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 max_count, /* max expected count */
91 bytes_left, /* bytes left to be sent */
92 io_start, /* start of IO */
93 error; /* any reported error */
94 struct completion completion; /* wait for i/o completion */
97 struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo */
98 struct pnfs_ds_commit_info ds_cinfo; /* Storage for cinfo */
99 struct work_struct work;
101 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
102 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
103 struct nfs_writeverf verf; /* unstable write verifier */
106 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
107 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
108 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
109 static void nfs_direct_write_schedule_work(struct work_struct *work);
111 static inline void get_dreq(struct nfs_direct_req *dreq)
113 atomic_inc(&dreq->io_count);
116 static inline int put_dreq(struct nfs_direct_req *dreq)
118 return atomic_dec_and_test(&dreq->io_count);
122 nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
127 WARN_ON_ONCE(dreq->count >= dreq->max_count);
129 if (dreq->mirror_count == 1) {
130 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
131 dreq->count += hdr->good_bytes;
133 /* mirrored writes */
134 count = dreq->mirrors[hdr->pgio_mirror_idx].count;
135 if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
136 count = hdr->io_start + hdr->good_bytes - dreq->io_start;
137 dreq->mirrors[hdr->pgio_mirror_idx].count = count;
139 /* update the dreq->count by finding the minimum agreed count from all
141 count = dreq->mirrors[0].count;
143 for (i = 1; i < dreq->mirror_count; i++)
144 count = min(count, dreq->mirrors[i].count);
151 * nfs_direct_select_verf - select the right verifier
152 * @dreq - direct request possibly spanning multiple servers
153 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
154 * @commit_idx - commit bucket index for the DS
156 * returns the correct verifier to use given the role of the server
158 static struct nfs_writeverf *
159 nfs_direct_select_verf(struct nfs_direct_req *dreq,
160 struct nfs_client *ds_clp,
163 struct nfs_writeverf *verfp = &dreq->verf;
165 #ifdef CONFIG_NFS_V4_1
167 * pNFS is in use, use the DS verf except commit_through_mds is set
168 * for layout segment where nbuckets is zero.
170 if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
171 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
172 verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
182 * nfs_direct_set_hdr_verf - set the write/commit verifier
183 * @dreq - direct request possibly spanning multiple servers
184 * @hdr - pageio header to validate against previously seen verfs
186 * Set the server's (MDS or DS) "seen" verifier
188 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
189 struct nfs_pgio_header *hdr)
191 struct nfs_writeverf *verfp;
193 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
194 WARN_ON_ONCE(verfp->committed >= 0);
195 memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
196 WARN_ON_ONCE(verfp->committed < 0);
199 static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
200 const struct nfs_writeverf *v2)
202 return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
206 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
207 * @dreq - direct request possibly spanning multiple servers
208 * @hdr - pageio header to validate against previously seen verf
210 * set the server's "seen" verf if not initialized.
211 * returns result of comparison between @hdr->verf and the "seen"
212 * verf of the server used by @hdr (DS or MDS)
214 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
215 struct nfs_pgio_header *hdr)
217 struct nfs_writeverf *verfp;
219 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
220 if (verfp->committed < 0) {
221 nfs_direct_set_hdr_verf(dreq, hdr);
224 return nfs_direct_cmp_verf(verfp, &hdr->verf);
228 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
229 * @dreq - direct request possibly spanning multiple servers
230 * @data - commit data to validate against previously seen verf
232 * returns result of comparison between @data->verf and the verf of
233 * the server used by @data (DS or MDS)
235 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
236 struct nfs_commit_data *data)
238 struct nfs_writeverf *verfp;
240 verfp = nfs_direct_select_verf(dreq, data->ds_clp,
241 data->ds_commit_index);
243 /* verifier not set so always fail */
244 if (verfp->committed < 0)
247 return nfs_direct_cmp_verf(verfp, &data->verf);
251 * nfs_direct_IO - NFS address space operation for direct I/O
252 * @iocb: target I/O control block
253 * @iov: array of vectors that define I/O buffer
254 * @pos: offset in file to begin the operation
255 * @nr_segs: size of iovec array
257 * The presence of this routine in the address space ops vector means
258 * the NFS client supports direct I/O. However, for most direct IO, we
259 * shunt off direct read and write requests before the VFS gets them,
260 * so this method is only ever called for swap.
