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