2 * linux/fs/nfs/direct.c
4 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
6 * High-performance uncached I/O for the Linux NFS client
8 * There are important applications whose performance or correctness
9 * depends on uncached access to file data. Database clusters
10 * (multiple copies of the same instance running on separate hosts)
11 * implement their own cache coherency protocol that subsumes file
12 * system cache protocols. Applications that process datasets
13 * considerably larger than the client's memory do not always benefit
14 * from a local cache. A streaming video server, for instance, has no
15 * need to cache the contents of a file.
17 * When an application requests uncached I/O, all read and write requests
18 * are made directly to the server; data stored or fetched via these
19 * requests is not cached in the Linux page cache. The client does not
20 * correct unaligned requests from applications. All requested bytes are
21 * held on permanent storage before a direct write system call returns to
24 * Solaris implements an uncached I/O facility called directio() that
25 * is used for backups and sequential I/O to very large files. Solaris
26 * also supports uncaching whole NFS partitions with "-o forcedirectio,"
27 * an undocumented mount option.
29 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
30 * help from Andrew Morton.
32 * 18 Dec 2001 Initial implementation for 2.4 --cel
33 * 08 Jul 2002 Version for 2.4.19, with bug fixes --trondmy
34 * 08 Jun 2003 Port to 2.5 APIs --cel
35 * 31 Mar 2004 Handle direct I/O without VFS support --cel
36 * 15 Sep 2004 Parallel async reads --cel
37 * 04 May 2005 support O_DIRECT with aio --cel
41 #include <linux/errno.h>
42 #include <linux/sched.h>
43 #include <linux/kernel.h>
44 #include <linux/file.h>
45 #include <linux/pagemap.h>
46 #include <linux/kref.h>
48 #include <linux/nfs_fs.h>
49 #include <linux/nfs_page.h>
50 #include <linux/sunrpc/clnt.h>
52 #include <asm/system.h>
53 #include <asm/uaccess.h>
54 #include <asm/atomic.h>
59 #define NFSDBG_FACILITY NFSDBG_VFS
61 static struct kmem_cache *nfs_direct_cachep;
64 * This represents a set of asynchronous requests that we're waiting on
66 struct nfs_direct_req {
67 struct kref kref; /* release manager */
70 struct nfs_open_context *ctx; /* file open context info */
71 struct kiocb * iocb; /* controlling i/o request */
72 struct inode * inode; /* target file of i/o */
74 /* completion state */
75 atomic_t io_count; /* i/os we're waiting for */
76 spinlock_t lock; /* protect completion state */
77 ssize_t count, /* bytes actually processed */
78 error; /* any reported error */
79 struct completion completion; /* wait for i/o completion */
82 struct list_head rewrite_list; /* saved nfs_write_data structs */
83 struct nfs_write_data * commit_data; /* special write_data for commits */
85 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
86 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
87 struct nfs_writeverf verf; /* unstable write verifier */
90 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
91 static const struct rpc_call_ops nfs_write_direct_ops;
93 static inline void get_dreq(struct nfs_direct_req *dreq)
95 atomic_inc(&dreq->io_count);
98 static inline int put_dreq(struct nfs_direct_req *dreq)
100 return atomic_dec_and_test(&dreq->io_count);
104 * nfs_direct_IO - NFS address space operation for direct I/O
105 * @rw: direction (read or write)
106 * @iocb: target I/O control block
107 * @iov: array of vectors that define I/O buffer
108 * @pos: offset in file to begin the operation
109 * @nr_segs: size of iovec array
111 * The presence of this routine in the address space ops vector means
112 * the NFS client supports direct I/O. However, we shunt off direct
113 * read and write requests before the VFS gets them, so this method
114 * should never be called.
