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/*
2 * linux/fs/nfs/direct.c
3 *
4 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
5 *
6 * High-performance uncached I/O for the Linux NFS client
7 *
8 * There are important applications whose performance or correctness
9 * depends on uncached access to file data. Database clusters
10 * (multiple copies of the same instance running on separate hosts)
11 * implement their own cache coherency protocol that subsumes file
12 * system cache protocols. Applications that process datasets
13 * considerably larger than the client's memory do not always benefit
14 * from a local cache. A streaming video server, for instance, has no
15 * need to cache the contents of a file.
16 *
17 * When an application requests uncached I/O, all read and write requests
18 * are made directly to the server; data stored or fetched via these
19 * requests is not cached in the Linux page cache. The client does not
20 * correct unaligned requests from applications. All requested bytes are
21 * held on permanent storage before a direct write system call returns to
22 * an application.
23 *
24 * Solaris implements an uncached I/O facility called directio() that
25 * is used for backups and sequential I/O to very large files. Solaris
26 * also supports uncaching whole NFS partitions with "-o forcedirectio,"
27 * an undocumented mount option.
28 *
29 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
30 * help from Andrew Morton.
31 *
32 * 18 Dec 2001 Initial implementation for 2.4 --cel
33 * 08 Jul 2002 Version for 2.4.19, with bug fixes --trondmy
34 * 08 Jun 2003 Port to 2.5 APIs --cel
35 * 31 Mar 2004 Handle direct I/O without VFS support --cel
36 * 15 Sep 2004 Parallel async reads --cel
37 * 04 May 2005 support O_DIRECT with aio --cel
38 *
39 */
40
41#include <linux/errno.h>
42#include <linux/sched.h>
43#include <linux/kernel.h>
44#include <linux/file.h>
45#include <linux/pagemap.h>
46#include <linux/kref.h>
47#include <linux/slab.h>
48#include <linux/task_io_accounting_ops.h>
49
50#include <linux/nfs_fs.h>
51#include <linux/nfs_page.h>
52#include <linux/sunrpc/clnt.h>
53
54#include <asm/system.h>
55#include <asm/uaccess.h>
56#include <linux/atomic.h>
57
58#include "internal.h"
59#include "iostat.h"
60
61#define NFSDBG_FACILITY NFSDBG_VFS
62
63static struct kmem_cache *nfs_direct_cachep;
64
65/*
66 * This represents a set of asynchronous requests that we're waiting on
67 */
68struct nfs_direct_req {
69 struct kref kref; /* release manager */
70
71 /* I/O parameters */
72 struct nfs_open_context *ctx; /* file open context info */
73 struct nfs_lock_context *l_ctx; /* Lock context info */
74 struct kiocb * iocb; /* controlling i/o request */
75 struct inode * inode; /* target file of i/o */
76
77 /* completion state */
78 atomic_t io_count; /* i/os we're waiting for */
79 spinlock_t lock; /* protect completion state */
80 ssize_t count, /* bytes actually processed */
81 error; /* any reported error */
82 struct completion completion; /* wait for i/o completion */
83
84 /* commit state */
85 struct list_head rewrite_list; /* saved nfs_write_data structs */
86 struct nfs_write_data * commit_data; /* special write_data for commits */
87 int flags;
88#define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
89#define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
90 struct nfs_writeverf verf; /* unstable write verifier */
91};
92
93static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
94static const struct rpc_call_ops nfs_write_direct_ops;
95
96static inline void get_dreq(struct nfs_direct_req *dreq)
97{
98 atomic_inc(&dreq->io_count);
99}
100
101static inline int put_dreq(struct nfs_direct_req *dreq)
102{
103 return atomic_dec_and_test(&dreq->io_count);
104}
105
106/**
107 * nfs_direct_IO - NFS address space operation for direct I/O
108 * @rw: direction (read or write)
109 * @iocb: target I/O control block
110 * @iov: array of vectors that define I/O buffer
111 * @pos: offset in file to begin the operation
112 * @nr_segs: size of iovec array
113 *
114 * The presence of this routine in the address space ops vector means
115 * the NFS client supports direct I/O. However, we shunt off direct
116 * read and write requests before the VFS gets them, so this method
117 * should never be called.
