1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/fs/nfs/direct.c
5 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
7 * High-performance uncached I/O for the Linux NFS client
9 * There are important applications whose performance or correctness
10 * depends on uncached access to file data. Database clusters
11 * (multiple copies of the same instance running on separate hosts)
12 * implement their own cache coherency protocol that subsumes file
13 * system cache protocols. Applications that process datasets
14 * considerably larger than the client's memory do not always benefit
15 * from a local cache. A streaming video server, for instance, has no
16 * need to cache the contents of a file.
18 * When an application requests uncached I/O, all read and write requests
19 * are made directly to the server; data stored or fetched via these
20 * requests is not cached in the Linux page cache. The client does not
21 * correct unaligned requests from applications. All requested bytes are
22 * held on permanent storage before a direct write system call returns to
25 * Solaris implements an uncached I/O facility called directio() that
26 * is used for backups and sequential I/O to very large files. Solaris
27 * also supports uncaching whole NFS partitions with "-o forcedirectio,"
28 * an undocumented mount option.
30 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
31 * help from Andrew Morton.
33 * 18 Dec 2001 Initial implementation for 2.4 --cel
34 * 08 Jul 2002 Version for 2.4.19, with bug fixes --trondmy
35 * 08 Jun 2003 Port to 2.5 APIs --cel
36 * 31 Mar 2004 Handle direct I/O without VFS support --cel
37 * 15 Sep 2004 Parallel async reads --cel
38 * 04 May 2005 support O_DIRECT with aio --cel
42 #include <linux/errno.h>
43 #include <linux/sched.h>
44 #include <linux/kernel.h>
45 #include <linux/file.h>
46 #include <linux/pagemap.h>
47 #include <linux/kref.h>
48 #include <linux/slab.h>
49 #include <linux/task_io_accounting_ops.h>
50 #include <linux/module.h>
52 #include <linux/nfs_fs.h>
53 #include <linux/nfs_page.h>
54 #include <linux/sunrpc/clnt.h>
56 #include <linux/uaccess.h>
57 #include <linux/atomic.h>
63 #define NFSDBG_FACILITY NFSDBG_VFS
65 static struct kmem_cache
*nfs_direct_cachep
;
67 struct nfs_direct_req
{
68 struct kref kref
; /* release manager */
71 struct nfs_open_context
*ctx
; /* file open context info */
72 struct nfs_lock_context
*l_ctx
; /* Lock context info */
73 struct kiocb
* iocb
; /* controlling i/o request */
74 struct inode
* inode
; /* target file of i/o */
76 /* completion state */
77 atomic_t io_count
; /* i/os we're waiting for */
78 spinlock_t lock
; /* protect completion state */
80 loff_t io_start
; /* Start offset for I/O */
81 ssize_t count
, /* bytes actually processed */
82 max_count
, /* max expected count */
83 bytes_left
, /* bytes left to be sent */
84 error
; /* any reported error */
85 struct completion completion
; /* wait for i/o completion */
88 struct nfs_mds_commit_info mds_cinfo
; /* Storage for cinfo */
89 struct pnfs_ds_commit_info ds_cinfo
; /* Storage for cinfo */
90 struct work_struct work
;
93 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
94 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
96 #define NFS_ODIRECT_SHOULD_DIRTY (3) /* dirty user-space page after read */
97 struct nfs_writeverf verf
; /* unstable write verifier */
100 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops
;
101 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops
;
102 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
);
103 static void nfs_direct_write_schedule_work(struct work_struct
*work
);
105 static inline void get_dreq(struct nfs_direct_req
*dreq
)
107 atomic_inc(&dreq
->io_count
);
110 static inline int put_dreq(struct nfs_direct_req
*dreq
)
112 return atomic_dec_and_test(&dreq
->io_count
);
116 nfs_direct_handle_truncated(struct nfs_direct_req
*dreq
,
117 const struct nfs_pgio_header
*hdr
,
120 if (!(test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
) ||
121 test_bit(NFS_IOHDR_EOF
, &hdr
->flags
)))
123 if (dreq
->max_count
>= dreq_len
