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>
30 #include "delegation.h"
39 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
41 #define MIN_POOL_WRITE (32)
42 #define MIN_POOL_COMMIT (4)
44 struct nfs_io_completion
{
45 void (*complete
)(void *data
);
51 * Local function declarations
53 static void nfs_redirty_request(struct nfs_page
*req
);
54 static const struct rpc_call_ops nfs_commit_ops
;
55 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops
;
56 static const struct nfs_commit_completion_ops nfs_commit_completion_ops
;
57 static const struct nfs_rw_ops nfs_rw_write_ops
;
58 static void nfs_clear_request_commit(struct nfs_page
*req
);
59 static void nfs_init_cinfo_from_inode(struct nfs_commit_info
*cinfo
,
61 static struct nfs_page
*
62 nfs_page_search_commits_for_head_request_locked(struct nfs_inode
*nfsi
,
65 static struct kmem_cache
*nfs_wdata_cachep
;
66 static mempool_t
*nfs_wdata_mempool
;
67 static struct kmem_cache
*nfs_cdata_cachep
;
68 static mempool_t
*nfs_commit_mempool
;
70 struct nfs_commit_data
*nfs_commitdata_alloc(bool never_fail
)
72 struct nfs_commit_data
*p
;
75 p
= mempool_alloc(nfs_commit_mempool
, GFP_NOIO
);
77 /* It is OK to do some reclaim, not no safe to wait
78 * for anything to be returned to the pool.
79 * mempool_alloc() cannot handle that particular combination,
80 * so we need two separate attempts.
82 p
= mempool_alloc(nfs_commit_mempool
, GFP_NOWAIT
);
84 p
= kmem_cache_alloc(nfs_cdata_cachep
, GFP_NOIO
|
85 __GFP_NOWARN
| __GFP_NORETRY
);
90 memset(p
, 0, sizeof(*p
));
91 INIT_LIST_HEAD(&p
->pages
);
94 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc
);
96 void nfs_commit_free(struct nfs_commit_data
*p
)
98 mempool_free(p
, nfs_commit_mempool
);
100 EXPORT_SYMBOL_GPL(nfs_commit_free
);
102 static struct nfs_pgio_header
*nfs_writehdr_alloc(void)
104 struct nfs_pgio_header
*p
= mempool_alloc(nfs_wdata_mempool
, GFP_NOIO
);
106 memset(p
, 0, sizeof(*p
));
107 p
->rw_mode
= FMODE_WRITE
;
111 static void nfs_writehdr_free(struct nfs_pgio_header
*hdr
)
113 mempool_free(hdr
, nfs_wdata_mempool
);
116 static struct nfs_io_completion
*nfs_io_completion_alloc(gfp_t gfp_flags
)
118 return kmalloc(sizeof(struct nfs_io_completion
), gfp_flags
);
121 static void nfs_io_completion_init(struct nfs_io_completion
*ioc
,
122 void (*complete
)(void *), void *data
)
124 ioc
->complete
= complete
;
126 kref_init(&ioc
->refcount
);
129 static void nfs_io_completion_release(struct kref
*kref
)
131 struct nfs_io_completion
*ioc
= container_of(kref
,
132 struct nfs_io_completion
, refcount
);
133 ioc
->complete(ioc
->data
);
137 static void nfs_io_completion_get(struct nfs_io_completion
*ioc
)
140 kref_get(&ioc
->refcount
);
143 static void nfs_io_completion_put(struct nfs_io_completion
*ioc
)
146 kref_put(&ioc
->refcount
, nfs_io_completion_release
);
149 static struct nfs_page
*
150 nfs_page_private_request(struct page
*page
)
152 if (!PagePrivate(page
))
154 return (struct nfs_page
*)page_private(page
);
158 * nfs_page_find_head_request_locked - find head request associated with @page
160 * must be called while holding the inode lock.
162 * returns matching head request with reference held, or NULL if not found.
164 static struct nfs_page
*
165 nfs_page_find_private_request(struct page
*page
)
167 struct address_space
*mapping
= page_file_mapping(page
);
168 struct nfs_page
*req
;
170 if (!PagePrivate(page
))
172 spin_lock(&mapping
->private_lock
);
173 req
= nfs_page_private_request(page
);
175 WARN_ON_ONCE(req
->wb_head
!= req
);
176 kref_get(&req
->wb_kref
);
178 spin_unlock(&mapping
->private_lock
);
182 static struct nfs_page
*
183 nfs_page_find_swap_request(struct page
*page
)
185 struct inode
*inode
= page_file_mapping(page
)->host
;
186 struct nfs_inode
*nfsi
= NFS_I(inode
);
187 struct nfs_page
*req
= NULL
;
188 if (!PageSwapCache(page
))
190 mutex_lock(&nfsi
->commit_mutex
);
191 if (PageSwapCache(page
)) {
192 req
= nfs_page_search_commits_for_head_request_locked(nfsi
,
195 WARN_ON_ONCE(req
->wb_head
!= req
);
196 kref_get(&req
->wb_kref
);
199 mutex_unlock(&nfsi
->commit_mutex
);
204 * nfs_page_find_head_request - find head request associated with @page
206 * returns matching head request with reference held, or NULL if not found.
208 static struct nfs_page
*nfs_page_find_head_request(struct page
*page
)
210 struct nfs_page
*req
;
212 req
= nfs_page_find_private_request(page
);
214 req
= nfs_page_find_swap_request(page
);
218 /* Adjust the file length if we're writing beyond the end */
219 static void nfs_grow_file(struct page
*page
, unsigned int offset
, unsigned int count
)
221 struct inode
*inode
= page_file_mapping(page
)->host
;
225 spin_lock(&inode
->i_lock
);
226 i_size
= i_size_read(inode
);
227 end_index
= (i_size
- 1) >> PAGE_SHIFT
;
228 if (i_size
> 0 && page_index(page
) < end_index
)
230 end
= page_file_offset(page
) + ((loff_t
)offset
+count
);
233 i_size_write(inode
, end
);
234 NFS_I(inode
)->cache_validity
&= ~NFS_INO_INVALID_SIZE
;
235 nfs_inc_stats(inode
, NFSIOS_EXTENDWRITE
);
237 spin_unlock(&inode
->i_lock
);
240 /* A writeback failed: mark the page as bad, and invalidate the page cache */
241 static void nfs_set_pageerror(struct page
*page
)
243 nfs_zap_mapping(page_file_mapping(page
)->host
, page_file_mapping(page
));
247 * nfs_page_group_search_locked
248 * @head - head request of page group
249 * @page_offset - offset into page
251 * Search page group with head @head to find a request that contains the
252 * page offset @page_offset.
254 * Returns a pointer to the first matching nfs request, or NULL if no
257 * Must be called with the page group lock held
259 static struct nfs_page
*
260 nfs_page_group_search_locked(struct nfs_page
*head
, unsigned int page_offset
)
262 struct nfs_page
*req
;
266 if (page_offset
>= req
->wb_pgbase
&&
267 page_offset
< (req
->wb_pgbase
+ req
->wb_bytes
))
270 req
= req
->wb_this_page
;
271 } while (req
!= head
);
277 * nfs_page_group_covers_page
278 * @head - head request of page group
280 * Return true if the page group with head @head covers the whole page,
281 * returns false otherwise
283 static bool nfs_page_group_covers_page(struct nfs_page
*req
)
285 struct nfs_page
*tmp
;
286 unsigned int pos
= 0;
287 unsigned int len
= nfs_page_length(req
->wb_page
);
289 nfs_page_group_lock(req
);
292 tmp
= nfs_page_group_search_locked(req
->wb_head
, pos
);
295 pos
= tmp
->wb_pgbase
+ tmp
->wb_bytes
;
298 nfs_page_group_unlock(req
);
302 /* We can set the PG_uptodate flag if we see that a write request
303 * covers the full page.
