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>
26 #include <linux/iversion.h>
28 #include <linux/uaccess.h>
29 #include <linux/sched/mm.h>
31 #include "delegation.h"
40 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
42 #define MIN_POOL_WRITE (32)
43 #define MIN_POOL_COMMIT (4)
45 struct nfs_io_completion
{
46 void (*complete
)(void *data
);
52 * Local function declarations
54 static void nfs_redirty_request(struct nfs_page
*req
);
55 static const struct rpc_call_ops nfs_commit_ops
;
56 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops
;
57 static const struct nfs_commit_completion_ops nfs_commit_completion_ops
;
58 static const struct nfs_rw_ops nfs_rw_write_ops
;
59 static void nfs_clear_request_commit(struct nfs_page
*req
);
60 static void nfs_init_cinfo_from_inode(struct nfs_commit_info
*cinfo
,
62 static struct nfs_page
*
63 nfs_page_search_commits_for_head_request_locked(struct nfs_inode
*nfsi
,
66 static struct kmem_cache
*nfs_wdata_cachep
;
67 static mempool_t
*nfs_wdata_mempool
;
68 static struct kmem_cache
*nfs_cdata_cachep
;
69 static mempool_t
*nfs_commit_mempool
;
71 struct nfs_commit_data
*nfs_commitdata_alloc(bool never_fail
)
73 struct nfs_commit_data
*p
;
76 p
= mempool_alloc(nfs_commit_mempool
, GFP_NOIO
);
78 /* It is OK to do some reclaim, not no safe to wait
79 * for anything to be returned to the pool.
80 * mempool_alloc() cannot handle that particular combination,
81 * so we need two separate attempts.
83 p
= mempool_alloc(nfs_commit_mempool
, GFP_NOWAIT
);
85 p
= kmem_cache_alloc(nfs_cdata_cachep
, GFP_NOIO
|
86 __GFP_NOWARN
| __GFP_NORETRY
);
91 memset(p
, 0, sizeof(*p
));
92 INIT_LIST_HEAD(&p
->pages
);
95 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc
);
97 void nfs_commit_free(struct nfs_commit_data
*p
)
99 mempool_free(p
, nfs_commit_mempool
);
101 EXPORT_SYMBOL_GPL(nfs_commit_free
);
103 static struct nfs_pgio_header
*nfs_writehdr_alloc(void)
105 struct nfs_pgio_header
*p
= mempool_alloc(nfs_wdata_mempool
, GFP_NOIO
);
107 memset(p
, 0, sizeof(*p
));
108 p
->rw_mode
= FMODE_WRITE
;
112 static void nfs_writehdr_free(struct nfs_pgio_header
*hdr
)
114 mempool_free(hdr
, nfs_wdata_mempool
);
117 static struct nfs_io_completion
*nfs_io_completion_alloc(gfp_t gfp_flags
)
119 return kmalloc(sizeof(struct nfs_io_completion
), gfp_flags
);
122 static void nfs_io_completion_init(struct nfs_io_completion
*ioc
,
123 void (*complete
)(void *), void *data
)
125 ioc
->complete
= complete
;
127 kref_init(&ioc
->refcount
);
130 static void nfs_io_completion_release(struct kref
*kref
)
132 struct nfs_io_completion
*ioc
= container_of(kref
,
133 struct nfs_io_completion
, refcount
);
134 ioc
->complete(ioc
->data
);
138 static void nfs_io_completion_get(struct nfs_io_completion
*ioc
)
141 kref_get(&ioc
->refcount
);
144 static void nfs_io_completion_put(struct nfs_io_completion
*ioc
)
147 kref_put(&ioc
->refcount
, nfs_io_completion_release
);
150 static struct nfs_page
*
151 nfs_page_private_request(struct page
*page
)
153 if (!PagePrivate(page
))
155 return (struct nfs_page
*)page_private(page
);
159 * nfs_page_find_head_request_locked - find head request associated with @page
161 * must be called while holding the inode lock.
163 * returns matching head request with reference held, or NULL if not found.
165 static struct nfs_page
*
166 nfs_page_find_private_request(struct page
*page
)
168 struct address_space
*mapping
= page_file_mapping(page
);
169 struct nfs_page
*req
;
171 if (!PagePrivate(page
))
173 spin_lock(&mapping
->private_lock
);
174 req
= nfs_page_private_request(page
);
176 WARN_ON_ONCE(req
->wb_head
!= req
);
177 kref_get(&req
->wb_kref
);
179 spin_unlock(&mapping
->private_lock
);
183 static struct nfs_page
*
184 nfs_page_find_swap_request(struct page
*page
)
186 struct inode
*inode
= page_file_mapping(page
)->host
;
187 struct nfs_inode
*nfsi
= NFS_I(inode
);
188 struct nfs_page
*req
= NULL
;
189 if (!PageSwapCache(page
))
191 mutex_lock(&nfsi
->commit_mutex
);
192 if (PageSwapCache(page
)) {
193 req
= nfs_page_search_commits_for_head_request_locked(nfsi
,
196 WARN_ON_ONCE(req
->wb_head
!= req
);
197 kref_get(&req
->wb_kref
);
200 mutex_unlock(&nfsi
->commit_mutex
);
205 * nfs_page_find_head_request - find head request associated with @page
207 * returns matching head request with reference held, or NULL if not found.
209 static struct nfs_page
*nfs_page_find_head_request(struct page
*page
)
211 struct nfs_page
*req
;
213 req
= nfs_page_find_private_request(page
);
215 req
= nfs_page_find_swap_request(page
);
219 /* Adjust the file length if we're writing beyond the end */
220 static void nfs_grow_file(struct page
*page
, unsigned int offset
, unsigned int count
)
222 struct inode
*inode
= page_file_mapping(page
)->host
;
226 spin_lock(&inode
->i_lock
);
227 i_size
= i_size_read(inode
);
228 end_index
= (i_size
- 1) >> PAGE_SHIFT
;
229 if (i_size
> 0 && page_index(page
) < end_index
)
231 end
= page_file_offset(page
) + ((loff_t
)offset
+count
);
234 i_size_write(inode
, end
);
235 NFS_I(inode
)->cache_validity
&= ~NFS_INO_INVALID_SIZE
;
236 nfs_inc_stats(inode
, NFSIOS_EXTENDWRITE
);
238 spin_unlock(&inode
->i_lock
);
241 /* A writeback failed: mark the page as bad, and invalidate the page cache */
242 static void nfs_set_pageerror(struct address_space
*mapping
)
244 nfs_zap_mapping(mapping
->host
, mapping
);
248 * nfs_page_group_search_locked
249 * @head - head request of page group
250 * @page_offset - offset into page
252 * Search page group with head @head to find a request that contains the
253 * page offset @page_offset.
255 * Returns a pointer to the first matching nfs request, or NULL if no
258 * Must be called with the page group lock held
260 static struct nfs_page
*
261 nfs_page_group_search_locked(struct nfs_page
*head
, unsigned int page_offset
)
263 struct nfs_page
*req
;
267 if (page_offset
>= req
->wb_pgbase
&&
268 page_offset
< (req
->wb_pgbase
+ req
->wb_bytes
))
271 req
= req
->wb_this_page
;
272 } while (req
!= head
);
278 * nfs_page_group_covers_page
279 * @head - head request of page group
281 * Return true if the page group with head @head covers the whole page,
282 * returns false otherwise
284 static bool nfs_page_group_covers_page(struct nfs_page
*req
)
286 struct nfs_page
*tmp
;
287 unsigned int pos
= 0;
288 unsigned int len
= nfs_page_length(req
->wb_page
);
290 nfs_page_group_lock(req
);
293 tmp
= nfs_page_group_search_locked(req
->wb_head
, pos
);
296 pos
= tmp
->wb_pgbase
+ tmp
->wb_bytes
;
299 nfs_page_group_unlock(req
);
303 /* We can set the PG_uptodate flag if we see that a write request
304 * covers the full page.