118 */
119ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
120{
121 dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
122 iocb->ki_filp->f_path.dentry->d_name.name,
123 (long long) pos, nr_segs);
124
125 return -EINVAL;
126}
127
128static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
129{
130 unsigned int npages;
131 unsigned int i;
132
133 if (count == 0)
134 return;
135 pages += (pgbase >> PAGE_SHIFT);
136 npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
137 for (i = 0; i < npages; i++) {
138 struct page *page = pages[i];
139 if (!PageCompound(page))
140 set_page_dirty(page);
141 }
142}
143
144static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
145{
146 unsigned int i;
147 for (i = 0; i < npages; i++)
148 page_cache_release(pages[i]);
149}
150
151static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
152{
153 struct nfs_direct_req *dreq;
154
155 dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
156 if (!dreq)
157 return NULL;
158
159 kref_init(&dreq->kref);
160 kref_get(&dreq->kref);
161 init_completion(&dreq->completion);
162 INIT_LIST_HEAD(&dreq->rewrite_list);
163 dreq->iocb = NULL;
164 dreq->ctx = NULL;
165 dreq->l_ctx = NULL;
166 spin_lock_init(&dreq->lock);
167 atomic_set(&dreq->io_count, 0);
168 dreq->count = 0;
169 dreq->error = 0;
170 dreq->flags = 0;
171
172 return dreq;
173}
174
175static void nfs_direct_req_free(struct kref *kref)
176{
177 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
178
179 if (dreq->l_ctx != NULL)
180 nfs_put_lock_context(dreq->l_ctx);
181 if (dreq->ctx != NULL)
182 put_nfs_open_context(dreq->ctx);
183 kmem_cache_free(nfs_direct_cachep, dreq);
184}
185
186static void nfs_direct_req_release(struct nfs_direct_req *dreq)
187{
188 kref_put(&dreq->kref, nfs_direct_req_free);
189}
190
191/*
192 * Collects and returns the final error value/byte-count.
193 */
194static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
195{
196 ssize_t result = -EIOCBQUEUED;
197
198 /* Async requests don't wait here */
199 if (dreq->iocb)
200 goto out;
201
202 result = wait_for_completion_killable(&dreq->completion);
203
204 if (!result)
205 result = dreq->error;
206 if (!result)
207 result = dreq->count;
208
209out:
210 return (ssize_t) result;
211}
212
213/*
214 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
215 * the iocb is still valid here if this is a synchronous request.
216 */
217static void nfs_direct_complete(struct nfs_direct_req *dreq)
218{
219 if (dreq->iocb) {
220 long res = (long) dreq->error;
221 if (!res)
222 res = (long) dreq->count;
223 aio_complete(dreq->iocb, res, 0);
224 }
225 complete_all(&dreq->completion);
226
227 nfs_direct_req_release(dreq);
228}
229
230/*
231 * We must hold a reference to all the pages in this direct read request
232 * until the RPCs complete. This could be long *after* we are woken up in
233 * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
234 */
235static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
236{
237 struct nfs_read_data *data = calldata;
238
239 nfs_readpage_result(task, data);
240}
241
242static void nfs_direct_read_release(void *calldata)
243{
244
245 struct nfs_read_data *data = calldata;
246 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
247 int status = data->task.tk_status;
248
249 spin_lock(&dreq->lock);
250 if (unlikely(status < 0)) {
251 dreq->error = status;
252 spin_unlock(&dreq->lock);
253 } else {
254 dreq->count += data->res.count;
255 spin_unlock(&dreq->lock);
256 nfs_direct_dirty_pages(data->pagevec,
257 data->args.pgbase,
258 data->res.count);
259 }
260 nfs_direct_release_pages(data->pagevec, data->npages);
261
262 if (put_dreq(dreq))
263 nfs_direct_complete(dreq);
264 nfs_readdata_free(data);
265}
266
267static const struct rpc_call_ops nfs_read_direct_ops = {
268#if defined(CONFIG_NFS_V4_1)
269 .rpc_call_prepare = nfs_read_prepare,
270#endif /* CONFIG_NFS_V4_1 */
271 .rpc_call_done = nfs_direct_read_result,
272 .rpc_release = nfs_direct_read_release,
273};
274
275/*
276 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
277 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
278 * bail and stop sending more reads. Read length accounting is
279 * handled automatically by nfs_direct_read_result(). Otherwise, if
280 * no requests have been sent, just return an error.
