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
40 #include <linux/config.h>
41 #include <linux/errno.h>
42 #include <linux/sched.h>
43 #include <linux/kernel.h>
44 #include <linux/smp_lock.h>
45 #include <linux/file.h>
46 #include <linux/pagemap.h>
47 #include <linux/kref.h>
49 #include <linux/nfs_fs.h>
50 #include <linux/nfs_page.h>
51 #include <linux/sunrpc/clnt.h>
53 #include <asm/system.h>
54 #include <asm/uaccess.h>
55 #include <asm/atomic.h>
57 #define NFSDBG_FACILITY NFSDBG_VFS
58 #define MAX_DIRECTIO_SIZE (4096UL << PAGE_SHIFT)
60 static kmem_cache_t
*nfs_direct_cachep
;
63 * This represents a set of asynchronous requests that we're waiting on
65 struct nfs_direct_req
{
66 struct kref kref
; /* release manager */
67 struct list_head list
; /* nfs_read_data structs */
68 wait_queue_head_t wait
; /* wait for i/o completion */
69 struct page
** pages
; /* pages in our buffer */
70 unsigned int npages
; /* count of pages */
71 atomic_t complete
, /* i/os we're waiting for */
72 count
, /* bytes actually processed */
73 error
; /* any reported error */
78 * nfs_get_user_pages - find and set up pages underlying user's buffer
79 * rw: direction (read or write)
80 * user_addr: starting address of this segment of user's buffer
81 * count: size of this segment
82 * @pages: returned array of page struct pointers underlying user's buffer
85 nfs_get_user_pages(int rw
, unsigned long user_addr
, size_t size
,
89 unsigned long page_count
;
92 /* set an arbitrary limit to prevent type overflow */
93 /* XXX: this can probably be as large as INT_MAX */
94 if (size
> MAX_DIRECTIO_SIZE
) {
99 page_count
= (user_addr
+ size
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
100 page_count
-= user_addr
>> PAGE_SHIFT
;
102 array_size
= (page_count
* sizeof(struct page
*));
103 *pages
= kmalloc(array_size
, GFP_KERNEL
);
105 down_read(¤t
->mm
->mmap_sem
);
106 result
= get_user_pages(current
, current
->mm
, user_addr
,
107 page_count
, (rw
== READ
), 0,
109 up_read(¤t
->mm
->mmap_sem
);
115 * nfs_free_user_pages - tear down page struct array
116 * @pages: array of page struct pointers underlying target buffer
117 * @npages: number of pages in the array
118 * @do_dirty: dirty the pages as we release them
121 nfs_free_user_pages(struct page
**pages
, int npages
, int do_dirty
)
124 for (i
= 0; i
< npages
; i
++) {
126 set_page_dirty_lock(pages
[i
]);
127 page_cache_release(pages
[i
]);
133 * nfs_direct_req_release - release nfs_direct_req structure for direct read
134 * @kref: kref object embedded in an nfs_direct_req structure
137 static void nfs_direct_req_release(struct kref
*kref
)
139 struct nfs_direct_req
*dreq
= container_of(kref
, struct nfs_direct_req
, kref
);
140 kmem_cache_free(nfs_direct_cachep
, dreq
);
144 * nfs_direct_read_alloc - allocate nfs_read_data structures for direct read
145 * @count: count of bytes for the read request
146 * @rsize: local rsize setting
148 * Note we also set the number of requests we have in the dreq when we are
149 * done. This prevents races with I/O completion so we will always wait
150 * until all requests have been dispatched and completed.
152 static struct nfs_direct_req
*nfs_direct_read_alloc(size_t nbytes
, unsigned int rsize
)
154 struct list_head
*list
;
155 struct nfs_direct_req
*dreq
;
156 unsigned int reads
= 0;
158 dreq
= kmem_cache_alloc(nfs_direct_cachep
, SLAB_KERNEL
);
162 kref_init(&dreq
->kref
);
163 init_waitqueue_head(&dreq
->wait
);
164 INIT_LIST_HEAD(&dreq
->list
);
165 atomic_set(&dreq
->count
, 0);
166 atomic_set(&dreq
->error
, 0);
170 struct nfs_read_data
*data
= nfs_readdata_alloc();
172 if (unlikely(!data
)) {
173 while (!list_empty(list
)) {
174 data
= list_entry(list
->next
,
175 struct nfs_read_data
, pages
);
176 list_del(&data
->pages
);
177 nfs_readdata_free(data
);
179 kref_put(&dreq
->kref
, nfs_direct_req_release
);
183 INIT_LIST_HEAD(&data
->pages
);
184 list_add(&data
->pages
, list
);
186 data
->req
= (struct nfs_page
*) dreq
;
192 kref_get(&dreq
->kref
);
193 atomic_set(&dreq
->complete
, reads
);
198 * nfs_direct_read_result - handle a read reply for a direct read request
199 * @data: address of NFS READ operation control block
200 * @status: status of this NFS READ operation
202 * We must hold a reference to all the pages in this direct read request
203 * until the RPCs complete. This could be long *after* we are woken up in
204 * nfs_direct_read_wait (for instance, if someone hits ^C on a slow server).
