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 max_count
, /* max expected count */
91 bytes_left
, /* bytes left to be sent */
92 io_start
, /* start of IO */
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
;
101 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
102 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
103 struct nfs_writeverf verf
; /* unstable write verifier */
106 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops
;
107 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops
;
108 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
, struct inode
*inode
);
109 static void nfs_direct_write_schedule_work(struct work_struct
*work
);
111 static inline void get_dreq(struct nfs_direct_req
*dreq
)
113 atomic_inc(&dreq
->io_count
);
116 static inline int put_dreq(struct nfs_direct_req
*dreq
)
118 return atomic_dec_and_test(&dreq
->io_count
);
122 nfs_direct_good_bytes(struct nfs_direct_req
*dreq
, struct nfs_pgio_header
*hdr
)
127 WARN_ON_ONCE(dreq
->count
>= dreq
->max_count
);
129 if (dreq
->mirror_count
== 1) {
130 dreq
->mirrors
[hdr
->pgio_mirror_idx
].count
+= hdr
->good_bytes
;
131 dreq
->count
+= hdr
->good_bytes
;
133 /* mirrored writes */
134 count
= dreq
->mirrors
[hdr
->pgio_mirror_idx
].count
;
135 if (count
+ dreq
->io_start
< hdr
->io_start
+ hdr
->good_bytes
) {
136 count
= hdr
->io_start
+ hdr
->good_bytes
- dreq
->io_start
;
137 dreq
->mirrors
[hdr
->pgio_mirror_idx
].count
= count
;
139 /* update the dreq->count by finding the minimum agreed count from all
141 count
= dreq
->mirrors
[0].count
;
143 for (i
= 1; i
< dreq
->mirror_count
; i
++)
144 count
= min(count
, dreq
->mirrors
[i
].count
);
151 * nfs_direct_select_verf - select the right verifier
152 * @dreq - direct request possibly spanning multiple servers
153 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
154 * @commit_idx - commit bucket index for the DS
156 * returns the correct verifier to use given the role of the server
158 static struct nfs_writeverf
*
159 nfs_direct_select_verf(struct nfs_direct_req
*dreq
,
160 struct nfs_client
*ds_clp
,
163 struct nfs_writeverf
*verfp
= &dreq
->verf
;
165 #ifdef CONFIG_NFS_V4_1
167 * pNFS is in use, use the DS verf except commit_through_mds is set
168 * for layout segment where nbuckets is zero.
170 if (ds_clp
&& dreq
->ds_cinfo
.nbuckets
> 0) {
171 if (commit_idx
>= 0 && commit_idx
< dreq
->ds_cinfo
.nbuckets
)
172 verfp
= &dreq
->ds_cinfo
.buckets
[commit_idx
].direct_verf
;
182 * nfs_direct_set_hdr_verf - set the write/commit verifier
183 * @dreq - direct request possibly spanning multiple servers
184 * @hdr - pageio header to validate against previously seen verfs
186 * Set the server's (MDS or DS) "seen" verifier
188 static void nfs_direct_set_hdr_verf(struct nfs_direct_req
*dreq
,
189 struct nfs_pgio_header
*hdr
)
191 struct nfs_writeverf
*verfp
;
193 verfp
= nfs_direct_select_verf(dreq
, hdr
->ds_clp
, hdr
->ds_commit_idx
);
194 WARN_ON_ONCE(verfp
->committed
>= 0);
195 memcpy(verfp
, &hdr
->verf
, sizeof(struct nfs_writeverf
));
196 WARN_ON_ONCE(verfp
->committed
< 0);
200 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
201 * @dreq - direct request possibly spanning multiple servers
202 * @hdr - pageio header to validate against previously seen verf
204 * set the server's "seen" verf if not initialized.
