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);
121 nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
126 if (dreq->mirror_count == 1) {
127 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
128 dreq->count += hdr->good_bytes;
130 /* mirrored writes */
131 count = dreq->mirrors[hdr->pgio_mirror_idx].count;
132 if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
133 count = hdr->io_start + hdr->good_bytes - dreq->io_start;
134 dreq->mirrors[hdr->pgio_mirror_idx].count = count;
136 /* update the dreq->count by finding the minimum agreed count from all
138 count = dreq->mirrors[0].count;
140 for (i = 1; i < dreq->mirror_count; i++)
141 count = min(count, dreq->mirrors[i].count);
148 * nfs_direct_select_verf - select the right verifier
149 * @dreq - direct request possibly spanning multiple servers
150 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
151 * @commit_idx - commit bucket index for the DS
153 * returns the correct verifier to use given the role of the server
155 static struct nfs_writeverf *
156 nfs_direct_select_verf(struct nfs_direct_req *dreq,
157 struct nfs_client *ds_clp,
160 struct nfs_writeverf *verfp = &dreq->verf;
162 #ifdef CONFIG_NFS_V4_1
164 * pNFS is in use, use the DS verf except commit_through_mds is set
165 * for layout segment where nbuckets is zero.
167 if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
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 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
263 if (iov_iter_rw(iter) == READ)
264 return nfs_file_direct_read(iocb, iter, pos);
265 return nfs_file_direct_write(iocb, iter);
268 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
271 for (i = 0; i < npages; i++)
272 page_cache_release(pages[i]);
275 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
276 struct nfs_direct_req *dreq)
278 cinfo->lock = &dreq->inode->i_lock;
279 cinfo->mds = &dreq->mds_cinfo;
280 cinfo->ds = &dreq->ds_cinfo;
282 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
285 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
286 struct nfs_pageio_descriptor *pgio,
287 struct nfs_page *req)
289 int mirror_count = 1;
291 if (pgio->pg_ops->pg_get_mirror_count)
292 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
294 dreq->mirror_count = mirror_count;
297 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
299 struct nfs_direct_req *dreq;
301 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
305 kref_init(&dreq->kref);
306 kref_get(&dreq->kref);
307 init_completion(&dreq->completion);
308 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
309 dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */
310 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
311 dreq->mirror_count = 1;
312 spin_lock_init(&dreq->lock);
317 static void nfs_direct_req_free(struct kref *kref)
319 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
321 nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
322 if (dreq->l_ctx != NULL)
323 nfs_put_lock_context(dreq->l_ctx);
324 if (dreq->ctx != NULL)
325 put_nfs_open_context(dreq->ctx);
326 kmem_cache_free(nfs_direct_cachep, dreq);
329 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
331 kref_put(&dreq->kref, nfs_direct_req_free);
334 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
336 return dreq->bytes_left;
338 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
341 * Collects and returns the final error value/byte-count.
343 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
345 ssize_t result = -EIOCBQUEUED;
347 /* Async requests don't wait here */
351 result = wait_for_completion_killable(&dreq->completion);
354 result = dreq->error;
356 result = dreq->count;
359 return (ssize_t) result;
363 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
364 * the iocb is still valid here if this is a synchronous request.
366 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
368 struct inode *inode = dreq->inode;
370 if (dreq->iocb && write) {
371 loff_t pos = dreq->iocb->ki_pos + dreq->count;
373 spin_lock(&inode->i_lock);
374 if (i_size_read(inode) < pos)
375 i_size_write(inode, pos);
376 spin_unlock(&inode->i_lock);
380 nfs_zap_mapping(inode, inode->i_mapping);
382 inode_dio_end(inode);
385 long res = (long) dreq->error;
387 res = (long) dreq->count;
388 dreq->iocb->ki_complete(dreq->iocb, res, 0);
391 complete_all(&dreq->completion);
393 nfs_direct_req_release(dreq);
396 static void nfs_direct_readpage_release(struct nfs_page *req)
398 dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
399 d_inode(req->wb_context->dentry)->i_sb->s_id,
400 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
402 (long long)req_offset(req));
403 nfs_release_request(req);
406 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
408 unsigned long bytes = 0;
409 struct nfs_direct_req *dreq = hdr->dreq;
411 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
414 spin_lock(&dreq->lock);
415 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
416 dreq->error = hdr->error;
418 nfs_direct_good_bytes(dreq, hdr);
420 spin_unlock(&dreq->lock);
422 while (!list_empty(&hdr->pages)) {
423 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
424 struct page *page = req->wb_page;
426 if (!PageCompound(page) && bytes < hdr->good_bytes)
427 set_page_dirty(page);
428 bytes += req->wb_bytes;
429 nfs_list_remove_request(req);
430 nfs_direct_readpage_release(req);
434 nfs_direct_complete(dreq, false);
438 static void nfs_read_sync_pgio_error(struct list_head *head)
440 struct nfs_page *req;
442 while (!list_empty(head)) {
443 req = nfs_list_entry(head->next);
444 nfs_list_remove_request(req);
445 nfs_release_request(req);
449 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
454 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
455 .error_cleanup = nfs_read_sync_pgio_error,
456 .init_hdr = nfs_direct_pgio_init,
457 .completion = nfs_direct_read_completion,
461 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
462 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
463 * bail and stop sending more reads. Read length accounting is
464 * handled automatically by nfs_direct_read_result(). Otherwise, if
465 * no requests have been sent, just return an error.
