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>
49 #include <linux/module.h>
51 #include <linux/nfs_fs.h>
52 #include <linux/nfs_page.h>
53 #include <linux/sunrpc/clnt.h>
55 #include <asm/uaccess.h>
56 #include <linux/atomic.h>
62 #define NFSDBG_FACILITY NFSDBG_VFS
64 static struct kmem_cache
*nfs_direct_cachep
;
67 * This represents a set of asynchronous requests that we're waiting on
69 struct nfs_direct_mirror
{
73 struct nfs_direct_req
{
74 struct kref kref
; /* release manager */
77 struct nfs_open_context
*ctx
; /* file open context info */
78 struct nfs_lock_context
*l_ctx
; /* Lock context info */
79 struct kiocb
* iocb
; /* controlling i/o request */
80 struct inode
* inode
; /* target file of i/o */
82 /* completion state */
83 atomic_t io_count
; /* i/os we're waiting for */
84 spinlock_t lock
; /* protect completion state */
86 struct nfs_direct_mirror mirrors
[NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX
];
89 ssize_t count
, /* bytes actually processed */
90 bytes_left
, /* bytes left to be sent */
91 io_start
, /* start of IO */
92 error
; /* any reported error */
93 struct completion completion
; /* wait for i/o completion */
96 struct nfs_mds_commit_info mds_cinfo
; /* Storage for cinfo */
97 struct pnfs_ds_commit_info ds_cinfo
; /* Storage for cinfo */
98 struct work_struct work
;
100 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
101 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
102 struct nfs_writeverf verf
; /* unstable write verifier */
105 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops
;
106 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops
;
107 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
, struct inode
*inode
);
108 static void nfs_direct_write_schedule_work(struct work_struct
*work
);
110 static inline void get_dreq(struct nfs_direct_req
*dreq
)
112 atomic_inc(&dreq
->io_count
);
115 static inline int put_dreq(struct nfs_direct_req
*dreq
)
117 return atomic_dec_and_test(&dreq
->io_count
);
120 void nfs_direct_set_resched_writes(struct nfs_direct_req
*dreq
)
122 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
124 EXPORT_SYMBOL_GPL(nfs_direct_set_resched_writes
);
127 nfs_direct_good_bytes(struct nfs_direct_req
*dreq
, struct nfs_pgio_header
*hdr
)
132 WARN_ON_ONCE(hdr
->pgio_mirror_idx
>= dreq
->mirror_count
);
134 count
= dreq
->mirrors
[hdr
->pgio_mirror_idx
].count
;
135 if (count
+ dreq
->io_start
< hdr
->io_start
+ hdr
->good_bytes
) {
136 count
= hdr
->io_start
+ hdr
->good_bytes
- dreq
->io_start
;
137 dreq
->mirrors
[hdr
->pgio_mirror_idx
].count
= count
;
140 /* update the dreq->count by finding the minimum agreed count from all
142 count
= dreq
->mirrors
[0].count
;
144 for (i
= 1; i
< dreq
->mirror_count
; i
++)
145 count
= min(count
, dreq
->mirrors
[i
].count
);
151 * nfs_direct_select_verf - select the right verifier
152 * @dreq - direct request possibly spanning multiple servers
153 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
154 * @commit_idx - commit bucket index for the DS
156 * returns the correct verifier to use given the role of the server
158 static struct nfs_writeverf
*
159 nfs_direct_select_verf(struct nfs_direct_req
*dreq
,
160 struct nfs_client
*ds_clp
,
163 struct nfs_writeverf
*verfp
= &dreq
->verf
;
165 #ifdef CONFIG_NFS_V4_1
167 /* pNFS is in use, use the DS verf */
168 if (commit_idx
>= 0 && commit_idx
< dreq
->ds_cinfo
.nbuckets
)
169 verfp
= &dreq
->ds_cinfo
.buckets
[commit_idx
].direct_verf
;
179 * nfs_direct_set_hdr_verf - set the write/commit verifier
180 * @dreq - direct request possibly spanning multiple servers
181 * @hdr - pageio header to validate against previously seen verfs
183 * Set the server's (MDS or DS) "seen" verifier
185 static void nfs_direct_set_hdr_verf(struct nfs_direct_req
*dreq
,
186 struct nfs_pgio_header
*hdr
)
188 struct nfs_writeverf
*verfp
;
190 verfp
= nfs_direct_select_verf(dreq
, hdr
->ds_clp
, hdr
->ds_commit_idx
);
191 WARN_ON_ONCE(verfp
->committed
>= 0);
192 memcpy(verfp
, &hdr
->verf
, sizeof(struct nfs_writeverf
));
193 WARN_ON_ONCE(verfp
->committed
< 0);
197 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
198 * @dreq - direct request possibly spanning multiple servers
199 * @hdr - pageio header to validate against previously seen verf
201 * set the server's "seen" verf if not initialized.