306 static void nfs_mark_uptodate(struct nfs_page
*req
)
308 if (PageUptodate(req
->wb_page
))
310 if (!nfs_page_group_covers_page(req
))
312 SetPageUptodate(req
->wb_page
);
315 static int wb_priority(struct writeback_control
*wbc
)
319 if (wbc
->sync_mode
== WB_SYNC_ALL
)
320 ret
= FLUSH_COND_STABLE
;
325 * NFS congestion control
328 int nfs_congestion_kb
;
330 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
331 #define NFS_CONGESTION_OFF_THRESH \
332 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
334 static void nfs_set_page_writeback(struct page
*page
)
336 struct inode
*inode
= page_file_mapping(page
)->host
;
337 struct nfs_server
*nfss
= NFS_SERVER(inode
);
338 int ret
= test_set_page_writeback(page
);
340 WARN_ON_ONCE(ret
!= 0);
342 if (atomic_long_inc_return(&nfss
->writeback
) >
343 NFS_CONGESTION_ON_THRESH
)
344 set_bdi_congested(inode_to_bdi(inode
), BLK_RW_ASYNC
);
347 static void nfs_end_page_writeback(struct nfs_page
*req
)
349 struct inode
*inode
= page_file_mapping(req
->wb_page
)->host
;
350 struct nfs_server
*nfss
= NFS_SERVER(inode
);
353 is_done
= nfs_page_group_sync_on_bit(req
, PG_WB_END
);
354 nfs_unlock_request(req
);
358 end_page_writeback(req
->wb_page
);
359 if (atomic_long_dec_return(&nfss
->writeback
) < NFS_CONGESTION_OFF_THRESH
)
360 clear_bdi_congested(inode_to_bdi(inode
), BLK_RW_ASYNC
);
364 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req
366 * this is a helper function for nfs_lock_and_join_requests
368 * @inode - inode associated with request page group, must be holding inode lock
369 * @head - head request of page group, must be holding head lock
370 * @req - request that couldn't lock and needs to wait on the req bit lock
372 * NOTE: this must be called holding page_group bit lock
373 * which will be released before returning.
375 * returns 0 on success, < 0 on error.
378 nfs_unroll_locks(struct inode
*inode
, struct nfs_page
*head
,
379 struct nfs_page
*req
)
381 struct nfs_page
*tmp
;
383 /* relinquish all the locks successfully grabbed this run */
384 for (tmp
= head
->wb_this_page
; tmp
!= req
; tmp
= tmp
->wb_this_page
) {
385 if (!kref_read(&tmp
->wb_kref
))
387 nfs_unlock_and_release_request(tmp
);
392 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
394 * @destroy_list - request list (using wb_this_page) terminated by @old_head
395 * @old_head - the old head of the list
397 * All subrequests must be locked and removed from all lists, so at this point
398 * they are only "active" in this function, and possibly in nfs_wait_on_request
399 * with a reference held by some other context.
402 nfs_destroy_unlinked_subrequests(struct nfs_page
*destroy_list
,
403 struct nfs_page
*old_head
,
406 while (destroy_list
) {
407 struct nfs_page
*subreq
= destroy_list
;
409 destroy_list
= (subreq
->wb_this_page
== old_head
) ?
410 NULL
: subreq
->wb_this_page
;
412 WARN_ON_ONCE(old_head
!= subreq
->wb_head
);
414 /* make sure old group is not used */
415 subreq
->wb_this_page
= subreq
;
417 clear_bit(PG_REMOVE
, &subreq
->wb_flags
);
419 /* Note: races with nfs_page_group_destroy() */
420 if (!kref_read(&subreq
->wb_kref
)) {
421 /* Check if we raced with nfs_page_group_destroy() */
422 if (test_and_clear_bit(PG_TEARDOWN
, &subreq
->wb_flags
))
423 nfs_free_request(subreq
);
427 subreq
->wb_head
= subreq
;
429 if (test_and_clear_bit(PG_INODE_REF
, &subreq
->wb_flags
)) {
430 nfs_release_request(subreq
);
431 atomic_long_dec(&NFS_I(inode
)->nrequests
);
434 /* subreq is now totally disconnected from page group or any
435 * write / commit lists. last chance to wake any waiters */
436 nfs_unlock_and_release_request(subreq
);
441 * nfs_lock_and_join_requests - join all subreqs to the head req and return
442 * a locked reference, cancelling any pending
443 * operations for this page.
445 * @page - the page used to lookup the "page group" of nfs_page structures
447 * This function joins all sub requests to the head request by first
448 * locking all requests in the group, cancelling any pending operations
449 * and finally updating the head request to cover the whole range covered by
450 * the (former) group. All subrequests are removed from any write or commit
451 * lists, unlinked from the group and destroyed.
453 * Returns a locked, referenced pointer to the head request - which after
454 * this call is guaranteed to be the only request associated with the page.
455 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
456 * error was encountered.
458 static struct nfs_page
*
459 nfs_lock_and_join_requests(struct page
*page
)
461 struct inode
*inode
= page_file_mapping(page
)->host
;
462 struct nfs_page
*head
, *subreq
;
463 struct nfs_page
*destroy_list
= NULL
;
464 unsigned int total_bytes
;
469 * A reference is taken only on the head request which acts as a
470 * reference to the whole page group - the group will not be destroyed
471 * until the head reference is released.
473 head
= nfs_page_find_head_request(page
);
477 /* lock the page head first in order to avoid an ABBA inefficiency */
478 if (!nfs_lock_request(head
)) {
479 ret
= nfs_wait_on_request(head
);
480 nfs_release_request(head
);
486 /* Ensure that nobody removed the request before we locked it */
487 if (head
!= nfs_page_private_request(page
) && !PageSwapCache(page
)) {
488 nfs_unlock_and_release_request(head
);
492 ret
= nfs_page_group_lock(head
);
494 goto release_request
;
496 /* lock each request in the page group */
497 total_bytes
= head
->wb_bytes
;
498 for (subreq
= head
->wb_this_page
; subreq
!= head
;
499 subreq
= subreq
->wb_this_page
) {
501 if (!kref_get_unless_zero(&subreq
->wb_kref
)) {
502 if (subreq
->wb_offset
== head
->wb_offset
+ total_bytes
)
503 total_bytes
+= subreq
->wb_bytes
;
507 while (!nfs_lock_request(subreq
)) {
509 * Unlock page to allow nfs_page_group_sync_on_bit()
512 nfs_page_group_unlock(head
);
513 ret
= nfs_wait_on_request(subreq
);
515 ret
= nfs_page_group_lock(head
);
517 nfs_unroll_locks(inode
, head
, subreq
);
518 nfs_release_request(subreq
);
519 goto release_request
;
523 * Subrequests are always contiguous, non overlapping
524 * and in order - but may be repeated (mirrored writes).
526 if (subreq
->wb_offset
== (head
->wb_offset
+ total_bytes
)) {
527 /* keep track of how many bytes this group covers */
528 total_bytes
+= subreq
->wb_bytes
;
529 } else if (WARN_ON_ONCE(subreq
->wb_offset
< head
->wb_offset
||
530 ((subreq
->wb_offset
+ subreq
->wb_bytes
) >
531 (head
->wb_offset
+ total_bytes
)))) {
532 nfs_page_group_unlock(head
);
533 nfs_unroll_locks(inode
, head
, subreq
);
534 nfs_unlock_and_release_request(subreq
);
536 goto release_request
;
540 /* Now that all requests are locked, make sure they aren't on any list.
541 * Commit list removal accounting is done after locks are dropped */
544 nfs_clear_request_commit(subreq
);
545 subreq
= subreq
->wb_this_page
;
546 } while (subreq
!= head
);
548 /* unlink subrequests from head, destroy them later */
549 if (head
->wb_this_page
!= head
) {
550 /* destroy list will be terminated by head */
551 destroy_list
= head
->wb_this_page
;
552 head
->wb_this_page
= head
;
554 /* change head request to cover whole range that
555 * the former page group covered */
556 head
->wb_bytes
= total_bytes
;
559 /* Postpone destruction of this request */
560 if (test_and_clear_bit(PG_REMOVE
, &head
->wb_flags
)) {
561 set_bit(PG_INODE_REF
, &head
->wb_flags
);
562 kref_get(&head
->wb_kref
);
563 atomic_long_inc(&NFS_I(inode
)->nrequests
);
566 nfs_page_group_unlock(head
);
568 nfs_destroy_unlinked_subrequests(destroy_list
, head
, inode
);
570 /* Did we lose a race with nfs_inode_remove_request()? */
571 if (!(PagePrivate(page
) || PageSwapCache(page
))) {
572 nfs_unlock_and_release_request(head
);
576 /* still holds ref on head from nfs_page_find_head_request
577 * and still has lock on head from lock loop */
581 nfs_unlock_and_release_request(head
);
585 static void nfs_write_error_remove_page(struct nfs_page
*req
)
587 nfs_end_page_writeback(req
);
588 generic_error_remove_page(page_file_mapping(req
->wb_page
),
590 nfs_release_request(req
);
594 nfs_error_is_fatal_on_server(int err
)
602 return nfs_error_is_fatal(err
);
606 * Find an associated nfs write request, and prepare to flush it out
607 * May return an error if the user signalled nfs_wait_on_request().