468 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
469 struct iov_iter *iter,
472 struct nfs_pageio_descriptor desc;
473 struct inode *inode = dreq->inode;
474 ssize_t result = -EINVAL;
475 size_t requested_bytes = 0;
476 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
478 nfs_pageio_init_read(&desc, dreq->inode, false,
479 &nfs_direct_read_completion_ops);
482 inode_dio_begin(inode);
484 while (iov_iter_count(iter)) {
485 struct page **pagevec;
490 result = iov_iter_get_pages_alloc(iter, &pagevec,
496 iov_iter_advance(iter, bytes);
497 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
498 for (i = 0; i < npages; i++) {
499 struct nfs_page *req;
500 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
501 /* XXX do we need to do the eof zeroing found in async_filler? */
502 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
505 result = PTR_ERR(req);
508 req->wb_index = pos >> PAGE_SHIFT;
509 req->wb_offset = pos & ~PAGE_MASK;
510 if (!nfs_pageio_add_request(&desc, req)) {
511 result = desc.pg_error;
512 nfs_release_request(req);
517 requested_bytes += req_len;
519 dreq->bytes_left -= req_len;
521 nfs_direct_release_pages(pagevec, npages);
527 nfs_pageio_complete(&desc);
530 * If no bytes were started, return the error, and let the
531 * generic layer handle the completion.
533 if (requested_bytes == 0) {
534 inode_dio_end(inode);
535 nfs_direct_req_release(dreq);
536 return result < 0 ? result : -EIO;
540 nfs_direct_complete(dreq, false);
545 * nfs_file_direct_read - file direct read operation for NFS files
546 * @iocb: target I/O control block
547 * @iter: vector of user buffers into which to read data
548 * @pos: byte offset in file where reading starts
550 * We use this function for direct reads instead of calling
551 * generic_file_aio_read() in order to avoid gfar's check to see if
552 * the request starts before the end of the file. For that check
553 * to work, we must generate a GETATTR before each direct read, and
554 * even then there is a window between the GETATTR and the subsequent
555 * READ where the file size could change. Our preference is simply
556 * to do all reads the application wants, and the server will take
557 * care of managing the end of file boundary.
559 * This function also eliminates unnecessarily updating the file's
560 * atime locally, as the NFS server sets the file's atime, and this
561 * client must read the updated atime from the server back into its
564 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
567 struct file *file = iocb->ki_filp;
568 struct address_space *mapping = file->f_mapping;
569 struct inode *inode = mapping->host;
570 struct nfs_direct_req *dreq;
571 struct nfs_lock_context *l_ctx;
572 ssize_t result = -EINVAL;
573 size_t count = iov_iter_count(iter);
574 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
576 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
577 file, count, (long long) pos);
583 mutex_lock(&inode->i_mutex);
584 result = nfs_sync_mapping(mapping);
588 task_io_account_read(count);
591 dreq = nfs_direct_req_alloc();
596 dreq->bytes_left = count;
597 dreq->io_start = pos;
598 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
599 l_ctx = nfs_get_lock_context(dreq->ctx);
601 result = PTR_ERR(l_ctx);
605 if (!is_sync_kiocb(iocb))
608 NFS_I(inode)->read_io += count;
609 result = nfs_direct_read_schedule_iovec(dreq, iter, pos);
611 mutex_unlock(&inode->i_mutex);
614 result = nfs_direct_wait(dreq);
616 iocb->ki_pos = pos + result;
619 nfs_direct_req_release(dreq);
623 nfs_direct_req_release(dreq);
625 mutex_unlock(&inode->i_mutex);
631 nfs_direct_write_scan_commit_list(struct inode *inode,
632 struct list_head *list,
633 struct nfs_commit_info *cinfo)
635 spin_lock(cinfo->lock);
636 #ifdef CONFIG_NFS_V4_1
637 if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
638 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
640 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
641 spin_unlock(cinfo->lock);
644 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
646 struct nfs_pageio_descriptor desc;
647 struct nfs_page *req, *tmp;
649 struct nfs_commit_info cinfo;
653 nfs_init_cinfo_from_dreq(&cinfo, dreq);
654 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
657 for (i = 0; i < dreq->mirror_count; i++)
658 dreq->mirrors[i].count = 0;
661 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
662 &nfs_direct_write_completion_ops);
665 req = nfs_list_entry(reqs.next);
666 nfs_direct_setup_mirroring(dreq, &desc, req);
668 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
669 if (!nfs_pageio_add_request(&desc, req)) {
670 nfs_list_remove_request(req);
671 nfs_list_add_request(req, &failed);
672 spin_lock(cinfo.lock);
675 spin_unlock(cinfo.lock);
677 nfs_release_request(req);
