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_killable(&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;
233 nfs_readpage_result(task, data);
236 static void nfs_direct_read_release(void *calldata)
239 struct nfs_read_data *data = calldata;
240 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
241 int status = data->task.tk_status;
243 spin_lock(&dreq->lock);
244 if (unlikely(status < 0)) {
245 dreq->error = status;
246 spin_unlock(&dreq->lock);
248 dreq->count += data->res.count;
249 spin_unlock(&dreq->lock);
250 nfs_direct_dirty_pages(data->pagevec,
254 nfs_direct_release_pages(data->pagevec, data->npages);
257 nfs_direct_complete(dreq);
258 nfs_readdata_release(calldata);
261 static const struct rpc_call_ops nfs_read_direct_ops = {
262 .rpc_call_done = nfs_direct_read_result,
263 .rpc_release = nfs_direct_read_release,
267 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
268 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
269 * bail and stop sending more reads. Read length accounting is
270 * handled automatically by nfs_direct_read_result(). Otherwise, if
271 * no requests have been sent, just return an error.
273 static ssize_t nfs_direct_read_schedule_segment(struct nfs_direct_req *dreq,
274 const struct iovec *iov,
277 struct nfs_open_context *ctx = dreq->ctx;
278 struct inode *inode = ctx->path.dentry->d_inode;
279 unsigned long user_addr = (unsigned long)iov->iov_base;
280 size_t count = iov->iov_len;
281 size_t rsize = NFS_SERVER(inode)->rsize;
282 struct rpc_task *task;
283 struct rpc_message msg = {
284 .rpc_cred = ctx->cred,
286 struct rpc_task_setup task_setup_data = {
287 .rpc_client = NFS_CLIENT(inode),
289 .callback_ops = &nfs_read_direct_ops,
290 .workqueue = nfsiod_workqueue,
291 .flags = RPC_TASK_ASYNC,
298 struct nfs_read_data *data;
301 pgbase = user_addr & ~PAGE_MASK;
302 bytes = min(rsize,count);
305 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
309 down_read(¤t->mm->mmap_sem);
310 result = get_user_pages(current, current->mm, user_addr,
311 data->npages, 1, 0, data->pagevec, NULL);
312 up_read(¤t->mm->mmap_sem);
314 nfs_readdata_release(data);
317 if ((unsigned)result < data->npages) {
318 bytes = result * PAGE_SIZE;
319 if (bytes <= pgbase) {
320 nfs_direct_release_pages(data->pagevec, result);
321 nfs_readdata_release(data);
325 data->npages = result;
330 data->req = (struct nfs_page *) dreq;
332 data->cred = msg.rpc_cred;
333 data->args.fh = NFS_FH(inode);
334 data->args.context = get_nfs_open_context(ctx);
335 data->args.offset = pos;
336 data->args.pgbase = pgbase;
337 data->args.pages = data->pagevec;
338 data->args.count = bytes;
339 data->res.fattr = &data->fattr;
341 data->res.count = bytes;
342 msg.rpc_argp = &data->args;
343 msg.rpc_resp = &data->res;
345 task_setup_data.task = &data->task;
346 task_setup_data.callback_data = data;
347 NFS_PROTO(inode)->read_setup(data, &msg);
349 task = rpc_run_task(&task_setup_data);
353 dprintk("NFS: %5u initiated direct read call "
354 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
357 (long long)NFS_FILEID(inode),
359 (unsigned long long)data->args.offset);
364 /* FIXME: Remove this unnecessary math from final patch */
366 pgbase &= ~PAGE_MASK;
367 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
370 } while (count != 0);
374 return result < 0 ? (ssize_t) result : -EFAULT;
377 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
378 const struct iovec *iov,
379 unsigned long nr_segs,
382 ssize_t result = -EINVAL;
383 size_t requested_bytes = 0;
388 for (seg = 0; seg < nr_segs; seg++) {
389 const struct iovec *vec = &iov[seg];
390 result = nfs_direct_read_schedule_segment(dreq, vec, pos);
393 requested_bytes += result;
394 if ((size_t)result < vec->iov_len)
400 nfs_direct_complete(dreq);
402 if (requested_bytes != 0)
410 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
411 unsigned long nr_segs, loff_t pos)
414 struct inode *inode = iocb->ki_filp->f_mapping->host;
415 struct nfs_direct_req *dreq;
417 dreq = nfs_direct_req_alloc();
422 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
423 if (!is_sync_kiocb(iocb))
426 result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
428 result = nfs_direct_wait(dreq);
429 nfs_direct_req_release(dreq);
434 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
436 while (!list_empty(&dreq->rewrite_list)) {
437 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
438 list_del(&data->pages);
439 nfs_direct_release_pages(data->pagevec, data->npages);
440 nfs_writedata_release(data);
444 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
445 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
447 struct inode *inode = dreq->inode;
449 struct nfs_write_data *data;
450 struct rpc_task *task;
451 struct rpc_message msg = {
452 .rpc_cred = dreq->ctx->cred,
