2 * linux/fs/nfs/direct.c
4 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
6 * High-performance uncached I/O for the Linux NFS client
8 * There are important applications whose performance or correctness
9 * depends on uncached access to file data. Database clusters
10 * (multiple copies of the same instance running on separate hosts)
11 * implement their own cache coherency protocol that subsumes file
12 * system cache protocols. Applications that process datasets
13 * considerably larger than the client's memory do not always benefit
14 * from a local cache. A streaming video server, for instance, has no
15 * need to cache the contents of a file.
17 * When an application requests uncached I/O, all read and write requests
18 * are made directly to the server; data stored or fetched via these
19 * requests is not cached in the Linux page cache. The client does not
20 * correct unaligned requests from applications. All requested bytes are
21 * held on permanent storage before a direct write system call returns to
24 * Solaris implements an uncached I/O facility called directio() that
25 * is used for backups and sequential I/O to very large files. Solaris
26 * also supports uncaching whole NFS partitions with "-o forcedirectio,"
27 * an undocumented mount option.
29 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
30 * help from Andrew Morton.
32 * 18 Dec 2001 Initial implementation for 2.4 --cel
33 * 08 Jul 2002 Version for 2.4.19, with bug fixes --trondmy
34 * 08 Jun 2003 Port to 2.5 APIs --cel
35 * 31 Mar 2004 Handle direct I/O without VFS support --cel
36 * 15 Sep 2004 Parallel async reads --cel
37 * 04 May 2005 support O_DIRECT with aio --cel
41 #include <linux/errno.h>
42 #include <linux/sched.h>
43 #include <linux/kernel.h>
44 #include <linux/file.h>
45 #include <linux/pagemap.h>
46 #include <linux/kref.h>
47 #include <linux/slab.h>
48 #include <linux/task_io_accounting_ops.h>
49 #include <linux/module.h>
51 #include <linux/nfs_fs.h>
52 #include <linux/nfs_page.h>
53 #include <linux/sunrpc/clnt.h>
55 #include <linux/uaccess.h>
56 #include <linux/atomic.h>
62 #define NFSDBG_FACILITY NFSDBG_VFS
64 static struct kmem_cache
*nfs_direct_cachep
;
67 * This represents a set of asynchronous requests that we're waiting on
69 struct nfs_direct_mirror
{
73 struct nfs_direct_req
{
74 struct kref kref
; /* release manager */
77 struct nfs_open_context
*ctx
; /* file open context info */
78 struct nfs_lock_context
*l_ctx
; /* Lock context info */
79 struct kiocb
* iocb
; /* controlling i/o request */
80 struct inode
* inode
; /* target file of i/o */
82 /* completion state */
83 atomic_t io_count
; /* i/os we're waiting for */
84 spinlock_t lock
; /* protect completion state */
86 struct nfs_direct_mirror mirrors
[NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX
];
89 loff_t io_start
; /* Start offset for I/O */
90 ssize_t count
, /* bytes actually processed */
91 max_count
, /* max expected count */
92 bytes_left
, /* bytes left to be sent */
93 error
; /* any reported error */
94 struct completion completion
; /* wait for i/o completion */
97 struct nfs_mds_commit_info mds_cinfo
; /* Storage for cinfo */
98 struct pnfs_ds_commit_info ds_cinfo
; /* Storage for cinfo */
99 struct work_struct work
;
102 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
103 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
105 #define NFS_ODIRECT_SHOULD_DIRTY (3) /* dirty user-space page after read */
106 struct nfs_writeverf verf
; /* unstable write verifier */
109 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops
;
110 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops
;
111 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
);
112 static void nfs_direct_write_schedule_work(struct work_struct
*work
);
114 static inline void get_dreq(struct nfs_direct_req
*dreq
)
116 atomic_inc(&dreq
->io_count
);
119 static inline int put_dreq(struct nfs_direct_req
*dreq
)
121 return atomic_dec_and_test(&dreq
->io_count
);
125 nfs_direct_good_bytes(struct nfs_direct_req
*dreq
