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 #define NFS_ODIRECT_DONE INT_MAX /* write verification failed */
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_IO - NFS address space operation for direct I/O
156 * @iocb: target I/O control block
159 * The presence of this routine in the address space ops vector means
160 * the NFS client supports direct I/O. However, for most direct IO, we
161 * shunt off direct read and write requests before the VFS gets them,
162 * so this method is only ever called for swap.
164 ssize_t
nfs_direct_IO(struct kiocb
*iocb
, struct iov_iter
*iter
)
166 struct inode
*inode
= iocb
->ki_filp
->f_mapping
->host
;
168 /* we only support swap file calling nfs_direct_IO */
169 if (!IS_SWAPFILE(inode
))
172 VM_BUG_ON(iov_iter_count(iter
) != PAGE_SIZE
);
174 if (iov_iter_rw(iter
) == READ
)
175 return nfs_file_direct_read(iocb
, iter
, true);
176 return nfs_file_direct_write(iocb
, iter
, true);
179 static void nfs_direct_release_pages(struct page
**pages
, unsigned int npages
)
182 for (i
= 0; i
< npages
; i
++)
186 void nfs_init_cinfo_from_dreq(struct nfs_commit_info
*cinfo
,
187 struct nfs_direct_req
*dreq
)
189 cinfo
->inode
= dreq
->inode
;
190 cinfo
->mds
= &dreq
->mds_cinfo
;
191 cinfo
->ds
= &dreq
->ds_cinfo
;
193 cinfo
->completion_ops
= &nfs_direct_commit_completion_ops
;
196 static inline struct nfs_direct_req
*nfs_direct_req_alloc(void)
198 struct nfs_direct_req
*dreq
;
200 dreq
= kmem_cache_zalloc(nfs_direct_cachep
, GFP_KERNEL
);
204 kref_init(&dreq
->kref
);
205 kref_get(&dreq
->kref
);
206 init_completion(&dreq
->completion
);
207 INIT_LIST_HEAD(&dreq
->mds_cinfo
.list
);
208 pnfs_init_ds_commit_info(&dreq
->ds_cinfo
);
209 INIT_WORK(&dreq
->work
, nfs_direct_write_schedule_work
);
210 spin_lock_init(&dreq
->lock
);
215 static void nfs_direct_req_free(struct kref
*kref
)
217 struct nfs_direct_req
*dreq
= container_of(kref
, struct nfs_direct_req
, kref
);
219 pnfs_release_ds_info(&dreq
->ds_cinfo
, dreq
->inode
);
220 if (dreq
->l_ctx
!= NULL
)
221 nfs_put_lock_context(dreq
->l_ctx
);
222 if (dreq
->ctx
!= NULL
)
223 put_nfs_open_context(dreq
->ctx
);
224 kmem_cache_free(nfs_direct_cachep
, dreq
);
227 static void nfs_direct_req_release(struct nfs_direct_req
*dreq
)
229 kref_put(&dreq
->kref
, nfs_direct_req_free
);
232 ssize_t
nfs_dreq_bytes_left(struct nfs_direct_req
*dreq
)
234 return dreq
->bytes_left
;
236 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left
);
239 * Collects and returns the final error value/byte-count.
241 static ssize_t
nfs_direct_wait(struct nfs_direct_req
*dreq
)
243 ssize_t result
= -EIOCBQUEUED
;
245 /* Async requests don't wait here */
249 result
= wait_for_completion_killable(&dreq
->completion
);
252 result
= dreq
->count
;
253 WARN_ON_ONCE(dreq
->count
< 0);
256 result
= dreq
->error
;
259 return (ssize_t
) result
;
263 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
264 * the iocb is still valid here if this is a synchronous request.
