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>
59 #define NFSDBG_FACILITY NFSDBG_VFS
60 #define MAX_DIRECTIO_SIZE (4096UL << PAGE_SHIFT)
62 static void nfs_free_user_pages(struct page
**pages
, int npages
, int do_dirty
);
63 static kmem_cache_t
*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 */
70 struct list_head list
; /* nfs_read_data structs */
71 struct file
* filp
; /* file descriptor */
72 struct kiocb
* iocb
; /* controlling i/o request */
73 wait_queue_head_t wait
; /* wait for i/o completion */
74 struct inode
* inode
; /* target file of I/O */
75 struct page
** pages
; /* pages in our buffer */
76 unsigned int npages
; /* count of pages */
77 atomic_t complete
, /* i/os we're waiting for */
78 count
, /* bytes actually processed */
79 error
; /* any reported error */
84 * nfs_direct_IO - NFS address space operation for direct I/O
85 * @rw: direction (read or write)
86 * @iocb: target I/O control block
87 * @iov: array of vectors that define I/O buffer
88 * @pos: offset in file to begin the operation
89 * @nr_segs: size of iovec array
91 * The presence of this routine in the address space ops vector means
92 * the NFS client supports direct I/O. However, we shunt off direct
93 * read and write requests before the VFS gets them, so this method
94 * should never be called.
96 ssize_t
nfs_direct_IO(int rw
, struct kiocb
*iocb
, const struct iovec
*iov
, loff_t pos
, unsigned long nr_segs
)
98 struct dentry
*dentry
= iocb
->ki_filp
->f_dentry
;
100 dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
101 dentry
->d_name
.name
, (long long) pos
, nr_segs
);
106 static inline int nfs_get_user_pages(int rw
, unsigned long user_addr
, size_t size
, struct page
***pages
)
108 int result
= -ENOMEM
;
109 unsigned long page_count
;
112 /* set an arbitrary limit to prevent type overflow */
113 /* XXX: this can probably be as large as INT_MAX */
114 if (size
> MAX_DIRECTIO_SIZE
) {
119 page_count
= (user_addr
+ size
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
120 page_count
-= user_addr
>> PAGE_SHIFT
;
122 array_size
= (page_count
* sizeof(struct page
*));
123 *pages
= kmalloc(array_size
, GFP_KERNEL
);
125 down_read(¤t
->mm
->mmap_sem
);
126 result
= get_user_pages(current
, current
->mm
, user_addr
,
127 page_count
, (rw
== READ
), 0,
129 up_read(¤t
->mm
->mmap_sem
);
131 * If we got fewer pages than expected from get_user_pages(),
132 * the user buffer runs off the end of a mapping; return EFAULT.
134 if (result
>= 0 && result
< page_count
) {
135 nfs_free_user_pages(*pages
, result
, 0);
143 static void nfs_free_user_pages(struct page
**pages
, int npages
, int do_dirty
)
146 for (i
= 0; i
< npages
; i
++) {
147 struct page
*page
= pages
[i
];
148 if (do_dirty
&& !PageCompound(page
))
149 set_page_dirty_lock(page
);
150 page_cache_release(page
);
155 static inline struct nfs_direct_req
*nfs_direct_req_alloc(void)
157 struct nfs_direct_req
*dreq
;
159 dreq
= kmem_cache_alloc(nfs_direct_cachep
, SLAB_KERNEL
);
163 kref_init(&dreq
->kref
);
164 init_waitqueue_head(&dreq
->wait
);
165 INIT_LIST_HEAD(&dreq
->list
);
167 atomic_set(&dreq
->count
, 0);
168 atomic_set(&dreq
->error
, 0);
173 static void nfs_direct_req_release(struct kref
*kref
)
175 struct nfs_direct_req
*dreq
= container_of(kref
, struct nfs_direct_req
, kref
);
176 kmem_cache_free(nfs_direct_cachep
, dreq
);
180 * Collects and returns the final error value/byte-count.