) {
124 dreq
->max_count
= dreq_len
;
125 if (dreq
->count
> dreq_len
)
126 dreq
->count
= dreq_len
;
128 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
))
129 dreq
->error
= hdr
->error
;
130 else /* Clear outstanding error if this is EOF */
136 nfs_direct_count_bytes(struct nfs_direct_req
*dreq
,
137 const struct nfs_pgio_header
*hdr
)
139 loff_t hdr_end
= hdr
->io_start
+ hdr
->good_bytes
;
140 ssize_t dreq_len
= 0;
142 if (hdr_end
> dreq
->io_start
)
143 dreq_len
= hdr_end
- dreq
->io_start
;
145 nfs_direct_handle_truncated(dreq
, hdr
, dreq_len
);
147 if (dreq_len
> dreq
->max_count
)
148 dreq_len
= dreq
->max_count
;
150 if (dreq
->count
< dreq_len
)
151 dreq
->count
= dreq_len
;
155 * nfs_direct_select_verf - select the right verifier
156 * @dreq - direct request possibly spanning multiple servers
157 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
158 * @commit_idx - commit bucket index for the DS
160 * returns the correct verifier to use given the role of the server
162 static struct nfs_writeverf
*
163 nfs_direct_select_verf(struct nfs_direct_req
*dreq
,
164 struct nfs_client
*ds_clp
,
167 struct nfs_writeverf
*verfp
= &dreq
->verf
;
169 #ifdef CONFIG_NFS_V4_1
171 * pNFS is in use, use the DS verf except commit_through_mds is set
172 * for layout segment where nbuckets is zero.
174 if (ds_clp
&& dreq
->ds_cinfo
.nbuckets
> 0) {
175 if (commit_idx
>= 0 && commit_idx
< dreq
->ds_cinfo
.nbuckets
)
176 verfp
= &dreq
->ds_cinfo
.buckets
[commit_idx
].direct_verf
;
186 * nfs_direct_set_hdr_verf - set the write/commit verifier
187 * @dreq - direct request possibly spanning multiple servers
188 * @hdr - pageio header to validate against previously seen verfs
190 * Set the server's (MDS or DS) "seen" verifier
192 static void nfs_direct_set_hdr_verf(struct nfs_direct_req
*dreq
,
193 struct nfs_pgio_header
*hdr
)
195 struct nfs_writeverf
*verfp
;
197 verfp
= nfs_direct_select_verf(dreq
, hdr
->ds_clp
, hdr
->ds_commit_idx
);
198 WARN_ON_ONCE(verfp
->committed
>= 0);
199 memcpy(verfp
, &hdr
->verf
, sizeof(struct nfs_writeverf
));
200 WARN_ON_ONCE(verfp
->committed
< 0);
203 static int nfs_direct_cmp_verf(const struct nfs_writeverf
*v1
,
204 const struct nfs_writeverf
*v2
)
206 return nfs_write_verifier_cmp(&v1
->verifier
, &v2
->verifier
);
210 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
211 * @dreq - direct request possibly spanning multiple servers
212 * @hdr - pageio header to validate against previously seen verf
214 * set the server's "seen" verf if not initialized.
215 * returns result of comparison between @hdr->verf and the "seen"
216 * verf of the server used by @hdr (DS or MDS)
218 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req
*dreq
,
219 struct nfs_pgio_header
*hdr
)
221 struct nfs_writeverf
*verfp
;
223 verfp
= nfs_direct_select_verf(dreq
, hdr
->ds_clp
, hdr
->ds_commit_idx
);
224 if (verfp
->committed
< 0) {
225 nfs_direct_set_hdr_verf(dreq
, hdr
);
228 return nfs_direct_cmp_verf(verfp
, &hdr
->verf
);
232 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
233 * @dreq - direct request possibly spanning multiple servers
234 * @data - commit data to validate against previously seen verf
236 * returns result of comparison between @data->verf and the verf of
237 * the server used by @data (DS or MDS)
239 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req
*dreq
,
240 struct nfs_commit_data
*data
)
242 struct nfs_writeverf
*verfp
;
244 verfp
= nfs_direct_select_verf(dreq
, data
->ds_clp
,
245 data
->ds_commit_index
);
247 /* verifier not set so always fail */
248 if (verfp
->committed
< 0 || data
->res
.verf
->committed
<= NFS_UNSTABLE
)
251 return nfs_direct_cmp_verf(verfp
, data
->res
.verf
);
255 * nfs_direct_IO - NFS address space operation for direct I/O
256 * @iocb: target I/O control block
259 * The presence of this routine in the address space ops vector means
260 * the NFS client supports direct I/O. However, for most direct IO, we
261 * shunt off direct read and write requests before the VFS gets them,
262 * so this method is only ever called for swap.
