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>
50 #include <linux/nfs_fs.h>
51 #include <linux/nfs_page.h>
52 #include <linux/sunrpc/clnt.h>
54 #include <asm/uaccess.h>
55 #include <linux/atomic.h>
61 #define NFSDBG_FACILITY NFSDBG_VFS
63 static struct kmem_cache
*nfs_direct_cachep
;
66 * This represents a set of asynchronous requests that we're waiting on
68 struct nfs_direct_req
{
69 struct kref kref
; /* release manager */
72 struct nfs_open_context
*ctx
; /* file open context info */
73 struct nfs_lock_context
*l_ctx
; /* Lock context info */
74 struct kiocb
* iocb
; /* controlling i/o request */
75 struct inode
* inode
; /* target file of i/o */
77 /* completion state */
78 atomic_t io_count
; /* i/os we're waiting for */
79 spinlock_t lock
; /* protect completion state */
80 ssize_t count
, /* bytes actually processed */
81 error
; /* any reported error */
82 struct completion completion
; /* wait for i/o completion */
85 struct nfs_mds_commit_info mds_cinfo
; /* Storage for cinfo */
86 struct pnfs_ds_commit_info ds_cinfo
; /* Storage for cinfo */
87 struct work_struct work
;
89 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
90 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
91 struct nfs_writeverf verf
; /* unstable write verifier */
94 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops
;
95 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops
;
96 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
, struct inode
*inode
);
97 static void nfs_direct_write_schedule_work(struct work_struct
*work
);
99 static inline void get_dreq(struct nfs_direct_req
*dreq
)
101 atomic_inc(&dreq
->io_count
);
104 static inline int put_dreq(struct nfs_direct_req
*dreq
)
106 return atomic_dec_and_test(&dreq
->io_count
);
110 * nfs_direct_IO - NFS address space operation for direct I/O
111 * @rw: direction (read or write)
112 * @iocb: target I/O control block
113 * @iov: array of vectors that define I/O buffer
114 * @pos: offset in file to begin the operation
115 * @nr_segs: size of iovec array
117 * The presence of this routine in the address space ops vector means
118 * the NFS client supports direct I/O. However, we shunt off direct
119 * read and write requests before the VFS gets them, so this method
120 * should never be called.
122 ssize_t
nfs_direct_IO(int rw
, struct kiocb
*iocb
, const struct iovec
*iov
, loff_t pos
, unsigned long nr_segs
)
124 dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
125 iocb
->ki_filp
->f_path
.dentry
->d_name
.name
,
126 (long long) pos
, nr_segs
);
131 static void nfs_direct_release_pages(struct page
**pages
, unsigned int npages
)
134 for (i
= 0; i
< npages
; i
++)
135 page_cache_release(pages
[i
]);
138 void nfs_init_cinfo_from_dreq(struct nfs_commit_info
*cinfo
,
139 struct nfs_direct_req
*dreq
)
141 cinfo
->lock
= &dreq
->lock
;
142 cinfo
->mds
= &dreq
->mds_cinfo
;
143 cinfo
->ds
= &dreq
->ds_cinfo
;
145 cinfo
->completion_ops
= &nfs_direct_commit_completion_ops
;
148 static inline struct nfs_direct_req
*nfs_direct_req_alloc(void)
150 struct nfs_direct_req
*dreq
;
152 dreq
= kmem_cache_zalloc(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
->mds_cinfo
.list
);
160 INIT_WORK(&dreq
->work
, nfs_direct_write_schedule_work
);
161 spin_lock_init(&dreq
->lock
);
166 static void nfs_direct_req_free(struct kref
*kref
)
168 struct nfs_direct_req
*dreq
= container_of(kref
, struct nfs_direct_req
, kref
);
170 if (dreq
->l_ctx
!= NULL
)
171 nfs_put_lock_context(dreq
->l_ctx
);
172 if (dreq
->ctx
!= NULL
)
173 put_nfs_open_context(dreq
->ctx
);
174 kmem_cache_free(nfs_direct_cachep
, dreq
);
177 static void nfs_direct_req_release(struct nfs_direct_req
*dreq
)
179 kref_put(&dreq
->kref
, nfs_direct_req_free
);
183 * Collects and returns the final error value/byte-count.