182 static ssize_t
nfs_direct_wait(struct nfs_direct_req
*dreq
)
184 int result
= -EIOCBQUEUED
;
186 /* Async requests don't wait here */
190 result
= wait_event_interruptible(dreq
->wait
,
191 (atomic_read(&dreq
->complete
) == 0));
194 result
= atomic_read(&dreq
->error
);
196 result
= atomic_read(&dreq
->count
);
199 kref_put(&dreq
->kref
, nfs_direct_req_release
);
200 return (ssize_t
) result
;
204 * We must hold a reference to all the pages in this direct read request
205 * until the RPCs complete. This could be long *after* we are woken up in
206 * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
208 * In addition, synchronous I/O uses a stack-allocated iocb. Thus we
209 * can't trust the iocb is still valid here if this is a synchronous
210 * request. If the waiter is woken prematurely, the iocb is long gone.
212 static void nfs_direct_complete(struct nfs_direct_req
*dreq
)
214 nfs_free_user_pages(dreq
->pages
, dreq
->npages
, 1);
217 long res
= atomic_read(&dreq
->error
);
219 res
= atomic_read(&dreq
->count
);
220 aio_complete(dreq
->iocb
, res
, 0);
222 wake_up(&dreq
->wait
);
224 kref_put(&dreq
->kref
, nfs_direct_req_release
);
228 * Note we also set the number of requests we have in the dreq when we are
229 * done. This prevents races with I/O completion so we will always wait
230 * until all requests have been dispatched and completed.
232 static struct nfs_direct_req
*nfs_direct_read_alloc(size_t nbytes
, size_t rsize
)
234 struct list_head
*list
;
235 struct nfs_direct_req
*dreq
;
236 unsigned int reads
= 0;
237 unsigned int rpages
= (rsize
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
239 dreq
= nfs_direct_req_alloc();
245 struct nfs_read_data
*data
= nfs_readdata_alloc(rpages
);
247 if (unlikely(!data
)) {
248 while (!list_empty(list
)) {
249 data
= list_entry(list
->next
,
250 struct nfs_read_data
, pages
);
251 list_del(&data
->pages
);
252 nfs_readdata_free(data
);
254 kref_put(&dreq
->kref
, nfs_direct_req_release
);
258 INIT_LIST_HEAD(&data
->pages
);
259 list_add(&data
->pages
, list
);
261 data
->req
= (struct nfs_page
*) dreq
;
267 kref_get(&dreq
->kref
);
268 atomic_set(&dreq
->complete
, reads
);
272 static void nfs_direct_read_result(struct rpc_task
*task
, void *calldata
)
274 struct nfs_read_data
*data
= calldata
;
275 struct nfs_direct_req
*dreq
= (struct nfs_direct_req
*) data
->req
;
277 if (nfs_readpage_result(task
, data
) != 0)
279 if (likely(task
->tk_status
>= 0))
280 atomic_add(data
->res
.count
, &dreq
->count
);
282 atomic_set(&dreq
->error
, task
->tk_status
);
284 if (unlikely(atomic_dec_and_test(&dreq
->complete
)))
285 nfs_direct_complete(dreq
);
288 static const struct rpc_call_ops nfs_read_direct_ops
= {
289 .rpc_call_done
= nfs_direct_read_result
,
290 .rpc_release
= nfs_readdata_release
,
294 * For each nfs_read_data struct that was allocated on the list, dispatch
295 * an NFS READ operation
297 static void nfs_direct_read_schedule(struct nfs_direct_req
*dreq
, unsigned long user_addr
, size_t count
, loff_t file_offset
)
299 struct file
*file
= dreq
->filp
;
300 struct inode
*inode
= file
->f_mapping
->host
;
301 struct nfs_open_context
*ctx
= (struct nfs_open_context
*)
303 struct list_head
*list
= &dreq
->list
;
304 struct page
**pages
= dreq
->pages
;
305 size_t rsize
= NFS_SERVER(inode
)->rsize
;
306 unsigned int curpage
, pgbase
;
309 pgbase
= user_addr
& ~PAGE_MASK
;
311 struct nfs_read_data
*data
;
318 data
= list_entry(list
->next
, struct nfs_read_data
