Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/fs/nfs/direct.c
4 *
5 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
6 *
7 * High-performance uncached I/O for the Linux NFS client
8 *
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.
17 *
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
23 * an application.
24 *
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.
29 *
30 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
31 * help from Andrew Morton.
32 *
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
39 *
40 */
41
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>
51
52#include <linux/nfs_fs.h>
53#include <linux/nfs_page.h>
54#include <linux/sunrpc/clnt.h>
55
56#include <linux/uaccess.h>
57#include <linux/atomic.h>
58
59#include "delegation.h"
60#include "internal.h"
61#include "iostat.h"
62#include "pnfs.h"
63#include "fscache.h"
64#include "nfstrace.h"
65
66#define NFSDBG_FACILITY NFSDBG_VFS
67
68static struct kmem_cache *nfs_direct_cachep;
69
70static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
71static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
72static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
73static void nfs_direct_write_schedule_work(struct work_struct *work);
74
75static inline void get_dreq(struct nfs_direct_req *dreq)
76{
77 atomic_inc(&dreq->io_count);
78}
79
80static inline int put_dreq(struct nfs_direct_req *dreq)
81{
82 return atomic_dec_and_test(&dreq->io_count);
83}
84
85static void
86nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
87 const struct nfs_pgio_header *hdr,
88 ssize_t dreq_len)
89{
90 if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
91 test_bit(NFS_IOHDR_EOF, &hdr->flags)))
92 return;
93 if (dreq->max_count >= dreq_len) {
94 dreq->max_count = dreq_len;
95 if (dreq->count > dreq_len)
96 dreq->count = dreq_len;
97 }
98
99 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && !dreq->error)
100 dreq->error = hdr->error;
101}
102
103static void
104nfs_direct_count_bytes(struct nfs_direct_req *dreq,
105 const struct nfs_pgio_header *hdr)
106{
107 loff_t hdr_end = hdr->io_start + hdr->good_bytes;
108 ssize_t dreq_len = 0;
109
110 if (hdr_end > dreq->io_start)
111 dreq_len = hdr_end - dreq->io_start;
112
113 nfs_direct_handle_truncated(dreq, hdr, dreq_len);
114
115 if (dreq_len > dreq->max_count)
116 dreq_len = dreq->max_count;
117
118 if (dreq->count < dreq_len)
119 dreq->count = dreq_len;
120}
121
122static void nfs_direct_truncate_request(struct nfs_direct_req *dreq,
123 struct nfs_page *req)
124{
125 loff_t offs = req_offset(req);
126 size_t req_start = (size_t)(offs - dreq->io_start);
127
128 if (req_start < dreq->max_count)
129 dreq->max_count = req_start;
130 if (req_start < dreq->count)
131 dreq->count = req_start;
132}
133
134static void nfs_direct_file_adjust_size_locked(struct inode *inode,
135 loff_t offset, size_t count)
136{
137 loff_t newsize = offset + (loff_t)count;
138 loff_t oldsize = i_size_read(inode);
139
140 if (newsize > oldsize) {
141 i_size_write(inode, newsize);
142 NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
143 trace_nfs_size_grow(inode, newsize);
144 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
145 }
146}
147
148/**
149 * nfs_swap_rw - NFS address space operation for swap I/O
150 * @iocb: target I/O control block
151 * @iter: I/O buffer
152 *
153 * Perform IO to the swap-file. This is much like direct IO.
154 */
155int nfs_swap_rw(struct kiocb *iocb, struct iov_iter *iter)
156{
157 ssize_t ret;
158
159 if (iov_iter_rw(iter) == READ)
160 ret = nfs_file_direct_read(iocb, iter, true);
161 else
162 ret = nfs_file_direct_write(iocb, iter, true);
163 if (ret < 0)
164 return ret;
165 return 0;
166}
167
168static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
169{
170 unsigned int i;
171 for (i = 0; i < npages; i++)
172 put_page(pages[i]);
173}
174
175void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
176 struct nfs_direct_req *dreq)
177{
178 cinfo->inode = dreq->inode;
179 cinfo->mds = &dreq->mds_cinfo;
180 cinfo->ds = &dreq->ds_cinfo;
181 cinfo->dreq = dreq;
182 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
183}
184
185static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
186{
187 struct nfs_direct_req *dreq;
188
189 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
190 if (!dreq)
191 return NULL;
192
193 kref_init(&dreq->kref);
194 kref_get(&dreq->kref);
195 init_completion(&dreq->completion);
196 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
197 pnfs_init_ds_commit_info(&dreq->ds_cinfo);
198 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
199 spin_lock_init(&dreq->lock);
200
201 return dreq;
202}
203
204static void nfs_direct_req_free(struct kref *kref)
205{
206 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
207
208 pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode);
209 if (dreq->l_ctx != NULL)
210 nfs_put_lock_context(dreq->l_ctx);
211 if (dreq->ctx != NULL)
212 put_nfs_open_context(dreq->ctx);
213 kmem_cache_free(nfs_direct_cachep, dreq);
214}
215
216static void nfs_direct_req_release(struct nfs_direct_req *dreq)
217{
218 kref_put(&dreq->kref, nfs_direct_req_free);
219}
220
221ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq, loff_t offset)
222{
223 loff_t start = offset - dreq->io_start;
224 return dreq->max_count - start;
225}
226EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
227
228/*
229 * Collects and returns the final error value/byte-count.
230 */
231static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
232{
233 ssize_t result = -EIOCBQUEUED;
234
235 /* Async requests don't wait here */
236 if (dreq->iocb)
237 goto out;
238
239 result = wait_for_completion_killable(&dreq->completion);
240
241 if (!result) {
242 result = dreq->count;
243 WARN_ON_ONCE(dreq->count < 0);
244 }
245 if (!result)
246 result = dreq->error;
247
248out:
249 return (ssize_t) result;
250}
251
252/*
253 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
254 * the iocb is still valid here if this is a synchronous request.
255 */
256static void nfs_direct_complete(struct nfs_direct_req *dreq)
257{
258 struct inode *inode = dreq->inode;
259
260 inode_dio_end(inode);
261
262 if (dreq->iocb) {
263 long res = (long) dreq->error;
264 if (dreq->count != 0) {
265 res = (long) dreq->count;
266 WARN_ON_ONCE(dreq->count < 0);
267 }
268 dreq->iocb->ki_complete(dreq->iocb, res);
269 }
270
271 complete(&dreq->completion);
272
273 nfs_direct_req_release(dreq);
274}
275
276static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
277{
278 unsigned long bytes = 0;
279 struct nfs_direct_req *dreq = hdr->dreq;
280
281 spin_lock(&dreq->lock);
282 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
283 spin_unlock(&dreq->lock);
284 goto out_put;
285 }
286
287 nfs_direct_count_bytes(dreq, hdr);
288 spin_unlock(&dreq->lock);
289
290 nfs_update_delegated_atime(dreq->inode);
291
292 while (!list_empty(&hdr->pages)) {
293 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
294 struct page *page = req->wb_page;
295
296 if (!PageCompound(page) && bytes < hdr->good_bytes &&
297 (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
298 set_page_dirty(page);
299 bytes += req->wb_bytes;
300 nfs_list_remove_request(req);
301 nfs_release_request(req);
302 }
303out_put:
304 if (put_dreq(dreq))
305 nfs_direct_complete(dreq);
306 hdr->release(hdr);
307}
308
309static void nfs_read_sync_pgio_error(struct list_head *head, int error)
310{
311 struct nfs_page *req;
312
313 while (!list_empty(head)) {
314 req = nfs_list_entry(head->next);
315 nfs_list_remove_request(req);
316 nfs_release_request(req);
317 }
318}
319
320static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
321{
322 get_dreq(hdr->dreq);
323}
324
325static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
326 .error_cleanup = nfs_read_sync_pgio_error,
327 .init_hdr = nfs_direct_pgio_init,
328 .completion = nfs_direct_read_completion,
329};
330
331/*
332 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
333 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
334 * bail and stop sending more reads. Read length accounting is
335 * handled automatically by nfs_direct_read_result(). Otherwise, if
336 * no requests have been sent, just return an error.
