struct nfs_page
*req
)
1703 __set_page_dirty_nobuffers(req
->wb_page
);
1707 * Commit dirty pages
1710 nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
,
1711 struct nfs_commit_info
*cinfo
)
1713 struct nfs_commit_data
*data
;
1715 /* another commit raced with us */
1716 if (list_empty(head
))
1719 data
= nfs_commitdata_alloc();
1724 /* Set up the argument struct */
1725 nfs_init_commit(data
, head
, NULL
, cinfo
);
1726 atomic_inc(&cinfo
->mds
->rpcs_out
);
1727 return nfs_initiate_commit(NFS_CLIENT(inode
), data
, NFS_PROTO(inode
),
1728 data
->mds_ops
, how
, 0);
1730 nfs_retry_commit(head
, NULL
, cinfo
, 0);
1734 int nfs_commit_file(struct file
*file
, struct nfs_write_verifier
*verf
)
1736 struct inode
*inode
= file_inode(file
);
1737 struct nfs_open_context
*open
;
1738 struct nfs_commit_info cinfo
;
1739 struct nfs_page
*req
;
1742 open
= get_nfs_open_context(nfs_file_open_context(file
));
1743 req
= nfs_create_request(open
, NULL
, NULL
, 0, i_size_read(inode
));
1749 nfs_init_cinfo_from_inode(&cinfo
, inode
);
1751 memcpy(&req
->wb_verf
, verf
, sizeof(struct nfs_write_verifier
));
1752 nfs_request_add_commit_list(req
, &cinfo
);
1753 ret
= nfs_commit_inode(inode
, FLUSH_SYNC
);
1757 nfs_free_request(req
);
1759 put_nfs_open_context(open
);
1762 EXPORT_SYMBOL_GPL(nfs_commit_file
);
1765 * COMMIT call returned
1767 static void nfs_commit_done(struct rpc_task
*task
, void *calldata
)
1769 struct nfs_commit_data
*data
= calldata
;
1771 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1772 task
->tk_pid
, task
->tk_status
);
1774 /* Call the NFS version-specific code */
1775 NFS_PROTO(data
->inode
)->commit_done(task
, data
);
1778 static void nfs_commit_release_pages(struct nfs_commit_data
*data
)
1780 struct nfs_page
*req
;
1781 int status
= data
->task
.tk_status
;
1782 struct nfs_commit_info cinfo
;
1783 struct nfs_server
*nfss
;
1785 while (!list_empty(&data
->pages
)) {
1786 req
= nfs_list_entry(data
->pages
.next
);
1787 nfs_list_remove_request(req
);
1789 nfs_clear_page_commit(req
->wb_page
);
1791 dprintk("NFS: commit (%s/%llu %d@%lld)",
1792 req
->wb_context
->dentry
->d_sb
->s_id
,
1793 (unsigned long long)NFS_FILEID(d_inode(req
->wb_context
->dentry
)),
1795 (long long)req_offset(req
));
1797 nfs_context_set_write_error(req
->wb_context
, status
);
1798 nfs_inode_remove_request(req
);
1799 dprintk(", error = %d\n", status
);
1803 /* Okay, COMMIT succeeded, apparently. Check the verifier
1804 * returned by the server against all stored verfs. */
1805 if (!memcmp(&req
->wb_verf
, &data
->verf
.verifier
, sizeof(req
->wb_verf
))) {
1806 /* We have a match */
1807 nfs_inode_remove_request(req
);
1811 /* We have a mismatch. Write the page again */
1812 dprintk(" mismatch\n");
1813 nfs_mark_request_dirty(req
);
1814 set_bit(NFS_CONTEXT_RESEND_WRITES
, &req
->wb_context
->flags
);
1816 nfs_unlock_and_release_request(req
);
1818 nfss
= NFS_SERVER(data
->inode
);
1819 if (atomic_long_read(&nfss
->writeback
) < NFS_CONGESTION_OFF_THRESH
)
1820 clear_bdi_congested(&nfss
->backing_dev_info
, BLK_RW_ASYNC
);
1822 nfs_init_cinfo(&cinfo
, data
->inode
, data
->dreq
);
1823 nfs_commit_end(cinfo
.mds
);
1826 static void nfs_commit_release(void *calldata
)
1828 struct nfs_commit_data
*data
= calldata
;
1830 data