185 static ssize_t
nfs_direct_wait(struct nfs_direct_req
*dreq
)
187 ssize_t result
= -EIOCBQUEUED
;
189 /* Async requests don't wait here */
193 result
= wait_for_completion_killable(&dreq
->completion
);
196 result
= dreq
->error
;
198 result
= dreq
->count
;
201 return (ssize_t
) result
;
205 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
206 * the iocb is still valid here if this is a synchronous request.
208 static void nfs_direct_complete(struct nfs_direct_req
*dreq
)
211 long res
= (long) dreq
->error
;
213 res
= (long) dreq
->count
;
214 aio_complete(dreq
->iocb
, res
, 0);
216 complete_all(&dreq
->completion
);
218 nfs_direct_req_release(dreq
);
221 static void nfs_direct_readpage_release(struct nfs_page
*req
)
223 dprintk("NFS: direct read done (%s/%lld %d@%lld)\n",
224 req
->wb_context
->dentry
->d_inode
->i_sb
->s_id
,
225 (long long)NFS_FILEID(req
->wb_context
->dentry
->d_inode
),
227 (long long)req_offset(req
));
228 nfs_release_request(req
);
231 static void nfs_direct_read_completion(struct nfs_pgio_header
*hdr
)
233 unsigned long bytes
= 0;
234 struct nfs_direct_req
*dreq
= hdr
->dreq
;
236 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
239 spin_lock(&dreq
->lock
);
240 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
) && (hdr
->good_bytes
== 0))
241 dreq
->error
= hdr
->error
;
243 dreq
->count
+= hdr
->good_bytes
;
244 spin_unlock(&dreq
->lock
);
246 while (!list_empty(&hdr
->pages
)) {
247 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
248 struct page
*page
= req
->wb_page
;
250 if (test_bit(NFS_IOHDR_EOF
, &hdr
->flags
)) {
251 if (bytes
> hdr
->good_bytes
)
252 zero_user(page
, 0, PAGE_SIZE
);
253 else if (hdr
->good_bytes
- bytes
< PAGE_SIZE
)
254 zero_user_segment(page
,
255 hdr
->good_bytes
& ~PAGE_MASK
,
258 if (!PageCompound(page
)) {
259 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
)) {
260 if (bytes
< hdr
->good_bytes
)
261 set_page_dirty(page
);
263 set_page_dirty(page
);
265 bytes
+= req
->wb_bytes
;
266 nfs_list_remove_request(req
);
267 nfs_direct_readpage_release(req
);
271 nfs_direct_complete(dreq
);
275 static void nfs_read_sync_pgio_error(struct list_head
*head
)
277 struct nfs_page
*req
;
279 while (!list_empty(head
)) {
280 req
= nfs_list_entry(head
->next
);
281 nfs_list_remove_request(req
);
282 nfs_release_request(req
);
286 static void nfs_direct_pgio_init(struct nfs_pgio_header
*hdr
)
291 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops
= {
292 .error_cleanup
= nfs_read_sync_pgio_error
,
293 .init_hdr
= nfs_direct_pgio_init
,
294 .completion
= nfs_direct_read_completion
,
298 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
299 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
300 * bail and stop sending more reads. Read length accounting is
301 * handled automatically by nfs_direct_read_result(). Otherwise, if
302 * no requests have been sent, just return an error.