337 */
338
339static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
340 struct iov_iter *iter,
341 loff_t pos)
342{
343 struct nfs_pageio_descriptor desc;
344 struct inode *inode = dreq->inode;
345 ssize_t result = -EINVAL;
346 size_t requested_bytes = 0;
347 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
348
349 nfs_pageio_init_read(&desc, dreq->inode, false,
350 &nfs_direct_read_completion_ops);
351 get_dreq(dreq);
352 desc.pg_dreq = dreq;
353 inode_dio_begin(inode);
354
355 while (iov_iter_count(iter)) {
356 struct page **pagevec;
357 size_t bytes;
358 size_t pgbase;
359 unsigned npages, i;
360
361 result = iov_iter_get_pages_alloc2(iter, &pagevec,
362 rsize, &pgbase);
363 if (result < 0)
364 break;
365
366 bytes = result;
367 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
368 for (i = 0; i < npages; i++) {
369 struct nfs_page *req;
370 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
371 /* XXX do we need to do the eof zeroing found in async_filler? */
372 req = nfs_page_create_from_page(dreq->ctx, pagevec[i],
373 pgbase, pos, req_len);
374 if (IS_ERR(req)) {
375 result = PTR_ERR(req);
376 break;
377 }
378 if (!nfs_pageio_add_request(&desc, req)) {
379 result = desc.pg_error;
380 nfs_release_request(req);
381 break;
382 }
383 pgbase = 0;
384 bytes -= req_len;
385 requested_bytes += req_len;
386 pos += req_len;
387 }
388 nfs_direct_release_pages(pagevec, npages);
389 kvfree(pagevec);
390 if (result < 0)
391 break;
392 }
393
394 nfs_pageio_complete(&desc);
395
396 /*
397 * If no bytes were started, return the error, and let the
398 * generic layer handle the completion.
399 */
400 if (requested_bytes == 0) {
401 inode_dio_end(inode);
402 nfs_direct_req_release(dreq);
403 return result < 0 ? result : -EIO;
404 }
405
406 if (put_dreq(dreq))
407 nfs_direct_complete(dreq);
408 return requested_bytes;
409}
410
411/**
412 * nfs_file_direct_read - file direct read operation for NFS files
413 * @iocb: target I/O control block
414 * @iter: vector of user buffers into which to read data
415 * @swap: flag indicating this is swap IO, not O_DIRECT IO
416 *
417 * We use this function for direct reads instead of calling
418 * generic_file_aio_read() in order to avoid gfar's check to see if
419 * the request starts before the end of the file. For that check
420 * to work, we must generate a GETATTR before each direct read, and
421 * even then there is a window between the GETATTR and the subsequent
422 * READ where the file size could change. Our preference is simply
423 * to do all reads the application wants, and the server will take
424 * care of managing the end of file boundary.
425 *
426 * This function also eliminates unnecessarily updating the file's
427 * atime locally, as the NFS server sets the file's atime, and this
428 * client must read the updated atime from the server back into its
429 * cache.
430 */
431ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
432 bool swap)
433{
434 struct file *file = iocb->ki_filp;
435 struct address_space *mapping = file->f_mapping;
436 struct inode *inode = mapping->host;
437 struct nfs_direct_req *dreq;
438 struct nfs_lock_context *l_ctx;
439 ssize_t result, requested;
440 size_t count = iov_iter_count(iter);
441 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
442
443 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
444 file, count, (long long) iocb->ki_pos);
445
446 result = 0;
447 if (!count)
448 goto out;
449
450 task_io_account_read(count);
451
452 result = -ENOMEM;
453 dreq = nfs_direct_req_alloc();
454 if (dreq == NULL)
455 goto out;
456
457 dreq->inode = inode;
458 dreq->max_count = count;
459 dreq->io_start = iocb->ki_pos;
460 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
461 l_ctx = nfs_get_lock_context(dreq->ctx);
462 if (IS_ERR(l_ctx)) {
463 result = PTR_ERR(l_ctx);
464 nfs_direct_req_release(dreq);
465 goto out_release;
466 }
467 dreq->l_ctx = l_ctx;
468 if (!is_sync_kiocb(iocb))
469 dreq->iocb = iocb;
470
471 if (user_backed_iter(iter))
472 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
473
474 if (!swap) {
475 result = nfs_start_io_direct(inode);
476 if (result) {
477 /* release the reference that would usually be
478 * consumed by nfs_direct_read_schedule_iovec()
479 */
480 nfs_direct_req_release(dreq);
481 goto out_release;
482 }
483 }
484
485 NFS_I(inode)->read_io += count;
486 requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
487
488 if (!swap)
489 nfs_end_io_direct(inode);
490
491 if (requested > 0) {
492 result = nfs_direct_wait(dreq);
493 if (result > 0) {
494 requested -= result;
495 iocb->ki_pos += result;
496 }
497 iov_iter_revert(iter, requested);
498 } else {
499 result = requested;
500 }
501
502out_release:
503 nfs_direct_req_release(dreq);
504out:
505 return result;
506}
507
508static void nfs_direct_add_page_head(struct list_head *list,
509 struct nfs_page *req)
510{
511 struct nfs_page *head = req->wb_head;
512
513 if (!list_empty(&head->wb_list) || !nfs_lock_request(head))
514 return;
515 if (!list_empty(&head->wb_list)) {
516 nfs_unlock_request(head);
517 return;
518 }
519 list_add(&head->wb_list, list);
520 kref_get(&head->wb_kref);
521 kref_get(&head->wb_kref);
522}
523
524static void nfs_direct_join_group(struct list_head *list,
525 struct nfs_commit_info *cinfo,
526 struct inode *inode)
527{
528 struct nfs_page *req, *subreq;
529
530 list_for_each_entry(req, list, wb_list) {
531 if (req->wb_head != req) {
532 nfs_direct_add_page_head(&req->wb_list, req);
533 continue;
534 }
535 subreq = req->wb_this_page;
536 if (subreq == req)
537 continue;
538 do {
539 /*
540 * Remove subrequests from this list before freeing
541 * them in the call to nfs_join_page_group().