264 ssize_t
nfs_direct_IO(struct kiocb
*iocb
, struct iov_iter
*iter
)
266 struct inode
*inode
= iocb
->ki_filp
->f_mapping
->host
;
268 /* we only support swap file calling nfs_direct_IO */
269 if (!IS_SWAPFILE(inode
))
272 VM_BUG_ON(iov_iter_count(iter
) != PAGE_SIZE
);
274 if (iov_iter_rw(iter
) == READ
)
275 return nfs_file_direct_read(iocb
, iter
);
276 return nfs_file_direct_write(iocb
, iter
);
279 static void nfs_direct_release_pages(struct page
**pages
, unsigned int npages
)
282 for (i
= 0; i
< npages
; i
++)
286 void nfs_init_cinfo_from_dreq(struct nfs_commit_info
*cinfo
,
287 struct nfs_direct_req
*dreq
)
289 cinfo
->inode
= dreq
->inode
;
290 cinfo
->mds
= &dreq
->mds_cinfo
;
291 cinfo
->ds
= &dreq
->ds_cinfo
;
293 cinfo
->completion_ops
= &nfs_direct_commit_completion_ops
;
296 static inline struct nfs_direct_req
*nfs_direct_req_alloc(void)
298 struct nfs_direct_req
*dreq
;
300 dreq
= kmem_cache_zalloc(nfs_direct_cachep
, GFP_KERNEL
);
304 kref_init(&dreq
->kref
);
305 kref_get(&dreq
->kref
);
306 init_completion(&dreq
->completion
);
307 INIT_LIST_HEAD(&dreq
->mds_cinfo
.list
);
308 dreq
->verf
.committed
= NFS_INVALID_STABLE_HOW
; /* not set yet */
309 INIT_WORK(&dreq
->work
, nfs_direct_write_schedule_work
);
310 spin_lock_init(&dreq
->lock
);
315 static void nfs_direct_req_free(struct kref
*kref
)
317 struct nfs_direct_req
*dreq
= container_of(kref
, struct nfs_direct_req
, kref
);
319 nfs_free_pnfs_ds_cinfo(&dreq
->ds_cinfo
);
320 if (dreq
->l_ctx
!= NULL
)
321 nfs_put_lock_context(dreq
->l_ctx
);
322 if (dreq
->ctx
!= NULL
)
323 put_nfs_open_context(dreq
->ctx
);
324 kmem_cache_free(nfs_direct_cachep
, dreq
);
327 static void nfs_direct_req_release(struct nfs_direct_req
*dreq
)
329 kref_put(&dreq
->kref
, nfs_direct_req_free
);
332 ssize_t
nfs_dreq_bytes_left(struct nfs_direct_req
*dreq
)
334 return dreq
->bytes_left
;
336 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left
);
339 * Collects and returns the final error value/byte-count.
