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 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, struct inode *inode);
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_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
130 WARN_ON_ONCE(dreq->count >= dreq->max_count);
132 if (dreq->mirror_count == 1) {
133 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
134 dreq->count += hdr->good_bytes;
136 /* mirrored writes */
137 count = dreq->mirrors[hdr->pgio_mirror_idx].count;
138 if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
139 count = hdr->io_start + hdr->good_bytes - dreq->io_start;
140 dreq->mirrors[hdr->pgio_mirror_idx].count = count;
142 /* update the dreq->count by finding the minimum agreed count from all
144 count = dreq->mirrors[0].count;
146 for (i = 1; i < dreq->mirror_count; i++)
147 count = min(count, dreq->mirrors[i].count);
154 * nfs_direct_select_verf - select the right verifier
155 * @dreq - direct request possibly spanning multiple servers
156 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
157 * @commit_idx - commit bucket index for the DS
159 * returns the correct verifier to use given the role of the server
161 static struct nfs_writeverf *
162 nfs_direct_select_verf(struct nfs_direct_req *dreq,
163 struct nfs_client *ds_clp,
166 struct nfs_writeverf *verfp = &dreq->verf;
168 #ifdef CONFIG_NFS_V4_1
170 * pNFS is in use, use the DS verf except commit_through_mds is set
171 * for layout segment where nbuckets is zero.
173 if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
174 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
175 verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
185 * nfs_direct_set_hdr_verf - set the write/commit verifier
186 * @dreq - direct request possibly spanning multiple servers
187 * @hdr - pageio header to validate against previously seen verfs
189 * Set the server's (MDS or DS) "seen" verifier
191 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
192 struct nfs_pgio_header *hdr)
194 struct nfs_writeverf *verfp;
196 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
197 WARN_ON_ONCE(verfp->committed >= 0);
198 memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
199 WARN_ON_ONCE(verfp->committed < 0);
202 static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
203 const struct nfs_writeverf *v2)
205 return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
209 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
210 * @dreq - direct request possibly spanning multiple servers
211 * @hdr - pageio header to validate against previously seen verf
213 * set the server's "seen" verf if not initialized.
214 * returns result of comparison between @hdr->verf and the "seen"
215 * verf of the server used by @hdr (DS or MDS)
217 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
218 struct nfs_pgio_header *hdr)
220 struct nfs_writeverf *verfp;
222 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
223 if (verfp->committed < 0) {
224 nfs_direct_set_hdr_verf(dreq, hdr);
227 return nfs_direct_cmp_verf(verfp, &hdr->verf);
231 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
232 * @dreq - direct request possibly spanning multiple servers
233 * @data - commit data to validate against previously seen verf
235 * returns result of comparison between @data->verf and the verf of
236 * the server used by @data (DS or MDS)
238 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
239 struct nfs_commit_data *data)
241 struct nfs_writeverf *verfp;
243 verfp = nfs_direct_select_verf(dreq, data->ds_clp,
244 data->ds_commit_index);
246 /* verifier not set so always fail */
247 if (verfp->committed < 0)
250 return nfs_direct_cmp_verf(verfp, &data->verf);
254 * nfs_direct_IO - NFS address space operation for direct I/O
255 * @iocb: target I/O control block
258 * The presence of this routine in the address space ops vector means
259 * the NFS client supports direct I/O. However, for most direct IO, we
260 * shunt off direct read and write requests before the VFS gets them,
261 * so this method is only ever called for swap.
