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_req {
70 struct kref kref; /* release manager */
73 struct nfs_open_context *ctx; /* file open context info */
74 struct nfs_lock_context *l_ctx; /* Lock context info */
75 struct kiocb * iocb; /* controlling i/o request */
76 struct inode * inode; /* target file of i/o */
78 /* completion state */
79 atomic_t io_count; /* i/os we're waiting for */
80 spinlock_t lock; /* protect completion state */
81 ssize_t count, /* bytes actually processed */
82 bytes_left, /* bytes left to be sent */
83 error; /* any reported error */
84 struct completion completion; /* wait for i/o completion */
87 struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo */
88 struct pnfs_ds_commit_info ds_cinfo; /* Storage for cinfo */
89 struct work_struct work;
91 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
92 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
93 struct nfs_writeverf verf; /* unstable write verifier */
96 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
97 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
98 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
99 static void nfs_direct_write_schedule_work(struct work_struct *work);
101 static inline void get_dreq(struct nfs_direct_req *dreq)
103 atomic_inc(&dreq->io_count);
106 static inline int put_dreq(struct nfs_direct_req *dreq)
108 return atomic_dec_and_test(&dreq->io_count);
112 * nfs_direct_IO - NFS address space operation for direct I/O
113 * @rw: direction (read or write)
114 * @iocb: target I/O control block
115 * @iov: array of vectors that define I/O buffer
116 * @pos: offset in file to begin the operation
117 * @nr_segs: size of iovec array
119 * The presence of this routine in the address space ops vector means
120 * the NFS client supports direct I/O. However, for most direct IO, we
121 * shunt off direct read and write requests before the VFS gets them,
122 * so this method is only ever called for swap.
124 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
126 #ifndef CONFIG_NFS_SWAP
127 dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n",
128 iocb->ki_filp, (long long) pos, nr_segs);
132 VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
134 if (rw == READ || rw == KERNEL_READ)
135 return nfs_file_direct_read(iocb, iov, nr_segs, pos,
136 rw == READ ? true : false);
137 return nfs_file_direct_write(iocb, iov, nr_segs, pos,
138 rw == WRITE ? true : false);
139 #endif /* CONFIG_NFS_SWAP */
142 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
145 for (i = 0; i < npages; i++)
146 page_cache_release(pages[i]);
149 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
150 struct nfs_direct_req *dreq)
152 cinfo->lock = &dreq->lock;
153 cinfo->mds = &dreq->mds_cinfo;
154 cinfo->ds = &dreq->ds_cinfo;
156 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
159 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
161 struct nfs_direct_req *dreq;
163 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
167 kref_init(&dreq->kref);
168 kref_get(&dreq->kref);
169 init_completion(&dreq->completion);
170 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
171 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
172 spin_lock_init(&dreq->lock);
177 static void nfs_direct_req_free(struct kref *kref)
179 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
181 if (dreq->l_ctx != NULL)
182 nfs_put_lock_context(dreq->l_ctx);
183 if (dreq->ctx != NULL)
184 put_nfs_open_context(dreq->ctx);
185 kmem_cache_free(nfs_direct_cachep, dreq);
188 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
190 kref_put(&dreq->kref, nfs_direct_req_free);
193 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
195 return dreq->bytes_left;
197 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
200 * Collects and returns the final error value/byte-count.
202 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
204 ssize_t result = -EIOCBQUEUED;
206 /* Async requests don't wait here */
210 result = wait_for_completion_killable(&dreq->completion);
213 result = dreq->error;
215 result = dreq->count;
218 return (ssize_t) result;
222 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
223 * the iocb is still valid here if this is a synchronous request.
