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 struct nfs_writeverf verf; /* unstable write verifier */
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_select_verf - select the right verifier
156 * @dreq - direct request possibly spanning multiple servers
157 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
158 * @commit_idx - commit bucket index for the DS
159 *
160 * returns the correct verifier to use given the role of the server
161 */
162static struct nfs_writeverf *
163nfs_direct_select_verf(struct nfs_direct_req *dreq,
164 struct nfs_client *ds_clp,
165 int commit_idx)
166{
167 struct nfs_writeverf *verfp = &dreq->verf;
168
169#ifdef CONFIG_NFS_V4_1
170 /*
171 * pNFS is in use, use the DS verf except commit_through_mds is set
172 * for layout segment where nbuckets is zero.
173 */
174 if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
175 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
176 verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
177 else
178 WARN_ON_ONCE(1);
179 }
180#endif
181 return verfp;
182}
183
184
185/*
186 * nfs_direct_set_hdr_verf - set the write/commit verifier
187 * @dreq - direct request possibly spanning multiple servers
188 * @hdr - pageio header to validate against previously seen verfs
189 *
190 * Set the server's (MDS or DS) "seen" verifier
191 */
192static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
193 struct nfs_pgio_header *hdr)
194{
195 struct nfs_writeverf *verfp;
196
197 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
198 WARN_ON_ONCE(verfp->committed >= 0);
199 memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
200 WARN_ON_ONCE(verfp->committed < 0);
201}
202
203static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
204 const struct nfs_writeverf *v2)
205{
206 return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
207}
208
209/*
210 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
211 * @dreq - direct request possibly spanning multiple servers
212 * @hdr - pageio header to validate against previously seen verf
213 *
214 * set the server's "seen" verf if not initialized.
215 * returns result of comparison between @hdr->verf and the "seen"
216 * verf of the server used by @hdr (DS or MDS)
217 */
218static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
219 struct nfs_pgio_header *hdr)
220{
221 struct nfs_writeverf *verfp;
222
223 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
224 if (verfp->committed < 0) {
225 nfs_direct_set_hdr_verf(dreq, hdr);
226 return 0;
227 }
228 return nfs_direct_cmp_verf(verfp, &hdr->verf);
229}
230
231/*
232 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
233 * @dreq - direct request possibly spanning multiple servers
234 * @data - commit data to validate against previously seen verf
235 *
236 * returns result of comparison between @data->verf and the verf of
237 * the server used by @data (DS or MDS)
238 */
239static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
240 struct nfs_commit_data *data)
241{
242 struct nfs_writeverf *verfp;
243
244 verfp = nfs_direct_select_verf(dreq, data->ds_clp,
245 data->ds_commit_index);
246
247 /* verifier not set so always fail */
248 if (verfp->committed < 0)
249 return 1;
250
251 return nfs_direct_cmp_verf(verfp, &data->verf);
252}
253
254/**
255 * nfs_direct_IO - NFS address space operation for direct I/O
256 * @iocb: target I/O control block
257 * @iter: I/O buffer
258 *
259 * The presence of this routine in the address space ops vector means
260 * the NFS client supports direct I/O. However, for most direct IO, we
261 * shunt off direct read and write requests before the VFS gets them,
262 * so this method is only ever called for swap.
263 */
264ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
265{
266 struct inode *inode = iocb->ki_filp->f_mapping->host;
267
268 /* we only support swap file calling nfs_direct_IO */
269 if (!IS_SWAPFILE(inode))
270 return 0;
271
272 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
273
274 if (iov_iter_rw(iter) == READ)
275 return nfs_file_direct_read(iocb, iter);
276 return nfs_file_direct_write(iocb, iter);
277}
278
279static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
280{
281 unsigned int i;
282 for (i = 0; i < npages; i++)
283 put_page(pages[i]);
284}
285
286void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
287 struct nfs_direct_req *dreq)
288{
289 cinfo->inode = dreq->inode;
290 cinfo->mds = &dreq->mds_cinfo;
291 cinfo->ds = &dreq->ds_cinfo;
292 cinfo->dreq = dreq;
293 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
294}
295
296static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
297{
298 struct nfs_direct_req *dreq;
299
300 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
301 if (!dreq)
302 return NULL;
303
304 kref_init(&dreq->kref);
305 kref_get(&dreq->kref);
306 init_completion(&dreq->completion);
307 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
308 dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */
309 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
310 spin_lock_init(&dreq->lock);
311
312 return dreq;
313}
314
315static void nfs_direct_req_free(struct kref *kref)
316{
317 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
318
319 nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
320 if (dreq->l_ctx != NULL)
321 nfs_put_lock_context(dreq->l_ctx);
322 if (dreq->ctx != NULL)
323 put_nfs_open_context(dreq->ctx);
324 kmem_cache_free(nfs_direct_cachep, dreq);
325}
326
327static void nfs_direct_req_release(struct nfs_direct_req *dreq)
328{
329 kref_put(&dreq->kref, nfs_direct_req_free);
330}
331
332ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
333{
334 return dreq->bytes_left;
335}
336EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
337
338/*
339 * Collects and returns the final error value/byte-count.