116 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
118 dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
119 iocb->ki_filp->f_path.dentry->d_name.name,
120 (long long) pos, nr_segs);
125 static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
132 pages += (pgbase >> PAGE_SHIFT);
133 npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
134 for (i = 0; i < npages; i++) {
135 struct page *page = pages[i];
136 if (!PageCompound(page))
137 set_page_dirty(page);
141 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
144 for (i = 0; i < npages; i++)
145 page_cache_release(pages[i]);
148 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
150 struct nfs_direct_req *dreq;
152 dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
156 kref_init(&dreq->kref);
157 kref_get(&dreq->kref);
158 init_completion(&dreq->completion);
159 INIT_LIST_HEAD(&dreq->rewrite_list);
162 spin_lock_init(&dreq->lock);
163 atomic_set(&dreq->io_count, 0);
171 static void nfs_direct_req_free(struct kref *kref)
173 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
175 if (dreq->ctx != NULL)
176 put_nfs_open_context(dreq->ctx);
177 kmem_cache_free(nfs_direct_cachep, dreq);
180 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
182 kref_put(&dreq->kref, nfs_direct_req_free);
186 * Collects and returns the final error value/byte-count.
188 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
190 ssize_t result = -EIOCBQUEUED;
192 /* Async requests don't wait here */
196 result = wait_for_completion_interruptible(&dreq->completion);
199 result = dreq->error;
201 result = dreq->count;
204 return (ssize_t) result;
208 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
209 * the iocb is still valid here if this is a synchronous request.
211 static void nfs_direct_complete(struct nfs_direct_req *dreq)
214 long res = (long) dreq->error;
216 res = (long) dreq->count;
217 aio_complete(dreq->iocb, res, 0);
219 complete_all(&dreq->completion);
221 nfs_direct_req_release(dreq);
225 * We must hold a reference to all the pages in this direct read request
226 * until the RPCs complete. This could be long *after* we are woken up in
227 * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
229 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
231 struct nfs_read_data *data = calldata;
232 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
234 if (nfs_readpage_result(task, data) != 0)
237 spin_lock(&dreq->lock);
238 if (unlikely(task->tk_status < 0)) {
239 dreq->error = task->tk_status;
240 spin_unlock(&dreq->lock);
242 dreq->count += data->res.count;
243 spin_unlock(&dreq->lock);
244 nfs_direct_dirty_pages(data->pagevec,
248 nfs_direct_release_pages(data->pagevec, data->npages);
251 nfs_direct_complete(dreq);
254 static const struct rpc_call_ops nfs_read_direct_ops = {
255 .rpc_call_done = nfs_direct_read_result,
256 .rpc_release = nfs_readdata_release,
260 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
261 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
262 * bail and stop sending more reads. Read length accounting is
263 * handled automatically by nfs_direct_read_result(). Otherwise, if
264 * no requests have been sent, just return an error.
266 static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
268 struct nfs_open_context *ctx = dreq->ctx;
269 struct inode *inode = ctx->path.dentry->d_inode;
270 size_t rsize = NFS_SERVER(inode)->rsize;
276 struct nfs_read_data *data;
279 pgbase = user_addr & ~PAGE_MASK;
280 bytes = min(rsize,count);
283 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
287 down_read(¤t->mm->mmap_sem);
288 result = get_user_pages(current, current->mm, user_addr,
289 data->npages, 1, 0, data->pagevec, NULL);
290 up_read(¤t->mm->mmap_sem);
292 nfs_readdata_release(data);
295 if ((unsigned)result < data->npages) {
296 bytes = result * PAGE_SIZE;
297 if (bytes <= pgbase) {
298 nfs_direct_release_pages(data->pagevec, result);
299 nfs_readdata_release(data);
303 data->npages = result;
308 data->req = (struct nfs_page *) dreq;
310 data->cred = ctx->cred;
311 data->args.fh = NFS_FH(inode);
312 data->args.context = ctx;
313 data->args.offset = pos;
314 data->args.pgbase = pgbase;
315 data->args.pages = data->pagevec;
316 data->args.count = bytes;
317 data->res.fattr = &data->fattr;
319 data->res.count = bytes;
321 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
322 &nfs_read_direct_ops, data);
323 NFS_PROTO(inode)->read_setup(data);
325 data->task.tk_cookie = (unsigned long) inode;
327 rpc_execute(&data->task);
329 dprintk("NFS: %5u initiated direct read call "
330 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
333 (long long)NFS_FILEID(inode),
335 (unsigned long long)data->args.offset);
340 /* FIXME: Remove this unnecessary math from final patch */
342 pgbase &= ~PAGE_MASK;
343 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
346 } while (count != 0);
350 return result < 0 ? (ssize_t) result : -EFAULT;
353 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
354 const struct iovec *iov,
355 unsigned long nr_segs,
358 ssize_t result = -EINVAL;
359 size_t requested_bytes = 0;
364 for (seg = 0; seg < nr_segs; seg++) {
365 const struct iovec *vec = &iov[seg];
366 result = nfs_direct_read_schedule(dreq,
367 (unsigned long)vec->iov_base,
371 requested_bytes += result;
372 if ((size_t)result < vec->iov_len)
378 nfs_direct_complete(dreq);
380 if (requested_bytes != 0)