542 */
543 if (!list_empty(&subreq->wb_list)) {
544 nfs_list_remove_request(subreq);
545 nfs_release_request(subreq);
546 }
547 } while ((subreq = subreq->wb_this_page) != req);
548 nfs_join_page_group(req, cinfo, inode);
549 }
550}
551
552static void
553nfs_direct_write_scan_commit_list(struct inode *inode,
554 struct list_head *list,
555 struct nfs_commit_info *cinfo)
556{
557 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
558 pnfs_recover_commit_reqs(list, cinfo);
559 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
560 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
561}
562
563static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
564{
565 struct nfs_pageio_descriptor desc;
566 struct nfs_page *req;
567 LIST_HEAD(reqs);
568 struct nfs_commit_info cinfo;
569
570 nfs_init_cinfo_from_dreq(&cinfo, dreq);
571 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
572
573 nfs_direct_join_group(&reqs, &cinfo, dreq->inode);
574
575 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
576 get_dreq(dreq);
577
578 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
579 &nfs_direct_write_completion_ops);
580 desc.pg_dreq = dreq;
581
582 while (!list_empty(&reqs)) {
583 req = nfs_list_entry(reqs.next);
584 /* Bump the transmission count */
585 req->wb_nio++;
586 if (!nfs_pageio_add_request(&desc, req)) {
587 spin_lock(&dreq->lock);
588 if (dreq->error < 0) {
589 desc.pg_error = dreq->error;
590 } else if (desc.pg_error != -EAGAIN) {
591 dreq->flags = 0;
592 if (!desc.pg_error)
593 desc.pg_error = -EIO;
594 dreq->error = desc.pg_error;
595 } else
596 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
597 spin_unlock(&dreq->lock);
598 break;
599 }
600 nfs_release_request(req);
601 }
602 nfs_pageio_complete(&desc);
603
604 while (!list_empty(&reqs)) {
605 req = nfs_list_entry(reqs.next);
606 nfs_list_remove_request(req);
607 nfs_unlock_and_release_request(req);
608 if (desc.pg_error == -EAGAIN) {
609 nfs_mark_request_commit(req, NULL, &cinfo, 0);
610 } else {
611 spin_lock(&dreq->lock);
612 nfs_direct_truncate_request(dreq, req);
613 spin_unlock(&dreq->lock);
614 nfs_release_request(req);
615 }
616 }
617
618 if (put_dreq(dreq))
619 nfs_direct_write_complete(dreq);
620}
621
622static void nfs_direct_commit_complete(struct nfs_commit_data *data)
623{
624 const struct nfs_writeverf *verf = data->res.verf;
625 struct nfs_direct_req *dreq = data->dreq;
626 struct nfs_commit_info cinfo;
627 struct nfs_page *req;
628 int status = data->task.tk_status;
629
630 trace_nfs_direct_commit_complete(dreq);
631
632 spin_lock(&dreq->lock);
633 if (status < 0) {
634 /* Errors in commit are fatal */
635 dreq->error = status;
636 dreq->flags = NFS_ODIRECT_DONE;
637 } else {
638 status = dreq->error;
639 }
640 spin_unlock(&dreq->lock);
641
642 nfs_init_cinfo_from_dreq(&cinfo, dreq);
643
644 while (!list_empty(&data->pages)) {
645 req = nfs_list_entry(data->pages.next);
646 nfs_list_remove_request(req);
647 if (status < 0) {
648 spin_lock(&dreq->lock);
649 nfs_direct_truncate_request(dreq, req);
650 spin_unlock(&dreq->lock);
651 nfs_release_request(req);
652 } else if (!nfs_write_match_verf(verf, req)) {
653 spin_lock(&dreq->lock);
654 if (dreq->flags == 0)
655 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
656 spin_unlock(&dreq->lock);
657 /*
658 * Despite the reboot, the write was successful,
659 * so reset wb_nio.
660 */
661 req->wb_nio = 0;
662 nfs_mark_request_commit(req, NULL, &cinfo, 0);
663 } else
664 nfs_release_request(req);
665 nfs_unlock_and_release_request(req);
666 }
667
668 if (nfs_commit_end(cinfo.mds))
669 nfs_direct_write_complete(dreq);
670}
671
672static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
673 struct nfs_page *req)
674{
675 struct nfs_direct_req *dreq = cinfo->dreq;
676
677 trace_nfs_direct_resched_write(dreq);
678
679 spin_lock(&dreq->lock);
680 if (dreq->flags != NFS_ODIRECT_DONE)
681 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
682 spin_unlock(&dreq->lock);
683 nfs_mark_request_commit(req, NULL, cinfo, 0);
684}
685
686static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
687 .completion = nfs_direct_commit_complete,
688 .resched_write = nfs_direct_resched_write,
689};
690
691static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
692{
693 int res;
694 struct nfs_commit_info cinfo;
695 LIST_HEAD(mds_list);
696
697 nfs_init_cinfo_from_dreq(&cinfo, dreq);
698 nfs_commit_begin(cinfo.mds);
699 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
700 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
701 if (res < 0) { /* res == -ENOMEM */
702 spin_lock(&dreq->lock);
703 if (dreq->flags == 0)
704 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
705 spin_unlock(&dreq->lock);
706 }
707 if (nfs_commit_end(cinfo.mds))
708 nfs_direct_write_complete(dreq);
709}
710
711static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
712{
713 struct nfs_commit_info cinfo;
714 struct nfs_page *req;
715 LIST_HEAD(reqs);
716
717 nfs_init_cinfo_from_dreq(&cinfo, dreq);
718 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
719
720 while (!list_empty(&reqs)) {
721 req = nfs_list_entry(reqs.next);
722 nfs_list_remove_request(req);
723 nfs_direct_truncate_request(dreq, req);
724 nfs_release_request(req);
725 nfs_unlock_and_release_request(req);
726 }
727}
728
729static void nfs_direct_write_schedule_work(struct work_struct *work)
730{
731 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
732 int flags = dreq->flags;
733
734 dreq->flags = 0;
735 switch (flags) {
736 case NFS_ODIRECT_DO_COMMIT:
737 nfs_direct_commit_schedule(dreq);
738 break;
739 case NFS_ODIRECT_RESCHED_WRITES:
740 nfs_direct_write_reschedule(dreq);
741 break;
742 default:
743 nfs_direct_write_clear_reqs(dreq);
744 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
745 nfs_direct_complete(dreq);
746 }
747}
748
749static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
750{
751 trace_nfs_direct_write_complete(dreq);
752 queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
753}
754
755static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
756{
757 struct nfs_direct_req *dreq = hdr->dreq;
758 struct nfs_commit_info cinfo;
759 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
760 struct inode *inode = dreq->inode;
761 int flags = NFS_ODIRECT_DONE;
762
763 trace_nfs_direct_write_completion(dreq);
764
765 nfs_init_cinfo_from_dreq(&cinfo, dreq);
766
767 spin_lock(&dreq->lock);
768 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
769 spin_unlock(&dreq->lock);
770 goto out_put;
771 }
772
773 nfs_direct_count_bytes(dreq, hdr);
774 if (test_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags) &&
775 !test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
776 if (!dreq->flags)
777 dreq->flags = NFS_ODIRECT_DO_COMMIT;
778 flags = dreq->flags;
779 }
780 spin_unlock(&dreq->lock);
781
782 spin_lock(&inode->i_lock);
783 nfs_direct_file_adjust_size_locked(inode, dreq->io_start, dreq->count);
784 nfs_update_delegated_mtime_locked(dreq->inode);
785 spin_unlock(&inode->i_lock);
786
787 while (!list_empty(&hdr->pages)) {
788
789 req = nfs_list_entry(hdr->pages.next);
790 nfs_list_remove_request(req);
791 if (flags == NFS_ODIRECT_DO_COMMIT) {
792 kref_get(&req->wb_kref);
793 memcpy(&req->wb_verf, &hdr->verf.verifier,
794 sizeof(req->wb_verf));
795 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
796 hdr->ds_commit_idx);
797 } else if (flags == NFS_ODIRECT_RESCHED_WRITES) {
798 kref_get(&req->wb_kref);
799 nfs_mark_request_commit(req, NULL, &cinfo, 0);
800 }
801 nfs_unlock_and_release_request(req);
802 }
803
804out_put:
805 if (put_dreq(dreq))
806 nfs_direct_write_complete(dreq);
807 hdr->release(hdr);
808}
809
810static void nfs_write_sync_pgio_error(struct list_head *head, int error)
811{
812 struct nfs_page *req;
813
814 while (!list_empty(head)) {
815 req = nfs_list_entry(head->next);
816 nfs_list_remove_request(req);
817 nfs_unlock_and_release_request(req);
818 }
819}
820
821static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
822{
823 struct nfs_direct_req *dreq = hdr->dreq;
824 struct nfs_page *req;
825 struct nfs_commit_info cinfo;
826
827 trace_nfs_direct_write_reschedule_io(dreq);
828
829 nfs_init_cinfo_from_dreq(&cinfo, dreq);
830 spin_lock(&dreq->lock);
831 if (dreq->error == 0)
832 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
833 set_bit(NFS_IOHDR_REDO, &hdr->flags);
834 spin_unlock(&dreq->lock);
835 while (!list_empty(&hdr->pages)) {
836 req = nfs_list_entry(hdr->pages.next);
837 nfs_list_remove_request(req);
838 nfs_unlock_request(req);
839 nfs_mark_request_commit(req, NULL, &cinfo, 0);
840 }
841}
842
843static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
844 .error_cleanup = nfs_write_sync_pgio_error,
845 .init_hdr = nfs_direct_pgio_init,
846 .completion = nfs_direct_write_completion,
847 .reschedule_io = nfs_direct_write_reschedule_io,
848};
849
850
851/*
852 * NB: Return the value of the first error return code. Subsequent
853 * errors after the first one are ignored.