, struct nfs_pgio_header
*hdr
)
130 WARN_ON_ONCE(dreq
->count
>= dreq
->max_count
);
132 if (dreq
->mirror_count
== 1) {
133 dreq
->mirrors
[hdr
->pgio_mirror_idx
].count
+= hdr
->good_bytes
;
134 dreq
->count
+= hdr
->good_bytes
;
136 /* mirrored writes */
137 count
= dreq
->mirrors
[hdr
->pgio_mirror_idx
].count
;
138 if (count
+ dreq
->io_start
< hdr
->io_start
+ hdr
->good_bytes
) {
139 count
= hdr
->io_start
+ hdr
->good_bytes
- dreq
->io_start
;
140 dreq
->mirrors
[hdr
->pgio_mirror_idx
].count
= count
;
142 /* update the dreq->count by finding the minimum agreed count from all
144 count
= dreq
->mirrors
[0].count
;
146 for (i
= 1; i
< dreq
->mirror_count
; i
++)
147 count
= min(count
, dreq
->mirrors
[i
].count
);
154 * nfs_direct_select_verf - select the right verifier
155 * @dreq - direct request possibly spanning multiple servers
156 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
157 * @commit_idx - commit bucket index for the DS
159 * returns the correct verifier to use given the role of the server
161 static struct nfs_writeverf
*
162 nfs_direct_select_verf(struct nfs_direct_req
*dreq
,
163 struct nfs_client
*ds_clp
,
166 struct nfs_writeverf
*verfp
= &dreq
->verf
;
168 #ifdef CONFIG_NFS_V4_1
170 * pNFS is in use, use the DS verf except commit_through_mds is set
171 * for layout segment where nbuckets is zero.
173 if (ds_clp
&& dreq
->ds_cinfo
.nbuckets
> 0) {
174 if (commit_idx
>= 0 && commit_idx
< dreq
->ds_cinfo
.nbuckets
)
175 verfp
= &dreq
->ds_cinfo
.buckets
[commit_idx
].direct_verf
;
185 * nfs_direct_set_hdr_verf - set the write/commit verifier
186 * @dreq - direct request possibly spanning multiple servers
187 * @hdr - pageio header to validate against previously seen verfs
189 * Set the server's (MDS or DS) "seen" verifier
191 static void nfs_direct_set_hdr_verf(struct nfs_direct_req
*dreq
,
192 struct nfs_pgio_header
*hdr
)
194 struct nfs_writeverf
*verfp
;
196 verfp
= nfs_direct_select_verf(dreq
, hdr
->ds_clp
, hdr
->ds_commit_idx
);
197 WARN_ON_ONCE(verfp
->committed
>= 0);
198 memcpy(verfp
, &hdr
->verf
, sizeof(struct nfs_writeverf
));
199 WARN_ON_ONCE(verfp
->committed
< 0);
202 static int nfs_direct_cmp_verf(const struct nfs_writeverf
*v1
,
203 const struct nfs_writeverf
*v2
)
205 return nfs_write_verifier_cmp(&v1
->verifier
, &v2
->verifier
);
209 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
210 * @dreq - direct request possibly spanning multiple servers
211 * @hdr - pageio header to validate against previously seen verf
213 * set the server's "seen" verf if not initialized.
214 * returns result of comparison between @hdr->verf and the "seen"
215 * verf of the server used by @hdr (DS or MDS)
217 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req
*dreq
,
218 struct nfs_pgio_header
*hdr
)
220 struct nfs_writeverf
*verfp
;
222 verfp
= nfs_direct_select_verf(dreq
, hdr
->ds_clp
, hdr
->ds_commit_idx
);
223 if (verfp
->committed
< 0) {
224 nfs_direct_set_hdr_verf(dreq
, hdr
);
227 return nfs_direct_cmp_verf(verfp
, &hdr
->verf
);
231 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
232 * @dreq - direct request possibly spanning multiple servers
233 * @data - commit data to validate against previously seen verf
235 * returns result of comparison between @data->verf and the verf of
236 * the server used by @data (DS or MDS)
238 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req
*dreq
,
239 struct nfs_commit_data
*data
)
241 struct nfs_writeverf
*verfp
;
243 verfp
= nfs_direct_select_verf(dreq
, data
->ds_clp
,
244 data
->ds_commit_index
);
246 /* verifier not set so always fail */
247 if (verfp
->committed
< 0)
250 return nfs_direct_cmp_verf(verfp
, &data
->verf
);
254 * nfs_direct_IO - NFS address space operation for direct I/O
255 * @iocb: target I/O control block
258 * The presence of this routine in the address space ops vector means
259 * the NFS client supports direct I/O. However, for most direct IO, we
260 * shunt off direct read and write requests before the VFS gets them,
261 * so this method is only ever called for swap.