281 */
282static ssize_t nfs_direct_read_schedule_segment(struct nfs_direct_req *dreq,
283 const struct iovec *iov,
284 loff_t pos)
285{
286 struct nfs_open_context *ctx = dreq->ctx;
287 struct inode *inode = ctx->dentry->d_inode;
288 unsigned long user_addr = (unsigned long)iov->iov_base;
289 size_t count = iov->iov_len;
290 size_t rsize = NFS_SERVER(inode)->rsize;
291 struct rpc_task *task;
292 struct rpc_message msg = {
293 .rpc_cred = ctx->cred,
294 };
295 struct rpc_task_setup task_setup_data = {
296 .rpc_client = NFS_CLIENT(inode),
297 .rpc_message = &msg,
298 .callback_ops = &nfs_read_direct_ops,
299 .workqueue = nfsiod_workqueue,
300 .flags = RPC_TASK_ASYNC,
301 };
302 unsigned int pgbase;
303 int result;
304 ssize_t started = 0;
305
306 do {
307 struct nfs_read_data *data;
308 size_t bytes;
309
310 pgbase = user_addr & ~PAGE_MASK;
311 bytes = min(rsize,count);
312
313 result = -ENOMEM;
314 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
315 if (unlikely(!data))
316 break;
317
318 down_read(¤t->mm->mmap_sem);
319 result = get_user_pages(current, current->mm, user_addr,
320 data->npages, 1, 0, data->pagevec, NULL);
321 up_read(¤t->mm->mmap_sem);
322 if (result < 0) {
323 nfs_readdata_free(data);
324 break;
325 }
326 if ((unsigned)result < data->npages) {
327 bytes = result * PAGE_SIZE;
328 if (bytes <= pgbase) {
329 nfs_direct_release_pages(data->pagevec, result);
330 nfs_readdata_free(data);
331 break;
332 }
333 bytes -= pgbase;
334 data->npages = result;
335 }
336
337 get_dreq(dreq);
338
339 data->req = (struct nfs_page *) dreq;
340 data->inode = inode;
341 data->cred = msg.rpc_cred;
342 data->args.fh = NFS_FH(inode);
343 data->args.context = ctx;
344 data->args.lock_context = dreq->l_ctx;
345 data->args.offset = pos;
346 data->args.pgbase = pgbase;
347 data->args.pages = data->pagevec;
348 data->args.count = bytes;
349 data->res.fattr = &data->fattr;
350 data->res.eof = 0;
351 data->res.count = bytes;
352 nfs_fattr_init(&data->fattr);
353 msg.rpc_argp = &data->args;
354 msg.rpc_resp = &data->res;
355
356 task_setup_data.task = &data->task;
357 task_setup_data.callback_data = data;
358 NFS_PROTO(inode)->read_setup(data, &msg);
359
360 task = rpc_run_task(&task_setup_data);
361 if (IS_ERR(task))
362 break;
363 rpc_put_task(task);
364
365 dprintk("NFS: %5u initiated direct read call "
366 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
367 data->task.tk_pid,
368 inode->i_sb->s_id,
369 (long long)NFS_FILEID(inode),
370 bytes,
371 (unsigned long long)data->args.offset);
372
373 started += bytes;
374 user_addr += bytes;
375 pos += bytes;
376 /* FIXME: Remove this unnecessary math from final patch */
377 pgbase += bytes;
378 pgbase &= ~PAGE_MASK;
379 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
380
381 count -= bytes;
382 } while (count != 0);
383
384 if (started)
385 return started;
386 return result < 0 ? (ssize_t) result : -EFAULT;
387}
388
389static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