454 struct rpc_task_setup task_setup_data = {
455 .rpc_client = NFS_CLIENT(inode),
456 .callback_ops = &nfs_write_direct_ops,
457 .workqueue = nfsiod_workqueue,
458 .flags = RPC_TASK_ASYNC,
464 list_for_each(p, &dreq->rewrite_list) {
465 data = list_entry(p, struct nfs_write_data, pages);
469 /* Use stable writes */
470 data->args.stable = NFS_FILE_SYNC;
475 nfs_fattr_init(&data->fattr);
476 data->res.count = data->args.count;
477 memset(&data->verf, 0, sizeof(data->verf));
480 * Reuse data->task; data->args should not have changed
481 * since the original request was sent.
483 task_setup_data.task = &data->task;
484 task_setup_data.callback_data = data;
485 msg.rpc_argp = &data->args;
486 msg.rpc_resp = &data->res;
487 NFS_PROTO(inode)->write_setup(data, &msg);
490 * We're called via an RPC callback, so BKL is already held.
492 task = rpc_run_task(&task_setup_data);
496 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
499 (long long)NFS_FILEID(inode),
501 (unsigned long long)data->args.offset);
505 nfs_direct_write_complete(dreq, inode);
508 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
510 struct nfs_write_data *data = calldata;
511 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
513 /* Call the NFS version-specific code */
514 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
516 if (unlikely(task->tk_status < 0)) {
517 dprintk("NFS: %5u commit failed with error %d.\n",
518 task->tk_pid, task->tk_status);
519 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
520 } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
521 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
522 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
525 dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
526 nfs_direct_write_complete(dreq, data->inode);
529 static const struct rpc_call_ops nfs_commit_direct_ops = {
530 .rpc_call_done = nfs_direct_commit_result,
531 .rpc_release = nfs_commit_release,
534 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
536 struct nfs_write_data *data = dreq->commit_data;
537 struct rpc_task *task;
538 struct rpc_message msg = {
539 .rpc_argp = &data->args,
540 .rpc_resp = &data->res,
541 .rpc_cred = dreq->ctx->cred,
543 struct rpc_task_setup task_setup_data = {
545 .rpc_client = NFS_CLIENT(dreq->inode),
547 .callback_ops = &nfs_commit_direct_ops,
548 .callback_data = data,
549 .workqueue = nfsiod_workqueue,
550 .flags = RPC_TASK_ASYNC,
553 data->inode = dreq->inode;
554 data->cred = msg.rpc_cred;
556 data->args.fh = NFS_FH(data->inode);
557 data->args.offset = 0;
558 data->args.count = 0;
559 data->args.context = get_nfs_open_context(dreq->ctx);
561 data->res.fattr = &data->fattr;
562 data->res.verf = &data->verf;
564 NFS_PROTO(data->inode)->commit_setup(data, &msg);
566 /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
567 dreq->commit_data = NULL;
569 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
571 task = rpc_run_task(&task_setup_data);
576 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
578 int flags = dreq->flags;
582 case NFS_ODIRECT_DO_COMMIT:
583 nfs_direct_commit_schedule(dreq);
585 case NFS_ODIRECT_RESCHED_WRITES:
586 nfs_direct_write_reschedule(dreq);
589 if (dreq->commit_data != NULL)
590 nfs_commit_free(dreq->commit_data);
591 nfs_direct_free_writedata(dreq);
592 nfs_zap_mapping(inode, inode->i_mapping);
593 nfs_direct_complete(dreq);
597 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
599 dreq->commit_data = nfs_commit_alloc();
600 if (dreq->commit_data != NULL)
601 dreq->commit_data->req = (struct nfs_page *) dreq;
604 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
606 dreq->commit_data = NULL;
609 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
611 nfs_direct_free_writedata(dreq);
612 nfs_zap_mapping(inode, inode->i_mapping);
613 nfs_direct_complete(dreq);
617 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
619 struct nfs_write_data *data = calldata;
620 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
621 int status = task->tk_status;
623 if (nfs_writeback_done(task, data) != 0)
626 spin_lock(&dreq->lock);
628 if (unlikely(status < 0)) {
629 /* An error has occurred, so we should not commit */
631 dreq->error = status;
633 if (unlikely(dreq->error != 0))
636 dreq->count += data->res.count;
638 if (data->res.verf->committed != NFS_FILE_SYNC) {
639 switch (dreq->flags) {
641 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
642 dreq->flags = NFS_ODIRECT_DO_COMMIT;
644 case NFS_ODIRECT_DO_COMMIT:
645 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
646 dprintk("NFS: %5u write verify failed\n", task->tk_pid);
647 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
652 spin_unlock(&dreq->lock);
656 * NB: Return the value of the first error return code. Subsequent
657 * errors after the first one are ignored.
