1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/fs/nfs/direct.c
5 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
7 * High-performance uncached I/O for the Linux NFS client
9 * There are important applications whose performance or correctness
10 * depends on uncached access to file data. Database clusters
11 * (multiple copies of the same instance running on separate hosts)
12 * implement their own cache coherency protocol that subsumes file
13 * system cache protocols. Applications that process datasets
14 * considerably larger than the client's memory do not always benefit
15 * from a local cache. A streaming video server, for instance, has no
16 * need to cache the contents of a file.
18 * When an application requests uncached I/O, all read and write requests
19 * are made directly to the server; data stored or fetched via these
20 * requests is not cached in the Linux page cache. The client does not
21 * correct unaligned requests from applications. All requested bytes are
22 * held on permanent storage before a direct write system call returns to
25 * Solaris implements an uncached I/O facility called directio() that
26 * is used for backups and sequential I/O to very large files. Solaris
27 * also supports uncaching whole NFS partitions with "-o forcedirectio,"
28 * an undocumented mount option.
30 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
31 * help from Andrew Morton.
33 * 18 Dec 2001 Initial implementation for 2.4 --cel
34 * 08 Jul 2002 Version for 2.4.19, with bug fixes --trondmy
35 * 08 Jun 2003 Port to 2.5 APIs --cel
36 * 31 Mar 2004 Handle direct I/O without VFS support --cel
37 * 15 Sep 2004 Parallel async reads --cel
38 * 04 May 2005 support O_DIRECT with aio --cel
42 #include <linux/errno.h>
43 #include <linux/sched.h>
44 #include <linux/kernel.h>
45 #include <linux/file.h>
46 #include <linux/pagemap.h>
47 #include <linux/kref.h>
48 #include <linux/slab.h>
49 #include <linux/task_io_accounting_ops.h>
50 #include <linux/module.h>
52 #include <linux/nfs_fs.h>
53 #include <linux/nfs_page.h>
54 #include <linux/sunrpc/clnt.h>
56 #include <linux/uaccess.h>
57 #include <linux/atomic.h>
63 #define NFSDBG_FACILITY NFSDBG_VFS
65 static struct kmem_cache *nfs_direct_cachep;
67 struct nfs_direct_req {
68 struct kref kref; /* release manager */
71 struct nfs_open_context *ctx; /* file open context info */
72 struct nfs_lock_context *l_ctx; /* Lock context info */
73 struct kiocb * iocb; /* controlling i/o request */
74 struct inode * inode; /* target file of i/o */
76 /* completion state */
77 atomic_t io_count; /* i/os we're waiting for */
78 spinlock_t lock; /* protect completion state */
80 loff_t io_start; /* Start offset for I/O */
81 ssize_t count, /* bytes actually processed */
82 max_count, /* max expected count */
83 bytes_left, /* bytes left to be sent */
84 error; /* any reported error */
85 struct completion completion; /* wait for i/o completion */
88 struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo */
89 struct pnfs_ds_commit_info ds_cinfo; /* Storage for cinfo */
90 struct work_struct work;
93 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
94 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
96 #define NFS_ODIRECT_SHOULD_DIRTY (3) /* dirty user-space page after read */
97 struct nfs_writeverf verf; /* unstable write verifier */
100 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
101 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
102 static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
103 static void nfs_direct_write_schedule_work(struct work_struct *work);
105 static inline void get_dreq(struct nfs_direct_req *dreq)
107 atomic_inc(&dreq->io_count);
110 static inline int put_dreq(struct nfs_direct_req *dreq)
112 return atomic_dec_and_test(&dreq->io_count);
116 nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
117 const struct nfs_pgio_header *hdr,
120 if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
121 test_bit(NFS_IOHDR_EOF, &hdr->flags)))
123 if (dreq->max_count >= dreq_len) {
124 dreq->max_count = dreq_len;
125 if (dreq->count > dreq_len)
126 dreq->count = dreq_len;
128 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
129 dreq->error = hdr->error;
130 else /* Clear outstanding error if this is EOF */
136 nfs_direct_count_bytes(struct nfs_direct_req *dreq,
137 const struct nfs_pgio_header *hdr)
139 loff_t hdr_end = hdr->io_start + hdr->good_bytes;
140 ssize_t dreq_len = 0;
142 if (hdr_end > dreq->io_start)
143 dreq_len = hdr_end - dreq->io_start;
145 nfs_direct_handle_truncated(dreq, hdr, dreq_len);
147 if (dreq_len > dreq->max_count)
148 dreq_len = dreq->max_count;
150 if (dreq->count < dreq_len)
151 dreq->count = dreq_len;
155 * nfs_direct_select_verf - select the right verifier
156 * @dreq - direct request possibly spanning multiple servers
157 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
158 * @commit_idx - commit bucket index for the DS
160 * returns the correct verifier to use given the role of the server
162 static struct nfs_writeverf *
163 nfs_direct_select_verf(struct nfs_direct_req *dreq,
164 struct nfs_client *ds_clp,
167 struct nfs_writeverf *verfp = &dreq->verf;
169 #ifdef CONFIG_NFS_V4_1
171 * pNFS is in use, use the DS verf except commit_through_mds is set
172 * for layout segment where nbuckets is zero.