262 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
264 struct inode *inode = iocb->ki_filp->f_mapping->host;
266 /* we only support swap file calling nfs_direct_IO */
267 if (!IS_SWAPFILE(inode))
270 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
272 if (iov_iter_rw(iter) == READ)
273 return nfs_file_direct_read(iocb, iter);
274 return nfs_file_direct_write(iocb, iter);
277 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
280 for (i = 0; i < npages; i++)
284 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
285 struct nfs_direct_req *dreq)
287 cinfo->inode = dreq->inode;
288 cinfo->mds = &dreq->mds_cinfo;
289 cinfo->ds = &dreq->ds_cinfo;
291 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
294 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
295 struct nfs_pageio_descriptor *pgio,
296 struct nfs_page *req)
298 int mirror_count = 1;
300 if (pgio->pg_ops->pg_get_mirror_count)
301 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
303 dreq->mirror_count = mirror_count;
306 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
308 struct nfs_direct_req *dreq;
310 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
314 kref_init(&dreq->kref);
315 kref_get(&dreq->kref);
316 init_completion(&dreq->completion);
317 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
318 dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */
319 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
320 dreq->mirror_count = 1;
321 spin_lock_init(&dreq->lock);
326 static void nfs_direct_req_free(struct kref *kref)
328 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
330 nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
331 if (dreq->l_ctx != NULL)
332 nfs_put_lock_context(dreq->l_ctx);
333 if (dreq->ctx != NULL)
334 put_nfs_open_context(dreq->ctx);
335 kmem_cache_free(nfs_direct_cachep, dreq);
338 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
340 kref_put(&dreq->kref, nfs_direct_req_free);
343 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
345 return dreq->bytes_left;
347 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
350 * Collects and returns the final error value/byte-count.
352 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
354 ssize_t result = -EIOCBQUEUED;
356 /* Async requests don't wait here */
360 result = wait_for_completion_killable(&dreq->completion);
363 result = dreq->count;
364 WARN_ON_ONCE(dreq->count < 0);
367 result = dreq->error;
370 return (ssize_t) result;
374 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
375 * the iocb is still valid here if this is a synchronous request.
377 static void nfs_direct_complete(struct nfs_direct_req *dreq)
379 struct inode *inode = dreq->inode;
381 inode_dio_end(inode);
384 long res = (long) dreq->error;
385 if (dreq->count != 0) {
386 res = (long) dreq->count;
387 WARN_ON_ONCE(dreq->count < 0);
389 dreq->iocb->ki_complete(dreq->iocb, res, 0);
392 complete_all(&dreq->completion);
394 nfs_direct_req_release(dreq);
397 static void nfs_direct_readpage_release(struct nfs_page *req)
399 dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
400 req->wb_context->dentry->d_sb->s_id,
401 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
403 (long long)req_offset(req));
404 nfs_release_request(req);
407 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
409 unsigned long bytes = 0;
410 struct nfs_direct_req *dreq = hdr->dreq;
412 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
415 spin_lock(&dreq->lock);
416 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
417 dreq->error = hdr->error;
419 nfs_direct_good_bytes(dreq, hdr);
421 spin_unlock(&dreq->lock);
423 while (!list_empty(&hdr->pages)) {
424 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
425 struct page *page = req->wb_page;
427 if (!PageCompound(page) && bytes < hdr->good_bytes)
428 set_page_dirty(page);
429 bytes += req->wb_bytes;
430 nfs_list_remove_request(req);
431 nfs_direct_readpage_release(req);
435 nfs_direct_complete(dreq);
439 static void nfs_read_sync_pgio_error(struct list_head *head)
441 struct nfs_page *req;
443 while (!list_empty(head)) {
444 req = nfs_list_entry(head->next);
445 nfs_list_remove_request(req);
446 nfs_release_request(req);
450 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
455 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
456 .error_cleanup = nfs_read_sync_pgio_error,
457 .init_hdr = nfs_direct_pgio_init,
458 .completion = nfs_direct_read_completion,
462 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
463 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
464 * bail and stop sending more reads. Read length accounting is
465 * handled automatically by nfs_direct_read_result(). Otherwise, if
466 * no requests have been sent, just return an error.