206 static void nfs_direct_read_result(struct nfs_read_data
*data
, int status
)
208 struct nfs_direct_req
*dreq
= (struct nfs_direct_req
*) data
->req
;
210 if (likely(status
>= 0))
211 atomic_add(data
->res
.count
, &dreq
->count
);
213 atomic_set(&dreq
->error
, status
);
215 if (unlikely(atomic_dec_and_test(&dreq
->complete
))) {
216 nfs_free_user_pages(dreq
->pages
, dreq
->npages
, 1);
217 wake_up(&dreq
->wait
);
218 kref_put(&dreq
->kref
, nfs_direct_req_release
);
223 * nfs_direct_read_schedule - dispatch NFS READ operations for a direct read
224 * @dreq: address of nfs_direct_req struct for this request
225 * @inode: target inode
226 * @ctx: target file open context
227 * @user_addr: starting address of this segment of user's buffer
228 * @count: size of this segment
229 * @file_offset: offset in file to begin the operation
231 * For each nfs_read_data struct that was allocated on the list, dispatch
232 * an NFS READ operation
234 static void nfs_direct_read_schedule(struct nfs_direct_req
*dreq
,
235 struct inode
*inode
, struct nfs_open_context
*ctx
,
236 unsigned long user_addr
, size_t count
, loff_t file_offset
)
238 struct list_head
*list
= &dreq
->list
;
239 struct page
**pages
= dreq
->pages
;
240 unsigned int curpage
, pgbase
;
241 unsigned int rsize
= NFS_SERVER(inode
)->rsize
;
244 pgbase
= user_addr
& ~PAGE_MASK
;
246 struct nfs_read_data
*data
;
253 data
= list_entry(list
->next
, struct nfs_read_data
, pages
);
254 list_del_init(&data
->pages
);
257 data
->cred
= ctx
->cred
;
258 data
->args
.fh
= NFS_FH(inode
);
259 data
->args
.context
= ctx
;
260 data
->args
.offset
= file_offset
;
261 data
->args
.pgbase
= pgbase
;
262 data
->args
.pages
= &pages
[curpage
];
263 data
->args
.count
= bytes
;
264 data
->res
.fattr
= &data
->fattr
;
266 data
->res
.count
= bytes
;
268 NFS_PROTO(inode
)->read_setup(data
);
270 data
->task
.tk_cookie
= (unsigned long) inode
;
271 data
->complete
= nfs_direct_read_result
;
274 rpc_execute(&data
->task
);
277 dfprintk(VFS
, "NFS: %4d initiated direct read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
280 (long long)NFS_FILEID(inode
),
282 (unsigned long long)data
->args
.offset
);
284 file_offset
+= bytes
;
286 curpage
+= pgbase
>> PAGE_SHIFT
;
287 pgbase
&= ~PAGE_MASK
;
290 } while (count
!= 0);
294 * nfs_direct_read_wait - wait for I/O completion for direct reads
295 * @dreq: request on which we are to wait
296 * @intr: whether or not this wait can be interrupted
298 * Collects and returns the final error value/byte-count.
300 static ssize_t
nfs_direct_read_wait(struct nfs_direct_req
*dreq
, int intr
)
305 result
= wait_event_interruptible(dreq
->wait
,
306 (atomic_read(&dreq
->complete
) == 0));
308 wait_event(dreq
->wait
, (atomic_read(&dreq
->complete
) == 0));
312 result
= atomic_read(&dreq
->error
);
314 result
= atomic_read(&dreq
->count
);
316 kref_put(&dreq
->kref
, nfs_direct_req_release
);
317 return (ssize_t
) result
;
321 * nfs_direct_read_seg - Read in one iov segment. Generate separate
322 * read RPCs for each "rsize" bytes.