263 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
265 struct inode *inode = iocb->ki_filp->f_mapping->host;
267 /* we only support swap file calling nfs_direct_IO */
268 if (!IS_SWAPFILE(inode))
271 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
273 if (iov_iter_rw(iter) == READ)
274 return nfs_file_direct_read(iocb, iter);
275 return nfs_file_direct_write(iocb, iter);
278 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
281 for (i = 0; i < npages; i++)
285 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
286 struct nfs_direct_req *dreq)
288 cinfo->inode = dreq->inode;
289 cinfo->mds = &dreq->mds_cinfo;
290 cinfo->ds = &dreq->ds_cinfo;
292 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
295 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
296 struct nfs_pageio_descriptor *pgio,
297 struct nfs_page *req)
299 int mirror_count = 1;
301 if (pgio->pg_ops->pg_get_mirror_count)
302 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
304 dreq->mirror_count = mirror_count;
307 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
309 struct nfs_direct_req *dreq;
311 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
315 kref_init(&dreq->kref);
316 kref_get(&dreq->kref);
317 init_completion(&dreq->completion);
318 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
319 dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */
320 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
321 dreq->mirror_count = 1;
322 spin_lock_init(&dreq->lock);
327 static void nfs_direct_req_free(struct kref *kref)
329 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
331 nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
332 if (dreq->l_ctx != NULL)
333 nfs_put_lock_context(dreq->l_ctx);
334 if (dreq->ctx != NULL)
335 put_nfs_open_context(dreq->ctx);
336 kmem_cache_free(nfs_direct_cachep, dreq);
339 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
341 kref_put(&dreq->kref, nfs_direct_req_free);
344 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
346 return dreq->bytes_left;
348 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
351 * Collects and returns the final error value/byte-count.
353 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
355 ssize_t result = -EIOCBQUEUED;
357 /* Async requests don't wait here */
361 result = wait_for_completion_killable(&dreq->completion);
364 result = dreq->count;
365 WARN_ON_ONCE(dreq->count < 0);
368 result = dreq->error;
371 return (ssize_t) result;
375 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
376 * the iocb is still valid here if this is a synchronous request.
378 static void nfs_direct_complete(struct nfs_direct_req *dreq)
380 struct inode *inode = dreq->inode;
382 inode_dio_end(inode);
385 long res = (long) dreq->error;
386 if (dreq->count != 0) {
387 res = (long) dreq->count;
388 WARN_ON_ONCE(dreq->count < 0);
390 dreq->iocb->ki_complete(dreq->iocb, res, 0);
393 complete(&dreq->completion);
395 nfs_direct_req_release(dreq);
398 static void nfs_direct_readpage_release(struct nfs_page *req)
400 dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
401 req->wb_context->dentry->d_sb->s_id,
402 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
404 (long long)req_offset(req));
405 nfs_release_request(req);
408 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
410 unsigned long bytes = 0;
411 struct nfs_direct_req *dreq = hdr->dreq;
413 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
416 spin_lock(&dreq->lock);
417 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
418 dreq->error = hdr->error;
420 nfs_direct_good_bytes(dreq, hdr);
422 spin_unlock(&dreq->lock);
424 while (!list_empty(&hdr->pages)) {
425 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
426 struct page *page = req->wb_page;
428 if (!PageCompound(page) && bytes < hdr->good_bytes &&
429 (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
430 set_page_dirty(page);
431 bytes += req->wb_bytes;
432 nfs_list_remove_request(req);
433 nfs_direct_readpage_release(req);
437 nfs_direct_complete(dreq);
441 static void nfs_read_sync_pgio_error(struct list_head *head)
443 struct nfs_page *req;
445 while (!list_empty(head)) {
446 req = nfs_list_entry(head->next);
447 nfs_list_remove_request(req);
448 nfs_release_request(req);
452 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
457 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
458 .error_cleanup = nfs_read_sync_pgio_error,
459 .init_hdr = nfs_direct_pgio_init,
460 .completion = nfs_direct_read_completion,
464 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
465 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
466 * bail and stop sending more reads. Read length accounting is
467 * handled automatically by nfs_direct_read_result(). Otherwise, if
468 * no requests have been sent, just return an error.
471 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
472 struct iov_iter *iter,
475 struct nfs_pageio_descriptor desc;
476 struct inode *inode = dreq->inode;
477 ssize_t result = -EINVAL;
478 size_t requested_bytes = 0;
479 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
481 nfs_pageio_init_read(&desc, dreq->inode, false,
482 &nfs_direct_read_completion_ops);
485 inode_dio_begin(inode);
487 while (iov_iter_count(iter)) {
488 struct page **pagevec;
493 result = iov_iter_get_pages_alloc(iter, &pagevec,
499 iov_iter_advance(iter, bytes);
500 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
501 for (i = 0; i < npages; i++) {
502 struct nfs_page *req;
503 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
504 /* XXX do we need to do the eof zeroing found in async_filler? */
505 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
508 result = PTR_ERR(req);
511 req->wb_index = pos >> PAGE_SHIFT;
512 req->wb_offset = pos & ~PAGE_MASK;
513 if (!nfs_pageio_add_request(&desc, req)) {
514 result = desc.pg_error;
515 nfs_release_request(req);
520 requested_bytes += req_len;
522 dreq->bytes_left -= req_len;
524 nfs_direct_release_pages(pagevec, npages);
530 nfs_pageio_complete(&desc);
533 * If no bytes were started, return the error, and let the
534 * generic layer handle the completion.