469 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
470 struct iov_iter *iter,
473 struct nfs_pageio_descriptor desc;
474 struct inode *inode = dreq->inode;
475 ssize_t result = -EINVAL;
476 size_t requested_bytes = 0;
477 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
479 nfs_pageio_init_read(&desc, dreq->inode, false,
480 &nfs_direct_read_completion_ops);
483 inode_dio_begin(inode);
485 while (iov_iter_count(iter)) {
486 struct page **pagevec;
491 result = iov_iter_get_pages_alloc(iter, &pagevec,
497 iov_iter_advance(iter, bytes);
498 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
499 for (i = 0; i < npages; i++) {
500 struct nfs_page *req;
501 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
502 /* XXX do we need to do the eof zeroing found in async_filler? */
503 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
506 result = PTR_ERR(req);
509 req->wb_index = pos >> PAGE_SHIFT;
510 req->wb_offset = pos & ~PAGE_MASK;
511 if (!nfs_pageio_add_request(&desc, req)) {
512 result = desc.pg_error;
513 nfs_release_request(req);
518 requested_bytes += req_len;
520 dreq->bytes_left -= req_len;
522 nfs_direct_release_pages(pagevec, npages);
528 nfs_pageio_complete(&desc);
531 * If no bytes were started, return the error, and let the
532 * generic layer handle the completion.
534 if (requested_bytes == 0) {
535 inode_dio_end(inode);
536 nfs_direct_req_release(dreq);
537 return result < 0 ? result : -EIO;
541 nfs_direct_complete(dreq);
546 * nfs_file_direct_read - file direct read operation for NFS files
547 * @iocb: target I/O control block
548 * @iter: vector of user buffers into which to read data
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)
566 struct file *file = iocb->ki_filp;
567 struct address_space *mapping = file->f_mapping;
568 struct inode *inode = mapping->host;
569 struct nfs_direct_req *dreq;
570 struct nfs_lock_context *l_ctx;
571 ssize_t result = -EINVAL;
572 size_t count = iov_iter_count(iter);
573 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
575 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
576 file, count, (long long) iocb->ki_pos);
582 task_io_account_read(count);
585 dreq = nfs_direct_req_alloc();
590 dreq->bytes_left = dreq->max_count = count;
591 dreq->io_start = iocb->ki_pos;
592 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
593 l_ctx = nfs_get_lock_context(dreq->ctx);
595 result = PTR_ERR(l_ctx);
599 if (!is_sync_kiocb(iocb))
602 nfs_start_io_direct(inode);
604 NFS_I(inode)->read_io += count;
605 result = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
607 nfs_end_io_direct(inode);
610 result = nfs_direct_wait(dreq);
612 iocb->ki_pos += result;
616 nfs_direct_req_release(dreq);
622 nfs_direct_write_scan_commit_list(struct inode *inode,
623 struct list_head *list,
624 struct nfs_commit_info *cinfo)
626 spin_lock(&cinfo->inode->i_lock);
627 #ifdef CONFIG_NFS_V4_1
628 if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
629 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
631 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
632 spin_unlock(&cinfo->inode->i_lock);
635 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
637 struct nfs_pageio_descriptor desc;
638 struct nfs_page *req, *tmp;
640 struct nfs_commit_info cinfo;
644 nfs_init_cinfo_from_dreq(&cinfo, dreq);
645 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
648 dreq->verf.committed = NFS_INVALID_STABLE_HOW;
649 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
650 for (i = 0; i < dreq->mirror_count; i++)
651 dreq->mirrors[i].count = 0;
654 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
655 &nfs_direct_write_completion_ops);
658 req = nfs_list_entry(reqs.next);
659 nfs_direct_setup_mirroring(dreq, &desc, req);
660 if (desc.pg_error < 0) {
661 list_splice_init(&reqs, &failed);
665 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
666 if (!nfs_pageio_add_request(&desc, req)) {
667 nfs_list_remove_request(req);
668 nfs_list_add_request(req, &failed);
669 spin_lock(&cinfo.inode->i_lock);
671 if (desc.pg_error < 0)
672 dreq->error = desc.pg_error;
675 spin_unlock(&cinfo.inode->i_lock);
677 nfs_release_request(req);
679 nfs_pageio_complete(&desc);
682 while (!list_empty(&failed)) {
683 req = nfs_list_entry(failed.next);
684 nfs_list_remove_request(req);
685 nfs_unlock_and_release_request(req);
689 nfs_direct_write_complete(dreq, dreq->inode);
692 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
694 struct nfs_direct_req *dreq = data->dreq;
695 struct nfs_commit_info cinfo;
696 struct nfs_page *req;
697 int status = data->task.tk_status;
699 nfs_init_cinfo_from_dreq(&cinfo, dreq);
701 dprintk("NFS: %5u commit failed with error %d.\n",
