4 * Write file data over NFS.
6 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
9 #include <linux/types.h>
10 #include <linux/slab.h>
12 #include <linux/pagemap.h>
13 #include <linux/file.h>
14 #include <linux/writeback.h>
15 #include <linux/swap.h>
16 #include <linux/migrate.h>
18 #include <linux/sunrpc/clnt.h>
19 #include <linux/nfs_fs.h>
20 #include <linux/nfs_mount.h>
21 #include <linux/nfs_page.h>
22 #include <linux/backing-dev.h>
23 #include <linux/export.h>
24 #include <linux/freezer.h>
25 #include <linux/wait.h>
27 #include <linux/uaccess.h>
29 #include "delegation.h"
38 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
40 #define MIN_POOL_WRITE (32)
41 #define MIN_POOL_COMMIT (4)
44 * Local function declarations
46 static void nfs_redirty_request(struct nfs_page
*req
);
47 static const struct rpc_call_ops nfs_commit_ops
;
48 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops
;
49 static const struct nfs_commit_completion_ops nfs_commit_completion_ops
;
50 static const struct nfs_rw_ops nfs_rw_write_ops
;
51 static void nfs_clear_request_commit(struct nfs_page
*req
);
52 static void nfs_init_cinfo_from_inode(struct nfs_commit_info
*cinfo
,
54 static struct nfs_page
*
55 nfs_page_search_commits_for_head_request_locked(struct nfs_inode
*nfsi
,
58 static struct kmem_cache
*nfs_wdata_cachep
;
59 static mempool_t
*nfs_wdata_mempool
;
60 static struct kmem_cache
*nfs_cdata_cachep
;
61 static mempool_t
*nfs_commit_mempool
;
63 struct nfs_commit_data
*nfs_commitdata_alloc(bool never_fail
)
65 struct nfs_commit_data
*p
;
68 p
= mempool_alloc(nfs_commit_mempool
, GFP_NOIO
);
70 /* It is OK to do some reclaim, not no safe to wait
71 * for anything to be returned to the pool.
72 * mempool_alloc() cannot handle that particular combination,
73 * so we need two separate attempts.
75 p
= mempool_alloc(nfs_commit_mempool
, GFP_NOWAIT
);
77 p
= kmem_cache_alloc(nfs_cdata_cachep
, GFP_NOIO
|
78 __GFP_NOWARN
| __GFP_NORETRY
);
83 memset(p
, 0, sizeof(*p
));
84 INIT_LIST_HEAD(&p
->pages
);
87 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc
);
89 void nfs_commit_free(struct nfs_commit_data
*p
)
91 mempool_free(p
, nfs_commit_mempool
);
93 EXPORT_SYMBOL_GPL(nfs_commit_free
);
95 static struct nfs_pgio_header
*nfs_writehdr_alloc(void)
97 struct nfs_pgio_header
*p
= mempool_alloc(nfs_wdata_mempool
, GFP_NOIO
);
100 memset(p
, 0, sizeof(*p
));
101 p
->rw_mode
= FMODE_WRITE
;
106 static void nfs_writehdr_free(struct nfs_pgio_header
*hdr
)
108 mempool_free(hdr
, nfs_wdata_mempool
);
111 static void nfs_context_set_write_error(struct nfs_open_context
*ctx
, int error
)
115 set_bit(NFS_CONTEXT_ERROR_WRITE
, &ctx
->flags
);
119 * nfs_page_find_head_request_locked - find head request associated with @page
121 * must be called while holding the inode lock.
123 * returns matching head request with reference held, or NULL if not found.
125 static struct nfs_page
*
126 nfs_page_find_head_request_locked(struct nfs_inode
*nfsi
, struct page
*page
)
128 struct nfs_page
*req
= NULL
;
130 if (PagePrivate(page
))
131 req
= (struct nfs_page
*)page_private(page
);
132 else if (unlikely(PageSwapCache(page
)))
133 req
= nfs_page_search_commits_for_head_request_locked(nfsi
,
137 WARN_ON_ONCE(req
->wb_head
!= req
);
138 kref_get(&req
->wb_kref
);
145 * nfs_page_find_head_request - find head request associated with @page
147 * returns matching head request with reference held, or NULL if not found.
149 static struct nfs_page
*nfs_page_find_head_request(struct page
*page
)
151 struct inode
*inode
= page_file_mapping(page
)->host
;
152 struct nfs_page
*req
= NULL
;
154 spin_lock(&inode
->i_lock
);
155 req
= nfs_page_find_head_request_locked(NFS_I(inode
), page
);
156 spin_unlock(&inode
->i_lock
);
160 /* Adjust the file length if we're writing beyond the end */
161 static void nfs_grow_file(struct page
*page
, unsigned int offset
, unsigned int count
)
163 struct inode
*inode
= page_file_mapping(page
)->host
;
167 spin_lock(&inode
->i_lock
);
168 i_size
= i_size_read(inode
);
169 end_index
= (i_size
- 1) >> PAGE_SHIFT
;
170 if (i_size
> 0 && page_index(page
) < end_index
)
172 end
= page_file_offset(page
) + ((loff_t
)offset
+count
);
175 i_size_write(inode
, end
);
176 nfs_inc_stats(inode
, NFSIOS_EXTENDWRITE
);
178 spin_unlock(&inode
->i_lock
);
181 /* A writeback failed: mark the page as bad, and invalidate the page cache */
182 static void nfs_set_pageerror(struct page
*page
)
184 nfs_zap_mapping(page_file_mapping(page
)->host
, page_file_mapping(page
));
188 * nfs_page_group_search_locked
189 * @head - head request of page group
190 * @page_offset - offset into page
192 * Search page group with head @head to find a request that contains the
193 * page offset @page_offset.
195 * Returns a pointer to the first matching nfs request, or NULL if no
198 * Must be called with the page group lock held
200 static struct nfs_page
*
201 nfs_page_group_search_locked(struct nfs_page
*head
, unsigned int page_offset
)
203 struct nfs_page
*req
;
205 WARN_ON_ONCE(head
!= head
->wb_head
);
206 WARN_ON_ONCE(!test_bit(PG_HEADLOCK
, &head
->wb_head
->wb_flags
));
210 if (page_offset
>= req
->wb_pgbase
&&
211 page_offset
< (req
->wb_pgbase
+ req
->wb_bytes
))
214 req
= req
->wb_this_page
;
215 } while (req
!= head
);
221 * nfs_page_group_covers_page
222 * @head - head request of page group
224 * Return true if the page group with head @head covers the whole page,
225 * returns false otherwise
227 static bool nfs_page_group_covers_page(struct nfs_page
*req
)
229 struct nfs_page
*tmp
;
230 unsigned int pos
= 0;
231 unsigned int len
= nfs_page_length(req
->wb_page
);
233 nfs_page_group_lock(req
, false);
236 tmp
= nfs_page_group_search_locked(req
->wb_head
, pos
);
238 /* no way this should happen */
239 WARN_ON_ONCE(tmp
->wb_pgbase
!= pos
);
240 pos
+= tmp
->wb_bytes
- (pos
- tmp
->wb_pgbase
);
242 } while (tmp
&& pos
< len
);
244 nfs_page_group_unlock(req
);
245 WARN_ON_ONCE(pos
> len
);
249 /* We can set the PG_uptodate flag if we see that a write request
250 * covers the full page.