, pages
);
319 list_del_init(&data
->pages
);
322 data
->cred
= ctx
->cred
;
323 data
->args
.fh
= NFS_FH(inode
);
324 data
->args
.context
= ctx
;
325 data
->args
.offset
= file_offset
;
326 data
->args
.pgbase
= pgbase
;
327 data
->args
.pages
= &pages
[curpage
];
328 data
->args
.count
= bytes
;
329 data
->res
.fattr
= &data
->fattr
;
331 data
->res
.count
= bytes
;
333 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), RPC_TASK_ASYNC
,
334 &nfs_read_direct_ops
, data
);
335 NFS_PROTO(inode
)->read_setup(data
);
337 data
->task
.tk_cookie
= (unsigned long) inode
;
340 rpc_execute(&data
->task
);
343 dfprintk(VFS
, "NFS: %4d initiated direct read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
346 (long long)NFS_FILEID(inode
),
348 (unsigned long long)data
->args
.offset
);
350 file_offset
+= bytes
;
352 curpage
+= pgbase
>> PAGE_SHIFT
;
353 pgbase
&= ~PAGE_MASK
;
356 } while (count
!= 0);
359 static ssize_t
nfs_direct_read(struct kiocb
*iocb
, unsigned long user_addr
, size_t count
, loff_t file_offset
, struct page
**pages
, unsigned int nr_pages
)
363 struct inode
*inode
= iocb
->ki_filp
->f_mapping
->host
;
364 struct rpc_clnt
*clnt
= NFS_CLIENT(inode
);
365 struct nfs_direct_req
*dreq
;
367 dreq
= nfs_direct_read_alloc(count
, NFS_SERVER(inode
)->rsize
);
372 dreq
->npages
= nr_pages
;
374 dreq
->filp
= iocb
->ki_filp
;
375 if (!is_sync_kiocb(iocb
))
378 nfs_add_stats(inode
, NFSIOS_DIRECTREADBYTES
, count
);
379 rpc_clnt_sigmask(clnt
, &oldset
);
380 nfs_direct_read_schedule(dreq
, user_addr
, count
, file_offset
);
381 result
= nfs_direct_wait(dreq
);
382 rpc_clnt_sigunmask(clnt
, &oldset
);
387 static struct nfs_direct_req
*nfs_direct_write_alloc(size_t nbytes
, size_t wsize
)
389 struct list_head
*list
;
390 struct nfs_direct_req
*dreq
;
391 unsigned int writes
= 0;
392 unsigned int wpages
= (wsize
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
394 dreq
= nfs_direct_req_alloc();
400 struct nfs_write_data
*data
= nfs_writedata_alloc(wpages
);
402 if (unlikely(!data
)) {
403 while (!list_empty(list
)) {
404 data
= list_entry(list
->next
,
405 struct nfs_write_data
, pages
);
406 list_del(&data
->pages
);
407 nfs_writedata_free(data
);
409 kref_put(&dreq
->kref
, nfs_direct_req_release
);
413 INIT_LIST_HEAD(&data
->pages
);
414 list_add(&data
->pages
, list
);
416 data
->req
= (struct nfs_page
*) dreq
;
422 kref_get(&dreq
->kref
);
423 atomic_set(&dreq
->complete
, writes
);
427 static void nfs_direct_write_result(struct rpc_task
*task
, void *calldata
)
429 struct nfs_write_data
*data
= calldata
;
430 struct nfs_direct_req
*dreq
= (struct nfs_direct_req
*) data
->req
;
431 int status
= task
->tk_status
;
433 if (nfs_writeback_done(task
, data
) != 0)
435 /* If the server fell back to an UNSTABLE write, it's an error. */
436 if (unlikely(data
->res
.verf
->committed
!= NFS_FILE_SYNC
))
439 if (likely(status
>= 0))
440 atomic_add(data
->res
.count
, &dreq
->count
);
442 atomic_set(&dreq
->error
, status
);
444 if (unlikely(atomic_dec_and_test(&dreq
->complete
))) {
445 nfs_end_data_update(data
->inode
);
446 nfs_direct_complete(dreq
);
450 static const struct rpc_call_ops nfs_write_direct_ops
= {
451 .rpc_call_done
= nfs_direct_write_result
,
452 .rpc_release
= nfs_writedata_release
,
456 * For each nfs_write_data struct that was allocated on the list, dispatch
457 * an NFS WRITE operation
459 * XXX: For now, support only FILE_SYNC writes. Later we may add
460 * support for UNSTABLE + COMMIT.