174 if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
175 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
176 verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
186 * nfs_direct_set_hdr_verf - set the write/commit verifier
187 * @dreq - direct request possibly spanning multiple servers
188 * @hdr - pageio header to validate against previously seen verfs
190 * Set the server's (MDS or DS) "seen" verifier
192 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
193 struct nfs_pgio_header *hdr)
195 struct nfs_writeverf *verfp;
197 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
198 WARN_ON_ONCE(verfp->committed >= 0);
199 memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
200 WARN_ON_ONCE(verfp->committed < 0);
203 static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
204 const struct nfs_writeverf *v2)
206 return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
210 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
211 * @dreq - direct request possibly spanning multiple servers
212 * @hdr - pageio header to validate against previously seen verf
214 * set the server's "seen" verf if not initialized.
215 * returns result of comparison between @hdr->verf and the "seen"
216 * verf of the server used by @hdr (DS or MDS)
218 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
219 struct nfs_pgio_header *hdr)
221 struct nfs_writeverf *verfp;
223 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
224 if (verfp->committed < 0) {
225 nfs_direct_set_hdr_verf(dreq, hdr);
228 return nfs_direct_cmp_verf(verfp, &hdr->verf);
232 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
233 * @dreq - direct request possibly spanning multiple servers
234 * @data - commit data to validate against previously seen verf
236 * returns result of comparison between @data->verf and the verf of
237 * the server used by @data (DS or MDS)
239 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
240 struct nfs_commit_data *data)
242 struct nfs_writeverf *verfp;
244 verfp = nfs_direct_select_verf(dreq, data->ds_clp,
245 data->ds_commit_index);
247 /* verifier not set so always fail */
248 if (verfp->committed < 0 || data->res.verf->committed <= NFS_UNSTABLE)
251 return nfs_direct_cmp_verf(verfp, data->res.verf);
255 * nfs_direct_IO - NFS address space operation for direct I/O
256 * @iocb: target I/O control block
259 * The presence of this routine in the address space ops vector means
260 * the NFS client supports direct I/O. However, for most direct IO, we
261 * shunt off direct read and write requests before the VFS gets them,
262 * so this method is only ever called for swap.
264 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
266 struct inode *inode = iocb->ki_filp->f_mapping->host;
268 /* we only support swap file calling nfs_direct_IO */
269 if (!IS_SWAPFILE(inode))
272 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
274 if (iov_iter_rw(iter) == READ)
275 return nfs_file_direct_read(iocb, iter, true);
276 return nfs_file_direct_write(iocb, iter, true);
279 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
282 for (i = 0; i < npages; i++)
286 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
287 struct nfs_direct_req *dreq)
289 cinfo->inode = dreq->inode;
290 cinfo->mds = &dreq->mds_cinfo;
291 cinfo->ds = &dreq->ds_cinfo;
293 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
296 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
298 struct nfs_direct_req *dreq;
300 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
304 kref_init(&dreq->kref);
305 kref_get(&dreq->kref);
306 init_completion(&dreq->completion);
307 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
308 dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */
309 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
310 spin_lock_init(&dreq->lock);
315 static void nfs_direct_req_free(struct kref *kref)
317 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
319 nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
320 if (dreq->l_ctx != NULL)
321 nfs_put_lock_context(dreq->l_ctx);
322 if (dreq->ctx != NULL)
323 put_nfs_open_context(dreq->ctx);
324 kmem_cache_free(nfs_direct_cachep, dreq);
327 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
329 kref_put(&dreq->kref, nfs_direct_req_free);
332 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
334 return dreq->bytes_left;
336 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
339 * Collects and returns the final error value/byte-count.