252 static void nfs_mark_uptodate(struct nfs_page
*req
)
254 if (PageUptodate(req
->wb_page
))
256 if (!nfs_page_group_covers_page(req
))
258 SetPageUptodate(req
->wb_page
);
261 static int wb_priority(struct writeback_control
*wbc
)
265 if (wbc
->sync_mode
== WB_SYNC_ALL
)
266 ret
= FLUSH_COND_STABLE
;
271 * NFS congestion control
274 int nfs_congestion_kb
;
276 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10))
277 #define NFS_CONGESTION_OFF_THRESH \
278 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
280 static void nfs_set_page_writeback(struct page
*page
)
282 struct inode
*inode
= page_file_mapping(page
)->host
;
283 struct nfs_server
*nfss
= NFS_SERVER(inode
);
284 int ret
= test_set_page_writeback(page
);
286 WARN_ON_ONCE(ret
!= 0);
288 if (atomic_long_inc_return(&nfss
->writeback
) >
289 NFS_CONGESTION_ON_THRESH
)
290 set_bdi_congested(inode_to_bdi(inode
), BLK_RW_ASYNC
);
293 static void nfs_end_page_writeback(struct nfs_page
*req
)
295 struct inode
*inode
= page_file_mapping(req
->wb_page
)->host
;
296 struct nfs_server
*nfss
= NFS_SERVER(inode
);
298 if (!nfs_page_group_sync_on_bit(req
, PG_WB_END
))
301 end_page_writeback(req
->wb_page
);
302 if (atomic_long_dec_return(&nfss
->writeback
) < NFS_CONGESTION_OFF_THRESH
)
303 clear_bdi_congested(inode_to_bdi(inode
), BLK_RW_ASYNC
);
307 /* nfs_page_group_clear_bits
308 * @req - an nfs request
309 * clears all page group related bits from @req
312 nfs_page_group_clear_bits(struct nfs_page
*req
)
314 clear_bit(PG_TEARDOWN
, &req
->wb_flags
);
315 clear_bit(PG_UNLOCKPAGE
, &req
->wb_flags
);
316 clear_bit(PG_UPTODATE
, &req
->wb_flags
);
317 clear_bit(PG_WB_END
, &req
->wb_flags
);
318 clear_bit(PG_REMOVE
, &req
->wb_flags
);
323 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req
325 * this is a helper function for nfs_lock_and_join_requests
327 * @inode - inode associated with request page group, must be holding inode lock
328 * @head - head request of page group, must be holding head lock
329 * @req - request that couldn't lock and needs to wait on the req bit lock
330 * @nonblock - if true, don't actually wait
332 * NOTE: this must be called holding page_group bit lock and inode spin lock
333 * and BOTH will be released before returning.
335 * returns 0 on success, < 0 on error.
338 nfs_unroll_locks_and_wait(struct inode
*inode
, struct nfs_page
*head
,
339 struct nfs_page
*req
, bool nonblock
)
340 __releases(&inode
->i_lock
)
342 struct nfs_page
*tmp
;
345 /* relinquish all the locks successfully grabbed this run */
346 for (tmp
= head
; tmp
!= req
; tmp
= tmp
->wb_this_page
)
347 nfs_unlock_request(tmp
);
349 WARN_ON_ONCE(test_bit(PG_TEARDOWN
, &req
->wb_flags
));
351 /* grab a ref on the request that will be waited on */
352 kref_get(&req
->wb_kref
);
354 nfs_page_group_unlock(head
);
355 spin_unlock(&inode
->i_lock
);
357 /* release ref from nfs_page_find_head_request_locked */
358 nfs_release_request(head
);
361 ret
= nfs_wait_on_request(req
);
364 nfs_release_request(req
);
370 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
372 * @destroy_list - request list (using wb_this_page) terminated by @old_head
373 * @old_head - the old head of the list
375 * All subrequests must be locked and removed from all lists, so at this point
376 * they are only "active" in this function, and possibly in nfs_wait_on_request
377 * with a reference held by some other context.
380 nfs_destroy_unlinked_subrequests(struct nfs_page
*destroy_list
,
381 struct nfs_page
*old_head
)
383 while (destroy_list
) {
384 struct nfs_page
*subreq
= destroy_list
;
386 destroy_list
= (subreq
->wb_this_page
== old_head
) ?
387 NULL
: subreq
->wb_this_page
;
389 WARN_ON_ONCE(old_head
!= subreq
->wb_head
);
391 /* make sure old group is not used */
392 subreq
->wb_head
= subreq
;
393 subreq
->wb_this_page
= subreq
;
395 /* subreq is now totally disconnected from page group or any
396 * write / commit lists. last chance to wake any waiters */
397 nfs_unlock_request(subreq
);
399 if (!test_bit(PG_TEARDOWN
, &subreq
->wb_flags
)) {
400 /* release ref on old head request */
401 nfs_release_request(old_head
);
403 nfs_page_group_clear_bits(subreq
);
405 /* release the PG_INODE_REF reference */
406 if (test_and_clear_bit(PG_INODE_REF
, &subreq
->wb_flags
))
407 nfs_release_request(subreq
);
411 WARN_ON_ONCE(test_bit(PG_CLEAN
, &subreq
->wb_flags
));
412 /* zombie requests have already released the last
413 * reference and were waiting on the rest of the
414 * group to complete. Since it's no longer part of a
415 * group, simply free the request */
416 nfs_page_group_clear_bits(subreq
);
417 nfs_free_request(subreq
);
423 * nfs_lock_and_join_requests - join all subreqs to the head req and return
424 * a locked reference, cancelling any pending
425 * operations for this page.
427 * @page - the page used to lookup the "page group" of nfs_page structures
428 * @nonblock - if true, don't block waiting for request locks
430 * This function joins all sub requests to the head request by first
431 * locking all requests in the group, cancelling any pending operations
432 * and finally updating the head request to cover the whole range covered by
433 * the (former) group. All subrequests are removed from any write or commit
434 * lists, unlinked from the group and destroyed.
436 * Returns a locked, referenced pointer to the head request - which after
437 * this call is guaranteed to be the only request associated with the page.
438 * Returns NULL if no requests are found for @page, or a ERR_PTR if an
439 * error was encountered.
441 static struct nfs_page
*
442 nfs_lock_and_join_requests(struct page
*page
, bool nonblock
)
444 struct inode
*inode
= page_file_mapping(page
)->host
;
445 struct nfs_page
*head
, *subreq
;
446 struct nfs_page
*destroy_list
= NULL
;
447 unsigned int total_bytes
;
453 WARN_ON_ONCE(destroy_list
);
455 spin_lock(&inode
->i_lock
);
458 * A reference is taken only on the head request which acts as a
459 * reference to the whole page group - the group will not be destroyed
460 * until the head reference is released.
462 head
= nfs_page_find_head_request_locked(NFS_I(inode
), page
);
465 spin_unlock(&inode
->i_lock
);
469 /* holding inode lock, so always make a non-blocking call to try the
471 ret
= nfs_page_group_lock(head
, true);
473 spin_unlock(&inode
->i_lock
);
475 if (!nonblock
&& ret
== -EAGAIN
) {
476 nfs_page_group_lock_wait(head
);
477 nfs_release_request(head
);
481 nfs_release_request(head
);
485 /* lock each request in the page group */
489 * Subrequests are always contiguous, non overlapping
490 * and in order - but may be repeated (mirrored writes).
492 if (subreq
->wb_offset
== (head
->wb_offset
+ total_bytes
)) {
493 /* keep track of how many bytes this group covers */
494 total_bytes
+= subreq
->wb_bytes
;
495 } else if (WARN_ON_ONCE(subreq
->wb_offset
< head
->wb_offset
||
496 ((subreq
->wb_offset
+ subreq
->wb_bytes
) >
497 (head
->wb_offset
+ total_bytes
)))) {
498 nfs_page_group_unlock(head
);
499 spin_unlock(&inode
->i_lock
);
500 return ERR_PTR(-EIO
);
503 if (!nfs_lock_request(subreq
)) {
504 /* releases page group bit lock and
505 * inode spin lock and all references */
506 ret
= nfs_unroll_locks_and_wait(inode
, head
,
515 subreq
= subreq
->wb_this_page
;
516 } while (subreq
!= head
);
518 /* Now that all requests are locked, make sure they aren't on any list.
519 * Commit list removal accounting is done after locks are dropped */
522 nfs_clear_request_commit(subreq
);
523 subreq
= subreq
->wb_this_page
;
524 } while (subreq
!= head
);
526 /* unlink subrequests from head, destroy them later */
527 if (head
->wb_this_page
!= head
) {
528 /* destroy list will be terminated by head */
529 destroy_list
= head
->wb_this_page
;
530 head
->wb_this_page
= head
;
532 /* change head request to cover whole range that
533 * the former page group covered */
534 head
->wb_bytes
= total_bytes
;
538 * prepare head request to be added to new pgio descriptor
540 nfs_page_group_clear_bits(head
);
543 * some part of the group was still on the inode list - otherwise
544 * the group wouldn't be involved in async write.
545 * grab a reference for the head request, iff it needs one.