266 static void nfs_direct_complete(struct nfs_direct_req
*dreq
)
268 struct inode
*inode
= dreq
->inode
;
270 inode_dio_end(inode
);
273 long res
= (long) dreq
->error
;
274 if (dreq
->count
!= 0) {
275 res
= (long) dreq
->count
;
276 WARN_ON_ONCE(dreq
->count
< 0);
278 dreq
->iocb
->ki_complete(dreq
->iocb
, res
, 0);
281 complete(&dreq
->completion
);
283 nfs_direct_req_release(dreq
);
286 static void nfs_direct_read_completion(struct nfs_pgio_header
*hdr
)
288 unsigned long bytes
= 0;
289 struct nfs_direct_req
*dreq
= hdr
->dreq
;
291 spin_lock(&dreq
->lock
);
292 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
293 spin_unlock(&dreq
->lock
);
297 nfs_direct_count_bytes(dreq
, hdr
);
298 spin_unlock(&dreq
->lock
);
300 while (!list_empty(&hdr
->pages
)) {
301 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
302 struct page
*page
= req
->wb_page
;
304 if (!PageCompound(page
) && bytes
< hdr
->good_bytes
&&
305 (dreq
->flags
== NFS_ODIRECT_SHOULD_DIRTY
))
306 set_page_dirty(page
);
307 bytes
+= req
->wb_bytes
;
308 nfs_list_remove_request(req
);
309 nfs_release_request(req
);
313 nfs_direct_complete(dreq
);
317 static void nfs_read_sync_pgio_error(struct list_head
*head
, int error
)
319 struct nfs_page
*req
;
321 while (!list_empty(head
)) {
322 req
= nfs_list_entry(head
->next
);
323 nfs_list_remove_request(req
);
324 nfs_release_request(req
);
328 static void nfs_direct_pgio_init(struct nfs_pgio_header
*hdr
)
333 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops
= {
334 .error_cleanup
= nfs_read_sync_pgio_error
,
335 .init_hdr
= nfs_direct_pgio_init
,
336 .completion
= nfs_direct_read_completion
,
340 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
341 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
342 * bail and stop sending more reads. Read length accounting is
343 * handled automatically by nfs_direct_read_result(). Otherwise, if
344 * no requests have been sent, just return an error.
347 static ssize_t
nfs_direct_read_schedule_iovec(struct nfs_direct_req
*dreq
,
348 struct iov_iter
*iter
,
351 struct nfs_pageio_descriptor desc
;
352 struct inode
*inode
= dreq
->inode
;
353 ssize_t result
= -EINVAL
;
354 size_t requested_bytes
= 0;
355 size_t rsize
= max_t(size_t, NFS_SERVER(inode
)->rsize
, PAGE_SIZE
);
357 nfs_pageio_init_read(&desc
, dreq
->inode
, false,
358 &nfs_direct_read_completion_ops
);
361 inode_dio_begin(inode
);
363 while (iov_iter_count(iter
)) {
364 struct page
**pagevec
;
369 result
= iov_iter_get_pages_alloc(iter
, &pagevec
,
375 iov_iter_advance(iter
, bytes
);
376 npages
= (result
+ pgbase
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
377 for (i
= 0; i
< npages
; i
++) {
378 struct nfs_page
*req
;
379 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
380 /* XXX do we need to do the eof zeroing found in async_filler? */
381 req
= nfs_create_request(dreq
->ctx
, pagevec
[i
],
384 result
= PTR_ERR(req
);
387 req
->wb_index
= pos
>> PAGE_SHIFT
;
388 req
->wb_offset
= pos
& ~PAGE_MASK
;
389 if (!nfs_pageio_add_request(&desc
, req
)) {
390 result
= desc
.pg_error
;
391 nfs_release_request(req
);
396 requested_bytes
+= req_len
;
398 dreq
->bytes_left
-= req_len
;
400 nfs_direct_release_pages(pagevec
, npages
);
406 nfs_pageio_complete(&desc
);
409 * If no bytes were started, return the error, and let the
410 * generic layer handle the completion.
412 if (requested_bytes
== 0) {
413 inode_dio_end(inode
);
414 nfs_direct_req_release(dreq
);
415 return result
< 0 ? result
: -EIO
;
419 nfs_direct_complete(dreq
);
420 return requested_bytes
;
424 * nfs_file_direct_read - file direct read operation for NFS files
425 * @iocb: target I/O control block
426 * @iter: vector of user buffers into which to read data
427 * @swap: flag indicating this is swap IO, not O_DIRECT IO
429 * We use this function for direct reads instead of calling
430 * generic_file_aio_read() in order to avoid gfar's check to see if
431 * the request starts before the end of the file. For that check
432 * to work, we must generate a GETATTR before each direct read, and
433 * even then there is a window between the GETATTR and the subsequent
434 * READ where the file size could change. Our preference is simply
435 * to do all reads the application wants, and the server will take
436 * care of managing the end of file boundary.