304 static ssize_t
nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor
*desc
,
305 const struct iovec
*iov
,
308 struct nfs_direct_req
*dreq
= desc
->pg_dreq
;
309 struct nfs_open_context
*ctx
= dreq
->ctx
;
310 struct inode
*inode
= ctx
->dentry
->d_inode
;
311 unsigned long user_addr
= (unsigned long)iov
->iov_base
;
312 size_t count
= iov
->iov_len
;
313 size_t rsize
= NFS_SERVER(inode
)->rsize
;
317 struct page
**pagevec
= NULL
;
324 pgbase
= user_addr
& ~PAGE_MASK
;
325 bytes
= min(max_t(size_t, rsize
, PAGE_SIZE
), count
);
328 npages
= nfs_page_array_len(pgbase
, bytes
);
330 pagevec
= kmalloc(npages
* sizeof(struct page
*),
334 down_read(¤t
->mm
->mmap_sem
);
335 result
= get_user_pages(current
, current
->mm
, user_addr
,
336 npages
, 1, 0, pagevec
, NULL
);
337 up_read(¤t
->mm
->mmap_sem
);
340 if ((unsigned)result
< npages
) {
341 bytes
= result
* PAGE_SIZE
;
342 if (bytes
<= pgbase
) {
343 nfs_direct_release_pages(pagevec
, result
);
350 for (i
= 0; i
< npages
; i
++) {
351 struct nfs_page
*req
;
352 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
353 /* XXX do we need to do the eof zeroing found in async_filler? */
354 req
= nfs_create_request(dreq
->ctx
, dreq
->inode
,
358 result
= PTR_ERR(req
);
361 req
->wb_index
= pos
>> PAGE_SHIFT
;
362 req
->wb_offset
= pos
& ~PAGE_MASK
;
363 if (!nfs_pageio_add_request(desc
, req
)) {
364 result
= desc
->pg_error
;
365 nfs_release_request(req
);
371 user_addr
+= req_len
;
375 /* The nfs_page now hold references to these pages */
376 nfs_direct_release_pages(pagevec
, npages
);
377 } while (count
!= 0 && result
>= 0);
383 return result
< 0 ? (ssize_t
) result
: -EFAULT
;
386 static ssize_t
nfs_direct_read_schedule_iovec(struct nfs_direct_req
*dreq
,
387 const struct iovec
*iov
,
388 unsigned long nr_segs
,
391 struct nfs_pageio_descriptor desc
;
392 ssize_t result
= -EINVAL
;
393 size_t requested_bytes
= 0;
396 nfs_pageio_init_read(&desc
, dreq
->inode
,
397 &nfs_direct_read_completion_ops
);
401 for (seg
= 0; seg
< nr_segs
; seg
++) {
402 const struct iovec
*vec
= &iov
[seg
];
403 result
= nfs_direct_read_schedule_segment(&desc
, vec
, pos
);
406 requested_bytes
+= result
;
407 if ((size_t)result
< vec
->iov_len
)
412 nfs_pageio_complete(&desc
);
415 * If no bytes were started, return the error, and let the
416 * generic layer handle the completion.
418 if (requested_bytes
== 0) {
419 nfs_direct_req_release(dreq
);
420 return result
< 0 ? result
: -EIO
;
424 nfs_direct_complete(dreq
);
428 static ssize_t
nfs_direct_read(struct kiocb
*iocb
, const struct iovec
*iov
,
429 unsigned long nr_segs
, loff_t pos
)
431 ssize_t result
= -ENOMEM
;
432 struct inode
*inode
= iocb
->ki_filp
->f_mapping
->host
;
433 struct nfs_direct_req
*dreq
;
435 dreq
= nfs_direct_req_alloc();
440 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
441 dreq
->l_ctx
= nfs_get_lock_context(dreq
->ctx
);
442 if (dreq
->l_ctx
== NULL
)
444 if (!is_sync_kiocb(iocb
))
447 result
= nfs_direct_read_schedule_iovec(dreq
, iov
, nr_segs
, pos
);
449 result
= nfs_direct_wait(dreq
);
450 NFS_I(inode
)->read_io
+= result
;
452 nfs_direct_req_release(dreq
);
457 static void nfs_inode_dio_write_done(struct inode
*inode
)
459 nfs_zap_mapping(inode
, inode
->i_mapping
);
460 inode_dio_done(inode
);
463 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
464 static void nfs_direct_write_reschedule(struct nfs_direct_req
*dreq
)
466 struct nfs_pageio_descriptor desc
;
467 struct nfs_page
*req
, *tmp
;