341 static ssize_t
nfs_direct_wait(struct nfs_direct_req
*dreq
)
343 ssize_t result
= -EIOCBQUEUED
;
345 /* Async requests don't wait here */
349 result
= wait_for_completion_killable(&dreq
->completion
);
352 result
= dreq
->count
;
353 WARN_ON_ONCE(dreq
->count
< 0);
356 result
= dreq
->error
;
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
)
368 struct inode
*inode
= dreq
->inode
;
370 inode_dio_end(inode
);
373 long res
= (long) dreq
->error
;
374 if (dreq
->count
!= 0) {
375 res
= (long) dreq
->count
;
376 WARN_ON_ONCE(dreq
->count
< 0);
378 dreq
->iocb
->ki_complete(dreq
->iocb
, res
, 0);
381 complete(&dreq
->completion
);
383 nfs_direct_req_release(dreq
);
386 static void nfs_direct_read_completion(struct nfs_pgio_header
*hdr
)
388 unsigned long bytes
= 0;
389 struct nfs_direct_req
*dreq
= hdr
->dreq
;
391 spin_lock(&dreq
->lock
);
392 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
393 spin_unlock(&dreq
->lock
);
397 nfs_direct_count_bytes(dreq
, hdr
);
398 spin_unlock(&dreq
->lock
);
400 while (!list_empty(&hdr
->pages
)) {
401 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
402 struct page
*page
= req
->wb_page
;
404 if (!PageCompound(page
) && bytes
< hdr
->good_bytes
&&
405 (dreq
->flags
== NFS_ODIRECT_SHOULD_DIRTY
))
406 set_page_dirty(page
);
407 bytes
+= req
->wb_bytes
;
408 nfs_list_remove_request(req
);
409 nfs_release_request(req
);
413 nfs_direct_complete(dreq
);
417 static void nfs_read_sync_pgio_error(struct list_head
*head
, int error
)
419 struct nfs_page
*req
;
421 while (!list_empty(head
)) {
422 req
= nfs_list_entry(head
->next
);
423 nfs_list_remove_request(req
);
424 nfs_release_request(req
);
428 static void nfs_direct_pgio_init(struct nfs_pgio_header
*hdr
)
433 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops
= {
434 .error_cleanup
= nfs_read_sync_pgio_error
,
435 .init_hdr
= nfs_direct_pgio_init
,
436 .completion
= nfs_direct_read_completion
,
440 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
441 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
442 * bail and stop sending more reads. Read length accounting is
443 * handled automatically by nfs_direct_read_result(). Otherwise, if
444 * no requests have been sent, just return an error.
447 static ssize_t
nfs_direct_read_schedule_iovec(struct nfs_direct_req
*dreq
,
448 struct iov_iter
*iter
,
451 struct nfs_pageio_descriptor desc
;
452 struct inode
*inode
= dreq
->inode
;
453 ssize_t result
= -EINVAL
;
454 size_t requested_bytes
= 0;
455 size_t rsize
= max_t(size_t, NFS_SERVER(inode
)->rsize
, PAGE_SIZE
);
457 nfs_pageio_init_read(&desc
, dreq
->inode
, false,
458 &nfs_direct_read_completion_ops
);
461 inode_dio_begin(inode
);
463 while (iov_iter_count(iter
)) {
464 struct page
**pagevec
;
469 result
= iov_iter_get_pages_alloc(iter
, &pagevec
,
475 iov_iter_advance(iter
, bytes
);
476 npages
= (result
+ pgbase
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
477 for (i
= 0; i
< npages
; i
++) {
478 struct nfs_page
*req
;
479 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
480 /* XXX do we need to do the eof zeroing found in async_filler? */
481 req
= nfs_create_request(dreq
->ctx
, pagevec
[i
],
484 result
= PTR_ERR(req
);
487 req
->wb_index
= pos
>> PAGE_SHIFT
;
488 req
->wb_offset
= pos
& ~PAGE_MASK
;
489 if (!nfs_pageio_add_request(&desc
, req
)) {
490 result
= desc
.pg_error
;
491 nfs_release_request(req
);
496 requested_bytes
+= req_len
;
498 dreq
->bytes_left
-= req_len
;
500 nfs_direct_release_pages(pagevec
, npages
);
506 nfs_pageio_complete(&desc
);
509 * If no bytes were started, return the error, and let the
510 * generic layer handle the completion.
