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 <asm/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 ssize_t count
, /* bytes actually processed */
90 bytes_left
, /* bytes left to be sent */
91 io_start
, /* start of IO */
92 error
; /* any reported error */
93 struct completion completion
; /* wait for i/o completion */
96 struct nfs_mds_commit_info mds_cinfo
; /* Storage for cinfo */
97 struct pnfs_ds_commit_info ds_cinfo
; /* Storage for cinfo */
98 struct work_struct work
;
100 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
101 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
102 struct nfs_writeverf verf
; /* unstable write verifier */
105 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops
;
106 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops
;
107 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
, struct inode
*inode
);
108 static void nfs_direct_write_schedule_work(struct work_struct
*work
);
110 static inline void get_dreq(struct nfs_direct_req
*dreq
)
112 atomic_inc(&dreq
->io_count
);
115 static inline int put_dreq(struct nfs_direct_req
*dreq
)
117 return atomic_dec_and_test(&dreq
->io_count
);
121 nfs_direct_good_bytes(struct nfs_direct_req
*dreq
, struct nfs_pgio_header
*hdr
)
126 if (dreq
->mirror_count
== 1) {
127 dreq
->mirrors
[hdr
->pgio_mirror_idx
].count
+= hdr
->good_bytes
;
128 dreq
->count
+= hdr
->good_bytes
;
130 /* mirrored writes */
131 count
= dreq
->mirrors
[hdr
->pgio_mirror_idx
].count
;
132 if (count
+ dreq
->io_start
< hdr
->io_start
+ hdr
->good_bytes
) {
133 count
= hdr
->io_start
+ hdr
->good_bytes
- dreq
->io_start
;
134 dreq
->mirrors
[hdr
->pgio_mirror_idx
].count
= count
;
136 /* update the dreq->count by finding the minimum agreed count from all
138 count
= dreq
->mirrors
[0].count
;
140 for (i
= 1; i
< dreq
->mirror_count
; i
++)
141 count
= min(count
, dreq
->mirrors
[i
].count
);
148 * nfs_direct_select_verf - select the right verifier
149 * @dreq - direct request possibly spanning multiple servers
150 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
151 * @commit_idx - commit bucket index for the DS
153 * returns the correct verifier to use given the role of the server
155 static struct nfs_writeverf
*
156 nfs_direct_select_verf(struct nfs_direct_req
*dreq
,
157 struct nfs_client
*ds_clp
,
160 struct nfs_writeverf
*verfp
= &dreq
->verf
;
162 #ifdef CONFIG_NFS_V4_1
164 * pNFS is in use, use the DS verf except commit_through_mds is set
165 * for layout segment where nbuckets is zero.
167 if (ds_clp
&& dreq
->ds_cinfo
.nbuckets
> 0) {
168 if (commit_idx
>= 0 && commit_idx
< dreq
->ds_cinfo
.nbuckets
)
169 verfp
= &dreq
->ds_cinfo
.buckets
[commit_idx
].direct_verf
;
179 * nfs_direct_set_hdr_verf - set the write/commit verifier
180 * @dreq - direct request possibly spanning multiple servers
181 * @hdr - pageio header to validate against previously seen verfs
183 * Set the server's (MDS or DS) "seen" verifier
185 static void nfs_direct_set_hdr_verf(struct nfs_direct_req
*dreq
,
186 struct nfs_pgio_header
*hdr
)
188 struct nfs_writeverf
*verfp
;
190 verfp
= nfs_direct_select_verf(dreq
, hdr
->ds_clp
, hdr
->ds_commit_idx
);
191 WARN_ON_ONCE(verfp
->committed
>= 0);
192 memcpy(verfp
, &hdr
->verf
, sizeof(struct nfs_writeverf
));
193 WARN_ON_ONCE(verfp
->committed
< 0);
197 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
198 * @dreq - direct request possibly spanning multiple servers
199 * @hdr - pageio header to validate against previously seen verf
201 * set the server's "seen" verf if not initialized.