388 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
389 unsigned long nr_segs, loff_t pos)
393 struct inode *inode = iocb->ki_filp->f_mapping->host;
394 struct rpc_clnt *clnt = NFS_CLIENT(inode);
395 struct nfs_direct_req *dreq;
397 dreq = nfs_direct_req_alloc();
402 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
403 if (!is_sync_kiocb(iocb))
406 rpc_clnt_sigmask(clnt, &oldset);
407 result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
409 result = nfs_direct_wait(dreq);
410 rpc_clnt_sigunmask(clnt, &oldset);
411 nfs_direct_req_release(dreq);
416 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
418 while (!list_empty(&dreq->rewrite_list)) {
419 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
420 list_del(&data->pages);
421 nfs_direct_release_pages(data->pagevec, data->npages);
422 nfs_writedata_release(data);
426 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
427 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
429 struct inode *inode = dreq->inode;
431 struct nfs_write_data *data;
436 list_for_each(p, &dreq->rewrite_list) {
437 data = list_entry(p, struct nfs_write_data, pages);
444 nfs_fattr_init(&data->fattr);
445 data->res.count = data->args.count;
446 memset(&data->verf, 0, sizeof(data->verf));
449 * Reuse data->task; data->args should not have changed
450 * since the original request was sent.
452 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
453 &nfs_write_direct_ops, data);
454 NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
456 data->task.tk_priority = RPC_PRIORITY_NORMAL;
457 data->task.tk_cookie = (unsigned long) inode;
460 * We're called via an RPC callback, so BKL is already held.
462 rpc_execute(&data->task);
464 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
467 (long long)NFS_FILEID(inode),
469 (unsigned long long)data->args.offset);
473 nfs_direct_write_complete(dreq, inode);
476 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
478 struct nfs_write_data *data = calldata;
479 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
481 /* Call the NFS version-specific code */
482 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
484 if (unlikely(task->tk_status < 0)) {
485 dprintk("NFS: %5u commit failed with error %d.\n",
486 task->tk_pid, task->tk_status);
487 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
488 } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
489 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
490 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
493 dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
494 nfs_direct_write_complete(dreq, data->inode);
497 static const struct rpc_call_ops nfs_commit_direct_ops = {
498 .rpc_call_done = nfs_direct_commit_result,
499 .rpc_release = nfs_commit_release,
502 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
504 struct nfs_write_data *data = dreq->commit_data;
506 data->inode = dreq->inode;
507 data->cred = dreq->ctx->cred;
509 data->args.fh = NFS_FH(data->inode);
510 data->args.offset = 0;
511 data->args.count = 0;
513 data->res.fattr = &data->fattr;
514 data->res.verf = &data->verf;
516 rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
517 &nfs_commit_direct_ops, data);
518 NFS_PROTO(data->inode)->commit_setup(data, 0);
520 data->task.tk_priority = RPC_PRIORITY_NORMAL;
521 data->task.tk_cookie = (unsigned long)data->inode;
522 /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
523 dreq->commit_data = NULL;
525 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
527 rpc_execute(&data->task);
530 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
532 int flags = dreq->flags;
536 case NFS_ODIRECT_DO_COMMIT:
537 nfs_direct_commit_schedule(dreq);
539 case NFS_ODIRECT_RESCHED_WRITES:
540 nfs_direct_write_reschedule(dreq);
543 if (dreq->commit_data != NULL)
544 nfs_commit_free(dreq->commit_data);
545 nfs_direct_free_writedata(dreq);
546 nfs_zap_mapping(inode, inode->i_mapping);
547 nfs_direct_complete(dreq);
551 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
553 dreq->commit_data = nfs_commit_alloc();
554 if (dreq->commit_data != NULL)
555 dreq->commit_data->req = (struct nfs_page *) dreq;
558 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
560 dreq->commit_data = NULL;
563 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
565 nfs_direct_free_writedata(dreq);
566 nfs_zap_mapping(inode, inode->i_mapping);
567 nfs_direct_complete(dreq);
571 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
573 struct nfs_write_data *data = calldata;
574 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
575 int status = task->tk_status;
577 if (nfs_writeback_done(task, data) != 0)
580 spin_lock(&dreq->lock);
582 if (unlikely(status < 0)) {
583 /* An error has occurred, so we should not commit */
585 dreq->error = status;
587 if (unlikely(dreq->error != 0))
590 dreq->count += data->res.count;
592 if (data->res.verf->committed != NFS_FILE_SYNC) {
593 switch (dreq->flags) {
595 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
596 dreq->flags = NFS_ODIRECT_DO_COMMIT;
598 case NFS_ODIRECT_DO_COMMIT:
599 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
600 dprintk("NFS: %5u write verify failed\n", task->tk_pid);
601 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
606 spin_unlock(&dreq->lock);
610 * NB: Return the value of the first error return code. Subsequent
611 * errors after the first one are ignored.