438 * This function also eliminates unnecessarily updating the file's
439 * atime locally, as the NFS server sets the file's atime, and this
440 * client must read the updated atime from the server back into its
443 ssize_t
nfs_file_direct_read(struct kiocb
*iocb
, struct iov_iter
*iter
,
446 struct file
*file
= iocb
->ki_filp
;
447 struct address_space
*mapping
= file
->f_mapping
;
448 struct inode
*inode
= mapping
->host
;
449 struct nfs_direct_req
*dreq
;
450 struct nfs_lock_context
*l_ctx
;
451 ssize_t result
, requested
;
452 size_t count
= iov_iter_count(iter
);
453 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTREADBYTES
, count
);
455 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
456 file
, count
, (long long) iocb
->ki_pos
);
462 task_io_account_read(count
);
465 dreq
= nfs_direct_req_alloc();
470 dreq
->bytes_left
= dreq
->max_count
= count
;
471 dreq
->io_start
= iocb
->ki_pos
;
472 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
473 l_ctx
= nfs_get_lock_context(dreq
->ctx
);
475 result
= PTR_ERR(l_ctx
);
476 nfs_direct_req_release(dreq
);
480 if (!is_sync_kiocb(iocb
))
483 if (iter_is_iovec(iter
))
484 dreq
->flags
= NFS_ODIRECT_SHOULD_DIRTY
;
487 nfs_start_io_direct(inode
);
489 NFS_I(inode
)->read_io
+= count
;
490 requested
= nfs_direct_read_schedule_iovec(dreq
, iter
, iocb
->ki_pos
);
493 nfs_end_io_direct(inode
);
496 result
= nfs_direct_wait(dreq
);
499 iocb
->ki_pos
+= result
;
501 iov_iter_revert(iter
, requested
);
507 nfs_direct_req_release(dreq
);
513 nfs_direct_join_group(struct list_head
*list
, struct inode
*inode
)
515 struct nfs_page
*req
, *next
;
517 list_for_each_entry(req
, list
, wb_list
) {
518 if (req
->wb_head
!= req
|| req
->wb_this_page
== req
)
520 for (next
= req
->wb_this_page
;
521 next
!= req
->wb_head
;
522 next
= next
->wb_this_page
) {
523 nfs_list_remove_request(next
);
524 nfs_release_request(next
);
526 nfs_join_page_group(req
, inode
);
531 nfs_direct_write_scan_commit_list(struct inode
*inode
,
532 struct list_head
*list
,
533 struct nfs_commit_info
*cinfo
)
535 mutex_lock(&NFS_I(cinfo
->inode
)->commit_mutex
);
536 pnfs_recover_commit_reqs(list
, cinfo
);
537 nfs_scan_commit_list(&cinfo
->mds
->list
, list
, cinfo
, 0);
538 mutex_unlock(&NFS_I(cinfo
->inode
)->commit_mutex
);
541 static void nfs_direct_write_reschedule(struct nfs_direct_req
*dreq
)
543 struct nfs_pageio_descriptor desc
;
544 struct nfs_page
*req
, *tmp
;
546 struct nfs_commit_info cinfo
;
549 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
550 nfs_direct_write_scan_commit_list(dreq
->inode
, &reqs
, &cinfo
);
552 nfs_direct_join_group(&reqs
, dreq
->inode
);
556 list_for_each_entry(req
, &reqs
, wb_list
)
557 dreq
->max_count
+= req
->wb_bytes
;
558 nfs_clear_pnfs_ds_commit_verifiers(&dreq
->ds_cinfo
);
561 nfs_pageio_init_write(&desc
, dreq
->inode
, FLUSH_STABLE
, false,
562 &nfs_direct_write_completion_ops
);
565 list_for_each_entry_safe(req
, tmp
, &reqs
, wb_list
) {
566 /* Bump the transmission count */
568 if (!nfs_pageio_add_request(&desc
, req
)) {
569 nfs_list_move_request(req
, &failed
);
570 spin_lock(&cinfo
.inode
->i_lock
);
572 if (desc
.pg_error
< 0)
573 dreq
->error
= desc
.pg_error
;
576 spin_unlock(&cinfo
.inode
->i_lock
);
578 nfs_release_request(req
);
580 nfs_pageio_complete(&desc
);
582 while (!list_empty(&failed
)) {
583 req
= nfs_list_entry(failed
.next
);
584 nfs_list_remove_request(req
);
585 nfs_unlock_and_release_request(req
);
589 nfs_direct_write_complete(dreq
);