340 */
341static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
342{
343 ssize_t result = -EIOCBQUEUED;
344
345 /* Async requests don't wait here */
346 if (dreq->iocb)
347 goto out;
348
349 result = wait_for_completion_killable(&dreq->completion);
350
351 if (!result) {
352 result = dreq->count;
353 WARN_ON_ONCE(dreq->count < 0);
354 }
355 if (!result)
356 result = dreq->error;
357
358out:
359 return (ssize_t) result;
360}
361
362/*
363 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
364 * the iocb is still valid here if this is a synchronous request.
365 */
366static void nfs_direct_complete(struct nfs_direct_req *dreq)
367{
368 struct inode *inode = dreq->inode;
369
370 inode_dio_end(inode);
371
372 if (dreq->iocb) {
373 long res = (long) dreq->error;
374 if (dreq->count != 0) {
375 res = (long) dreq->count;
376 WARN_ON_ONCE(dreq->count < 0);
377 }
378 dreq->iocb->ki_complete(dreq->iocb, res, 0);
379 }
380
381 complete(&dreq->completion);
382
383 nfs_direct_req_release(dreq);
384}
385
386static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
387{
388 unsigned long bytes = 0;
389 struct nfs_direct_req *dreq = hdr->dreq;
390
391 spin_lock(&dreq->lock);
392 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
393 spin_unlock(&dreq->lock);
394 goto out_put;
395 }
396
397 nfs_direct_count_bytes(dreq, hdr);
398 spin_unlock(&dreq->lock);
399
400 while (!list_empty(&hdr->pages)) {
401 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
402 struct page *page = req->wb_page;
403
404 if (!PageCompound(page) && bytes < hdr->good_bytes &&
405 (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
406 set_page_dirty(page);
407 bytes += req->wb_bytes;
408 nfs_list_remove_request(req);
409 nfs_release_request(req);
410 }
411out_put:
412 if (put_dreq(dreq))
413 nfs_direct_complete(dreq);
414 hdr->release(hdr);
415}
416
417static void nfs_read_sync_pgio_error(struct list_head *head, int error)
418{
419 struct nfs_page *req;
420
421 while (!list_empty(head)) {
422 req = nfs_list_entry(head->next);
423 nfs_list_remove_request(req);
424 nfs_release_request(req);
425 }
426}
427
428static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
429{
430 get_dreq(hdr->dreq);
431}
432
433static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
434 .error_cleanup = nfs_read_sync_pgio_error,
435 .init_hdr = nfs_direct_pgio_init,
436 .completion = nfs_direct_read_completion,
437};
438
439/*
440 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
441 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
442 * bail and stop sending more reads. Read length accounting is
443 * handled automatically by nfs_direct_read_result(). Otherwise, if
444 * no requests have been sent, just return an error.