613 static void nfs_direct_write_release(void *calldata)
615 struct nfs_write_data *data = calldata;
616 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
619 nfs_direct_write_complete(dreq, data->inode);
622 static const struct rpc_call_ops nfs_write_direct_ops = {
623 .rpc_call_done = nfs_direct_write_result,
624 .rpc_release = nfs_direct_write_release,
628 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
629 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
630 * bail and stop sending more writes. Write length accounting is
631 * handled automatically by nfs_direct_write_result(). Otherwise, if
632 * no requests have been sent, just return an error.
634 static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
636 struct nfs_open_context *ctx = dreq->ctx;
637 struct inode *inode = ctx->path.dentry->d_inode;
638 size_t wsize = NFS_SERVER(inode)->wsize;
644 struct nfs_write_data *data;
647 pgbase = user_addr & ~PAGE_MASK;
648 bytes = min(wsize,count);
651 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
655 down_read(¤t->mm->mmap_sem);
656 result = get_user_pages(current, current->mm, user_addr,
657 data->npages, 0, 0, data->pagevec, NULL);
658 up_read(¤t->mm->mmap_sem);
660 nfs_writedata_release(data);
663 if ((unsigned)result < data->npages) {
664 bytes = result * PAGE_SIZE;
665 if (bytes <= pgbase) {
666 nfs_direct_release_pages(data->pagevec, result);
667 nfs_writedata_release(data);
671 data->npages = result;
676 list_move_tail(&data->pages, &dreq->rewrite_list);
678 data->req = (struct nfs_page *) dreq;
680 data->cred = ctx->cred;
681 data->args.fh = NFS_FH(inode);
682 data->args.context = ctx;
683 data->args.offset = pos;
684 data->args.pgbase = pgbase;
685 data->args.pages = data->pagevec;
686 data->args.count = bytes;
687 data->res.fattr = &data->fattr;
688 data->res.count = bytes;
689 data->res.verf = &data->verf;
691 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
692 &nfs_write_direct_ops, data);
693 NFS_PROTO(inode)->write_setup(data, sync);
695 data->task.tk_priority = RPC_PRIORITY_NORMAL;
696 data->task.tk_cookie = (unsigned long) inode;
698 rpc_execute(&data->task);
700 dprintk("NFS: %5u initiated direct write call "
701 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
704 (long long)NFS_FILEID(inode),
706 (unsigned long long)data->args.offset);
712 /* FIXME: Remove this useless math from the final patch */
714 pgbase &= ~PAGE_MASK;
715 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
718 } while (count != 0);
722 return result < 0 ? (ssize_t) result : -EFAULT;
725 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
726 const struct iovec *iov,
727 unsigned long nr_segs,
728 loff_t pos, int sync)
731 size_t requested_bytes = 0;
736 for (seg = 0; seg < nr_segs; seg++) {
737 const struct iovec *vec = &iov[seg];
738 result = nfs_direct_write_schedule(dreq,
739 (unsigned long)vec->iov_base,
744 requested_bytes += result;
745 if ((size_t)result < vec->iov_len)
751 nfs_direct_write_complete(dreq, dreq->inode);
753 if (requested_bytes != 0)
761 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
762 unsigned long nr_segs, loff_t pos,
767 struct inode *inode = iocb->ki_filp->f_mapping->host;
768 struct rpc_clnt *clnt = NFS_CLIENT(inode);
769 struct nfs_direct_req *dreq;
770 size_t wsize = NFS_SERVER(inode)->wsize;
773 dreq = nfs_direct_req_alloc();
776 nfs_alloc_commit_data(dreq);
778 if (dreq->commit_data == NULL || count < wsize)
782 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
783 if (!is_sync_kiocb(iocb))
786 rpc_clnt_sigmask(clnt, &oldset);
787 result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
789 result = nfs_direct_wait(dreq);
790 rpc_clnt_sigunmask(clnt, &oldset);
791 nfs_direct_req_release(dreq);
