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>
48 #include <linux/nfs_fs.h>
49 #include <linux/nfs_page.h>
50 #include <linux/sunrpc/clnt.h>
52 #include <asm/system.h>
53 #include <asm/uaccess.h>
54 #include <asm/atomic.h>
59 #define NFSDBG_FACILITY NFSDBG_VFS
61 static struct kmem_cache
*nfs_direct_cachep
;
64 * This represents a set of asynchronous requests that we're waiting on
66 struct nfs_direct_req
{
67 struct kref kref
; /* release manager */
70 struct nfs_open_context
*ctx
; /* file open context info */
71 struct kiocb
* iocb
; /* controlling i/o request */
72 struct inode
* inode
; /* target file of i/o */
74 /* completion state */
75 atomic_t io_count
; /* i/os we're waiting for */
76 spinlock_t lock
; /* protect completion state */
77 ssize_t count
, /* bytes actually processed */
78 error
; /* any reported error */
79 struct completion completion
; /* wait for i/o completion */
82 struct list_head rewrite_list
; /* saved nfs_write_data structs */
83 struct nfs_write_data
* commit_data
; /* special write_data for commits */
85 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
86 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
87 struct nfs_writeverf verf
; /* unstable write verifier */
90 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
, struct inode
*inode
);
91 static const struct rpc_call_ops nfs_write_direct_ops
;
93 static inline void get_dreq(struct nfs_direct_req
*dreq
)
95 atomic_inc(&dreq
->io_count
);
98 static inline int put_dreq(struct nfs_direct_req
*dreq
)
100 return atomic_dec_and_test(&dreq
->io_count
);
104 * nfs_direct_IO - NFS address space operation for direct I/O
105 * @rw: direction (read or write)
106 * @iocb: target I/O control block
107 * @iov: array of vectors that define I/O buffer
108 * @pos: offset in file to begin the operation
109 * @nr_segs: size of iovec array
111 * The presence of this routine in the address space ops vector means
112 * the NFS client supports direct I/O. However, we shunt off direct
113 * read and write requests before the VFS gets them, so this method
114 * should never be called.
116 ssize_t
nfs_direct_IO(int rw
, struct kiocb
*iocb
, const struct iovec
*iov
, loff_t pos
, unsigned long nr_segs
)
118 dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
119 iocb
->ki_filp
->f_path
.dentry
->d_name
.name
,
120 (long long) pos
, nr_segs
);
125 static void nfs_direct_dirty_pages(struct page
**pages
, unsigned int pgbase
, size_t count
)
132 pages
+= (pgbase
>> PAGE_SHIFT
);
133 npages
= (count
+ (pgbase
& ~PAGE_MASK
) + PAGE_SIZE
- 1) >> PAGE_SHIFT
;
134 for (i
= 0; i
< npages
; i
++) {
135 struct page
*page
= pages
[i
];
136 if (!PageCompound(page
))
137 set_page_dirty(page
);
141 static void nfs_direct_release_pages(struct page
**pages
, unsigned int npages
)
144 for (i
= 0; i
< npages
; i
++)
145 page_cache_release(pages
[i
]);
148 static inline struct nfs_direct_req
*nfs_direct_req_alloc(void)
150 struct nfs_direct_req
*dreq
;
152 dreq
= kmem_cache_alloc(nfs_direct_cachep
, GFP_KERNEL
);
156 kref_init(&dreq
->kref
);
157 kref_get(&dreq
->kref
);
158 init_completion(&dreq
->completion
);
159 INIT_LIST_HEAD(&dreq
->rewrite_list
);
162 spin_lock_init(&dreq
->lock
);
163 atomic_set(&dreq
->io_count
, 0);
171 static void nfs_direct_req_free(struct kref
*kref
)
173 struct nfs_direct_req
*dreq
= container_of(kref
, struct nfs_direct_req
, kref
);
175 if (dreq
->ctx
!= NULL
)
176 put_nfs_open_context(dreq
->ctx
);
177 kmem_cache_free(nfs_direct_cachep
, dreq
);
180 static void nfs_direct_req_release(struct nfs_direct_req
*dreq
)
182 kref_put(&dreq
->kref
, nfs_direct_req_free
);
186 * Collects and returns the final error value/byte-count.