305 static void nfs_mark_uptodate(struct nfs_page
*req
)
307 if (PageUptodate(req
->wb_page
))
309 if (!nfs_page_group_covers_page(req
))
311 SetPageUptodate(req
->wb_page
);
314 static int wb_priority(struct writeback_control
*wbc
)
318 if (wbc
->sync_mode
== WB_SYNC_ALL
)
319 ret
= FLUSH_COND_STABLE
;
324 * NFS congestion control
327 int nfs_congestion_kb
;
329 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
330 #define NFS_CONGESTION_OFF_THRESH \
331 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
333 static void nfs_set_page_writeback(struct page
*page
)
335 struct inode
*inode
= page_file_mapping(page
)->host
;
336 struct nfs_server
*nfss
= NFS_SERVER(inode
);
337 int ret
= test_set_page_writeback(page
);
339 WARN_ON_ONCE(ret
!= 0);
341 if (atomic_long_inc_return(&nfss
->writeback
) >
342 NFS_CONGESTION_ON_THRESH
)
343 set_bdi_congested(inode_to_bdi(inode
), BLK_RW_ASYNC
);
346 static void nfs_end_page_writeback(struct nfs_page
*req
)
348 struct inode
*inode
= page_file_mapping(req
->wb_page
)->host
;
349 struct nfs_server
*nfss
= NFS_SERVER(inode
);
352 is_done
= nfs_page_group_sync_on_bit(req
, PG_WB_END
);
353 nfs_unlock_request(req
);
357 end_page_writeback(req
->wb_page
);
358 if (atomic_long_dec_return(&nfss
->writeback
) < NFS_CONGESTION_OFF_THRESH
)
359 clear_bdi_congested(inode_to_bdi(inode
), BLK_RW_ASYNC
);
363 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req
365 * this is a helper function for nfs_lock_and_join_requests
367 * @inode - inode associated with request page group, must be holding inode lock
368 * @head - head request of page group, must be holding head lock
369 * @req - request that couldn't lock and needs to wait on the req bit lock
371 * NOTE: this must be called holding page_group bit lock
372 * which will be released before returning.
374 * returns 0 on success, < 0 on error.
377 nfs_unroll_locks(struct inode
*inode
, struct nfs_page
*head
,
378 struct nfs_page
*req
)
380 struct nfs_page
*tmp
;
382 /* relinquish all the locks successfully grabbed this run */
383 for (tmp
= head
->wb_this_page
; tmp
!= req
; tmp
= tmp
->wb_this_page
) {
384 if (!kref_read(&tmp
->wb_kref
))
386 nfs_unlock_and_release_request(tmp
);
391 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
393 * @destroy_list - request list (using wb_this_page) terminated by @old_head
394 * @old_head - the old head of the list
396 * All subrequests must be locked and removed from all lists, so at this point
397 * they are only "active" in this function, and possibly in nfs_wait_on_request
398 * with a reference held by some other context.
401 nfs_destroy_unlinked_subrequests(struct nfs_page
*destroy_list
,
402 struct nfs_page
*old_head
,
405 while (destroy_list
) {
406 struct nfs_page
*subreq
= destroy_list
;
408 destroy_list
= (subreq
->wb_this_page
== old_head
) ?
409 NULL
: subreq
->wb_this_page
;
411 WARN_ON_ONCE(old_head
!= subreq
->wb_head
);
413 /* make sure old group is not used */
414 subreq
->wb_this_page
= subreq
;
416 clear_bit(PG_REMOVE
, &subreq
->wb_flags
);
418 /* Note: races with nfs_page_group_destroy() */
419 if (!kref_read(&subreq
->wb_kref
)) {
420 /* Check if we raced with nfs_page_group_destroy() */
421 if (test_and_clear_bit(PG_TEARDOWN
, &subreq
->wb_flags
))
422 nfs_free_request(subreq
);
426 subreq
->wb_head
= subreq
;
428 if (test_and_clear_bit(PG_INODE_REF
, &subreq
->wb_flags
)) {
429 nfs_release_request(subreq
);
430 atomic_long_dec(&NFS_I(inode
)->nrequests
);
433 /* subreq is now totally disconnected from page group or any
434 * write / commit lists. last chance to wake any waiters */
435 nfs_unlock_and_release_request(subreq
);
440 * nfs_lock_and_join_requests - join all subreqs to the head req and return
441 * a locked reference, cancelling any pending
442 * operations for this page.
444 * @page - the page used to lookup the "page group" of nfs_page structures
446 * This function joins all sub requests to the head request by first
447 * locking all requests in the group, cancelling any pending operations
448 * and finally updating the head request to cover the whole range covered by
449 * the (former) group. All subrequests are removed from any write or commit
450 * lists, unlinked from the group and destroyed.
452 * Returns a locked, referenced pointer to the head request - which after
453 * this call is guaranteed to be the only request associated with the page.
454 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
455 * error was encountered.
457 static struct nfs_page
*
458 nfs_lock_and_join_requests(struct page
*page
)
460 struct inode
*inode
= page_file_mapping(page
)->host
;
461 struct nfs_page
*head
, *subreq
;
462 struct nfs_page
*destroy_list
= NULL
;
463 unsigned int total_bytes
;
468 * A reference is taken only on the head request which acts as a
469 * reference to the whole page group - the group will not be destroyed
470 * until the head reference is released.
472 head
= nfs_page_find_head_request(page
);
476 /* lock the page head first in order to avoid an ABBA inefficiency */
477 if (!nfs_lock_request(head
)) {
478 ret
= nfs_wait_on_request(head
);
479 nfs_release_request(head
);
485 /* Ensure that nobody removed the request before we locked it */
486 if (head
!= nfs_page_private_request(page
) && !PageSwapCache(page
)) {
487 nfs_unlock_and_release_request(head
);
491 ret
= nfs_page_group_lock(head
);
493 goto release_request
;
495 /* lock each request in the page group */
496 total_bytes
= head
->wb_bytes
;
497 for (subreq
= head
->wb_this_page
; subreq
!= head
;
498 subreq
= subreq
->wb_this_page
) {
500 if (!kref_get_unless_zero(&subreq
->wb_kref
)) {
501 if (subreq
->wb_offset
== head
->wb_offset
+ total_bytes
)
502 total_bytes
+= subreq
->wb_bytes
;
506 while (!nfs_lock_request(subreq
)) {
508 * Unlock page to allow nfs_page_group_sync_on_bit()
511 nfs_page_group_unlock(head
);
512 ret
= nfs_wait_on_request(subreq
);
514 ret
= nfs_page_group_lock(head
);
516 nfs_unroll_locks(inode
, head
, subreq
);
517 nfs_release_request(subreq
);
518 goto release_request
;
522 * Subrequests are always contiguous, non overlapping
523 * and in order - but may be repeated (mirrored writes).
525 if (subreq
->wb_offset
== (head
->wb_offset
+ total_bytes
)) {
526 /* keep track of how many bytes this group covers */
527 total_bytes
+= subreq
->wb_bytes
;
528 } else if (WARN_ON_ONCE(subreq
->wb_offset
< head
->wb_offset
||
529 ((subreq
->wb_offset
+ subreq
->wb_bytes
) >
530 (head
->wb_offset
+ total_bytes
)))) {
531 nfs_page_group_unlock(head
);
532 nfs_unroll_locks(inode
, head
, subreq
);
533 nfs_unlock_and_release_request(subreq
);
535 goto release_request
;
539 /* Now that all requests are locked, make sure they aren't on any list.
540 * Commit list removal accounting is done after locks are dropped */
543 nfs_clear_request_commit(subreq
);
544 subreq
= subreq
->wb_this_page
;
545 } while (subreq
!= head
);
547 /* unlink subrequests from head, destroy them later */
548 if (head
->wb_this_page
!= head
) {
549 /* destroy list will be terminated by head */
550 destroy_list
= head
->wb_this_page
;
551 head
->wb_this_page
= head
;
553 /* change head request to cover whole range that
554 * the former page group covered */
555 head
->wb_bytes
= total_bytes
;
558 /* Postpone destruction of this request */
559 if (test_and_clear_bit(PG_REMOVE
, &head
->wb_flags
)) {
560 set_bit(PG_INODE_REF
, &head
->wb_flags
);
561 kref_get(&head
->wb_kref
);
562 atomic_long_inc(&NFS_I(inode
)->nrequests
);
565 nfs_page_group_unlock(head
);
567 nfs_destroy_unlinked_subrequests(destroy_list
, head
, inode
);
569 /* Did we lose a race with nfs_inode_remove_request()? */
570 if (!(PagePrivate(page
) || PageSwapCache(page
))) {
571 nfs_unlock_and_release_request(head
);
575 /* still holds ref on head from nfs_page_find_head_request
576 * and still has lock on head from lock loop */
580 nfs_unlock_and_release_request(head
);
584 static void nfs_write_error_remove_page(struct nfs_page
*req
)
586 nfs_end_page_writeback(req
);
587 generic_error_remove_page(page_file_mapping(req
->wb_page
),
589 nfs_release_request(req
);
593 nfs_error_is_fatal_on_server(int err
)
601 return nfs_error_is_fatal(err
);
605 * Find an associated nfs write request, and prepare to flush it out
606 * May return an error if the user signalled nfs_wait_on_request().