202 * returns result of comparison between @hdr->verf and the "seen"
203 * verf of the server used by @hdr (DS or MDS)
205 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req
*dreq
,
206 struct nfs_pgio_header
*hdr
)
208 struct nfs_writeverf
*verfp
;
210 verfp
= nfs_direct_select_verf(dreq
, hdr
->ds_clp
, hdr
->ds_commit_idx
);
211 if (verfp
->committed
< 0) {
212 nfs_direct_set_hdr_verf(dreq
, hdr
);
215 return memcmp(verfp
, &hdr
->verf
, sizeof(struct nfs_writeverf
));
219 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
220 * @dreq - direct request possibly spanning multiple servers
221 * @data - commit data to validate against previously seen verf
223 * returns result of comparison between @data->verf and the verf of
224 * the server used by @data (DS or MDS)
226 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req
*dreq
,
227 struct nfs_commit_data
*data
)
229 struct nfs_writeverf
*verfp
;
231 verfp
= nfs_direct_select_verf(dreq
, data
->ds_clp
,
232 data
->ds_commit_index
);
234 /* verifier not set so always fail */
235 if (verfp
->committed
< 0)
238 return memcmp(verfp
, &data
->verf
, sizeof(struct nfs_writeverf
));
242 * nfs_direct_IO - NFS address space operation for direct I/O
243 * @iocb: target I/O control block
244 * @iov: array of vectors that define I/O buffer
245 * @pos: offset in file to begin the operation
246 * @nr_segs: size of iovec array
248 * The presence of this routine in the address space ops vector means
249 * the NFS client supports direct I/O. However, for most direct IO, we
250 * shunt off direct read and write requests before the VFS gets them,
251 * so this method is only ever called for swap.
253 ssize_t
nfs_direct_IO(struct kiocb
*iocb
, struct iov_iter
*iter
)
255 struct inode
*inode
= iocb
->ki_filp
->f_mapping
->host
;
257 /* we only support swap file calling nfs_direct_IO */
258 if (!IS_SWAPFILE(inode
))
261 VM_BUG_ON(iov_iter_count(iter
) != PAGE_SIZE
);
263 if (iov_iter_rw(iter
) == READ
)
264 return nfs_file_direct_read(iocb
, iter
);
265 return nfs_file_direct_write(iocb
, iter
);
268 static void nfs_direct_release_pages(struct page
**pages
, unsigned int npages
)
271 for (i
= 0; i
< npages
; i
++)
275 void nfs_init_cinfo_from_dreq(struct nfs_commit_info
*cinfo
,
276 struct nfs_direct_req
*dreq
)
278 cinfo
->lock
= &dreq
->inode
->i_lock
;
279 cinfo
->mds
= &dreq
->mds_cinfo
;
280 cinfo
->ds
= &dreq
->ds_cinfo
;
282 cinfo
->completion_ops
= &nfs_direct_commit_completion_ops
;
285 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req
*dreq
,
286 struct nfs_pageio_descriptor
*pgio
,
287 struct nfs_page
*req
)
289 int mirror_count
= 1;
291 if (pgio
->pg_ops
->pg_get_mirror_count
)
292 mirror_count
= pgio
->pg_ops
->pg_get_mirror_count(pgio
, req
);
294 dreq
->mirror_count
= mirror_count
;
297 static inline struct nfs_direct_req
*nfs_direct_req_alloc(void)
299 struct nfs_direct_req
*dreq
;
301 dreq
= kmem_cache_zalloc(nfs_direct_cachep
, GFP_KERNEL
);
305 kref_init(&dreq
->kref
);
306 kref_get(&dreq
->kref
);
307 init_completion(&dreq
->completion
);
308 INIT_LIST_HEAD(&dreq
->mds_cinfo
.list
);
309 dreq
->verf
.committed
= NFS_INVALID_STABLE_HOW
; /* not set yet */
310 INIT_WORK(&dreq
->work
, nfs_direct_write_schedule_work
);
311 dreq
->mirror_count
= 1;
312 spin_lock_init(&dreq
->lock
);
317 static void nfs_direct_req_free(struct kref
*kref
)
319 struct nfs_direct_req
*dreq
= container_of(kref
, struct nfs_direct_req
, kref
);
321 nfs_free_pnfs_ds_cinfo(&dreq
->ds_cinfo
);
322 if (dreq
->l_ctx
!= NULL
)
323 nfs_put_lock_context(dreq
->l_ctx
);
324 if (dreq
->ctx
!= NULL
)
325 put_nfs_open_context(dreq
->ctx
);
326 kmem_cache_free(nfs_direct_cachep
, dreq
);
329 static void nfs_direct_req_release(struct nfs_direct_req
*dreq
)
331 kref_put(&dreq
->kref
, nfs_direct_req_free
);
334 ssize_t
nfs_dreq_bytes_left(struct nfs_direct_req
*dreq
)
336 return dreq
->bytes_left
;
338 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left
);
341 * Collects and returns the final error value/byte-count.