547 if (!test_and_set_bit(PG_INODE_REF
, &head
->wb_flags
))
548 kref_get(&head
->wb_kref
);
550 nfs_page_group_unlock(head
);
552 /* drop lock to clean uprequests on destroy list */
553 spin_unlock(&inode
->i_lock
);
555 nfs_destroy_unlinked_subrequests(destroy_list
, head
);
557 /* still holds ref on head from nfs_page_find_head_request_locked
558 * and still has lock on head from lock loop */
562 static void nfs_write_error_remove_page(struct nfs_page
*req
)
564 nfs_unlock_request(req
);
565 nfs_end_page_writeback(req
);
566 generic_error_remove_page(page_file_mapping(req
->wb_page
),
568 nfs_release_request(req
);
572 nfs_error_is_fatal_on_server(int err
)
580 return nfs_error_is_fatal(err
);
584 * Find an associated nfs write request, and prepare to flush it out
585 * May return an error if the user signalled nfs_wait_on_request().
587 static int nfs_page_async_flush(struct nfs_pageio_descriptor
*pgio
,
588 struct page
*page
, bool nonblock
)
590 struct nfs_page
*req
;
593 req
= nfs_lock_and_join_requests(page
, nonblock
);
600 nfs_set_page_writeback(page
);
601 WARN_ON_ONCE(test_bit(PG_CLEAN
, &req
->wb_flags
));
604 /* If there is a fatal error that covers this write, just exit */
605 if (nfs_error_is_fatal_on_server(req
->wb_context
->error
))
608 if (!nfs_pageio_add_request(pgio
, req
)) {
609 ret
= pgio
->pg_error
;
611 * Remove the problematic req upon fatal errors on the server
613 if (nfs_error_is_fatal(ret
)) {
614 nfs_context_set_write_error(req
->wb_context
, ret
);
615 if (nfs_error_is_fatal_on_server(ret
))
618 nfs_redirty_request(req
);
621 nfs_add_stats(page_file_mapping(page
)->host
,
622 NFSIOS_WRITEPAGES
, 1);
626 nfs_write_error_remove_page(req
);
630 static int nfs_do_writepage(struct page
*page
, struct writeback_control
*wbc
,
631 struct nfs_pageio_descriptor
*pgio
)
635 nfs_pageio_cond_complete(pgio
, page_index(page
));
636 ret
= nfs_page_async_flush(pgio
, page
, wbc
->sync_mode
== WB_SYNC_NONE
);
637 if (ret
== -EAGAIN
) {
638 redirty_page_for_writepage(wbc
, page
);
645 * Write an mmapped page to the server.
647 static int nfs_writepage_locked(struct page
*page
,
648 struct writeback_control
*wbc
)
650 struct nfs_pageio_descriptor pgio
;
651 struct inode
*inode
= page_file_mapping(page
)->host
;
654 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGE
);
655 nfs_pageio_init_write(&pgio
, inode
, 0,
656 false, &nfs_async_write_completion_ops
);
657 err
= nfs_do_writepage(page
, wbc
, &pgio
);
658 nfs_pageio_complete(&pgio
);
661 if (pgio
.pg_error
< 0)
662 return pgio
.pg_error
;
666 int nfs_writepage(struct page
*page
, struct writeback_control
*wbc
)
670 ret
= nfs_writepage_locked(page
, wbc
);
675 static int nfs_writepages_callback(struct page
*page
, struct writeback_control
*wbc
, void *data
)
679 ret
= nfs_do_writepage(page
, wbc
, data
);
684 int nfs_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
686 struct inode
*inode
= mapping
->host
;
687 struct nfs_pageio_descriptor pgio
;
690 nfs_inc_stats(inode
, NFSIOS_VFSWRITEPAGES
);
692 nfs_pageio_init_write(&pgio
, inode
, wb_priority(wbc
), false,
693 &nfs_async_write_completion_ops
);
694 err
= write_cache_pages(mapping
, wbc
, nfs_writepages_callback
, &pgio
);
695 nfs_pageio_complete(&pgio
);
708 * Insert a write request into an inode
710 static void nfs_inode_add_request(struct inode
*inode
, struct nfs_page
*req
)
712 struct nfs_inode
*nfsi
= NFS_I(inode
);
714 WARN_ON_ONCE(req
->wb_this_page
!= req
);
716 /* Lock the request! */
717 nfs_lock_request(req
);
719 spin_lock(&inode
->i_lock
);
720 if (!nfsi
->nrequests
&&
721 NFS_PROTO(inode
)->have_delegation(inode
, FMODE_WRITE
))
724 * Swap-space should not get truncated. Hence no need to plug the race
725 * with invalidate/truncate.
727 if (likely(!PageSwapCache(req
->wb_page
))) {
728 set_bit(PG_MAPPED
, &req
->wb_flags
);
729 SetPagePrivate(req
->wb_page
);
730 set_page_private(req
->wb_page
, (unsigned long)req
);
733 /* this a head request for a page group - mark it as having an
734 * extra reference so sub groups can follow suit.
735 * This flag also informs pgio layer when to bump nrequests when
736 * adding subrequests. */
737 WARN_ON(test_and_set_bit(PG_INODE_REF
, &req
->wb_flags
));
738 kref_get(&req
->wb_kref
);
739 spin_unlock(&inode
->i_lock
);
743 * Remove a write request from an inode
745 static void nfs_inode_remove_request(struct nfs_page
*req
)
747 struct inode
*inode
= d_inode(req
->wb_context
->dentry
);
748 struct nfs_inode
*nfsi
= NFS_I(inode
);
749 struct nfs_page
*head
;
751 if (nfs_page_group_sync_on_bit(req
, PG_REMOVE
)) {
754 spin_lock(&inode
->i_lock
);
755 if (likely(head
->wb_page
&& !PageSwapCache(head
->wb_page
))) {
756 set_page_private(head
->wb_page
, 0);
757 ClearPagePrivate(head
->wb_page
);
758 clear_bit(PG_MAPPED
, &head
->wb_flags
);
761 spin_unlock(&inode
->i_lock
);
763 spin_lock(&inode
->i_lock
);
765 spin_unlock(&inode
->i_lock
);
768 if (test_and_clear_bit(PG_INODE_REF
, &req
->wb_flags
))
769 nfs_release_request(req
);
773 nfs_mark_request_dirty(struct nfs_page
*req
)
776 __set_page_dirty_nobuffers(req
->wb_page
);
780 * nfs_page_search_commits_for_head_request_locked
782 * Search through commit lists on @inode for the head request for @page.
783 * Must be called while holding the inode (which is cinfo) lock.
785 * Returns the head request if found, or NULL if not found.
787 static struct nfs_page
*
788 nfs_page_search_commits_for_head_request_locked(struct nfs_inode
*nfsi
,
791 struct nfs_page
*freq
, *t
;
792 struct nfs_commit_info cinfo
;
793 struct inode
*inode
= &nfsi
->vfs_inode
;
795 nfs_init_cinfo_from_inode(&cinfo
, inode
);
797 /* search through pnfs commit lists */
798 freq
= pnfs_search_commit_reqs(inode
, &cinfo
, page
);
800 return freq
->wb_head
;
802 /* Linearly search the commit list for the correct request */
803 list_for_each_entry_safe(freq
, t
, &cinfo
.mds
->list
, wb_list
) {
804 if (freq
->wb_page
== page
)
805 return freq
->wb_head
;
812 * nfs_request_add_commit_list_locked - add request to a commit list
813 * @req: pointer to a struct nfs_page
814 * @dst: commit list head
815 * @cinfo: holds list lock and accounting info
817 * This sets the PG_CLEAN bit, updates the cinfo count of
818 * number of outstanding requests requiring a commit as well as
821 * The caller must hold cinfo->inode->i_lock, and the nfs_page lock.
824 nfs_request_add_commit_list_locked(struct nfs_page
*req
, struct list_head
*dst
,
825 struct nfs_commit_info
*cinfo
)
827 set_bit(PG_CLEAN
, &req
->wb_flags
);
828 nfs_list_add_request(req
, dst
);
829 cinfo
->mds
->ncommit
++;
831 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked
);
834 * nfs_request_add_commit_list - add request to a commit list
835 * @req: pointer to a struct nfs_page
836 * @dst: commit list head
837 * @cinfo: holds list lock and accounting info
839 * This sets the PG_CLEAN bit, updates the cinfo count of
840 * number of outstanding requests requiring a commit as well as
843 * The caller must _not_ hold the cinfo->lock, but must be
844 * holding the nfs_page lock.