609 static int nfs_page_async_flush(struct nfs_pageio_descriptor
*pgio
,
612 struct nfs_page
*req
;
615 req
= nfs_lock_and_join_requests(page
);
622 nfs_set_page_writeback(page
);
623 WARN_ON_ONCE(test_bit(PG_CLEAN
, &req
->wb_flags
));
625 ret
= req
->wb_context
->error
;
626 /* If there is a fatal error that covers this write, just exit */
627 if (nfs_error_is_fatal_on_server(ret
))
631 if (!nfs_pageio_add_request(pgio
, req
)) {
632 ret
= pgio
->pg_error
;
634 * Remove the problematic req upon fatal errors on the server
636 if (nfs_error_is_fatal(ret
)) {
637 nfs_context_set_write_error(req
->wb_context
, ret
);
638 if (nfs_error_is_fatal_on_server(ret
))
642 nfs_redirty_request(req
);
644 nfs_add_stats(page_file_mapping(page
)->host
,
645 NFSIOS_WRITEPAGES
, 1);
649 nfs_write_error_remove_page(req
);
653 static int nfs_do_writepage(struct page
*page
, struct writeback_control
*wbc
,
654 struct nfs_pageio_descriptor
*pgio
)
658 nfs_pageio_cond_complete(pgio
, page_index(page
));
659 ret
= nfs_page_async_flush(pgio
, page
);
660 if (ret
== -EAGAIN
) {
661 redirty_page_for_writepage(wbc
, page
);
668 * Write an mmapped page to the server.
670 static int nfs_writepage_locked(struct page
*page
,
671 struct writeback_control
*wbc
)
673 struct nfs_pageio_descriptor pgio
;
674 struct inode
*inode
= page_file_mapping(page
)->host
;
677 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGE
);
678 nfs_pageio_init_write(&pgio
, inode
, 0,
679 false, &nfs_async_write_completion_ops
);
680 err
= nfs_do_writepage(page
, wbc
, &pgio
);
681 nfs_pageio_complete(&pgio
);
684 if (pgio
.pg_error
< 0)
685 return pgio
.pg_error
;
689 int nfs_writepage(struct page
*page
, struct writeback_control
*wbc
)
693 ret
= nfs_writepage_locked(page
, wbc
);
698 static int nfs_writepages_callback(struct page
*page
, struct writeback_control
*wbc
, void *data
)
702 ret
= nfs_do_writepage(page
, wbc
, data
);
707 static void nfs_io_completion_commit(void *inode
)
709 nfs_commit_inode(inode
, 0);
712 int nfs_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
714 struct inode
*inode
= mapping
->host
;
715 struct nfs_pageio_descriptor pgio
;
716 struct nfs_io_completion
*ioc
;
717 unsigned int pflags
= memalloc_nofs_save();
720 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGES
);
722 ioc
= nfs_io_completion_alloc(GFP_NOFS
);
724 nfs_io_completion_init(ioc
, nfs_io_completion_commit
, inode
);
726 nfs_pageio_init_write(&pgio
, inode
, wb_priority(wbc
), false,
727 &nfs_async_write_completion_ops
);
728 pgio
.pg_io_completion
= ioc
;
729 err
= write_cache_pages(mapping
, wbc
, nfs_writepages_callback
, &pgio
);
730 nfs_pageio_complete(&pgio
);
731 nfs_io_completion_put(ioc
);
733 memalloc_nofs_restore(pflags
);
746 * Insert a write request into an inode
748 static void nfs_inode_add_request(struct inode
*inode
, struct nfs_page
*req
)
750 struct address_space
*mapping
= page_file_mapping(req
->wb_page
);
751 struct nfs_inode
*nfsi
= NFS_I(inode
);
753 WARN_ON_ONCE(req
->wb_this_page
!= req
);
755 /* Lock the request! */
756 nfs_lock_request(req
);
759 * Swap-space should not get truncated. Hence no need to plug the race
760 * with invalidate/truncate.
762 spin_lock(&mapping
->private_lock
);
763 if (!nfs_have_writebacks(inode
) &&
764 NFS_PROTO(inode
)->have_delegation(inode
, FMODE_WRITE
))
765 inode_inc_iversion_raw(inode
);
766 if (likely(!PageSwapCache(req
->wb_page
))) {
767 set_bit(PG_MAPPED
, &req
->wb_flags
);
768 SetPagePrivate(req
->wb_page
);
769 set_page_private(req
->wb_page
, (unsigned long)req
);
771 spin_unlock(&mapping
->private_lock
);
772 atomic_long_inc(&nfsi
->nrequests
);
773 /* this a head request for a page group - mark it as having an
774 * extra reference so sub groups can follow suit.
775 * This flag also informs pgio layer when to bump nrequests when
776 * adding subrequests. */
777 WARN_ON(test_and_set_bit(PG_INODE_REF
, &req
->wb_flags
));
778 kref_get(&req
->wb_kref
);
782 * Remove a write request from an inode
784 static void nfs_inode_remove_request(struct nfs_page
*req
)
786 struct address_space
*mapping
= page_file_mapping(req
->wb_page
);
787 struct inode
*inode
= mapping
->host
;
788 struct nfs_inode
*nfsi
= NFS_I(inode
);
789 struct nfs_page
*head
;
791 atomic_long_dec(&nfsi
->nrequests
);
792 if (nfs_page_group_sync_on_bit(req
, PG_REMOVE
)) {
795 spin_lock(&mapping
->private_lock
);
796 if (likely(head
->wb_page
&& !PageSwapCache(head
->wb_page
))) {
797 set_page_private(head
->wb_page
, 0);
798 ClearPagePrivate(head
->wb_page
);
799 clear_bit(PG_MAPPED
, &head
->wb_flags
);
801 spin_unlock(&mapping
->private_lock
);
804 if (test_and_clear_bit(PG_INODE_REF
, &req
->wb_flags
))
805 nfs_release_request(req
);
809 nfs_mark_request_dirty(struct nfs_page
*req
)
812 __set_page_dirty_nobuffers(req
->wb_page
);
816 * nfs_page_search_commits_for_head_request_locked
818 * Search through commit lists on @inode for the head request for @page.
819 * Must be called while holding the inode (which is cinfo) lock.
821 * Returns the head request if found, or NULL if not found.
823 static struct nfs_page
*
824 nfs_page_search_commits_for_head_request_locked(struct nfs_inode
*nfsi
,
827 struct nfs_page
*freq
, *t
;
828 struct nfs_commit_info cinfo
;
829 struct inode
*inode
= &nfsi
->vfs_inode
;
831 nfs_init_cinfo_from_inode(&cinfo
, inode
);
833 /* search through pnfs commit lists */
834 freq
= pnfs_search_commit_reqs(inode
, &cinfo
, page
);
836 return freq
->wb_head
;
838 /* Linearly search the commit list for the correct request */
839 list_for_each_entry_safe(freq
, t
, &cinfo
.mds
->list
, wb_list
) {
840 if (freq
->wb_page
== page
)
841 return freq
->wb_head
;
848 * nfs_request_add_commit_list_locked - add request to a commit list
849 * @req: pointer to a struct nfs_page
850 * @dst: commit list head
851 * @cinfo: holds list lock and accounting info
853 * This sets the PG_CLEAN bit, updates the cinfo count of
854 * number of outstanding requests requiring a commit as well as
857 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
861 nfs_request_add_commit_list_locked(struct nfs_page
*req
, struct list_head
*dst
,
862 struct nfs_commit_info
*cinfo
)
864 set_bit(PG_CLEAN
, &req
->wb_flags
);
865 nfs_list_add_request(req
, dst
);
866 atomic_long_inc(&cinfo
->mds
->ncommit
);
868 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked
);
871 * nfs_request_add_commit_list - add request to a commit list
872 * @req: pointer to a struct nfs_page
873 * @cinfo: holds list lock and accounting info
875 * This sets the PG_CLEAN bit, updates the cinfo count of
876 * number of outstanding requests requiring a commit as well as
879 * The caller must _not_ hold the cinfo->lock, but must be
880 * holding the nfs_page lock.
883 nfs_request_add_commit_list(struct nfs_page
*req
, struct nfs_commit_info
*cinfo
)
885 mutex_lock(&NFS_I(cinfo
->inode
)->commit_mutex
);
886 nfs_request_add_commit_list_locked(req
, &cinfo
->mds
->list
, cinfo
);
887 mutex_unlock(&NFS_I(cinfo
->inode
)->commit_mutex
);
889 nfs_mark_page_unstable(req
->wb_page
, cinfo
);
891 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list
);
894 * nfs_request_remove_commit_list - Remove request from a commit list
895 * @req: pointer to a nfs_page
896 * @cinfo: holds list lock and accounting info
898 * This clears the PG_CLEAN bit, and updates the cinfo's count of
899 * number of outstanding requests requiring a commit
900 * It does not update the MM page stats.