341 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
343 ssize_t result = -EIOCBQUEUED;
345 /* Async requests don't wait here */
349 result = wait_for_completion_killable(&dreq->completion);
352 result = dreq->count;
353 WARN_ON_ONCE(dreq->count < 0);
356 result = dreq->error;
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)
368 struct inode *inode = dreq->inode;
370 inode_dio_end(inode);
373 long res = (long) dreq->error;
374 if (dreq->count != 0) {
375 res = (long) dreq->count;
376 WARN_ON_ONCE(dreq->count < 0);
378 dreq->iocb->ki_complete(dreq->iocb, res, 0);
381 complete(&dreq->completion);
383 nfs_direct_req_release(dreq);
386 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
388 unsigned long bytes = 0;
389 struct nfs_direct_req *dreq = hdr->dreq;
391 spin_lock(&dreq->lock);
392 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
393 spin_unlock(&dreq->lock);
397 nfs_direct_count_bytes(dreq, hdr);
398 spin_unlock(&dreq->lock);
400 while (!list_empty(&hdr->pages)) {
401 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
402 struct page *page = req->wb_page;
404 if (!PageCompound(page) && bytes < hdr->good_bytes &&
405 (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
406 set_page_dirty(page);
407 bytes += req->wb_bytes;
408 nfs_list_remove_request(req);
409 nfs_release_request(req);
413 nfs_direct_complete(dreq);
417 static void nfs_read_sync_pgio_error(struct list_head *head, int error)
419 struct nfs_page *req;
421 while (!list_empty(head)) {
422 req = nfs_list_entry(head->next);
423 nfs_list_remove_request(req);
424 nfs_release_request(req);
428 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
433 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
434 .error_cleanup = nfs_read_sync_pgio_error,
435 .init_hdr = nfs_direct_pgio_init,
436 .completion = nfs_direct_read_completion,
440 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
441 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
442 * bail and stop sending more reads. Read length accounting is
443 * handled automatically by nfs_direct_read_result(). Otherwise, if
444 * no requests have been sent, just return an error.
447 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
448 struct iov_iter *iter,
451 struct nfs_pageio_descriptor desc;
452 struct inode *inode = dreq->inode;
453 ssize_t result = -EINVAL;
454 size_t requested_bytes = 0;
455 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
457 nfs_pageio_init_read(&desc, dreq->inode, false,
458 &nfs_direct_read_completion_ops);
461 inode_dio_begin(inode);
463 while (iov_iter_count(iter)) {
464 struct page **pagevec;
469 result = iov_iter_get_pages_alloc(iter, &pagevec,
475 iov_iter_advance(iter, bytes);
476 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
477 for (i = 0; i < npages; i++) {
478 struct nfs_page *req;
479 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
480 /* XXX do we need to do the eof zeroing found in async_filler? */
481 req = nfs_create_request(dreq->ctx, pagevec[i],
484 result = PTR_ERR(req);
487 req->wb_index = pos >> PAGE_SHIFT;
488 req->wb_offset = pos & ~PAGE_MASK;
489 if (!nfs_pageio_add_request(&desc, req)) {
490 result = desc.pg_error;
491 nfs_release_request(req);
496 requested_bytes += req_len;
498 dreq->bytes_left -= req_len;
500 nfs_direct_release_pages(pagevec, npages);
506 nfs_pageio_complete(&desc);
509 * If no bytes were started, return the error, and let the
510 * generic layer handle the completion.