679 nfs_pageio_complete(&desc);
681 while (!list_empty(&failed)) {
682 req = nfs_list_entry(failed.next);
683 nfs_list_remove_request(req);
684 nfs_unlock_and_release_request(req);
688 nfs_direct_write_complete(dreq, dreq->inode);
691 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
693 struct nfs_direct_req *dreq = data->dreq;
694 struct nfs_commit_info cinfo;
695 struct nfs_page *req;
696 int status = data->task.tk_status;
698 nfs_init_cinfo_from_dreq(&cinfo, dreq);
700 dprintk("NFS: %5u commit failed with error %d.\n",
701 data->task.tk_pid, status);
702 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
703 } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
704 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
705 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
708 dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
709 while (!list_empty(&data->pages)) {
710 req = nfs_list_entry(data->pages.next);
711 nfs_list_remove_request(req);
712 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
713 /* Note the rewrite will go through mds */
714 nfs_mark_request_commit(req, NULL, &cinfo, 0);
716 nfs_release_request(req);
717 nfs_unlock_and_release_request(req);
720 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
721 nfs_direct_write_complete(dreq, data->inode);
724 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
725 .completion = nfs_direct_commit_complete,
728 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
731 struct nfs_commit_info cinfo;
734 nfs_init_cinfo_from_dreq(&cinfo, dreq);
735 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
736 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
737 if (res < 0) /* res == -ENOMEM */
738 nfs_direct_write_reschedule(dreq);
741 static void nfs_direct_write_schedule_work(struct work_struct *work)
743 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
744 int flags = dreq->flags;
748 case NFS_ODIRECT_DO_COMMIT:
749 nfs_direct_commit_schedule(dreq);
751 case NFS_ODIRECT_RESCHED_WRITES:
752 nfs_direct_write_reschedule(dreq);
755 nfs_direct_complete(dreq, true);
759 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
761 schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
764 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
766 struct nfs_direct_req *dreq = hdr->dreq;
767 struct nfs_commit_info cinfo;
768 bool request_commit = false;
769 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
771 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
774 nfs_init_cinfo_from_dreq(&cinfo, dreq);
776 spin_lock(&dreq->lock);
778 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
780 dreq->error = hdr->error;
782 if (dreq->error == 0) {
783 nfs_direct_good_bytes(dreq, hdr);
784 if (nfs_write_need_commit(hdr)) {
785 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
786 request_commit = true;
787 else if (dreq->flags == 0) {
788 nfs_direct_set_hdr_verf(dreq, hdr);
789 request_commit = true;
790 dreq->flags = NFS_ODIRECT_DO_COMMIT;
791 } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
792 request_commit = true;
793 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
795 NFS_ODIRECT_RESCHED_WRITES;
799 spin_unlock(&dreq->lock);
801 while (!list_empty(&hdr->pages)) {
803 req = nfs_list_entry(hdr->pages.next);
804 nfs_list_remove_request(req);
805 if (request_commit) {
806 kref_get(&req->wb_kref);
807 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
810 nfs_unlock_and_release_request(req);
815 nfs_direct_write_complete(dreq, hdr->inode);
819 static void nfs_write_sync_pgio_error(struct list_head *head)
821 struct nfs_page *req;
823 while (!list_empty(head)) {
824 req = nfs_list_entry(head->next);
825 nfs_list_remove_request(req);
826 nfs_unlock_and_release_request(req);
830 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
832 struct nfs_direct_req *dreq = hdr->dreq;
834 spin_lock(&dreq->lock);
835 if (dreq->error == 0) {
836 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
837 /* fake unstable write to let common nfs resend pages */
838 hdr->verf.committed = NFS_UNSTABLE;
839 hdr->good_bytes = hdr->args.count;
841 spin_unlock(&dreq->lock);
844 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
845 .error_cleanup = nfs_write_sync_pgio_error,
846 .init_hdr = nfs_direct_pgio_init,
847 .completion = nfs_direct_write_completion,
848 .reschedule_io = nfs_direct_write_reschedule_io,
853 * NB: Return the value of the first error return code. Subsequent
854 * errors after the first one are ignored.
857 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
858 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
859 * bail and stop sending more writes. Write length accounting is
860 * handled automatically by nfs_direct_write_result(). Otherwise, if
861 * no requests have been sent, just return an error.