205 * returns result of comparison between @hdr->verf and the "seen"
206 * verf of the server used by @hdr (DS or MDS)
208 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req
*dreq
,
209 struct nfs_pgio_header
*hdr
)
211 struct nfs_writeverf
*verfp
;
213 verfp
= nfs_direct_select_verf(dreq
, hdr
->ds_clp
, hdr
->ds_commit_idx
);
214 if (verfp
->committed
< 0) {
215 nfs_direct_set_hdr_verf(dreq
, hdr
);
218 return memcmp(verfp
, &hdr
->verf
, sizeof(struct nfs_writeverf
));
222 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
223 * @dreq - direct request possibly spanning multiple servers
224 * @data - commit data to validate against previously seen verf
226 * returns result of comparison between @data->verf and the verf of
227 * the server used by @data (DS or MDS)
229 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req
*dreq
,
230 struct nfs_commit_data
*data
)
232 struct nfs_writeverf
*verfp
;
234 verfp
= nfs_direct_select_verf(dreq
, data
->ds_clp
,
235 data
->ds_commit_index
);
237 /* verifier not set so always fail */
238 if (verfp
->committed
< 0)
241 return memcmp(verfp
, &data
->verf
, sizeof(struct nfs_writeverf
));
245 * nfs_direct_IO - NFS address space operation for direct I/O
246 * @iocb: target I/O control block
247 * @iov: array of vectors that define I/O buffer
248 * @pos: offset in file to begin the operation
249 * @nr_segs: size of iovec array
251 * The presence of this routine in the address space ops vector means
252 * the NFS client supports direct I/O. However, for most direct IO, we
253 * shunt off direct read and write requests before the VFS gets them,
254 * so this method is only ever called for swap.
256 ssize_t
nfs_direct_IO(struct kiocb
*iocb
, struct iov_iter
*iter
)
258 struct inode
*inode
= iocb
->ki_filp
->f_mapping
->host
;
260 /* we only support swap file calling nfs_direct_IO */
261 if (!IS_SWAPFILE(inode
))
264 VM_BUG_ON(iov_iter_count(iter
) != PAGE_SIZE
);
266 if (iov_iter_rw(iter
) == READ
)
267 return nfs_file_direct_read(iocb
, iter
);
268 return nfs_file_direct_write(iocb
, iter
);
271 static void nfs_direct_release_pages(struct page
**pages
, unsigned int npages
)
274 for (i
= 0; i
< npages
; i
++)
278 void nfs_init_cinfo_from_dreq(struct nfs_commit_info
*cinfo
,
279 struct nfs_direct_req
*dreq
)
281 cinfo
->inode
= dreq
->inode
;
282 cinfo
->mds
= &dreq
->mds_cinfo
;
283 cinfo
->ds
= &dreq
->ds_cinfo
;
285 cinfo
->completion_ops
= &nfs_direct_commit_completion_ops
;
288 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req
*dreq
,
289 struct nfs_pageio_descriptor
*pgio
,
290 struct nfs_page
*req
)
292 int mirror_count
= 1;
294 if (pgio
->pg_ops
->pg_get_mirror_count
)
295 mirror_count
= pgio
->pg_ops
->pg_get_mirror_count(pgio
, req
);
297 dreq
->mirror_count
= mirror_count
;
300 static inline struct nfs_direct_req
*nfs_direct_req_alloc(void)
302 struct nfs_direct_req
*dreq
;
304 dreq
= kmem_cache_zalloc(nfs_direct_cachep
, GFP_KERNEL
);
308 kref_init(&dreq
->kref
);
309 kref_get(&dreq
->kref
);
310 init_completion(&dreq
->completion
);
311 INIT_LIST_HEAD(&dreq
->mds_cinfo
.list
);
312 dreq
->verf
.committed
= NFS_INVALID_STABLE_HOW
; /* not set yet */
313 INIT_WORK(&dreq
->work
, nfs_direct_write_schedule_work
);
314 dreq
->mirror_count
= 1;
315 spin_lock_init(&dreq
->lock
);
320 static void nfs_direct_req_free(struct kref
*kref
)
322 struct nfs_direct_req
*dreq
= container_of(kref
, struct nfs_direct_req
, kref
);
324 nfs_free_pnfs_ds_cinfo(&dreq
->ds_cinfo
);
325 if (dreq
->l_ctx
!= NULL
)
326 nfs_put_lock_context(dreq
->l_ctx
);
327 if (dreq
->ctx
!= NULL
)
328 put_nfs_open_context(dreq
->ctx
);
329 kmem_cache_free(nfs_direct_cachep
, dreq
);
332 static void nfs_direct_req_release(struct nfs_direct_req
*dreq
)
334 kref_put(&dreq
->kref
, nfs_direct_req_free
);
337 ssize_t
nfs_dreq_bytes_left(struct nfs_direct_req
*dreq
)
339 return dreq
->bytes_left
;
341 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left
);
344 * Collects and returns the final error value/byte-count.