512 if (requested_bytes == 0) {
513 inode_dio_end(inode);
514 nfs_direct_req_release(dreq);
515 return result < 0 ? result : -EIO;
519 nfs_direct_complete(dreq);
520 return requested_bytes;
524 * nfs_file_direct_read - file direct read operation for NFS files
525 * @iocb: target I/O control block
526 * @iter: vector of user buffers into which to read data
527 * @swap: flag indicating this is swap IO, not O_DIRECT IO
529 * We use this function for direct reads instead of calling
530 * generic_file_aio_read() in order to avoid gfar's check to see if
531 * the request starts before the end of the file. For that check
532 * to work, we must generate a GETATTR before each direct read, and
533 * even then there is a window between the GETATTR and the subsequent
534 * READ where the file size could change. Our preference is simply
535 * to do all reads the application wants, and the server will take
536 * care of managing the end of file boundary.
538 * This function also eliminates unnecessarily updating the file's
539 * atime locally, as the NFS server sets the file's atime, and this
540 * client must read the updated atime from the server back into its
543 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
546 struct file *file = iocb->ki_filp;
547 struct address_space *mapping = file->f_mapping;
548 struct inode *inode = mapping->host;
549 struct nfs_direct_req *dreq;
550 struct nfs_lock_context *l_ctx;
551 ssize_t result = -EINVAL, requested;
552 size_t count = iov_iter_count(iter);
553 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
555 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
556 file, count, (long long) iocb->ki_pos);
562 task_io_account_read(count);
565 dreq = nfs_direct_req_alloc();
570 dreq->bytes_left = dreq->max_count = count;
571 dreq->io_start = iocb->ki_pos;
572 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
573 l_ctx = nfs_get_lock_context(dreq->ctx);
575 result = PTR_ERR(l_ctx);
576 nfs_direct_req_release(dreq);
580 if (!is_sync_kiocb(iocb))
583 if (iter_is_iovec(iter))
584 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
587 nfs_start_io_direct(inode);
589 NFS_I(inode)->read_io += count;
590 requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
593 nfs_end_io_direct(inode);
596 result = nfs_direct_wait(dreq);
599 iocb->ki_pos += result;
601 iov_iter_revert(iter, requested);
607 nfs_direct_req_release(dreq);
613 nfs_direct_write_scan_commit_list(struct inode *inode,
614 struct list_head *list,
615 struct nfs_commit_info *cinfo)
617 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
618 #ifdef CONFIG_NFS_V4_1
619 if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
620 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
622 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
623 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
626 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
628 struct nfs_pageio_descriptor desc;
629 struct nfs_page *req, *tmp;
631 struct nfs_commit_info cinfo;
634 nfs_init_cinfo_from_dreq(&cinfo, dreq);
635 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
639 list_for_each_entry(req, &reqs, wb_list)
640 dreq->max_count += req->wb_bytes;
641 dreq->verf.committed = NFS_INVALID_STABLE_HOW;
642 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
645 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
646 &nfs_direct_write_completion_ops);
649 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
650 /* Bump the transmission count */
652 if (!nfs_pageio_add_request(&desc, req)) {
653 nfs_list_move_request(req, &failed);
654 spin_lock(&cinfo.inode->i_lock);
656 if (desc.pg_error < 0)
657 dreq->error = desc.pg_error;
660 spin_unlock(&cinfo.inode->i_lock);
662 nfs_release_request(req);
664 nfs_pageio_complete(&desc);
666 while (!list_empty(&failed)) {
667 req = nfs_list_entry(failed.next);
668 nfs_list_remove_request(req);
669 nfs_unlock_and_release_request(req);
673 nfs_direct_write_complete(dreq);
676 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
678 struct nfs_direct_req *dreq = data->dreq;
679 struct nfs_commit_info cinfo;
680 struct nfs_page *req;
681 int status = data->task.tk_status;
683 nfs_init_cinfo_from_dreq(&cinfo, dreq);
684 if (status < 0 || nfs_direct_cmp_commit_data_verf(dreq, data))
685 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