343 static ssize_t
nfs_direct_wait(struct nfs_direct_req
*dreq
)
345 ssize_t result
= -EIOCBQUEUED
;
347 /* Async requests don't wait here */
351 result
= wait_for_completion_killable(&dreq
->completion
);
354 result
= dreq
->error
;
356 result
= dreq
->count
;
359 return (ssize_t
) result
;
363 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
364 * the iocb is still valid here if this is a synchronous request.
366 static void nfs_direct_complete(struct nfs_direct_req
*dreq
, bool write
)
368 struct inode
*inode
= dreq
->inode
;
370 if (dreq
->iocb
&& write
) {
371 loff_t pos
= dreq
->iocb
->ki_pos
+ dreq
->count
;
373 spin_lock(&inode
->i_lock
);
374 if (i_size_read(inode
) < pos
)
375 i_size_write(inode
, pos
);
376 spin_unlock(&inode
->i_lock
);
380 nfs_zap_mapping(inode
, inode
->i_mapping
);
382 inode_dio_end(inode
);
385 long res
= (long) dreq
->error
;
387 res
= (long) dreq
->count
;
388 dreq
->iocb
->ki_complete(dreq
->iocb
, res
, 0);
391 complete_all(&dreq
->completion
);
393 nfs_direct_req_release(dreq
);
396 static void nfs_direct_readpage_release(struct nfs_page
*req
)
398 dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
399 req
->wb_context
->dentry
->d_sb
->s_id
,
400 (unsigned long long)NFS_FILEID(d_inode(req
->wb_context
->dentry
)),
402 (long long)req_offset(req
));
403 nfs_release_request(req
);
406 static void nfs_direct_read_completion(struct nfs_pgio_header
*hdr
)
408 unsigned long bytes
= 0;
409 struct nfs_direct_req
*dreq
= hdr
->dreq
;
411 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
414 spin_lock(&dreq
->lock
);
415 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
) && (hdr
->good_bytes
== 0))
416 dreq
->error
= hdr
->error
;
418 nfs_direct_good_bytes(dreq
, hdr
);
420 spin_unlock(&dreq
->lock
);
422 while (!list_empty(&hdr
->pages
)) {
423 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
424 struct page
*page
= req
->wb_page
;
426 if (!PageCompound(page
) && bytes
< hdr
->good_bytes
)
427 set_page_dirty(page
);
428 bytes
+= req
->wb_bytes
;
429 nfs_list_remove_request(req
);
430 nfs_direct_readpage_release(req
);
434 nfs_direct_complete(dreq
, false);
438 static void nfs_read_sync_pgio_error(struct list_head
*head
)
440 struct nfs_page
*req
;
442 while (!list_empty(head
)) {
443 req
= nfs_list_entry(head
->next
);
444 nfs_list_remove_request(req
);
445 nfs_release_request(req
);
449 static void nfs_direct_pgio_init(struct nfs_pgio_header
*hdr
)
454 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops
= {
455 .error_cleanup
= nfs_read_sync_pgio_error
,
456 .init_hdr
= nfs_direct_pgio_init
,
457 .completion
= nfs_direct_read_completion
,
461 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
462 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
463 * bail and stop sending more reads. Read length accounting is
464 * handled automatically by nfs_direct_read_result(). Otherwise, if
465 * no requests have been sent, just return an error.
468 static ssize_t
nfs_direct_read_schedule_iovec(struct nfs_direct_req
*dreq
,
469 struct iov_iter
*iter
,
472 struct nfs_pageio_descriptor desc
;
473 struct inode
*inode
= dreq
->inode
;
474 ssize_t result
= -EINVAL
;
475 size_t requested_bytes
= 0;
476 size_t rsize
= max_t(size_t, NFS_SERVER(inode
)->rsize
, PAGE_SIZE
);
478 nfs_pageio_init_read(&desc
, dreq
->inode
, false,
479 &nfs_direct_read_completion_ops
);
482 inode_dio_begin(inode
);
484 while (iov_iter_count(iter
)) {
485 struct page
**pagevec
;
490 result
= iov_iter_get_pages_alloc(iter
, &pagevec
,
496 iov_iter_advance(iter
, bytes
);
497 npages
= (result
+ pgbase
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
498 for (i
= 0; i
< npages
; i
++) {
499 struct nfs_page
*req
;
500 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
501 /* XXX do we need to do the eof zeroing found in async_filler? */
502 req
= nfs_create_request(dreq
->ctx
, pagevec
[i
], NULL
,
505 result
= PTR_ERR(req
);
508 req
->wb_index
= pos
>> PAGE_SHIFT
;
509 req
->wb_offset
= pos
& ~PAGE_MASK
;
510 if (!nfs_pageio_add_request(&desc
, req
)) {
511 result
= desc
.pg_error
;
512 nfs_release_request(req
);
517 requested_bytes
+= req_len
;
519 dreq
->bytes_left
-= req_len
;
521 nfs_direct_release_pages(pagevec
, npages
);
527 nfs_pageio_complete(&desc
);
530 * If no bytes were started, return the error, and let the
531 * generic layer handle the completion.