346 static ssize_t
nfs_direct_wait(struct nfs_direct_req
*dreq
)
348 ssize_t result
= -EIOCBQUEUED
;
350 /* Async requests don't wait here */
354 result
= wait_for_completion_killable(&dreq
->completion
);
357 result
= dreq
->error
;
359 result
= dreq
->count
;
362 return (ssize_t
) result
;
366 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
367 * the iocb is still valid here if this is a synchronous request.
369 static void nfs_direct_complete(struct nfs_direct_req
*dreq
, bool write
)
371 struct inode
*inode
= dreq
->inode
;
373 if (dreq
->iocb
&& write
) {
374 loff_t pos
= dreq
->iocb
->ki_pos
+ dreq
->count
;
376 spin_lock(&inode
->i_lock
);
377 if (i_size_read(inode
) < pos
)
378 i_size_write(inode
, pos
);
379 spin_unlock(&inode
->i_lock
);
383 nfs_zap_mapping(inode
, inode
->i_mapping
);
385 inode_dio_end(inode
);
388 long res
= (long) dreq
->error
;
390 res
= (long) dreq
->count
;
391 dreq
->iocb
->ki_complete(dreq
->iocb
, res
, 0);
394 complete_all(&dreq
->completion
);
396 nfs_direct_req_release(dreq
);
399 static void nfs_direct_readpage_release(struct nfs_page
*req
)
401 dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
402 req
->wb_context
->dentry
->d_sb
->s_id
,
403 (unsigned long long)NFS_FILEID(d_inode(req
->wb_context
->dentry
)),
405 (long long)req_offset(req
));
406 nfs_release_request(req
);
409 static void nfs_direct_read_completion(struct nfs_pgio_header
*hdr
)
411 unsigned long bytes
= 0;
412 struct nfs_direct_req
*dreq
= hdr
->dreq
;
414 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
417 spin_lock(&dreq
->lock
);
418 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
) && (hdr
->good_bytes
== 0))
419 dreq
->error
= hdr
->error
;
421 nfs_direct_good_bytes(dreq
, hdr
);
423 spin_unlock(&dreq
->lock
);
425 while (!list_empty(&hdr
->pages
)) {
426 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
427 struct page
*page
= req
->wb_page
;
429 if (!PageCompound(page
) && bytes
< hdr
->good_bytes
)
430 set_page_dirty(page
);
431 bytes
+= req
->wb_bytes
;
432 nfs_list_remove_request(req
);
433 nfs_direct_readpage_release(req
);
437 nfs_direct_complete(dreq
, false);
441 static void nfs_read_sync_pgio_error(struct list_head
*head
)
443 struct nfs_page
*req
;
445 while (!list_empty(head
)) {
446 req
= nfs_list_entry(head
->next
);
447 nfs_list_remove_request(req
);
448 nfs_release_request(req
);
452 static void nfs_direct_pgio_init(struct nfs_pgio_header
*hdr
)
457 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops
= {
458 .error_cleanup
= nfs_read_sync_pgio_error
,
459 .init_hdr
= nfs_direct_pgio_init
,
460 .completion
= nfs_direct_read_completion
,
464 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
465 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
466 * bail and stop sending more reads. Read length accounting is
467 * handled automatically by nfs_direct_read_result(). Otherwise, if
468 * no requests have been sent, just return an error.
471 static ssize_t
nfs_direct_read_schedule_iovec(struct nfs_direct_req
*dreq
,
472 struct iov_iter
*iter
,
475 struct nfs_pageio_descriptor desc
;
476 struct inode
*inode
= dreq
->inode
;
477 ssize_t result
= -EINVAL
;
478 size_t requested_bytes
= 0;
479 size_t rsize
= max_t(size_t, NFS_SERVER(inode
)->rsize
, PAGE_SIZE
);
481 nfs_pageio_init_read(&desc
, dreq
->inode
, false,
482 &nfs_direct_read_completion_ops
);
485 inode_dio_begin(inode
);
487 while (iov_iter_count(iter
)) {
488 struct page
**pagevec
;
493 result
= iov_iter_get_pages_alloc(iter
, &pagevec
,
499 iov_iter_advance(iter
, bytes
);
500 npages
= (result
+ pgbase
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
501 for (i
= 0; i
< npages
; i
++) {
502 struct nfs_page
*req
;
503 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
504 /* XXX do we need to do the eof zeroing found in async_filler? */
505 req
= nfs_create_request(dreq
->ctx
, pagevec
[i
], NULL
,
508 result
= PTR_ERR(req
);
511 req
->wb_index
= pos
>> PAGE_SHIFT
;
512 req
->wb_offset
= pos
& ~PAGE_MASK
;
513 if (!nfs_pageio_add_request(&desc
, req
)) {
514 result
= desc
.pg_error
;
515 nfs_release_request(req
);
520 requested_bytes
+= req_len
;
522 dreq
->bytes_left
-= req_len
;
524 nfs_direct_release_pages(pagevec
, npages
);
530 nfs_pageio_complete(&desc
);
533 * If no bytes were started, return the error, and let the
534 * generic layer handle the completion.