536 if (requested_bytes == 0) {
537 inode_dio_end(inode);
538 nfs_direct_req_release(dreq);
539 return result < 0 ? result : -EIO;
543 nfs_direct_complete(dreq);
548 * nfs_file_direct_read - file direct read operation for NFS files
549 * @iocb: target I/O control block
550 * @iter: vector of user buffers into which to read data
552 * We use this function for direct reads instead of calling
553 * generic_file_aio_read() in order to avoid gfar's check to see if
554 * the request starts before the end of the file. For that check
555 * to work, we must generate a GETATTR before each direct read, and
556 * even then there is a window between the GETATTR and the subsequent
557 * READ where the file size could change. Our preference is simply
558 * to do all reads the application wants, and the server will take
559 * care of managing the end of file boundary.
561 * This function also eliminates unnecessarily updating the file's
562 * atime locally, as the NFS server sets the file's atime, and this
563 * client must read the updated atime from the server back into its
566 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
568 struct file *file = iocb->ki_filp;
569 struct address_space *mapping = file->f_mapping;
570 struct inode *inode = mapping->host;
571 struct nfs_direct_req *dreq;
572 struct nfs_lock_context *l_ctx;
573 ssize_t result = -EINVAL;
574 size_t count = iov_iter_count(iter);
575 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
577 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
578 file, count, (long long) iocb->ki_pos);
584 task_io_account_read(count);
587 dreq = nfs_direct_req_alloc();
592 dreq->bytes_left = dreq->max_count = count;
593 dreq->io_start = iocb->ki_pos;
594 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
595 l_ctx = nfs_get_lock_context(dreq->ctx);
597 result = PTR_ERR(l_ctx);
598 nfs_direct_req_release(dreq);
602 if (!is_sync_kiocb(iocb))
605 if (iter_is_iovec(iter))
606 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
608 nfs_start_io_direct(inode);
610 NFS_I(inode)->read_io += count;
611 result = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
613 nfs_end_io_direct(inode);
616 result = nfs_direct_wait(dreq);
618 iocb->ki_pos += result;
622 nfs_direct_req_release(dreq);
628 nfs_direct_write_scan_commit_list(struct inode *inode,
629 struct list_head *list,
630 struct nfs_commit_info *cinfo)
632 spin_lock(&cinfo->inode->i_lock);
633 #ifdef CONFIG_NFS_V4_1
634 if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
635 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
637 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
638 spin_unlock(&cinfo->inode->i_lock);
641 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
643 struct nfs_pageio_descriptor desc;
644 struct nfs_page *req, *tmp;
646 struct nfs_commit_info cinfo;
650 nfs_init_cinfo_from_dreq(&cinfo, dreq);
651 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
654 dreq->verf.committed = NFS_INVALID_STABLE_HOW;
655 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
656 for (i = 0; i < dreq->mirror_count; i++)
657 dreq->mirrors[i].count = 0;
660 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
661 &nfs_direct_write_completion_ops);
664 req = nfs_list_entry(reqs.next);
665 nfs_direct_setup_mirroring(dreq, &desc, req);
666 if (desc.pg_error < 0) {
667 list_splice_init(&reqs, &failed);
671 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
672 if (!nfs_pageio_add_request(&desc, req)) {
673 nfs_list_remove_request(req);
674 nfs_list_add_request(req, &failed);
675 spin_lock(&cinfo.inode->i_lock);
677 if (desc.pg_error < 0)
678 dreq->error = desc.pg_error;
681 spin_unlock(&cinfo.inode->i_lock);
683 nfs_release_request(req);
685 nfs_pageio_complete(&desc);
688 while (!list_empty(&failed)) {
689 req = nfs_list_entry(failed.next);
690 nfs_list_remove_request(req);
691 nfs_unlock_and_release_request(req);
695 nfs_direct_write_complete(dreq, dreq->inode);
698 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
700 struct nfs_direct_req *dreq = data->dreq;
701 struct nfs_commit_info cinfo;
702 struct nfs_page *req;
703 int status = data->task.tk_status;
705 nfs_init_cinfo_from_dreq(&cinfo, dreq);
707 dprintk("NFS: %5u commit failed with error %d.\n",
708 data->task.tk_pid, status);
709 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
710 } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
711 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