847 nfs_request_add_commit_list(struct nfs_page
*req
, struct nfs_commit_info
*cinfo
)
849 spin_lock(&cinfo
->inode
->i_lock
);
850 nfs_request_add_commit_list_locked(req
, &cinfo
->mds
->list
, cinfo
);
851 spin_unlock(&cinfo
->inode
->i_lock
);
853 nfs_mark_page_unstable(req
->wb_page
, cinfo
);
855 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list
);
858 * nfs_request_remove_commit_list - Remove request from a commit list
859 * @req: pointer to a nfs_page
860 * @cinfo: holds list lock and accounting info
862 * This clears the PG_CLEAN bit, and updates the cinfo's count of
863 * number of outstanding requests requiring a commit
864 * It does not update the MM page stats.
866 * The caller _must_ hold the cinfo->lock and the nfs_page lock.
869 nfs_request_remove_commit_list(struct nfs_page
*req
,
870 struct nfs_commit_info
*cinfo
)
872 if (!test_and_clear_bit(PG_CLEAN
, &(req
)->wb_flags
))
874 nfs_list_remove_request(req
);
875 cinfo
->mds
->ncommit
--;
877 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list
);
879 static void nfs_init_cinfo_from_inode(struct nfs_commit_info
*cinfo
,
882 cinfo
->inode
= inode
;
883 cinfo
->mds
= &NFS_I(inode
)->commit_info
;
884 cinfo
->ds
= pnfs_get_ds_info(inode
);
886 cinfo
->completion_ops
= &nfs_commit_completion_ops
;
889 void nfs_init_cinfo(struct nfs_commit_info
*cinfo
,
891 struct nfs_direct_req
*dreq
)
894 nfs_init_cinfo_from_dreq(cinfo
, dreq
);
896 nfs_init_cinfo_from_inode(cinfo
, inode
);
898 EXPORT_SYMBOL_GPL(nfs_init_cinfo
);
901 * Add a request to the inode's commit list.
904 nfs_mark_request_commit(struct nfs_page
*req
, struct pnfs_layout_segment
*lseg
,
905 struct nfs_commit_info
*cinfo
, u32 ds_commit_idx
)
907 if (pnfs_mark_request_commit(req
, lseg
, cinfo
, ds_commit_idx
))
909 nfs_request_add_commit_list(req
, cinfo
);
913 nfs_clear_page_commit(struct page
*page
)
915 dec_node_page_state(page
, NR_UNSTABLE_NFS
);
916 dec_wb_stat(&inode_to_bdi(page_file_mapping(page
)->host
)->wb
,
920 /* Called holding inode (/cinfo) lock */
922 nfs_clear_request_commit(struct nfs_page
*req
)
924 if (test_bit(PG_CLEAN
, &req
->wb_flags
)) {
925 struct inode
*inode
= d_inode(req
->wb_context
->dentry
);
926 struct nfs_commit_info cinfo
;
928 nfs_init_cinfo_from_inode(&cinfo
, inode
);
929 if (!pnfs_clear_request_commit(req
, &cinfo
)) {
930 nfs_request_remove_commit_list(req
, &cinfo
);
932 nfs_clear_page_commit(req
->wb_page
);
936 int nfs_write_need_commit(struct nfs_pgio_header
*hdr
)
938 if (hdr
->verf
.committed
== NFS_DATA_SYNC
)
939 return hdr
->lseg
== NULL
;
940 return hdr
->verf
.committed
!= NFS_FILE_SYNC
;
943 static void nfs_write_completion(struct nfs_pgio_header
*hdr
)
945 struct nfs_commit_info cinfo
;
946 unsigned long bytes
= 0;
948 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
950 nfs_init_cinfo_from_inode(&cinfo
, hdr
->inode
);
951 while (!list_empty(&hdr
->pages
)) {
952 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
954 bytes
+= req
->wb_bytes
;
955 nfs_list_remove_request(req
);
956 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
) &&
957 (hdr
->good_bytes
< bytes
)) {
958 nfs_set_pageerror(req
->wb_page
);
959 nfs_context_set_write_error(req
->wb_context
, hdr
->error
);
962 if (nfs_write_need_commit(hdr
)) {
963 memcpy(&req
->wb_verf
, &hdr
->verf
.verifier
, sizeof(req
->wb_verf
));
964 nfs_mark_request_commit(req
, hdr
->lseg
, &cinfo
,
965 hdr
->pgio_mirror_idx
);
969 nfs_inode_remove_request(req
);
971 nfs_unlock_request(req
);
972 nfs_end_page_writeback(req
);
973 nfs_release_request(req
);
980 nfs_reqs_to_commit(struct nfs_commit_info
*cinfo
)
982 return cinfo
->mds
->ncommit
;
985 /* cinfo->inode->i_lock held by caller */
987 nfs_scan_commit_list(struct list_head
*src
, struct list_head
*dst
,
988 struct nfs_commit_info
*cinfo
, int max
)
990 struct nfs_page
*req
, *tmp
;
993 list_for_each_entry_safe(req
, tmp
, src
, wb_list
) {
994 if (!nfs_lock_request(req
))
996 kref_get(&req
->wb_kref
);
997 if (cond_resched_lock(&cinfo
->inode
->i_lock
))
998 list_safe_reset_next(req
, tmp
, wb_list
);
999 nfs_request_remove_commit_list(req
, cinfo
);
1000 nfs_list_add_request(req
, dst
);
1002 if ((ret
== max
) && !cinfo
->dreq
)
1009 * nfs_scan_commit - Scan an inode for commit requests
1010 * @inode: NFS inode to scan
1011 * @dst: mds destination list
1012 * @cinfo: mds and ds lists of reqs ready to commit
1014 * Moves requests from the inode's 'commit' request list.
1015 * The requests are *not* checked to ensure that they form a contiguous set.
1018 nfs_scan_commit(struct inode
*inode
, struct list_head
*dst
,
1019 struct nfs_commit_info
*cinfo
)
1023 spin_lock(&cinfo
->inode
->i_lock
);
1024 if (cinfo
->mds
->ncommit
> 0) {
1025 const int max
= INT_MAX
;
1027 ret
= nfs_scan_commit_list(&cinfo
->mds
->list
, dst
,
1029 ret
+= pnfs_scan_commit_lists(inode
, cinfo
, max
- ret
);
1031 spin_unlock(&cinfo
->inode
->i_lock
);
1036 * Search for an existing write request, and attempt to update
1037 * it to reflect a new dirty region on a given page.
1039 * If the attempt fails, then the existing request is flushed out
1042 static struct nfs_page
*nfs_try_to_update_request(struct inode
*inode
,
1044 unsigned int offset
,
1047 struct nfs_page
*req
;
1052 if (!PagePrivate(page
))
1055 end
= offset
+ bytes
;
1056 spin_lock(&inode
->i_lock
);
1059 req
= nfs_page_find_head_request_locked(NFS_I(inode
), page
);
1063 /* should be handled by nfs_flush_incompatible */
1064 WARN_ON_ONCE(req
->wb_head
!= req
);
1065 WARN_ON_ONCE(req
->wb_this_page
!= req
);
1067 rqend
= req
->wb_offset
+ req
->wb_bytes
;
1069 * Tell the caller to flush out the request if
1070 * the offsets are non-contiguous.
1071 * Note: nfs_flush_incompatible() will already
1072 * have flushed out requests having wrong owners.
1075 || end
< req
->wb_offset
)
1078 if (nfs_lock_request(req
))
1081 /* The request is locked, so wait and then retry */
1082 spin_unlock(&inode
->i_lock
);
1083 error
= nfs_wait_on_request(req
);
1084 nfs_release_request(req
);
1087 spin_lock(&inode
->i_lock
);
1090 /* Okay, the request matches. Update the region */
1091 if (offset
< req
->wb_offset
) {
1092 req
->wb_offset
= offset
;
1093 req
->wb_pgbase
= offset
;
1096 req
->wb_bytes
= end
- req
->wb_offset
;
1098 req
->wb_bytes
= rqend
- req
->wb_offset
;
1101 nfs_clear_request_commit(req
);
1102 spin_unlock(&inode
->i_lock
);
1105 spin_unlock(&inode
->i_lock
);
1106 nfs_release_request(req
);
1107 error
= nfs_wb_page(inode
, page
);
1109 return ERR_PTR(error
);
1113 * Try to update an existing write request, or create one if there is none.