608 static int nfs_page_async_flush(struct nfs_pageio_descriptor
*pgio
,
611 struct nfs_page
*req
;
614 req
= nfs_lock_and_join_requests(page
);
621 nfs_set_page_writeback(page
);
622 WARN_ON_ONCE(test_bit(PG_CLEAN
, &req
->wb_flags
));
624 ret
= req
->wb_context
->error
;
625 /* If there is a fatal error that covers this write, just exit */
626 if (nfs_error_is_fatal_on_server(ret
))
630 if (!nfs_pageio_add_request(pgio
, req
)) {
631 ret
= pgio
->pg_error
;
633 * Remove the problematic req upon fatal errors on the server
635 if (nfs_error_is_fatal(ret
)) {
636 nfs_context_set_write_error(req
->wb_context
, ret
);
637 if (nfs_error_is_fatal_on_server(ret
))
641 nfs_redirty_request(req
);
643 nfs_add_stats(page_file_mapping(page
)->host
,
644 NFSIOS_WRITEPAGES
, 1);
648 nfs_write_error_remove_page(req
);
652 static int nfs_do_writepage(struct page
*page
, struct writeback_control
*wbc
,
653 struct nfs_pageio_descriptor
*pgio
)
657 nfs_pageio_cond_complete(pgio
, page_index(page
));
658 ret
= nfs_page_async_flush(pgio
, page
);
659 if (ret
== -EAGAIN
) {
660 redirty_page_for_writepage(wbc
, page
);
667 * Write an mmapped page to the server.
669 static int nfs_writepage_locked(struct page
*page
,
670 struct writeback_control
*wbc
)
672 struct nfs_pageio_descriptor pgio
;
673 struct inode
*inode
= page_file_mapping(page
)->host
;
676 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGE
);
677 nfs_pageio_init_write(&pgio
, inode
, 0,
678 false, &nfs_async_write_completion_ops
);
679 err
= nfs_do_writepage(page
, wbc
, &pgio
);
680 nfs_pageio_complete(&pgio
);
683 if (pgio
.pg_error
< 0)
684 return pgio
.pg_error
;
688 int nfs_writepage(struct page
*page
, struct writeback_control
*wbc
)
692 ret
= nfs_writepage_locked(page
, wbc
);
697 static int nfs_writepages_callback(struct page
*page
, struct writeback_control
*wbc
, void *data
)
701 ret
= nfs_do_writepage(page
, wbc
, data
);
706 static void nfs_io_completion_commit(void *inode
)
708 nfs_commit_inode(inode
, 0);
711 int nfs_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
713 struct inode
*inode
= mapping
->host
;
714 struct nfs_pageio_descriptor pgio
;
715 struct nfs_io_completion
*ioc
= nfs_io_completion_alloc(GFP_NOFS
);
718 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGES
);
721 nfs_io_completion_init(ioc
, nfs_io_completion_commit
, inode
);
723 nfs_pageio_init_write(&pgio
, inode
, wb_priority(wbc
), false,
724 &nfs_async_write_completion_ops
);
725 pgio
.pg_io_completion
= ioc
;
726 err
= write_cache_pages(mapping
, wbc
, nfs_writepages_callback
, &pgio
);
727 nfs_pageio_complete(&pgio
);
728 nfs_io_completion_put(ioc
);
741 * Insert a write request into an inode
743 static void nfs_inode_add_request(struct inode
*inode
, struct nfs_page
*req
)
745 struct address_space
*mapping
= page_file_mapping(req
->wb_page
);
746 struct nfs_inode
*nfsi
= NFS_I(inode
);
748 WARN_ON_ONCE(req
->wb_this_page
!= req
);
750 /* Lock the request! */
751 nfs_lock_request(req
);
754 * Swap-space should not get truncated. Hence no need to plug the race
755 * with invalidate/truncate.
757 spin_lock(&mapping
->private_lock
);
758 if (!nfs_have_writebacks(inode
) &&
759 NFS_PROTO(inode
)->have_delegation(inode
, FMODE_WRITE
))
760 inode_inc_iversion_raw(inode
);
761 if (likely(!PageSwapCache(req
->wb_page
))) {
762 set_bit(PG_MAPPED
, &req
->wb_flags
);
763 SetPagePrivate(req
->wb_page
);
764 set_page_private(req
->wb_page
, (unsigned long)req
);
766 spin_unlock(&mapping
->private_lock
);
767 atomic_long_inc(&nfsi
->nrequests
);
768 /* this a head request for a page group - mark it as having an
769 * extra reference so sub groups can follow suit.
770 * This flag also informs pgio layer when to bump nrequests when
771 * adding subrequests. */
772 WARN_ON(test_and_set_bit(PG_INODE_REF
, &req
->wb_flags
));
773 kref_get(&req
->wb_kref
);
777 * Remove a write request from an inode
779 static void nfs_inode_remove_request(struct nfs_page
*req
)
781 struct address_space
*mapping
= page_file_mapping(req
->wb_page
);
782 struct inode
*inode
= mapping
->host
;
783 struct nfs_inode
*nfsi
= NFS_I(inode
);
784 struct nfs_page
*head
;
786 atomic_long_dec(&nfsi
->nrequests
);
787 if (nfs_page_group_sync_on_bit(req
, PG_REMOVE
)) {
790 spin_lock(&mapping
->private_lock
);
791 if (likely(head
->wb_page
&& !PageSwapCache(head
->wb_page
))) {
792 set_page_private(head
->wb_page
, 0);
793 ClearPagePrivate(head
->wb_page
);
794 clear_bit(PG_MAPPED
, &head
->wb_flags
);
796 spin_unlock(&mapping
->private_lock
);
799 if (test_and_clear_bit(PG_INODE_REF
, &req
->wb_flags
))
800 nfs_release_request(req
);
804 nfs_mark_request_dirty(struct nfs_page
*req
)
807 __set_page_dirty_nobuffers(req
->wb_page
);
811 * nfs_page_search_commits_for_head_request_locked
813 * Search through commit lists on @inode for the head request for @page.
814 * Must be called while holding the inode (which is cinfo) lock.
816 * Returns the head request if found, or NULL if not found.
818 static struct nfs_page
*
819 nfs_page_search_commits_for_head_request_locked(struct nfs_inode
*nfsi
,
822 struct nfs_page
*freq
, *t
;
823 struct nfs_commit_info cinfo
;
824 struct inode
*inode
= &nfsi
->vfs_inode
;
826 nfs_init_cinfo_from_inode(&cinfo
, inode
);
828 /* search through pnfs commit lists */
829 freq
= pnfs_search_commit_reqs(inode
, &cinfo
, page
);
831 return freq
->wb_head
;
833 /* Linearly search the commit list for the correct request */
834 list_for_each_entry_safe(freq
, t
, &cinfo
.mds
->list
, wb_list
) {
835 if (freq
->wb_page
== page
)
836 return freq
->wb_head
;
843 * nfs_request_add_commit_list_locked - add request to a commit list
844 * @req: pointer to a struct nfs_page
845 * @dst: commit list head
846 * @cinfo: holds list lock and accounting info
848 * This sets the PG_CLEAN bit, updates the cinfo count of
849 * number of outstanding requests requiring a commit as well as
852 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
856 nfs_request_add_commit_list_locked(struct nfs_page
*req
, struct list_head
*dst
,
857 struct nfs_commit_info
*cinfo
)
859 set_bit(PG_CLEAN
, &req
->wb_flags
);
860 nfs_list_add_request(req
, dst
);
861 atomic_long_inc(&cinfo
->mds
->ncommit
);
863 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked
);
866 * nfs_request_add_commit_list - add request to a commit list
867 * @req: pointer to a struct nfs_page
868 * @dst: commit list head
869 * @cinfo: holds list lock and accounting info
871 * This sets the PG_CLEAN bit, updates the cinfo count of
872 * number of outstanding requests requiring a commit as well as
875 * The caller must _not_ hold the cinfo->lock, but must be
876 * holding the nfs_page lock.
879 nfs_request_add_commit_list(struct nfs_page
*req
, struct nfs_commit_info
*cinfo
)
881 mutex_lock(&NFS_I(cinfo
->inode
)->commit_mutex
);
882 nfs_request_add_commit_list_locked(req
, &cinfo
->mds
->list
, cinfo
);
883 mutex_unlock(&NFS_I(cinfo
->inode
)->commit_mutex
);
885 nfs_mark_page_unstable(req
->wb_page
, cinfo
);
887 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list
);
890 * nfs_request_remove_commit_list - Remove request from a commit list
891 * @req: pointer to a nfs_page
892 * @cinfo: holds list lock and accounting info
894 * This clears the PG_CLEAN bit, and updates the cinfo's count of
895 * number of outstanding requests requiring a commit
896 * It does not update the MM page stats.