323 * @inode: target inode
324 * @ctx: target file open context
325 * @user_addr: starting address of this segment of user's buffer
326 * @count: size of this segment
327 * @file_offset: offset in file to begin the operation
328 * @pages: array of addresses of page structs defining user's buffer
329 * @nr_pages: number of pages in the array
332 static ssize_t
nfs_direct_read_seg(struct inode
*inode
,
333 struct nfs_open_context
*ctx
, unsigned long user_addr
,
334 size_t count
, loff_t file_offset
, struct page
**pages
,
335 unsigned int nr_pages
)
339 struct rpc_clnt
*clnt
= NFS_CLIENT(inode
);
340 struct nfs_direct_req
*dreq
;
342 dreq
= nfs_direct_read_alloc(count
, NFS_SERVER(inode
)->rsize
);
347 dreq
->npages
= nr_pages
;
349 rpc_clnt_sigmask(clnt
, &oldset
);
350 nfs_direct_read_schedule(dreq
, inode
, ctx
, user_addr
, count
,
352 result
= nfs_direct_read_wait(dreq
, clnt
->cl_intr
);
353 rpc_clnt_sigunmask(clnt
, &oldset
);
359 * nfs_direct_read - For each iov segment, map the user's buffer
360 * then generate read RPCs.
361 * @inode: target inode
362 * @ctx: target file open context
363 * @iov: array of vectors that define I/O buffer
364 * file_offset: offset in file to begin the operation
365 * nr_segs: size of iovec array
367 * We've already pushed out any non-direct writes so that this read
368 * will see them when we read from the server.
371 nfs_direct_read(struct inode
*inode
, struct nfs_open_context
*ctx
,
372 const struct iovec
*iov
, loff_t file_offset
,
373 unsigned long nr_segs
)
375 ssize_t tot_bytes
= 0;
376 unsigned long seg
= 0;
378 while ((seg
< nr_segs
) && (tot_bytes
>= 0)) {
382 const struct iovec
*vec
= &iov
[seg
++];
383 unsigned long user_addr
= (unsigned long) vec
->iov_base
;
384 size_t size
= vec
->iov_len
;
386 page_count
= nfs_get_user_pages(READ
, user_addr
, size
, &pages
);
387 if (page_count
< 0) {
388 nfs_free_user_pages(pages
, 0, 0);
394 result
= nfs_direct_read_seg(inode
, ctx
, user_addr
, size
,
395 file_offset
, pages
, page_count
);
403 file_offset
+= result
;
412 * nfs_direct_write_seg - Write out one iov segment. Generate separate
413 * write RPCs for each "wsize" bytes, then commit.
414 * @inode: target inode
415 * @ctx: target file open context
416 * user_addr: starting address of this segment of user's buffer
417 * count: size of this segment
418 * file_offset: offset in file to begin the operation
419 * @pages: array of addresses of page structs defining user's buffer
420 * nr_pages: size of pages array
422 static ssize_t
nfs_direct_write_seg(struct inode
*inode
,
423 struct nfs_open_context
*ctx
, unsigned long user_addr
,
424 size_t count
, loff_t file_offset
, struct page
**pages
,
427 const unsigned int wsize
= NFS_SERVER(inode
)->wsize
;
429 int curpage
, need_commit
;
430 ssize_t result
, tot_bytes
;
431 struct nfs_writeverf first_verf
;
432 struct nfs_write_data
*wdata
;
434 wdata
= nfs_writedata_alloc();
438 wdata
->inode
= inode
;
439 wdata
->cred
= ctx
->cred
;
440 wdata
->args
.fh
= NFS_FH(inode
);
441 wdata
->args
.context
= ctx
;
442 wdata
->args
.stable
= NFS_UNSTABLE
;
443 if (IS_SYNC(inode
) || NFS_PROTO(inode
)->version
== 2 || count
<= wsize
)
444 wdata
->args
.stable
= NFS_FILE_SYNC
;
445 wdata
->res
.fattr
= &wdata
->fattr
;
446 wdata
->res
.verf
= &wdata
->verf
;
448 nfs_begin_data_update(inode
);
454 wdata
->args
.pgbase
= user_addr
& ~PAGE_MASK
;
455 wdata
->args
.offset
= file_offset
;
457 wdata
->args
.count
= request
;
458 if (wdata
->args
.count
> wsize
)
459 wdata
->args
.count
= wsize
;
460 wdata
->args
.pages
= &pages
[curpage
];
462 dprintk("NFS: direct write: c=%u o=%Ld ua=%lu, pb=%u, cp=%u\n",
463 wdata
->args
.count
, (long long) wdata
->args
.offset
,
464 user_addr
+ tot_bytes
, wdata
->args
.pgbase
, curpage
);
467 result
= NFS_PROTO(inode
)->write(wdata
);
477 memcpy(&first_verf
.verifier
, &wdata
->verf
.verifier
,
478 sizeof(first_verf
.verifier
));
479 if (wdata
->verf
.committed
!= NFS_FILE_SYNC
) {
481 if (memcmp(&first_verf
.verifier
, &wdata
->verf
.verifier
,
482 sizeof(first_verf
.verifier
)));
488 /* in case of a short write: stop now, let the app recover */
489 if (result
< wdata
->args
.count
)
492 wdata
->args
.offset
+= result
;
493 wdata
->args
.pgbase
+= result
;
494 curpage
+= wdata
->args
.pgbase
>> PAGE_SHIFT
;
495 wdata
->args
.pgbase
&= ~PAGE_MASK
;
497 } while (request
!= 0);
500 * Commit data written so far, even in the event of an error
503 wdata
->args
.count
= tot_bytes
;
504 wdata
->args
.offset
= file_offset
;
507 result
= NFS_PROTO(inode
)->commit(wdata
);
510 if (result
< 0 || memcmp(&first_verf
.verifier
,
511 &wdata
->verf
.verifier
,
512 sizeof(first_verf
.verifier
)) != 0)
518 nfs_end_data_update(inode
);
519 nfs_writedata_free(wdata
);
523 wdata
->args
.stable
= NFS_FILE_SYNC
;
528 * nfs_direct_write - For each iov segment, map the user's buffer
529 * then generate write and commit RPCs.