512 if (requested_bytes
== 0) {
513 inode_dio_end(inode
);
514 nfs_direct_req_release(dreq
);
515 return result
< 0 ? result
: -EIO
;
519 nfs_direct_complete(dreq
);
520 return requested_bytes
;
524 * nfs_file_direct_read - file direct read operation for NFS files
525 * @iocb: target I/O control block
526 * @iter: vector of user buffers into which to read data
528 * We use this function for direct reads instead of calling
529 * generic_file_aio_read() in order to avoid gfar's check to see if
530 * the request starts before the end of the file. For that check
531 * to work, we must generate a GETATTR before each direct read, and
532 * even then there is a window between the GETATTR and the subsequent
533 * READ where the file size could change. Our preference is simply
534 * to do all reads the application wants, and the server will take
535 * care of managing the end of file boundary.
537 * This function also eliminates unnecessarily updating the file's
538 * atime locally, as the NFS server sets the file's atime, and this
539 * client must read the updated atime from the server back into its
542 ssize_t
nfs_file_direct_read(struct kiocb
*iocb
, struct iov_iter
*iter
)
544 struct file
*file
= iocb
->ki_filp
;
545 struct address_space
*mapping
= file
->f_mapping
;
546 struct inode
*inode
= mapping
->host
;
547 struct nfs_direct_req
*dreq
;
548 struct nfs_lock_context
*l_ctx
;
549 ssize_t result
= -EINVAL
, requested
;
550 size_t count
= iov_iter_count(iter
);
551 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTREADBYTES
, count
);
553 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
554 file
, count
, (long long) iocb
->ki_pos
);
560 task_io_account_read(count
);
563 dreq
= nfs_direct_req_alloc();
568 dreq
->bytes_left
= dreq
->max_count
= count
;
569 dreq
->io_start
= iocb
->ki_pos
;
570 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
571 l_ctx
= nfs_get_lock_context(dreq
->ctx
);
573 result
= PTR_ERR(l_ctx
);
577 if (!is_sync_kiocb(iocb
))
580 if (iter_is_iovec(iter
))
581 dreq
->flags
= NFS_ODIRECT_SHOULD_DIRTY
;
583 nfs_start_io_direct(inode
);
585 NFS_I(inode
)->read_io
+= count
;
586 requested
= nfs_direct_read_schedule_iovec(dreq
, iter
, iocb
->ki_pos
);
588 nfs_end_io_direct(inode
);
591 result
= nfs_direct_wait(dreq
);
594 iocb
->ki_pos
+= result
;
596 iov_iter_revert(iter
, requested
);
602 nfs_direct_req_release(dreq
);
608 nfs_direct_write_scan_commit_list(struct inode
*inode
,
609 struct list_head
*list
,
610 struct nfs_commit_info
*cinfo
)
612 mutex_lock(&NFS_I(cinfo
->inode
)->commit_mutex
);
613 #ifdef CONFIG_NFS_V4_1
614 if (cinfo
->ds
!= NULL
&& cinfo
->ds
->nwritten
!= 0)
615 NFS_SERVER(inode
)->pnfs_curr_ld
->recover_commit_reqs(list
, cinfo
);
617 nfs_scan_commit_list(&cinfo
->mds
->list
, list
, cinfo
, 0);
618 mutex_unlock(&NFS_I(cinfo
->inode
)->commit_mutex
);
621 static void nfs_direct_write_reschedule(struct nfs_direct_req
*dreq
)
623 struct nfs_pageio_descriptor desc
;
624 struct nfs_page
*req
, *tmp
;
626 struct nfs_commit_info cinfo
;
629 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
630 nfs_direct_write_scan_commit_list(dreq
->inode
, &reqs
, &cinfo
);
634 list_for_each_entry(req
, &reqs
, wb_list
)
635 dreq
->max_count
+= req
->wb_bytes
;
636 dreq
->verf
.committed
= NFS_INVALID_STABLE_HOW
;
637 nfs_clear_pnfs_ds_commit_verifiers(&dreq
->ds_cinfo
);
640 nfs_pageio_init_write(&desc
, dreq
->inode
, FLUSH_STABLE
, false,
641 &nfs_direct_write_completion_ops
);
644 list_for_each_entry_safe(req
, tmp
, &reqs