536 if (requested_bytes
== 0) {
537 inode_dio_end(inode
);
538 nfs_direct_req_release(dreq
);
539 return result
< 0 ? result
: -EIO
;
543 nfs_direct_complete(dreq
, false);
548 * nfs_file_direct_read - file direct read operation for NFS files
549 * @iocb: target I/O control block
550 * @iter: vector of user buffers into which to read data
552 * We use this function for direct reads instead of calling
553 * generic_file_aio_read() in order to avoid gfar's check to see if
554 * the request starts before the end of the file. For that check
555 * to work, we must generate a GETATTR before each direct read, and
556 * even then there is a window between the GETATTR and the subsequent
557 * READ where the file size could change. Our preference is simply
558 * to do all reads the application wants, and the server will take
559 * care of managing the end of file boundary.
561 * This function also eliminates unnecessarily updating the file's
562 * atime locally, as the NFS server sets the file's atime, and this
563 * client must read the updated atime from the server back into its
566 ssize_t
nfs_file_direct_read(struct kiocb
*iocb
, struct iov_iter
*iter
)
568 struct file
*file
= iocb
->ki_filp
;
569 struct address_space
*mapping
= file
->f_mapping
;
570 struct inode
*inode
= mapping
->host
;
571 struct nfs_direct_req
*dreq
;
572 struct nfs_lock_context
*l_ctx
;
573 ssize_t result
= -EINVAL
;
574 size_t count
= iov_iter_count(iter
);
575 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTREADBYTES
, count
);
577 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
578 file
, count
, (long long) iocb
->ki_pos
);
585 result
= nfs_sync_mapping(mapping
);
589 task_io_account_read(count
);
592 dreq
= nfs_direct_req_alloc();
597 dreq
->bytes_left
= dreq
->max_count
= count
;
598 dreq
->io_start
= iocb
->ki_pos
;
599 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
600 l_ctx
= nfs_get_lock_context(dreq
->ctx
);
602 result
= PTR_ERR(l_ctx
);
606 if (!is_sync_kiocb(iocb
))
609 NFS_I(inode
)->read_io
+= count
;
610 result
= nfs_direct_read_schedule_iovec(dreq
, iter
, iocb
->ki_pos
);
615 result
= nfs_direct_wait(dreq
);
617 iocb
->ki_pos
+= result
;
620 nfs_direct_req_release(dreq
);
624 nfs_direct_req_release(dreq
);
632 nfs_direct_write_scan_commit_list(struct inode
*inode
,
633 struct list_head
*list
,
634 struct nfs_commit_info
*cinfo
)
636 spin_lock(&cinfo
->inode
->i_lock
);
637 #ifdef CONFIG_NFS_V4_1
638 if (cinfo
->ds
!= NULL
&& cinfo
->ds
->nwritten
!= 0)
639 NFS_SERVER(inode
)->pnfs_curr_ld
->recover_commit_reqs(list
, cinfo
);
641 nfs_scan_commit_list(&cinfo
->mds
->list
, list
, cinfo
, 0);
642 spin_unlock(&cinfo
->inode
->i_lock
);
645 static void nfs_direct_write_reschedule(struct nfs_direct_req
*dreq
)
647 struct nfs_pageio_descriptor desc
;
648 struct nfs_page
*req
, *tmp
;
650 struct nfs_commit_info cinfo
;
654 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
655 nfs_direct_write_scan_commit_list(dreq
->inode
, &reqs
, &cinfo
);
658 for (i
= 0; i
< dreq
->mirror_count
; i
++)
659 dreq
->mirrors
[i
].count
= 0;
662 nfs_pageio_init_write(&desc
, dreq
->inode
, FLUSH_STABLE
, false,
663 &nfs_direct_write_completion_ops
);
666 req
= nfs_list_entry(reqs
.next
);
667 nfs_direct_setup_mirroring(dreq
, &desc
, req
);
668 if (desc
.pg_error
< 0) {
669 list_splice_init(&reqs
, &failed
);
673 list_for_each_entry_safe(req
, tmp
, &reqs
, wb_list
) {
674 if (!nfs_pageio_add_request(&desc
, req
)) {
675 nfs_list_remove_request(req
);
676 nfs_list_add_request(req
, &failed
);
677 spin_lock(&cinfo
.inode
->i_lock
);
679 if (desc