659 static void nfs_direct_write_release(void *calldata)
661 struct nfs_write_data *data = calldata;
662 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
665 nfs_direct_write_complete(dreq, data->inode);
668 static const struct rpc_call_ops nfs_write_direct_ops = {
669 .rpc_call_done = nfs_direct_write_result,
670 .rpc_release = nfs_direct_write_release,
674 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
675 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
676 * bail and stop sending more writes. Write length accounting is
677 * handled automatically by nfs_direct_write_result(). Otherwise, if
678 * no requests have been sent, just return an error.
680 static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
681 const struct iovec *iov,
682 loff_t pos, int sync)
684 struct nfs_open_context *ctx = dreq->ctx;
685 struct inode *inode = ctx->path.dentry->d_inode;
686 unsigned long user_addr = (unsigned long)iov->iov_base;
687 size_t count = iov->iov_len;
688 struct rpc_task *task;
689 struct rpc_message msg = {
690 .rpc_cred = ctx->cred,
692 struct rpc_task_setup task_setup_data = {
693 .rpc_client = NFS_CLIENT(inode),
695 .callback_ops = &nfs_write_direct_ops,
696 .workqueue = nfsiod_workqueue,
697 .flags = RPC_TASK_ASYNC,
699 size_t wsize = NFS_SERVER(inode)->wsize;
705 struct nfs_write_data *data;
708 pgbase = user_addr & ~PAGE_MASK;
709 bytes = min(wsize,count);
712 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
716 down_read(¤t->mm->mmap_sem);
717 result = get_user_pages(current, current->mm, user_addr,
718 data->npages, 0, 0, data->pagevec, NULL);
719 up_read(¤t->mm->mmap_sem);
721 nfs_writedata_release(data);
724 if ((unsigned)result < data->npages) {
725 bytes = result * PAGE_SIZE;
726 if (bytes <= pgbase) {
727 nfs_direct_release_pages(data->pagevec, result);
728 nfs_writedata_release(data);
732 data->npages = result;
737 list_move_tail(&data->pages, &dreq->rewrite_list);
739 data->req = (struct nfs_page *) dreq;
741 data->cred = msg.rpc_cred;
742 data->args.fh = NFS_FH(inode);
743 data->args.context = get_nfs_open_context(ctx);
744 data->args.offset = pos;
745 data->args.pgbase = pgbase;
746 data->args.pages = data->pagevec;
747 data->args.count = bytes;
748 data->args.stable = sync;
749 data->res.fattr = &data->fattr;
750 data->res.count = bytes;
751 data->res.verf = &data->verf;
753 task_setup_data.task = &data->task;
754 task_setup_data.callback_data = data;
755 msg.rpc_argp = &data->args;
756 msg.rpc_resp = &data->res;
757 NFS_PROTO(inode)->write_setup(data, &msg);
759 task = rpc_run_task(&task_setup_data);
763 dprintk("NFS: %5u initiated direct write call "
764 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
767 (long long)NFS_FILEID(inode),
769 (unsigned long long)data->args.offset);
775 /* FIXME: Remove this useless math from the final patch */
777 pgbase &= ~PAGE_MASK;
778 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
781 } while (count != 0);
785 return result < 0 ? (ssize_t) result : -EFAULT;
788 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
789 const struct iovec *iov,
790 unsigned long nr_segs,
791 loff_t pos, int sync)
794 size_t requested_bytes = 0;
799 for (seg = 0; seg < nr_segs; seg++) {
800 const struct iovec *vec = &iov[seg];
801 result = nfs_direct_write_schedule_segment(dreq, vec,
805 requested_bytes += result;
806 if ((size_t)result < vec->iov_len)
812 nfs_direct_write_complete(dreq, dreq->inode);
814 if (requested_bytes != 0)
822 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
823 unsigned long nr_segs, loff_t pos,
827 struct inode *inode = iocb->ki_filp->f_mapping->host;
828 struct nfs_direct_req *dreq;
829 size_t wsize = NFS_SERVER(inode)->wsize;
830 int sync = NFS_UNSTABLE;
832 dreq = nfs_direct_req_alloc();
835 nfs_alloc_commit_data(dreq);