854 */
855/*
856 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
857 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
858 * bail and stop sending more writes. Write length accounting is
859 * handled automatically by nfs_direct_write_result(). Otherwise, if
860 * no requests have been sent, just return an error.
861 */
862static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
863 struct iov_iter *iter,
864 loff_t pos, int ioflags)
865{
866 struct nfs_pageio_descriptor desc;
867 struct inode *inode = dreq->inode;
868 struct nfs_commit_info cinfo;
869 ssize_t result = 0;
870 size_t requested_bytes = 0;
871 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
872 bool defer = false;
873
874 trace_nfs_direct_write_schedule_iovec(dreq);
875
876 nfs_pageio_init_write(&desc, inode, ioflags, false,
877 &nfs_direct_write_completion_ops);
878 desc.pg_dreq = dreq;
879 get_dreq(dreq);
880 inode_dio_begin(inode);
881
882 NFS_I(inode)->write_io += iov_iter_count(iter);
883 while (iov_iter_count(iter)) {
884 struct page **pagevec;
885 size_t bytes;
886 size_t pgbase;
887 unsigned npages, i;
888
889 result = iov_iter_get_pages_alloc2(iter, &pagevec,
890 wsize, &pgbase);
891 if (result < 0)
892 break;
893
894 bytes = result;
895 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
896 for (i = 0; i < npages; i++) {
897 struct nfs_page *req;
898 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
899
900 req = nfs_page_create_from_page(dreq->ctx, pagevec[i],
901 pgbase, pos, req_len);
902 if (IS_ERR(req)) {
903 result = PTR_ERR(req);
904 break;
905 }
906
907 if (desc.pg_error < 0) {
908 nfs_free_request(req);
909 result = desc.pg_error;
910 break;
911 }
912
913 pgbase = 0;
914 bytes -= req_len;
915 requested_bytes += req_len;
916 pos += req_len;
917
918 if (defer) {
919 nfs_mark_request_commit(req, NULL, &cinfo, 0);
920 continue;
921 }
922
923 nfs_lock_request(req);
924 if (nfs_pageio_add_request(&desc, req))
925 continue;
926
927 /* Exit on hard errors */
928 if (desc.pg_error < 0 && desc.pg_error != -EAGAIN) {
929 result = desc.pg_error;
930 nfs_unlock_and_release_request(req);
931 break;
932 }
933
934 /* If the error is soft, defer remaining requests */
935 nfs_init_cinfo_from_dreq(&cinfo, dreq);
936 spin_lock(&dreq->lock);
937 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
938 spin_unlock(&dreq->lock);
939 nfs_unlock_request(req);
940 nfs_mark_request_commit(req, NULL, &cinfo, 0);
941 desc.pg_error = 0;
942 defer = true;
943 }
944 nfs_direct_release_pages(pagevec, npages);
945 kvfree(pagevec);
946 if (result < 0)
947 break;
948 }
949 nfs_pageio_complete(&desc);
950
951 /*
952 * If no bytes were started, return the error, and let the
953 * generic layer handle the completion.
954 */
955 if (requested_bytes == 0) {
956 inode_dio_end(inode);
957 nfs_direct_req_release(dreq);
958 return result < 0 ? result : -EIO;
959 }
960
961 if (put_dreq(dreq))
962 nfs_direct_write_complete(dreq);
963 return requested_bytes;
964}
965
966/**
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
970 * @swap: flag indicating this is swap IO, not O_DIRECT IO
971 *
972 * We use this function for direct writes instead of calling
973 * generic_file_aio_write() in order to avoid taking the inode
974 * semaphore and updating the i_size. The NFS server will set
975 * the new i_size and this client must read the updated size
976 * back into its cache. We let the server do generic write
977 * parameter checking and report problems.
978 *
979 * We eliminate local atime updates, see direct read above.
980 *
981 * We avoid unnecessary page cache invalidations for normal cached
982 * readers of this file.
983 *
984 * Note that O_APPEND is not supported for NFS direct writes, as there
985 * is no atomic O_APPEND write facility in the NFS protocol.