469 struct nfs_commit_info cinfo
;
472 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
473 pnfs_recover_commit_reqs(dreq
->inode
, &reqs
, &cinfo
);
474 spin_lock(cinfo
.lock
);
475 nfs_scan_commit_list(&cinfo
.mds
->list
, &reqs
, &cinfo
, 0);
476 spin_unlock(cinfo
.lock
);
481 nfs_pageio_init_write(&desc
, dreq
->inode
, FLUSH_STABLE
,
482 &nfs_direct_write_completion_ops
);
485 list_for_each_entry_safe(req
, tmp
, &reqs
, wb_list
) {
486 if (!nfs_pageio_add_request(&desc
, req
)) {
487 nfs_list_add_request(req
, &failed
);
488 spin_lock(cinfo
.lock
);
491 spin_unlock(cinfo
.lock
);
494 nfs_pageio_complete(&desc
);
496 while (!list_empty(&failed
))
497 nfs_unlock_and_release_request(req
);
500 nfs_direct_write_complete(dreq
, dreq
->inode
);
503 static void nfs_direct_commit_complete(struct nfs_commit_data
*data
)
505 struct nfs_direct_req
*dreq
= data
->dreq
;
506 struct nfs_commit_info cinfo
;
507 struct nfs_page
*req
;
508 int status
= data
->task
.tk_status
;
510 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
512 dprintk("NFS: %5u commit failed with error %d.\n",
513 data
->task
.tk_pid
, status
);
514 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
515 } else if (memcmp(&dreq
->verf
, &data
->verf
, sizeof(data
->verf
))) {
516 dprintk("NFS: %5u commit verify failed\n", data
->task
.tk_pid
);
517 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
520 dprintk("NFS: %5u commit returned %d\n", data
->task
.tk_pid
, status
);
521 while (!list_empty(&data
->pages
)) {
522 req
= nfs_list_entry(data
->pages
.next
);
523 nfs_list_remove_request(req
);
524 if (dreq
->flags
== NFS_ODIRECT_RESCHED_WRITES
) {
525 /* Note the rewrite will go through mds */
526 kref_get(&req
->wb_kref
);
527 nfs_mark_request_commit(req
, NULL
, &cinfo
);
529 nfs_unlock_and_release_request(req
);
532 if (atomic_dec_and_test(&cinfo
.mds
->rpcs_out
))
533 nfs_direct_write_complete(dreq
, data
->inode
);
536 static void nfs_direct_error_cleanup(struct nfs_inode
*nfsi
)
538 /* There is no lock to clear */
541 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops
= {
542 .completion
= nfs_direct_commit_complete
,
543 .error_cleanup
= nfs_direct_error_cleanup
,
546 static void nfs_direct_commit_schedule(struct nfs_direct_req
*dreq
)
549 struct nfs_commit_info cinfo
;
552 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
553 nfs_scan_commit(dreq
->inode
, &mds_list
, &cinfo
);
554 res
= nfs_generic_commit_list(dreq
->inode
, &mds_list
, 0, &cinfo
);
555 if (res
< 0) /* res == -ENOMEM */
556 nfs_direct_write_reschedule(dreq
);
559 static void nfs_direct_write_schedule_work(struct work_struct
*work
)
561 struct nfs_direct_req
*dreq
= container_of(work
, struct nfs_direct_req
, work
);
562 int flags
= dreq
->flags
;
566 case NFS_ODIRECT_DO_COMMIT
:
567 nfs_direct_commit_schedule(dreq
);
569 case NFS_ODIRECT_RESCHED_WRITES
:
570 nfs_direct_write_reschedule(dreq
);
573 nfs_inode_dio_write_done(dreq
->inode
);
574 nfs_direct_complete(dreq
);
578 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
, struct inode
*inode
)
580 schedule_work(&dreq
->work
); /* Calls nfs_direct_write_schedule_work */
584 static void nfs_direct_write_schedule_work(struct work_struct
*work
)
588 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
, struct inode
*inode
)
590 nfs_inode_dio_write_done(inode
);
591 nfs_direct_complete(dreq
);
596 * NB: Return the value of the first error return code. Subsequent
597 * errors after the first one are ignored.