837 if (dreq->commit_data == NULL || count < wsize)
838 sync = NFS_FILE_SYNC;
841 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
842 if (!is_sync_kiocb(iocb))
845 result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
847 result = nfs_direct_wait(dreq);
848 nfs_direct_req_release(dreq);
854 * nfs_file_direct_read - file direct read operation for NFS files
855 * @iocb: target I/O control block
856 * @iov: vector of user buffers into which to read data
857 * @nr_segs: size of iov vector
858 * @pos: byte offset in file where reading starts
860 * We use this function for direct reads instead of calling
861 * generic_file_aio_read() in order to avoid gfar's check to see if
862 * the request starts before the end of the file. For that check
863 * to work, we must generate a GETATTR before each direct read, and
864 * even then there is a window between the GETATTR and the subsequent
865 * READ where the file size could change. Our preference is simply
866 * to do all reads the application wants, and the server will take
867 * care of managing the end of file boundary.
869 * This function also eliminates unnecessarily updating the file's
870 * atime locally, as the NFS server sets the file's atime, and this
871 * client must read the updated atime from the server back into its
874 ssize_t nfs_file_direct_read(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_DIRECTREADBYTES, count);
885 dprintk("nfs: direct read(%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);
894 retval = nfs_sync_mapping(mapping);
898 retval = nfs_direct_read(iocb, iov, nr_segs, pos);
900 iocb->ki_pos = pos + retval;
907 * nfs_file_direct_write - file direct write operation for NFS files
908 * @iocb: target I/O control block
909 * @iov: vector of user buffers from which to write data
910 * @nr_segs: size of iov vector
911 * @pos: byte offset in file where writing starts
913 * We use this function for direct writes instead of calling
914 * generic_file_aio_write() in order to avoid taking the inode
915 * semaphore and updating the i_size. The NFS server will set
916 * the new i_size and this client must read the updated size
917 * back into its cache. We let the server do generic write
918 * parameter checking and report problems.
920 * We also avoid an unnecessary invocation of generic_osync_inode(),
921 * as it is fairly meaningless to sync the metadata of an NFS file.
923 * We eliminate local atime updates, see direct read above.
925 * We avoid unnecessary page cache invalidations for normal cached
926 * readers of this file.
928 * Note that O_APPEND is not supported for NFS direct writes, as there
929 * is no atomic O_APPEND write facility in the NFS protocol.
931 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
932 unsigned long nr_segs, loff_t pos)
934 ssize_t retval = -EINVAL;
935 struct file *file = iocb->ki_filp;
936 struct address_space *mapping = file->f_mapping;
939 count = iov_length(iov, nr_segs);
940 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
942 dfprintk(VFS, "nfs: direct write(%s/%s, %zd@%Ld)\n",
943 file->f_path.dentry->d_parent->d_name.name,
944 file->f_path.dentry->d_name.name,
945 count, (long long) pos);
947 retval = generic_write_checks(file, &pos, &count, 0);
952 if ((ssize_t) count < 0)
958 retval = nfs_sync_mapping(mapping);
962 retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
965 iocb->ki_pos = pos + retval;
972 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
975 int __init nfs_init_directcache(void)
977 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
978 sizeof(struct nfs_direct_req),
979 0, (SLAB_RECLAIM_ACCOUNT|
982 if (nfs_direct_cachep == NULL)
989 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
992 void nfs_destroy_directcache(void)
994 kmem_cache_destroy(nfs_direct_cachep);