445 */
446
447static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
448 struct iov_iter *iter,
449 loff_t pos)
450{
451 struct nfs_pageio_descriptor desc;
452 struct inode *inode = dreq->inode;
453 ssize_t result = -EINVAL;
454 size_t requested_bytes = 0;
455 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
456
457 nfs_pageio_init_read(&desc, dreq->inode, false,
458 &nfs_direct_read_completion_ops);
459 get_dreq(dreq);
460 desc.pg_dreq = dreq;
461 inode_dio_begin(inode);
462
463 while (iov_iter_count(iter)) {
464 struct page **pagevec;
465 size_t bytes;
466 size_t pgbase;
467 unsigned npages, i;
468
469 result = iov_iter_get_pages_alloc(iter, &pagevec,
470 rsize, &pgbase);
471 if (result < 0)
472 break;
473
474 bytes = result;
475 iov_iter_advance(iter, bytes);
476 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
477 for (i = 0; i < npages; i++) {
478 struct nfs_page *req;
479 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
480 /* XXX do we need to do the eof zeroing found in async_filler? */
481 req = nfs_create_request(dreq->ctx, pagevec[i],
482 pgbase, req_len);
483 if (IS_ERR(req)) {
484 result = PTR_ERR(req);
485 break;
486 }
487 req->wb_index = pos >> PAGE_SHIFT;
488 req->wb_offset = pos & ~PAGE_MASK;
489 if (!nfs_pageio_add_request(&desc, req)) {
490 result = desc.pg_error;
491 nfs_release_request(req);
492 break;
493 }
494 pgbase = 0;
495 bytes -= req_len;
496 requested_bytes += req_len;
497 pos += req_len;
498 dreq->bytes_left -= req_len;
499 }
500 nfs_direct_release_pages(pagevec, npages);
501 kvfree(pagevec);
502 if (result < 0)
503 break;
504 }
505
506 nfs_pageio_complete(&desc);
507
508 /*
509 * If no bytes were started, return the error, and let the
510 * generic layer handle the completion.
511 */
512 if (requested_bytes == 0) {
513 inode_dio_end(inode);
514 nfs_direct_req_release(dreq);
515 return result < 0 ? result : -EIO;
516 }
517
518 if (put_dreq(dreq))
519 nfs_direct_complete(dreq);
520 return requested_bytes;
521}
522
523/**
524 * nfs_file_direct_read - file direct read operation for NFS files
525 * @iocb: target I/O control block
526 * @iter: vector of user buffers into which to read data
527 *
528 * We use this function for direct reads instead of calling
529 * generic_file_aio_read() in order to avoid gfar's check to see if
530 * the request starts before the end of the file. For that check
531 * to work, we must generate a GETATTR before each direct read, and
532 * even then there is a window between the GETATTR and the subsequent
533 * READ where the file size could change. Our preference is simply
534 * to do all reads the application wants, and the server will take
535 * care of managing the end of file boundary.
536 *
537 * This function also eliminates unnecessarily updating the file's
538 * atime locally, as the NFS server sets the file's atime, and this
539 * client must read the updated atime from the server back into its
540 * cache.
541 */
542ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
543{
544 struct file *file = iocb->ki_filp;
545 struct address_space *mapping = file->f_mapping;
546 struct inode *inode = mapping->host;
547 struct nfs_direct_req *dreq;
548 struct nfs_lock_context *l_ctx;
549 ssize_t result = -EINVAL, requested;
550 size_t count = iov_iter_count(iter);
551 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
552
553 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
554 file, count, (long long) iocb->ki_pos);
555
556 result = 0;
557 if (!count)
558 goto out;
559
560 task_io_account_read(count);
561
562 result = -ENOMEM;
563 dreq = nfs_direct_req_alloc();
564 if (dreq == NULL)
565 goto out;
566
567 dreq->inode = inode;
568 dreq->bytes_left = dreq->max_count = count;
569 dreq->io_start = iocb->ki_pos;
570 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
571 l_ctx = nfs_get_lock_context(dreq->ctx);
572 if (IS_ERR(l_ctx)) {
573 result = PTR_ERR(l_ctx);
574 goto out_release;
575 }
576 dreq->l_ctx = l_ctx;
577 if (!is_sync_kiocb(iocb))
578 dreq->iocb = iocb;
579
580 if (iter_is_iovec(iter))
581 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
582
583 nfs_start_io_direct(inode);
584
585 NFS_I(inode)->read_io += count;
586 requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
587
588 nfs_end_io_direct(inode);
589
590 if (requested > 0) {
591 result = nfs_direct_wait(dreq);
592 if (result > 0) {
593 requested -= result;
594 iocb->ki_pos += result;
595 }
596 iov_iter_revert(iter, requested);
597 } else {
598 result = requested;
599 }