687 while (!list_empty(&data->pages)) {
688 req = nfs_list_entry(data->pages.next);
689 nfs_list_remove_request(req);
690 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
692 * Despite the reboot, the write was successful,
696 /* Note the rewrite will go through mds */
697 nfs_mark_request_commit(req, NULL, &cinfo, 0);
699 nfs_release_request(req);
700 nfs_unlock_and_release_request(req);
703 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
704 nfs_direct_write_complete(dreq);
707 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
708 struct nfs_page *req)
710 struct nfs_direct_req *dreq = cinfo->dreq;
712 spin_lock(&dreq->lock);
713 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
714 spin_unlock(&dreq->lock);
715 nfs_mark_request_commit(req, NULL, cinfo, 0);
718 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
719 .completion = nfs_direct_commit_complete,
720 .resched_write = nfs_direct_resched_write,
723 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
726 struct nfs_commit_info cinfo;
729 nfs_init_cinfo_from_dreq(&cinfo, dreq);
730 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
731 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
732 if (res < 0) /* res == -ENOMEM */
733 nfs_direct_write_reschedule(dreq);
736 static void nfs_direct_write_schedule_work(struct work_struct *work)
738 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
739 int flags = dreq->flags;
743 case NFS_ODIRECT_DO_COMMIT:
744 nfs_direct_commit_schedule(dreq);
746 case NFS_ODIRECT_RESCHED_WRITES:
747 nfs_direct_write_reschedule(dreq);
750 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
751 nfs_direct_complete(dreq);
755 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
757 queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
760 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
762 struct nfs_direct_req *dreq = hdr->dreq;
763 struct nfs_commit_info cinfo;
764 bool request_commit = false;
765 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
767 nfs_init_cinfo_from_dreq(&cinfo, dreq);
769 spin_lock(&dreq->lock);
770 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
771 spin_unlock(&dreq->lock);
775 nfs_direct_count_bytes(dreq, hdr);
776 if (hdr->good_bytes != 0) {
777 if (nfs_write_need_commit(hdr)) {
778 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
779 request_commit = true;
780 else if (dreq->flags == 0) {
781 nfs_direct_set_hdr_verf(dreq, hdr);
782 request_commit = true;
783 dreq->flags = NFS_ODIRECT_DO_COMMIT;
784 } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
785 request_commit = true;
786 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
788 NFS_ODIRECT_RESCHED_WRITES;
792 spin_unlock(&dreq->lock);
794 while (!list_empty(&hdr->pages)) {
796 req = nfs_list_entry(hdr->pages.next);
797 nfs_list_remove_request(req);
798 if (request_commit) {
799 kref_get(&req->wb_kref);
800 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
803 nfs_unlock_and_release_request(req);
808 nfs_direct_write_complete(dreq);
812 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
814 struct nfs_page *req;
816 while (!list_empty(head)) {
817 req = nfs_list_entry(head->next);
818 nfs_list_remove_request(req);
819 nfs_unlock_and_release_request(req);
823 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
825 struct nfs_direct_req *dreq = hdr->dreq;
827 spin_lock(&dreq->lock);
828 if (dreq->error == 0) {
829 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
830 /* fake unstable write to let common nfs resend pages */
831 hdr->verf.committed = NFS_UNSTABLE;
832 hdr->good_bytes = hdr->args.count;
834 spin_unlock(&dreq->lock);
837 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
838 .error_cleanup = nfs_write_sync_pgio_error,
839 .init_hdr = nfs_direct_pgio_init,
840 .completion = nfs_direct_write_completion,
841 .reschedule_io = nfs_direct_write_reschedule_io,
846 * NB: Return the value of the first error return code. Subsequent
847 * errors after the first one are ignored.
850 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
851 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
852 * bail and stop sending more writes. Write length accounting is
853 * handled automatically by nfs_direct_write_result(). Otherwise, if
854 * no requests have been sent, just return an error.