202 * returns result of comparison between @hdr->verf and the "seen"
203 * verf of the server used by @hdr (DS or MDS)
205 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req
*dreq
,
206 struct nfs_pgio_header
*hdr
)
208 struct nfs_writeverf
*verfp
;
210 verfp
= nfs_direct_select_verf(dreq
, hdr
->ds_clp
, hdr
->ds_commit_idx
);
211 if (verfp
->committed
< 0) {
212 nfs_direct_set_hdr_verf(dreq
, hdr
);
215 return memcmp(verfp
, &hdr
->verf
, sizeof(struct nfs_writeverf
));
219 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
220 * @dreq - direct request possibly spanning multiple servers
221 * @data - commit data to validate against previously seen verf
223 * returns result of comparison between @data->verf and the verf of
224 * the server used by @data (DS or MDS)
226 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req
*dreq
,
227 struct nfs_commit_data
*data
)
229 struct nfs_writeverf
*verfp
;
231 verfp
= nfs_direct_select_verf(dreq
, data
->ds_clp
,
232 data
->ds_commit_index
);
234 /* verifier not set so always fail */
235 if (verfp
->committed
< 0)
238 return memcmp(verfp
, &data
->verf
, sizeof(struct nfs_writeverf
));
242 * nfs_direct_IO - NFS address space operation for direct I/O
243 * @iocb: target I/O control block
244 * @iov: array of vectors that define I/O buffer
245 * @pos: offset in file to begin the operation
246 * @nr_segs: size of iovec array
248 * The presence of this routine in the address space ops vector means
249 * the NFS client supports direct I/O. However, for most direct IO, we
250 * shunt off direct read and write requests before the VFS gets them,
251 * so this method is only ever called for swap.
253 ssize_t
nfs_direct_IO(struct kiocb
*iocb
, struct iov_iter
*iter
, loff_t pos
)
255 struct inode
*inode
= iocb
->ki_filp
->f_mapping
->host
;
257 /* we only support swap file calling nfs_direct_IO */
258 if (!IS_SWAPFILE(inode
))
261 #ifndef CONFIG_NFS_SWAP
262 dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n",
263 iocb
->ki_filp
, (long long) pos
, iter
->nr_segs
);
267 VM_BUG_ON(iov_iter_count(iter
) != PAGE_SIZE
);
269 if (iov_iter_rw(iter
) == READ
)
270 return nfs_file_direct_read(iocb
, iter
, pos
);
271 return nfs_file_direct_write(iocb
, iter
);
272 #endif /* CONFIG_NFS_SWAP */
275 static void nfs_direct_release_pages(struct page
**pages
, unsigned int npages
)
278 for (i
= 0; i
< npages
; i
++)
279 page_cache_release(pages
[i
]);
282 void nfs_init_cinfo_from_dreq(struct nfs_commit_info
*cinfo
,
283 struct nfs_direct_req
*dreq
)
285 cinfo
->lock
= &dreq
->inode
->i_lock
;
286 cinfo
->mds
= &dreq
->mds_cinfo
;
287 cinfo
->ds
= &dreq
->ds_cinfo
;
289 cinfo
->completion_ops
= &nfs_direct_commit_completion_ops
;
292 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req
*dreq
,
293 struct nfs_pageio_descriptor
*pgio
,
294 struct nfs_page
*req
)
296 int mirror_count
= 1;
298 if (pgio
->pg_ops
->pg_get_mirror_count
)
299 mirror_count
= pgio
->pg_ops
->pg_get_mirror_count(pgio
, req
);
301 dreq
->mirror_count
= mirror_count
;
304 static inline struct nfs_direct_req
*nfs_direct_req_alloc(void)
306 struct nfs_direct_req
*dreq
;
308 dreq
= kmem_cache_zalloc(nfs_direct_cachep
, GFP_KERNEL