->completion_ops
->completion(data
);
1831 nfs_commitdata_release(calldata
);
1834 static const struct rpc_call_ops nfs_commit_ops
= {
1835 .rpc_call_prepare
= nfs_commit_prepare
,
1836 .rpc_call_done
= nfs_commit_done
,
1837 .rpc_release
= nfs_commit_release
,
1840 static const struct nfs_commit_completion_ops nfs_commit_completion_ops
= {
1841 .completion
= nfs_commit_release_pages
,
1842 .resched_write
= nfs_commit_resched_write
,
1845 int nfs_generic_commit_list(struct inode
*inode
, struct list_head
*head
,
1846 int how
, struct nfs_commit_info
*cinfo
)
1850 status
= pnfs_commit_list(inode
, head
, how
, cinfo
);
1851 if (status
== PNFS_NOT_ATTEMPTED
)
1852 status
= nfs_commit_list(inode
, head
, how
, cinfo
);
1856 int nfs_commit_inode(struct inode
*inode
, int how
)
1859 struct nfs_commit_info cinfo
;
1860 int may_wait
= how
& FLUSH_SYNC
;
1864 nfs_init_cinfo_from_inode(&cinfo
, inode
);
1865 nfs_commit_begin(cinfo
.mds
);
1866 res
= nfs_scan_commit(inode
, &head
, &cinfo
);
1868 error
= nfs_generic_commit_list(inode
, &head
, how
, &cinfo
);
1869 nfs_commit_end(cinfo
.mds
);
1873 goto out_mark_dirty
;
1874 error
= wait_on_commit(cinfo
.mds
);
1880 /* Note: If we exit without ensuring that the commit is complete,
1881 * we must mark the inode as dirty. Otherwise, future calls to
1882 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1883 * that the data is on the disk.
1886 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
1889 EXPORT_SYMBOL_GPL(nfs_commit_inode
);
1891 int nfs_write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1893 struct nfs_inode
*nfsi
= NFS_I(inode
);
1894 int flags
= FLUSH_SYNC
;
1897 /* no commits means nothing needs to be done */
1898 if (!nfsi
->commit_info
.ncommit
)
1901 if (wbc
->sync_mode
== WB_SYNC_NONE
) {
1902 /* Don't commit yet if this is a non-blocking flush and there
1903 * are a lot of outstanding writes for this mapping.
1905 if (nfsi
->commit_info
.ncommit
<= (nfsi
->nrequests
>> 1))
1906 goto out_mark_dirty
;
1908 /* don't wait for the COMMIT response */
1912 ret
= nfs_commit_inode(inode
, flags
);
1914 if (wbc
->sync_mode
== WB_SYNC_NONE
) {
1915 if (ret
< wbc
->nr_to_write
)
1916 wbc
->nr_to_write
-= ret
;
1918 wbc
->nr_to_write
= 0;
1923 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
1926 EXPORT_SYMBOL_GPL(nfs_write_inode
);
1929 * flush the inode to disk.
1931 int nfs_wb_all(struct inode
*inode
)
1935 trace_nfs_writeback_inode_enter(inode
);
1937 ret
= filemap_write_and_wait(inode
->i_mapping
);
1940 ret
= nfs_commit_inode(inode
, FLUSH_SYNC
);
1943 pnfs_sync_inode(inode
, true);
1947 trace_nfs_writeback_inode_exit(inode
, ret
);
1950 EXPORT_SYMBOL_GPL(nfs_wb_all
);
1952 int nfs_wb_page_cancel(struct inode
*inode
, struct page
*page
)
1954 struct nfs_page
*req
;
1957 wait_on_page_writeback(page
);
1959 /* blocking call to cancel all requests and join to a single (head)
1961 req
= nfs_lock_and_join_requests(page
, false);
1966 /* all requests from this page have been cancelled by
1967 * nfs_lock_and_join_requests, so just remove the head
1968 * request from the inode / page_private pointer and
1970 nfs_inode_remove_request(req
);
1971 nfs_unlock_and_release_request(req
);
1978 * Write back all requests on one page - we do this before reading it.