600 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
601 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
602 * bail and stop sending more writes. Write length accounting is
603 * handled automatically by nfs_direct_write_result(). Otherwise, if
604 * no requests have been sent, just return an error.
606 static ssize_t
nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor
*desc
,
607 const struct iovec
*iov
,
610 struct nfs_direct_req
*dreq
= desc
->pg_dreq
;
611 struct nfs_open_context
*ctx
= dreq
->ctx
;
612 struct inode
*inode
= ctx
->dentry
->d_inode
;
613 unsigned long user_addr
= (unsigned long)iov
->iov_base
;
614 size_t count
= iov
->iov_len
;
615 size_t wsize
= NFS_SERVER(inode
)->wsize
;
619 struct page
**pagevec
= NULL
;
626 pgbase
= user_addr
& ~PAGE_MASK
;
627 bytes
= min(max_t(size_t, wsize
, PAGE_SIZE
), count
);
630 npages
= nfs_page_array_len(pgbase
, bytes
);
632 pagevec
= kmalloc(npages
* sizeof(struct page
*), GFP_KERNEL
);
636 down_read(¤t
->mm
->mmap_sem
);
637 result
= get_user_pages(current
, current
->mm
, user_addr
,
638 npages
, 0, 0, pagevec
, NULL
);
639 up_read(¤t
->mm
->mmap_sem
);
643 if ((unsigned)result
< npages
) {
644 bytes
= result
* PAGE_SIZE
;
645 if (bytes
<= pgbase
) {
646 nfs_direct_release_pages(pagevec
, result
);
653 for (i
= 0; i
< npages
; i
++) {
654 struct nfs_page
*req
;
655 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
657 req
= nfs_create_request(dreq
->ctx
, dreq
->inode
,
661 result
= PTR_ERR(req
);
664 nfs_lock_request(req
);
665 req
->wb_index
= pos
>> PAGE_SHIFT
;
666 req
->wb_offset
= pos
& ~PAGE_MASK
;
667 if (!nfs_pageio_add_request(desc
, req
)) {
668 result
= desc
->pg_error
;
669 nfs_unlock_and_release_request(req
);
675 user_addr
+= req_len
;
679 /* The nfs_page now hold references to these pages */
680 nfs_direct_release_pages(pagevec
, npages
);
681 } while (count
!= 0 && result
>= 0);
687 return result
< 0 ? (ssize_t
) result
: -EFAULT
;
690 static void nfs_direct_write_completion(struct nfs_pgio_header
*hdr
)
692 struct nfs_direct_req
*dreq
= hdr
->dreq
;
693 struct nfs_commit_info cinfo
;
695 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
697 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
700 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
702 spin_lock(&dreq
->lock
);
704 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
)) {
706 dreq
->error
= hdr
->error
;
708 if (dreq
->error
!= 0)
709 bit
= NFS_IOHDR_ERROR
;
711 dreq
->count
+= hdr
->good_bytes
;
712 if (test_bit(NFS_IOHDR_NEED_RESCHED
, &hdr
->flags
)) {
713 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
714 bit
= NFS_IOHDR_NEED_RESCHED
;
715 } else if (test_bit(NFS_IOHDR_NEED_COMMIT
, &hdr
->flags
)) {
716 if (dreq
->flags
== NFS_ODIRECT_RESCHED_WRITES
)
717 bit
= NFS_IOHDR_NEED_RESCHED
;
718 else if (dreq
->flags
== 0) {
719 memcpy(&dreq
->verf
, &req
->wb_verf
,
721 bit
= NFS_IOHDR_NEED_COMMIT
;
722 dreq
->flags
= NFS_ODIRECT_DO_COMMIT
;
723 } else if (dreq
->flags
== NFS_ODIRECT_DO_COMMIT
) {
724 if (memcmp(&dreq
->verf
, &req
->wb_verf
, sizeof(dreq
->verf
))) {
725 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
726 bit
= NFS_IOHDR_NEED_RESCHED
;
728 bit
= NFS_IOHDR_NEED_COMMIT
;
732 spin_unlock(&dreq