600
601out_release:
602 nfs_direct_req_release(dreq);
603out:
604 return result;
605}
606
607static void
608nfs_direct_write_scan_commit_list(struct inode *inode,
609 struct list_head *list,
610 struct nfs_commit_info *cinfo)
611{
612 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
613#ifdef CONFIG_NFS_V4_1
614 if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
615 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
616#endif
617 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
618 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
619}
620
621static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
622{
623 struct nfs_pageio_descriptor desc;
624 struct nfs_page *req, *tmp;
625 LIST_HEAD(reqs);
626 struct nfs_commit_info cinfo;
627 LIST_HEAD(failed);
628
629 nfs_init_cinfo_from_dreq(&cinfo, dreq);
630 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
631
632 dreq->count = 0;
633 dreq->max_count = 0;
634 list_for_each_entry(req, &reqs, wb_list)
635 dreq->max_count += req->wb_bytes;
636 dreq->verf.committed = NFS_INVALID_STABLE_HOW;
637 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
638 get_dreq(dreq);
639
640 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
641 &nfs_direct_write_completion_ops);
642 desc.pg_dreq = dreq;
643
644 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
645 /* Bump the transmission count */
646 req->wb_nio++;
647 if (!nfs_pageio_add_request(&desc, req)) {
648 nfs_list_move_request(req, &failed);
649 spin_lock(&cinfo.inode->i_lock);
650 dreq->flags = 0;
651 if (desc.pg_error < 0)
652 dreq->error = desc.pg_error;
653 else
654 dreq->error = -EIO;
655 spin_unlock(&cinfo.inode->i_lock);
656 }
657 nfs_release_request(req);
658 }
659 nfs_pageio_complete(&desc);
660
661 while (!list_empty(&failed)) {
662 req = nfs_list_entry(failed.next);
663 nfs_list_remove_request(req);
664 nfs_unlock_and_release_request(req);
665 }
666
667 if (put_dreq(dreq))
668 nfs_direct_write_complete(dreq);
669}
670
671static void nfs_direct_commit_complete(struct nfs_commit_data *data)
672{
673 struct nfs_direct_req *dreq = data->dreq;
674 struct nfs_commit_info cinfo;
675 struct nfs_page *req;
676 int status = data->task.tk_status;
677
678 nfs_init_cinfo_from_dreq(&cinfo, dreq);
679 if (status < 0 || nfs_direct_cmp_commit_data_verf(dreq, data))
680 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
681
682 while (!list_empty(&data->pages)) {
683 req = nfs_list_entry(data->pages.next);
684 nfs_list_remove_request(req);
685 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
686 /*
687 * Despite the reboot, the write was successful,
688 * so reset wb_nio.
689 */
690 req->wb_nio = 0;
691 /* Note the rewrite will go through mds */
692 nfs_mark_request_commit(req, NULL, &cinfo, 0);
693 } else
694 nfs_release_request(req);
695 nfs_unlock_and_release_request(req);
696 }
697
698 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
699 nfs_direct_write_complete(dreq);
700}
701
702static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
703 struct nfs_page *req)
704{
705 struct nfs_direct_req *dreq = cinfo->dreq;
706
707 spin_lock(&dreq->lock);
708 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
709 spin_unlock(&dreq->lock);
710 nfs_mark_request_commit(req, NULL, cinfo, 0);
711}
712
713static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
714 .completion = nfs_direct_commit_complete,
715 .resched_write = nfs_direct_resched_write,
716};
717
718static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
719{
720 int res;
721 struct nfs_commit_info cinfo;
722 LIST_HEAD(mds_list);
723
724 nfs_init_cinfo_from_dreq(&cinfo, dreq);
725 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
726 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
727 if (res < 0) /* res == -ENOMEM */
728 nfs_direct_write_reschedule(dreq);
729}
730
731static void nfs_direct_write_schedule_work(struct work_struct *work)
732{
733 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
734 int flags = dreq->flags;
735
736 dreq->flags = 0;
737 switch (flags) {
738 case NFS_ODIRECT_DO_COMMIT:
739 nfs_direct_commit_schedule(dreq);
740 break;
741 case NFS_ODIRECT_RESCHED_WRITES:
742 nfs_direct_write_reschedule(dreq);
743 break;
744 default:
745 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
746 nfs_direct_complete(dreq);
747 }
748}
749
750static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
751{
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 bool request_commit = false;