1980 int nfs_wb_single_page(struct inode
*inode
, struct page
*page
, bool launder
)
1982 loff_t range_start
= page_file_offset(page
);
1983 loff_t range_end
= range_start
+ (loff_t
)(PAGE_SIZE
- 1);
1984 struct writeback_control wbc
= {
1985 .sync_mode
= WB_SYNC_ALL
,
1987 .range_start
= range_start
,
1988 .range_end
= range_end
,
1992 trace_nfs_writeback_page_enter(inode
);
1995 wait_on_page_writeback(page
);
1996 if (clear_page_dirty_for_io(page
)) {
1997 ret
= nfs_writepage_locked(page
, &wbc
, launder
);
2003 if (!PagePrivate(page
))
2005 ret
= nfs_commit_inode(inode
, FLUSH_SYNC
);
2010 trace_nfs_writeback_page_exit(inode
, ret
);
2014 #ifdef CONFIG_MIGRATION
2015 int nfs_migrate_page(struct address_space
*mapping
, struct page
*newpage
,
2016 struct page
*page
, enum migrate_mode mode
)
2019 * If PagePrivate is set, then the page is currently associated with
2020 * an in-progress read or write request. Don't try to migrate it.
2022 * FIXME: we could do this in principle, but we'll need a way to ensure
2023 * that we can safely release the inode reference while holding
2026 if (PagePrivate(page
))
2029 if (!nfs_fscache_release_page(page
, GFP_KERNEL
))
2032 return migrate_page(mapping
, newpage
, page
, mode
);
2036 int __init
nfs_init_writepagecache(void)
2038 nfs_wdata_cachep
= kmem_cache_create("nfs_write_data",
2039 sizeof(struct nfs_pgio_header
),
2040 0, SLAB_HWCACHE_ALIGN
,
2042 if (nfs_wdata_cachep
== NULL
)
2045 nfs_wdata_mempool
= mempool_create_slab_pool(MIN_POOL_WRITE
,
2047 if (nfs_wdata_mempool
== NULL
)
2048 goto out_destroy_write_cache
;
2050 nfs_cdata_cachep
= kmem_cache_create("nfs_commit_data",
2051 sizeof(struct nfs_commit_data
),
2052 0, SLAB_HWCACHE_ALIGN
,
2054 if (nfs_cdata_cachep
== NULL
)
2055 goto out_destroy_write_mempool
;
2057 nfs_commit_mempool
= mempool_create_slab_pool(MIN_POOL_COMMIT
,
2059 if (nfs_commit_mempool
== NULL
)
2060 goto out_destroy_commit_cache
;
2063 * NFS congestion size, scale with available memory.
2075 * This allows larger machines to have larger/more transfers.
2076 * Limit the default to 256M
2078 nfs_congestion_kb
= (16*int_sqrt(totalram_pages
)) << (PAGE_SHIFT
-10);
2079 if (nfs_congestion_kb
> 256*1024)
2080 nfs_congestion_kb
= 256*1024;
2084 out_destroy_commit_cache
:
2085 kmem_cache_destroy(nfs_cdata_cachep
);
2086 out_destroy_write_mempool
:
2087 mempool_destroy(nfs_wdata_mempool
);
2088 out_destroy_write_cache
:
2089 kmem_cache_destroy(nfs_wdata_cachep
);
2093 void nfs_destroy_writepagecache(void)
2095 mempool_destroy(nfs_commit_mempool
);
2096 kmem_cache_destroy(nfs_cdata_cachep
);
2097 mempool_destroy(nfs_wdata_mempool
);
2098 kmem_cache_destroy(nfs_wdata_cachep
);
2101 static const struct nfs_rw_ops nfs_rw_write_ops
= {
2102 .rw_mode
= FMODE_WRITE
,
2103 .rw_alloc_header
= nfs_writehdr_alloc
,
2104 .rw_free_header
= nfs_writehdr_free
,
2105 .rw_done
= nfs_writeback_done
,
2106 .rw_result
= nfs_writeback_result
,
2107 .rw_initiate
= nfs_initiate_write
,