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 <linux/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 loff_t io_start; /* Start offset for I/O */
90 ssize_t count, /* bytes actually processed */
91 max_count, /* max expected count */
92 bytes_left, /* bytes left to be sent */
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;
102 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
103 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
105 #define NFS_ODIRECT_SHOULD_DIRTY (3) /* dirty user-space page after read */
106 struct nfs_writeverf verf; /* unstable write verifier */
109 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
110 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
111 static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
112 static void nfs_direct_write_schedule_work(struct work_struct *work);
114 static inline void get_dreq(struct nfs_direct_req *dreq)
116 atomic_inc(&dreq->io_count);
119 static inline int put_dreq(struct nfs_direct_req *dreq)
121 return atomic_dec_and_test(&dreq->io_count);
125 nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
126 const struct nfs_pgio_header *hdr,
129 struct nfs_direct_mirror *mirror = &dreq->mirrors[hdr->pgio_mirror_idx];
131 if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
132 test_bit(NFS_IOHDR_EOF, &hdr->flags)))
134 if (dreq->max_count >= dreq_len) {
135 dreq->max_count = dreq_len;
136 if (dreq->count > dreq_len)
137 dreq->count = dreq_len;
139 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
140 dreq->error = hdr->error;
141 else /* Clear outstanding error if this is EOF */
144 if (mirror->count > dreq_len)
145 mirror->count = dreq_len;
149 nfs_direct_count_bytes(struct nfs_direct_req *dreq,
150 const struct nfs_pgio_header *hdr)
152 struct nfs_direct_mirror *mirror = &dreq->mirrors[hdr->pgio_mirror_idx];
153 loff_t hdr_end = hdr->io_start + hdr->good_bytes;
154 ssize_t dreq_len = 0;
156 if (hdr_end > dreq->io_start)
157 dreq_len = hdr_end - dreq->io_start;
159 nfs_direct_handle_truncated(dreq, hdr, dreq_len);
161 if (dreq_len > dreq->max_count)
162 dreq_len = dreq->max_count;
164 if (mirror->count < dreq_len)
165 mirror->count = dreq_len;
166 if (dreq->count < dreq_len)
167 dreq->count = dreq_len;
171 * nfs_direct_select_verf - select the right verifier
172 * @dreq - direct request possibly spanning multiple servers
173 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
174 * @commit_idx - commit bucket index for the DS
176 * returns the correct verifier to use given the role of the server
178 static struct nfs_writeverf *
179 nfs_direct_select_verf(struct nfs_direct_req *dreq,
180 struct nfs_client *ds_clp,
183 struct nfs_writeverf *verfp = &dreq->verf;
185 #ifdef CONFIG_NFS_V4_1
187 * pNFS is in use, use the DS verf except commit_through_mds is set
188 * for layout segment where nbuckets is zero.
190 if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
191 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
192 verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
202 * nfs_direct_set_hdr_verf - set the write/commit verifier
203 * @dreq - direct request possibly spanning multiple servers
204 * @hdr - pageio header to validate against previously seen verfs
206 * Set the server's (MDS or DS) "seen" verifier
208 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
209 struct nfs_pgio_header *hdr)
211 struct nfs_writeverf *verfp;
213 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
214 WARN_ON_ONCE(verfp->committed >= 0);
215 memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
216 WARN_ON_ONCE(verfp->committed < 0);
219 static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
220 const struct nfs_writeverf *v2)
222 return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
226 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
227 * @dreq - direct request possibly spanning multiple servers
228 * @hdr - pageio header to validate against previously seen verf
230 * set the server's "seen" verf if not initialized.
231 * returns result of comparison between @hdr->verf and the "seen"
232 * verf of the server used by @hdr (DS or MDS)
234 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
235 struct nfs_pgio_header *hdr)
237 struct nfs_writeverf *verfp;
239 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
240 if (verfp->committed < 0) {
241 nfs_direct_set_hdr_verf(dreq, hdr);
244 return nfs_direct_cmp_verf(verfp, &hdr->verf);
248 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
249 * @dreq - direct request possibly spanning multiple servers
250 * @data - commit data to validate against previously seen verf
252 * returns result of comparison between @data->verf and the verf of
253 * the server used by @data (DS or MDS)
255 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
256 struct nfs_commit_data *data)
258 struct nfs_writeverf *verfp;
260 verfp = nfs_direct_select_verf(dreq, data->ds_clp,
261 data->ds_commit_index);
263 /* verifier not set so always fail */
264 if (verfp->committed < 0 || data->res.verf->committed <= NFS_UNSTABLE)
267 return nfs_direct_cmp_verf(verfp, data->res.verf);
271 * nfs_direct_IO - NFS address space operation for direct I/O
272 * @iocb: target I/O control block
275 * The presence of this routine in the address space ops vector means
276 * the NFS client supports direct I/O. However, for most direct IO, we
277 * shunt off direct read and write requests before the VFS gets them,
278 * so this method is only ever called for swap.