898 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
901 nfs_request_remove_commit_list(struct nfs_page
*req
,
902 struct nfs_commit_info
*cinfo
)
904 if (!test_and_clear_bit(PG_CLEAN
, &(req
)->wb_flags
))
906 nfs_list_remove_request(req
);
907 atomic_long_dec(&cinfo
->mds
->ncommit
);
909 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list
);
911 static void nfs_init_cinfo_from_inode(struct nfs_commit_info
*cinfo
,
914 cinfo
->inode
= inode
;
915 cinfo
->mds
= &NFS_I(inode
)->commit_info
;
916 cinfo
->ds
= pnfs_get_ds_info(inode
);
918 cinfo
->completion_ops
= &nfs_commit_completion_ops
;
921 void nfs_init_cinfo(struct nfs_commit_info
*cinfo
,
923 struct nfs_direct_req
*dreq
)
926 nfs_init_cinfo_from_dreq(cinfo
, dreq
);
928 nfs_init_cinfo_from_inode(cinfo
, inode
);
930 EXPORT_SYMBOL_GPL(nfs_init_cinfo
);
933 * Add a request to the inode's commit list.
936 nfs_mark_request_commit(struct nfs_page
*req
, struct pnfs_layout_segment
*lseg
,
937 struct nfs_commit_info
*cinfo
, u32 ds_commit_idx
)
939 if (pnfs_mark_request_commit(req
, lseg
, cinfo
, ds_commit_idx
))
941 nfs_request_add_commit_list(req
, cinfo
);
945 nfs_clear_page_commit(struct page
*page
)
947 dec_node_page_state(page
, NR_UNSTABLE_NFS
);
948 dec_wb_stat(&inode_to_bdi(page_file_mapping(page
)->host
)->wb
,
952 /* Called holding the request lock on @req */
954 nfs_clear_request_commit(struct nfs_page
*req
)
956 if (test_bit(PG_CLEAN
, &req
->wb_flags
)) {
957 struct inode
*inode
= d_inode(req
->wb_context
->dentry
);
958 struct nfs_commit_info cinfo
;
960 nfs_init_cinfo_from_inode(&cinfo
, inode
);
961 mutex_lock(&NFS_I(inode
)->commit_mutex
);
962 if (!pnfs_clear_request_commit(req
, &cinfo
)) {
963 nfs_request_remove_commit_list(req
, &cinfo
);
965 mutex_unlock(&NFS_I(inode
)->commit_mutex
);
966 nfs_clear_page_commit(req
->wb_page
);
970 int nfs_write_need_commit(struct nfs_pgio_header
*hdr
)
972 if (hdr
->verf
.committed
== NFS_DATA_SYNC
)
973 return hdr
->lseg
== NULL
;
974 return hdr
->verf
.committed
!= NFS_FILE_SYNC
;
977 static void nfs_async_write_init(struct nfs_pgio_header
*hdr
)
979 nfs_io_completion_get(hdr
->io_completion
);
982 static void nfs_write_completion(struct nfs_pgio_header
*hdr
)
984 struct nfs_commit_info cinfo
;
985 unsigned long bytes
= 0;
987 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
989 nfs_init_cinfo_from_inode(&cinfo
, hdr
->inode
);
990 while (!list_empty(&hdr
->pages
)) {
991 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
993 bytes
+= req
->wb_bytes
;
994 nfs_list_remove_request(req
);
995 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
) &&
996 (hdr
->good_bytes
< bytes
)) {
997 nfs_set_pageerror(req
->wb_page
);
998 nfs_context_set_write_error(req
->wb_context
, hdr
->error
);
1001 if (nfs_write_need_commit(hdr
)) {
1002 memcpy(&req
->wb_verf
, &hdr
->verf
.verifier
, sizeof(req
->wb_verf
));
1003 nfs_mark_request_commit(req
, hdr
->lseg
, &cinfo
,
1004 hdr
->pgio_mirror_idx
);
1008 nfs_inode_remove_request(req
);
1010 nfs_end_page_writeback(req
);
1011 nfs_release_request(req
);
1014 nfs_io_completion_put(hdr
->io_completion
);
1019 nfs_reqs_to_commit(struct nfs_commit_info
*cinfo
)
1021 return atomic_long_read(&cinfo
->mds
->ncommit
);
1024 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1026 nfs_scan_commit_list(struct list_head
*src
, struct list_head
*dst
,
1027 struct nfs_commit_info
*cinfo
, int max
)
1029 struct nfs_page
*req
, *tmp
;
1033 list_for_each_entry_safe(req
, tmp
, src
, wb_list
) {
1034 kref_get(&req
->wb_kref
);
1035 if (!nfs_lock_request(req
)) {
1038 /* Prevent deadlock with nfs_lock_and_join_requests */
1039 if (!list_empty(dst
)) {
1040 nfs_release_request(req
);
1043 /* Ensure we make progress to prevent livelock */
1044 mutex_unlock(&NFS_I(cinfo
->inode
)->commit_mutex
);
1045 status
= nfs_wait_on_request(req
);
1046 nfs_release_request(req
);
1047 mutex_lock(&NFS_I(cinfo
->inode
)->commit_mutex
);
1052 nfs_request_remove_commit_list(req
, cinfo
);
1053 clear_bit(PG_COMMIT_TO_DS
, &req
->wb_flags
);
1054 nfs_list_add_request(req
, dst
);
1056 if ((ret
== max
) && !cinfo
->dreq
)
1062 EXPORT_SYMBOL_GPL(nfs_scan_commit_list
);
1065 * nfs_scan_commit - Scan an inode for commit requests
1066 * @inode: NFS inode to scan
1067 * @dst: mds destination list
1068 * @cinfo: mds and ds lists of reqs ready to commit
1070 * Moves requests from the inode's 'commit' request list.
1071 * The requests are *not* checked to ensure that they form a contiguous set.
1074 nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
,
1075 struct nfs_commit_info
*cinfo
)
1079 if (!atomic_long_read(&cinfo
->mds
->ncommit
))
1081 mutex_lock(&NFS_I(cinfo
->inode
)->commit_mutex
);
1082 if (atomic_long_read(&cinfo
->mds
->ncommit
) > 0) {
1083 const int max
= INT_MAX
;
1085 ret
= nfs_scan_commit_list(&cinfo
->mds
->list
, dst
,
1087 ret
+= pnfs_scan_commit_lists(inode
, cinfo
, max
- ret
);
1089 mutex_unlock(&NFS_I(cinfo
->inode
)->commit_mutex
);
1094 * Search for an existing write request, and attempt to update
1095 * it to reflect a new dirty region on a given page.
1097 * If the attempt fails, then the existing request is flushed out
1100 static struct nfs_page
*nfs_try_to_update_request(struct inode
*inode
,
1102 unsigned int offset
,
1105 struct nfs_page
*req
;
1110 end
= offset
+ bytes
;
1112 req
= nfs_lock_and_join_requests(page
);
1113 if (IS_ERR_OR_NULL(req
))
1116 rqend
= req
->wb_offset
+ req
->wb_bytes
;
1118 * Tell the caller to flush out the request if
1119 * the offsets are non-contiguous.
1120 * Note: nfs_flush_incompatible() will already
1121 * have flushed out requests having wrong owners.
1123 if (offset
> rqend
|| end
< req
->wb_offset
)
1126 /* Okay, the request matches. Update the region */
1127 if (offset
< req
->wb_offset
) {
1128 req
->wb_offset
= offset
;
1129 req
->wb_pgbase
= offset
;
1132 req
->wb_bytes
= end
- req
->wb_offset
;
1134 req
->wb_bytes
= rqend
- req
->wb_offset
;
1138 * Note: we mark the request dirty here because
1139 * nfs_lock_and_join_requests() cannot preserve
1140 * commit flags, so we have to replay the write.
1142 nfs_mark_request_dirty(req
);
1143 nfs_unlock_and_release_request(req
);
1144 error
= nfs_wb_page(inode
, page
);
1145 return (error
< 0) ? ERR_PTR(error
) : NULL
;
1149 * Try to update an existing write request, or create one if there is none.
1151 * Note: Should always be called with the Page Lock held to prevent races
1152 * if we have to add a new request. Also assumes that the caller has
1153 * already called nfs_flush_incompatible() if necessary.