.pg_error
< 0)
680 dreq
->error
= desc
.pg_error
;
683 spin_unlock(&cinfo
.inode
->i_lock
);
685 nfs_release_request(req
);
687 nfs_pageio_complete(&desc
);
690 while (!list_empty(&failed
)) {
691 req
= nfs_list_entry(failed
.next
);
692 nfs_list_remove_request(req
);
693 nfs_unlock_and_release_request(req
);
697 nfs_direct_write_complete(dreq
, dreq
->inode
);
700 static void nfs_direct_commit_complete(struct nfs_commit_data
*data
)
702 struct nfs_direct_req
*dreq
= data
->dreq
;
703 struct nfs_commit_info cinfo
;
704 struct nfs_page
*req
;
705 int status
= data
->task
.tk_status
;
707 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
709 dprintk("NFS: %5u commit failed with error %d.\n",
710 data
->task
.tk_pid
, status
);
711 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
712 } else if (nfs_direct_cmp_commit_data_verf(dreq
, data
)) {
713 dprintk("NFS: %5u commit verify failed\n", data
->task
.tk_pid
);
714 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
717 dprintk("NFS: %5u commit returned %d\n", data
->task
.tk_pid
, status
);
718 while (!list_empty(&data
->pages
)) {
719 req
= nfs_list_entry(data
->pages
.next
);
720 nfs_list_remove_request(req
);
721 if (dreq
->flags
== NFS_ODIRECT_RESCHED_WRITES
) {
722 /* Note the rewrite will go through mds */
723 nfs_mark_request_commit(req
, NULL
, &cinfo
, 0);
725 nfs_release_request(req
);
726 nfs_unlock_and_release_request(req
);
729 if (atomic_dec_and_test(&cinfo
.mds
->rpcs_out
))
730 nfs_direct_write_complete(dreq
, data
->inode
);
733 static void nfs_direct_resched_write(struct nfs_commit_info
*cinfo
,
734 struct nfs_page
*req
)
736 struct nfs_direct_req
*dreq
= cinfo
->dreq
;
738 spin_lock(&dreq
->lock
);
739 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
740 spin_unlock(&dreq
->lock
);
741 nfs_mark_request_commit(req
, NULL
, cinfo
, 0);
744 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops
= {
745 .completion
= nfs_direct_commit_complete
,
746 .resched_write
= nfs_direct_resched_write
,
749 static void nfs_direct_commit_schedule(struct nfs_direct_req
*dreq
)
752 struct nfs_commit_info cinfo
;
755 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
756 nfs_scan_commit(dreq
->inode
, &mds_list
, &cinfo
);
757 res
= nfs_generic_commit_list(dreq
->inode
, &mds_list
, 0, &cinfo
);
758 if (res
< 0) /* res == -ENOMEM */
759 nfs_direct_write_reschedule(dreq
);
762 static void nfs_direct_write_schedule_work(struct work_struct
*work
)
764 struct nfs_direct_req
*dreq
= container_of(work
, struct nfs_direct_req
, work
);
765 int flags
= dreq
->flags
;
769 case NFS_ODIRECT_DO_COMMIT
:
770 nfs_direct_commit_schedule(dreq
);
772 case NFS_ODIRECT_RESCHED_WRITES
:
773 nfs_direct_write_reschedule(dreq
);
776 nfs_direct_complete(dreq
, true);
780 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
, struct inode
*inode
)
782 schedule_work(&dreq
->work
); /* Calls nfs_direct_write_schedule_work */
785 static void nfs_direct_write_completion(struct nfs_pgio_header
*hdr
)
787 struct nfs_direct_req
*dreq
= hdr
->dreq
;
788 struct nfs_commit_info cinfo
;
789 bool request_commit
= false;
790 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
792 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
795 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
797 spin_lock(&dreq
->lock
);
799 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
)) {
801 dreq
->error
= hdr
->error
;
803 if (dreq
->error
== 0) {
804 nfs_direct_good_bytes(dreq
, hdr
);
805 if (nfs_write_need_commit(hdr
)) {
806 if (dreq
->flags
== NFS_ODIRECT_RESCHED_WRITES
)
807 request_commit