986 */
987ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
988 bool swap)
989{
990 ssize_t result, requested;
991 size_t count;
992 struct file *file = iocb->ki_filp;
993 struct address_space *mapping = file->f_mapping;
994 struct inode *inode = mapping->host;
995 struct nfs_direct_req *dreq;
996 struct nfs_lock_context *l_ctx;
997 loff_t pos, end;
998
999 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
1000 file, iov_iter_count(iter), (long long) iocb->ki_pos);
1001
1002 if (swap)
1003 /* bypass generic checks */
1004 result = iov_iter_count(iter);
1005 else
1006 result = generic_write_checks(iocb, iter);
1007 if (result <= 0)
1008 return result;
1009 count = result;
1010 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
1011
1012 pos = iocb->ki_pos;
1013 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1014
1015 task_io_account_write(count);
1016
1017 result = -ENOMEM;
1018 dreq = nfs_direct_req_alloc();
1019 if (!dreq)
1020 goto out;
1021
1022 dreq->inode = inode;
1023 dreq->max_count = count;
1024 dreq->io_start = pos;
1025 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1026 l_ctx = nfs_get_lock_context(dreq->ctx);
1027 if (IS_ERR(l_ctx)) {
1028 result = PTR_ERR(l_ctx);
1029 nfs_direct_req_release(dreq);
1030 goto out_release;
1031 }
1032 dreq->l_ctx = l_ctx;
1033 if (!is_sync_kiocb(iocb))
1034 dreq->iocb = iocb;
1035 pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode);
1036
1037 if (swap) {
1038 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1039 FLUSH_STABLE);
1040 } else {
1041 result = nfs_start_io_direct(inode);
1042 if (result) {
1043 /* release the reference that would usually be
1044 * consumed by nfs_direct_write_schedule_iovec()
1045 */
1046 nfs_direct_req_release(dreq);
1047 goto out_release;
1048 }
1049
1050 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1051 FLUSH_COND_STABLE);
1052
1053 if (mapping->nrpages) {
1054 invalidate_inode_pages2_range(mapping,
1055 pos >> PAGE_SHIFT, end);
1056 }
1057
1058 nfs_end_io_direct(inode);
1059 }
1060
1061 if (requested > 0) {
1062 result = nfs_direct_wait(dreq);
1063 if (result > 0) {
1064 requested -= result;
1065 iocb->ki_pos = pos + result;
1066 /* XXX: should check the generic_write_sync retval */
1067 generic_write_sync(iocb, result);
1068 }
1069 iov_iter_revert(iter, requested);
1070 } else {
1071 result = requested;
1072 }
1073 nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE);
1074out_release:
1075 nfs_direct_req_release(dreq);
1076out:
1077 return result;
1078}
1079
1080/**
1081 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1082 *
1083 */
1084int __init nfs_init_directcache(void)
1085{
1086 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1087 sizeof(struct nfs_direct_req),
1088 0, SLAB_RECLAIM_ACCOUNT,
1089 NULL);
1090 if (nfs_direct_cachep == NULL)
1091 return -ENOMEM;
1092
1093 return 0;
1094}
1095
1096/**
1097 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1098 *
1099 */
1100void nfs_destroy_directcache(void)
1101{
1102 kmem_cache_destroy(nfs_direct_cachep);
1103}
1/*
2 * linux/fs/nfs/direct.c
3 *
4 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
5 *
6 * High-performance uncached I/O for the Linux NFS client
7 *
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.
16 *
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
22 * an application.
23 *
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.
28 *
29 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
30 * help from Andrew Morton.
31 *
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
38 *
39 */
40
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>
50
51#include <linux/nfs_fs.h>
52#include <linux/nfs_page.h>
53#include <linux/sunrpc/clnt.h>
54
55#include <linux/uaccess.h>
56#include <linux/atomic.h>
57
58#include "internal.h"
59#include "iostat.h"
60#include "pnfs.h"
61
62#define NFSDBG_FACILITY NFSDBG_VFS
63
64static struct kmem_cache *nfs_direct_cachep;
65
66/*
67 * This represents a set of asynchronous requests that we're waiting on
68 */
69struct nfs_direct_mirror {
70 ssize_t count;
71};
72
73struct nfs_direct_req {
74 struct kref kref; /* release manager */
75
76 /* I/O parameters */
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 */
81
82 /* completion state */
83 atomic_t io_count; /* i/os we're waiting for */
84 spinlock_t lock; /* protect completion state */
85
86 struct nfs_direct_mirror mirrors[NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX];
87 int mirror_count;
88
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 */
95
96 /* commit state */
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;
100 int flags;
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 */
104};
105
106static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
107static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
108static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
109static void nfs_direct_write_schedule_work(struct work_struct *work);
110
111static inline void get_dreq(struct nfs_direct_req *dreq)
112{
113 atomic_inc(&dreq->io_count);
114}
115
116static inline int put_dreq(struct nfs_direct_req *dreq)
117{
118 return atomic_dec_and_test(&dreq->io_count);
119}
120
121static void
122nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
123{
124 int i;
125 ssize_t count;
126
127 WARN_ON_ONCE(dreq->count >= dreq->max_count);
128
129 if (dreq->mirror_count == 1) {
130 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
131 dreq->count += hdr->good_bytes;
132 } else {
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;
138 }
139 /* update the dreq->count by finding the minimum agreed count from all
140 * mirrors */
141 count = dreq->mirrors[0].count;
142
143 for (i = 1; i < dreq->mirror_count; i++)
144 count = min(count, dreq->mirrors[i].count);
145
146 dreq->count = count;
147 }
148}
149
150/*
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
155 *
156 * returns the correct verifier to use given the role of the server
157 */
158static struct nfs_writeverf *
159nfs_direct_select_verf(struct nfs_direct_req *dreq,
160 struct nfs_client *ds_clp,
161 int commit_idx)
162{
163 struct nfs_writeverf *verfp = &dreq->verf;
164
165#ifdef CONFIG_NFS_V4_1
166 /*
167 * pNFS is in use, use the DS verf except commit_through_mds is set
168 * for layout segment where nbuckets is zero.
169 */
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;
173 else
174 WARN_ON_ONCE(1);
175 }
176#endif
177 return verfp;
178}
179
180
181/*
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
185 *
186 * Set the server's (MDS or DS) "seen" verifier
187 */
188static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
189 struct nfs_pgio_header *hdr)
190{
191 struct nfs_writeverf *verfp;
192
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);
197}
198
199static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
200 const struct nfs_writeverf *v2)
201{
202 return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
203}
204
205/*
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
209 *
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)
213 */
214static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
215 struct nfs_pgio_header *hdr)
216{
217 struct nfs_writeverf *verfp;
218
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);
222 return 0;
223 }
224 return nfs_direct_cmp_verf(verfp, &hdr->verf);
225}
226
227/*
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
231 *
232 * returns result of comparison between @data->verf and the verf of
233 * the server used by @data (DS or MDS)
234 */
235static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
236 struct nfs_commit_data *data)
237{
238 struct nfs_writeverf *verfp;
239
240 verfp = nfs_direct_select_verf(dreq, data->ds_clp,
241 data->ds_commit_index);
242
243 /* verifier not set so always fail */
244 if (verfp->committed < 0)
245 return 1;
246
247 return nfs_direct_cmp_verf(verfp, &data->verf);
248}
249
250/**
251 * nfs_direct_IO - NFS address space operation for direct I/O
252 * @iocb: target I/O control block
253 * @iter: I/O buffer
254 *
255 * The presence of this routine in the address space ops vector means
256 * the NFS client supports direct I/O. However, for most direct IO, we
257 * shunt off direct read and write requests before the VFS gets them,
258 * so this method is only ever called for swap.