, wb_list
) {
645 /* Bump the transmission count */
647 if (!nfs_pageio_add_request(&desc
, req
)) {
648 nfs_list_move_request(req
, &failed
);
649 spin_lock(&cinfo
.inode
->i_lock
);
651 if (desc
.pg_error
< 0)
652 dreq
->error
= desc
.pg_error
;
655 spin_unlock(&cinfo
.inode
->i_lock
);
657 nfs_release_request(req
);
659 nfs_pageio_complete(&desc
);
661 while (!list_empty(&failed
)) {
662 req
= nfs_list_entry(failed
.next
);
663 nfs_list_remove_request(req
);
664 nfs_unlock_and_release_request(req
);
668 nfs_direct_write_complete(dreq
);
671 static void nfs_direct_commit_complete(struct nfs_commit_data
*data
)
673 struct nfs_direct_req
*dreq
= data
->dreq
;
674 struct nfs_commit_info cinfo
;
675 struct nfs_page
*req
;
676 int status
= data
->task
.tk_status
;
678 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
679 if (status
< 0 || nfs_direct_cmp_commit_data_verf(dreq
, data
))
680 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
682 while (!list_empty(&data
->pages
)) {
683 req
= nfs_list_entry(data
->pages
.next
);
684 nfs_list_remove_request(req
);
685 if (dreq
->flags
== NFS_ODIRECT_RESCHED_WRITES
) {
687 * Despite the reboot, the write was successful,
691 /* Note the rewrite will go through mds */
692 nfs_mark_request_commit(req
, NULL
, &cinfo
, 0);
694 nfs_release_request(req
);
695 nfs_unlock_and_release_request(req
);
698 if (atomic_dec_and_test(&cinfo
.mds
->rpcs_out
))
699 nfs_direct_write_complete(dreq
);
702 static void nfs_direct_resched_write(struct nfs_commit_info
*cinfo
,
703 struct nfs_page
*req
)
705 struct nfs_direct_req
*dreq
= cinfo
->dreq
;
707 spin_lock(&dreq
->lock
);
708 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
709 spin_unlock(&dreq
->lock
);
710 nfs_mark_request_commit(req
, NULL
, cinfo
, 0);
713 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops
= {
714 .completion
= nfs_direct_commit_complete
,
715 .resched_write
= nfs_direct_resched_write
,
718 static void nfs_direct_commit_schedule(struct nfs_direct_req
*dreq
)
721 struct nfs_commit_info cinfo
;
724 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
725 nfs_scan_commit(dreq
->inode
, &mds_list
, &cinfo
);
726 res
= nfs_generic_commit_list(dreq
->inode
, &mds_list
, 0, &cinfo
);
727 if (res
< 0) /* res == -ENOMEM */
728 nfs_direct_write_reschedule(dreq
);
731 static void nfs_direct_write_schedule_work(struct work_struct
*work
)
733 struct nfs_direct_req
*dreq
= container_of(work
, struct nfs_direct_req
, work
);
734 int flags
= dreq
->flags
;
738 case NFS_ODIRECT_DO_COMMIT
:
739 nfs_direct_commit_schedule(dreq
);
741 case NFS_ODIRECT_RESCHED_WRITES
:
742 nfs_direct_write_reschedule(dreq
);
745 nfs_zap_mapping(dreq
->inode
, dreq
->inode
->i_mapping
);
746 nfs_direct_complete(dreq
);
750 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
)
752 queue_work(nfsiod_workqueue
, &dreq
->work
); /* Calls nfs_direct_write_schedule_work */
755 static void nfs_direct_write_completion(struct nfs_pgio_header
*hdr
)
757 struct nfs_direct_req
*dreq
= hdr
->dreq
;
758 struct nfs_commit_info cinfo
;
759 bool request_commit
= false;
760 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
762 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
764 spin_lock(&dreq
->lock
);
765 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
766 spin_unlock(&dreq
->lock
);
770 nfs_direct_count_bytes(dreq
, hdr
);
771 if (hdr
->good_bytes
!= 0) {
772 if (nfs_write_need_commit(hdr
)) {
773 if (dreq
->flags
== NFS_ODIRECT_RESCHED_WRITES
)
774 request_commit
= true;
775 else if (dreq