390 const struct iovec *iov,
391 unsigned long nr_segs,
392 loff_t pos)
393{
394 ssize_t result = -EINVAL;
395 size_t requested_bytes = 0;
396 unsigned long seg;
397
398 get_dreq(dreq);
399
400 for (seg = 0; seg < nr_segs; seg++) {
401 const struct iovec *vec = &iov[seg];
402 result = nfs_direct_read_schedule_segment(dreq, vec, pos);
403 if (result < 0)
404 break;
405 requested_bytes += result;
406 if ((size_t)result < vec->iov_len)
407 break;
408 pos += vec->iov_len;
409 }
410
411 /*
412 * If no bytes were started, return the error, and let the
413 * generic layer handle the completion.
414 */
415 if (requested_bytes == 0) {
416 nfs_direct_req_release(dreq);
417 return result < 0 ? result : -EIO;
418 }
419
420 if (put_dreq(dreq))
421 nfs_direct_complete(dreq);
422 return 0;
423}
424
425static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
426 unsigned long nr_segs, loff_t pos)
427{
428 ssize_t result = -ENOMEM;
429 struct inode *inode = iocb->ki_filp->f_mapping->host;
430 struct nfs_direct_req *dreq;
431
432 dreq = nfs_direct_req_alloc();
433 if (dreq == NULL)
434 goto out;
435
436 dreq->inode = inode;
437 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
438 dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
439 if (dreq->l_ctx == NULL)
440 goto out_release;
441 if (!is_sync_kiocb(iocb))
442 dreq->iocb = iocb;
443
444 result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
445 if (!result)
446 result = nfs_direct_wait(dreq);
447out_release:
448 nfs_direct_req_release(dreq);
449out:
450 return result;
451}
452
453static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
454{
455 while (!list_empty(&dreq->rewrite_list)) {
456 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
457 list_del(&data->pages);
458 nfs_direct_release_pages(data->pagevec, data->npages);
459 nfs_writedata_free(data);
460 }
461}
462
463#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
464static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
465{
466 struct inode *inode = dreq->inode;
467 struct list_head *p;
468 struct nfs_write_data *data;
469 struct rpc_task *task;
470 struct rpc_message msg = {
471 .rpc_cred = dreq->ctx->cred,
472 };
473 struct rpc_task_setup task_setup_data = {
474 .rpc_client = NFS_CLIENT(inode),
475 .rpc_message = &msg,
476 .callback_ops = &nfs_write_direct_ops,
477 .workqueue = nfsiod_workqueue,
478 .flags = RPC_TASK_ASYNC,
479 };
480
481 dreq->count = 0;
482 get_dreq(dreq);
483
484 list_for_each(p, &dreq->rewrite_list) {
485 data = list_entry(p, struct nfs_write_data, pages);
486
487 get_dreq(dreq);
488
489 /* Use stable writes */
490 data->args.stable = NFS_FILE_SYNC;
491
492 /*
493 * Reset data->res.
494 */
495 nfs_fattr_init(&data->fattr);
496 data->res.count = data->args.count;
497 memset(&data->verf, 0, sizeof(data->verf));
498
499 /*
500 * Reuse data->task; data->args should not have changed
501 * since the original request was sent.