);
312 kref_init(&dreq
->kref
);
313 kref_get(&dreq
->kref
);
314 init_completion(&dreq
->completion
);
315 INIT_LIST_HEAD(&dreq
->mds_cinfo
.list
);
316 dreq
->verf
.committed
= NFS_INVALID_STABLE_HOW
; /* not set yet */
317 INIT_WORK(&dreq
->work
, nfs_direct_write_schedule_work
);
318 dreq
->mirror_count
= 1;
319 spin_lock_init(&dreq
->lock
);
324 static void nfs_direct_req_free(struct kref
*kref
)
326 struct nfs_direct_req
*dreq
= container_of(kref
, struct nfs_direct_req
, kref
);
328 nfs_free_pnfs_ds_cinfo(&dreq
->ds_cinfo
);
329 if (dreq
->l_ctx
!= NULL
)
330 nfs_put_lock_context(dreq
->l_ctx
);
331 if (dreq
->ctx
!= NULL
)
332 put_nfs_open_context(dreq
->ctx
);
333 kmem_cache_free(nfs_direct_cachep
, dreq
);
336 static void nfs_direct_req_release(struct nfs_direct_req
*dreq
)
338 kref_put(&dreq
->kref
, nfs_direct_req_free
);
341 ssize_t
nfs_dreq_bytes_left(struct nfs_direct_req
*dreq
)
343 return dreq
->bytes_left
;
345 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left
);
348 * Collects and returns the final error value/byte-count.
350 static ssize_t
nfs_direct_wait(struct nfs_direct_req
*dreq
)
352 ssize_t result
= -EIOCBQUEUED
;
354 /* Async requests don't wait here */
358 result
= wait_for_completion_killable(&dreq
->completion
);
361 result
= dreq
->error
;
363 result
= dreq
->count
;
366 return (ssize_t
) result
;
370 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
371 * the iocb is still valid here if this is a synchronous request.
373 static void nfs_direct_complete(struct nfs_direct_req
*dreq
, bool write
)
375 struct inode
*inode
= dreq
->inode
;
377 if (dreq
->iocb
&& write
) {
378 loff_t pos
= dreq
->iocb
->ki_pos
+ dreq
->count
;
380 spin_lock(&inode
->i_lock
);
381 if (i_size_read(inode
) < pos
)
382 i_size_write(inode
, pos
);
383 spin_unlock(&inode
->i_lock
);
387 nfs_zap_mapping(inode
, inode
->i_mapping
);
389 inode_dio_done(inode
);
392 long res
= (long) dreq
->error
;
394 res
= (long) dreq
->count
;
395 dreq
->iocb
->ki_complete(dreq
->iocb
, res
, 0);
398 complete_all(&dreq
->completion
);
400 nfs_direct_req_release(dreq
);
403 static void nfs_direct_readpage_release(struct nfs_page
*req
)
405 dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
406 d_inode(req
->wb_context
->dentry
)->i_sb
->s_id
,
407 (unsigned long long)NFS_FILEID(d_inode(req
->wb_context
->dentry
)),
409 (long long)req_offset(req
));
410 nfs_release_request(req
);
413 static void nfs_direct_read_completion(struct nfs_pgio_header
*hdr
)
415 unsigned long bytes
= 0;
416 struct nfs_direct_req
*dreq
= hdr
->dreq
;
418 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
421 spin_lock(&dreq
->lock
);
422 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
) && (hdr
->good_bytes
== 0))
423 dreq
->error
= hdr
->error
;
425 nfs_direct_good_bytes(dreq
, hdr
);
427 spin_unlock(&dreq
->lock
);
429 while (!list_empty(&hdr
->pages
)) {
430 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
431 struct page
*page
= req
->wb_page
;
433 if (!PageCompound(page
) && bytes
< hdr
->good_bytes
)
434 set_page_dirty(page
);
435 bytes
+= req
->wb_bytes
;
436 nfs_list_remove_request(req
);
437 nfs_direct_readpage_release(req
);