225 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
227 struct inode *inode = dreq->inode;
229 if (dreq->iocb && write) {
230 loff_t pos = dreq->iocb->ki_pos + dreq->count;
232 spin_lock(&inode->i_lock);
233 if (i_size_read(inode) < pos)
234 i_size_write(inode, pos);
235 spin_unlock(&inode->i_lock);
239 nfs_zap_mapping(inode, inode->i_mapping);
241 inode_dio_done(inode);
244 long res = (long) dreq->error;
246 res = (long) dreq->count;
247 aio_complete(dreq->iocb, res, 0);
250 complete_all(&dreq->completion);
252 nfs_direct_req_release(dreq);
255 static void nfs_direct_readpage_release(struct nfs_page *req)
257 dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
258 req->wb_context->dentry->d_inode->i_sb->s_id,
259 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
261 (long long)req_offset(req));
262 nfs_release_request(req);
265 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
267 unsigned long bytes = 0;
268 struct nfs_direct_req *dreq = hdr->dreq;
270 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
273 spin_lock(&dreq->lock);
274 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
275 dreq->error = hdr->error;
277 dreq->count += hdr->good_bytes;
278 spin_unlock(&dreq->lock);
280 while (!list_empty(&hdr->pages)) {
281 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
282 struct page *page = req->wb_page;
284 if (!PageCompound(page) && bytes < hdr->good_bytes)
285 set_page_dirty(page);
286 bytes += req->wb_bytes;
287 nfs_list_remove_request(req);
288 nfs_direct_readpage_release(req);
292 nfs_direct_complete(dreq, false);
296 static void nfs_read_sync_pgio_error(struct list_head *head)
298 struct nfs_page *req;
300 while (!list_empty(head)) {
301 req = nfs_list_entry(head->next);
302 nfs_list_remove_request(req);
303 nfs_release_request(req);
307 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
312 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
313 .error_cleanup = nfs_read_sync_pgio_error,
314 .init_hdr = nfs_direct_pgio_init,
315 .completion = nfs_direct_read_completion,
319 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
320 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
321 * bail and stop sending more reads. Read length accounting is
322 * handled automatically by nfs_direct_read_result(). Otherwise, if
323 * no requests have been sent, just return an error.
325 static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc,
326 const struct iovec *iov,
327 loff_t pos, bool uio)
329 struct nfs_direct_req *dreq = desc->pg_dreq;
330 struct nfs_open_context *ctx = dreq->ctx;
331 struct inode *inode = ctx->dentry->d_inode;
332 unsigned long user_addr = (unsigned long)iov->iov_base;
333 size_t count = iov->iov_len;
334 size_t rsize = NFS_SERVER(inode)->rsize;
338 struct page **pagevec = NULL;
345 pgbase = user_addr & ~PAGE_MASK;
346 bytes = min(max_t(size_t, rsize, PAGE_SIZE), count);
349 npages = nfs_page_array_len(pgbase, bytes);
351 pagevec = kmalloc(npages * sizeof(struct page *),
356 down_read(¤t->mm->mmap_sem);
357 result = get_user_pages(current, current->mm, user_addr,
358 npages, 1, 0, pagevec, NULL);
359 up_read(¤t->mm->mmap_sem);
363 WARN_ON(npages != 1);
364 result = get_kernel_page(user_addr, 1, pagevec);
365 if (WARN_ON(result != 1))
369 if ((unsigned)result < npages) {
370 bytes = result * PAGE_SIZE;
371 if (bytes <= pgbase) {
372 nfs_direct_release_pages(pagevec, result);
379 for (i = 0; i < npages; i++) {
380 struct nfs_page *req;
381 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
382 /* XXX do we need to do the eof zeroing found in async_filler? */
383 req = nfs_create_request(dreq->ctx, dreq->inode,
387 result = PTR_ERR(req);
390 req->wb_index = pos >> PAGE_SHIFT;
391 req->wb_offset = pos & ~PAGE_MASK;
392 if (!nfs_pageio_add_request(desc, req)) {
393 result = desc->pg_error;
394 nfs_release_request(req);
400 user_addr += req_len;
403 dreq->bytes_left -= req_len;
405 /* The nfs_page now hold references to these pages */
406 nfs_direct_release_pages(pagevec, npages);
407 } while (count != 0 && result >= 0);
413 return result < 0 ? (ssize_t) result : -EFAULT;
416 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
417 const struct iovec *iov,
418 unsigned long nr_segs,
419 loff_t pos, bool uio)
421 struct nfs_pageio_descriptor desc;
422 struct inode *inode = dreq->inode;
423 ssize_t result = -EINVAL;
424 size_t requested_bytes = 0;
427 NFS_PROTO(dreq->inode)->read_pageio_init(&desc, dreq->inode,
428 &nfs_direct_read_completion_ops);
431 atomic_inc(&inode->i_dio_count);
433 for (seg = 0; seg < nr_segs; seg++) {
434 const struct iovec *vec = &iov[seg];
435 result = nfs_direct_read_schedule_segment(&desc, vec, pos, uio);
438 requested_bytes += result;
439 if ((size_t)result < vec->iov_len)
444 nfs_pageio_complete(&desc);
447 * If no bytes were started, return the error, and let the
448 * generic layer handle the completion.