702 data->task.tk_pid, status);
703 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
704 } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
705 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
706 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
709 dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
710 while (!list_empty(&data->pages)) {
711 req = nfs_list_entry(data->pages.next);
712 nfs_list_remove_request(req);
713 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
714 /* Note the rewrite will go through mds */
715 nfs_mark_request_commit(req, NULL, &cinfo, 0);
717 nfs_release_request(req);
718 nfs_unlock_and_release_request(req);
721 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
722 nfs_direct_write_complete(dreq, data->inode);
725 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
726 struct nfs_page *req)
728 struct nfs_direct_req *dreq = cinfo->dreq;
730 spin_lock(&dreq->lock);
731 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
732 spin_unlock(&dreq->lock);
733 nfs_mark_request_commit(req, NULL, cinfo, 0);
736 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
737 .completion = nfs_direct_commit_complete,
738 .resched_write = nfs_direct_resched_write,
741 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
744 struct nfs_commit_info cinfo;
747 nfs_init_cinfo_from_dreq(&cinfo, dreq);
748 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
749 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
750 if (res < 0) /* res == -ENOMEM */
751 nfs_direct_write_reschedule(dreq);
754 static void nfs_direct_write_schedule_work(struct work_struct *work)
756 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
757 int flags = dreq->flags;
761 case NFS_ODIRECT_DO_COMMIT:
762 nfs_direct_commit_schedule(dreq);
764 case NFS_ODIRECT_RESCHED_WRITES:
765 nfs_direct_write_reschedule(dreq);
768 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
769 nfs_direct_complete(dreq);
773 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
775 schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
778 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
780 struct nfs_direct_req *dreq = hdr->dreq;
781 struct nfs_commit_info cinfo;
782 bool request_commit = false;
783 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
785 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
788 nfs_init_cinfo_from_dreq(&cinfo, dreq);
790 spin_lock(&dreq->lock);
792 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
794 dreq->error = hdr->error;
796 if (dreq->error == 0) {
797 nfs_direct_good_bytes(dreq, hdr);
798 if (nfs_write_need_commit(hdr)) {
799 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
800 request_commit = true;
801 else if (dreq->flags == 0) {
802 nfs_direct_set_hdr_verf(dreq, hdr);
803 request_commit = true;
804 dreq->flags = NFS_ODIRECT_DO_COMMIT;
805 } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
806 request_commit = true;
807 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
809 NFS_ODIRECT_RESCHED_WRITES;
813 spin_unlock(&dreq->lock);
815 while (!list_empty(&hdr->pages)) {
817 req = nfs_list_entry(hdr->pages.next);
818 nfs_list_remove_request(req);
819 if (request_commit) {
820 kref_get(&req->wb_kref);
821 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
824 nfs_unlock_and_release_request(req);
829 nfs_direct_write_complete(dreq, hdr->inode);
833 static void nfs_write_sync_pgio_error(struct list_head *head)
835 struct nfs_page *req;
837 while (!list_empty(head)) {
838 req = nfs_list_entry(head->next);
839 nfs_list_remove_request(req);
840 nfs_unlock_and_release_request(req);
844 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
846 struct nfs_direct_req *dreq = hdr->dreq;
848 spin_lock(&dreq->lock);
849 if (dreq->error == 0) {
850 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
851 /* fake unstable write to let common nfs resend pages */
852 hdr->verf.committed = NFS_UNSTABLE;
853 hdr->good_bytes = hdr->args.count;
855 spin_unlock(&dreq->lock);
858 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
859 .error_cleanup = nfs_write_sync_pgio_error,
860 .init_hdr = nfs_direct_pgio_init,
861 .completion = nfs_direct_write_completion,
862 .reschedule_io = nfs_direct_write_reschedule_io,
867 * NB: Return the value of the first error return code. Subsequent
868 * errors after the first one are ignored.
871 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
872 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
873 * bail and stop sending more writes. Write length accounting is
874 * handled automatically by nfs_direct_write_result(). Otherwise, if
875 * no requests have been sent, just return an error.