280 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
282 struct inode *inode = iocb->ki_filp->f_mapping->host;
284 /* we only support swap file calling nfs_direct_IO */
285 if (!IS_SWAPFILE(inode))
288 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
290 if (iov_iter_rw(iter) == READ)
291 return nfs_file_direct_read(iocb, iter);
292 return nfs_file_direct_write(iocb, iter);
295 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
298 for (i = 0; i < npages; i++)
302 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
303 struct nfs_direct_req *dreq)
305 cinfo->inode = dreq->inode;
306 cinfo->mds = &dreq->mds_cinfo;
307 cinfo->ds = &dreq->ds_cinfo;
309 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
312 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
313 struct nfs_pageio_descriptor *pgio,
314 struct nfs_page *req)
316 int mirror_count = 1;
318 if (pgio->pg_ops->pg_get_mirror_count)
319 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
321 dreq->mirror_count = mirror_count;
324 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
326 struct nfs_direct_req *dreq;
328 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
332 kref_init(&dreq->kref);
333 kref_get(&dreq->kref);
334 init_completion(&dreq->completion);
335 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
336 dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */
337 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
338 dreq->mirror_count = 1;
339 spin_lock_init(&dreq->lock);
344 static void nfs_direct_req_free(struct kref *kref)
346 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
348 nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
349 if (dreq->l_ctx != NULL)
350 nfs_put_lock_context(dreq->l_ctx);
351 if (dreq->ctx != NULL)
352 put_nfs_open_context(dreq->ctx);
353 kmem_cache_free(nfs_direct_cachep, dreq);
356 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
358 kref_put(&dreq->kref, nfs_direct_req_free);
361 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
363 return dreq->bytes_left;
365 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
368 * Collects and returns the final error value/byte-count.
370 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
372 ssize_t result = -EIOCBQUEUED;
374 /* Async requests don't wait here */
378 result = wait_for_completion_killable(&dreq->completion);
381 result = dreq->count;
382 WARN_ON_ONCE(dreq->count < 0);
385 result = dreq->error;
388 return (ssize_t) result;
392 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
393 * the iocb is still valid here if this is a synchronous request.
395 static void nfs_direct_complete(struct nfs_direct_req *dreq)
397 struct inode *inode = dreq->inode;
399 inode_dio_end(inode);
402 long res = (long) dreq->error;
403 if (dreq->count != 0) {
404 res = (long) dreq->count;
405 WARN_ON_ONCE(dreq->count < 0);
407 dreq->iocb->ki_complete(dreq->iocb, res, 0);
410 complete(&dreq->completion);
412 nfs_direct_req_release(dreq);
415 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
417 unsigned long bytes = 0;
418 struct nfs_direct_req *dreq = hdr->dreq;
420 spin_lock(&dreq->lock);
421 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
422 spin_unlock(&dreq->lock);
426 nfs_direct_count_bytes(dreq, hdr);
427 spin_unlock(&dreq->lock);
429 while (!list_empty(&hdr->pages)) {
430 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
431 struct page *page = req->wb_page;
433 if (!PageCompound(page) && bytes < hdr->good_bytes &&
434 (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
435 set_page_dirty(page);
436 bytes += req->wb_bytes;
437 nfs_list_remove_request(req);
438 nfs_release_request(req);
442 nfs_direct_complete(dreq);
446 static void nfs_read_sync_pgio_error(struct list_head *head, int error)
448 struct nfs_page *req;
450 while (!list_empty(head)) {
451 req = nfs_list_entry(head->next);
452 nfs_list_remove_request(req);
453 nfs_release_request(req);
457 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
462 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
463 .error_cleanup = nfs_read_sync_pgio_error,
464 .init_hdr = nfs_direct_pgio_init,
465 .completion = nfs_direct_read_completion,
469 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
470 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
471 * bail and stop sending more reads. Read length accounting is
472 * handled automatically by nfs_direct_read_result(). Otherwise, if
473 * no requests have been sent, just return an error.