->flags
== 0) {
776 nfs_direct_set_hdr_verf(dreq
, hdr
);
777 request_commit
= true;
778 dreq
->flags
= NFS_ODIRECT_DO_COMMIT
;
779 } else if (dreq
->flags
== NFS_ODIRECT_DO_COMMIT
) {
780 request_commit
= true;
781 if (nfs_direct_set_or_cmp_hdr_verf(dreq
, hdr
))
783 NFS_ODIRECT_RESCHED_WRITES
;
787 spin_unlock(&dreq
->lock
);
789 while (!list_empty(&hdr
->pages
)) {
791 req
= nfs_list_entry(hdr
->pages
.next
);
792 nfs_list_remove_request(req
);
793 if (request_commit
) {
794 kref_get(&req
->wb_kref
);
795 nfs_mark_request_commit(req
, hdr
->lseg
, &cinfo
,
798 nfs_unlock_and_release_request(req
);
803 nfs_direct_write_complete(dreq
);
807 static void nfs_write_sync_pgio_error(struct list_head
*head
, int error
)
809 struct nfs_page
*req
;
811 while (!list_empty(head
)) {
812 req
= nfs_list_entry(head
->next
);
813 nfs_list_remove_request(req
);
814 nfs_unlock_and_release_request(req
);
818 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header
*hdr
)
820 struct nfs_direct_req
*dreq
= hdr
->dreq
;
822 spin_lock(&dreq
->lock
);
823 if (dreq
->error
== 0) {
824 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
825 /* fake unstable write to let common nfs resend pages */
826 hdr
->verf
.committed
= NFS_UNSTABLE
;
827 hdr
->good_bytes
= hdr
->args
.offset
+ hdr
->args
.count
-
830 spin_unlock(&dreq
->lock
);
833 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops
= {
834 .error_cleanup
= nfs_write_sync_pgio_error
,
835 .init_hdr
= nfs_direct_pgio_init
,
836 .completion
= nfs_direct_write_completion
,
837 .reschedule_io
= nfs_direct_write_reschedule_io
,
842 * NB: Return the value of the first error return code. Subsequent
843 * errors after the first one are ignored.
846 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
847 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
848 * bail and stop sending more writes. Write length accounting is
849 * handled automatically by nfs_direct_write_result(). Otherwise, if
850 * no requests have been sent, just return an error.
852 static ssize_t
nfs_direct_write_schedule_iovec(struct nfs_direct_req
*dreq
,
853 struct iov_iter
*iter
,
856 struct nfs_pageio_descriptor desc
;
857 struct inode
*inode
= dreq
->inode
;
859 size_t requested_bytes
= 0;
860 size_t wsize
= max_t(size_t, NFS_SERVER(inode
)->wsize
, PAGE_SIZE
);
862 nfs_pageio_init_write(&desc
, inode
, FLUSH_COND_STABLE
, false,
863 &nfs_direct_write_completion_ops
);
866 inode_dio_begin(inode
);
868 NFS_I(inode
)->write_io
+= iov_iter_count(iter
);
869 while (iov_iter_count(iter
)) {
870 struct page
**pagevec
;
875 result
= iov_iter_get_pages_alloc(iter
, &pagevec
,
881 iov_iter_advance(iter
, bytes
);
882 npages
= (result
+ pgbase
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
883 for (i
= 0; i
< npages
; i
++) {
884 struct nfs_page
*req
;
885 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
887 req
= nfs_create_request(dreq
->ctx
, pagevec
[i
],
890 result
= PTR_ERR(req
);
894 if (desc
.pg_error
< 0) {
895 nfs_free_request(req
);
896 result
= desc
.pg_error
;
900 nfs_lock_request(req
);
901 req
->wb_index
= pos
>> PAGE_SHIFT
;
902 req
->wb_offset
= pos
& ~PAGE_MASK
;
903 if (!nfs_pageio_add_request(&desc
, req
)) {
904 result
= desc
.pg_error
;
905 nfs_unlock_and_release_request(req
);
910 requested_bytes
+= req_len
;
912 dreq
->bytes_left
-= req_len
;
914 nfs_direct_release_pages(pagevec
, npages
);
919 nfs_pageio_complete(&desc
);
922 * If no bytes were started, return the error, and let the
923 * generic layer handle the completion.