530 * @inode: target inode
531 * @ctx: target file open context
532 * @iov: array of vectors that define I/O buffer
533 * file_offset: offset in file to begin the operation
534 * nr_segs: size of iovec array
536 * Upon return, generic_file_direct_IO invalidates any cached pages
537 * that non-direct readers might access, so they will pick up these
538 * writes immediately.
540 static ssize_t
nfs_direct_write(struct inode
*inode
,
541 struct nfs_open_context
*ctx
, const struct iovec
*iov
,
542 loff_t file_offset
, unsigned long nr_segs
)
544 ssize_t tot_bytes
= 0;
545 unsigned long seg
= 0;
547 while ((seg
< nr_segs
) && (tot_bytes
>= 0)) {
551 const struct iovec
*vec
= &iov
[seg
++];
552 unsigned long user_addr
= (unsigned long) vec
->iov_base
;
553 size_t size
= vec
->iov_len
;
555 page_count
= nfs_get_user_pages(WRITE
, user_addr
, size
, &pages
);
556 if (page_count
< 0) {
557 nfs_free_user_pages(pages
, 0, 0);
563 result
= nfs_direct_write_seg(inode
, ctx
, user_addr
, size
,
564 file_offset
, pages
, page_count
);
565 nfs_free_user_pages(pages
, page_count
, 0);
573 file_offset
+= result
;
581 * nfs_direct_IO - NFS address space operation for direct I/O
582 * rw: direction (read or write)
583 * @iocb: target I/O control block
584 * @iov: array of vectors that define I/O buffer
585 * file_offset: offset in file to begin the operation
586 * nr_segs: size of iovec array
590 nfs_direct_IO(int rw
, struct kiocb
*iocb
, const struct iovec
*iov
,
591 loff_t file_offset
, unsigned long nr_segs
)
593 ssize_t result
= -EINVAL
;
594 struct file
*file
= iocb
->ki_filp
;
595 struct nfs_open_context
*ctx
;
596 struct dentry
*dentry
= file
->f_dentry
;
597 struct inode
*inode
= dentry
->d_inode
;
600 * No support for async yet
602 if (!is_sync_kiocb(iocb
))
605 ctx
= (struct nfs_open_context
*)file
->private_data
;
608 dprintk("NFS: direct_IO(read) (%s) off/no(%Lu/%lu)\n",
609 dentry
->d_name
.name
, file_offset
, nr_segs
);
611 result
= nfs_direct_read(inode
, ctx
, iov
,
612 file_offset
, nr_segs
);
615 dprintk("NFS: direct_IO(write) (%s) off/no(%Lu/%lu)\n",
616 dentry
->d_name
.name
, file_offset
, nr_segs
);
618 result
= nfs_direct_write(inode
, ctx
, iov
,
619 file_offset
, nr_segs
);
628 * nfs_file_direct_read - file direct read operation for NFS files
629 * @iocb: target I/O control block
630 * @buf: user's buffer into which to read data
631 * count: number of bytes to read
632 * pos: byte offset in file where reading starts
634 * We use this function for direct reads instead of calling
635 * generic_file_aio_read() in order to avoid gfar's check to see if
636 * the request starts before the end of the file. For that check
637 * to work, we must generate a GETATTR before each direct read, and
638 * even then there is a window between the GETATTR and the subsequent
639 * READ where the file size could change. So our preference is simply
640 * to do all reads the application wants, and the server will take
641 * care of managing the end of file boundary.