533 if (requested_bytes
== 0) {
534 inode_dio_end(inode
);
535 nfs_direct_req_release(dreq
);
536 return result
< 0 ? result
: -EIO
;
540 nfs_direct_complete(dreq
, false);
545 * nfs_file_direct_read - file direct read operation for NFS files
546 * @iocb: target I/O control block
547 * @iter: vector of user buffers into which to read data
549 * We use this function for direct reads instead of calling
550 * generic_file_aio_read() in order to avoid gfar's check to see if
551 * the request starts before the end of the file. For that check
552 * to work, we must generate a GETATTR before each direct read, and
553 * even then there is a window between the GETATTR and the subsequent
554 * READ where the file size could change. Our preference is simply
555 * to do all reads the application wants, and the server will take
556 * care of managing the end of file boundary.
558 * This function also eliminates unnecessarily updating the file's
559 * atime locally, as the NFS server sets the file's atime, and this
560 * client must read the updated atime from the server back into its
563 ssize_t
nfs_file_direct_read(struct kiocb
*iocb
, struct iov_iter
*iter
)
565 struct file
*file
= iocb
->ki_filp
;
566 struct address_space
*mapping
= file
->f_mapping
;
567 struct inode
*inode
= mapping
->host
;
568 struct nfs_direct_req
*dreq
;
569 struct nfs_lock_context
*l_ctx
;
570 ssize_t result
= -EINVAL
;
571 size_t count
= iov_iter_count(iter
);
572 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTREADBYTES
, count
);
574 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
575 file
, count
, (long long) iocb
->ki_pos
);
582 result
= nfs_sync_mapping(mapping
);
586 task_io_account_read(count
);
589 dreq
= nfs_direct_req_alloc();
594 dreq
->bytes_left
= count
;
595 dreq
->io_start
= iocb
->ki_pos
;
596 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
597 l_ctx
= nfs_get_lock_context(dreq
->ctx
);
599 result
= PTR_ERR(l_ctx
);
603 if (!is_sync_kiocb(iocb
))
606 NFS_I(inode
)->read_io
+= count
;
607 result
= nfs_direct_read_schedule_iovec(dreq
, iter
, iocb
->ki_pos
);
612 result
= nfs_direct_wait(dreq
);
614 iocb
->ki_pos
+= result
;
617 nfs_direct_req_release(dreq
);
621 nfs_direct_req_release(dreq
);
629 nfs_direct_write_scan_commit_list(struct inode
*inode
,
630 struct list_head
*list
,
631 struct nfs_commit_info
*cinfo
)
633 spin_lock(cinfo
->lock
);
634 #ifdef CONFIG_NFS_V4_1
635 if (cinfo
->ds
!= NULL
&& cinfo
->ds
->nwritten
!= 0)
636 NFS_SERVER(inode
)->pnfs_curr_ld
->recover_commit_reqs(list
, cinfo
);
638 nfs_scan_commit_list(&cinfo
->mds
->list
, list
, cinfo
, 0);
639 spin_unlock(cinfo
->lock
);
642 static void nfs_direct_write_reschedule(struct nfs_direct_req
*dreq
)
644 struct nfs_pageio_descriptor desc
;
645 struct nfs_page
*req
, *tmp
;
647 struct nfs_commit_info cinfo
;
651 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
652 nfs_direct_write_scan_commit_list(dreq
->inode
, &reqs
, &cinfo
);
655 for (i
= 0; i
< dreq
->mirror_count
; i
++)
656 dreq
->mirrors
[i
].count
= 0;
659 nfs_pageio_init_write(&desc
, dreq
->inode
, FLUSH_STABLE
, false,
660 &nfs_direct_write_completion_ops
);
663 req
= nfs_list_entry(reqs
.next
);
664 nfs_direct_setup_mirroring(dreq
, &desc
, req
);
665 if (desc
.pg_error
< 0) {
666 list_splice_init(&reqs
, &failed
);
670 list_for_each_entry_safe(req
, tmp
, &reqs
, wb_list
) {
671 if (!nfs_pageio_add_request(&desc
, req
)) {
672 nfs_list_remove_request(req
);
673 nfs_list_add_request(req
, &failed
);
674 spin_lock(cinfo
.lock
);
676 if (desc
.pg_error
< 0)
677 dreq
->error