476 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
477 struct iov_iter *iter,
480 struct nfs_pageio_descriptor desc;
481 struct inode *inode = dreq->inode;
482 ssize_t result = -EINVAL;
483 size_t requested_bytes = 0;
484 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
486 nfs_pageio_init_read(&desc, dreq->inode, false,
487 &nfs_direct_read_completion_ops);
490 inode_dio_begin(inode);
492 while (iov_iter_count(iter)) {
493 struct page **pagevec;
498 result = iov_iter_get_pages_alloc(iter, &pagevec,
504 iov_iter_advance(iter, bytes);
505 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
506 for (i = 0; i < npages; i++) {
507 struct nfs_page *req;
508 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
509 /* XXX do we need to do the eof zeroing found in async_filler? */
510 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
513 result = PTR_ERR(req);
516 req->wb_index = pos >> PAGE_SHIFT;
517 req->wb_offset = pos & ~PAGE_MASK;
518 if (!nfs_pageio_add_request(&desc, req)) {
519 result = desc.pg_error;
520 nfs_release_request(req);
525 requested_bytes += req_len;
527 dreq->bytes_left -= req_len;
529 nfs_direct_release_pages(pagevec, npages);
535 nfs_pageio_complete(&desc);
538 * If no bytes were started, return the error, and let the
539 * generic layer handle the completion.
541 if (requested_bytes == 0) {
542 inode_dio_end(inode);
543 nfs_direct_req_release(dreq);
544 return result < 0 ? result : -EIO;
548 nfs_direct_complete(dreq);
549 return requested_bytes;
553 * nfs_file_direct_read - file direct read operation for NFS files
554 * @iocb: target I/O control block
555 * @iter: vector of user buffers into which to read data
557 * We use this function for direct reads instead of calling
558 * generic_file_aio_read() in order to avoid gfar's check to see if
559 * the request starts before the end of the file. For that check
560 * to work, we must generate a GETATTR before each direct read, and
561 * even then there is a window between the GETATTR and the subsequent
562 * READ where the file size could change. Our preference is simply
563 * to do all reads the application wants, and the server will take
564 * care of managing the end of file boundary.
566 * This function also eliminates unnecessarily updating the file's
567 * atime locally, as the NFS server sets the file's atime, and this
568 * client must read the updated atime from the server back into its
571 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
573 struct file *file = iocb->ki_filp;
574 struct address_space *mapping = file->f_mapping;
575 struct inode *inode = mapping->host;
576 struct nfs_direct_req *dreq;
577 struct nfs_lock_context *l_ctx;
578 ssize_t result = -EINVAL, requested;
579 size_t count = iov_iter_count(iter);
580 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
582 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
583 file, count, (long long) iocb->ki_pos);
589 task_io_account_read(count);
592 dreq = nfs_direct_req_alloc();
597 dreq->bytes_left = dreq->max_count = count;
598 dreq->io_start = iocb->ki_pos;
599 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
600 l_ctx = nfs_get_lock_context(dreq->ctx);
602 result = PTR_ERR(l_ctx);
603 nfs_direct_req_release(dreq);
607 if (!is_sync_kiocb(iocb))
610 if (iter_is_iovec(iter))
611 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
613 nfs_start_io_direct(inode);
615 NFS_I(inode)->read_io += count;
616 requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
618 nfs_end_io_direct(inode);
621 result = nfs_direct_wait(dreq);
624 iocb->ki_pos += result;
626 iov_iter_revert(iter, requested);
632 nfs_direct_req_release(dreq);
638 nfs_direct_write_scan_commit_list(struct inode *inode,
639 struct list_head *list,
640 struct nfs_commit_info *cinfo)
642 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
643 #ifdef CONFIG_NFS_V4_1
644 if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
645 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
647 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
648 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
651 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
653 struct nfs_pageio_descriptor desc;
654 struct nfs_page *req, *tmp;
656 struct nfs_commit_info cinfo;
660 nfs_init_cinfo_from_dreq(&cinfo, dreq);
661 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
665 list_for_each_entry(req, &reqs, wb_list)
666 dreq->max_count += req->wb_bytes;
667 dreq->verf.committed = NFS_INVALID_STABLE_HOW;
668 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
669 for (i = 0; i < dreq->mirror_count; i++)
670 dreq->mirrors[i].count = 0;
673 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
674 &nfs_direct_write_completion_ops);
677 req = nfs_list_entry(reqs.next);
678 nfs_direct_setup_mirroring(dreq, &desc, req);
679 if (desc.pg_error < 0) {
680 list_splice_init(&reqs, &failed);