1155 static struct nfs_page
* nfs_setup_write_request(struct nfs_open_context
* ctx
,
1156 struct page
*page
, unsigned int offset
, unsigned int bytes
)
1158 struct inode
*inode
= page_file_mapping(page
)->host
;
1159 struct nfs_page
*req
;
1161 req
= nfs_try_to_update_request(inode
, page
, offset
, bytes
);
1164 req
= nfs_create_request(ctx
, page
, NULL
, offset
, bytes
);
1167 nfs_inode_add_request(inode
, req
);
1172 static int nfs_writepage_setup(struct nfs_open_context
*ctx
, struct page
*page
,
1173 unsigned int offset
, unsigned int count
)
1175 struct nfs_page
*req
;
1177 req
= nfs_setup_write_request(ctx
, page
, offset
, count
);
1179 return PTR_ERR(req
);
1180 /* Update file length */
1181 nfs_grow_file(page
, offset
, count
);
1182 nfs_mark_uptodate(req
);
1183 nfs_mark_request_dirty(req
);
1184 nfs_unlock_and_release_request(req
);
1188 int nfs_flush_incompatible(struct file
*file
, struct page
*page
)
1190 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
1191 struct nfs_lock_context
*l_ctx
;
1192 struct file_lock_context
*flctx
= file_inode(file
)->i_flctx
;
1193 struct nfs_page
*req
;
1194 int do_flush
, status
;
1196 * Look for a request corresponding to this page. If there
1197 * is one, and it belongs to another file, we flush it out
1198 * before we try to copy anything into the page. Do this
1199 * due to the lack of an ACCESS-type call in NFSv2.
1200 * Also do the same if we find a request from an existing
1204 req
= nfs_page_find_head_request(page
);
1207 l_ctx
= req
->wb_lock_context
;
1208 do_flush
= req
->wb_page
!= page
||
1209 !nfs_match_open_context(req
->wb_context
, ctx
);
1210 if (l_ctx
&& flctx
&&
1211 !(list_empty_careful(&flctx
->flc_posix
) &&
1212 list_empty_careful(&flctx
->flc_flock
))) {
1213 do_flush
|= l_ctx
->lockowner
!= current
->files
;
1215 nfs_release_request(req
);
1218 status
= nfs_wb_page(page_file_mapping(page
)->host
, page
);
1219 } while (status
== 0);
1224 * Avoid buffered writes when a open context credential's key would
1227 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1229 * Return 0 and set a credential flag which triggers the inode to flush
1230 * and performs NFS_FILE_SYNC writes if the key will expired within
1231 * RPC_KEY_EXPIRE_TIMEO.
1234 nfs_key_timeout_notify(struct file
*filp
, struct inode
*inode
)
1236 struct nfs_open_context
*ctx
= nfs_file_open_context(filp
);
1238 if (nfs_ctx_key_to_expire(ctx
, inode
) &&
1240 /* Already expired! */
1246 * Test if the open context credential key is marked to expire soon.
1248 bool nfs_ctx_key_to_expire(struct nfs_open_context
*ctx
, struct inode
*inode
)
1250 struct rpc_auth
*auth
= NFS_SERVER(inode
)->client
->cl_auth
;
1251 struct rpc_cred
*cred
= ctx
->ll_cred
;
1252 struct auth_cred acred
= {
1256 if (cred
&& !cred
->cr_ops
->crmatch(&acred
, cred
, 0)) {
1258 ctx
->ll_cred
= NULL
;
1262 cred
= auth
->au_ops
->lookup_cred(auth
, &acred
, 0);
1263 if (!cred
|| IS_ERR(cred
))
1265 ctx
->ll_cred
= cred
;
1266 return !!(cred
->cr_ops
->crkey_timeout
&&
1267 cred
->cr_ops
->crkey_timeout(cred
));
1271 * If the page cache is marked as unsafe or invalid, then we can't rely on
1272 * the PageUptodate() flag. In this case, we will need to turn off
1273 * write optimisations that depend on the page contents being correct.
1275 static bool nfs_write_pageuptodate(struct page
*page
, struct inode
*inode
)
1277 struct nfs_inode
*nfsi
= NFS_I(inode
);
1279 if (nfs_have_delegated_attributes(inode
))
1281 if (nfsi
->cache_validity
& NFS_INO_REVAL_PAGECACHE
)
1284 if (test_bit(NFS_INO_INVALIDATING
, &nfsi
->flags
))
1287 if (nfsi
->cache_validity
& NFS_INO_INVALID_DATA
)
1289 return PageUptodate(page
) != 0;
1293 is_whole_file_wrlock(struct file_lock
*fl
)
1295 return fl
->fl_start
== 0 && fl
->fl_end
== OFFSET_MAX
&&
1296 fl
->fl_type
== F_WRLCK
;
1299 /* If we know the page is up to date, and we're not using byte range locks (or
1300 * if we have the whole file locked for writing), it may be more efficient to
1301 * extend the write to cover the entire page in order to avoid fragmentation
1304 * If the file is opened for synchronous writes then we can just skip the rest
1307 static int nfs_can_extend_write(struct file
*file
, struct page
*page
, struct inode
*inode
)
1310 struct file_lock_context
*flctx
= inode
->i_flctx
;
1311 struct file_lock
*fl
;
1313 if (file
->f_flags
& O_DSYNC
)
1315 if (!nfs_write_pageuptodate(page
, inode
))
1317 if (NFS_PROTO(inode
)->have_delegation(inode
, FMODE_WRITE
))
1319 if (!flctx
|| (list_empty_careful(&flctx
->flc_flock
) &&
1320 list_empty_careful(&flctx
->flc_posix
)))
1323 /* Check to see if there are whole file write locks */
1325 spin_lock(&flctx
->flc_lock
);
1326 if (!list_empty(&flctx
->flc_posix
)) {
1327 fl
= list_first_entry(&flctx
->flc_posix
, struct file_lock
,
1329 if (is_whole_file_wrlock(fl
))
1331 } else if (!list_empty(&flctx
->flc_flock
)) {
1332 fl
= list_first_entry(&flctx
->flc_flock
, struct file_lock
,
1334 if (fl
->fl_type
== F_WRLCK
)
1337 spin_unlock(&flctx
->flc_lock
);
1342 * Update and possibly write a cached page of an NFS file.
1344 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1345 * things with a page scheduled for an RPC call (e.g. invalidate it).
1347 int nfs_updatepage(struct file
*file
, struct page
*page
,
1348 unsigned int offset
, unsigned int count
)
1350 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
1351 struct inode
*inode
= page_file_mapping(page
)->host
;
1354 nfs_inc_stats(inode
, NFSIOS_VFSUPDATEPAGE
);
1356 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1357 file
, count
, (long long)(page_file_offset(page
) + offset
));
1362 if (nfs_can_extend_write(file
, page
, inode
)) {
1363 count
= max(count
+ offset
, nfs_page_length(page
));
1367 status
= nfs_writepage_setup(ctx
, page
, offset
, count
);
1369 nfs_set_pageerror(page
);
1371 __set_page_dirty_nobuffers(page
);
1373 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1374 status
, (long long)i_size_read(inode
));
1378 static int flush_task_priority(int how
)
1380 switch (how
& (FLUSH_HIGHPRI
|FLUSH_LOWPRI
)) {
1382 return RPC_PRIORITY_HIGH
;
1384 return RPC_PRIORITY_LOW
;
1386 return RPC_PRIORITY_NORMAL
;
1389 static void nfs_initiate_write(struct nfs_pgio_header
*hdr
,
1390 struct rpc_message
*msg
,
1391 const struct nfs_rpc_ops
*rpc_ops
,
1392 struct rpc_task_setup
*task_setup_data
, int how
)
1394 int priority
= flush_task_priority(how
);
1396 task_setup_data
->priority
= priority
;
1397 rpc_ops
->write_setup(hdr
, msg
, &task_setup_data
->rpc_client
);
1398 trace_nfs_initiate_write(hdr
->inode
, hdr
->io_start
, hdr
->good_bytes
,
1402 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1403 * call this on each, which will prepare them to be retried on next
1404 * writeback using standard nfs.