= true;
808 else if (dreq
->flags
== 0) {
809 nfs_direct_set_hdr_verf(dreq
, hdr
);
810 request_commit
= true;
811 dreq
->flags
= NFS_ODIRECT_DO_COMMIT
;
812 } else if (dreq
->flags
== NFS_ODIRECT_DO_COMMIT
) {
813 request_commit
= true;
814 if (nfs_direct_set_or_cmp_hdr_verf(dreq
, hdr
))
816 NFS_ODIRECT_RESCHED_WRITES
;
820 spin_unlock(&dreq
->lock
);
822 while (!list_empty(&hdr
->pages
)) {
824 req
= nfs_list_entry(hdr
->pages
.next
);
825 nfs_list_remove_request(req
);
826 if (request_commit
) {
827 kref_get(&req
->wb_kref
);
828 nfs_mark_request_commit(req
, hdr
->lseg
, &cinfo
,
831 nfs_unlock_and_release_request(req
);
836 nfs_direct_write_complete(dreq
, hdr
->inode
);
840 static void nfs_write_sync_pgio_error(struct list_head
*head
)
842 struct nfs_page
*req
;
844 while (!list_empty(head
)) {
845 req
= nfs_list_entry(head
->next
);
846 nfs_list_remove_request(req
);
847 nfs_unlock_and_release_request(req
);
851 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header
*hdr
)
853 struct nfs_direct_req
*dreq
= hdr
->dreq
;
855 spin_lock(&dreq
->lock
);
856 if (dreq
->error
== 0) {
857 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
858 /* fake unstable write to let common nfs resend pages */
859 hdr
->verf
.committed
= NFS_UNSTABLE
;
860 hdr
->good_bytes
= hdr
->args
.count
;
862 spin_unlock(&dreq
->lock
);
865 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops
= {
866 .error_cleanup
= nfs_write_sync_pgio_error
,
867 .init_hdr
= nfs_direct_pgio_init
,
868 .completion
= nfs_direct_write_completion
,
869 .reschedule_io
= nfs_direct_write_reschedule_io
,
874 * NB: Return the value of the first error return code. Subsequent
875 * errors after the first one are ignored.
878 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
879 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
880 * bail and stop sending more writes. Write length accounting is
881 * handled automatically by nfs_direct_write_result(). Otherwise, if
882 * no requests have been sent, just return an error.
884 static ssize_t
nfs_direct_write_schedule_iovec(struct nfs_direct_req
*dreq
,
885 struct iov_iter
*iter
,
888 struct nfs_pageio_descriptor desc
;
889 struct inode
*inode
= dreq
->inode
;
891 size_t requested_bytes
= 0;
892 size_t wsize
= max_t(size_t, NFS_SERVER(inode
)->wsize
, PAGE_SIZE
);
894 nfs_pageio_init_write(&desc
, inode
, FLUSH_COND_STABLE
, false,
895 &nfs_direct_write_completion_ops
);
898 inode_dio_begin(inode
);
900 NFS_I(inode
)->write_io
+= iov_iter_count(iter
);
901 while (iov_iter_count(iter
)) {
902 struct page
**pagevec
;
907 result
= iov_iter_get_pages_alloc(iter
, &pagevec
,
913 iov_iter_advance(iter
, bytes
);
914 npages
= (result
+ pgbase
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
915 for (i
= 0; i
< npages
; i
++) {
916 struct nfs_page
*req
;
917 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
919 req
= nfs_create_request(dreq
->ctx
, pagevec
[i
], NULL
,
922 result
= PTR_ERR(req
);
926 nfs_direct_setup_mirroring(dreq
, &desc
, req
);
927 if (desc
.pg_error
< 0) {
928 nfs_free_request(req
);
929 result
= desc
.pg_error
;
933 nfs_lock_request(req
);
934 req
->wb_index
= pos
>> PAGE_SHIFT
;
935 req
->wb_offset
= pos
& ~PAGE_MASK
;
936 if (!nfs_pageio_add_request(&desc
, req
)) {
937 result
= desc
.pg_error
;
938 nfs_unlock_and_release_request(req
);
943 requested_bytes
+= req_len
;
945 dreq
->bytes_left
-= req_len
;
947 nfs_direct_release_pages(pagevec
, npages
);
952 nfs_pageio_complete(&desc
);
955 * If no bytes were started, return the error, and let the
956 * generic layer handle the completion.