684 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
685 if (!nfs_pageio_add_request(&desc, req)) {
686 nfs_list_move_request(req, &failed);
687 spin_lock(&cinfo.inode->i_lock);
689 if (desc.pg_error < 0)
690 dreq->error = desc.pg_error;
693 spin_unlock(&cinfo.inode->i_lock);
695 nfs_release_request(req);
697 nfs_pageio_complete(&desc);
700 while (!list_empty(&failed)) {
701 req = nfs_list_entry(failed.next);
702 nfs_list_remove_request(req);
703 nfs_unlock_and_release_request(req);
707 nfs_direct_write_complete(dreq);
710 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
712 struct nfs_direct_req *dreq = data->dreq;
713 struct nfs_commit_info cinfo;
714 struct nfs_page *req;
715 int status = data->task.tk_status;
717 nfs_init_cinfo_from_dreq(&cinfo, dreq);
718 if (status < 0 || nfs_direct_cmp_commit_data_verf(dreq, data))
719 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
721 while (!list_empty(&data->pages)) {
722 req = nfs_list_entry(data->pages.next);
723 nfs_list_remove_request(req);
724 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
725 /* Note the rewrite will go through mds */
726 nfs_mark_request_commit(req, NULL, &cinfo, 0);
728 nfs_release_request(req);
729 nfs_unlock_and_release_request(req);
732 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
733 nfs_direct_write_complete(dreq);
736 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
737 struct nfs_page *req)
739 struct nfs_direct_req *dreq = cinfo->dreq;
741 spin_lock(&dreq->lock);
742 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
743 spin_unlock(&dreq->lock);
744 nfs_mark_request_commit(req, NULL, cinfo, 0);
747 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
748 .completion = nfs_direct_commit_complete,
749 .resched_write = nfs_direct_resched_write,
752 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
755 struct nfs_commit_info cinfo;
758 nfs_init_cinfo_from_dreq(&cinfo, dreq);
759 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
760 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
761 if (res < 0) /* res == -ENOMEM */
762 nfs_direct_write_reschedule(dreq);
765 static void nfs_direct_write_schedule_work(struct work_struct *work)
767 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
768 int flags = dreq->flags;
772 case NFS_ODIRECT_DO_COMMIT:
773 nfs_direct_commit_schedule(dreq);
775 case NFS_ODIRECT_RESCHED_WRITES:
776 nfs_direct_write_reschedule(dreq);
779 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
780 nfs_direct_complete(dreq);
784 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
786 queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
789 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
791 struct nfs_direct_req *dreq = hdr->dreq;
792 struct nfs_commit_info cinfo;
793 bool request_commit = false;
794 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
796 nfs_init_cinfo_from_dreq(&cinfo, dreq);
798 spin_lock(&dreq->lock);
799 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
800 spin_unlock(&dreq->lock);
804 nfs_direct_count_bytes(dreq, hdr);
805 if (hdr->good_bytes != 0) {
806 if (nfs_write_need_commit(hdr)) {
807 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
808 request_commit = true;
809 else if (dreq->flags == 0) {
810 nfs_direct_set_hdr_verf(dreq, hdr);
811 request_commit = true;
812 dreq->flags = NFS_ODIRECT_DO_COMMIT;
813 } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
814 request_commit = true;
815 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
817 NFS_ODIRECT_RESCHED_WRITES;
821 spin_unlock(&dreq->lock);
823 while (!list_empty(&hdr->pages)) {
825 req = nfs_list_entry(hdr->pages.next);
826 nfs_list_remove_request(req);
827 if (request_commit) {
828 kref_get(&req->wb_kref);
829 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
832 nfs_unlock_and_release_request(req);
837 nfs_direct_write_complete(dreq);
841 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
843 struct nfs_page *req;
845 while (!list_empty(head)) {
846 req = nfs_list_entry(head->next);
847 nfs_list_remove_request(req);
848 nfs_unlock_and_release_request(req);
852 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
854 struct nfs_direct_req *dreq = hdr->dreq;
856 spin_lock(&dreq->lock);
857 if (dreq->error == 0) {
858 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
859 /* fake unstable write to let common nfs resend pages */
860 hdr->verf.committed = NFS_UNSTABLE;
861 hdr->good_bytes = hdr->args.count;
863 spin_unlock(&dreq->lock);
866 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
867 .error_cleanup = nfs_write_sync_pgio_error,
868 .init_hdr = nfs_direct_pgio_init,
869 .completion = nfs_direct_write_completion,
870 .reschedule_io = nfs_direct_write_reschedule_io,
875 * NB: Return the value of the first error return code. Subsequent
876 * errors after the first one are ignored.