263 ssize_t
nfs_direct_IO(struct kiocb
*iocb
, struct iov_iter
*iter
)
265 struct inode
*inode
= iocb
->ki_filp
->f_mapping
->host
;
267 /* we only support swap file calling nfs_direct_IO */
268 if (!IS_SWAPFILE(inode
))
271 VM_BUG_ON(iov_iter_count(iter
) != PAGE_SIZE
);
273 if (iov_iter_rw(iter
) == READ
)
274 return nfs_file_direct_read(iocb
, iter
);
275 return nfs_file_direct_write(iocb
, iter
);
278 static void nfs_direct_release_pages(struct page
**pages
, unsigned int npages
)
281 for (i
= 0; i
< npages
; i
++)
285 void nfs_init_cinfo_from_dreq(struct nfs_commit_info
*cinfo
,
286 struct nfs_direct_req
*dreq
)
288 cinfo
->inode
= dreq
->inode
;
289 cinfo
->mds
= &dreq
->mds_cinfo
;
290 cinfo
->ds
= &dreq
->ds_cinfo
;
292 cinfo
->completion_ops
= &nfs_direct_commit_completion_ops
;
295 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req
*dreq
,
296 struct nfs_pageio_descriptor
*pgio
,
297 struct nfs_page
*req
)
299 int mirror_count
= 1;
301 if (pgio
->pg_ops
->pg_get_mirror_count
)
302 mirror_count
= pgio
->pg_ops
->pg_get_mirror_count(pgio
, req
);
304 dreq
->mirror_count
= mirror_count
;
307 static inline struct nfs_direct_req
*nfs_direct_req_alloc(void)
309 struct nfs_direct_req
*dreq
;
311 dreq
= kmem_cache_zalloc(nfs_direct_cachep
, GFP_KERNEL
);
315 kref_init(&dreq
->kref
);
316 kref_get(&dreq
->kref
);
317 init_completion(&dreq
->completion
);
318 INIT_LIST_HEAD(&dreq
->mds_cinfo
.list
);
319 dreq
->verf
.committed
= NFS_INVALID_STABLE_HOW
; /* not set yet */
320 INIT_WORK(&dreq
->work
, nfs_direct_write_schedule_work
);
321 dreq
->mirror_count
= 1;
322 spin_lock_init(&dreq
->lock
);
327 static void nfs_direct_req_free(struct kref
*kref
)
329 struct nfs_direct_req
*dreq
= container_of(kref
, struct nfs_direct_req
, kref
);
331 nfs_free_pnfs_ds_cinfo(&dreq
->ds_cinfo
);
332 if (dreq
->l_ctx
!= NULL
)
333 nfs_put_lock_context(dreq
->l_ctx
);
334 if (dreq
->ctx
!= NULL
)
335 put_nfs_open_context(dreq
->ctx
);
336 kmem_cache_free(nfs_direct_cachep
, dreq
);
339 static void nfs_direct_req_release(struct nfs_direct_req
*dreq
)
341 kref_put(&dreq
->kref
, nfs_direct_req_free
);
344 ssize_t
nfs_dreq_bytes_left(struct nfs_direct_req
*dreq
)
346 return dreq
->bytes_left
;
348 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left
);
351 * Collects and returns the final error value/byte-count.
353 static ssize_t
nfs_direct_wait(struct nfs_direct_req
*dreq
)
355 ssize_t result
= -EIOCBQUEUED
;
357 /* Async requests don't wait here */
361 result
= wait_for_completion_killable(&dreq
->completion
);
364 result
= dreq
->count
;
365 WARN_ON_ONCE(dreq
->count
< 0);
368 result
= dreq
->error
;
371 return (ssize_t
) result
;
375 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
376 * the iocb is still valid here if this is a synchronous request.