712 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
715 dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
716 while (!list_empty(&data->pages)) {
717 req = nfs_list_entry(data->pages.next);
718 nfs_list_remove_request(req);
719 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
720 /* Note the rewrite will go through mds */
721 nfs_mark_request_commit(req, NULL, &cinfo, 0);
723 nfs_release_request(req);
724 nfs_unlock_and_release_request(req);
727 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
728 nfs_direct_write_complete(dreq, data->inode);
731 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
732 struct nfs_page *req)
734 struct nfs_direct_req *dreq = cinfo->dreq;
736 spin_lock(&dreq->lock);
737 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
738 spin_unlock(&dreq->lock);
739 nfs_mark_request_commit(req, NULL, cinfo, 0);
742 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
743 .completion = nfs_direct_commit_complete,
744 .resched_write = nfs_direct_resched_write,
747 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
750 struct nfs_commit_info cinfo;
753 nfs_init_cinfo_from_dreq(&cinfo, dreq);
754 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
755 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
756 if (res < 0) /* res == -ENOMEM */
757 nfs_direct_write_reschedule(dreq);
760 static void nfs_direct_write_schedule_work(struct work_struct *work)
762 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
763 int flags = dreq->flags;
767 case NFS_ODIRECT_DO_COMMIT:
768 nfs_direct_commit_schedule(dreq);
770 case NFS_ODIRECT_RESCHED_WRITES:
771 nfs_direct_write_reschedule(dreq);
774 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
775 nfs_direct_complete(dreq);
779 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
781 schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
784 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
786 struct nfs_direct_req *dreq = hdr->dreq;
787 struct nfs_commit_info cinfo;
788 bool request_commit = false;
789 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
791 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
794 nfs_init_cinfo_from_dreq(&cinfo, dreq);
796 spin_lock(&dreq->lock);
798 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
799 dreq->error = hdr->error;
800 if (dreq->error == 0) {
801 nfs_direct_good_bytes(dreq, hdr);
802 if (nfs_write_need_commit(hdr)) {
803 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
804 request_commit = true;
805 else if (dreq->flags == 0) {
806 nfs_direct_set_hdr_verf(dreq, hdr);
807 request_commit = true;
808 dreq->flags = NFS_ODIRECT_DO_COMMIT;
809 } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
810 request_commit = true;
811 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
813 NFS_ODIRECT_RESCHED_WRITES;
817 spin_unlock(&dreq->lock);
819 while (!list_empty(&hdr->pages)) {
821 req = nfs_list_entry(hdr->pages.next);
822 nfs_list_remove_request(req);
823 if (request_commit) {
824 kref_get(&req->wb_kref);
825 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
828 nfs_unlock_and_release_request(req);
833 nfs_direct_write_complete(dreq, hdr->inode);
837 static void nfs_write_sync_pgio_error(struct list_head *head)
839 struct nfs_page *req;
841 while (!list_empty(head)) {
842 req = nfs_list_entry(head->next);
843 nfs_list_remove_request(req);
844 nfs_unlock_and_release_request(req);
848 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
850 struct nfs_direct_req *dreq = hdr->dreq;
852 spin_lock(&dreq->lock);
853 if (dreq->error == 0) {
854 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
855 /* fake unstable write to let common nfs resend pages */
856 hdr->verf.committed = NFS_UNSTABLE;
857 hdr->good_bytes = hdr->args.count;
859 spin_unlock(&dreq->lock);
862 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
863 .error_cleanup = nfs_write_sync_pgio_error,
864 .init_hdr = nfs_direct_pgio_init,
865 .completion = nfs_direct_write_completion,
866 .reschedule_io = nfs_direct_write_reschedule_io,
871 * NB: Return the value of the first error return code. Subsequent
872 * errors after the first one are ignored.
875 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
876 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
877 * bail and stop sending more writes. Write length accounting is
878 * handled automatically by nfs_direct_write_result(). Otherwise, if
879 * no requests have been sent, just return an error.