958 if (requested_bytes
== 0) {
959 inode_dio_end(inode
);
960 nfs_direct_req_release(dreq
);
961 return result
< 0 ? result
: -EIO
;
965 nfs_direct_write_complete(dreq
, dreq
->inode
);
970 * nfs_file_direct_write - file direct write operation for NFS files
971 * @iocb: target I/O control block
972 * @iter: vector of user buffers from which to write data
974 * We use this function for direct writes instead of calling
975 * generic_file_aio_write() in order to avoid taking the inode
976 * semaphore and updating the i_size. The NFS server will set
977 * the new i_size and this client must read the updated size
978 * back into its cache. We let the server do generic write
979 * parameter checking and report problems.
981 * We eliminate local atime updates, see direct read above.
983 * We avoid unnecessary page cache invalidations for normal cached
984 * readers of this file.
986 * Note that O_APPEND is not supported for NFS direct writes, as there
987 * is no atomic O_APPEND write facility in the NFS protocol.
989 ssize_t
nfs_file_direct_write(struct kiocb
*iocb
, struct iov_iter
*iter
)
991 ssize_t result
= -EINVAL
;
992 struct file
*file
= iocb
->ki_filp
;
993 struct address_space
*mapping
= file
->f_mapping
;
994 struct inode
*inode
= mapping
->host
;
995 struct nfs_direct_req
*dreq
;
996 struct nfs_lock_context
*l_ctx
;
999 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
1000 file
, iov_iter_count(iter
), (long long) iocb
->ki_pos
);
1002 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTWRITTENBYTES
,
1003 iov_iter_count(iter
));
1006 end
= (pos
+ iov_iter_count(iter
) - 1) >> PAGE_SHIFT
;
1010 result
= nfs_sync_mapping(mapping
);
1014 if (mapping
->nrpages
) {
1015 result
= invalidate_inode_pages2_range(mapping
,
1016 pos
>> PAGE_SHIFT
, end
);
1021 task_io_account_write(iov_iter_count(iter
));
1024 dreq
= nfs_direct_req_alloc();
1028 dreq
->inode
= inode
;
1029 dreq
->bytes_left
= dreq
->max_count
= iov_iter_count(iter
);
1030 dreq
->io_start
= pos
;
1031 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
1032 l_ctx
= nfs_get_lock_context(dreq
->ctx
);
1033 if (IS_ERR(l_ctx
)) {
1034 result
= PTR_ERR(l_ctx
);
1037 dreq
->l_ctx
= l_ctx
;
1038 if (!is_sync_kiocb(iocb
))
1041 result
= nfs_direct_write_schedule_iovec(dreq
, iter
, pos
);
1043 if (mapping
->nrpages
) {
1044 invalidate_inode_pages2_range(mapping
,
1045 pos
>> PAGE_SHIFT
, end
);
1048 inode_unlock(inode
);
1051 result
= nfs_direct_wait(dreq
);
1053 struct inode
*inode
= mapping
->host
;
1055 iocb
->ki_pos
= pos
+ result
;
1056 spin_lock(&inode
->i_lock
);
1057 if (i_size_read(inode
) < iocb
->ki_pos
)
1058 i_size_write(inode
, iocb
->ki_pos
);
1059 spin_unlock(&inode
->i_lock
);
1061 /* XXX: should check the generic_write_sync retval */
1062 generic_write_sync(iocb
, result
);
1065 nfs_direct_req_release(dreq
);
1069 nfs_direct_req_release(dreq
);
1071 inode_unlock(inode
);
1076 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1079 int __init
nfs_init_directcache(void)
1081 nfs_direct_cachep
= kmem_cache_create("nfs_direct_cache",
1082 sizeof(struct nfs_direct_req
),
1083 0, (SLAB_RECLAIM_ACCOUNT
|
1086 if (nfs_direct_cachep
== NULL
)
1093 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1096 void nfs_destroy_directcache(void)
1098 kmem_cache_destroy(nfs_direct_cachep
);