378 static void nfs_direct_complete(struct nfs_direct_req
*dreq
)
380 struct inode
*inode
= dreq
->inode
;
382 inode_dio_end(inode
);
385 long res
= (long) dreq
->error
;
386 if (dreq
->count
!= 0) {
387 res
= (long) dreq
->count
;
388 WARN_ON_ONCE(dreq
->count
< 0);
390 dreq
->iocb
->ki_complete(dreq
->iocb
, res
, 0);
393 complete(&dreq
->completion
);
395 nfs_direct_req_release(dreq
);
398 static void nfs_direct_read_completion(struct nfs_pgio_header
*hdr
)
400 unsigned long bytes
= 0;
401 struct nfs_direct_req
*dreq
= hdr
->dreq
;
403 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
406 spin_lock(&dreq
->lock
);
407 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
) && (hdr
->good_bytes
== 0))
408 dreq
->error
= hdr
->error
;
410 nfs_direct_good_bytes(dreq
, hdr
);
412 spin_unlock(&dreq
->lock
);
414 while (!list_empty(&hdr
->pages
)) {
415 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
416 struct page
*page
= req
->wb_page
;
418 if (!PageCompound(page
) && bytes
< hdr
->good_bytes
&&
419 (dreq
->flags
== NFS_ODIRECT_SHOULD_DIRTY
))
420 set_page_dirty(page
);
421 bytes
+= req
->wb_bytes
;
422 nfs_list_remove_request(req
);
423 nfs_release_request(req
);
427 nfs_direct_complete(dreq
);
431 static void nfs_read_sync_pgio_error(struct list_head
*head
)
433 struct nfs_page
*req
;
435 while (!list_empty(head
)) {
436 req
= nfs_list_entry(head
->next
);
437 nfs_list_remove_request(req
);
438 nfs_release_request(req
);
442 static void nfs_direct_pgio_init(struct nfs_pgio_header
*hdr
)
447 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops
= {
448 .error_cleanup
= nfs_read_sync_pgio_error
,
449 .init_hdr
= nfs_direct_pgio_init
,
450 .completion
= nfs_direct_read_completion
,
454 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
455 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
456 * bail and stop sending more reads. Read length accounting is
457 * handled automatically by nfs_direct_read_result(). Otherwise, if
458 * no requests have been sent, just return an error.
461 static ssize_t
nfs_direct_read_schedule_iovec(struct nfs_direct_req
*dreq
,
462 struct iov_iter
*iter
,
465 struct nfs_pageio_descriptor desc
;
466 struct inode
*inode
= dreq
->inode
;
467 ssize_t result
= -EINVAL
;
468 size_t requested_bytes
= 0;
469 size_t rsize
= max_t(size_t, NFS_SERVER(inode
)->rsize
, PAGE_SIZE
);
471 nfs_pageio_init_read(&desc
, dreq
->inode
, false,
472 &nfs_direct_read_completion_ops
);
475 inode_dio_begin(inode
);
477 while (iov_iter_count(iter
)) {
478 struct page
**pagevec
;
483 result
= iov_iter_get_pages_alloc(iter
, &pagevec
,
489 iov_iter_advance(iter
, bytes
);
490 npages
= (result
+ pgbase
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
491 for (i
= 0; i
< npages
; i
++) {
492 struct nfs_page
*req
;
493 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
494 /* XXX do we need to do the eof zeroing found in async_filler? */
495 req
= nfs_create_request(dreq
->ctx
, pagevec
[i
], NULL
,
498 result
= PTR_ERR(req
);
501 req
->wb_index
= pos
>> PAGE_SHIFT
;
502 req
->wb_offset
= pos
& ~PAGE_MASK
;
503 if (!nfs_pageio_add_request(&desc
, req
)) {
504 result
= desc
.pg_error
;
505 nfs_release_request(req
);
510 requested_bytes
+= req_len
;
512 dreq
->bytes_left
-= req_len
;
514 nfs_direct_release_pages(pagevec
, npages
);
520 nfs_pageio_complete(&desc
);
523 * If no bytes were started, return the error, and let the
524 * generic layer handle the completion.