925 if (requested_bytes
== 0) {
926 inode_dio_end(inode
);
927 nfs_direct_req_release(dreq
);
928 return result
< 0 ? result
: -EIO
;
932 nfs_direct_write_complete(dreq
);
933 return requested_bytes
;
937 * nfs_file_direct_write - file direct write operation for NFS files
938 * @iocb: target I/O control block
939 * @iter: vector of user buffers from which to write data
941 * We use this function for direct writes instead of calling
942 * generic_file_aio_write() in order to avoid taking the inode
943 * semaphore and updating the i_size. The NFS server will set
944 * the new i_size and this client must read the updated size
945 * back into its cache. We let the server do generic write
946 * parameter checking and report problems.
948 * We eliminate local atime updates, see direct read above.
950 * We avoid unnecessary page cache invalidations for normal cached
951 * readers of this file.
953 * Note that O_APPEND is not supported for NFS direct writes, as there
954 * is no atomic O_APPEND write facility in the NFS protocol.
956 ssize_t
nfs_file_direct_write(struct kiocb
*iocb
, struct iov_iter
*iter
)
958 ssize_t result
= -EINVAL
, requested
;
960 struct file
*file
= iocb
->ki_filp
;
961 struct address_space
*mapping
= file
->f_mapping
;
962 struct inode
*inode
= mapping
->host
;
963 struct nfs_direct_req
*dreq
;
964 struct nfs_lock_context
*l_ctx
;
967 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
968 file
, iov_iter_count(iter
), (long long) iocb
->ki_pos
);
970 result
= generic_write_checks(iocb
, iter
);
974 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTWRITTENBYTES
, count
);
977 end
= (pos
+ iov_iter_count(iter
) - 1) >> PAGE_SHIFT
;
979 task_io_account_write(count
);
982 dreq
= nfs_direct_req_alloc();
987 dreq
->bytes_left
= dreq
->max_count
= count
;
988 dreq
->io_start
= pos
;
989 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
990 l_ctx
= nfs_get_lock_context(dreq
->ctx
);
992 result
= PTR_ERR(l_ctx
);
996 if (!is_sync_kiocb(iocb
))
999 nfs_start_io_direct(inode
);
1001 requested
= nfs_direct_write_schedule_iovec(dreq
, iter
, pos
);
1003 if (mapping
->nrpages
) {
1004 invalidate_inode_pages2_range(mapping
,
1005 pos
>> PAGE_SHIFT
, end
);
1008 nfs_end_io_direct(inode
);
1010 if (requested
> 0) {
1011 result
= nfs_direct_wait(dreq
);
1013 requested
-= result
;
1014 iocb
->ki_pos
= pos
+ result
;
1015 /* XXX: should check the generic_write_sync retval */
1016 generic_write_sync(iocb
, result
);
1018 iov_iter_revert(iter
, requested
);
1023 nfs_direct_req_release(dreq
);
1029 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1032 int __init
nfs_init_directcache(void)
1034 nfs_direct_cachep
= kmem_cache_create("nfs_direct_cache",
1035 sizeof(struct nfs_direct_req
),
1036 0, (SLAB_RECLAIM_ACCOUNT
|
1039 if (nfs_direct_cachep
== NULL
)
1046 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1049 void nfs_destroy_directcache(void)
1051 kmem_cache_destroy(nfs_direct_cachep
);