1115 * Note: Should always be called with the Page Lock held to prevent races
1116 * if we have to add a new request. Also assumes that the caller has
1117 * already called nfs_flush_incompatible() if necessary.
1119 static struct nfs_page
* nfs_setup_write_request(struct nfs_open_context
* ctx
,
1120 struct page
*page
, unsigned int offset
, unsigned int bytes
)
1122 struct inode
*inode
= page_file_mapping(page
)->host
;
1123 struct nfs_page
*req
;
1125 req
= nfs_try_to_update_request(inode
, page
, offset
, bytes
);
1128 req
= nfs_create_request(ctx
, page
, NULL
, offset
, bytes
);
1131 nfs_inode_add_request(inode
, req
);
1136 static int nfs_writepage_setup(struct nfs_open_context
*ctx
, struct page
*page
,
1137 unsigned int offset
, unsigned int count
)
1139 struct nfs_page
*req
;
1141 req
= nfs_setup_write_request(ctx
, page
, offset
, count
);
1143 return PTR_ERR(req
);
1144 /* Update file length */
1145 nfs_grow_file(page
, offset
, count
);
1146 nfs_mark_uptodate(req
);
1147 nfs_mark_request_dirty(req
);
1148 nfs_unlock_and_release_request(req
);
1152 int nfs_flush_incompatible(struct file
*file
, struct page
*page
)
1154 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
1155 struct nfs_lock_context
*l_ctx
;
1156 struct file_lock_context
*flctx
= file_inode(file
)->i_flctx
;
1157 struct nfs_page
*req
;
1158 int do_flush
, status
;
1160 * Look for a request corresponding to this page. If there
1161 * is one, and it belongs to another file, we flush it out
1162 * before we try to copy anything into the page. Do this
1163 * due to the lack of an ACCESS-type call in NFSv2.
1164 * Also do the same if we find a request from an existing
1168 req
= nfs_page_find_head_request(page
);
1171 l_ctx
= req
->wb_lock_context
;
1172 do_flush
= req
->wb_page
!= page
||
1173 !nfs_match_open_context(req
->wb_context
, ctx
);
1174 /* for now, flush if more than 1 request in page_group */
1175 do_flush
|= req
->wb_this_page
!= req
;
1176 if (l_ctx
&& flctx
&&
1177 !(list_empty_careful(&flctx
->flc_posix
) &&
1178 list_empty_careful(&flctx
->flc_flock
))) {
1179 do_flush
|= l_ctx
->lockowner
!= current
->files
;
1181 nfs_release_request(req
);
1184 status
= nfs_wb_page(page_file_mapping(page
)->host
, page
);
1185 } while (status
== 0);
1190 * Avoid buffered writes when a open context credential's key would
1193 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1195 * Return 0 and set a credential flag which triggers the inode to flush
1196 * and performs NFS_FILE_SYNC writes if the key will expired within
1197 * RPC_KEY_EXPIRE_TIMEO.
1200 nfs_key_timeout_notify(struct file
*filp
, struct inode
*inode
)
1202 struct nfs_open_context
*ctx
= nfs_file_open_context(filp
);
1203 struct rpc_auth
*auth
= NFS_SERVER(inode
)->client
->cl_auth
;
1205 return rpcauth_key_timeout_notify(auth
, ctx
->cred
);
1209 * Test if the open context credential key is marked to expire soon.
1211 bool nfs_ctx_key_to_expire(struct nfs_open_context
*ctx
, struct inode
*inode
)
1213 struct rpc_auth
*auth
= NFS_SERVER(inode
)->client
->cl_auth
;
1215 return rpcauth_cred_key_to_expire(auth
, ctx
->cred
);
1219 * If the page cache is marked as unsafe or invalid, then we can't rely on
1220 * the PageUptodate() flag. In this case, we will need to turn off
1221 * write optimisations that depend on the page contents being correct.
1223 static bool nfs_write_pageuptodate(struct page
*page
, struct inode
*inode
)
1225 struct nfs_inode
*nfsi
= NFS_I(inode
);
1227 if (nfs_have_delegated_attributes(inode
))
1229 if (nfsi
->cache_validity
& NFS_INO_REVAL_PAGECACHE
)
1232 if (test_bit(NFS_INO_INVALIDATING
, &nfsi
->flags
))
1235 if (nfsi
->cache_validity
& NFS_INO_INVALID_DATA
)
1237 return PageUptodate(page
) != 0;
1241 is_whole_file_wrlock(struct file_lock
*fl
)
1243 return fl
->fl_start
== 0 && fl
->fl_end
== OFFSET_MAX
&&
1244 fl
->fl_type
== F_WRLCK
;
1247 /* If we know the page is up to date, and we're not using byte range locks (or
1248 * if we have the whole file locked for writing), it may be more efficient to
1249 * extend the write to cover the entire page in order to avoid fragmentation
1252 * If the file is opened for synchronous writes then we can just skip the rest
1255 static int nfs_can_extend_write(struct file
*file
, struct page
*page
, struct inode
*inode
)
1258 struct file_lock_context
*flctx
= inode
->i_flctx
;
1259 struct file_lock
*fl
;
1261 if (file
->f_flags
& O_DSYNC
)
1263 if (!nfs_write_pageuptodate(page
, inode
))
1265 if (NFS_PROTO(inode
)->have_delegation(inode
, FMODE_WRITE
))
1267 if (!flctx
|| (list_empty_careful(&flctx
->flc_flock
) &&
1268 list_empty_careful(&flctx
->flc_posix
)))
1271 /* Check to see if there are whole file write locks */
1273 spin_lock(&flctx
->flc_lock
);
1274 if (!list_empty(&flctx
->flc_posix
)) {
1275 fl
= list_first_entry(&flctx
->flc_posix
, struct file_lock
,
1277 if (is_whole_file_wrlock(fl
))
1279 } else if (!list_empty(&flctx
->flc_flock
)) {
1280 fl
= list_first_entry(&flctx
->flc_flock
, struct file_lock
,
1282 if (fl
->fl_type
== F_WRLCK
)
1285 spin_unlock(&flctx
->flc_lock
);
1290 * Update and possibly write a cached page of an NFS file.
1292 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1293 * things with a page scheduled for an RPC call (e.g. invalidate it).
1295 int nfs_updatepage(struct file
*file
, struct page
*page
,
1296 unsigned int offset
, unsigned int count
)
1298 struct nfs_open_context
*ctx
= nfs_file_open_context(file
);
1299 struct inode
*inode
= page_file_mapping(page
)->host
;
1302 nfs_inc_stats(inode
, NFSIOS_VFSUPDATEPAGE
);
1304 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n",
1305 file
, count
, (long long)(page_file_offset(page
) + offset
));
1310 if (nfs_can_extend_write(file
, page
, inode
)) {
1311 count
= max(count
+ offset
, nfs_page_length(page
));
1315 status
= nfs_writepage_setup(ctx
, page
, offset
, count
);
1317 nfs_set_pageerror(page
);
1319 __set_page_dirty_nobuffers(page
);
1321 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n",
1322 status
, (long long)i_size_read(inode
));
1326 static int flush_task_priority(int how
)
1328 switch (how
& (FLUSH_HIGHPRI
|FLUSH_LOWPRI
)) {
1330 return RPC_PRIORITY_HIGH
;
1332 return RPC_PRIORITY_LOW
;
1334 return RPC_PRIORITY_NORMAL
;
1337 static void nfs_initiate_write(struct nfs_pgio_header
*hdr
,
1338 struct rpc_message
*msg
,
1339 const struct nfs_rpc_ops
*rpc_ops
,
1340 struct rpc_task_setup
*task_setup_data
, int how
)
1342 int priority
= flush_task_priority(how
);
1344 task_setup_data
->priority
= priority
;
1345 rpc_ops
->write_setup(hdr
, msg
);
1347 nfs4_state_protect_write(NFS_SERVER(hdr
->inode
)->nfs_client
,
1348 &task_setup_data
->rpc_client
, msg
, hdr
);
1351 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1352 * call this on each, which will prepare them to be retried on next
1353 * writeback using standard nfs.