902 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
905 nfs_request_remove_commit_list(struct nfs_page
*req
,
906 struct nfs_commit_info
*cinfo
)
908 if (!test_and_clear_bit(PG_CLEAN
, &(req
)->wb_flags
))
910 nfs_list_remove_request(req
);
911 atomic_long_dec(&cinfo
->mds
->ncommit
);
913 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list
);
915 static void nfs_init_cinfo_from_inode(struct nfs_commit_info
*cinfo
,
918 cinfo
->inode
= inode
;
919 cinfo
->mds
= &NFS_I(inode
)->commit_info
;
920 cinfo
->ds
= pnfs_get_ds_info(inode
);
922 cinfo
->completion_ops
= &nfs_commit_completion_ops
;
925 void nfs_init_cinfo(struct nfs_commit_info
*cinfo
,
927 struct nfs_direct_req
*dreq
)
930 nfs_init_cinfo_from_dreq(cinfo
, dreq
);
932 nfs_init_cinfo_from_inode(cinfo
, inode
);
934 EXPORT_SYMBOL_GPL(nfs_init_cinfo
);
937 * Add a request to the inode's commit list.
940 nfs_mark_request_commit(struct nfs_page
*req
, struct pnfs_layout_segment
*lseg
,
941 struct nfs_commit_info
*cinfo
, u32 ds_commit_idx
)
943 if (pnfs_mark_request_commit(req
, lseg
, cinfo
, ds_commit_idx
))
945 nfs_request_add_commit_list(req
, cinfo
);
949 nfs_clear_page_commit(struct page
*page
)
951 dec_node_page_state(page
, NR_UNSTABLE_NFS
);
952 dec_wb_stat(&inode_to_bdi(page_file_mapping(page
)->host
)->wb
,
956 /* Called holding the request lock on @req */
958 nfs_clear_request_commit(struct nfs_page
*req
)
960 if (test_bit(PG_CLEAN
, &req
->wb_flags
)) {
961 struct inode
*inode
= d_inode(req
->wb_context
->dentry
);
962 struct nfs_commit_info cinfo
;
964 nfs_init_cinfo_from_inode(&cinfo
, inode
);
965 mutex_lock(&NFS_I(inode
)->commit_mutex
);
966 if (!pnfs_clear_request_commit(req
, &cinfo
)) {
967 nfs_request_remove_commit_list(req
, &cinfo
);
969 mutex_unlock(&NFS_I(inode
)->commit_mutex
);
970 nfs_clear_page_commit(req
->wb_page
);
974 int nfs_write_need_commit(struct nfs_pgio_header
*hdr
)
976 if (hdr
->verf
.committed
== NFS_DATA_SYNC
)
977 return hdr
->lseg
== NULL
;
978 return hdr
->verf
.committed
!= NFS_FILE_SYNC
;
981 static void nfs_async_write_init(struct nfs_pgio_header
*hdr
)
983 nfs_io_completion_get(hdr
->io_completion
);
986 static void nfs_write_completion(struct nfs_pgio_header
*hdr
)
988 struct nfs_commit_info cinfo
;
989 unsigned long bytes
= 0;
991 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
993 nfs_init_cinfo_from_inode(&cinfo
, hdr
->inode
);
994 while (!list_empty(&hdr
->pages
)) {
995 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
997 bytes
+= req
->wb_bytes
;
998 nfs_list_remove_request(req
);
999 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
) &&
1000 (hdr
->good_bytes
< bytes
)) {
1001 nfs_set_pageerror(page_file_mapping(req
->wb_page
));
1002 nfs_context_set_write_error(req
->wb_context
, hdr
->error
);
1005 if (nfs_write_need_commit(hdr
)) {
1006 memcpy(&req
->wb_verf
, &hdr
->verf
.verifier
, sizeof(req
->wb_verf
));
1007 nfs_mark_request_commit(req
, hdr
->lseg
, &cinfo
,
1008 hdr
->pgio_mirror_idx
);
1012 nfs_inode_remove_request(req
);
1014 nfs_end_page_writeback(req
);
1015 nfs_release_request(req
);
1018 nfs_io_completion_put(hdr
->io_completion
);
1023 nfs_reqs_to_commit(struct nfs_commit_info
*cinfo
)
1025 return atomic_long_read(&cinfo
->mds
->ncommit
);
1028 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1030 nfs_scan_commit_list(struct list_head
*src
, struct list_head
*dst
,
1031 struct nfs_commit_info
*cinfo
, int max
)
1033 struct nfs_page
*req
, *tmp
;
1037 list_for_each_entry_safe(req
, tmp
, src
, wb_list
) {
1038 kref_get(&req
->wb_kref
);
1039 if (!nfs_lock_request(req
)) {
1042 /* Prevent deadlock with nfs_lock_and_join_requests */
1043 if (!list_empty(dst
)) {
1044 nfs_release_request(req
);
1047 /* Ensure we make progress to prevent livelock */
1048 mutex_unlock(&NFS_I(cinfo
->inode
)->commit_mutex
);
1049 status
= nfs_wait_on_request(req
);
1050 nfs_release_request(req
);
1051 mutex_lock(&NFS_I(cinfo
->inode
)->commit_mutex
);
1056 nfs_request_remove_commit_list(req
, cinfo
);
1057 clear_bit(PG_COMMIT_TO_DS
, &req
->wb_flags
);
1058 nfs_list_add_request(req
, dst
);
1060 if ((ret
== max
) && !cinfo
->dreq
)
1066 EXPORT_SYMBOL_GPL(nfs_scan_commit_list
);
1069 * nfs_scan_commit - Scan an inode for commit requests
1070 * @inode: NFS inode to scan
1071 * @dst: mds destination list
1072 * @cinfo: mds and ds lists of reqs ready to commit
1074 * Moves requests from the inode's 'commit' request list.
1075 * The requests are *not* checked to ensure that they form a contiguous set.
1078 nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
,
1079 struct nfs_commit_info
*cinfo
)
1083 if (!atomic_long_read(&cinfo
->mds
->ncommit
))
1085 mutex_lock(&NFS_I(cinfo
->inode
)->commit_mutex
);
1086 if (atomic_long_read(&cinfo
->mds
->ncommit
) > 0) {
1087 const int max
= INT_MAX
;
1089 ret
= nfs_scan_commit_list(&cinfo
->mds
->list
, dst
,
1091 ret
+= pnfs_scan_commit_lists(inode
, cinfo
, max
- ret
);
1093 mutex_unlock(&NFS_I(cinfo
->inode
)->commit_mutex
);
1098 * Search for an existing write request, and attempt to update
1099 * it to reflect a new dirty region on a given page.
1101 * If the attempt fails, then the existing request is flushed out
1104 static struct nfs_page
*nfs_try_to_update_request(struct inode
*inode
,
1106 unsigned int offset
,
1109 struct nfs_page
*req
;
1114 end
= offset
+ bytes
;
1116 req
= nfs_lock_and_join_requests(page
);
1117 if (IS_ERR_OR_NULL(req
))
1120 rqend
= req
->wb_offset
+ req
->wb_bytes
;
1122 * Tell the caller to flush out the request if
1123 * the offsets are non-contiguous.
1124 * Note: nfs_flush_incompatible() will already
1125 * have flushed out requests having wrong owners.
1127 if (offset
> rqend
|| end
< req
->wb_offset
)
1130 /* Okay, the request matches. Update the region */
1131 if (offset
< req
->wb_offset
) {
1132 req
->wb_offset
= offset
;
1133 req
->wb_pgbase
= offset
;
1136 req
->wb_bytes
= end
- req
->wb_offset
;
1138 req
->wb_bytes
= rqend
- req
->wb_offset
;
1142 * Note: we mark the request dirty here because
1143 * nfs_lock_and_join_requests() cannot preserve
1144 * commit flags, so we have to replay the write.
1146 nfs_mark_request_dirty(req
);
1147 nfs_unlock_and_release_request(req
);
1148 error
= nfs_wb_page(inode
, page
);
1149 return (error
< 0) ? ERR_PTR(error
) : NULL
;
1153 * Try to update an existing write request, or create one if there is none.
1155 * Note: Should always be called with the Page Lock held to prevent races
1156 * if we have to add a new request. Also assumes that the caller has
1157 * already called nfs_flush_incompatible() if necessary.