592 static void nfs_direct_commit_complete(struct nfs_commit_data
*data
)
594 const struct nfs_writeverf
*verf
= data
->res
.verf
;
595 struct nfs_direct_req
*dreq
= data
->dreq
;
596 struct nfs_commit_info cinfo
;
597 struct nfs_page
*req
;
598 int status
= data
->task
.tk_status
;
601 /* Errors in commit are fatal */
602 dreq
->error
= status
;
605 dreq
->flags
= NFS_ODIRECT_DONE
;
606 } else if (dreq
->flags
== NFS_ODIRECT_DONE
)
607 status
= dreq
->error
;
609 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
611 while (!list_empty(&data
->pages
)) {
612 req
= nfs_list_entry(data
->pages
.next
);
613 nfs_list_remove_request(req
);
614 if (status
>= 0 && !nfs_write_match_verf(verf
, req
)) {
615 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
617 * Despite the reboot, the write was successful,
621 nfs_mark_request_commit(req
, NULL
, &cinfo
, 0);
622 } else /* Error or match */
623 nfs_release_request(req
);
624 nfs_unlock_and_release_request(req
);
627 if (nfs_commit_end(cinfo
.mds
))
628 nfs_direct_write_complete(dreq
);
631 static void nfs_direct_resched_write(struct nfs_commit_info
*cinfo
,
632 struct nfs_page
*req
)
634 struct nfs_direct_req
*dreq
= cinfo
->dreq
;
636 spin_lock(&dreq
->lock
);
637 if (dreq
->flags
!= NFS_ODIRECT_DONE
)
638 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
639 spin_unlock(&dreq
->lock
);
640 nfs_mark_request_commit(req
, NULL
, cinfo
, 0);
643 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops
= {
644 .completion
= nfs_direct_commit_complete
,
645 .resched_write
= nfs_direct_resched_write
,
648 static void nfs_direct_commit_schedule(struct nfs_direct_req
*dreq
)
651 struct nfs_commit_info cinfo
;
654 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
655 nfs_scan_commit(dreq
->inode
, &mds_list
, &cinfo
);
656 res
= nfs_generic_commit_list(dreq
->inode
, &mds_list
, 0, &cinfo
);
657 if (res
< 0) /* res == -ENOMEM */
658 nfs_direct_write_reschedule(dreq
);
661 static void nfs_direct_write_clear_reqs(struct nfs_direct_req
*dreq
)
663 struct nfs_commit_info cinfo
;
664 struct nfs_page
*req
;
667 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
668 nfs_direct_write_scan_commit_list(dreq
->inode
, &reqs
, &cinfo
);
670 while (!list_empty(&reqs
)) {
671 req
= nfs_list_entry(reqs
.next
);
672 nfs_list_remove_request(req
);
673 nfs_release_request(req
);
674 nfs_unlock_and_release_request(req
);
678 static void nfs_direct_write_schedule_work(struct work_struct
*work
)
680 struct nfs_direct_req
*dreq
= container_of(work
, struct nfs_direct_req
, work
);
681 int flags
= dreq
->flags
;
685 case NFS_ODIRECT_DO_COMMIT
:
686 nfs_direct_commit_schedule(dreq
);
688 case NFS_ODIRECT_RESCHED_WRITES
:
689 nfs_direct_write_reschedule(dreq
);
692 nfs_direct_write_clear_reqs(dreq
);
693 nfs_zap_mapping(dreq
->inode
, dreq
->inode
->i_mapping
);
694 nfs_direct_complete(dreq
);
698 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
)
700 queue_work(nfsiod_workqueue
, &dreq
->work
); /* Calls nfs_direct_write_schedule_work */
703 static void nfs_direct_write_completion(struct nfs_pgio_header
*hdr
)
705 struct nfs_direct_req
*dreq
= hdr
->dreq
;
706 struct nfs_commit_info cinfo
;
707 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
708 int flags
= NFS_ODIRECT_DONE
;
710 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
712 spin_lock(&dreq
->lock
);
713 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
714 spin_unlock(&dreq
->lock
);
718 nfs_direct_count_bytes(dreq
, hdr
);
719 if (hdr
->good_bytes
!= 0 && nfs_write_need_commit(hdr
)) {