->lock
);
734 while (!list_empty(&hdr
->pages
)) {
735 req
= nfs_list_entry(hdr
->pages
.next
);
736 nfs_list_remove_request(req
);
738 case NFS_IOHDR_NEED_RESCHED
:
739 case NFS_IOHDR_NEED_COMMIT
:
740 kref_get(&req
->wb_kref
);
741 nfs_mark_request_commit(req
, hdr
->lseg
, &cinfo
);
743 nfs_unlock_and_release_request(req
);
748 nfs_direct_write_complete(dreq
, hdr
->inode
);
752 static void nfs_write_sync_pgio_error(struct list_head
*head
)
754 struct nfs_page
*req
;
756 while (!list_empty(head
)) {
757 req
= nfs_list_entry(head
->next
);
758 nfs_list_remove_request(req
);
759 nfs_unlock_and_release_request(req
);
763 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops
= {
764 .error_cleanup
= nfs_write_sync_pgio_error
,
765 .init_hdr
= nfs_direct_pgio_init
,
766 .completion
= nfs_direct_write_completion
,
769 static ssize_t
nfs_direct_write_schedule_iovec(struct nfs_direct_req
*dreq
,
770 const struct iovec
*iov
,
771 unsigned long nr_segs
,
774 struct nfs_pageio_descriptor desc
;
775 struct inode
*inode
= dreq
->inode
;
777 size_t requested_bytes
= 0;
780 nfs_pageio_init_write(&desc
, inode
, FLUSH_COND_STABLE
,
781 &nfs_direct_write_completion_ops
);
784 atomic_inc(&inode
->i_dio_count
);
786 for (seg
= 0; seg
< nr_segs
; seg
++) {
787 const struct iovec
*vec
= &iov
[seg
];
788 result
= nfs_direct_write_schedule_segment(&desc
, vec
, pos
);
791 requested_bytes
+= result
;
792 if ((size_t)result
< vec
->iov_len
)
796 nfs_pageio_complete(&desc
);
797 NFS_I(dreq
->inode
)->write_io
+= desc
.pg_bytes_written
;
800 * If no bytes were started, return the error, and let the
801 * generic layer handle the completion.
803 if (requested_bytes
== 0) {
804 inode_dio_done(inode
);
805 nfs_direct_req_release(dreq
);
806 return result
< 0 ? result
: -EIO
;
810 nfs_direct_write_complete(dreq
, dreq
->inode
);
814 static ssize_t
nfs_direct_write(struct kiocb
*iocb
, const struct iovec
*iov
,
815 unsigned long nr_segs
, loff_t pos
,
818 ssize_t result
= -ENOMEM
;
819 struct inode
*inode
= iocb
->ki_filp
->f_mapping
->host
;
820 struct nfs_direct_req
*dreq
;
822 dreq
= nfs_direct_req_alloc();
827 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
828 dreq
->l_ctx
= nfs_get_lock_context(dreq
->ctx
);
829 if (dreq
->l_ctx
== NULL
)
831 if (!is_sync_kiocb(iocb
))
834 result
= nfs_direct_write_schedule_iovec(dreq
, iov
, nr_segs
, pos
);
836 result
= nfs_direct_wait(dreq
);
838 nfs_direct_req_release(dreq
);
844 * nfs_file_direct_read - file direct read operation for NFS files
845 * @iocb: target I/O control block
846 * @iov: vector of user buffers into which to read data
847 * @nr_segs: size of iov vector
848 * @pos: byte offset in file where reading starts
850 * We use this function for direct reads instead of calling
851 * generic_file_aio_read() in order to avoid gfar's check to see if
852 * the request starts before the end of the file. For that check
853 * to work, we must generate a GETATTR before each direct read, and
854 * even then there is a window between the GETATTR and the subsequent
855 * READ where the file size could change. Our preference is simply
856 * to do all reads the application wants, and the server will take
857 * care of managing the end of file boundary.