450 if (requested_bytes == 0) {
451 inode_dio_done(inode);
452 nfs_direct_req_release(dreq);
453 return result < 0 ? result : -EIO;
457 nfs_direct_complete(dreq, false);
461 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
462 unsigned long nr_segs, loff_t pos, bool uio)
464 ssize_t result = -ENOMEM;
465 struct inode *inode = iocb->ki_filp->f_mapping->host;
466 struct nfs_direct_req *dreq;
467 struct nfs_lock_context *l_ctx;
469 dreq = nfs_direct_req_alloc();
474 dreq->bytes_left = iov_length(iov, nr_segs);
475 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
476 l_ctx = nfs_get_lock_context(dreq->ctx);
478 result = PTR_ERR(l_ctx);
482 if (!is_sync_kiocb(iocb))
485 NFS_I(inode)->read_io += iov_length(iov, nr_segs);
486 result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos, uio);
488 result = nfs_direct_wait(dreq);
490 nfs_direct_req_release(dreq);
495 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
496 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
498 struct nfs_pageio_descriptor desc;
499 struct nfs_page *req, *tmp;
501 struct nfs_commit_info cinfo;
504 nfs_init_cinfo_from_dreq(&cinfo, dreq);
505 pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
506 spin_lock(cinfo.lock);
507 nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
508 spin_unlock(cinfo.lock);
513 NFS_PROTO(dreq->inode)->write_pageio_init(&desc, dreq->inode, FLUSH_STABLE,
514 &nfs_direct_write_completion_ops);
517 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
518 if (!nfs_pageio_add_request(&desc, req)) {
519 nfs_list_remove_request(req);
520 nfs_list_add_request(req, &failed);
521 spin_lock(cinfo.lock);
524 spin_unlock(cinfo.lock);
526 nfs_release_request(req);
528 nfs_pageio_complete(&desc);
530 while (!list_empty(&failed)) {
531 req = nfs_list_entry(failed.next);
532 nfs_list_remove_request(req);
533 nfs_unlock_and_release_request(req);
537 nfs_direct_write_complete(dreq, dreq->inode);
540 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
542 struct nfs_direct_req *dreq = data->dreq;
543 struct nfs_commit_info cinfo;
544 struct nfs_page *req;
545 int status = data->task.tk_status;
547 nfs_init_cinfo_from_dreq(&cinfo, dreq);
549 dprintk("NFS: %5u commit failed with error %d.\n",
550 data->task.tk_pid, status);
551 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
552 } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
553 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
554 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
557 dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
558 while (!list_empty(&data->pages)) {
559 req = nfs_list_entry(data->pages.next);
560 nfs_list_remove_request(req);
561 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
562 /* Note the rewrite will go through mds */
563 nfs_mark_request_commit(req, NULL, &cinfo);
565 nfs_release_request(req);
566 nfs_unlock_and_release_request(req);
569 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
570 nfs_direct_write_complete(dreq, data->inode);
573 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
575 /* There is no lock to clear */
578 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
579 .completion = nfs_direct_commit_complete,
580 .error_cleanup = nfs_direct_error_cleanup,
583 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
586 struct nfs_commit_info cinfo;
589 nfs_init_cinfo_from_dreq(&cinfo, dreq);
590 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
591 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
592 if (res < 0) /* res == -ENOMEM */
593 nfs_direct_write_reschedule(dreq);
596 static void nfs_direct_write_schedule_work(struct work_struct *work)
598 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
599 int flags = dreq->flags;
603 case NFS_ODIRECT_DO_COMMIT:
604 nfs_direct_commit_schedule(dreq);
606 case NFS_ODIRECT_RESCHED_WRITES:
607 nfs_direct_write_reschedule(dreq);
610 nfs_direct_complete(dreq, true);
614 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
616 schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
620 static void nfs_direct_write_schedule_work(struct work_struct *work)
624 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
626 nfs_direct_complete(dreq, true);
631 * NB: Return the value of the first error return code. Subsequent
632 * errors after the first one are ignored.
635 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
636 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
637 * bail and stop sending more writes. Write length accounting is
638 * handled automatically by nfs_direct_write_result(). Otherwise, if
639 * no requests have been sent, just return an error.