502 */
503 task_setup_data.task = &data->task;
504 task_setup_data.callback_data = data;
505 msg.rpc_argp = &data->args;
506 msg.rpc_resp = &data->res;
507 NFS_PROTO(inode)->write_setup(data, &msg);
508
509 /*
510 * We're called via an RPC callback, so BKL is already held.
511 */
512 task = rpc_run_task(&task_setup_data);
513 if (!IS_ERR(task))
514 rpc_put_task(task);
515
516 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
517 data->task.tk_pid,
518 inode->i_sb->s_id,
519 (long long)NFS_FILEID(inode),
520 data->args.count,
521 (unsigned long long)data->args.offset);
522 }
523
524 if (put_dreq(dreq))
525 nfs_direct_write_complete(dreq, inode);
526}
527
528static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
529{
530 struct nfs_write_data *data = calldata;
531
532 /* Call the NFS version-specific code */
533 NFS_PROTO(data->inode)->commit_done(task, data);
534}
535
536static void nfs_direct_commit_release(void *calldata)
537{
538 struct nfs_write_data *data = calldata;
539 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
540 int status = data->task.tk_status;
541
542 if (status < 0) {
543 dprintk("NFS: %5u commit failed with error %d.\n",
544 data->task.tk_pid, status);
545 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
546 } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
547 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
548 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
549 }
550
551 dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
552 nfs_direct_write_complete(dreq, data->inode);
553 nfs_commit_free(data);
554}
555
556static const struct rpc_call_ops nfs_commit_direct_ops = {
557#if defined(CONFIG_NFS_V4_1)
558 .rpc_call_prepare = nfs_write_prepare,
559#endif /* CONFIG_NFS_V4_1 */
560 .rpc_call_done = nfs_direct_commit_result,
561 .rpc_release = nfs_direct_commit_release,
562};
563
564static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
565{
566 struct nfs_write_data *data = dreq->commit_data;
567 struct rpc_task *task;
568 struct rpc_message msg = {
569 .rpc_argp = &data->args,
570 .rpc_resp = &data->res,
571 .rpc_cred = dreq->ctx->cred,
572 };
573 struct rpc_task_setup task_setup_data = {
574 .task = &data->task,
575 .rpc_client = NFS_CLIENT(dreq->inode),
576 .rpc_message = &msg,
577 .callback_ops = &nfs_commit_direct_ops,
578 .callback_data = data,
579 .workqueue = nfsiod_workqueue,
580 .flags = RPC_TASK_ASYNC,
581 };
582
583 data->inode = dreq->inode;
584 data->cred = msg.rpc_cred;
585
586 data->args.fh = NFS_FH(data->inode);
587 data->args.offset = 0;
588 data->args.count = 0;
589 data->args.context = dreq->ctx;
590 data->args.lock_context = dreq->l_ctx;
591 data->res.count = 0;
592 data->res.fattr = &data->fattr;
593 data->res.verf = &data->verf;
594 nfs_fattr_init(&data->fattr);
595
596 NFS_PROTO(data->inode)->commit_setup(data, &msg);
597
598 /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
599 dreq->commit_data = NULL;
600
601 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
602
603 task = rpc_run_task(&task_setup_data);
604 if (!IS_ERR(task))
605 rpc_put_task(task);
606}
607
608static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
609{
610 int flags = dreq->flags;
611
612 dreq->flags = 0;
613 switch (flags) {
614 case NFS_ODIRECT_DO_COMMIT:
615 nfs_direct_commit_schedule(dreq);
616 break;
617 case NFS_ODIRECT_RESCHED_WRITES:
618 nfs_direct_write_reschedule(dreq);
619 break;
620 default:
621 if (dreq->commit_data != NULL)
622 nfs_commit_free(dreq->commit_data);
623 nfs_direct_free_writedata(dreq);
624 nfs_zap_mapping(inode, inode->i_mapping);
625 nfs_direct_complete(dreq);
626 }
627}
628
629static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
630{
631 dreq->commit_data = nfs_commitdata_alloc();
632 if (dreq->commit_data != NULL)
633 dreq->commit_data->req = (struct nfs_page *) dreq;
634}
635#else
636static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
637{
638 dreq->commit_data = NULL;
639}
640
641static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
642{
643 nfs_direct_free_writedata(dreq);
644 nfs_zap_mapping(inode, inode->i_mapping);
645 nfs_direct_complete(dreq);
646}
647#endif
648
649static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
650{
651 struct nfs_write_data *data = calldata;
652
653 nfs_writeback_done(task, data);
654}
655
656/*
657 * NB: Return the value of the first error return code. Subsequent
658 * errors after the first one are ignored.