1159 static struct nfs_page
* nfs_setup_write_request(struct nfs_open_context
* ctx
,
1160 struct page
*page
, unsigned int offset
, unsigned int bytes
)
1162 struct inode
*inode
= page_file_mapping(page
)->host
;
1163 struct nfs_page
*req
;
1165 req
= nfs_try_to_update_request(inode
, page
, offset
, bytes
);
1168 req
= nfs_create_request(ctx
, page
, NULL
, offset
, bytes
);
1171 nfs_inode_add_request(inode
, req
);
1176 static int nfs_writepage_setup(struct nfs_open_context
*ctx
, struct page
*page
,
1177 unsigned int offset
, unsigned int count
)
1179 struct nfs_page
*req
;
1181 req
= nfs_setup_write_request(ctx
, page
, offset
, count
);
1183 return PTR_ERR(req
);
1184 /* Update file length */
1185 nfs_grow_file(page
, offset
, count
);
1186 nfs_mark_uptodate(req
);
1187 nfs_mark_request_dirty(req
);
1188 nfs_unlock_and_release_request(req
);
1192 int nfs_flush_incompatible(struct file
*file
, struct page
*page
)
1194 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
1195 struct nfs_lock_context
*l_ctx
;
1196 struct file_lock_context
*flctx
= file_inode(file
)->i_flctx
;
1197 struct nfs_page
*req
;
1198 int do_flush
, status
;
1200 * Look for a request corresponding to this page. If there
1201 * is one, and it belongs to another file, we flush it out
1202 * before we try to copy anything into the page. Do this
1203 * due to the lack of an ACCESS-type call in NFSv2.
1204 * Also do the same if we find a request from an existing
1208 req
= nfs_page_find_head_request(page
);
1211 l_ctx
= req
->wb_lock_context
;
1212 do_flush
= req
->wb_page
!= page
||
1213 !nfs_match_open_context(req
->wb_context
, ctx
);
1214 if (l_ctx
&& flctx
&&
1215 !(list_empty_careful(&flctx
->flc_posix
) &&
1216 list_empty_careful(&flctx
->flc_flock
))) {
1217 do_flush
|= l_ctx
->lockowner
!= current
->files
;
1219 nfs_release_request(req
);
1222 status
= nfs_wb_page(page_file_mapping(page
)->host
, page
);
1223 } while (status
== 0);
1228 * Avoid buffered writes when a open context credential's key would
1231 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1233 * Return 0 and set a credential flag which triggers the inode to flush
1234 * and performs NFS_FILE_SYNC writes if the key will expired within
1235 * RPC_KEY_EXPIRE_TIMEO.
1238 nfs_key_timeout_notify(struct file
*filp
, struct inode
*inode
)
1240 struct nfs_open_context
*ctx
= nfs_file_open_context(filp
);
1242 if (nfs_ctx_key_to_expire(ctx
, inode
) &&
1244 /* Already expired! */
1250 * Test if the open context credential key is marked to expire soon.
1252 bool nfs_ctx_key_to_expire(struct nfs_open_context
*ctx
, struct inode
*inode
)
1254 struct rpc_auth
*auth
= NFS_SERVER(inode
)->client
->cl_auth
;
1255 struct rpc_cred
*cred
= ctx
->ll_cred
;
1256 struct auth_cred acred
= {
1260 if (cred
&& !cred
->cr_ops
->crmatch(&acred
, cred
, 0)) {
1262 ctx
->ll_cred
= NULL
;
1266 cred
= auth
->au_ops
->lookup_cred(auth
, &acred
, 0);
1267 if (!cred
|| IS_ERR(cred
))
1269 ctx
->ll_cred
= cred
;
1270 return !!(cred
->cr_ops
->crkey_timeout
&&
1271 cred
->cr_ops
->crkey_timeout(cred
));
1275 * If the page cache is marked as unsafe or invalid, then we can't rely on
1276 * the PageUptodate() flag. In this case, we will need to turn off
1277 * write optimisations that depend on the page contents being correct.
1279 static bool nfs_write_pageuptodate(struct page
*page
, struct inode
*inode
)
1281 struct nfs_inode
*nfsi
= NFS_I(inode
);
1283 if (nfs_have_delegated_attributes(inode
))
1285 if (nfsi
->cache_validity
& NFS_INO_REVAL_PAGECACHE
)
1288 if (test_bit(NFS_INO_INVALIDATING
, &nfsi
->flags
))
1291 if (nfsi
->cache_validity
& NFS_INO_INVALID_DATA
)
1293 return PageUptodate(page
) != 0;
1297 is_whole_file_wrlock(struct file_lock
*fl
)
1299 return fl
->fl_start
== 0 && fl
->fl_end
== OFFSET_MAX
&&
1300 fl
->fl_type
== F_WRLCK
;
1303 /* If we know the page is up to date, and we're not using byte range locks (or
1304 * if we have the whole file locked for writing), it may be more efficient to
1305 * extend the write to cover the entire page in order to avoid fragmentation
1308 * If the file is opened for synchronous writes then we can just skip the rest
1311 static int nfs_can_extend_write(struct file
*file
, struct page
*page
, struct inode
*inode
)
1314 struct file_lock_context
*flctx
= inode
->i_flctx
;
1315 struct file_lock
*fl
;
1317 if (file
->f_flags
& O_DSYNC
)
1319 if (!nfs_write_pageuptodate(page
, inode
))
1321 if (NFS_PROTO(inode
)->have_delegation(inode
, FMODE_WRITE
))
1323 if (!flctx
|| (list_empty_careful(&flctx
->flc_flock
) &&
1324 list_empty_careful(&flctx
->flc_posix
)))
1327 /* Check to see if there are whole file write locks */
1329 spin_lock(&flctx
->flc_lock
);
1330 if (!list_empty(&flctx
->flc_posix
)) {
1331 fl
= list_first_entry(&flctx
->flc_posix
, struct file_lock
,
1333 if (is_whole_file_wrlock(fl
))
1335 } else if (!list_empty(&flctx
->flc_flock
)) {
1336 fl
= list_first_entry(&flctx
->flc_flock
, struct file_lock
,
1338 if (fl
->fl_type
== F_WRLCK
)
1341 spin_unlock(&flctx
->flc_lock
);
1346 * Update and possibly write a cached page of an NFS file.
1348 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1349 * things with a page scheduled for an RPC call (e.g. invalidate it).
1351 int nfs_updatepage(struct file
*file
, struct page
*page
,
1352 unsigned int offset
, unsigned int count
)
1354 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
1355 struct address_space
*mapping
= page_file_mapping(page
);
1356 struct inode
*inode
= mapping
->host
;
1359 nfs_inc_stats(inode
, NFSIOS_VFSUPDATEPAGE
);
1361 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1362 file
, count
, (long long)(page_file_offset(page
) + offset
));
1367 if (nfs_can_extend_write(file
, page
, inode
)) {
1368 count
= max(count
+ offset
, nfs_page_length(page
));
1372 status
= nfs_writepage_setup(ctx
, page
, offset
, count
);
1374 nfs_set_pageerror(mapping
);
1376 __set_page_dirty_nobuffers(page
);
1378 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1379 status
, (long long)i_size_read(inode
));
1383 static int flush_task_priority(int how
)
1385 switch (how
& (FLUSH_HIGHPRI
|FLUSH_LOWPRI
)) {
1387 return RPC_PRIORITY_HIGH
;
1389 return RPC_PRIORITY_LOW
;
1391 return RPC_PRIORITY_NORMAL
;
1394 static void nfs_initiate_write(struct nfs_pgio_header
*hdr
,
1395 struct rpc_message
*msg
,
1396 const struct nfs_rpc_ops
*rpc_ops
,
1397 struct rpc_task_setup
*task_setup_data
, int how
)
1399 int priority
= flush_task_priority(how
);
1401 task_setup_data
->priority
= priority
;
1402 rpc_ops
->write_setup(hdr
, msg
, &task_setup_data
->rpc_client
);
1403 trace_nfs_initiate_write(hdr
->inode
, hdr
->io_start
, hdr
->good_bytes
,
1407 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1408 * call this on each, which will prepare them to be retried on next
1409 * writeback using standard nfs.