526 if (requested_bytes
== 0) {
527 inode_dio_end(inode
);
528 nfs_direct_req_release(dreq
);
529 return result
< 0 ? result
: -EIO
;
533 nfs_direct_complete(dreq
);
534 return requested_bytes
;
538 * nfs_file_direct_read - file direct read operation for NFS files
539 * @iocb: target I/O control block
540 * @iter: vector of user buffers into which to read data
542 * We use this function for direct reads instead of calling
543 * generic_file_aio_read() in order to avoid gfar's check to see if
544 * the request starts before the end of the file. For that check
545 * to work, we must generate a GETATTR before each direct read, and
546 * even then there is a window between the GETATTR and the subsequent
547 * READ where the file size could change. Our preference is simply
548 * to do all reads the application wants, and the server will take
549 * care of managing the end of file boundary.
551 * This function also eliminates unnecessarily updating the file's
552 * atime locally, as the NFS server sets the file's atime, and this
553 * client must read the updated atime from the server back into its
556 ssize_t
nfs_file_direct_read(struct kiocb
*iocb
, struct iov_iter
*iter
)
558 struct file
*file
= iocb
->ki_filp
;
559 struct address_space
*mapping
= file
->f_mapping
;
560 struct inode
*inode
= mapping
->host
;
561 struct nfs_direct_req
*dreq
;
562 struct nfs_lock_context
*l_ctx
;
563 ssize_t result
= -EINVAL
, requested
;
564 size_t count
= iov_iter_count(iter
);
565 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTREADBYTES
, count
);
567 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
568 file
, count
, (long long) iocb
->ki_pos
);
574 task_io_account_read(count
);
577 dreq
= nfs_direct_req_alloc();
582 dreq
->bytes_left
= dreq
->max_count
= count
;
583 dreq
->io_start
= iocb
->ki_pos
;
584 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
585 l_ctx
= nfs_get_lock_context(dreq
->ctx
);
587 result
= PTR_ERR(l_ctx
);
591 if (!is_sync_kiocb(iocb
))
594 if (iter_is_iovec(iter
))
595 dreq
->flags
= NFS_ODIRECT_SHOULD_DIRTY
;
597 nfs_start_io_direct(inode
);
599 NFS_I(inode
)->read_io
+= count
;
600 requested
= nfs_direct_read_schedule_iovec(dreq
, iter
, iocb
->ki_pos
);
602 nfs_end_io_direct(inode
);
605 result
= nfs_direct_wait(dreq
);
608 iocb
->ki_pos
+= result
;
610 iov_iter_revert(iter
, requested
);
616 nfs_direct_req_release(dreq
);
622 nfs_direct_write_scan_commit_list(struct inode
*inode
,
623 struct list_head
*list
,
624 struct nfs_commit_info
*cinfo
)
626 mutex_lock(&NFS_I(cinfo
->inode
)->commit_mutex
);
627 #ifdef CONFIG_NFS_V4_1
628 if (cinfo
->ds
!= NULL
&& cinfo
->ds
->nwritten
!= 0)
629 NFS_SERVER(inode
)->pnfs_curr_ld
->recover_commit_reqs(list
, cinfo
);
631 nfs_scan_commit_list(&cinfo
->mds
->list
, list
, cinfo
, 0);
632 mutex_unlock(&NFS_I(cinfo
->inode
)->commit_mutex
);
635 static void nfs_direct_write_reschedule(struct nfs_direct_req
*dreq
)
637 struct nfs_pageio_descriptor desc
;
638 struct nfs_page
*req
, *tmp
;
640 struct nfs_commit_info cinfo
;
644 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
645 nfs_direct_write_scan_commit_list(dreq
->inode
, &reqs
, &cinfo
);
648 dreq
->verf
.committed
= NFS_INVALID_STABLE_HOW
;
649 nfs_clear_pnfs_ds_commit_verifiers(&dreq
->ds_cinfo
);
650 for (i
= 0; i
< dreq
->mirror_count
; i
++)
651 dreq
->mirrors
[i
].count
= 0;
654 nfs_pageio_init_write(&desc
, dreq
->inode
, FLUSH_STABLE
, false,
655 &nfs_direct_write_completion_ops
);
658 req
= nfs_list_entry(reqs
.next
);
659 nfs_direct_setup_mirroring(dreq
, &desc
, req
);
660 if (desc
.pg_error