1355 static void nfs_redirty_request(struct nfs_page
*req
)
1357 nfs_mark_request_dirty(req
);
1358 set_bit(NFS_CONTEXT_RESEND_WRITES
, &req
->wb_context
->flags
);
1359 nfs_unlock_request(req
);
1360 nfs_end_page_writeback(req
);
1361 nfs_release_request(req
);
1364 static void nfs_async_write_error(struct list_head
*head
)
1366 struct nfs_page
*req
;
1368 while (!list_empty(head
)) {
1369 req
= nfs_list_entry(head
->next
);
1370 nfs_list_remove_request(req
);
1371 nfs_redirty_request(req
);
1375 static void nfs_async_write_reschedule_io(struct nfs_pgio_header
*hdr
)
1377 nfs_async_write_error(&hdr
->pages
);
1380 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops
= {
1381 .error_cleanup
= nfs_async_write_error
,
1382 .completion
= nfs_write_completion
,
1383 .reschedule_io
= nfs_async_write_reschedule_io
,
1386 void nfs_pageio_init_write(struct nfs_pageio_descriptor
*pgio
,
1387 struct inode
*inode
, int ioflags
, bool force_mds
,
1388 const struct nfs_pgio_completion_ops
*compl_ops
)
1390 struct nfs_server
*server
= NFS_SERVER(inode
);
1391 const struct nfs_pageio_ops
*pg_ops
= &nfs_pgio_rw_ops
;
1393 #ifdef CONFIG_NFS_V4_1
1394 if (server
->pnfs_curr_ld
&& !force_mds
)
1395 pg_ops
= server
->pnfs_curr_ld
->pg_write_ops
;
1397 nfs_pageio_init(pgio
, inode
, pg_ops
, compl_ops
, &nfs_rw_write_ops
,
1398 server
->wsize
, ioflags
, GFP_NOIO
);
1400 EXPORT_SYMBOL_GPL(nfs_pageio_init_write
);
1402 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor
*pgio
)
1404 struct nfs_pgio_mirror
*mirror
;
1406 if (pgio
->pg_ops
&& pgio
->pg_ops
->pg_cleanup
)
1407 pgio
->pg_ops
->pg_cleanup(pgio
);
1409 pgio
->pg_ops
= &nfs_pgio_rw_ops
;
1411 nfs_pageio_stop_mirroring(pgio
);
1413 mirror
= &pgio
->pg_mirrors
[0];
1414 mirror
->pg_bsize
= NFS_SERVER(pgio
->pg_inode
)->wsize
;
1416 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds
);
1419 void nfs_commit_prepare(struct rpc_task
*task
, void *calldata
)
1421 struct nfs_commit_data
*data
= calldata
;
1423 NFS_PROTO(data
->inode
)->commit_rpc_prepare(task
, data
);
1427 * Special version of should_remove_suid() that ignores capabilities.
1429 static int nfs_should_remove_suid(const struct inode
*inode
)
1431 umode_t mode
= inode
->i_mode
;
1434 /* suid always must be killed */
1435 if (unlikely(mode
& S_ISUID
))
1436 kill
= ATTR_KILL_SUID
;
1439 * sgid without any exec bits is just a mandatory locking mark; leave
1440 * it alone. If some exec bits are set, it's a real sgid; kill it.
1442 if (unlikely((mode
& S_ISGID
) && (mode
& S_IXGRP
)))
1443 kill
|= ATTR_KILL_SGID
;
1445 if (unlikely(kill
&& S_ISREG(mode
)))
1451 static void nfs_writeback_check_extend(struct nfs_pgio_header
*hdr
,
1452 struct nfs_fattr
*fattr
)
1454 struct nfs_pgio_args
*argp
= &hdr
->args
;
1455 struct nfs_pgio_res
*resp
= &hdr
->res
;
1456 u64 size
= argp
->offset
+ resp
->count
;
1458 if (!(fattr
->valid
& NFS_ATTR_FATTR_SIZE
))
1460 if (nfs_size_to_loff_t(fattr
->size
) < i_size_read(hdr
->inode
)) {
1461 fattr
->valid
&= ~NFS_ATTR_FATTR_SIZE
;
1464 if (size
!= fattr
->size
)
1466 /* Set attribute barrier */
1467 nfs_fattr_set_barrier(fattr
);
1468 /* ...and update size */
1469 fattr
->valid
|= NFS_ATTR_FATTR_SIZE
;
1472 void nfs_writeback_update_inode(struct nfs_pgio_header
*hdr
)
1474 struct nfs_fattr
*fattr
= &hdr
->fattr
;
1475 struct inode
*inode
= hdr
->inode
;
1477 spin_lock(&inode
->i_lock
);
1478 nfs_writeback_check_extend(hdr
, fattr
);
1479 nfs_post_op_update_inode_force_wcc_locked(inode
, fattr
);
1480 spin_unlock(&inode
->i_lock
);
1482 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode
);
1485 * This function is called when the WRITE call is complete.
1487 static int nfs_writeback_done(struct rpc_task
*task
,
1488 struct nfs_pgio_header
*hdr
,
1489 struct inode
*inode
)
1494 * ->write_done will attempt to use post-op attributes to detect
1495 * conflicting writes by other clients. A strict interpretation
1496 * of close-to-open would allow us to continue caching even if
1497 * another writer had changed the file, but some applications
1498 * depend on tighter cache coherency when writing.
1500 status
= NFS_PROTO(inode
)->write_done(task
, hdr
);
1503 nfs_add_stats(inode
, NFSIOS_SERVERWRITTENBYTES
, hdr
->res
.count
);
1505 if (hdr
->res
.verf
->committed
< hdr
->args
.stable
&&
1506 task
->tk_status
>= 0) {
1507 /* We tried a write call, but the server did not
1508 * commit data to stable storage even though we
1510 * Note: There is a known bug in Tru64 < 5.0 in which
1511 * the server reports NFS_DATA_SYNC, but performs
1512 * NFS_FILE_SYNC. We therefore implement this checking
1513 * as a dprintk() in order to avoid filling syslog.
1515 static unsigned long complain
;
1517 /* Note this will print the MDS for a DS write */
1518 if (time_before(complain
, jiffies
)) {
1519 dprintk("NFS: faulty NFS server %s:"
1520 " (committed = %d) != (stable = %d)\n",
1521 NFS_SERVER(inode
)->nfs_client
->cl_hostname
,
1522 hdr
->res
.verf
->committed
, hdr
->args
.stable
);
1523 complain
= jiffies
+ 300 * HZ
;
1527 /* Deal with the suid/sgid bit corner case */
1528 if (nfs_should_remove_suid(inode
))
1529 nfs_mark_for_revalidate(inode
);
1534 * This function is called when the WRITE call is complete.
1536 static void nfs_writeback_result(struct rpc_task
*task
,
1537 struct nfs_pgio_header
*hdr
)
1539 struct nfs_pgio_args
*argp
= &hdr
->args
;
1540 struct nfs_pgio_res
*resp
= &hdr
->res
;
1542 if (resp
->count
< argp
->count
) {
1543 static unsigned long complain
;
1545 /* This a short write! */
1546 nfs_inc_stats(hdr
->inode
, NFSIOS_SHORTWRITE
);
1548 /* Has the server at least made some progress? */
1549 if (resp
->count
== 0) {
1550 if (time_before(complain
, jiffies
)) {
1552 "NFS: Server wrote zero bytes, expected %u.\n",
1554 complain
= jiffies
+ 300 * HZ
;
1556 nfs_set_pgio_error(hdr
, -EIO
, argp
->offset
);
1557 task
->tk_status
= -EIO
;
1561 /* For non rpc-based layout drivers, retry-through-MDS */
1562 if (!task
->tk_ops
) {
1563 hdr
->pnfs_error
= -EAGAIN
;
1567 /* Was this an NFSv2 write or an NFSv3 stable write? */
1568 if (resp
->verf
->committed
!= NFS_UNSTABLE
) {
1569 /* Resend from where the server left off */
1570 hdr
->mds_offset
+= resp
->count
;
1571 argp
->offset
+= resp
->count
;
1572 argp
->pgbase
+= resp
->count
;
1573 argp
->count
-= resp
->count
;
1575 /* Resend as a stable write in order to avoid
1576 * headaches in the case of a server crash.