441 nfs_direct_complete(dreq
, false);
445 static void nfs_read_sync_pgio_error(struct list_head
*head
)
447 struct nfs_page
*req
;
449 while (!list_empty(head
)) {
450 req
= nfs_list_entry(head
->next
);
451 nfs_list_remove_request(req
);
452 nfs_release_request(req
);
456 static void nfs_direct_pgio_init(struct nfs_pgio_header
*hdr
)
461 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops
= {
462 .error_cleanup
= nfs_read_sync_pgio_error
,
463 .init_hdr
= nfs_direct_pgio_init
,
464 .completion
= nfs_direct_read_completion
,
468 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
469 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
470 * bail and stop sending more reads. Read length accounting is
471 * handled automatically by nfs_direct_read_result(). Otherwise, if
472 * no requests have been sent, just return an error.
475 static ssize_t
nfs_direct_read_schedule_iovec(struct nfs_direct_req
*dreq
,
476 struct iov_iter
*iter
,
479 struct nfs_pageio_descriptor desc
;
480 struct inode
*inode
= dreq
->inode
;
481 ssize_t result
= -EINVAL
;
482 size_t requested_bytes
= 0;
483 size_t rsize
= max_t(size_t, NFS_SERVER(inode
)->rsize
, PAGE_SIZE
);
485 nfs_pageio_init_read(&desc
, dreq
->inode
, false,
486 &nfs_direct_read_completion_ops
);
489 atomic_inc(&inode
->i_dio_count
);
491 while (iov_iter_count(iter
)) {
492 struct page
**pagevec
;
497 result
= iov_iter_get_pages_alloc(iter
, &pagevec
,
503 iov_iter_advance(iter
, bytes
);
504 npages
= (result
+ pgbase
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
505 for (i
= 0; i
< npages
; i
++) {
506 struct nfs_page
*req
;
507 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
508 /* XXX do we need to do the eof zeroing found in async_filler? */
509 req
= nfs_create_request(dreq
->ctx
, pagevec
[i
], NULL
,
512 result
= PTR_ERR(req
);
515 req
->wb_index
= pos
>> PAGE_SHIFT
;
516 req
->wb_offset
= pos
& ~PAGE_MASK
;
517 if (!nfs_pageio_add_request(&desc
, req
)) {
518 result
= desc
.pg_error
;
519 nfs_release_request(req
);
524 requested_bytes
+= req_len
;
526 dreq
->bytes_left
-= req_len
;
528 nfs_direct_release_pages(pagevec
, npages
);
534 nfs_pageio_complete(&desc
);
537 * If no bytes were started, return the error, and let the
538 * generic layer handle the completion.
540 if (requested_bytes
== 0) {
541 inode_dio_done(inode
);
542 nfs_direct_req_release(dreq
);
543 return result
< 0 ? result
: -EIO
;
547 nfs_direct_complete(dreq
, false);
552 * nfs_file_direct_read - file direct read operation for NFS files
553 * @iocb: target I/O control block
554 * @iter: vector of user buffers into which to read data
555 * @pos: byte offset in file where reading starts
557 * We use this function for direct reads instead of calling
558 * generic_file_aio_read() in order to avoid gfar's check to see if
559 * the request starts before the end of the file. For that check
560 * to work, we must generate a GETATTR before each direct read, and
561 * even then there is a window between the GETATTR and the subsequent
562 * READ where the file size could change. Our preference is simply
563 * to do all reads the application wants, and the server will take
564 * care of managing the end of file boundary.