< 0) {
661 list_splice_init(&reqs
, &failed
);
665 list_for_each_entry_safe(req
, tmp
, &reqs
, wb_list
) {
666 if (!nfs_pageio_add_request(&desc
, req
)) {
667 nfs_list_remove_request(req
);
668 nfs_list_add_request(req
, &failed
);
669 spin_lock(&cinfo
.inode
->i_lock
);
671 if (desc
.pg_error
< 0)
672 dreq
->error
= desc
.pg_error
;
675 spin_unlock(&cinfo
.inode
->i_lock
);
677 nfs_release_request(req
);
679 nfs_pageio_complete(&desc
);
682 while (!list_empty(&failed
)) {
683 req
= nfs_list_entry(failed
.next
);
684 nfs_list_remove_request(req
);
685 nfs_unlock_and_release_request(req
);
689 nfs_direct_write_complete(dreq
);
692 static void nfs_direct_commit_complete(struct nfs_commit_data
*data
)
694 struct nfs_direct_req
*dreq
= data
->dreq
;
695 struct nfs_commit_info cinfo
;
696 struct nfs_page
*req
;
697 int status
= data
->task
.tk_status
;
699 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
700 if (status
< 0 || nfs_direct_cmp_commit_data_verf(dreq
, data
))
701 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
703 while (!list_empty(&data
->pages
)) {
704 req
= nfs_list_entry(data
->pages
.next
);
705 nfs_list_remove_request(req
);
706 if (dreq
->flags
== NFS_ODIRECT_RESCHED_WRITES
) {
707 /* Note the rewrite will go through mds */
708 nfs_mark_request_commit(req
, NULL
, &cinfo
, 0);
710 nfs_release_request(req
);
711 nfs_unlock_and_release_request(req
);
714 if (atomic_dec_and_test(&cinfo
.mds
->rpcs_out
))
715 nfs_direct_write_complete(dreq
);
718 static void nfs_direct_resched_write(struct nfs_commit_info
*cinfo
,
719 struct nfs_page
*req
)
721 struct nfs_direct_req
*dreq
= cinfo
->dreq
;
723 spin_lock(&dreq
->lock
);
724 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
725 spin_unlock(&dreq
->lock
);
726 nfs_mark_request_commit(req
, NULL
, cinfo
, 0);
729 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops
= {
730 .completion
= nfs_direct_commit_complete
,
731 .resched_write
= nfs_direct_resched_write
,
734 static void nfs_direct_commit_schedule(struct nfs_direct_req
*dreq
)
737 struct nfs_commit_info cinfo
;
740 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
741 nfs_scan_commit(dreq
->inode
, &mds_list
, &cinfo
);
742 res
= nfs_generic_commit_list(dreq
->inode
, &mds_list
, 0, &cinfo
);
743 if (res
< 0) /* res == -ENOMEM */
744 nfs_direct_write_reschedule(dreq
);
747 static void nfs_direct_write_schedule_work(struct work_struct
*work
)
749 struct nfs_direct_req
*dreq
= container_of(work
, struct nfs_direct_req
, work
);
750 int flags
= dreq
->flags
;
754 case NFS_ODIRECT_DO_COMMIT
:
755 nfs_direct_commit_schedule(dreq
);
757 case NFS_ODIRECT_RESCHED_WRITES
:
758 nfs_direct_write_reschedule(dreq
);
761 nfs_zap_mapping(dreq
->inode
, dreq
->inode
->i_mapping
);
762 nfs_direct_complete(dreq
);
766 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
)
768 schedule_work(&dreq
->work
); /* Calls nfs_direct_write_schedule_work */
771 static void nfs_direct_write_completion(struct nfs_pgio_header
*hdr
)
773 struct nfs_direct_req
*dreq
= hdr
->dreq
;
774 struct nfs_commit_info cinfo
;
775 bool request_commit
= false;
776 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
778 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
781 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
783 spin_lock(&dreq
->lock
);
785 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
))
786 dreq
->error
= hdr
->error
;
787 if (dreq
->error
== 0) {
788 nfs_direct_good_bytes(dreq
, hdr
);
789 if (nfs_write_need_commit(hdr
)) {
790 if (dreq
->flags
== NFS_ODIRECT_RESCHED_WRITES
)
791 request_commit