877 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
878 struct iov_iter *iter,
881 struct nfs_pageio_descriptor desc;
882 struct inode *inode = dreq->inode;
884 size_t requested_bytes = 0;
885 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
887 nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
888 &nfs_direct_write_completion_ops);
891 inode_dio_begin(inode);
893 NFS_I(inode)->write_io += iov_iter_count(iter);
894 while (iov_iter_count(iter)) {
895 struct page **pagevec;
900 result = iov_iter_get_pages_alloc(iter, &pagevec,
906 iov_iter_advance(iter, bytes);
907 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
908 for (i = 0; i < npages; i++) {
909 struct nfs_page *req;
910 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
912 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
915 result = PTR_ERR(req);
919 nfs_direct_setup_mirroring(dreq, &desc, req);
920 if (desc.pg_error < 0) {
921 nfs_free_request(req);
922 result = desc.pg_error;
926 nfs_lock_request(req);
927 req->wb_index = pos >> PAGE_SHIFT;
928 req->wb_offset = pos & ~PAGE_MASK;
929 if (!nfs_pageio_add_request(&desc, req)) {
930 result = desc.pg_error;
931 nfs_unlock_and_release_request(req);
936 requested_bytes += req_len;
938 dreq->bytes_left -= req_len;
940 nfs_direct_release_pages(pagevec, npages);
945 nfs_pageio_complete(&desc);
948 * If no bytes were started, return the error, and let the
949 * generic layer handle the completion.
951 if (requested_bytes == 0) {
952 inode_dio_end(inode);
953 nfs_direct_req_release(dreq);
954 return result < 0 ? result : -EIO;
958 nfs_direct_write_complete(dreq, dreq->inode);
963 * nfs_file_direct_write - file direct write operation for NFS files
964 * @iocb: target I/O control block
965 * @iter: vector of user buffers from which to write data
967 * We use this function for direct writes instead of calling
968 * generic_file_aio_write() in order to avoid taking the inode
969 * semaphore and updating the i_size. The NFS server will set
970 * the new i_size and this client must read the updated size
971 * back into its cache. We let the server do generic write
972 * parameter checking and report problems.
974 * We eliminate local atime updates, see direct read above.
976 * We avoid unnecessary page cache invalidations for normal cached
977 * readers of this file.
979 * Note that O_APPEND is not supported for NFS direct writes, as there
980 * is no atomic O_APPEND write facility in the NFS protocol.
982 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
984 ssize_t result = -EINVAL;
986 struct file *file = iocb->ki_filp;
987 struct address_space *mapping = file->f_mapping;
988 struct inode *inode = mapping->host;
989 struct nfs_direct_req *dreq;
990 struct nfs_lock_context *l_ctx;
993 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
994 file, iov_iter_count(iter), (long long) iocb->ki_pos);
996 result = generic_write_checks(iocb, iter);
1000 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
1003 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1005 task_io_account_write(count);
1008 dreq = nfs_direct_req_alloc();
1012 dreq->inode = inode;
1013 dreq->bytes_left = dreq->max_count = count;
1014 dreq->io_start = pos;
1015 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1016 l_ctx = nfs_get_lock_context(dreq->ctx);
1017 if (IS_ERR(l_ctx)) {
1018 result = PTR_ERR(l_ctx);
1021 dreq->l_ctx = l_ctx;
1022 if (!is_sync_kiocb(iocb))
1025 nfs_start_io_direct(inode);
1027 result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1029 if (mapping->nrpages) {
1030 invalidate_inode_pages2_range(mapping,
1031 pos >> PAGE_SHIFT, end);
1034 nfs_end_io_direct(inode);
1037 result = nfs_direct_wait(dreq);
1039 iocb->ki_pos = pos + result;
1040 /* XXX: should check the generic_write_sync retval */
1041 generic_write_sync(iocb, result);
1045 nfs_direct_req_release(dreq);
1051 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1054 int __init nfs_init_directcache(void)
1056 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1057 sizeof(struct nfs_direct_req),
1058 0, (SLAB_RECLAIM_ACCOUNT|
1061 if (nfs_direct_cachep == NULL)
1068 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1071 void nfs_destroy_directcache(void)
1073 kmem_cache_destroy(nfs_direct_cachep);