1411 static void nfs_redirty_request(struct nfs_page
*req
)
1413 nfs_mark_request_dirty(req
);
1414 set_bit(NFS_CONTEXT_RESEND_WRITES
, &req
->wb_context
->flags
);
1415 nfs_end_page_writeback(req
);
1416 nfs_release_request(req
);
1419 static void nfs_async_write_error(struct list_head
*head
, int error
)
1421 struct nfs_page
*req
;
1423 while (!list_empty(head
)) {
1424 req
= nfs_list_entry(head
->next
);
1425 nfs_list_remove_request(req
);
1426 if (nfs_error_is_fatal(error
)) {
1427 nfs_context_set_write_error(req
->wb_context
, error
);
1428 if (nfs_error_is_fatal_on_server(error
)) {
1429 nfs_write_error_remove_page(req
);
1433 nfs_redirty_request(req
);
1437 static void nfs_async_write_reschedule_io(struct nfs_pgio_header
*hdr
)
1439 nfs_async_write_error(&hdr
->pages
, 0);
1440 filemap_fdatawrite_range(hdr
->inode
->i_mapping
, hdr
->args
.offset
,
1441 hdr
->args
.offset
+ hdr
->args
.count
- 1);
1444 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops
= {
1445 .init_hdr
= nfs_async_write_init
,
1446 .error_cleanup
= nfs_async_write_error
,
1447 .completion
= nfs_write_completion
,
1448 .reschedule_io
= nfs_async_write_reschedule_io
,
1451 void nfs_pageio_init_write(struct nfs_pageio_descriptor
*pgio
,
1452 struct inode
*inode
, int ioflags
, bool force_mds
,
1453 const struct nfs_pgio_completion_ops
*compl_ops
)
1455 struct nfs_server
*server
= NFS_SERVER(inode
);
1456 const struct nfs_pageio_ops
*pg_ops
= &nfs_pgio_rw_ops
;
1458 #ifdef CONFIG_NFS_V4_1
1459 if (server
->pnfs_curr_ld
&& !force_mds
)
1460 pg_ops
= server
->pnfs_curr_ld
->pg_write_ops
;
1462 nfs_pageio_init(pgio
, inode
, pg_ops
, compl_ops
, &nfs_rw_write_ops
,
1463 server
->wsize
, ioflags
);
1465 EXPORT_SYMBOL_GPL(nfs_pageio_init_write
);
1467 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor
*pgio
)
1469 struct nfs_pgio_mirror
*mirror
;
1471 if (pgio
->pg_ops
&& pgio
->pg_ops
->pg_cleanup
)
1472 pgio
->pg_ops
->pg_cleanup(pgio
);
1474 pgio
->pg_ops
= &nfs_pgio_rw_ops
;
1476 nfs_pageio_stop_mirroring(pgio
);
1478 mirror
= &pgio
->pg_mirrors
[0];
1479 mirror
->pg_bsize
= NFS_SERVER(pgio
->pg_inode
)->wsize
;
1481 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds
);
1484 void nfs_commit_prepare(struct rpc_task
*task
, void *calldata
)
1486 struct nfs_commit_data
*data
= calldata
;
1488 NFS_PROTO(data
->inode
)->commit_rpc_prepare(task
, data
);
1492 * Special version of should_remove_suid() that ignores capabilities.
1494 static int nfs_should_remove_suid(const struct inode
*inode
)
1496 umode_t mode
= inode
->i_mode
;
1499 /* suid always must be killed */
1500 if (unlikely(mode
& S_ISUID
))
1501 kill
= ATTR_KILL_SUID
;
1504 * sgid without any exec bits is just a mandatory locking mark; leave
1505 * it alone. If some exec bits are set, it's a real sgid; kill it.
1507 if (unlikely((mode
& S_ISGID
) && (mode
& S_IXGRP
)))
1508 kill
|= ATTR_KILL_SGID
;
1510 if (unlikely(kill
&& S_ISREG(mode
)))
1516 static void nfs_writeback_check_extend(struct nfs_pgio_header
*hdr
,
1517 struct nfs_fattr
*fattr
)
1519 struct nfs_pgio_args
*argp
= &hdr
->args
;
1520 struct nfs_pgio_res
*resp
= &hdr
->res
;
1521 u64 size
= argp
->offset
+ resp
->count
;
1523 if (!(fattr
->valid
& NFS_ATTR_FATTR_SIZE
))
1525 if (nfs_size_to_loff_t(fattr
->size
) < i_size_read(hdr
->inode
)) {
1526 fattr
->valid
&= ~NFS_ATTR_FATTR_SIZE
;
1529 if (size
!= fattr
->size
)
1531 /* Set attribute barrier */
1532 nfs_fattr_set_barrier(fattr
);
1533 /* ...and update size */
1534 fattr
->valid
|= NFS_ATTR_FATTR_SIZE
;
1537 void nfs_writeback_update_inode(struct nfs_pgio_header
*hdr
)
1539 struct nfs_fattr
*fattr
= &hdr
->fattr
;
1540 struct inode
*inode
= hdr
->inode
;
1542 spin_lock(&inode
->i_lock
);
1543 nfs_writeback_check_extend(hdr
, fattr
);
1544 nfs_post_op_update_inode_force_wcc_locked(inode
, fattr
);
1545 spin_unlock(&inode
->i_lock
);
1547 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode
);
1550 * This function is called when the WRITE call is complete.
1552 static int nfs_writeback_done(struct rpc_task
*task
,
1553 struct nfs_pgio_header
*hdr
,
1554 struct inode
*inode
)
1559 * ->write_done will attempt to use post-op attributes to detect
1560 * conflicting writes by other clients. A strict interpretation
1561 * of close-to-open would allow us to continue caching even if
1562 * another writer had changed the file, but some applications
1563 * depend on tighter cache coherency when writing.
1565 status
= NFS_PROTO(inode
)->write_done(task
, hdr
);
1569 nfs_add_stats(inode
, NFSIOS_SERVERWRITTENBYTES
, hdr
->res
.count
);
1570 trace_nfs_writeback_done(inode
, task
->tk_status
,
1571 hdr
->args
.offset
, hdr
->res
.verf
);
1573 if (hdr
->res
.verf
->committed
< hdr
->args
.stable
&&
1574 task
->tk_status
>= 0) {
1575 /* We tried a write call, but the server did not
1576 * commit data to stable storage even though we
1578 * Note: There is a known bug in Tru64 < 5.0 in which
1579 * the server reports NFS_DATA_SYNC, but performs
1580 * NFS_FILE_SYNC. We therefore implement this checking
1581 * as a dprintk() in order to avoid filling syslog.
1583 static unsigned long complain
;
1585 /* Note this will print the MDS for a DS write */
1586 if (time_before(complain
, jiffies
)) {
1587 dprintk("NFS: faulty NFS server %s:"
1588 " (committed = %d) != (stable = %d)\n",
1589 NFS_SERVER(inode
)->nfs_client
->cl_hostname
,
1590 hdr
->res
.verf
->committed
, hdr
->args
.stable
);
1591 complain
= jiffies
+ 300 * HZ
;
1595 /* Deal with the suid/sgid bit corner case */
1596 if (nfs_should_remove_suid(inode
)) {
1597 spin_lock(&inode
->i_lock
);
1598 NFS_I(inode
)->cache_validity
|= NFS_INO_INVALID_OTHER
;
1599 spin_unlock(&inode
->i_lock
);
1605 * This function is called when the WRITE call is complete.
1607 static void nfs_writeback_result(struct rpc_task
*task
,
1608 struct nfs_pgio_header
*hdr
)
1610 struct nfs_pgio_args
*argp
= &hdr
->args
;
1611 struct nfs_pgio_res
*resp
= &hdr
->res
;
1613 if (resp
->count
< argp
->count
) {
1614 static unsigned long complain
;
1616 /* This a short write! */
1617 nfs_inc_stats(hdr
->inode
, NFSIOS_SHORTWRITE
);
1619 /* Has the server at least made some progress? */
1620 if (resp
->count
== 0) {
1621 if (time_before(complain
, jiffies
)) {
1623 "NFS: Server wrote zero bytes, expected %u.\n",
1625 complain
= jiffies
+ 300 * HZ
;
1627 nfs_set_pgio_error(hdr
, -EIO
, argp
->offset
);
1628 task
->tk_status
= -EIO
;
1632 /* For non rpc-based layout drivers, retry-through-MDS */
1633 if (!task
->tk_ops
) {
1634 hdr
->pnfs_error
= -EAGAIN
;
1638 /* Was this an NFSv2 write or an NFSv3 stable write? */
1639 if (resp
->verf
->committed
!= NFS_UNSTABLE
) {
1640 /* Resend from where the server left off */
1641 hdr
->mds_offset
+= resp
->count
;
1642 argp
->offset
+= resp
->count
;
1643 argp
->pgbase
+= resp
->count
;
1644 argp
->count
-= resp
->count
;
1646 /* Resend as a stable write in order to avoid
1647 * headaches in the case of a server crash.