1406 static void nfs_redirty_request(struct nfs_page
*req
)
1408 nfs_mark_request_dirty(req
);
1409 set_bit(NFS_CONTEXT_RESEND_WRITES
, &req
->wb_context
->flags
);
1410 nfs_end_page_writeback(req
);
1411 nfs_release_request(req
);
1414 static void nfs_async_write_error(struct list_head
*head
)
1416 struct nfs_page
*req
;
1418 while (!list_empty(head
)) {
1419 req
= nfs_list_entry(head
->next
);
1420 nfs_list_remove_request(req
);
1421 nfs_redirty_request(req
);
1425 static void nfs_async_write_reschedule_io(struct nfs_pgio_header
*hdr
)
1427 nfs_async_write_error(&hdr
->pages
);
1428 filemap_fdatawrite_range(hdr
->inode
->i_mapping
, hdr
->args
.offset
,
1429 hdr
->args
.offset
+ hdr
->args
.count
- 1);
1432 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops
= {
1433 .init_hdr
= nfs_async_write_init
,
1434 .error_cleanup
= nfs_async_write_error
,
1435 .completion
= nfs_write_completion
,
1436 .reschedule_io
= nfs_async_write_reschedule_io
,
1439 void nfs_pageio_init_write(struct nfs_pageio_descriptor
*pgio
,
1440 struct inode
*inode
, int ioflags
, bool force_mds
,
1441 const struct nfs_pgio_completion_ops
*compl_ops
)
1443 struct nfs_server
*server
= NFS_SERVER(inode
);
1444 const struct nfs_pageio_ops
*pg_ops
= &nfs_pgio_rw_ops
;
1446 #ifdef CONFIG_NFS_V4_1
1447 if (server
->pnfs_curr_ld
&& !force_mds
)
1448 pg_ops
= server
->pnfs_curr_ld
->pg_write_ops
;
1450 nfs_pageio_init(pgio
, inode
, pg_ops
, compl_ops
, &nfs_rw_write_ops
,
1451 server
->wsize
, ioflags
);
1453 EXPORT_SYMBOL_GPL(nfs_pageio_init_write
);
1455 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor
*pgio
)
1457 struct nfs_pgio_mirror
*mirror
;
1459 if (pgio
->pg_ops
&& pgio
->pg_ops
->pg_cleanup
)
1460 pgio
->pg_ops
->pg_cleanup(pgio
);
1462 pgio
->pg_ops
= &nfs_pgio_rw_ops
;
1464 nfs_pageio_stop_mirroring(pgio
);
1466 mirror
= &pgio
->pg_mirrors
[0];
1467 mirror
->pg_bsize
= NFS_SERVER(pgio
->pg_inode
)->wsize
;
1469 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds
);
1472 void nfs_commit_prepare(struct rpc_task
*task
, void *calldata
)
1474 struct nfs_commit_data
*data
= calldata
;
1476 NFS_PROTO(data
->inode
)->commit_rpc_prepare(task
, data
);
1480 * Special version of should_remove_suid() that ignores capabilities.
1482 static int nfs_should_remove_suid(const struct inode
*inode
)
1484 umode_t mode
= inode
->i_mode
;
1487 /* suid always must be killed */
1488 if (unlikely(mode
& S_ISUID
))
1489 kill
= ATTR_KILL_SUID
;
1492 * sgid without any exec bits is just a mandatory locking mark; leave
1493 * it alone. If some exec bits are set, it's a real sgid; kill it.
1495 if (unlikely((mode
& S_ISGID
) && (mode
& S_IXGRP
)))
1496 kill
|= ATTR_KILL_SGID
;
1498 if (unlikely(kill
&& S_ISREG(mode
)))
1504 static void nfs_writeback_check_extend(struct nfs_pgio_header
*hdr
,
1505 struct nfs_fattr
*fattr
)
1507 struct nfs_pgio_args
*argp
= &hdr
->args
;
1508 struct nfs_pgio_res
*resp
= &hdr
->res
;
1509 u64 size
= argp
->offset
+ resp
->count
;
1511 if (!(fattr
->valid
& NFS_ATTR_FATTR_SIZE
))
1513 if (nfs_size_to_loff_t(fattr
->size
) < i_size_read(hdr
->inode
)) {
1514 fattr
->valid
&= ~NFS_ATTR_FATTR_SIZE
;
1517 if (size
!= fattr
->size
)
1519 /* Set attribute barrier */
1520 nfs_fattr_set_barrier(fattr
);
1521 /* ...and update size */
1522 fattr
->valid
|= NFS_ATTR_FATTR_SIZE
;
1525 void nfs_writeback_update_inode(struct nfs_pgio_header
*hdr
)
1527 struct nfs_fattr
*fattr
= &hdr
->fattr
;
1528 struct inode
*inode
= hdr
->inode
;
1530 spin_lock(&inode
->i_lock
);
1531 nfs_writeback_check_extend(hdr
, fattr
);
1532 nfs_post_op_update_inode_force_wcc_locked(inode
, fattr
);
1533 spin_unlock(&inode
->i_lock
);
1535 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode
);
1538 * This function is called when the WRITE call is complete.
1540 static int nfs_writeback_done(struct rpc_task
*task
,
1541 struct nfs_pgio_header
*hdr
,
1542 struct inode
*inode
)
1547 * ->write_done will attempt to use post-op attributes to detect
1548 * conflicting writes by other clients. A strict interpretation
1549 * of close-to-open would allow us to continue caching even if
1550 * another writer had changed the file, but some applications
1551 * depend on tighter cache coherency when writing.
1553 status
= NFS_PROTO(inode
)->write_done(task
, hdr
);
1557 nfs_add_stats(inode
, NFSIOS_SERVERWRITTENBYTES
, hdr
->res
.count
);
1558 trace_nfs_writeback_done(inode
, task
->tk_status
,
1559 hdr
->args
.offset
, hdr
->res
.verf
);
1561 if (hdr
->res
.verf
->committed
< hdr
->args
.stable
&&
1562 task
->tk_status
>= 0) {
1563 /* We tried a write call, but the server did not
1564 * commit data to stable storage even though we
1566 * Note: There is a known bug in Tru64 < 5.0 in which
1567 * the server reports NFS_DATA_SYNC, but performs
1568 * NFS_FILE_SYNC. We therefore implement this checking
1569 * as a dprintk() in order to avoid filling syslog.
1571 static unsigned long complain
;
1573 /* Note this will print the MDS for a DS write */
1574 if (time_before(complain
, jiffies
)) {
1575 dprintk("NFS: faulty NFS server %s:"
1576 " (committed = %d) != (stable = %d)\n",
1577 NFS_SERVER(inode
)->nfs_client
->cl_hostname
,
1578 hdr
->res
.verf
->committed
, hdr
->args
.stable
);
1579 complain
= jiffies
+ 300 * HZ
;
1583 /* Deal with the suid/sgid bit corner case */
1584 if (nfs_should_remove_suid(inode
)) {
1585 spin_lock(&inode
->i_lock
);
1586 NFS_I(inode
)->cache_validity
|= NFS_INO_INVALID_OTHER
;
1587 spin_unlock(&inode
->i_lock
);
1593 * This function is called when the WRITE call is complete.
1595 static void nfs_writeback_result(struct rpc_task
*task
,
1596 struct nfs_pgio_header
*hdr
)
1598 struct nfs_pgio_args
*argp
= &hdr
->args
;
1599 struct nfs_pgio_res
*resp
= &hdr
->res
;
1601 if (resp
->count
< argp
->count
) {
1602 static unsigned long complain
;
1604 /* This a short write! */
1605 nfs_inc_stats(hdr
->inode
, NFSIOS_SHORTWRITE
);
1607 /* Has the server at least made some progress? */
1608 if (resp
->count
== 0) {
1609 if (time_before(complain
, jiffies
)) {
1611 "NFS: Server wrote zero bytes, expected %u.\n",
1613 complain
= jiffies
+ 300 * HZ
;
1615 nfs_set_pgio_error(hdr
, -EIO
, argp
->offset
);
1616 task
->tk_status
= -EIO
;
1620 /* For non rpc-based layout drivers, retry-through-MDS */
1621 if (!task
->tk_ops
) {
1622 hdr
->pnfs_error
= -EAGAIN
;
1626 /* Was this an NFSv2 write or an NFSv3 stable write? */
1627 if (resp
->verf
->committed
!= NFS_UNSTABLE
) {
1628 /* Resend from where the server left off */
1629 hdr
->mds_offset
+= resp
->count
;
1630 argp
->offset
+= resp
->count
;
1631 argp
->pgbase
+= resp
->count
;
1632 argp
->count
-= resp
->count
;
1634 /* Resend as a stable write in order to avoid
1635 * headaches in the case of a server crash.