= desc
.pg_error
;
680 spin_unlock(cinfo
.lock
);
682 nfs_release_request(req
);
684 nfs_pageio_complete(&desc
);
687 while (!list_empty(&failed
)) {
688 req
= nfs_list_entry(failed
.next
);
689 nfs_list_remove_request(req
);
690 nfs_unlock_and_release_request(req
);
694 nfs_direct_write_complete(dreq
, dreq
->inode
);
697 static void nfs_direct_commit_complete(struct nfs_commit_data
*data
)
699 struct nfs_direct_req
*dreq
= data
->dreq
;
700 struct nfs_commit_info cinfo
;
701 struct nfs_page
*req
;
702 int status
= data
->task
.tk_status
;
704 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
706 dprintk("NFS: %5u commit failed with error %d.\n",
707 data
->task
.tk_pid
, status
);
708 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
709 } else if (nfs_direct_cmp_commit_data_verf(dreq
, data
)) {
710 dprintk("NFS: %5u commit verify failed\n", data
->task
.tk_pid
);
711 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
714 dprintk("NFS: %5u commit returned %d\n", data
->task
.tk_pid
, status
);
715 while (!list_empty(&data
->pages
)) {
716 req
= nfs_list_entry(data
->pages
.next
);
717 nfs_list_remove_request(req
);
718 if (dreq
->flags
== NFS_ODIRECT_RESCHED_WRITES
) {
719 /* Note the rewrite will go through mds */
720 nfs_mark_request_commit(req
, NULL
, &cinfo
, 0);
722 nfs_release_request(req
);
723 nfs_unlock_and_release_request(req
);
726 if (atomic_dec_and_test(&cinfo
.mds
->rpcs_out
))
727 nfs_direct_write_complete(dreq
, data
->inode
);
730 static void nfs_direct_resched_write(struct nfs_commit_info
*cinfo
,
731 struct nfs_page
*req
)
733 struct nfs_direct_req
*dreq
= cinfo
->dreq
;
735 spin_lock(&dreq
->lock
);
736 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
737 spin_unlock(&dreq
->lock
);
738 nfs_mark_request_commit(req
, NULL
, cinfo
, 0);
741 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops
= {
742 .completion
= nfs_direct_commit_complete
,
743 .resched_write
= nfs_direct_resched_write
,
746 static void nfs_direct_commit_schedule(struct nfs_direct_req
*dreq
)
749 struct nfs_commit_info cinfo
;
752 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
753 nfs_scan_commit(dreq
->inode
, &mds_list
, &cinfo
);
754 res
= nfs_generic_commit_list(dreq
->inode
, &mds_list
, 0, &cinfo
);
755 if (res
< 0) /* res == -ENOMEM */
756 nfs_direct_write_reschedule(dreq
);
759 static void nfs_direct_write_schedule_work(struct work_struct
*work
)
761 struct nfs_direct_req
*dreq
= container_of(work
, struct nfs_direct_req
, work
);
762 int flags
= dreq
->flags
;
766 case NFS_ODIRECT_DO_COMMIT
:
767 nfs_direct_commit_schedule(dreq
);
769 case NFS_ODIRECT_RESCHED_WRITES
:
770 nfs_direct_write_reschedule(dreq
);
773 nfs_direct_complete(dreq
, true);
777 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
, struct inode
*inode
)
779 schedule_work(&dreq
->work
); /* Calls nfs_direct_write_schedule_work */
782 static void nfs_direct_write_completion(struct nfs_pgio_header
*hdr
)
784 struct nfs_direct_req
*dreq
= hdr
->dreq
;
785 struct nfs_commit_info cinfo
;
786 bool request_commit
= false;
787 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
789 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
792 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
794 spin_lock(&dreq
->lock
);
796 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
)) {
798 dreq
->error
= hdr
->error
;
800 if (dreq
->error
== 0) {
801 nfs_direct_good_bytes(dreq
, hdr
);
802 if (nfs_write_need_commit(hdr
)) {
803 if (dreq
->flags
== NFS_ODIRECT_RESCHED_WRITES
)
804 request_commit
= true;
805 else if (dreq