721 dreq
->flags
= NFS_ODIRECT_DO_COMMIT
;
724 spin_unlock(&dreq
->lock
);
726 while (!list_empty(&hdr
->pages
)) {
728 req
= nfs_list_entry(hdr
->pages
.next
);
729 nfs_list_remove_request(req
);
730 if (flags
== NFS_ODIRECT_DO_COMMIT
) {
731 kref_get(&req
->wb_kref
);
732 memcpy(&req
->wb_verf
, &hdr
->verf
.verifier
,
733 sizeof(req
->wb_verf
));
734 nfs_mark_request_commit(req
, hdr
->lseg
, &cinfo
,
736 } else if (flags
== NFS_ODIRECT_RESCHED_WRITES
) {
737 kref_get(&req
->wb_kref
);
738 nfs_mark_request_commit(req
, NULL
, &cinfo
, 0);
740 nfs_unlock_and_release_request(req
);
745 nfs_direct_write_complete(dreq
);
749 static void nfs_write_sync_pgio_error(struct list_head
*head
, int error
)
751 struct nfs_page
*req
;
753 while (!list_empty(head
)) {
754 req
= nfs_list_entry(head
->next
);
755 nfs_list_remove_request(req
);
756 nfs_unlock_and_release_request(req
);
760 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header
*hdr
)
762 struct nfs_direct_req
*dreq
= hdr
->dreq
;
764 spin_lock(&dreq
->lock
);
765 if (dreq
->error
== 0) {
766 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
767 /* fake unstable write to let common nfs resend pages */
768 hdr
->verf
.committed
= NFS_UNSTABLE
;
769 hdr
->good_bytes
= hdr
->args
.offset
+ hdr
->args
.count
-
772 spin_unlock(&dreq
->lock
);
775 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops
= {
776 .error_cleanup
= nfs_write_sync_pgio_error
,
777 .init_hdr
= nfs_direct_pgio_init
,
778 .completion
= nfs_direct_write_completion
,
779 .reschedule_io
= nfs_direct_write_reschedule_io
,
784 * NB: Return the value of the first error return code. Subsequent
785 * errors after the first one are ignored.
788 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
789 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
790 * bail and stop sending more writes. Write length accounting is
791 * handled automatically by nfs_direct_write_result(). Otherwise, if
792 * no requests have been sent, just return an error.
794 static ssize_t
nfs_direct_write_schedule_iovec(struct nfs_direct_req
*dreq
,
795 struct iov_iter
*iter
,
796 loff_t pos
, int ioflags
)
798 struct nfs_pageio_descriptor desc
;
799 struct inode
*inode
= dreq
->inode
;
801 size_t requested_bytes
= 0;
802 size_t wsize
= max_t(size_t, NFS_SERVER(inode
)->wsize
, PAGE_SIZE
);
804 nfs_pageio_init_write(&desc
, inode
, ioflags
, false,
805 &nfs_direct_write_completion_ops
);
808 inode_dio_begin(inode
);
810 NFS_I(inode
)->write_io
+= iov_iter_count(iter
);
811 while (iov_iter_count(iter
)) {
812 struct page
**pagevec
;
817 result
= iov_iter_get_pages_alloc(iter
, &pagevec
,
823 iov_iter_advance(iter
, bytes
);
824 npages
= (result
+ pgbase
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
825 for (i
= 0; i
< npages
; i
++) {
826 struct nfs_page
*req
;
827 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
829 req
= nfs_create_request(dreq
->ctx
, pagevec
[i
],
832 result
= PTR_ERR(req
);
836 if (desc
.pg_error
< 0) {
837 nfs_free_request(req
);
838 result
= desc
.pg_error
;
842 nfs_lock_request(req
);
843 req
->wb_index
= pos
>> PAGE_SHIFT
;
844 req
->wb_offset
= pos
& ~PAGE_MASK
;
845 if (!nfs_pageio_add_request(&desc
, req
)) {
846 result
= desc
.pg_error
;
847 nfs_unlock_and_release_request(req
);
852 requested_bytes
+= req_len
;
854 dreq
->bytes_left
-= req_len
;
856 nfs_direct_release_pages(pagevec
, npages
);
861 nfs_pageio_complete(&desc
);
864 * If no bytes were started, return the error, and let the
865 * generic layer handle the completion.