566 * This function also eliminates unnecessarily updating the file's
567 * atime locally, as the NFS server sets the file's atime, and this
568 * client must read the updated atime from the server back into its
571 ssize_t
nfs_file_direct_read(struct kiocb
*iocb
, struct iov_iter
*iter
,
574 struct file
*file
= iocb
->ki_filp
;
575 struct address_space
*mapping
= file
->f_mapping
;
576 struct inode
*inode
= mapping
->host
;
577 struct nfs_direct_req
*dreq
;
578 struct nfs_lock_context
*l_ctx
;
579 ssize_t result
= -EINVAL
;
580 size_t count
= iov_iter_count(iter
);
581 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTREADBYTES
, count
);
583 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
584 file
, count
, (long long) pos
);
590 mutex_lock(&inode
->i_mutex
);
591 result
= nfs_sync_mapping(mapping
);
595 task_io_account_read(count
);
598 dreq
= nfs_direct_req_alloc();
603 dreq
->bytes_left
= count
;
604 dreq
->io_start
= pos
;
605 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
606 l_ctx
= nfs_get_lock_context(dreq
->ctx
);
608 result
= PTR_ERR(l_ctx
);
612 if (!is_sync_kiocb(iocb
))
615 NFS_I(inode
)->read_io
+= count
;
616 result
= nfs_direct_read_schedule_iovec(dreq
, iter
, pos
);
618 mutex_unlock(&inode
->i_mutex
);
621 result
= nfs_direct_wait(dreq
);
623 iocb
->ki_pos
= pos
+ result
;
626 nfs_direct_req_release(dreq
);
630 nfs_direct_req_release(dreq
);
632 mutex_unlock(&inode
->i_mutex
);
638 nfs_direct_write_scan_commit_list(struct inode
*inode
,
639 struct list_head
*list
,
640 struct nfs_commit_info
*cinfo
)
642 spin_lock(cinfo
->lock
);
643 #ifdef CONFIG_NFS_V4_1
644 if (cinfo
->ds
!= NULL
&& cinfo
->ds
->nwritten
!= 0)
645 NFS_SERVER(inode
)->pnfs_curr_ld
->recover_commit_reqs(list
, cinfo
);
647 nfs_scan_commit_list(&cinfo
->mds
->list
, list
, cinfo
, 0);
648 spin_unlock(cinfo
->lock
);
651 static void nfs_direct_write_reschedule(struct nfs_direct_req
*dreq
)
653 struct nfs_pageio_descriptor desc
;
654 struct nfs_page
*req
, *tmp
;
656 struct nfs_commit_info cinfo
;
660 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
661 nfs_direct_write_scan_commit_list(dreq
->inode
, &reqs
, &cinfo
);
664 for (i
= 0; i
< dreq
->mirror_count
; i
++)
665 dreq
->mirrors
[i
].count
= 0;
668 nfs_pageio_init_write(&desc
, dreq
->inode
, FLUSH_STABLE
, false,
669 &nfs_direct_write_completion_ops
);
672 req
= nfs_list_entry(reqs
.next
);
673 nfs_direct_setup_mirroring(dreq
, &desc
, req
);
675 list_for_each_entry_safe(req
, tmp
, &reqs
, wb_list
) {
676 if (!nfs_pageio_add_request(&desc
, req
)) {
677 nfs_list_remove_request(req
);
678 nfs_list_add_request(req
, &failed
);
679 spin_lock(cinfo
.lock
);
682 spin_unlock(cinfo
.lock
);
684 nfs_release_request(req
);
686 nfs_pageio_complete(&desc
);
688 while (!list_empty(&failed
)) {
689 req
= nfs_list_entry(failed
.next
);
690 nfs_list_remove_request(req
);
691 nfs_unlock_and_release_request(req
);
695 nfs_direct_write_complete(dreq
, dreq
->inode
);
698 static void nfs_direct_commit_complete(struct nfs_commit_data