1637 argp
->stable
= NFS_FILE_SYNC
;
1639 rpc_restart_call_prepare(task
);
1643 static int wait_on_commit(struct nfs_mds_commit_info
*cinfo
)
1645 return wait_var_event_killable(&cinfo
->rpcs_out
,
1646 !atomic_read(&cinfo
->rpcs_out
));
1649 static void nfs_commit_begin(struct nfs_mds_commit_info
*cinfo
)
1651 atomic_inc(&cinfo
->rpcs_out
);
1654 static void nfs_commit_end(struct nfs_mds_commit_info
*cinfo
)
1656 if (atomic_dec_and_test(&cinfo
->rpcs_out
))
1657 wake_up_var(&cinfo
->rpcs_out
);
1660 void nfs_commitdata_release(struct nfs_commit_data
*data
)
1662 put_nfs_open_context(data
->context
);
1663 nfs_commit_free(data
);
1665 EXPORT_SYMBOL_GPL(nfs_commitdata_release
);
1667 int nfs_initiate_commit(struct rpc_clnt
*clnt
, struct nfs_commit_data
*data
,
1668 const struct nfs_rpc_ops
*nfs_ops
,
1669 const struct rpc_call_ops
*call_ops
,
1672 struct rpc_task
*task
;
1673 int priority
= flush_task_priority(how
);
1674 struct rpc_message msg
= {
1675 .rpc_argp
= &data
->args
,
1676 .rpc_resp
= &data
->res
,
1677 .rpc_cred
= data
->cred
,
1679 struct rpc_task_setup task_setup_data
= {
1680 .task
= &data
->task
,
1682 .rpc_message
= &msg
,
1683 .callback_ops
= call_ops
,
1684 .callback_data
= data
,
1685 .workqueue
= nfsiod_workqueue
,
1686 .flags
= RPC_TASK_ASYNC
| flags
,
1687 .priority
= priority
,
1689 /* Set up the initial task struct. */
1690 nfs_ops
->commit_setup(data
, &msg
, &task_setup_data
.rpc_client
);
1691 trace_nfs_initiate_commit(data
);
1693 dprintk("NFS: initiated commit call\n");
1695 task
= rpc_run_task(&task_setup_data
);
1697 return PTR_ERR(task
);
1698 if (how
& FLUSH_SYNC
)
1699 rpc_wait_for_completion_task(task
);
1703 EXPORT_SYMBOL_GPL(nfs_initiate_commit
);
1705 static loff_t
nfs_get_lwb(struct list_head
*head
)
1708 struct nfs_page
*req
;
1710 list_for_each_entry(req
, head
, wb_list
)
1711 if (lwb
< (req_offset(req
) + req
->wb_bytes
))
1712 lwb
= req_offset(req
) + req
->wb_bytes
;
1718 * Set up the argument/result storage required for the RPC call.
1720 void nfs_init_commit(struct nfs_commit_data
*data
,
1721 struct list_head
*head
,
1722 struct pnfs_layout_segment
*lseg
,
1723 struct nfs_commit_info
*cinfo
)
1725 struct nfs_page
*first
= nfs_list_entry(head
->next
);
1726 struct inode
*inode
= d_inode(first
->wb_context
->dentry
);
1728 /* Set up the RPC argument and reply structs
1729 * NB: take care not to mess about with data->commit et al. */
1731 list_splice_init(head
, &data
->pages
);
1733 data
->inode
= inode
;
1734 data
->cred
= first
->wb_context
->cred
;
1735 data
->lseg
= lseg
; /* reference transferred */
1736 /* only set lwb for pnfs commit */
1738 data
->lwb
= nfs_get_lwb(&data
->pages
);
1739 data
->mds_ops
= &nfs_commit_ops
;
1740 data
->completion_ops
= cinfo
->completion_ops
;
1741 data
->dreq
= cinfo
->dreq
;
1743 data
->args
.fh
= NFS_FH(data
->inode
);
1744 /* Note: we always request a commit of the entire inode */
1745 data
->args
.offset
= 0;
1746 data
->args
.count
= 0;
1747 data
->context
= get_nfs_open_context(first
->wb_context
);
1748 data
->res
.fattr
= &data
->fattr
;
1749 data
->res
.verf
= &data
->verf
;
1750 nfs_fattr_init(&data
->fattr
);
1752 EXPORT_SYMBOL_GPL(nfs_init_commit
);
1754 void nfs_retry_commit(struct list_head
*page_list
,
1755 struct pnfs_layout_segment
*lseg
,
1756 struct nfs_commit_info
*cinfo
,
1759 struct nfs_page
*req
;
1761 while (!list_empty(page_list
)) {
1762 req
= nfs_list_entry(page_list
->next
);
1763 nfs_list_remove_request(req
);
1764 nfs_mark_request_commit(req
, lseg
, cinfo
, ds_commit_idx
);
1766 nfs_clear_page_commit(req
->wb_page
);
1767 nfs_unlock_and_release_request(req
);
1770 EXPORT_SYMBOL_GPL(nfs_retry_commit
);
1773 nfs_commit_resched_write(struct nfs_commit_info
*cinfo
,
1774 struct nfs_page
*req
)
1776 __set_page_dirty_nobuffers(req
->wb_page
);
1780 * Commit dirty pages
1783 nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
,
1784 struct nfs_commit_info
*cinfo
)
1786 struct nfs_commit_data
*data
;
1788 /* another commit raced with us */
1789 if (list_empty(head
))
1792 data
= nfs_commitdata_alloc(true);
1794 /* Set up the argument struct */
1795 nfs_init_commit(data
, head
, NULL
, cinfo
);
1796 atomic_inc(&cinfo
->mds
->rpcs_out
);
1797 return nfs_initiate_commit(NFS_CLIENT(inode
), data
, NFS_PROTO(inode
),
1798 data
->mds_ops
, how
, 0);
1802 * COMMIT call returned
1804 static void nfs_commit_done(struct rpc_task
*task
, void *calldata
)
1806 struct nfs_commit_data
*data
= calldata
;
1808 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1809 task
->tk_pid
, task
->tk_status
);
1811 /* Call the NFS version-specific code */
1812 NFS_PROTO(data
->inode
)->commit_done(task
, data
);
1813 trace_nfs_commit_done(data
);
1816 static void nfs_commit_release_pages(struct nfs_commit_data
*data
)
1818 struct nfs_page
*req
;
1819 int status
= data
->task
.tk_status
;
1820 struct nfs_commit_info cinfo
;
1821 struct nfs_server
*nfss
;
1823 while (!list_empty(&data
->pages
)) {
1824 req
= nfs_list_entry(data
->pages
.next
);
1825 nfs_list_remove_request(req
);
1827 nfs_clear_page_commit(req
->wb_page
);
1829 dprintk("NFS: commit (%s/%llu %d@%lld)",
1830 req
->wb_context
->dentry
->d_sb
->s_id
,
1831 (unsigned long long)NFS_FILEID(d_inode(req
->wb_context
->dentry
)),
1833 (long long)req_offset(req
));
1835 nfs_context_set_write_error(req
->wb_context
, status
);
1837 nfs_inode_remove_request(req
);
1838 dprintk_cont(", error = %d\n", status
);
1842 /* Okay, COMMIT succeeded, apparently. Check the verifier
1843 * returned by the server against all stored verfs. */
1844 if (!nfs_write_verifier_cmp(&req
->wb_verf
, &data
->verf
.verifier
)) {
1845 /* We have a match */
1847 nfs_inode_remove_request(req
);
1848 dprintk_cont(" OK\n");
1851 /* We have a mismatch. Write the page again */
1852 dprintk_cont(" mismatch\n");
1853 nfs_mark_request_dirty(req
);
1854 set_bit(NFS_CONTEXT_RESEND_WRITES
, &req
->wb_context
->flags
);
1856 nfs_unlock_and_release_request(req
);
1857 /* Latency breaker */
1860 nfss
= NFS_SERVER(data
->inode
);
1861 if (atomic_long_read(&nfss
->writeback
) < NFS_CONGESTION_OFF_THRESH
)
1862 clear_bdi_congested(inode_to_bdi(data
->inode
), BLK_RW_ASYNC
);
1864 nfs_init_cinfo(&cinfo
, data
->inode
, data
->dreq
);
1865 nfs_commit_end(cinfo
.mds
);
1868 static void nfs_commit_release(void *calldata
)
1870 struct nfs_commit_data
*data
= calldata
;
1872 data
->completion_ops
->completion(data
);
1873 nfs_commitdata_release(calldata
);
1876 static const struct rpc_call_ops nfs_commit_ops
= {
1877 .rpc_call_prepare
= nfs_commit_prepare
,
1878 .rpc_call_done
= nfs_commit_done
,
1879 .rpc_release
= nfs_commit_release
,
1882 static const struct nfs_commit_completion_ops nfs_commit_completion_ops
= {
1883 .completion
= nfs_commit_release_pages
,
1884 .resched_write
= nfs_commit_resched_write
,
1887 int nfs_generic_commit_list(struct inode
*inode
, struct list_head
*head
,
1888 int how
, struct nfs_commit_info
*cinfo
)
1892 status
= pnfs_commit_list(inode
, head
, how
, cinfo
);
1893 if (status
== PNFS_NOT_ATTEMPTED
)
1894 status
= nfs_commit_list(inode
, head
, how
, cinfo
);
1898 static int __nfs_commit_inode(struct inode
*inode
, int how
,
1899 struct writeback_control
*wbc
)
1902 struct nfs_commit_info cinfo
;
1903 int may_wait
= how
& FLUSH_SYNC
;
1906 nfs_init_cinfo_from_inode(&cinfo
, inode
);
1907 nfs_commit_begin(cinfo
.mds
);
1909 ret
= nscan
= nfs_scan_commit(inode
, &head
, &cinfo
);
1912 ret
= nfs_generic_commit_list(inode
, &head
, how
, &cinfo
);
1916 if (wbc
&& wbc
->sync_mode
== WB_SYNC_NONE
) {
1917 if (nscan
< wbc
->nr_to_write
)
1918 wbc
->nr_to_write
-= nscan
;
1920 wbc
->nr_to_write
= 0;
1922 if (nscan
< INT_MAX
)
1926 nfs_commit_end(cinfo
.mds
);
1927 if (ret
|| !may_wait
)
1929 return wait_on_commit(cinfo
.mds
);
1932 int nfs_commit_inode(struct inode
*inode
, int how
)
1934 return __nfs_commit_inode(inode
, how
, NULL
);
1936 EXPORT_SYMBOL_GPL(nfs_commit_inode
);
1938 int nfs_write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1940 struct nfs_inode
*nfsi
= NFS_I(inode
);
1941 int flags
= FLUSH_SYNC
;
1944 if (wbc
->sync_mode
== WB_SYNC_NONE
) {
1945 /* no commits means nothing needs to be done */
1946 if (!atomic_long_read(&nfsi
->commit_info
.ncommit
))
1947 goto check_requests_outstanding
;
1949 /* Don't commit yet if this is a non-blocking flush and there
1950 * are a lot of outstanding writes for this mapping.