188 static ssize_t
nfs_direct_wait(struct nfs_direct_req
*dreq
)
190 ssize_t result
= -EIOCBQUEUED
;
192 /* Async requests don't wait here */
196 result
= wait_for_completion_killable(&dreq
->completion
);
199 result
= dreq
->error
;
201 result
= dreq
->count
;
204 return (ssize_t
) result
;
208 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
209 * the iocb is still valid here if this is a synchronous request.
211 static void nfs_direct_complete(struct nfs_direct_req
*dreq
)
214 long res
= (long) dreq
->error
;
216 res
= (long) dreq
->count
;
217 aio_complete(dreq
->iocb
, res
, 0);
219 complete_all(&dreq
->completion
);
221 nfs_direct_req_release(dreq
);
225 * We must hold a reference to all the pages in this direct read request
226 * until the RPCs complete. This could be long *after* we are woken up in
227 * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
229 static void nfs_direct_read_result(struct rpc_task
*task
, void *calldata
)
231 struct nfs_read_data
*data
= calldata
;
232 struct nfs_direct_req
*dreq
= (struct nfs_direct_req
*) data
->req
;
234 if (nfs_readpage_result(task
, data
) != 0)
237 spin_lock(&dreq
->lock
);
238 if (unlikely(task
->tk_status
< 0)) {
239 dreq
->error
= task
->tk_status
;
240 spin_unlock(&dreq
->lock
);
242 dreq
->count
+= data
->res
.count
;
243 spin_unlock(&dreq
->lock
);
244 nfs_direct_dirty_pages(data
->pagevec
,
248 nfs_direct_release_pages(data
->pagevec
, data
->npages
);
251 nfs_direct_complete(dreq
);
254 static const struct rpc_call_ops nfs_read_direct_ops
= {
255 .rpc_call_done
= nfs_direct_read_result
,
256 .rpc_release
= nfs_readdata_release
,
260 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
261 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
262 * bail and stop sending more reads. Read length accounting is
263 * handled automatically by nfs_direct_read_result(). Otherwise, if
264 * no requests have been sent, just return an error.
266 static ssize_t
nfs_direct_read_schedule_segment(struct nfs_direct_req
*dreq
,
267 const struct iovec
*iov
,
270 struct nfs_open_context
*ctx
= dreq
->ctx
;
271 struct inode
*inode
= ctx
->path
.dentry
->d_inode
;
272 unsigned long user_addr
= (unsigned long)iov
->iov_base
;
273 size_t count
= iov
->iov_len
;
274 size_t rsize
= NFS_SERVER(inode
)->rsize
;
280 struct nfs_read_data
*data
;
283 pgbase
= user_addr
& ~PAGE_MASK
;
284 bytes
= min(rsize
,count
);
287 data
= nfs_readdata_alloc(nfs_page_array_len(pgbase
, bytes
));
291 down_read(¤t
->mm
->mmap_sem
);
292 result
= get_user_pages(current
, current
->mm
, user_addr
,
293 data
->npages
, 1, 0, data
->pagevec
, NULL
);
294 up_read(¤t
->mm
->mmap_sem
);
296 nfs_readdata_release(data
);
299 if ((unsigned)result
< data
->npages
) {
300 bytes
= result
* PAGE_SIZE
;
301 if (bytes
<= pgbase
) {
302 nfs_direct_release_pages(data
->pagevec
, result
);
303 nfs_readdata_release(data
);
307 data
->npages
= result
;
312 data
->req
= (struct nfs_page
*) dreq
;
314 data
->cred
= ctx
->cred
;
315 data
->args
.fh
= NFS_FH(inode
);
316 data
->args
.context
= ctx
;
317 data
->args
.offset
= pos
;
318 data
->args
.pgbase
= pgbase
;
319 data
->args
.pages
= data
->pagevec
;
320 data
->args
.count
= bytes
;
321 data
->res
.fattr
= &data
->fattr
;
323 data
->res
.count
= bytes
;
325 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), RPC_TASK_ASYNC
,
326 &nfs_read_direct_ops
, data
);
327 NFS_PROTO(inode
)->read_setup(data
);
329 data
->task
.tk_cookie
= (unsigned long) inode
;
331 rpc_execute(&data
->task
);
333 dprintk("NFS: %5u initiated direct read call "
334 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
337 (long long)NFS_FILEID(inode