*data
)
700 struct nfs_direct_req
*dreq
= data
->dreq
;
701 struct nfs_commit_info cinfo
;
702 struct nfs_page
*req
;
703 int status
= data
->task
.tk_status
;
705 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
707 dprintk("NFS: %5u commit failed with error %d.\n",
708 data
->task
.tk_pid
, status
);
709 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
710 } else if (nfs_direct_cmp_commit_data_verf(dreq
, data
)) {
711 dprintk("NFS: %5u commit verify failed\n", data
->task
.tk_pid
);
712 dreq
->flags
= NFS_ODIRECT_RESCHED_WRITES
;
715 dprintk("NFS: %5u commit returned %d\n", data
->task
.tk_pid
, status
);
716 while (!list_empty(&data
->pages
)) {
717 req
= nfs_list_entry(data
->pages
.next
);
718 nfs_list_remove_request(req
);
719 if (dreq
->flags
== NFS_ODIRECT_RESCHED_WRITES
) {
720 /* Note the rewrite will go through mds */
721 nfs_mark_request_commit(req
, NULL
, &cinfo
, 0);
723 nfs_release_request(req
);
724 nfs_unlock_and_release_request(req
);
727 if (atomic_dec_and_test(&cinfo
.mds
->rpcs_out
))
728 nfs_direct_write_complete(dreq
, data
->inode
);
731 static void nfs_direct_error_cleanup(struct nfs_inode
*nfsi
)
733 /* There is no lock to clear */
736 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops
= {
737 .completion
= nfs_direct_commit_complete
,
738 .error_cleanup
= nfs_direct_error_cleanup
,
741 static void nfs_direct_commit_schedule(struct nfs_direct_req
*dreq
)
744 struct nfs_commit_info cinfo
;
747 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
748 nfs_scan_commit(dreq
->inode
, &mds_list
, &cinfo
);
749 res
= nfs_generic_commit_list(dreq
->inode
, &mds_list
, 0, &cinfo
);
750 if (res
< 0) /* res == -ENOMEM */
751 nfs_direct_write_reschedule(dreq
);
754 static void nfs_direct_write_schedule_work(struct work_struct
*work
)
756 struct nfs_direct_req
*dreq
= container_of(work
, struct nfs_direct_req
, work
);
757 int flags
= dreq
->flags
;
761 case NFS_ODIRECT_DO_COMMIT
:
762 nfs_direct_commit_schedule(dreq
);
764 case NFS_ODIRECT_RESCHED_WRITES
:
765 nfs_direct_write_reschedule(dreq
);
768 nfs_direct_complete(dreq
, true);
772 static void nfs_direct_write_complete(struct nfs_direct_req
*dreq
, struct inode
*inode
)
774 schedule_work(&dreq
->work
); /* Calls nfs_direct_write_schedule_work */
777 static void nfs_direct_write_completion(struct nfs_pgio_header
*hdr
)
779 struct nfs_direct_req
*dreq
= hdr
->dreq
;
780 struct nfs_commit_info cinfo
;
781 bool request_commit
= false;
782 struct nfs_page
*req
= nfs_list_entry(hdr
->pages
.next
);
784 if (test_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
787 nfs_init_cinfo_from_dreq(&cinfo
, dreq
);
789 spin_lock(&dreq
->lock
);
791 if (test_bit(NFS_IOHDR_ERROR
, &hdr
->flags
)) {
793 dreq
->error
= hdr
->error
;
795 if (dreq
->error
== 0) {
796 nfs_direct_good_bytes(dreq
, hdr
);
797 if (nfs_write_need_commit(hdr
)) {
798 if (dreq
->flags
== NFS_ODIRECT_RESCHED_WRITES
)
799 request_commit
= true;
800 else if (dreq
->flags
== 0) {