881 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
882 struct iov_iter *iter,
885 struct nfs_pageio_descriptor desc;
886 struct inode *inode = dreq->inode;
888 size_t requested_bytes = 0;
889 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
891 nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
892 &nfs_direct_write_completion_ops);
895 inode_dio_begin(inode);
897 NFS_I(inode)->write_io += iov_iter_count(iter);
898 while (iov_iter_count(iter)) {
899 struct page **pagevec;
904 result = iov_iter_get_pages_alloc(iter, &pagevec,
910 iov_iter_advance(iter, bytes);
911 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
912 for (i = 0; i < npages; i++) {
913 struct nfs_page *req;
914 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
916 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
919 result = PTR_ERR(req);
923 nfs_direct_setup_mirroring(dreq, &desc, req);
924 if (desc.pg_error < 0) {
925 nfs_free_request(req);
926 result = desc.pg_error;
930 nfs_lock_request(req);
931 req->wb_index = pos >> PAGE_SHIFT;
932 req->wb_offset = pos & ~PAGE_MASK;
933 if (!nfs_pageio_add_request(&desc, req)) {
934 result = desc.pg_error;
935 nfs_unlock_and_release_request(req);
940 requested_bytes += req_len;
942 dreq->bytes_left -= req_len;
944 nfs_direct_release_pages(pagevec, npages);
949 nfs_pageio_complete(&desc);
952 * If no bytes were started, return the error, and let the
953 * generic layer handle the completion.
955 if (requested_bytes == 0) {
956 inode_dio_end(inode);
957 nfs_direct_req_release(dreq);
958 return result < 0 ? result : -EIO;
962 nfs_direct_write_complete(dreq, dreq->inode);
967 * nfs_file_direct_write - file direct write operation for NFS files
968 * @iocb: target I/O control block
969 * @iter: vector of user buffers from which to write data
971 * We use this function for direct writes instead of calling
972 * generic_file_aio_write() in order to avoid taking the inode
973 * semaphore and updating the i_size. The NFS server will set
974 * the new i_size and this client must read the updated size
975 * back into its cache. We let the server do generic write
976 * parameter checking and report problems.
978 * We eliminate local atime updates, see direct read above.
980 * We avoid unnecessary page cache invalidations for normal cached
981 * readers of this file.
983 * Note that O_APPEND is not supported for NFS direct writes, as there
984 * is no atomic O_APPEND write facility in the NFS protocol.
986 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
988 ssize_t result = -EINVAL;
990 struct file *file = iocb->ki_filp;
991 struct address_space *mapping = file->f_mapping;
992 struct inode *inode = mapping->host;
993 struct nfs_direct_req *dreq;
994 struct nfs_lock_context *l_ctx;
997 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
998 file, iov_iter_count(iter), (long long) iocb->ki_pos);
1000 result = generic_write_checks(iocb, iter);
1004 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
1007 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1009 task_io_account_write(count);
1012 dreq = nfs_direct_req_alloc();
1016 dreq->inode = inode;
1017 dreq->bytes_left = dreq->max_count = count;
1018 dreq->io_start = pos;
1019 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1020 l_ctx = nfs_get_lock_context(dreq->ctx);
1021 if (IS_ERR(l_ctx)) {
1022 result = PTR_ERR(l_ctx);
1023 nfs_direct_req_release(dreq);
1026 dreq->l_ctx = l_ctx;
1027 if (!is_sync_kiocb(iocb))
1030 nfs_start_io_direct(inode);
1032 result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1034 if (mapping->nrpages) {
1035 invalidate_inode_pages2_range(mapping,
1036 pos >> PAGE_SHIFT, end);
1039 nfs_end_io_direct(inode);
1042 result = nfs_direct_wait(dreq);
1044 iocb->ki_pos = pos + result;
1045 /* XXX: should check the generic_write_sync retval */
1046 generic_write_sync(iocb, result);
1050 nfs_direct_req_release(dreq);
1056 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1059 int __init nfs_init_directcache(void)
1061 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1062 sizeof(struct nfs_direct_req),
1063 0, (SLAB_RECLAIM_ACCOUNT|
1066 if (nfs_direct_cachep == NULL)
1073 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1076 void nfs_destroy_directcache(void)
1078 kmem_cache_destroy(nfs_direct_cachep);