),
339 (unsigned long long)data
->args
.offset
);
344 /* FIXME: Remove this unnecessary math from final patch */
346 pgbase
&= ~PAGE_MASK
;
347 BUG_ON(pgbase
!= (user_addr
& ~PAGE_MASK
));
350 } while (count
!= 0);
354 return result
< 0 ? (ssize_t
) result
: -EFAULT
;
357 static ssize_t
nfs_direct_read_schedule_iovec(struct nfs_direct_req
*dreq
,
358 const struct iovec
*iov
,
359 unsigned long nr_segs
,
362 ssize_t result
= -EINVAL
;
363 size_t requested_bytes
= 0;
368 for (seg
= 0; seg
< nr_segs
; seg
++) {
369 const struct iovec
*vec
= &iov
[seg
];
370 result
= nfs_direct_read_schedule_segment(dreq
, vec
, pos
);
373 requested_bytes
+= result
;
374 if ((size_t)result
< vec
->iov_len
)
380 nfs_direct_complete(dreq
);
382 if (requested_bytes
!= 0)
390 static ssize_t
nfs_direct_read(struct kiocb
*iocb
, const struct iovec
*iov
,
391 unsigned long nr_segs
, loff_t pos
)
394 struct inode
*inode
= iocb
->ki_filp
->f_mapping
->host
;
395 struct nfs_direct_req
*dreq
;
397 dreq
= nfs_direct_req_alloc();
402 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
403 if (!is_sync_kiocb(iocb
))
406 result
= nfs_direct_read_schedule_iovec(dreq
, iov
, nr_segs
, pos
);
408 result
= nfs_direct_wait(dreq
);
409 nfs_direct_req_release(dreq
);
414 static void nfs_direct_free_writedata(struct nfs_direct_req
*dreq
)
416 while (!list_empty(&dreq
->rewrite_list
)) {
417 struct nfs_write_data
*data
= list_entry(dreq
->rewrite_list
.next
, struct nfs_write_data
, pages
);
418 list_del(&data
->pages
);
419 nfs_direct_release_pages(data
->pagevec
, data
->npages
);
420 nfs_writedata_release(data
);
424 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
425 static void nfs_direct_write_reschedule(struct nfs_direct_req
*dreq
)
427 struct inode
*inode
= dreq
->inode
;
429 struct nfs_write_data
*data
;
434 list_for_each(p
, &dreq
->rewrite_list
) {
435 data
= list_entry(p
, struct nfs_write_data
, pages
);
442 nfs_fattr_init(&data
->fattr
);
443 data
->res
.count
= data
->args
.count
;
444 memset(&data
->verf
, 0, sizeof(data
->verf
));
447 * Reuse data->task; data->args should not have changed
448 * since the original request was sent.
450 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), RPC_TASK_ASYNC
,
451 &nfs_write_direct_ops
, data
);
452 NFS_PROTO(inode
)->write_setup(data
, FLUSH_STABLE
);
454 data
->task
.tk_priority
= RPC_PRIORITY_NORMAL
;
455 data
->task
.tk_cookie
= (unsigned long) inode
;
458 * We're called via an RPC callback, so BKL is already held.
460 rpc_execute(&data
->task
);
462 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
465 (long long)NFS_FILEID(inode
),
467 (unsigned long long)data
->args
.offset
);
471 nfs_direct_write_complete(dreq
, inode
);
474 static void nfs_direct_commit_result(struct rpc_task
*task
, void *calldata
)
476 struct nfs_write_data
*data
= calldata
;
477 struct nfs_direct_req
*dreq
= (struct nfs_direct_req
*) data
->req
;
479 /* Call the NFS version-specific code */
480 if (NFS_PROTO(data
->inode
)->commit_done(task
, data
) != 0)
482 if (unlikely(task
->tk_status
< 0)) {
483 dprintk("NFS: %5u commit failed with error %d.\n",
484 task
->tk_pid
, task
->tk_status
);
485 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
486 } else if (memcmp(&dreq
->verf
, &data
->verf
, sizeof(data
->verf
))) {
487 dprintk("NFS: %5u commit verify failed\n", task
->tk_pid
);
488 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
491 dprintk("NFS: %5u commit returned %d\n", task
->tk_pid
, task
->tk_status
);
492 nfs_direct_write_complete(dreq
, data
->inode
);
495 static const struct rpc_call_ops nfs_commit_direct_ops
= {
496 .rpc_call_done
= nfs_direct_commit_result