859 * This function also eliminates unnecessarily updating the file's
860 * atime locally, as the NFS server sets the file's atime, and this
861 * client must read the updated atime from the server back into its
864 ssize_t
nfs_file_direct_read(struct kiocb
*iocb
, const struct iovec
*iov
,
865 unsigned long nr_segs
, loff_t pos
)
867 ssize_t retval
= -EINVAL
;
868 struct file
*file
= iocb
->ki_filp
;
869 struct address_space
*mapping
= file
->f_mapping
;
872 count
= iov_length(iov
, nr_segs
);
873 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTREADBYTES
, count
);
875 dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
876 file
->f_path
.dentry
->d_parent
->d_name
.name
,
877 file
->f_path
.dentry
->d_name
.name
,
878 count
, (long long) pos
);
884 retval
= nfs_sync_mapping(mapping
);
888 task_io_account_read(count
);
890 retval
= nfs_direct_read(iocb
, iov
, nr_segs
, pos
);
892 iocb
->ki_pos
= pos
+ retval
;
899 * nfs_file_direct_write - file direct write operation for NFS files
900 * @iocb: target I/O control block
901 * @iov: vector of user buffers from which to write data
902 * @nr_segs: size of iov vector
903 * @pos: byte offset in file where writing starts
905 * We use this function for direct writes instead of calling
906 * generic_file_aio_write() in order to avoid taking the inode
907 * semaphore and updating the i_size. The NFS server will set
908 * the new i_size and this client must read the updated size
909 * back into its cache. We let the server do generic write
910 * parameter checking and report problems.
912 * We eliminate local atime updates, see direct read above.
914 * We avoid unnecessary page cache invalidations for normal cached
915 * readers of this file.
917 * Note that O_APPEND is not supported for NFS direct writes, as there
918 * is no atomic O_APPEND write facility in the NFS protocol.
920 ssize_t
nfs_file_direct_write(struct kiocb
*iocb
, const struct iovec
*iov
,
921 unsigned long nr_segs
, loff_t pos
)
923 ssize_t retval
= -EINVAL
;
924 struct file
*file
= iocb
->ki_filp
;
925 struct address_space
*mapping
= file
->f_mapping
;
928 count
= iov_length(iov
, nr_segs
);
929 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTWRITTENBYTES
, count
);
931 dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
932 file
->f_path
.dentry
->d_parent
->d_name
.name
,
933 file
->f_path
.dentry
->d_name
.name
,
934 count
, (long long) pos
);
936 retval
= generic_write_checks(file
, &pos
, &count
, 0);
941 if ((ssize_t
) count
< 0)
947 retval
= nfs_sync_mapping(mapping
);
951 task_io_account_write(count
);
953 retval
= nfs_direct_write(iocb
, iov
, nr_segs
, pos
, count
);
955 struct inode
*inode
= mapping
->host
;
957 iocb
->ki_pos
= pos
+ retval
;
958 spin_lock(&inode
->i_lock
);
959 if (i_size_read(inode
) < iocb
->ki_pos
)
960 i_size_write(inode
, iocb
->ki_pos
);
961 spin_unlock(&inode
->i_lock
);
968 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
971 int __init
nfs_init_directcache(void)
973 nfs_direct_cachep
= kmem_cache_create("nfs_direct_cache",
974 sizeof(struct nfs_direct_req
),
975 0, (SLAB_RECLAIM_ACCOUNT
|
978 if (nfs_direct_cachep
== NULL
)
985 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
988 void nfs_destroy_directcache(void)
990 kmem_cache_destroy(nfs_direct_cachep
);