856 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
857 struct iov_iter *iter,
858 loff_t pos, int ioflags)
860 struct nfs_pageio_descriptor desc;
861 struct inode *inode = dreq->inode;
863 size_t requested_bytes = 0;
864 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
866 nfs_pageio_init_write(&desc, inode, ioflags, false,
867 &nfs_direct_write_completion_ops);
870 inode_dio_begin(inode);
872 NFS_I(inode)->write_io += iov_iter_count(iter);
873 while (iov_iter_count(iter)) {
874 struct page **pagevec;
879 result = iov_iter_get_pages_alloc(iter, &pagevec,
885 iov_iter_advance(iter, bytes);
886 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
887 for (i = 0; i < npages; i++) {
888 struct nfs_page *req;
889 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
891 req = nfs_create_request(dreq->ctx, pagevec[i],
894 result = PTR_ERR(req);
898 if (desc.pg_error < 0) {
899 nfs_free_request(req);
900 result = desc.pg_error;
904 nfs_lock_request(req);
905 req->wb_index = pos >> PAGE_SHIFT;
906 req->wb_offset = pos & ~PAGE_MASK;
907 if (!nfs_pageio_add_request(&desc, req)) {
908 result = desc.pg_error;
909 nfs_unlock_and_release_request(req);
914 requested_bytes += req_len;
916 dreq->bytes_left -= req_len;
918 nfs_direct_release_pages(pagevec, npages);
923 nfs_pageio_complete(&desc);
926 * If no bytes were started, return the error, and let the
927 * generic layer handle the completion.
929 if (requested_bytes == 0) {
930 inode_dio_end(inode);
931 nfs_direct_req_release(dreq);
932 return result < 0 ? result : -EIO;
936 nfs_direct_write_complete(dreq);
937 return requested_bytes;
941 * nfs_file_direct_write - file direct write operation for NFS files
942 * @iocb: target I/O control block
943 * @iter: vector of user buffers from which to write data
944 * @swap: flag indicating this is swap IO, not O_DIRECT IO
946 * We use this function for direct writes instead of calling
947 * generic_file_aio_write() in order to avoid taking the inode
948 * semaphore and updating the i_size. The NFS server will set
949 * the new i_size and this client must read the updated size
950 * back into its cache. We let the server do generic write
951 * parameter checking and report problems.
953 * We eliminate local atime updates, see direct read above.
955 * We avoid unnecessary page cache invalidations for normal cached
956 * readers of this file.
958 * Note that O_APPEND is not supported for NFS direct writes, as there
959 * is no atomic O_APPEND write facility in the NFS protocol.
961 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
964 ssize_t result = -EINVAL, requested;
966 struct file *file = iocb->ki_filp;
967 struct address_space *mapping = file->f_mapping;
968 struct inode *inode = mapping->host;
969 struct nfs_direct_req *dreq;
970 struct nfs_lock_context *l_ctx;
973 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
974 file, iov_iter_count(iter), (long long) iocb->ki_pos);
977 /* bypass generic checks */
978 result = iov_iter_count(iter);
980 result = generic_write_checks(iocb, iter);
984 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
987 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
989 task_io_account_write(count);
992 dreq = nfs_direct_req_alloc();
997 dreq->bytes_left = dreq->max_count = count;
998 dreq->io_start = pos;
999 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1000 l_ctx = nfs_get_lock_context(dreq->ctx);
1001 if (IS_ERR(l_ctx)) {
1002 result = PTR_ERR(l_ctx);
1003 nfs_direct_req_release(dreq);
1006 dreq->l_ctx = l_ctx;
1007 if (!is_sync_kiocb(iocb))
1011 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1014 nfs_start_io_direct(inode);
1016 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1019 if (mapping->nrpages) {
1020 invalidate_inode_pages2_range(mapping,
1021 pos >> PAGE_SHIFT, end);
1024 nfs_end_io_direct(inode);
1027 if (requested > 0) {
1028 result = nfs_direct_wait(dreq);
1030 requested -= result;
1031 iocb->ki_pos = pos + result;
1032 /* XXX: should check the generic_write_sync retval */
1033 generic_write_sync(iocb, result);
1035 iov_iter_revert(iter, requested);
1040 nfs_direct_req_release(dreq);
1046 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1049 int __init nfs_init_directcache(void)
1051 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1052 sizeof(struct nfs_direct_req),
1053 0, (SLAB_RECLAIM_ACCOUNT|
1056 if (nfs_direct_cachep == NULL)
1063 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1066 void nfs_destroy_directcache(void)
1068 kmem_cache_destroy(nfs_direct_cachep);