= true;
792 else if (dreq
->flags
== 0) {
793 nfs_direct_set_hdr_verf(dreq
, hdr
);
794 request_commit
= true;
795 dreq
->flags
= NFS_ODIRECT_DO_COMMIT
;
796 } else if (dreq
->flags
== NFS_ODIRECT_DO_COMMIT
) {
797 request_commit
= true;
798 if (nfs_direct_set_or_cmp_hdr_verf(dreq
, hdr
))
800 NFS_ODIRECT_RESCHED_WRITES
;
804 spin_unlock(&dreq
->lock
);
806 while (!list_empty(&hdr
->pages
)) {
808 req
= nfs_list_entry(hdr
->pages
.next
);
809 nfs_list_remove_request(req
);
810 if (request_commit
) {
811 kref_get(&req
->wb_kref
);
812 nfs_mark_request_commit(req
, hdr
->lseg
, &cinfo
,
815 nfs_unlock_and_release_request(req
);
820 nfs_direct_write_complete(dreq
);
824 static void nfs_write_sync_pgio_error(struct list_head
*head
)
826 struct nfs_page
*req
;
828 while (!list_empty(head
)) {
829 req
= nfs_list_entry(head
->next
);
830 nfs_list_remove_request(req
);
831 nfs_unlock_and_release_request(req
);
835 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header
*hdr
)
837 struct nfs_direct_req
*dreq
= hdr
->dreq
;
839 spin_lock(&dreq
->lock
);
840 if (dreq
->error
== 0) {
841 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
842 /* fake unstable write to let common nfs resend pages */
843 hdr
->verf
.committed
= NFS_UNSTABLE
;
844 hdr
->good_bytes
= hdr
->args
.count
;
846 spin_unlock(&dreq
->lock
);
849 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops
= {
850 .error_cleanup
= nfs_write_sync_pgio_error
,
851 .init_hdr
= nfs_direct_pgio_init
,
852 .completion
= nfs_direct_write_completion
,
853 .reschedule_io
= nfs_direct_write_reschedule_io
,
858 * NB: Return the value of the first error return code. Subsequent
859 * errors after the first one are ignored.
862 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
863 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
864 * bail and stop sending more writes. Write length accounting is
865 * handled automatically by nfs_direct_write_result(). Otherwise, if
866 * no requests have been sent, just return an error.
868 static ssize_t
nfs_direct_write_schedule_iovec(struct nfs_direct_req
*dreq
,
869 struct iov_iter
*iter
,
872 struct nfs_pageio_descriptor desc
;
873 struct inode
*inode
= dreq
->inode
;
875 size_t requested_bytes
= 0;
876 size_t wsize
= max_t(size_t, NFS_SERVER(inode
)->wsize
, PAGE_SIZE
);
878 nfs_pageio_init_write(&desc
, inode
, FLUSH_COND_STABLE
, false,
879 &nfs_direct_write_completion_ops
);
882 inode_dio_begin(inode
);
884 NFS_I(inode
)->write_io
+= iov_iter_count(iter
);
885 while (iov_iter_count(iter
)) {
886 struct page
**pagevec
;
891 result
= iov_iter_get_pages_alloc(iter
, &pagevec
,
897 iov_iter_advance(iter
, bytes
);
898 npages
= (result
+ pgbase
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
899 for (i
= 0; i
< npages
; i
++) {
900 struct nfs_page
*req
;
901 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
903 req
= nfs_create_request(dreq
->ctx
, pagevec
[i
], NULL
,
906 result
= PTR_ERR(req
);
910 nfs_direct_setup_mirroring(dreq
, &desc
, req
);
911 if (desc
.pg_error
< 0) {
912 nfs_free_request(req
);
913 result
= desc
.pg_error
;
917 nfs_lock_request(req
);
918 req
->wb_index
= pos
>> PAGE_SHIFT
;
919 req
->wb_offset
= pos
& ~PAGE_MASK
;
920 if (!nfs_pageio_add_request(&desc
, req
)) {
921 result
= desc
.pg_error
;
922 nfs_unlock_and_release_request(req
);
927 requested_bytes
+= req_len
;
929 dreq
->bytes_left
-= req_len
;
931 nfs_direct_release_pages(pagevec
, npages
);
936 nfs_pageio_complete(&desc
);
939 * If no bytes were started, return the error, and let the
940 * generic layer handle the completion.