259 */
260ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
261{
262 struct inode *inode = iocb->ki_filp->f_mapping->host;
263
264 /* we only support swap file calling nfs_direct_IO */
265 if (!IS_SWAPFILE(inode))
266 return 0;
267
268 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
269
270 if (iov_iter_rw(iter) == READ)
271 return nfs_file_direct_read(iocb, iter);
272 return nfs_file_direct_write(iocb, iter);
273}
274
275static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
276{
277 unsigned int i;
278 for (i = 0; i < npages; i++)
279 put_page(pages[i]);
280}
281
282void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
283 struct nfs_direct_req *dreq)
284{
285 cinfo->inode = dreq->inode;
286 cinfo->mds = &dreq->mds_cinfo;
287 cinfo->ds = &dreq->ds_cinfo;
288 cinfo->dreq = dreq;
289 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
290}
291
292static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
293 struct nfs_pageio_descriptor *pgio,
294 struct nfs_page *req)
295{
296 int mirror_count = 1;
297
298 if (pgio->pg_ops->pg_get_mirror_count)
299 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
300
301 dreq->mirror_count = mirror_count;
302}
303
304static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
305{
306 struct nfs_direct_req *dreq;
307
308 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
309 if (!dreq)
310 return NULL;
311
312 kref_init(&dreq->kref);
313 kref_get(&dreq->kref);
314 init_completion(&dreq->completion);
315 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
316 dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */
317 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
318 dreq->mirror_count = 1;
319 spin_lock_init(&dreq->lock);
320
321 return dreq;
322}
323
324static void nfs_direct_req_free(struct kref *kref)
325{
326 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
327
328 nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
329 if (dreq->l_ctx != NULL)
330 nfs_put_lock_context(dreq->l_ctx);
331 if (dreq->ctx != NULL)
332 put_nfs_open_context(dreq->ctx);
333 kmem_cache_free(nfs_direct_cachep, dreq);
334}
335
336static void nfs_direct_req_release(struct nfs_direct_req *dreq)
337{
338 kref_put(&dreq->kref, nfs_direct_req_free);
339}
340
341ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
342{
343 return dreq->bytes_left;
344}
345EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
346
347/*
348 * Collects and returns the final error value/byte-count.
349 */
350static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
351{
352 ssize_t result = -EIOCBQUEUED;
353
354 /* Async requests don't wait here */
355 if (dreq->iocb)
356 goto out;
357
358 result = wait_for_completion_killable(&dreq->completion);
359
360 if (!result) {
361 result = dreq->count;
362 WARN_ON_ONCE(dreq->count < 0);
363 }
364 if (!result)
365 result = dreq->error;
366
367out:
368 return (ssize_t) result;
369}
370
371/*
372 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
373 * the iocb is still valid here if this is a synchronous request.
374 */
375static void nfs_direct_complete(struct nfs_direct_req *dreq)
376{
377 struct inode *inode = dreq->inode;
378
379 inode_dio_end(inode);
380
381 if (dreq->iocb) {
382 long res = (long) dreq->error;
383 if (dreq->count != 0) {
384 res = (long) dreq->count;
385 WARN_ON_ONCE(dreq->count < 0);
386 }
387 dreq->iocb->ki_complete(dreq->iocb, res, 0);
388 }
389
390 complete(&dreq->completion);
391
392 nfs_direct_req_release(dreq);
393}
394
395static void nfs_direct_readpage_release(struct nfs_page *req)
396{
397 dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
398 req->wb_context->dentry->d_sb->s_id,
399 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
400 req->wb_bytes,
401 (long long)req_offset(req));
402 nfs_release_request(req);
403}
404
405static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
406{
407 unsigned long bytes = 0;
408 struct nfs_direct_req *dreq = hdr->dreq;
409
410 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
411 goto out_put;
412
413 spin_lock(&dreq->lock);
414 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
415 dreq->error = hdr->error;
416 else
417 nfs_direct_good_bytes(dreq, hdr);
418
419 spin_unlock(&dreq->lock);
420
421 while (!list_empty(&hdr->pages)) {
422 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
423 struct page *page = req->wb_page;
424
425 if (!PageCompound(page) && bytes < hdr->good_bytes)
426 set_page_dirty(page);
427 bytes += req->wb_bytes;
428 nfs_list_remove_request(req);
429 nfs_direct_readpage_release(req);
430 }
431out_put:
432 if (put_dreq(dreq))
433 nfs_direct_complete(dreq);
434 hdr->release(hdr);
435}
436
437static void nfs_read_sync_pgio_error(struct list_head *head)
438{
439 struct nfs_page *req;
440
441 while (!list_empty(head)) {
442 req = nfs_list_entry(head->next);
443 nfs_list_remove_request(req);
444 nfs_release_request(req);
445 }
446}
447
448static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
449{
450 get_dreq(hdr->dreq);
451}
452
453static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
454 .error_cleanup = nfs_read_sync_pgio_error,
455 .init_hdr = nfs_direct_pgio_init,
456 .completion = nfs_direct_read_completion,
457};
458
459/*
460 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
461 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
462 * bail and stop sending more reads. Read length accounting is
463 * handled automatically by nfs_direct_read_result(). Otherwise, if
464 * no requests have been sent, just return an error.
465 */
466
467static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
468 struct iov_iter *iter,
469 loff_t pos)
470{
471 struct nfs_pageio_descriptor desc;
472 struct inode *inode = dreq->inode;
473 ssize_t result = -EINVAL;
474 size_t requested_bytes = 0;
475 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
476
477 nfs_pageio_init_read(&desc, dreq->inode, false,
478 &nfs_direct_read_completion_ops);
479 get_dreq(dreq);
480 desc.pg_dreq = dreq;
481 inode_dio_begin(inode);
482
483 while (iov_iter_count(iter)) {
484 struct page **pagevec;
485 size_t bytes;
486 size_t pgbase;
487 unsigned npages, i;
488
489 result = iov_iter_get_pages_alloc(iter, &pagevec,
490 rsize, &pgbase);
491 if (result < 0)
492 break;
493
494 bytes = result;
495 iov_iter_advance(iter, bytes);
496 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
497 for (i = 0; i < npages; i++) {
498 struct nfs_page *req;
499 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
500 /* XXX do we need to do the eof zeroing found in async_filler? */
501 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
502 pgbase, req_len);
503 if (IS_ERR(req)) {
504 result = PTR_ERR(req);
505 break;
506 }
507 req->wb_index = pos >> PAGE_SHIFT;
508 req->wb_offset = pos & ~PAGE_MASK;
509 if (!nfs_pageio_add_request(&desc, req)) {
510 result = desc.pg_error;
511 nfs_release_request(req);
512 break;
513 }
514 pgbase = 0;
515 bytes -= req_len;
516 requested_bytes += req_len;
517 pos += req_len;
518 dreq->bytes_left -= req_len;
519 }
520 nfs_direct_release_pages(pagevec, npages);
521 kvfree(pagevec);
522 if (result < 0)
523 break;
524 }
525
526 nfs_pageio_complete(&desc);
527
528 /*
529 * If no bytes were started, return the error, and let the
530 * generic layer handle the completion.
531 */
532 if (requested_bytes == 0) {
533 inode_dio_end(inode);
534 nfs_direct_req_release(dreq);
535 return result < 0 ? result : -EIO;
536 }
537
538 if (put_dreq(dreq))
539 nfs_direct_complete(dreq);
540 return 0;
541}
542
543/**
544 * nfs_file_direct_read - file direct read operation for NFS files
545 * @iocb: target I/O control block
546 * @iter: vector of user buffers into which to read data
547 *
548 * We use this function for direct reads instead of calling
549 * generic_file_aio_read() in order to avoid gfar's check to see if
550 * the request starts before the end of the file. For that check
551 * to work, we must generate a GETATTR before each direct read, and
552 * even then there is a window between the GETATTR and the subsequent
553 * READ where the file size could change. Our preference is simply
554 * to do all reads the application wants, and the server will take
555 * care of managing the end of file boundary.
556 *
557 * This function also eliminates unnecessarily updating the file's
558 * atime locally, as the NFS server sets the file's atime, and this
559 * client must read the updated atime from the server back into its
560 * cache.