,
497 .rpc_release
= nfs_commit_release
,
500 static void nfs_direct_commit_schedule(struct nfs_direct_req
*dreq
)
502 struct nfs_write_data
*data
= dreq
->commit_data
;
504 data
->inode
= dreq
->inode
;
505 data
->cred
= dreq
->ctx
->cred
;
507 data
->args
.fh
= NFS_FH(data
->inode
);
508 data
->args
.offset
= 0;
509 data
->args
.count
= 0;
511 data
->res
.fattr
= &data
->fattr
;
512 data
->res
.verf
= &data
->verf
;
514 rpc_init_task(&data
->task
, NFS_CLIENT(dreq
->inode
), RPC_TASK_ASYNC
,
515 &nfs_commit_direct_ops
, data
);
516 NFS_PROTO(data
->inode
)->commit_setup(data
, 0);
518 data
->task
.tk_priority
= RPC_PRIORITY_NORMAL
;
519 data
->task
.tk_cookie
= (unsigned long)data
->inode
;
520 /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
521 dreq
->commit_data
= NULL
;
523 dprintk("NFS: %5u initiated commit call\n", data
->task
.tk_pid
);
525 rpc_execute(&data
->task
);
528 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
, struct inode
*inode
)
530 int flags
= dreq
->flags
;
534 case NFS_ODIRECT_DO_COMMIT
:
535 nfs_direct_commit_schedule(dreq
);
537 case NFS_ODIRECT_RESCHED_WRITES
:
538 nfs_direct_write_reschedule(dreq
);
541 if (dreq
->commit_data
!= NULL
)
542 nfs_commit_free(dreq
->commit_data
);
543 nfs_direct_free_writedata(dreq
);
544 nfs_zap_mapping(inode
, inode
->i_mapping
);
545 nfs_direct_complete(dreq
);
549 static void nfs_alloc_commit_data(struct nfs_direct_req
*dreq
)
551 dreq
->commit_data
= nfs_commit_alloc();
552 if (dreq
->commit_data
!= NULL
)
553 dreq
->commit_data
->req
= (struct nfs_page
*) dreq
;
556 static inline void nfs_alloc_commit_data(struct nfs_direct_req
*dreq
)
558 dreq
->commit_data
= NULL
;
561 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
, struct inode
*inode
)
563 nfs_direct_free_writedata(dreq
);
564 nfs_zap_mapping(inode
, inode
->i_mapping
);
565 nfs_direct_complete(dreq
);
569 static void nfs_direct_write_result(struct rpc_task
*task
, void *calldata
)
571 struct nfs_write_data
*data
= calldata
;
572 struct nfs_direct_req
*dreq
= (struct nfs_direct_req
*) data
->req
;
573 int status
= task
->tk_status
;
575 if (nfs_writeback_done(task
, data
) != 0)
578 spin_lock(&dreq
->lock
);
580 if (unlikely(status
< 0)) {
581 /* An error has occurred, so we should not commit */
583 dreq
->error
= status
;
585 if (unlikely(dreq
->error
!= 0))
588 dreq
->count
+= data
->res
.count
;
590 if (data
->res
.verf
->committed
!= NFS_FILE_SYNC
) {
591 switch (dreq
->flags
) {
593 memcpy(&dreq
->verf
, &data
->verf
, sizeof(dreq
->verf
));
594 dreq
->flags
= NFS_ODIRECT_DO_COMMIT
;
596 case NFS_ODIRECT_DO_COMMIT
:
597 if (memcmp(&dreq
->verf
, &data
->verf
, sizeof(dreq
->verf
))) {
598 dprintk("NFS: %5u write verify failed\n", task
->tk_pid
);
599 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
604 spin_unlock(&dreq
->lock
);
608 * NB: Return the value of the first error return code. Subsequent
609 * errors after the first one are ignored.
611 static void nfs_direct_write_release(void *calldata
)
613 struct nfs_write_data
*data
= calldata
;
614 struct nfs_direct_req
*dreq
= (struct nfs_direct_req
*) data
->req
;
617 nfs_direct_write_complete(dreq
, data
->inode
);
620 static const struct rpc_call_ops nfs_write_direct_ops
= {
621 .rpc_call_done
= nfs_direct_write_result
,
622 .rpc_release
= nfs_direct_write_release
,
626 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
627 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
628 * bail and stop sending more writes. Write length accounting is
629 * handled automatically by nfs_direct_write_result(). Otherwise, if
630 * no requests have been sent, just return an error.