801 nfs_direct_set_hdr_verf(dreq
, hdr
);
802 request_commit
= true;
803 dreq
->flags
= NFS_ODIRECT_DO_COMMIT
;
804 } else if (dreq
->flags
== NFS_ODIRECT_DO_COMMIT
) {
805 request_commit
= true;
806 if (nfs_direct_set_or_cmp_hdr_verf(dreq
, hdr
))
808 NFS_ODIRECT_RESCHED_WRITES
;
812 spin_unlock(&dreq
->lock
);
814 while (!list_empty(&hdr
->pages
)) {
816 req
= nfs_list_entry(hdr
->pages
.next
);
817 nfs_list_remove_request(req
);
818 if (request_commit
) {
819 kref_get(&req
->wb_kref
);
820 nfs_mark_request_commit(req
, hdr
->lseg
, &cinfo
,
823 nfs_unlock_and_release_request(req
);
828 nfs_direct_write_complete(dreq
, hdr
->inode
);
832 static void nfs_write_sync_pgio_error(struct list_head
*head
)
834 struct nfs_page
*req
;
836 while (!list_empty(head
)) {
837 req
= nfs_list_entry(head
->next
);
838 nfs_list_remove_request(req
);
839 nfs_unlock_and_release_request(req
);
843 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops
= {
844 .error_cleanup
= nfs_write_sync_pgio_error
,
845 .init_hdr
= nfs_direct_pgio_init
,
846 .completion
= nfs_direct_write_completion
,
851 * NB: Return the value of the first error return code. Subsequent
852 * errors after the first one are ignored.
855 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
856 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
857 * bail and stop sending more writes. Write length accounting is
858 * handled automatically by nfs_direct_write_result(). Otherwise, if
859 * no requests have been sent, just return an error.
861 static ssize_t
nfs_direct_write_schedule_iovec(struct nfs_direct_req
*dreq
,
862 struct iov_iter
*iter
,
865 struct nfs_pageio_descriptor desc
;
866 struct inode
*inode
= dreq
->inode
;
868 size_t requested_bytes
= 0;
869 size_t wsize
= max_t(size_t, NFS_SERVER(inode
)->wsize
, PAGE_SIZE
);
871 nfs_pageio_init_write(&desc
, inode
, FLUSH_COND_STABLE
, false,
872 &nfs_direct_write_completion_ops
);
875 atomic_inc(&inode
->i_dio_count
);
877 NFS_I(inode
)->write_io
+= iov_iter_count(iter
);
878 while (iov_iter_count(iter
)) {
879 struct page
**pagevec
;
884 result
= iov_iter_get_pages_alloc(iter
, &pagevec
,
890 iov_iter_advance(iter
, bytes
);
891 npages
= (result
+ pgbase
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
892 for (i
= 0; i
< npages
; i
++) {
893 struct nfs_page
*req
;
894 unsigned int req_len
= min_t(size_t, bytes
, PAGE_SIZE
- pgbase
);
896 req
= nfs_create_request(dreq
->ctx
, pagevec
[i
], NULL
,
899 result
= PTR_ERR(req
);
903 nfs_direct_setup_mirroring(dreq
, &desc
, req
);
905 nfs_lock_request(req
);
906 req
->wb_index
= pos
>> PAGE_SHIFT
;
907 req
->wb_offset
= pos
& ~PAGE_MASK
;
908 if (!nfs_pageio_add_request(&desc
, req
)) {
909 result
= desc
.pg_error
;
910 nfs_unlock_and_release_request(req
);
915 requested_bytes
+= req_len
;
917 dreq
->bytes_left
-= req_len
;
919 nfs_direct_release_pages(pagevec
, npages
);
924 nfs_pageio_complete(&desc
);
927 * If no bytes were started, return the error, and let the
928 * generic layer handle the completion.