561 */
562ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
563{
564 struct file *file = iocb->ki_filp;
565 struct address_space *mapping = file->f_mapping;
566 struct inode *inode = mapping->host;
567 struct nfs_direct_req *dreq;
568 struct nfs_lock_context *l_ctx;
569 ssize_t result = -EINVAL;
570 size_t count = iov_iter_count(iter);
571 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
572
573 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
574 file, count, (long long) iocb->ki_pos);
575
576 result = 0;
577 if (!count)
578 goto out;
579
580 task_io_account_read(count);
581
582 result = -ENOMEM;
583 dreq = nfs_direct_req_alloc();
584 if (dreq == NULL)
585 goto out;
586
587 dreq->inode = inode;
588 dreq->bytes_left = dreq->max_count = count;
589 dreq->io_start = iocb->ki_pos;
590 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
591 l_ctx = nfs_get_lock_context(dreq->ctx);
592 if (IS_ERR(l_ctx)) {
593 result = PTR_ERR(l_ctx);
594 goto out_release;
595 }
596 dreq->l_ctx = l_ctx;
597 if (!is_sync_kiocb(iocb))
598 dreq->iocb = iocb;
599
600 nfs_start_io_direct(inode);
601
602 NFS_I(inode)->read_io += count;
603 result = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
604
605 nfs_end_io_direct(inode);
606
607 if (!result) {
608 result = nfs_direct_wait(dreq);
609 if (result > 0)
610 iocb->ki_pos += result;
611 }
612
613out_release:
614 nfs_direct_req_release(dreq);
615out:
616 return result;
617}
618
619static void
620nfs_direct_write_scan_commit_list(struct inode *inode,
621 struct list_head *list,
622 struct nfs_commit_info *cinfo)
623{
624 spin_lock(&cinfo->inode->i_lock);
625#ifdef CONFIG_NFS_V4_1
626 if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
627 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
628#endif
629 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
630 spin_unlock(&cinfo->inode->i_lock);
631}
632
633static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
634{
635 struct nfs_pageio_descriptor desc;
636 struct nfs_page *req, *tmp;
637 LIST_HEAD(reqs);
638 struct nfs_commit_info cinfo;
639 LIST_HEAD(failed);
640 int i;
641
642 nfs_init_cinfo_from_dreq(&cinfo, dreq);
643 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
644
645 dreq->count = 0;
646 dreq->verf.committed = NFS_INVALID_STABLE_HOW;
647 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
648 for (i = 0; i < dreq->mirror_count; i++)
649 dreq->mirrors[i].count = 0;
650 get_dreq(dreq);
651
652 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
653 &nfs_direct_write_completion_ops);
654 desc.pg_dreq = dreq;
655
656 req = nfs_list_entry(reqs.next);
657 nfs_direct_setup_mirroring(dreq, &desc, req);
658 if (desc.pg_error < 0) {
659 list_splice_init(&reqs, &failed);
660 goto out_failed;
661 }
662
663 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
664 if (!nfs_pageio_add_request(&desc, req)) {
665 nfs_list_remove_request(req);
666 nfs_list_add_request(req, &failed);
667 spin_lock(&cinfo.inode->i_lock);
668 dreq->flags = 0;
669 if (desc.pg_error < 0)
670 dreq->error = desc.pg_error;
671 else
672 dreq->error = -EIO;
673 spin_unlock(&cinfo.inode->i_lock);
674 }
675 nfs_release_request(req);
676 }
677 nfs_pageio_complete(&desc);
678
679out_failed:
680 while (!list_empty(&failed)) {
681 req = nfs_list_entry(failed.next);
682 nfs_list_remove_request(req);
683 nfs_unlock_and_release_request(req);
684 }
685
686 if (put_dreq(dreq))
687 nfs_direct_write_complete(dreq);
688}
689
690static void nfs_direct_commit_complete(struct nfs_commit_data *data)
691{
692 struct nfs_direct_req *dreq = data->dreq;
693 struct nfs_commit_info cinfo;
694 struct nfs_page *req;
695 int status = data->task.tk_status;
696
697 nfs_init_cinfo_from_dreq(&cinfo, dreq);
698 if (status < 0) {
699 dprintk("NFS: %5u commit failed with error %d.\n",
700 data->task.tk_pid, status);
701 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
702 } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
703 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
704 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
705 }
706
707 dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
708 while (!list_empty(&data->pages)) {
709 req = nfs_list_entry(data->pages.next);
710 nfs_list_remove_request(req);
711 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
712 /* Note the rewrite will go through mds */
713 nfs_mark_request_commit(req, NULL, &cinfo, 0);
714 } else
715 nfs_release_request(req);
716 nfs_unlock_and_release_request(req);
717 }
718
719 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
720 nfs_direct_write_complete(dreq);
721}
722
723static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
724 struct nfs_page *req)
725{
726 struct nfs_direct_req *dreq = cinfo->dreq;
727
728 spin_lock(&dreq->lock);
729 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
730 spin_unlock(&dreq->lock);
731 nfs_mark_request_commit(req, NULL, cinfo, 0);
732}
733
734static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
735 .completion = nfs_direct_commit_complete,
736 .resched_write = nfs_direct_resched_write,
737};
738
739static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
740{
741 int res;
742 struct nfs_commit_info cinfo;
743 LIST_HEAD(mds_list);
744
745 nfs_init_cinfo_from_dreq(&cinfo, dreq);
746 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
747 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
748 if (res < 0) /* res == -ENOMEM */
749 nfs_direct_write_reschedule(dreq);
750}
751
752static void nfs_direct_write_schedule_work(struct work_struct *work)
753{
754 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
755 int flags = dreq->flags;
756
757 dreq->flags = 0;
758 switch (flags) {
759 case NFS_ODIRECT_DO_COMMIT:
760 nfs_direct_commit_schedule(dreq);
761 break;
762 case NFS_ODIRECT_RESCHED_WRITES:
763 nfs_direct_write_reschedule(dreq);
764 break;
765 default:
766 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
767 nfs_direct_complete(dreq);
768 }
769}
770
771static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
772{
773 schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
774}
775
776static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
777{
778 struct nfs_direct_req *dreq = hdr->dreq;
779 struct nfs_commit_info cinfo;
780 bool request_commit = false;
781 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
782
783 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
784 goto out_put;
785
786 nfs_init_cinfo_from_dreq(&cinfo, dreq);
787
788 spin_lock(&dreq->lock);
789
790 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
791 dreq->flags = 0;
792 dreq->error = hdr->error;
793 }
794 if (dreq->error == 0) {
795 nfs_direct_good_bytes(dreq, hdr);
796 if (nfs_write_need_commit(hdr)) {
797 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
798 request_commit = true;
799 else if (dreq->flags == 0) {
800 nfs_direct_set_hdr_verf(dreq, hdr);
801 request_commit = true;
802 dreq->flags = NFS_ODIRECT_DO_COMMIT;
803 } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
804 request_commit = true;
805 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
806 dreq->flags =
807 NFS_ODIRECT_RESCHED_WRITES;
808 }
809 }
810 }
811 spin_unlock(&dreq->lock);
812
813 while (!list_empty(&hdr->pages)) {
814
815 req = nfs_list_entry(hdr->pages.next);
816 nfs_list_remove_request(req);
817 if (request_commit) {
818 kref_get(&req->wb_kref);
819 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
820 hdr->ds_commit_idx);
821 }
822 nfs_unlock_and_release_request(req);
823 }
824
825out_put:
826 if (put_dreq(dreq))
827 nfs_direct_write_complete(dreq);
828 hdr->release(hdr);
829}
830
831static void nfs_write_sync_pgio_error(struct list_head *head)
832{
833 struct nfs_page *req;
834
835 while (!list_empty(head)) {
836 req = nfs_list_entry(head->next);
837 nfs_list_remove_request(req);
838 nfs_unlock_and_release_request(req);
839 }
840}
841
842static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
843{
844 struct nfs_direct_req *dreq = hdr->dreq;
845
846 spin_lock(&dreq->lock);
847 if (dreq->error == 0) {
848 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
849 /* fake unstable write to let common nfs resend pages */
850 hdr->verf.committed = NFS_UNSTABLE;
851 hdr->good_bytes = hdr->args.count;
852 }
853 spin_unlock(&dreq->lock);
854}
855
856static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
857 .error_cleanup = nfs_write_sync_pgio_error,
858 .init_hdr = nfs_direct_pgio_init,
859 .completion = nfs_direct_write_completion,
860 .reschedule_io = nfs_direct_write_reschedule_io,
861};
862
863
864/*
865 * NB: Return the value of the first error return code. Subsequent
866 * errors after the first one are ignored.