462 static void nfs_direct_write_schedule(struct nfs_direct_req
*dreq
, unsigned long user_addr
, size_t count
, loff_t file_offset
)
464 struct file
*file
= dreq
->filp
;
465 struct inode
*inode
= file
->f_mapping
->host
;
466 struct nfs_open_context
*ctx
= (struct nfs_open_context
*)
468 struct list_head
*list
= &dreq
->list
;
469 struct page
**pages
= dreq
->pages
;
470 size_t wsize
= NFS_SERVER(inode
)->wsize
;
471 unsigned int curpage
, pgbase
;
474 pgbase
= user_addr
& ~PAGE_MASK
;
476 struct nfs_write_data
*data
;
483 data
= list_entry(list
->next
, struct nfs_write_data
, pages
);
484 list_del_init(&data
->pages
);
487 data
->cred
= ctx
->cred
;
488 data
->args
.fh
= NFS_FH(inode
);
489 data
->args
.context
= ctx
;
490 data
->args
.offset
= file_offset
;
491 data
->args
.pgbase
= pgbase
;
492 data
->args
.pages
= &pages
[curpage
];
493 data
->args
.count
= bytes
;
494 data
->res
.fattr
= &data
->fattr
;
495 data
->res
.count
= bytes
;
496 data
->res
.verf
= &data
->verf
;
498 rpc_init_task(&data
->task
, NFS_CLIENT(inode
), RPC_TASK_ASYNC
,
499 &nfs_write_direct_ops
, data
);
500 NFS_PROTO(inode
)->write_setup(data
, FLUSH_STABLE
);
502 data
->task
.tk_priority
= RPC_PRIORITY_NORMAL
;
503 data
->task
.tk_cookie
= (unsigned long) inode
;
506 rpc_execute(&data
->task
);
509 dfprintk(VFS
, "NFS: %4d initiated direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
512 (long long)NFS_FILEID(inode
),
514 (unsigned long long)data
->args
.offset
);
516 file_offset
+= bytes
;
518 curpage
+= pgbase
>> PAGE_SHIFT
;
519 pgbase
&= ~PAGE_MASK
;
522 } while (count
!= 0);
525 static ssize_t
nfs_direct_write(struct kiocb
*iocb
, unsigned long user_addr
, size_t count
, loff_t file_offset
, struct page
**pages
, int nr_pages
)
529 struct inode
*inode
= iocb
->ki_filp
->f_mapping
->host
;
530 struct rpc_clnt
*clnt
= NFS_CLIENT(inode
);
531 struct nfs_direct_req
*dreq
;
533 dreq
= nfs_direct_write_alloc(count
, NFS_SERVER(inode
)->wsize
);
538 dreq
->npages
= nr_pages
;
540 dreq
->filp
= iocb
->ki_filp
;
541 if (!is_sync_kiocb(iocb
))
544 nfs_add_stats(inode
, NFSIOS_DIRECTWRITTENBYTES
, count
);
546 nfs_begin_data_update(inode
);
548 rpc_clnt_sigmask(clnt
, &oldset
);
549 nfs_direct_write_schedule(dreq
, user_addr
, count
, file_offset
);
550 result
= nfs_direct_wait(dreq
);
551 rpc_clnt_sigunmask(clnt
, &oldset
);
557 * nfs_file_direct_read - file direct read operation for NFS files
558 * @iocb: target I/O control block
559 * @buf: user's buffer into which to read data
560 * count: number of bytes to read
561 * pos: byte offset in file where reading starts
563 * We use this function for direct reads instead of calling
564 * generic_file_aio_read() in order to avoid gfar's check to see if
565 * the request starts before the end of the file. For that check
566 * to work, we must generate a GETATTR before each direct read, and
567 * even then there is a window between the GETATTR and the subsequent
568 * READ where the file size could change. So our preference is simply
569 * to do all reads the application wants, and the server will take
570 * care of managing the end of file boundary.