867 if (requested_bytes
== 0) {
868 inode_dio_end(inode
);
869 nfs_direct_req_release(dreq
);
870 return result
< 0 ? result
: -EIO
;
874 nfs_direct_write_complete(dreq
);
875 return requested_bytes
;
879 * nfs_file_direct_write - file direct write operation for NFS files
880 * @iocb: target I/O control block
881 * @iter: vector of user buffers from which to write data
882 * @swap: flag indicating this is swap IO, not O_DIRECT IO
884 * We use this function for direct writes instead of calling
885 * generic_file_aio_write() in order to avoid taking the inode
886 * semaphore and updating the i_size. The NFS server will set
887 * the new i_size and this client must read the updated size
888 * back into its cache. We let the server do generic write
889 * parameter checking and report problems.
891 * We eliminate local atime updates, see direct read above.
893 * We avoid unnecessary page cache invalidations for normal cached
894 * readers of this file.
896 * Note that O_APPEND is not supported for NFS direct writes, as there
897 * is no atomic O_APPEND write facility in the NFS protocol.
899 ssize_t
nfs_file_direct_write(struct kiocb
*iocb
, struct iov_iter
*iter
,
902 ssize_t result
, requested
;
904 struct file
*file
= iocb
->ki_filp
;
905 struct address_space
*mapping
= file
->f_mapping
;
906 struct inode
*inode
= mapping
->host
;
907 struct nfs_direct_req
*dreq
;
908 struct nfs_lock_context
*l_ctx
;
911 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
912 file
, iov_iter_count(iter
), (long long) iocb
->ki_pos
);
915 /* bypass generic checks */
916 result
= iov_iter_count(iter
);
918 result
= generic_write_checks(iocb
, iter
);
922 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTWRITTENBYTES
, count
);
925 end
= (pos
+ iov_iter_count(iter
) - 1) >> PAGE_SHIFT
;
927 task_io_account_write(count
);
930 dreq
= nfs_direct_req_alloc();
935 dreq
->bytes_left
= dreq
->max_count
= count
;
936 dreq
->io_start
= pos
;
937 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
938 l_ctx
= nfs_get_lock_context(dreq
->ctx
);
940 result
= PTR_ERR(l_ctx
);
941 nfs_direct_req_release(dreq
);
945 if (!is_sync_kiocb(iocb
))
947 pnfs_init_ds_commit_info_ops(&dreq
->ds_cinfo
, inode
);
950 requested
= nfs_direct_write_schedule_iovec(dreq
, iter
, pos
,
953 nfs_start_io_direct(inode
);
955 requested
= nfs_direct_write_schedule_iovec(dreq
, iter
, pos
,
958 if (mapping
->nrpages
) {
959 invalidate_inode_pages2_range(mapping
,
960 pos
>> PAGE_SHIFT
, end
);
963 nfs_end_io_direct(inode
);
967 result
= nfs_direct_wait(dreq
);
970 iocb
->ki_pos
= pos
+ result
;
971 /* XXX: should check the generic_write_sync retval */
972 generic_write_sync(iocb
, result
);
974 iov_iter_revert(iter
, requested
);
979 nfs_direct_req_release(dreq
);
985 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
988 int __init
nfs_init_directcache(void)
990 nfs_direct_cachep
= kmem_cache_create("nfs_direct_cache",
991 sizeof(struct nfs_direct_req
),
992 0, (SLAB_RECLAIM_ACCOUNT
|
995 if (nfs_direct_cachep
== NULL
)
1002 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1005 void nfs_destroy_directcache(void)
1007 kmem_cache_destroy(nfs_direct_cachep
);