760 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
761
762 nfs_init_cinfo_from_dreq(&cinfo, dreq);
763
764 spin_lock(&dreq->lock);
765 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
766 spin_unlock(&dreq->lock);
767 goto out_put;
768 }
769
770 nfs_direct_count_bytes(dreq, hdr);
771 if (hdr->good_bytes != 0) {
772 if (nfs_write_need_commit(hdr)) {
773 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
774 request_commit = true;
775 else if (dreq->flags == 0) {
776 nfs_direct_set_hdr_verf(dreq, hdr);
777 request_commit = true;
778 dreq->flags = NFS_ODIRECT_DO_COMMIT;
779 } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
780 request_commit = true;
781 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
782 dreq->flags =
783 NFS_ODIRECT_RESCHED_WRITES;
784 }
785 }
786 }
787 spin_unlock(&dreq->lock);
788
789 while (!list_empty(&hdr->pages)) {
790
791 req = nfs_list_entry(hdr->pages.next);
792 nfs_list_remove_request(req);
793 if (request_commit) {
794 kref_get(&req->wb_kref);
795 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
796 hdr->ds_commit_idx);
797 }
798 nfs_unlock_and_release_request(req);
799 }
800
801out_put:
802 if (put_dreq(dreq))
803 nfs_direct_write_complete(dreq);
804 hdr->release(hdr);
805}
806
807static void nfs_write_sync_pgio_error(struct list_head *head, int error)
808{
809 struct nfs_page *req;
810
811 while (!list_empty(head)) {
812 req = nfs_list_entry(head->next);
813 nfs_list_remove_request(req);
814 nfs_unlock_and_release_request(req);
815 }
816}
817
818static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
819{
820 struct nfs_direct_req *dreq = hdr->dreq;
821
822 spin_lock(&dreq->lock);
823 if (dreq->error == 0) {
824 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
825 /* fake unstable write to let common nfs resend pages */
826 hdr->verf.committed = NFS_UNSTABLE;
827 hdr->good_bytes = hdr->args.count;
828 }
829 spin_unlock(&dreq->lock);
830}
831
832static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
833 .error_cleanup = nfs_write_sync_pgio_error,
834 .init_hdr = nfs_direct_pgio_init,
835 .completion = nfs_direct_write_completion,
836 .reschedule_io = nfs_direct_write_reschedule_io,
837};
838
839
840/*
841 * NB: Return the value of the first error return code. Subsequent
842 * errors after the first one are ignored.
843 */
844/*
845 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
846 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
847 * bail and stop sending more writes. Write length accounting is
848 * handled automatically by nfs_direct_write_result(). Otherwise, if
849 * no requests have been sent, just return an error.
850 */
851static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
852 struct iov_iter *iter,
853 loff_t pos)
854{
855 struct nfs_pageio_descriptor desc;
856 struct inode *inode = dreq->inode;
857 ssize_t result = 0;
858 size_t requested_bytes = 0;
859 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
860
861 nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
862 &nfs_direct_write_completion_ops);
863 desc.pg_dreq = dreq;
864 get_dreq(dreq);
865 inode_dio_begin(inode);
866
867 NFS_I(inode)->write_io += iov_iter_count(iter);
868 while (iov_iter_count(iter)) {
869 struct page **pagevec;
870 size_t bytes;
871 size_t pgbase;
872 unsigned npages, i;
873
874 result = iov_iter_get_pages_alloc(iter, &pagevec,
875 wsize, &pgbase);
876 if (result < 0)
877 break;
878
879 bytes = result;
880 iov_iter_advance(iter, bytes);
881 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
882 for (i = 0; i < npages; i++) {
883 struct nfs_page *req;
884 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
885
886 req = nfs_create_request(dreq->ctx, pagevec[i],
887 pgbase, req_len);
888 if (IS_ERR(req)) {
889 result = PTR_ERR(req);
890 break;
891 }
892
893 if (desc.pg_error < 0) {
894 nfs_free_request(req);
895 result = desc.pg_error;
896 break;
897 }
898
899 nfs_lock_request(req);
900 req->wb_index = pos >> PAGE_SHIFT;
901 req->wb_offset = pos & ~PAGE_MASK;
902 if (!nfs_pageio_add_request(&desc, req)) {
903 result = desc.pg_error;
904 nfs_unlock_and_release_request(req);
905 break;
906 }
907 pgbase = 0;
908 bytes -= req_len;
909 requested_bytes += req_len;
910 pos += req_len;
911 dreq->bytes_left -= req_len;
912 }
913 nfs_direct_release_pages(pagevec, npages);
914 kvfree(pagevec);
915 if (result < 0)
916 break;
917 }
918 nfs_pageio_complete(&desc);
919
920 /*
921 * If no bytes were started, return the error, and let the
922 * generic layer handle the completion.