632 static ssize_t
nfs_direct_write_schedule_segment(struct nfs_direct_req
*dreq
,
633 const struct iovec
*iov
,
634 loff_t pos
, int sync
)
636 struct nfs_open_context
*ctx
= dreq
->ctx
;
637 struct inode
*inode
= ctx
->path
.dentry
->d_inode
;
638 unsigned long user_addr
= (unsigned long)iov
->iov_base
;
639 size_t count
= iov
->iov_len
;
640 size_t wsize
= NFS_SERVER(inode
)->wsize
;
646 struct nfs_write_data
*data
;
649 pgbase
= user_addr
& ~PAGE_MASK
;
650 bytes
= min(wsize
,count
);
653 data
= nfs_writedata_alloc(nfs_page_array_len(pgbase
, bytes
));
657 down_read(¤t
->mm
->mmap_sem
);
658 result
= get_user_pages(current
, current
->mm
, user_addr
,
659 data
->npages
, 0, 0, data
->pagevec
, NULL
);
660 up_read(¤t
->mm
->mmap_sem
);
662 nfs_writedata_release(data
);
665 if ((unsigned)result
< data
->npages
) {
666 bytes
= result
* PAGE_SIZE
;
667 if (bytes
<= pgbase
) {
668 nfs_direct_release_pages(data
->pagevec
, result
);
669 nfs_writedata_release(data
);
673 data
->npages
= result
;
678 list_move_tail(&data
->pages
, &dreq
->rewrite_list
);
680 data
->req
= (struct nfs_page
*) dreq
;
682 data
->cred
= ctx
->cred
;
683 data
->args
.fh
= NFS_FH(inode
);
684 data
->args
.context
= ctx
;
685 data
->args
.offset
= pos
;
686 data
->args
.pgbase
= pgbase
;
687 data
->args
.pages
= data
->pagevec
;
688 data
->args
.count
= bytes
;
689 data
->res
.fattr
= &data
->fattr
;
690 data
->res
.count
= bytes
;
691 data
->res
.verf
= &data
->verf
;
693 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), RPC_TASK_ASYNC
,
694 &nfs_write_direct_ops
, data
);
695 NFS_PROTO(inode
)->write_setup(data
, sync
);
697 data
->task
.tk_priority
= RPC_PRIORITY_NORMAL
;
698 data
->task
.tk_cookie
= (unsigned long) inode
;
700 rpc_execute(&data
->task
);
702 dprintk("NFS: %5u initiated direct write call "
703 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
706 (long long)NFS_FILEID(inode
),
708 (unsigned long long)data
->args
.offset
);
714 /* FIXME: Remove this useless math from the final patch */
716 pgbase
&= ~PAGE_MASK
;
717 BUG_ON(pgbase
!= (user_addr
& ~PAGE_MASK
));
720 } while (count
!= 0);
724 return result
< 0 ? (ssize_t
) result
: -EFAULT
;
727 static ssize_t
nfs_direct_write_schedule_iovec(struct nfs_direct_req
*dreq
,
728 const struct iovec
*iov
,
729 unsigned long nr_segs
,
730 loff_t pos
, int sync
)
733 size_t requested_bytes
= 0;
738 for (seg
= 0; seg
< nr_segs
; seg
++) {
739 const struct iovec
*vec
= &iov
[seg
];
740 result
= nfs_direct_write_schedule_segment(dreq
, vec
,
744 requested_bytes
+= result
;
745 if ((size_t)result
< vec
->iov_len
)
751 nfs_direct_write_complete(dreq
, dreq
->inode
);
753 if (requested_bytes
!= 0)
761 static ssize_t
nfs_direct_write(struct kiocb
*iocb
, const struct iovec
*iov
,
762 unsigned long nr_segs
, loff_t pos
,
766 struct inode
*inode
= iocb
->ki_filp
->f_mapping
->host
;
767 struct nfs_direct_req
*dreq
;
768 size_t wsize
= NFS_SERVER(inode
)->wsize
;
771 dreq
= nfs_direct_req_alloc();
774 nfs_alloc_commit_data(dreq
);
776 if (dreq
->commit_data
== NULL
|| count
< wsize
)
780 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
781 if (!is_sync_kiocb(iocb
))
784 result
= nfs_direct_write_schedule_iovec(dreq
, iov
, nr_segs
, pos
, sync
);
786 result
= nfs_direct_wait(dreq
);
787 nfs_direct_req_release(dreq
);
793 * nfs_file_direct_read - file direct read operation for NFS files
794 * @iocb: target I/O control block
795 * @iov: vector of user buffers into which to read data
796 * @nr_segs: size of iov vector
797 * @pos: byte offset in file where reading starts
799 * We use this function for direct reads instead of calling
800 * generic_file_aio_read() in order to avoid gfar's check to see if
801 * the request starts before the end of the file. For that check
802 * to work, we must generate a GETATTR before each direct read, and
803 * even then there is a window between the GETATTR and the subsequent
804 * READ where the file size could change. Our preference is simply
805 * to do all reads the application wants, and the server will take
806 * care of managing the end of file boundary.