1649 argp
->stable
= NFS_FILE_SYNC
;
1651 rpc_restart_call_prepare(task
);
1655 static int wait_on_commit(struct nfs_mds_commit_info
*cinfo
)
1657 return wait_var_event_killable(&cinfo
->rpcs_out
,
1658 !atomic_read(&cinfo
->rpcs_out
));
1661 static void nfs_commit_begin(struct nfs_mds_commit_info
*cinfo
)
1663 atomic_inc(&cinfo
->rpcs_out
);
1666 static void nfs_commit_end(struct nfs_mds_commit_info
*cinfo
)
1668 if (atomic_dec_and_test(&cinfo
->rpcs_out
))
1669 wake_up_var(&cinfo
->rpcs_out
);
1672 void nfs_commitdata_release(struct nfs_commit_data
*data
)
1674 put_nfs_open_context(data
->context
);
1675 nfs_commit_free(data
);
1677 EXPORT_SYMBOL_GPL(nfs_commitdata_release
);
1679 int nfs_initiate_commit(struct rpc_clnt
*clnt
, struct nfs_commit_data
*data
,
1680 const struct nfs_rpc_ops
*nfs_ops
,
1681 const struct rpc_call_ops
*call_ops
,
1684 struct rpc_task
*task
;
1685 int priority
= flush_task_priority(how
);
1686 struct rpc_message msg
= {
1687 .rpc_argp
= &data
->args
,
1688 .rpc_resp
= &data
->res
,
1689 .rpc_cred
= data
->cred
,
1691 struct rpc_task_setup task_setup_data
= {
1692 .task
= &data
->task
,
1694 .rpc_message
= &msg
,
1695 .callback_ops
= call_ops
,
1696 .callback_data
= data
,
1697 .workqueue
= nfsiod_workqueue
,
1698 .flags
= RPC_TASK_ASYNC
| flags
,
1699 .priority
= priority
,
1701 /* Set up the initial task struct. */
1702 nfs_ops
->commit_setup(data
, &msg
, &task_setup_data
.rpc_client
);
1703 trace_nfs_initiate_commit(data
);
1705 dprintk("NFS: initiated commit call\n");
1707 task
= rpc_run_task(&task_setup_data
);
1709 return PTR_ERR(task
);
1710 if (how
& FLUSH_SYNC
)
1711 rpc_wait_for_completion_task(task
);
1715 EXPORT_SYMBOL_GPL(nfs_initiate_commit
);
1717 static loff_t
nfs_get_lwb(struct list_head
*head
)
1720 struct nfs_page
*req
;
1722 list_for_each_entry(req
, head
, wb_list
)
1723 if (lwb
< (req_offset(req
) + req
->wb_bytes
))
1724 lwb
= req_offset(req
) + req
->wb_bytes
;
1730 * Set up the argument/result storage required for the RPC call.
1732 void nfs_init_commit(struct nfs_commit_data
*data
,
1733 struct list_head
*head
,
1734 struct pnfs_layout_segment
*lseg
,
1735 struct nfs_commit_info
*cinfo
)
1737 struct nfs_page
*first
= nfs_list_entry(head
->next
);
1738 struct inode
*inode
= d_inode(first
->wb_context
->dentry
);
1740 /* Set up the RPC argument and reply structs
1741 * NB: take care not to mess about with data->commit et al. */
1743 list_splice_init(head
, &data
->pages
);
1745 data
->inode
= inode
;
1746 data
->cred
= first
->wb_context
->cred
;
1747 data
->lseg
= lseg
; /* reference transferred */
1748 /* only set lwb for pnfs commit */
1750 data
->lwb
= nfs_get_lwb(&data
->pages
);
1751 data
->mds_ops
= &nfs_commit_ops
;
1752 data
->completion_ops
= cinfo
->completion_ops
;
1753 data
->dreq
= cinfo
->dreq
;
1755 data
->args
.fh
= NFS_FH(data
->inode
);
1756 /* Note: we always request a commit of the entire inode */
1757 data
->args
.offset
= 0;
1758 data
->args
.count
= 0;
1759 data
->context
= get_nfs_open_context(first
->wb_context
);
1760 data
->res
.fattr
= &data
->fattr
;
1761 data
->res
.verf
= &data
->verf
;
1762 nfs_fattr_init(&data
->fattr
);
1764 EXPORT_SYMBOL_GPL(nfs_init_commit
);
1766 void nfs_retry_commit(struct list_head
*page_list
,
1767 struct pnfs_layout_segment
*lseg
,
1768 struct nfs_commit_info
*cinfo
,
1771 struct nfs_page
*req
;
1773 while (!list_empty(page_list
)) {
1774 req
= nfs_list_entry(page_list
->next
);
1775 nfs_list_remove_request(req
);
1776 nfs_mark_request_commit(req
, lseg
, cinfo
, ds_commit_idx
);
1778 nfs_clear_page_commit(req
->wb_page
);
1779 nfs_unlock_and_release_request(req
);
1782 EXPORT_SYMBOL_GPL(nfs_retry_commit
);
1785 nfs_commit_resched_write(struct nfs_commit_info
*cinfo
,
1786 struct nfs_page
*req
)
1788 __set_page_dirty_nobuffers(req
->wb_page
);
1792 * Commit dirty pages
1795 nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
,
1796 struct nfs_commit_info
*cinfo
)
1798 struct nfs_commit_data
*data
;
1800 /* another commit raced with us */
1801 if (list_empty(head
))
1804 data
= nfs_commitdata_alloc(true);
1806 /* Set up the argument struct */
1807 nfs_init_commit(data
, head
, NULL
, cinfo
);
1808 atomic_inc(&cinfo
->mds
->rpcs_out
);
1809 return nfs_initiate_commit(NFS_CLIENT(inode
), data
, NFS_PROTO(inode
),
1810 data
->mds_ops
, how
, 0);
1814 * COMMIT call returned
1816 static void nfs_commit_done(struct rpc_task
*task
, void *calldata
)
1818 struct nfs_commit_data
*data
= calldata
;
1820 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1821 task
->tk_pid
, task
->tk_status
);
1823 /* Call the NFS version-specific code */
1824 NFS_PROTO(data
->inode
)->commit_done(task
, data
);
1825 trace_nfs_commit_done(data
);
1828 static void nfs_commit_release_pages(struct nfs_commit_data
*data
)
1830 struct nfs_page
*req
;
1831 int status
= data
->task
.tk_status
;
1832 struct nfs_commit_info cinfo
;
1833 struct nfs_server
*nfss
;
1835 while (!list_empty(&data
->pages
)) {
1836 req
= nfs_list_entry(data
->pages
.next
);
1837 nfs_list_remove_request(req
);
1839 nfs_clear_page_commit(req
->wb_page
);
1841 dprintk("NFS: commit (%s/%llu %d@%lld)",
1842 req
->wb_context
->dentry
->d_sb
->s_id
,
1843 (unsigned long long)NFS_FILEID(d_inode(req
->wb_context
->dentry
)),
1845 (long long)req_offset(req
));
1847 nfs_context_set_write_error(req
->wb_context
, status
);
1849 nfs_inode_remove_request(req
);
1850 dprintk_cont(", error = %d\n", status
);
1854 /* Okay, COMMIT succeeded, apparently. Check the verifier
1855 * returned by the server against all stored verfs. */
1856 if (!nfs_write_verifier_cmp(&req
->wb_verf
, &data
->verf
.verifier
)) {
1857 /* We have a match */
1859 nfs_inode_remove_request(req
);
1860 dprintk_cont(" OK\n");
1863 /* We have a mismatch. Write the page again */
1864 dprintk_cont(" mismatch\n");
1865 nfs_mark_request_dirty(req
);
1866 set_bit(NFS_CONTEXT_RESEND_WRITES
, &req
->wb_context
->flags
);
1868 nfs_unlock_and_release_request(req
);
1869 /* Latency breaker */
1872 nfss
= NFS_SERVER(data
->inode
);
1873 if (atomic_long_read(&nfss
->writeback
) < NFS_CONGESTION_OFF_THRESH
)
1874 clear_bdi_congested(inode_to_bdi(data
->inode
), BLK_RW_ASYNC
);
1876 nfs_init_cinfo(&cinfo
, data
->inode
, data
->dreq
);
1877 nfs_commit_end(cinfo
.mds
);
1880 static void nfs_commit_release(void *calldata
)
1882 struct nfs_commit_data
*data
= calldata
;
1884 data
->completion_ops
->completion(data
);
1885 nfs_commitdata_release(calldata
);
1888 static const struct rpc_call_ops nfs_commit_ops
= {
1889 .rpc_call_prepare
= nfs_commit_prepare
,
1890 .rpc_call_done