930 if (requested_bytes
== 0) {
931 inode_dio_done(inode
);
932 nfs_direct_req_release(dreq
);
933 return result
< 0 ? result
: -EIO
;
937 nfs_direct_write_complete(dreq
, dreq
->inode
);
942 * nfs_file_direct_write - file direct write operation for NFS files
943 * @iocb: target I/O control block
944 * @iter: vector of user buffers from which to write data
945 * @pos: byte offset in file where writing starts
947 * We use this function for direct writes instead of calling
948 * generic_file_aio_write() in order to avoid taking the inode
949 * semaphore and updating the i_size. The NFS server will set
950 * the new i_size and this client must read the updated size
951 * back into its cache. We let the server do generic write
952 * parameter checking and report problems.
954 * We eliminate local atime updates, see direct read above.
956 * We avoid unnecessary page cache invalidations for normal cached
957 * readers of this file.
959 * Note that O_APPEND is not supported for NFS direct writes, as there
960 * is no atomic O_APPEND write facility in the NFS protocol.
962 ssize_t
nfs_file_direct_write(struct kiocb
*iocb
, struct iov_iter
*iter
)
964 ssize_t result
= -EINVAL
;
965 struct file
*file
= iocb
->ki_filp
;
966 struct address_space
*mapping
= file
->f_mapping
;
967 struct inode
*inode
= mapping
->host
;
968 struct nfs_direct_req
*dreq
;
969 struct nfs_lock_context
*l_ctx
;
972 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
973 file
, iov_iter_count(iter
), (long long) iocb
->ki_pos
);
975 nfs_add_stats(mapping
->host
, NFSIOS_DIRECTWRITTENBYTES
,
976 iov_iter_count(iter
));
979 end
= (pos
+ iov_iter_count(iter
) - 1) >> PAGE_CACHE_SHIFT
;
981 mutex_lock(&inode
->i_mutex
);
983 result
= nfs_sync_mapping(mapping
);
987 if (mapping
->nrpages
) {
988 result
= invalidate_inode_pages2_range(mapping
,
989 pos
>> PAGE_CACHE_SHIFT
, end
);
994 task_io_account_write(iov_iter_count(iter
));
997 dreq
= nfs_direct_req_alloc();
1001 dreq
->inode
= inode
;
1002 dreq
->bytes_left
= iov_iter_count(iter
);
1003 dreq
->io_start
= pos
;
1004 dreq
->ctx
= get_nfs_open_context(nfs_file_open_context(iocb
->ki_filp
));
1005 l_ctx
= nfs_get_lock_context(dreq
->ctx
);
1006 if (IS_ERR(l_ctx
)) {
1007 result
= PTR_ERR(l_ctx
);
1010 dreq
->l_ctx
= l_ctx
;
1011 if (!is_sync_kiocb(iocb
))
1014 result
= nfs_direct_write_schedule_iovec(dreq
, iter
, pos
);
1016 if (mapping
->nrpages
) {
1017 invalidate_inode_pages2_range(mapping
,
1018 pos
>> PAGE_CACHE_SHIFT
, end
);
1021 mutex_unlock(&inode
->i_mutex
);
1024 result
= nfs_direct_wait(dreq
);
1026 struct inode
*inode
= mapping
->host
;
1028 iocb
->ki_pos
= pos
+ result
;
1029 spin_lock(&inode
->i_lock
);
1030 if (i_size_read(inode
) < iocb
->ki_pos
)
1031 i_size_write(inode
, iocb
->ki_pos
);
1032 spin_unlock(&inode
->i_lock
);
1035 nfs_direct_req_release(dreq
);
1039 nfs_direct_req_release(dreq
);
1041 mutex_unlock(&inode
->i_mutex
);
1046 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1049 int __init
nfs_init_directcache(void)
1051 nfs_direct_cachep
= kmem_cache_create("nfs_direct_cache",
1052 sizeof(struct nfs_direct_req
),
1053 0, (SLAB_RECLAIM_ACCOUNT
|
1056 if (nfs_direct_cachep
== NULL
)
1063 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1066 void nfs_destroy_directcache(void)
1068 kmem_cache_destroy(nfs_direct_cachep
);