->flags
== 0) {
806 nfs_direct_set_hdr_verf(dreq
, hdr
);
807 request_commit
= true;
808 dreq
->flags
= NFS_ODIRECT_DO_COMMIT
;
809 } else if (dreq
->flags
== NFS_ODIRECT_DO_COMMIT
) {
810 request_commit
= true;
811 if (nfs_direct_set_or_cmp_hdr_verf(dreq
, hdr
))
813 NFS_ODIRECT_RESCHED_WRITES
;
817 spin_unlock(&dreq
->lock
);
819 while (!list_empty(&hdr
->pages
)) {
821 req
= nfs_list_entry(hdr
->pages
.next
);
822 nfs_list_remove_request(req
);
823 if (request_commit
) {
824 kref_get(&req
->wb_kref
);
825 nfs_mark_request_commit(req
, hdr
->lseg
, &cinfo
,
828 nfs_unlock_and_release_request(req
);
833 nfs_direct_write_complete(dreq
, hdr
->inode
);
837 static void nfs_write_sync_pgio_error(struct list_head
*head
)
839 struct nfs_page
*req
;
841 while (!list_empty(head
)) {
842 req
= nfs_list_entry(head
->next
);
843 nfs_list_remove_request(req
);
844 nfs_unlock_and_release_request(req
);
848 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header
*hdr
)
850 struct nfs_direct_req
*dreq
= hdr
->dreq
;
852 spin_lock(&dreq
->lock
);
853 if (dreq
->error
== 0) {
854 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
855 /* fake unstable write to let common nfs resend pages */
856 hdr
->verf
.committed
= NFS_UNSTABLE
;
857 hdr
->good_bytes
= hdr
->args
.count
;
859 spin_unlock(&dreq
->lock
);
862 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops
= {
863 .error_cleanup
= nfs_write_sync_pgio_error
,
864 .init_hdr
= nfs_direct_pgio_init
,
865 .completion
= nfs_direct_write_completion
,
866 .reschedule_io
= nfs_direct_write_reschedule_io
,
871 * NB: Return the value of the first error return code. Subsequent
872 * errors after the first one are ignored.
875 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
876 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
877 * bail and stop sending more writes. Write length accounting is
878 * handled automatically by nfs_direct_write_result(). Otherwise, if
879 * no requests have been sent, just return an error.
881 static ssize_t
nfs_direct_write_schedule_iovec(struct nfs_direct_req
*dreq
,
882 struct iov_iter
*iter
,
885 struct nfs_pageio_descriptor desc
;
886 struct inode
*inode
= dreq
->inode
;
888 size_t requested_bytes
= 0;
889 size_t wsize
= max_t(size_t, NFS_SERVER(inode
)->wsize
, PAGE_SIZE
);
891 nfs_pageio_init_write(&desc
, inode
, FLUSH_COND_STABLE
, false,
892 &nfs_direct_write_completion_ops
);
895 inode_dio_begin(inode
);
897 NFS_I(inode
)->write_io
+= iov_iter_count(iter
);
898 while (iov_iter_count(iter
)) {
899 struct page
**pagevec
;
904 result
= iov_iter_get_pages_alloc(iter
, &pagevec
,
910 iov_iter_advance(iter
, bytes
);
911 npages
= (result
+ pgbase
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
912 for (i
= 0; i
< npages
; i
++) {
913 struct nfs_page
*req
;
914 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
916 req
= nfs_create_request(dreq
->ctx
, pagevec
[i
], NULL
,
919 result
= PTR_ERR(req
);
923 nfs_direct_setup_mirroring(dreq
, &desc
, req
);
924 if (desc
.pg_error
< 0) {
925 nfs_free_request(req
);
926 result
= desc
.pg_error
;
930 nfs_lock_request(req
);
931 req
->wb_index
= pos
>> PAGE_SHIFT
;
932 req
->wb_offset
= pos
& ~PAGE_MASK
;
933 if (!nfs_pageio_add_request(&desc
, req
)) {
934 result
= desc
.pg_error
;
935 nfs_unlock_and_release_request(req
);
940 requested_bytes
+= req_len
;
942 dreq
->bytes_left
-= req_len
;
944 nfs_direct_release_pages(pagevec
, npages
);
949 nfs_pageio_complete(&desc
);
952 * If no bytes were started, return the error, and let the
953 * generic layer handle the completion.