863 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
864 struct iov_iter *iter,
867 struct nfs_pageio_descriptor desc;
868 struct inode *inode = dreq->inode;
870 size_t requested_bytes = 0;
871 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
873 nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
874 &nfs_direct_write_completion_ops);
877 inode_dio_begin(inode);
879 NFS_I(inode)->write_io += iov_iter_count(iter);
880 while (iov_iter_count(iter)) {
881 struct page **pagevec;
886 result = iov_iter_get_pages_alloc(iter, &pagevec,
892 iov_iter_advance(iter, bytes);
893 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
894 for (i = 0; i < npages; i++) {
895 struct nfs_page *req;
896 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
898 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
901 result = PTR_ERR(req);
905 nfs_direct_setup_mirroring(dreq, &desc, req);
907 nfs_lock_request(req);
908 req->wb_index = pos >> PAGE_SHIFT;
909 req->wb_offset = pos & ~PAGE_MASK;
910 if (!nfs_pageio_add_request(&desc, req)) {
911 result = desc.pg_error;
912 nfs_unlock_and_release_request(req);
917 requested_bytes += req_len;
919 dreq->bytes_left -= req_len;
921 nfs_direct_release_pages(pagevec, npages);
926 nfs_pageio_complete(&desc);
929 * If no bytes were started, return the error, and let the
930 * generic layer handle the completion.
932 if (requested_bytes == 0) {
933 inode_dio_end(inode);
934 nfs_direct_req_release(dreq);
935 return result < 0 ? result : -EIO;
939 nfs_direct_write_complete(dreq, dreq->inode);
944 * nfs_file_direct_write - file direct write operation for NFS files
945 * @iocb: target I/O control block
946 * @iter: vector of user buffers from which to write data
947 * @pos: byte offset in file where writing starts
949 * We use this function for direct writes instead of calling
950 * generic_file_aio_write() in order to avoid taking the inode
951 * semaphore and updating the i_size. The NFS server will set
952 * the new i_size and this client must read the updated size
953 * back into its cache. We let the server do generic write
954 * parameter checking and report problems.
956 * We eliminate local atime updates, see direct read above.
958 * We avoid unnecessary page cache invalidations for normal cached
959 * readers of this file.
961 * Note that O_APPEND is not supported for NFS direct writes, as there
962 * is no atomic O_APPEND write facility in the NFS protocol.
964 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
966 ssize_t result = -EINVAL;
967 struct file *file = iocb->ki_filp;
968 struct address_space *mapping = file->f_mapping;
969 struct inode *inode = mapping->host;
970 struct nfs_direct_req *dreq;
971 struct nfs_lock_context *l_ctx;
974 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
975 file, iov_iter_count(iter), (long long) iocb->ki_pos);
977 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES,
978 iov_iter_count(iter));
981 end = (pos + iov_iter_count(iter) - 1) >> PAGE_CACHE_SHIFT;
983 mutex_lock(&inode->i_mutex);
985 result = nfs_sync_mapping(mapping);
989 if (mapping->nrpages) {
990 result = invalidate_inode_pages2_range(mapping,
991 pos >> PAGE_CACHE_SHIFT, end);
996 task_io_account_write(iov_iter_count(iter));
999 dreq = nfs_direct_req_alloc();
1003 dreq->inode = inode;
1004 dreq->bytes_left = iov_iter_count(iter);
1005 dreq->io_start = pos;
1006 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1007 l_ctx = nfs_get_lock_context(dreq->ctx);
1008 if (IS_ERR(l_ctx)) {
1009 result = PTR_ERR(l_ctx);
1012 dreq->l_ctx = l_ctx;
1013 if (!is_sync_kiocb(iocb))
1016 result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1018 if (mapping->nrpages) {
1019 invalidate_inode_pages2_range(mapping,
1020 pos >> PAGE_CACHE_SHIFT, end);
1023 mutex_unlock(&inode->i_mutex);
1026 result = nfs_direct_wait(dreq);
1028 struct inode *inode = mapping->host;
1030 iocb->ki_pos = pos + result;
1031 spin_lock(&inode->i_lock);
1032 if (i_size_read(inode) < iocb->ki_pos)
1033 i_size_write(inode, iocb->ki_pos);
1034 spin_unlock(&inode->i_lock);
1035 generic_write_sync(file, pos, result);
1038 nfs_direct_req_release(dreq);
1042 nfs_direct_req_release(dreq);
1044 mutex_unlock(&inode->i_mutex);
1049 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1052 int __init nfs_init_directcache(void)
1054 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1055 sizeof(struct nfs_direct_req),
1056 0, (SLAB_RECLAIM_ACCOUNT|
1059 if (nfs_direct_cachep == NULL)
1066 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1069 void nfs_destroy_directcache(void)
1071 kmem_cache_destroy(nfs_direct_cachep);