942 if (requested_bytes
== 0) {
943 inode_dio_end(inode
);
944 nfs_direct_req_release(dreq
);
945 return result
< 0 ? result
: -EIO
;
949 nfs_direct_write_complete(dreq
);
950 return requested_bytes
;
954 * nfs_file_direct_write - file direct write operation for NFS files
955 * @iocb: target I/O control block
956 * @iter: vector of user buffers from which to write data
958 * We use this function for direct writes instead of calling
959 * generic_file_aio_write() in order to avoid taking the inode
960 * semaphore and updating the i_size. The NFS server will set
961 * the new i_size and this client must read the updated size
962 * back into its cache. We let the server do generic write
963 * parameter checking and report problems.
965 * We eliminate local atime updates, see direct read above.
967 * We avoid unnecessary page cache invalidations for normal cached
968 * readers of this file.
970 * Note that O_APPEND is not supported for NFS direct writes, as there
971 * is no atomic O_APPEND write facility in the NFS protocol.
973 ssize_t
nfs_file_direct_write(struct kiocb
*iocb
, struct iov_iter
*iter
)
975 ssize_t result
= -EINVAL
, requested
;
977 struct file
*file
= iocb
->ki_filp
;
978 struct address_space
*mapping
= file
->f_mapping
;
979 struct inode
*inode
= mapping
->host
;
980 struct nfs_direct_req
*dreq
;
981 struct nfs_lock_context
*l_ctx
;
984 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
985 file
, iov_iter_count(iter
), (long long) iocb
->ki_pos
);
987 result
= generic_write_checks(iocb
, iter
);
991 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTWRITTENBYTES
, count
);
994 end
= (pos
+ iov_iter_count(iter
) - 1) >> PAGE_SHIFT
;
996 task_io_account_write(count
);
999 dreq
= nfs_direct_req_alloc();
1003 dreq
->inode
= inode
;
1004 dreq
->bytes_left
= dreq
->max_count
= count
;
1005 dreq
->io_start
= pos
;
1006 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
1007 l_ctx
= nfs_get_lock_context(dreq
->ctx
);
1008 if (IS_ERR(l_ctx
)) {
1009 result
= PTR_ERR(l_ctx
);
1012 dreq
->l_ctx
= l_ctx
;
1013 if (!is_sync_kiocb(iocb
))
1016 nfs_start_io_direct(inode
);
1018 requested
= nfs_direct_write_schedule_iovec(dreq
, iter
, pos
);
1020 if (mapping
->nrpages
) {
1021 invalidate_inode_pages2_range(mapping
,
1022 pos
>> PAGE_SHIFT
, end
);
1025 nfs_end_io_direct(inode
);
1027 if (requested
> 0) {
1028 result
= nfs_direct_wait(dreq
);
1030 requested
-= result
;
1031 iocb
->ki_pos
= pos
+ result
;
1032 /* XXX: should check the generic_write_sync retval */
1033 generic_write_sync(iocb
, result
);
1035 iov_iter_revert(iter
, requested
);
1040 nfs_direct_req_release(dreq
);
1046 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1049 int __init
nfs_init_directcache(void)
1051 nfs_direct_cachep
= kmem_cache_create("nfs_direct_cache",
1052 sizeof(struct nfs_direct_req
),
1053 0, (SLAB_RECLAIM_ACCOUNT
|
1056 if (nfs_direct_cachep
== NULL
)
1063 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1066 void nfs_destroy_directcache(void)
1068 kmem_cache_destroy(nfs_direct_cachep
);