641 static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc,
642 const struct iovec *iov,
643 loff_t pos, bool uio)
645 struct nfs_direct_req *dreq = desc->pg_dreq;
646 struct nfs_open_context *ctx = dreq->ctx;
647 struct inode *inode = ctx->dentry->d_inode;
648 unsigned long user_addr = (unsigned long)iov->iov_base;
649 size_t count = iov->iov_len;
650 size_t wsize = NFS_SERVER(inode)->wsize;
654 struct page **pagevec = NULL;
661 pgbase = user_addr & ~PAGE_MASK;
662 bytes = min(max_t(size_t, wsize, PAGE_SIZE), count);
665 npages = nfs_page_array_len(pgbase, bytes);
667 pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
672 down_read(¤t->mm->mmap_sem);
673 result = get_user_pages(current, current->mm, user_addr,
674 npages, 0, 0, pagevec, NULL);
675 up_read(¤t->mm->mmap_sem);
679 WARN_ON(npages != 1);
680 result = get_kernel_page(user_addr, 0, pagevec);
681 if (WARN_ON(result != 1))
685 if ((unsigned)result < npages) {
686 bytes = result * PAGE_SIZE;
687 if (bytes <= pgbase) {
688 nfs_direct_release_pages(pagevec, result);
695 for (i = 0; i < npages; i++) {
696 struct nfs_page *req;
697 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
699 req = nfs_create_request(dreq->ctx, dreq->inode,
703 result = PTR_ERR(req);
706 nfs_lock_request(req);
707 req->wb_index = pos >> PAGE_SHIFT;
708 req->wb_offset = pos & ~PAGE_MASK;
709 if (!nfs_pageio_add_request(desc, req)) {
710 result = desc->pg_error;
711 nfs_unlock_and_release_request(req);
717 user_addr += req_len;
720 dreq->bytes_left -= req_len;
722 /* The nfs_page now hold references to these pages */
723 nfs_direct_release_pages(pagevec, npages);
724 } while (count != 0 && result >= 0);
730 return result < 0 ? (ssize_t) result : -EFAULT;
733 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
735 struct nfs_direct_req *dreq = hdr->dreq;
736 struct nfs_commit_info cinfo;
738 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
740 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
743 nfs_init_cinfo_from_dreq(&cinfo, dreq);
745 spin_lock(&dreq->lock);
747 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
749 dreq->error = hdr->error;
751 if (dreq->error != 0)
752 bit = NFS_IOHDR_ERROR;
754 dreq->count += hdr->good_bytes;
755 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
756 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
757 bit = NFS_IOHDR_NEED_RESCHED;
758 } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
759 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
760 bit = NFS_IOHDR_NEED_RESCHED;
761 else if (dreq->flags == 0) {
762 memcpy(&dreq->verf, hdr->verf,
764 bit = NFS_IOHDR_NEED_COMMIT;
765 dreq->flags = NFS_ODIRECT_DO_COMMIT;
766 } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
767 if (memcmp(&dreq->verf, hdr->verf, sizeof(dreq->verf))) {
768 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
769 bit = NFS_IOHDR_NEED_RESCHED;
771 bit = NFS_IOHDR_NEED_COMMIT;
775 spin_unlock(&dreq->lock);
777 while (!list_empty(&hdr->pages)) {
778 req = nfs_list_entry(hdr->pages.next);
779 nfs_list_remove_request(req);
781 case NFS_IOHDR_NEED_RESCHED:
782 case NFS_IOHDR_NEED_COMMIT:
783 kref_get(&req->wb_kref);
784 nfs_mark_request_commit(req, hdr->lseg, &cinfo);
786 nfs_unlock_and_release_request(req);
791 nfs_direct_write_complete(dreq, hdr->inode);
795 static void nfs_write_sync_pgio_error(struct list_head *head)
797 struct nfs_page *req;
799 while (!list_empty(head)) {
800 req = nfs_list_entry(head->next);
801 nfs_list_remove_request(req);
802 nfs_unlock_and_release_request(req);
806 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
807 .error_cleanup = nfs_write_sync_pgio_error,
808 .init_hdr = nfs_direct_pgio_init,
809 .completion = nfs_direct_write_completion,
812 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
813 const struct iovec *iov,
814 unsigned long nr_segs,
815 loff_t pos, bool uio)
817 struct nfs_pageio_descriptor desc;
818 struct inode *inode = dreq->inode;
820 size_t requested_bytes = 0;
823 NFS_PROTO(inode)->write_pageio_init(&desc, inode, FLUSH_COND_STABLE,
824 &nfs_direct_write_completion_ops);
827 atomic_inc(&inode->i_dio_count);
829 NFS_I(dreq->inode)->write_io += iov_length(iov, nr_segs);
830 for (seg = 0; seg < nr_segs; seg++) {
831 const struct iovec *vec = &iov[seg];
832 result = nfs_direct_write_schedule_segment(&desc, vec, pos, uio);
835 requested_bytes += result;
836 if ((size_t)result < vec->iov_len)
840 nfs_pageio_complete(&desc);
843 * If no bytes were started, return the error, and let the
844 * generic layer handle the completion.