867 */
868/*
869 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
870 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
871 * bail and stop sending more writes. Write length accounting is
872 * handled automatically by nfs_direct_write_result(). Otherwise, if
873 * no requests have been sent, just return an error.
874 */
875static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
876 struct iov_iter *iter,
877 loff_t pos)
878{
879 struct nfs_pageio_descriptor desc;
880 struct inode *inode = dreq->inode;
881 ssize_t result = 0;
882 size_t requested_bytes = 0;
883 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
884
885 nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
886 &nfs_direct_write_completion_ops);
887 desc.pg_dreq = dreq;
888 get_dreq(dreq);
889 inode_dio_begin(inode);
890
891 NFS_I(inode)->write_io += iov_iter_count(iter);
892 while (iov_iter_count(iter)) {
893 struct page **pagevec;
894 size_t bytes;
895 size_t pgbase;
896 unsigned npages, i;
897
898 result = iov_iter_get_pages_alloc(iter, &pagevec,
899 wsize, &pgbase);
900 if (result < 0)
901 break;
902
903 bytes = result;
904 iov_iter_advance(iter, bytes);
905 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
906 for (i = 0; i < npages; i++) {
907 struct nfs_page *req;
908 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
909
910 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
911 pgbase, req_len);
912 if (IS_ERR(req)) {
913 result = PTR_ERR(req);
914 break;
915 }
916
917 nfs_direct_setup_mirroring(dreq, &desc, req);
918 if (desc.pg_error < 0) {
919 nfs_free_request(req);
920 result = desc.pg_error;
921 break;
922 }
923
924 nfs_lock_request(req);
925 req->wb_index = pos >> PAGE_SHIFT;
926 req->wb_offset = pos & ~PAGE_MASK;
927 if (!nfs_pageio_add_request(&desc, req)) {
928 result = desc.pg_error;
929 nfs_unlock_and_release_request(req);
930 break;
931 }
932 pgbase = 0;
933 bytes -= req_len;
934 requested_bytes += req_len;
935 pos += req_len;
936 dreq->bytes_left -= req_len;
937 }
938 nfs_direct_release_pages(pagevec, npages);
939 kvfree(pagevec);
940 if (result < 0)
941 break;
942 }
943 nfs_pageio_complete(&desc);
944
945 /*
946 * If no bytes were started, return the error, and let the
947 * generic layer handle the completion.
948 */
949 if (requested_bytes == 0) {
950 inode_dio_end(inode);
951 nfs_direct_req_release(dreq);
952 return result < 0 ? result : -EIO;
953 }
954
955 if (put_dreq(dreq))
956 nfs_direct_write_complete(dreq);
957 return 0;
958}
959
960/**
961 * nfs_file_direct_write - file direct write operation for NFS files
962 * @iocb: target I/O control block
963 * @iter: vector of user buffers from which to write data
964 *
965 * We use this function for direct writes instead of calling
966 * generic_file_aio_write() in order to avoid taking the inode
967 * semaphore and updating the i_size. The NFS server will set
968 * the new i_size and this client must read the updated size
969 * back into its cache. We let the server do generic write
970 * parameter checking and report problems.
971 *
972 * We eliminate local atime updates, see direct read above.
973 *
974 * We avoid unnecessary page cache invalidations for normal cached
975 * readers of this file.
976 *
977 * Note that O_APPEND is not supported for NFS direct writes, as there
978 * is no atomic O_APPEND write facility in the NFS protocol.
979 */
980ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
981{
982 ssize_t result = -EINVAL;
983 size_t count;
984 struct file *file = iocb->ki_filp;
985 struct address_space *mapping = file->f_mapping;
986 struct inode *inode = mapping->host;
987 struct nfs_direct_req *dreq;
988 struct nfs_lock_context *l_ctx;
989 loff_t pos, end;
990
991 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
992 file, iov_iter_count(iter), (long long) iocb->ki_pos);
993
994 result = generic_write_checks(iocb, iter);
995 if (result <= 0)
996 return result;
997 count = result;
998 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
999
1000 pos = iocb->ki_pos;
1001 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1002
1003 task_io_account_write(count);
1004
1005 result = -ENOMEM;
1006 dreq = nfs_direct_req_alloc();
1007 if (!dreq)
1008 goto out;
1009
1010 dreq->inode = inode;
1011 dreq->bytes_left = dreq->max_count = count;
1012 dreq->io_start = pos;
1013 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1014 l_ctx = nfs_get_lock_context(dreq->ctx);
1015 if (IS_ERR(l_ctx)) {
1016 result = PTR_ERR(l_ctx);
1017 goto out_release;
1018 }
1019 dreq->l_ctx = l_ctx;
1020 if (!is_sync_kiocb(iocb))
1021 dreq->iocb = iocb;
1022
1023 nfs_start_io_direct(inode);
1024
1025 result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1026
1027 if (mapping->nrpages) {
1028 invalidate_inode_pages2_range(mapping,
1029 pos >> PAGE_SHIFT, end);
1030 }
1031
1032 nfs_end_io_direct(inode);
1033
1034 if (!result) {
1035 result = nfs_direct_wait(dreq);
1036 if (result > 0) {
1037 iocb->ki_pos = pos + result;
1038 /* XXX: should check the generic_write_sync retval */
1039 generic_write_sync(iocb, result);
1040 }
1041 }
1042out_release:
1043 nfs_direct_req_release(dreq);
1044out:
1045 return result;
1046}
1047
1048/**
1049 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1050 *
1051 */
1052int __init nfs_init_directcache(void)
1053{
1054 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1055 sizeof(struct nfs_direct_req),
1056 0, (SLAB_RECLAIM_ACCOUNT|
1057 SLAB_MEM_SPREAD),
1058 NULL);
1059 if (nfs_direct_cachep == NULL)
1060 return -ENOMEM;
1061
1062 return 0;
1063}
1064
1065/**
1066 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1067 *
1068 */
1069void nfs_destroy_directcache(void)
1070{
1071 kmem_cache_destroy(nfs_direct_cachep);
1072}