955 if (requested_bytes
== 0) {
956 inode_dio_end(inode
);
957 nfs_direct_req_release(dreq
);
958 return result
< 0 ? result
: -EIO
;
962 nfs_direct_write_complete(dreq
, dreq
->inode
);
967 * nfs_file_direct_write - file direct write operation for NFS files
968 * @iocb: target I/O control block
969 * @iter: vector of user buffers from which to write data
971 * We use this function for direct writes instead of calling
972 * generic_file_aio_write() in order to avoid taking the inode
973 * semaphore and updating the i_size. The NFS server will set
974 * the new i_size and this client must read the updated size
975 * back into its cache. We let the server do generic write
976 * parameter checking and report problems.
978 * We eliminate local atime updates, see direct read above.
980 * We avoid unnecessary page cache invalidations for normal cached
981 * readers of this file.
983 * Note that O_APPEND is not supported for NFS direct writes, as there
984 * is no atomic O_APPEND write facility in the NFS protocol.
986 ssize_t
nfs_file_direct_write(struct kiocb
*iocb
, struct iov_iter
*iter
)
988 ssize_t result
= -EINVAL
;
989 struct file
*file
= iocb
->ki_filp
;
990 struct address_space
*mapping
= file
->f_mapping
;
991 struct inode
*inode
= mapping
->host
;
992 struct nfs_direct_req
*dreq
;
993 struct nfs_lock_context
*l_ctx
;
996 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
997 file
, iov_iter_count(iter
), (long long) iocb
->ki_pos
);
999 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTWRITTENBYTES
,
1000 iov_iter_count(iter
));
1003 end
= (pos
+ iov_iter_count(iter
) - 1) >> PAGE_SHIFT
;
1007 result
= nfs_sync_mapping(mapping
);
1011 if (mapping
->nrpages
) {
1012 result
= invalidate_inode_pages2_range(mapping
,
1013 pos
>> PAGE_SHIFT
, end
);
1018 task_io_account_write(iov_iter_count(iter
));
1021 dreq
= nfs_direct_req_alloc();
1025 dreq
->inode
= inode
;
1026 dreq
->bytes_left
= iov_iter_count(iter
);
1027 dreq
->io_start
= pos
;
1028 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
1029 l_ctx
= nfs_get_lock_context(dreq
->ctx
);
1030 if (IS_ERR(l_ctx
)) {
1031 result
= PTR_ERR(l_ctx
);
1034 dreq
->l_ctx
= l_ctx
;
1035 if (!is_sync_kiocb(iocb
))
1038 result
= nfs_direct_write_schedule_iovec(dreq
, iter
, pos
);
1040 if (mapping
->nrpages
) {
1041 invalidate_inode_pages2_range(mapping
,
1042 pos
>> PAGE_SHIFT
, end
);
1045 inode_unlock(inode
);
1048 result
= nfs_direct_wait(dreq
);
1050 struct inode
*inode
= mapping
->host
;
1052 iocb
->ki_pos
= pos
+ result
;
1053 spin_lock(&inode
->i_lock
);
1054 if (i_size_read(inode
) < iocb
->ki_pos
)
1055 i_size_write(inode
, iocb
->ki_pos
);
1056 spin_unlock(&inode
->i_lock
);
1058 /* XXX: should check the generic_write_sync retval */
1059 generic_write_sync(iocb
, result
);
1062 nfs_direct_req_release(dreq
);
1066 nfs_direct_req_release(dreq
);
1068 inode_unlock(inode
);
1073 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1076 int __init
nfs_init_directcache(void)
1078 nfs_direct_cachep
= kmem_cache_create("nfs_direct_cache",
1079 sizeof(struct nfs_direct_req
),
1080 0, (SLAB_RECLAIM_ACCOUNT
|
1083 if (nfs_direct_cachep
== NULL
)
1090 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1093 void nfs_destroy_directcache(void)
1095 kmem_cache_destroy(nfs_direct_cachep
);