797 * nfs_file_direct_read - file direct read operation for NFS files
798 * @iocb: target I/O control block
799 * @iov: vector of user buffers into which to read data
800 * @nr_segs: size of iov vector
801 * @pos: byte offset in file where reading starts
803 * We use this function for direct reads instead of calling
804 * generic_file_aio_read() in order to avoid gfar's check to see if
805 * the request starts before the end of the file. For that check
806 * to work, we must generate a GETATTR before each direct read, and
807 * even then there is a window between the GETATTR and the subsequent
808 * READ where the file size could change. Our preference is simply
809 * to do all reads the application wants, and the server will take
810 * care of managing the end of file boundary.
812 * This function also eliminates unnecessarily updating the file's
813 * atime locally, as the NFS server sets the file's atime, and this
814 * client must read the updated atime from the server back into its
817 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
818 unsigned long nr_segs, loff_t pos)
820 ssize_t retval = -EINVAL;
821 struct file *file = iocb->ki_filp;
822 struct address_space *mapping = file->f_mapping;
825 count = iov_length(iov, nr_segs);
826 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
828 dprintk("nfs: direct read(%s/%s, %zd@%Ld)\n",
829 file->f_path.dentry->d_parent->d_name.name,
830 file->f_path.dentry->d_name.name,
831 count, (long long) pos);
837 retval = nfs_sync_mapping(mapping);
841 retval = nfs_direct_read(iocb, iov, nr_segs, pos);
843 iocb->ki_pos = pos + retval;
850 * nfs_file_direct_write - file direct write operation for NFS files
851 * @iocb: target I/O control block
852 * @iov: vector of user buffers from which to write data
853 * @nr_segs: size of iov vector
854 * @pos: byte offset in file where writing starts
856 * We use this function for direct writes instead of calling
857 * generic_file_aio_write() in order to avoid taking the inode
858 * semaphore and updating the i_size. The NFS server will set
859 * the new i_size and this client must read the updated size
860 * back into its cache. We let the server do generic write
861 * parameter checking and report problems.
863 * We also avoid an unnecessary invocation of generic_osync_inode(),
864 * as it is fairly meaningless to sync the metadata of an NFS file.
866 * We eliminate local atime updates, see direct read above.
868 * We avoid unnecessary page cache invalidations for normal cached
869 * readers of this file.
871 * Note that O_APPEND is not supported for NFS direct writes, as there
872 * is no atomic O_APPEND write facility in the NFS protocol.
874 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
875 unsigned long nr_segs, loff_t pos)
877 ssize_t retval = -EINVAL;
878 struct file *file = iocb->ki_filp;
879 struct address_space *mapping = file->f_mapping;
882 count = iov_length(iov, nr_segs);
883 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
885 dfprintk(VFS, "nfs: direct write(%s/%s, %zd@%Ld)\n",
886 file->f_path.dentry->d_parent->d_name.name,
887 file->f_path.dentry->d_name.name,
888 count, (long long) pos);
890 retval = generic_write_checks(file, &pos, &count, 0);
894 goto out; /* return 0 */
897 if ((ssize_t) count < 0)
903 retval = nfs_sync_mapping(mapping);
907 retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
910 iocb->ki_pos = pos + retval;
917 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
920 int __init nfs_init_directcache(void)
922 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
923 sizeof(struct nfs_direct_req),
924 0, (SLAB_RECLAIM_ACCOUNT|
927 if (nfs_direct_cachep == NULL)
934 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
937 void nfs_destroy_directcache(void)
939 kmem_cache_destroy(nfs_direct_cachep);