643 * This function also eliminates unnecessarily updating the file's
644 * atime locally, as the NFS server sets the file's atime, and this
645 * client must read the updated atime from the server back into its
649 nfs_file_direct_read(struct kiocb
*iocb
, char __user
*buf
, size_t count
, loff_t pos
)
651 ssize_t retval
= -EINVAL
;
652 loff_t
*ppos
= &iocb
->ki_pos
;
653 struct file
*file
= iocb
->ki_filp
;
654 struct nfs_open_context
*ctx
=
655 (struct nfs_open_context
*) file
->private_data
;
656 struct address_space
*mapping
= file
->f_mapping
;
657 struct inode
*inode
= mapping
->host
;
663 dprintk("nfs: direct read(%s/%s, %lu@%lu)\n",
664 file
->f_dentry
->d_parent
->d_name
.name
,
665 file
->f_dentry
->d_name
.name
,
666 (unsigned long) count
, (unsigned long) pos
);
668 if (!is_sync_kiocb(iocb
))
673 if (!access_ok(VERIFY_WRITE
, iov
.iov_base
, iov
.iov_len
))
679 retval
= nfs_sync_mapping(mapping
);
683 retval
= nfs_direct_read(inode
, ctx
, &iov
, pos
, 1);
685 *ppos
= pos
+ retval
;
692 * nfs_file_direct_write - file direct write operation for NFS files
693 * @iocb: target I/O control block
694 * @buf: user's buffer from which to write data
695 * count: number of bytes to write
696 * pos: byte offset in file where writing starts
698 * We use this function for direct writes instead of calling
699 * generic_file_aio_write() in order to avoid taking the inode
700 * semaphore and updating the i_size. The NFS server will set
701 * the new i_size and this client must read the updated size
702 * back into its cache. We let the server do generic write
703 * parameter checking and report problems.
705 * We also avoid an unnecessary invocation of generic_osync_inode(),
706 * as it is fairly meaningless to sync the metadata of an NFS file.
708 * We eliminate local atime updates, see direct read above.
710 * We avoid unnecessary page cache invalidations for normal cached
711 * readers of this file.
713 * Note that O_APPEND is not supported for NFS direct writes, as there
714 * is no atomic O_APPEND write facility in the NFS protocol.
717 nfs_file_direct_write(struct kiocb
*iocb
, const char __user
*buf
, size_t count
, loff_t pos
)
719 ssize_t retval
= -EINVAL
;
720 loff_t
*ppos
= &iocb
->ki_pos
;
721 unsigned long limit
= current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
;
722 struct file
*file
= iocb
->ki_filp
;
723 struct nfs_open_context
*ctx
=
724 (struct nfs_open_context
*) file
->private_data
;
725 struct address_space
*mapping
= file
->f_mapping
;
726 struct inode
*inode
= mapping
->host
;
728 .iov_base
= (char __user
*)buf
,
732 dfprintk(VFS
, "nfs: direct write(%s/%s(%ld), %lu@%lu)\n",
733 file
->f_dentry
->d_parent
->d_name
.name
,
734 file
->f_dentry
->d_name
.name
, inode
->i_ino
,
735 (unsigned long) count
, (unsigned long) pos
);
737 if (!is_sync_kiocb(iocb
))
744 if (!access_ok(VERIFY_READ
, iov
.iov_base
, iov
.iov_len
))
747 if (limit
!= RLIM_INFINITY
) {
749 send_sig(SIGXFSZ
, current
, 0);
752 if (count
> limit
- (unsigned long) pos
)
753 count
= limit
- (unsigned long) pos
;
759 retval
= nfs_sync_mapping(mapping
);
763 retval
= nfs_direct_write(inode
, ctx
, &iov
, pos
, 1);
764 if (mapping
->nrpages
)
765 invalidate_inode_pages2(mapping
);
767 *ppos
= pos
+ retval
;
773 int nfs_init_directcache(void)
775 nfs_direct_cachep
= kmem_cache_create("nfs_direct_cache",
776 sizeof(struct nfs_direct_req
),
777 0, SLAB_RECLAIM_ACCOUNT
,
779 if (nfs_direct_cachep
== NULL
)
785 void nfs_destroy_directcache(void)
787 if (kmem_cache_destroy(nfs_direct_cachep
))
788 printk(KERN_INFO
"nfs_direct_cache: not all structures were freed\n");