846 if (requested_bytes == 0) {
847 inode_dio_done(inode);
848 nfs_direct_req_release(dreq);
849 return result < 0 ? result : -EIO;
853 nfs_direct_write_complete(dreq, dreq->inode);
857 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
858 unsigned long nr_segs, loff_t pos,
859 size_t count, bool uio)
861 ssize_t result = -ENOMEM;
862 struct inode *inode = iocb->ki_filp->f_mapping->host;
863 struct nfs_direct_req *dreq;
864 struct nfs_lock_context *l_ctx;
866 dreq = nfs_direct_req_alloc();
871 dreq->bytes_left = count;
872 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
873 l_ctx = nfs_get_lock_context(dreq->ctx);
875 result = PTR_ERR(l_ctx);
879 if (!is_sync_kiocb(iocb))
882 result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, uio);
884 result = nfs_direct_wait(dreq);
886 nfs_direct_req_release(dreq);
892 * nfs_file_direct_read - file direct read operation for NFS files
893 * @iocb: target I/O control block
894 * @iov: vector of user buffers into which to read data
895 * @nr_segs: size of iov vector
896 * @pos: byte offset in file where reading starts
898 * We use this function for direct reads instead of calling
899 * generic_file_aio_read() in order to avoid gfar's check to see if
900 * the request starts before the end of the file. For that check
901 * to work, we must generate a GETATTR before each direct read, and
902 * even then there is a window between the GETATTR and the subsequent
903 * READ where the file size could change. Our preference is simply
904 * to do all reads the application wants, and the server will take
905 * care of managing the end of file boundary.
907 * This function also eliminates unnecessarily updating the file's
908 * atime locally, as the NFS server sets the file's atime, and this
909 * client must read the updated atime from the server back into its
912 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
913 unsigned long nr_segs, loff_t pos, bool uio)
915 ssize_t retval = -EINVAL;
916 struct file *file = iocb->ki_filp;
917 struct address_space *mapping = file->f_mapping;
920 count = iov_length(iov, nr_segs);
921 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
923 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
924 file, count, (long long) pos);
930 retval = nfs_sync_mapping(mapping);
934 task_io_account_read(count);
936 retval = nfs_direct_read(iocb, iov, nr_segs, pos, uio);
938 iocb->ki_pos = pos + retval;
945 * nfs_file_direct_write - file direct write operation for NFS files
946 * @iocb: target I/O control block
947 * @iov: vector of user buffers from which to write data
948 * @nr_segs: size of iov vector
949 * @pos: byte offset in file where writing starts
951 * We use this function for direct writes instead of calling
952 * generic_file_aio_write() in order to avoid taking the inode
953 * semaphore and updating the i_size. The NFS server will set
954 * the new i_size and this client must read the updated size
955 * back into its cache. We let the server do generic write
956 * parameter checking and report problems.
958 * We eliminate local atime updates, see direct read above.
960 * We avoid unnecessary page cache invalidations for normal cached
961 * readers of this file.
963 * Note that O_APPEND is not supported for NFS direct writes, as there
964 * is no atomic O_APPEND write facility in the NFS protocol.
966 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
967 unsigned long nr_segs, loff_t pos, bool uio)
969 ssize_t retval = -EINVAL;
970 struct file *file = iocb->ki_filp;
971 struct address_space *mapping = file->f_mapping;
974 count = iov_length(iov, nr_segs);
975 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
977 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
978 file, count, (long long) pos);
980 retval = generic_write_checks(file, &pos, &count, 0);
985 if ((ssize_t) count < 0)
991 retval = nfs_sync_mapping(mapping);
995 task_io_account_write(count);
997 retval = nfs_direct_write(iocb, iov, nr_segs, pos, count, uio);
999 struct inode *inode = mapping->host;
1001 iocb->ki_pos = pos + retval;
1002 spin_lock(&inode->i_lock);
1003 if (i_size_read(inode) < iocb->ki_pos)
1004 i_size_write(inode, iocb->ki_pos);
1005 spin_unlock(&inode->i_lock);
1012 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1015 int __init nfs_init_directcache(void)
1017 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1018 sizeof(struct nfs_direct_req),
1019 0, (SLAB_RECLAIM_ACCOUNT|
1022 if (nfs_direct_cachep == NULL)
1029 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1032 void nfs_destroy_directcache(void)
1034 kmem_cache_destroy(nfs_direct_cachep);