808 * This function also eliminates unnecessarily updating the file's
809 * atime locally, as the NFS server sets the file's atime, and this
810 * client must read the updated atime from the server back into its
813 ssize_t
nfs_file_direct_read(struct kiocb
*iocb
, const struct iovec
*iov
,
814 unsigned long nr_segs
, loff_t pos
)
816 ssize_t retval
= -EINVAL
;
817 struct file
*file
= iocb
->ki_filp
;
818 struct address_space
*mapping
= file
->f_mapping
;
821 count
= iov_length(iov
, nr_segs
);
822 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTREADBYTES
, count
);
824 dprintk("nfs: direct read(%s/%s, %zd@%Ld)\n",
825 file
->f_path
.dentry
->d_parent
->d_name
.name
,
826 file
->f_path
.dentry
->d_name
.name
,
827 count
, (long long) pos
);
833 retval
= nfs_sync_mapping(mapping
);
837 retval
= nfs_direct_read(iocb
, iov
, nr_segs
, pos
);
839 iocb
->ki_pos
= pos
+ retval
;
846 * nfs_file_direct_write - file direct write operation for NFS files
847 * @iocb: target I/O control block
848 * @iov: vector of user buffers from which to write data
849 * @nr_segs: size of iov vector
850 * @pos: byte offset in file where writing starts
852 * We use this function for direct writes instead of calling
853 * generic_file_aio_write() in order to avoid taking the inode
854 * semaphore and updating the i_size. The NFS server will set
855 * the new i_size and this client must read the updated size
856 * back into its cache. We let the server do generic write
857 * parameter checking and report problems.
859 * We also avoid an unnecessary invocation of generic_osync_inode(),
860 * as it is fairly meaningless to sync the metadata of an NFS file.
862 * We eliminate local atime updates, see direct read above.
864 * We avoid unnecessary page cache invalidations for normal cached
865 * readers of this file.
867 * Note that O_APPEND is not supported for NFS direct writes, as there
868 * is no atomic O_APPEND write facility in the NFS protocol.
870 ssize_t
nfs_file_direct_write(struct kiocb
*iocb
, const struct iovec
*iov
,
871 unsigned long nr_segs
, loff_t pos
)
873 ssize_t retval
= -EINVAL
;
874 struct file
*file
= iocb
->ki_filp
;
875 struct address_space
*mapping
= file
->f_mapping
;
878 count
= iov_length(iov
, nr_segs
);
879 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTWRITTENBYTES
, count
);
881 dfprintk(VFS
, "nfs: direct write(%s/%s, %zd@%Ld)\n",
882 file
->f_path
.dentry
->d_parent
->d_name
.name
,
883 file
->f_path
.dentry
->d_name
.name
,
884 count
, (long long) pos
);
886 retval
= generic_write_checks(file
, &pos
, &count
, 0);
890 goto out
; /* return 0 */
893 if ((ssize_t
) count
< 0)
899 retval
= nfs_sync_mapping(mapping
);
903 retval
= nfs_direct_write(iocb
, iov
, nr_segs
, pos
, count
);
906 iocb
->ki_pos
= pos
+ retval
;
913 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
916 int __init
nfs_init_directcache(void)
918 nfs_direct_cachep
= kmem_cache_create("nfs_direct_cache",
919 sizeof(struct nfs_direct_req
),
920 0, (SLAB_RECLAIM_ACCOUNT
|
923 if (nfs_direct_cachep
== NULL
)
930 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
933 void nfs_destroy_directcache(void)
935 kmem_cache_destroy(nfs_direct_cachep
);