1578 argp
->stable
= NFS_FILE_SYNC
;
1580 rpc_restart_call_prepare(task
);
1584 static int wait_on_commit(struct nfs_mds_commit_info
*cinfo
)
1586 return wait_on_atomic_t(&cinfo
->rpcs_out
,
1587 nfs_wait_atomic_killable
, TASK_KILLABLE
);
1590 static void nfs_commit_begin(struct nfs_mds_commit_info
*cinfo
)
1592 atomic_inc(&cinfo
->rpcs_out
);
1595 static void nfs_commit_end(struct nfs_mds_commit_info
*cinfo
)
1597 if (atomic_dec_and_test(&cinfo
->rpcs_out
))
1598 wake_up_atomic_t(&cinfo
->rpcs_out
);
1601 void nfs_commitdata_release(struct nfs_commit_data
*data
)
1603 put_nfs_open_context(data
->context
);
1604 nfs_commit_free(data
);
1606 EXPORT_SYMBOL_GPL(nfs_commitdata_release
);
1608 int nfs_initiate_commit(struct rpc_clnt
*clnt
, struct nfs_commit_data
*data
,
1609 const struct nfs_rpc_ops
*nfs_ops
,
1610 const struct rpc_call_ops
*call_ops
,
1613 struct rpc_task
*task
;
1614 int priority
= flush_task_priority(how
);
1615 struct rpc_message msg
= {
1616 .rpc_argp
= &data
->args
,
1617 .rpc_resp
= &data
->res
,
1618 .rpc_cred
= data
->cred
,
1620 struct rpc_task_setup task_setup_data
= {
1621 .task
= &data
->task
,
1623 .rpc_message
= &msg
,
1624 .callback_ops
= call_ops
,
1625 .callback_data
= data
,
1626 .workqueue
= nfsiod_workqueue
,
1627 .flags
= RPC_TASK_ASYNC
| flags
,
1628 .priority
= priority
,
1630 /* Set up the initial task struct. */
1631 nfs_ops
->commit_setup(data
, &msg
);
1633 dprintk("NFS: initiated commit call\n");
1635 nfs4_state_protect(NFS_SERVER(data
->inode
)->nfs_client
,
1636 NFS_SP4_MACH_CRED_COMMIT
, &task_setup_data
.rpc_client
, &msg
);
1638 task
= rpc_run_task(&task_setup_data
);
1640 return PTR_ERR(task
);
1641 if (how
& FLUSH_SYNC
)
1642 rpc_wait_for_completion_task(task
);
1646 EXPORT_SYMBOL_GPL(nfs_initiate_commit
);
1648 static loff_t
nfs_get_lwb(struct list_head
*head
)
1651 struct nfs_page
*req
;
1653 list_for_each_entry(req
, head
, wb_list
)
1654 if (lwb
< (req_offset(req
) + req
->wb_bytes
))
1655 lwb
= req_offset(req
) + req
->wb_bytes
;
1661 * Set up the argument/result storage required for the RPC call.
1663 void nfs_init_commit(struct nfs_commit_data
*data
,
1664 struct list_head
*head
,
1665 struct pnfs_layout_segment
*lseg
,
1666 struct nfs_commit_info
*cinfo
)
1668 struct nfs_page
*first
= nfs_list_entry(head
->next
);
1669 struct inode
*inode
= d_inode(first
->wb_context
->dentry
);
1671 /* Set up the RPC argument and reply structs
1672 * NB: take care not to mess about with data->commit et al. */
1674 list_splice_init(head
, &data
->pages
);
1676 data
->inode
= inode
;
1677 data
->cred
= first
->wb_context
->cred
;
1678 data
->lseg
= lseg
; /* reference transferred */
1679 /* only set lwb for pnfs commit */
1681 data
->lwb
= nfs_get_lwb(&data
->pages
);
1682 data
->mds_ops
= &nfs_commit_ops
;
1683 data
->completion_ops
= cinfo
->completion_ops
;
1684 data
->dreq
= cinfo
->dreq
;
1686 data
->args
.fh
= NFS_FH(data
->inode
);
1687 /* Note: we always request a commit of the entire inode */
1688 data
->args
.offset
= 0;
1689 data
->args
.count
= 0;
1690 data
->context
= get_nfs_open_context(first
->wb_context
);
1691 data
->res
.fattr
= &data
->fattr
;
1692 data
->res
.verf
= &data
->verf
;
1693 nfs_fattr_init(&data
->fattr
);
1695 EXPORT_SYMBOL_GPL(nfs_init_commit
);
1697 void nfs_retry_commit(struct list_head
*page_list
,
1698 struct pnfs_layout_segment
*lseg
,
1699 struct nfs_commit_info
*cinfo
,
1702 struct nfs_page
*req
;
1704 while (!list_empty(page_list
)) {
1705 req
= nfs_list_entry(page_list
->next
);
1706 nfs_list_remove_request(req
);
1707 nfs_mark_request_commit(req
, lseg
, cinfo
, ds_commit_idx
);
1709 nfs_clear_page_commit(req
->wb_page
);
1710 nfs_unlock_and_release_request(req
);
1713 EXPORT_SYMBOL_GPL(nfs_retry_commit
);
1716 nfs_commit_resched_write(struct nfs_commit_info
*cinfo
,
1717 struct nfs_page
*req
)
1719 __set_page_dirty_nobuffers(req
->wb_page
);
1723 * Commit dirty pages
1726 nfs_commit_list(struct inode
*inode
, struct list_head
*head
, int how
,
1727 struct nfs_commit_info
*cinfo
)
1729 struct nfs_commit_data
*data
;
1731 /* another commit raced with us */
1732 if (list_empty(head
))
1735 data
= nfs_commitdata_alloc(true);
1737 /* Set up the argument struct */
1738 nfs_init_commit(data
, head
, NULL
, cinfo
);
1739 atomic_inc(&cinfo
->mds
->rpcs_out
);
1740 return nfs_initiate_commit(NFS_CLIENT(inode
), data
, NFS_PROTO(inode
),
1741 data
->mds_ops
, how
, 0);
1745 * COMMIT call returned
1747 static void nfs_commit_done(struct rpc_task
*task
, void *calldata
)
1749 struct nfs_commit_data
*data
= calldata
;
1751 dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1752 task
->tk_pid
, task
->tk_status
);
1754 /* Call the NFS version-specific code */
1755 NFS_PROTO(data
->inode
)->commit_done(task
, data
);
1758 static void nfs_commit_release_pages(struct nfs_commit_data
*data
)
1760 struct nfs_page
*req
;
1761 int status
= data
->task
.tk_status
;
1762 struct nfs_commit_info cinfo
;
1763 struct nfs_server
*nfss
;
1765 while (!list_empty(&data
->pages
)) {
1766 req
= nfs_list_entry(data
->pages
.next
);
1767 nfs_list_remove_request(req
);
1769 nfs_clear_page_commit(req
->wb_page
);
1771 dprintk("NFS: commit (%s/%llu %d@%lld)",
1772 req
->wb_context
->dentry
->d_sb
->s_id
,
1773 (unsigned long long)NFS_FILEID(d_inode(req
->wb_context
->dentry
)),
1775 (long long)req_offset(req
));
1777 nfs_context_set_write_error(req
->wb_context
, status
);
1779 nfs_inode_remove_request(req
);
1780 dprintk_cont(", error = %d\n", status
);
1784 /* Okay, COMMIT succeeded, apparently. Check the verifier
1785 * returned by the server against all stored verfs. */
1786 if (!nfs_write_verifier_cmp(&req
->wb_verf
, &data
->verf
.verifier
)) {
1787 /* We have a match */
1789 nfs_inode_remove_request(req
);
1790 dprintk_cont(" OK\n");
1793 /* We have a mismatch. Write the page again */
1794 dprintk_cont(" mismatch\n");
1795 nfs_mark_request_dirty(req
);
1796 set_bit(NFS_CONTEXT_RESEND_WRITES
, &req
->wb_context
->flags
);
1798 nfs_unlock_and_release_request(req
);
1800 nfss
= NFS_SERVER(data
->inode
);
1801 if (atomic_long_read(&nfss
->writeback
) < NFS_CONGESTION_OFF_THRESH
)
1802 clear_bdi_congested(inode_to_bdi(data
->inode
), BLK_RW_ASYNC
);
1804 nfs_init_cinfo(&cinfo
, data
->inode
, data
->dreq
);
1805 nfs_commit_end(cinfo
.mds
);
1808 static void nfs_commit_release(void *calldata
)
1810 struct nfs_commit_data
*data
= calldata
;
1812 data
->completion_ops
->completion(data
);
1813 nfs_commitdata_release(calldata
);
1816 static const struct rpc_call_ops nfs_commit_ops
= {
1817 .rpc_call_prepare
= nfs_commit_prepare
,
1818 .rpc_call_done
= nfs_commit_done
,
1819 .rpc_release
= nfs_commit_release
,
1822 static const struct nfs_commit_completion_ops nfs_commit_completion_ops
= {
1823 .completion
= nfs_commit_release_pages
,
1824 .resched_write
= nfs_commit_resched_write
,
1827 int nfs_generic_commit_list(struct inode
*inode
, struct list_head
*head
,
1828 int how
, struct nfs_commit_info
*cinfo
)
1832 status
= pnfs_commit_list(inode
, head
, how
, cinfo
);
1833 if (status
== PNFS_NOT_ATTEMPTED
)
1834 status
= nfs_commit_list(inode
, head
, how
, cinfo
);
1838 int nfs_commit_inode(struct inode
*inode
, int how
)
1841 struct nfs_commit_info cinfo
;
1842 int may_wait
= how
& FLUSH_SYNC
;
1846 nfs_init_cinfo_from_inode(&cinfo
, inode
);
1847 nfs_commit_begin(cinfo
.mds
);
1848 res
= nfs_scan_commit(inode
, &head
, &cinfo
);
1850 error
= nfs_generic_commit_list(inode
, &head
, how
, &cinfo
);
1851 nfs_commit_end(cinfo
.mds
);
1855 goto out_mark_dirty
;
1856 error
= wait_on_commit(cinfo
.mds
);
1862 /* Note: If we exit without ensuring that the commit is complete,
1863 * we must mark the inode as dirty. Otherwise, future calls to
1864 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1865 * that the data is on the disk.
