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