572 * This function also eliminates unnecessarily updating the file's
573 * atime locally, as the NFS server sets the file's atime, and this
574 * client must read the updated atime from the server back into its
577 ssize_t
nfs_file_direct_read(struct kiocb
*iocb
, char __user
*buf
, size_t count
, loff_t pos
)
579 ssize_t retval
= -EINVAL
;
582 struct file
*file
= iocb
->ki_filp
;
583 struct address_space
*mapping
= file
->f_mapping
;
585 dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
586 file
->f_dentry
->d_parent
->d_name
.name
,
587 file
->f_dentry
->d_name
.name
,
588 (unsigned long) count
, (long long) pos
);
593 if (!access_ok(VERIFY_WRITE
, buf
, count
))
599 retval
= nfs_sync_mapping(mapping
);
603 page_count
= nfs_get_user_pages(READ
, (unsigned long) buf
,
605 if (page_count
< 0) {
606 nfs_free_user_pages(pages
, 0, 0);
611 retval
= nfs_direct_read(iocb
, (unsigned long) buf
, count
, pos
,
614 iocb
->ki_pos
= pos
+ retval
;
621 * nfs_file_direct_write - file direct write operation for NFS files
622 * @iocb: target I/O control block
623 * @buf: user's buffer from which to write data
624 * count: number of bytes to write
625 * pos: byte offset in file where writing starts
627 * We use this function for direct writes instead of calling
628 * generic_file_aio_write() in order to avoid taking the inode
629 * semaphore and updating the i_size. The NFS server will set
630 * the new i_size and this client must read the updated size
631 * back into its cache. We let the server do generic write
632 * parameter checking and report problems.
634 * We also avoid an unnecessary invocation of generic_osync_inode(),
635 * as it is fairly meaningless to sync the metadata of an NFS file.
637 * We eliminate local atime updates, see direct read above.
639 * We avoid unnecessary page cache invalidations for normal cached
640 * readers of this file.
642 * Note that O_APPEND is not supported for NFS direct writes, as there
643 * is no atomic O_APPEND write facility in the NFS protocol.
645 ssize_t
nfs_file_direct_write(struct kiocb
*iocb
, const char __user
*buf
, size_t count
, loff_t pos
)
650 struct file
*file
= iocb
->ki_filp
;
651 struct address_space
*mapping
= file
->f_mapping
;
653 dfprintk(VFS
, "nfs: direct write(%s/%s, %lu@%Ld)\n",
654 file
->f_dentry
->d_parent
->d_name
.name
,
655 file
->f_dentry
->d_name
.name
,
656 (unsigned long) count
, (long long) pos
);
658 retval
= generic_write_checks(file
, &pos
, &count
, 0);
663 if ((ssize_t
) count
< 0)
670 if (!access_ok(VERIFY_READ
, buf
, count
))
673 retval
= nfs_sync_mapping(mapping
);
677 page_count
= nfs_get_user_pages(WRITE
, (unsigned long) buf
,
679 if (page_count
< 0) {
680 nfs_free_user_pages(pages
, 0, 0);
685 retval
= nfs_direct_write(iocb
, (unsigned long) buf
, count
,
686 pos
, pages
, page_count
);
689 * XXX: nfs_end_data_update() already ensures this file's
690 * cached data is subsequently invalidated. Do we really
691 * need to call invalidate_inode_pages2() again here?
693 * For aio writes, this invalidation will almost certainly
694 * occur before the writes complete. Kind of racey.
696 if (mapping
->nrpages
)
697 invalidate_inode_pages2(mapping
);
700 iocb
->ki_pos
= pos
+ retval
;
706 int nfs_init_directcache(void)
708 nfs_direct_cachep
= kmem_cache_create("nfs_direct_cache",
709 sizeof(struct nfs_direct_req
),
710 0, SLAB_RECLAIM_ACCOUNT
,
712 if (nfs_direct_cachep
== NULL
)
718 void nfs_destroy_directcache(void)
720 if (kmem_cache_destroy(nfs_direct_cachep
))
721 printk(KERN_INFO
"nfs_direct_cache: not all structures were freed\n");