1868 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
1871 EXPORT_SYMBOL_GPL(nfs_commit_inode
);
1873 int nfs_write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1875 struct nfs_inode
*nfsi
= NFS_I(inode
);
1876 int flags
= FLUSH_SYNC
;
1879 /* no commits means nothing needs to be done */
1880 if (!nfsi
->commit_info
.ncommit
)
1883 if (wbc
->sync_mode
== WB_SYNC_NONE
) {
1884 /* Don't commit yet if this is a non-blocking flush and there
1885 * are a lot of outstanding writes for this mapping.
1887 if (nfsi
->commit_info
.ncommit
<= (nfsi
->nrequests
>> 1))
1888 goto out_mark_dirty
;
1890 /* don't wait for the COMMIT response */
1894 ret
= nfs_commit_inode(inode
, flags
);
1896 if (wbc
->sync_mode
== WB_SYNC_NONE
) {
1897 if (ret
< wbc
->nr_to_write
)
1898 wbc
->nr_to_write
-= ret
;
1900 wbc
->nr_to_write
= 0;
1905 __mark_inode_dirty(inode
, I_DIRTY_DATASYNC
);
1908 EXPORT_SYMBOL_GPL(nfs_write_inode
);
1911 * Wrapper for filemap_write_and_wait_range()
1913 * Needed for pNFS in order to ensure data becomes visible to the
1916 int nfs_filemap_write_and_wait_range(struct address_space
*mapping
,
1917 loff_t lstart
, loff_t lend
)
1921 ret
= filemap_write_and_wait_range(mapping
, lstart
, lend
);
1923 ret
= pnfs_sync_inode(mapping
->host
, true);
1926 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range
);
1929 * flush the inode to disk.
1931 int nfs_wb_all(struct inode
*inode
)
1935 trace_nfs_writeback_inode_enter(inode
);
1937 ret
= filemap_write_and_wait(inode
->i_mapping
);
1940 ret
= nfs_commit_inode(inode
, FLUSH_SYNC
);
1943 pnfs_sync_inode(inode
, true);
1947 trace_nfs_writeback_inode_exit(inode
, ret
);
1950 EXPORT_SYMBOL_GPL(nfs_wb_all
);
1952 int nfs_wb_page_cancel(struct inode
*inode
, struct page
*page
)
1954 struct nfs_page
*req
;
1957 wait_on_page_writeback(page
);
1959 /* blocking call to cancel all requests and join to a single (head)
1961 req
= nfs_lock_and_join_requests(page
, false);
1966 /* all requests from this page have been cancelled by
1967 * nfs_lock_and_join_requests, so just remove the head
1968 * request from the inode / page_private pointer and
1970 nfs_inode_remove_request(req
);
1971 nfs_unlock_and_release_request(req
);
1978 * Write back all requests on one page - we do this before reading it.
1980 int nfs_wb_page(struct inode
*inode
, struct page
*page
)
1982 loff_t range_start
= page_file_offset(page
);
1983 loff_t range_end
= range_start
+ (loff_t
)(PAGE_SIZE
- 1);
1984 struct writeback_control wbc
= {
1985 .sync_mode
= WB_SYNC_ALL
,
1987 .range_start
= range_start
,
1988 .range_end
= range_end
,
1992 trace_nfs_writeback_page_enter(inode
);
1995 wait_on_page_writeback(page
);
1996 if (clear_page_dirty_for_io(page
)) {
1997 ret
= nfs_writepage_locked(page
, &wbc
);
2003 if (!PagePrivate(page
))
2005 ret
= nfs_commit_inode(inode
, FLUSH_SYNC
);
2010 trace_nfs_writeback_page_exit(inode
, ret
);
2014 #ifdef CONFIG_MIGRATION
2015 int nfs_migrate_page(struct address_space
*mapping
, struct page
*newpage
,
2016 struct page
*page
, enum migrate_mode mode
)
2019 * If PagePrivate is set, then the page is currently associated with
2020 * an in-progress read or write request. Don't try to migrate it.
2022 * FIXME: we could do this in principle, but we'll need a way to ensure
2023 * that we can safely release the inode reference while holding
2026 if (PagePrivate(page
))
2029 if (!nfs_fscache_release_page(page
, GFP_KERNEL
))
2032 return migrate_page(mapping
, newpage
, page
, mode
);
2036 int __init
nfs_init_writepagecache(void)
2038 nfs_wdata_cachep
= kmem_cache_create("nfs_write_data",
2039 sizeof(struct nfs_pgio_header
),
2040 0, SLAB_HWCACHE_ALIGN
,
2042 if (nfs_wdata_cachep
== NULL
)
2045 nfs_wdata_mempool
= mempool_create_slab_pool(MIN_POOL_WRITE
,
2047 if (nfs_wdata_mempool
== NULL
)
2048 goto out_destroy_write_cache
;
2050 nfs_cdata_cachep
= kmem_cache_create("nfs_commit_data",
2051 sizeof(struct nfs_commit_data
),
2052 0, SLAB_HWCACHE_ALIGN
,
2054 if (nfs_cdata_cachep
== NULL
)
2055 goto out_destroy_write_mempool
;
2057 nfs_commit_mempool
= mempool_create_slab_pool(MIN_POOL_COMMIT
,
2059 if (nfs_commit_mempool
== NULL
)
2060 goto out_destroy_commit_cache
;
2063 * NFS congestion size, scale with available memory.
2075 * This allows larger machines to have larger/more transfers.
2076 * Limit the default to 256M
2078 nfs_congestion_kb
= (16*int_sqrt(totalram_pages
)) << (PAGE_SHIFT
-10);
2079 if (nfs_congestion_kb
> 256*1024)
2080 nfs_congestion_kb
= 256*1024;
2084 out_destroy_commit_cache
:
2085 kmem_cache_destroy(nfs_cdata_cachep
);
2086 out_destroy_write_mempool
:
2087 mempool_destroy(nfs_wdata_mempool
);
2088 out_destroy_write_cache
:
2089 kmem_cache_destroy(nfs_wdata_cachep
);
2093 void nfs_destroy_writepagecache(void)
2095 mempool_destroy(nfs_commit_mempool
);
2096 kmem_cache_destroy(nfs_cdata_cachep
);
2097 mempool_destroy(nfs_wdata_mempool
);
2098 kmem_cache_destroy(nfs_wdata_cachep
);
2101 static const struct nfs_rw_ops nfs_rw_write_ops
= {
2102 .rw_alloc_header
= nfs_writehdr_alloc
,
2103 .rw_free_header
= nfs_writehdr_free
,
2104 .rw_done
= nfs_writeback_done
,
2105 .rw_result
= nfs_writeback_result
,
2106 .rw_initiate
= nfs_initiate_write
,