= nfs_commit_done
,
1891 .rpc_release
= nfs_commit_release
,
1894 static const struct nfs_commit_completion_ops nfs_commit_completion_ops
= {
1895 .completion
= nfs_commit_release_pages
,
1896 .resched_write
= nfs_commit_resched_write
,
1899 int nfs_generic_commit_list(struct inode
*inode
, struct list_head
*head
,
1900 int how
, struct nfs_commit_info
*cinfo
)
1904 status
= pnfs_commit_list(inode
, head
, how
, cinfo
);
1905 if (status
== PNFS_NOT_ATTEMPTED
)
1906 status
= nfs_commit_list(inode
, head
, how
, cinfo
);
1910 static int __nfs_commit_inode(struct inode
*inode
, int how
,
1911 struct writeback_control
*wbc
)
1914 struct nfs_commit_info cinfo
;
1915 int may_wait
= how
& FLUSH_SYNC
;
1918 nfs_init_cinfo_from_inode(&cinfo
, inode
);
1919 nfs_commit_begin(cinfo
.mds
);
1921 ret
= nscan
= nfs_scan_commit(inode
, &head
, &cinfo
);
1924 ret
= nfs_generic_commit_list(inode
, &head
, how
, &cinfo
);
1928 if (wbc
&& wbc
->sync_mode
== WB_SYNC_NONE
) {
1929 if (nscan
< wbc
->nr_to_write
)
1930 wbc
->nr_to_write
-= nscan
;
1932 wbc
->nr_to_write
= 0;
1934 if (nscan
< INT_MAX
)
1938 nfs_commit_end(cinfo
.mds
);
1939 if (ret
|| !may_wait
)
1941 return wait_on_commit(cinfo
.mds
);
1944 int nfs_commit_inode(struct inode
*inode
, int how
)
1946 return __nfs_commit_inode(inode
, how
, NULL
);
1948 EXPORT_SYMBOL_GPL(nfs_commit_inode
);
1950 int nfs_write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1952 struct nfs_inode
*nfsi
= NFS_I(inode
);
1953 int flags
= FLUSH_SYNC
;
1956 if (wbc
->sync_mode
== WB_SYNC_NONE
) {
1957 /* no commits means nothing needs to be done */
1958 if (!atomic_long_read(&nfsi
->commit_info
.ncommit
))
1959 goto check_requests_outstanding
;
1961 /* Don't commit yet if this is a non-blocking flush and there
1962 * are a lot of outstanding writes for this mapping.
1964 if (mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_WRITEBACK
))
1965 goto out_mark_dirty
;
1967 /* don't wait for the COMMIT response */
1971 ret
= __nfs_commit_inode(inode
, flags
, wbc
);
1973 if (flags
& FLUSH_SYNC
)
1975 } else if (atomic_long_read(&nfsi
->commit_info
.ncommit
))
1976 goto out_mark_dirty
;
1978 check_requests_outstanding
:
1979 if (!atomic_read(&nfsi
->commit_info
.rpcs_out
))
1982 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
1985 EXPORT_SYMBOL_GPL(nfs_write_inode
);
1988 * Wrapper for filemap_write_and_wait_range()
1990 * Needed for pNFS in order to ensure data becomes visible to the
1993 int nfs_filemap_write_and_wait_range(struct address_space
*mapping
,
1994 loff_t lstart
, loff_t lend
)
1998 ret
= filemap_write_and_wait_range(mapping
, lstart
, lend
);
2000 ret
= pnfs_sync_inode(mapping
->host
, true);
2003 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range
);
2006 * flush the inode to disk.
2008 int nfs_wb_all(struct inode
*inode
)
2012 trace_nfs_writeback_inode_enter(inode
);
2014 ret
= filemap_write_and_wait(inode
->i_mapping
);
2017 ret
= nfs_commit_inode(inode
, FLUSH_SYNC
);
2020 pnfs_sync_inode(inode
, true);
2024 trace_nfs_writeback_inode_exit(inode
, ret
);
2027 EXPORT_SYMBOL_GPL(nfs_wb_all
);
2029 int nfs_wb_page_cancel(struct inode
*inode
, struct page
*page
)
2031 struct nfs_page
*req
;
2034 wait_on_page_writeback(page
);
2036 /* blocking call to cancel all requests and join to a single (head)
2038 req
= nfs_lock_and_join_requests(page
);
2043 /* all requests from this page have been cancelled by
2044 * nfs_lock_and_join_requests, so just remove the head
2045 * request from the inode / page_private pointer and
2047 nfs_inode_remove_request(req
);
2048 nfs_unlock_and_release_request(req
);
2055 * Write back all requests on one page - we do this before reading it.
2057 int nfs_wb_page(struct inode
*inode
, struct page
*page
)
2059 loff_t range_start
= page_file_offset(page
);
2060 loff_t range_end
= range_start
+ (loff_t
)(PAGE_SIZE
- 1);
2061 struct writeback_control wbc
= {
2062 .sync_mode
= WB_SYNC_ALL
,
2064 .range_start
= range_start
,
2065 .range_end
= range_end
,
2069 trace_nfs_writeback_page_enter(inode
);
2072 wait_on_page_writeback(page
);
2073 if (clear_page_dirty_for_io(page
)) {
2074 ret
= nfs_writepage_locked(page
, &wbc
);
2080 if (!PagePrivate(page
))
2082 ret
= nfs_commit_inode(inode
, FLUSH_SYNC
);
2087 trace_nfs_writeback_page_exit(inode
, ret
);
2091 #ifdef CONFIG_MIGRATION
2092 int nfs_migrate_page(struct address_space
*mapping
, struct page
*newpage
,
2093 struct page
*page
, enum migrate_mode mode
)
2096 * If PagePrivate is set, then the page is currently associated with
2097 * an in-progress read or write request. Don't try to migrate it.
2099 * FIXME: we could do this in principle, but we'll need a way to ensure
2100 * that we can safely release the inode reference while holding
2103 if (PagePrivate(page
))
2106 if (!nfs_fscache_release_page(page
, GFP_KERNEL
))
2109 return migrate_page(mapping
, newpage
, page
, mode
);
2113 int __init
nfs_init_writepagecache(void)
2115 nfs_wdata_cachep
= kmem_cache_create("nfs_write_data",
2116 sizeof(struct nfs_pgio_header
),
2117 0, SLAB_HWCACHE_ALIGN
,
2119 if (nfs_wdata_cachep
== NULL
)
2122 nfs_wdata_mempool
= mempool_create_slab_pool(MIN_POOL_WRITE
,
2124 if (nfs_wdata_mempool
== NULL
)
2125 goto out_destroy_write_cache
;
2127 nfs_cdata_cachep
= kmem_cache_create("nfs_commit_data",
2128 sizeof(struct nfs_commit_data
),
2129 0, SLAB_HWCACHE_ALIGN
,
2131 if (nfs_cdata_cachep
== NULL
)
2132 goto out_destroy_write_mempool
;
2134 nfs_commit_mempool
= mempool_create_slab_pool(MIN_POOL_COMMIT
,
2136 if (nfs_commit_mempool
== NULL
)
2137 goto out_destroy_commit_cache
;
2140 * NFS congestion size, scale with available memory.
2152 * This allows larger machines to have larger/more transfers.
2153 * Limit the default to 256M
2155 nfs_congestion_kb
= (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT
-10);
2156 if (nfs_congestion_kb
> 256*1024)
2157 nfs_congestion_kb
= 256*1024;
2161 out_destroy_commit_cache
:
2162 kmem_cache_destroy(nfs_cdata_cachep
);
2163 out_destroy_write_mempool
:
2164 mempool_destroy(nfs_wdata_mempool
);
2165 out_destroy_write_cache
:
2166 kmem_cache_destroy(nfs_wdata_cachep
);
2170 void nfs_destroy_writepagecache(void)
2172 mempool_destroy(nfs_commit_mempool
);
2173 kmem_cache_destroy(nfs_cdata_cachep
);
2174 mempool_destroy(nfs_wdata_mempool
);
2175 kmem_cache_destroy(nfs_wdata_cachep
);
2178 static const struct nfs_rw_ops nfs_rw_write_ops
= {
2179 .rw_alloc_header
= nfs_writehdr_alloc
,
2180 .rw_free_header
= nfs_writehdr_free
,
2181 .rw_done
= nfs_writeback_done
,
2182 .rw_result
= nfs_writeback_result
,
2183 .rw_initiate
= nfs_initiate_write
,