923 */
924 if (requested_bytes == 0) {
925 inode_dio_end(inode);
926 nfs_direct_req_release(dreq);
927 return result < 0 ? result : -EIO;
928 }
929
930 if (put_dreq(dreq))
931 nfs_direct_write_complete(dreq);
932 return requested_bytes;
933}
934
935/**
936 * nfs_file_direct_write - file direct write operation for NFS files
937 * @iocb: target I/O control block
938 * @iter: vector of user buffers from which to write data
939 *
940 * We use this function for direct writes instead of calling
941 * generic_file_aio_write() in order to avoid taking the inode
942 * semaphore and updating the i_size. The NFS server will set
943 * the new i_size and this client must read the updated size
944 * back into its cache. We let the server do generic write
945 * parameter checking and report problems.
946 *
947 * We eliminate local atime updates, see direct read above.
948 *
949 * We avoid unnecessary page cache invalidations for normal cached
950 * readers of this file.
951 *
952 * Note that O_APPEND is not supported for NFS direct writes, as there
953 * is no atomic O_APPEND write facility in the NFS protocol.
954 */
955ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
956{
957 ssize_t result = -EINVAL, requested;
958 size_t count;
959 struct file *file = iocb->ki_filp;
960 struct address_space *mapping = file->f_mapping;
961 struct inode *inode = mapping->host;
962 struct nfs_direct_req *dreq;
963 struct nfs_lock_context *l_ctx;
964 loff_t pos, end;
965
966 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
967 file, iov_iter_count(iter), (long long) iocb->ki_pos);
968
969 result = generic_write_checks(iocb, iter);
970 if (result <= 0)
971 return result;
972 count = result;
973 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
974
975 pos = iocb->ki_pos;
976 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
977
978 task_io_account_write(count);
979
980 result = -ENOMEM;
981 dreq = nfs_direct_req_alloc();
982 if (!dreq)
983 goto out;
984
985 dreq->inode = inode;
986 dreq->bytes_left = dreq->max_count = count;
987 dreq->io_start = pos;
988 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
989 l_ctx = nfs_get_lock_context(dreq->ctx);
990 if (IS_ERR(l_ctx)) {
991 result = PTR_ERR(l_ctx);
992 goto out_release;
993 }
994 dreq->l_ctx = l_ctx;
995 if (!is_sync_kiocb(iocb))
996 dreq->iocb = iocb;
997
998 nfs_start_io_direct(inode);
999
1000 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1001
1002 if (mapping->nrpages) {
1003 invalidate_inode_pages2_range(mapping,
1004 pos >> PAGE_SHIFT, end);
1005 }
1006
1007 nfs_end_io_direct(inode);
1008
1009 if (requested > 0) {
1010 result = nfs_direct_wait(dreq);
1011 if (result > 0) {
1012 requested -= result;
1013 iocb->ki_pos = pos + result;
1014 /* XXX: should check the generic_write_sync retval */
1015 generic_write_sync(iocb, result);
1016 }
1017 iov_iter_revert(iter, requested);
1018 } else {
1019 result = requested;
1020 }
1021out_release:
1022 nfs_direct_req_release(dreq);
1023out:
1024 return result;
1025}
1026
1027/**
1028 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1029 *
1030 */
1031int __init nfs_init_directcache(void)
1032{
1033 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1034 sizeof(struct nfs_direct_req),
1035 0, (SLAB_RECLAIM_ACCOUNT|
1036 SLAB_MEM_SPREAD),
1037 NULL);
1038 if (nfs_direct_cachep == NULL)
1039 return -ENOMEM;
1040
1041 return 0;
1042}
1043
1044/**
1045 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1046 *
1047 */
1048void nfs_destroy_directcache(void)
1049{
1050 kmem_cache_destroy(nfs_direct_cachep);
1051}
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}