879 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
880 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
881 * bail and stop sending more writes. Write length accounting is
882 * handled automatically by nfs_direct_write_result(). Otherwise, if
883 * no requests have been sent, just return an error.
885 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
886 struct iov_iter *iter,
889 struct nfs_pageio_descriptor desc;
890 struct inode *inode = dreq->inode;
892 size_t requested_bytes = 0;
893 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
895 nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
896 &nfs_direct_write_completion_ops);
899 inode_dio_begin(inode);
901 NFS_I(inode)->write_io += iov_iter_count(iter);
902 while (iov_iter_count(iter)) {
903 struct page **pagevec;
908 result = iov_iter_get_pages_alloc(iter, &pagevec,
914 iov_iter_advance(iter, bytes);
915 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
916 for (i = 0; i < npages; i++) {
917 struct nfs_page *req;
918 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
920 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
923 result = PTR_ERR(req);
927 nfs_direct_setup_mirroring(dreq, &desc, req);
928 if (desc.pg_error < 0) {
929 nfs_free_request(req);
930 result = desc.pg_error;
934 nfs_lock_request(req);
935 req->wb_index = pos >> PAGE_SHIFT;
936 req->wb_offset = pos & ~PAGE_MASK;
937 if (!nfs_pageio_add_request(&desc, req)) {
938 result = desc.pg_error;
939 nfs_unlock_and_release_request(req);
944 requested_bytes += req_len;
946 dreq->bytes_left -= req_len;
948 nfs_direct_release_pages(pagevec, npages);
953 nfs_pageio_complete(&desc);
956 * If no bytes were started, return the error, and let the
957 * generic layer handle the completion.
959 if (requested_bytes == 0) {
960 inode_dio_end(inode);
961 nfs_direct_req_release(dreq);
962 return result < 0 ? result : -EIO;
966 nfs_direct_write_complete(dreq);
967 return requested_bytes;
971 * nfs_file_direct_write - file direct write operation for NFS files
972 * @iocb: target I/O control block
973 * @iter: vector of user buffers from which to write data
975 * We use this function for direct writes instead of calling
976 * generic_file_aio_write() in order to avoid taking the inode
977 * semaphore and updating the i_size. The NFS server will set
978 * the new i_size and this client must read the updated size
979 * back into its cache. We let the server do generic write
980 * parameter checking and report problems.
982 * We eliminate local atime updates, see direct read above.
984 * We avoid unnecessary page cache invalidations for normal cached
985 * readers of this file.
987 * Note that O_APPEND is not supported for NFS direct writes, as there
988 * is no atomic O_APPEND write facility in the NFS protocol.
990 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
992 ssize_t result = -EINVAL, requested;
994 struct file *file = iocb->ki_filp;
995 struct address_space *mapping = file->f_mapping;
996 struct inode *inode = mapping->host;
997 struct nfs_direct_req *dreq;
998 struct nfs_lock_context *l_ctx;
1001 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
1002 file, iov_iter_count(iter), (long long) iocb->ki_pos);
1004 result = generic_write_checks(iocb, iter);
1008 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
1011 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1013 task_io_account_write(count);
1016 dreq = nfs_direct_req_alloc();
1020 dreq->inode = inode;
1021 dreq->bytes_left = dreq->max_count = count;
1022 dreq->io_start = pos;
1023 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1024 l_ctx = nfs_get_lock_context(dreq->ctx);
1025 if (IS_ERR(l_ctx)) {
1026 result = PTR_ERR(l_ctx);
1027 nfs_direct_req_release(dreq);
1030 dreq->l_ctx = l_ctx;
1031 if (!is_sync_kiocb(iocb))
1034 nfs_start_io_direct(inode);
1036 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1038 if (mapping->nrpages) {
1039 invalidate_inode_pages2_range(mapping,
1040 pos >> PAGE_SHIFT, end);
1043 nfs_end_io_direct(inode);
1045 if (requested > 0) {
1046 result = nfs_direct_wait(dreq);
1048 requested -= result;
1049 iocb->ki_pos = pos + result;
1050 /* XXX: should check the generic_write_sync retval */
1051 generic_write_sync(iocb, result);
1053 iov_iter_revert(iter, requested);
1058 nfs_direct_req_release(dreq);
1064 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1067 int __init nfs_init_directcache(void)
1069 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1070 sizeof(struct nfs_direct_req),
1071 0, (SLAB_RECLAIM_ACCOUNT|
1074 if (nfs_direct_cachep == NULL)
1081 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1084 void nfs_destroy_directcache(void)
1086 kmem_cache_destroy(nfs_direct_cachep);