1952 if (mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_WRITEBACK
))
1953 goto out_mark_dirty
;
1955 /* don't wait for the COMMIT response */
1959 ret
= __nfs_commit_inode(inode
, flags
, wbc
);
1961 if (flags
& FLUSH_SYNC
)
1963 } else if (atomic_long_read(&nfsi
->commit_info
.ncommit
))
1964 goto out_mark_dirty
;
1966 check_requests_outstanding
:
1967 if (!atomic_read(&nfsi
->commit_info
.rpcs_out
))
1970 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
1973 EXPORT_SYMBOL_GPL(nfs_write_inode
);
1976 * Wrapper for filemap_write_and_wait_range()
1978 * Needed for pNFS in order to ensure data becomes visible to the
1981 int nfs_filemap_write_and_wait_range(struct address_space
*mapping
,
1982 loff_t lstart
, loff_t lend
)
1986 ret
= filemap_write_and_wait_range(mapping
, lstart
, lend
);
1988 ret
= pnfs_sync_inode(mapping
->host
, true);
1991 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range
);
1994 * flush the inode to disk.
1996 int nfs_wb_all(struct inode
*inode
)
2000 trace_nfs_writeback_inode_enter(inode
);
2002 ret
= filemap_write_and_wait(inode
->i_mapping
);
2005 ret
= nfs_commit_inode(inode
, FLUSH_SYNC
);
2008 pnfs_sync_inode(inode
, true);
2012 trace_nfs_writeback_inode_exit(inode
, ret
);
2015 EXPORT_SYMBOL_GPL(nfs_wb_all
);
2017 int nfs_wb_page_cancel(struct inode
*inode
, struct page
*page
)
2019 struct nfs_page
*req
;
2022 wait_on_page_writeback(page
);
2024 /* blocking call to cancel all requests and join to a single (head)
2026 req
= nfs_lock_and_join_requests(page
);
2031 /* all requests from this page have been cancelled by
2032 * nfs_lock_and_join_requests, so just remove the head
2033 * request from the inode / page_private pointer and
2035 nfs_inode_remove_request(req
);
2036 nfs_unlock_and_release_request(req
);
2043 * Write back all requests on one page - we do this before reading it.
2045 int nfs_wb_page(struct inode
*inode
, struct page
*page
)
2047 loff_t range_start
= page_file_offset(page
);
2048 loff_t range_end
= range_start
+ (loff_t
)(PAGE_SIZE
- 1);
2049 struct writeback_control wbc
= {
2050 .sync_mode
= WB_SYNC_ALL
,
2052 .range_start
= range_start
,
2053 .range_end
= range_end
,
2057 trace_nfs_writeback_page_enter(inode
);
2060 wait_on_page_writeback(page
);
2061 if (clear_page_dirty_for_io(page
)) {
2062 ret
= nfs_writepage_locked(page
, &wbc
);
2068 if (!PagePrivate(page
))
2070 ret
= nfs_commit_inode(inode
, FLUSH_SYNC
);
2075 trace_nfs_writeback_page_exit(inode
, ret
);
2079 #ifdef CONFIG_MIGRATION
2080 int nfs_migrate_page(struct address_space
*mapping
, struct page
*newpage
,
2081 struct page
*page
, enum migrate_mode mode
)
2084 * If PagePrivate is set, then the page is currently associated with
2085 * an in-progress read or write request. Don't try to migrate it.
2087 * FIXME: we could do this in principle, but we'll need a way to ensure
2088 * that we can safely release the inode reference while holding
2091 if (PagePrivate(page
))
2094 if (!nfs_fscache_release_page(page
, GFP_KERNEL
))
2097 return migrate_page(mapping
, newpage
, page
, mode
);
2101 int __init
nfs_init_writepagecache(void)
2103 nfs_wdata_cachep
= kmem_cache_create("nfs_write_data",
2104 sizeof(struct nfs_pgio_header
),
2105 0, SLAB_HWCACHE_ALIGN
,
2107 if (nfs_wdata_cachep
== NULL
)
2110 nfs_wdata_mempool
= mempool_create_slab_pool(MIN_POOL_WRITE
,
2112 if (nfs_wdata_mempool
== NULL
)
2113 goto out_destroy_write_cache
;
2115 nfs_cdata_cachep
= kmem_cache_create("nfs_commit_data",
2116 sizeof(struct nfs_commit_data
),
2117 0, SLAB_HWCACHE_ALIGN
,
2119 if (nfs_cdata_cachep
== NULL
)
2120 goto out_destroy_write_mempool
;
2122 nfs_commit_mempool
= mempool_create_slab_pool(MIN_POOL_COMMIT
,
2124 if (nfs_commit_mempool
== NULL
)
2125 goto out_destroy_commit_cache
;
2128 * NFS congestion size, scale with available memory.
2140 * This allows larger machines to have larger/more transfers.
2141 * Limit the default to 256M
2143 nfs_congestion_kb
= (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT
-10);
2144 if (nfs_congestion_kb
> 256*1024)
2145 nfs_congestion_kb
= 256*1024;
2149 out_destroy_commit_cache
:
2150 kmem_cache_destroy(nfs_cdata_cachep
);
2151 out_destroy_write_mempool
:
2152 mempool_destroy(nfs_wdata_mempool
);
2153 out_destroy_write_cache
:
2154 kmem_cache_destroy(nfs_wdata_cachep
);
2158 void nfs_destroy_writepagecache(void)
2160 mempool_destroy(nfs_commit_mempool
);
2161 kmem_cache_destroy(nfs_cdata_cachep
);
2162 mempool_destroy(nfs_wdata_mempool
);
2163 kmem_cache_destroy(nfs_wdata_cachep
);
2166 static const struct nfs_rw_ops nfs_rw_write_ops
= {
2167 .rw_alloc_header
= nfs_writehdr_alloc
,
2168 .rw_free_header
= nfs_writehdr_free
,
2169 .rw_done
= nfs_writeback_done
,
2170 .rw_result
= nfs_writeback_result
,
2171 .rw_initiate
= nfs_initiate_write
,