659 */
660static void nfs_direct_write_release(void *calldata)
661{
662 struct nfs_write_data *data = calldata;
663 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
664 int status = data->task.tk_status;
665
666 spin_lock(&dreq->lock);
667
668 if (unlikely(status < 0)) {
669 /* An error has occurred, so we should not commit */
670 dreq->flags = 0;
671 dreq->error = status;
672 }
673 if (unlikely(dreq->error != 0))
674 goto out_unlock;
675
676 dreq->count += data->res.count;
677
678 if (data->res.verf->committed != NFS_FILE_SYNC) {
679 switch (dreq->flags) {
680 case 0:
681 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
682 dreq->flags = NFS_ODIRECT_DO_COMMIT;
683 break;
684 case NFS_ODIRECT_DO_COMMIT:
685 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
686 dprintk("NFS: %5u write verify failed\n", data->task.tk_pid);
687 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
688 }
689 }
690 }
691out_unlock:
692 spin_unlock(&dreq->lock);
693
694 if (put_dreq(dreq))
695 nfs_direct_write_complete(dreq, data->inode);
696}
697
698static const struct rpc_call_ops nfs_write_direct_ops = {
699#if defined(CONFIG_NFS_V4_1)
700 .rpc_call_prepare = nfs_write_prepare,
701#endif /* CONFIG_NFS_V4_1 */
702 .rpc_call_done = nfs_direct_write_result,
703 .rpc_release = nfs_direct_write_release,
704};
705
706/*
707 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
708 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
709 * bail and stop sending more writes. Write length accounting is
710 * handled automatically by nfs_direct_write_result(). Otherwise, if
711 * no requests have been sent, just return an error.
712 */
713static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
714 const struct iovec *iov,
715 loff_t pos, int sync)
716{
717 struct nfs_open_context *ctx = dreq->ctx;
718 struct inode *inode = ctx->dentry->d_inode;
719 unsigned long user_addr = (unsigned long)iov->iov_base;
720 size_t count = iov->iov_len;
721 struct rpc_task *task;
722 struct rpc_message msg = {
723 .rpc_cred = ctx->cred,
724 };
725 struct rpc_task_setup task_setup_data = {
726 .rpc_client = NFS_CLIENT(inode),
727 .rpc_message = &msg,
728 .callback_ops = &nfs_write_direct_ops,
729 .workqueue = nfsiod_workqueue,
730 .flags = RPC_TASK_ASYNC,
731 };
732 size_t wsize = NFS_SERVER(inode)->wsize;
733 unsigned int pgbase;
734 int result;
735 ssize_t started = 0;
736
737 do {
738 struct nfs_write_data *data;
739 size_t bytes;
740
741 pgbase = user_addr & ~PAGE_MASK;
742 bytes = min(wsize,count);
743
744 result = -ENOMEM;
745 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
746 if (unlikely(!data))
747 break;
748
749 down_read(¤t->mm->mmap_sem);
750 result = get_user_pages(current, current->mm, user_addr,
751 data->npages, 0, 0, data->pagevec, NULL);
752 up_read(¤t->mm->mmap_sem);
753 if (result < 0) {
754 nfs_writedata_free(data);
755 break;
756 }
757 if ((unsigned)result < data->npages) {
758 bytes = result * PAGE_SIZE;
759 if (bytes <= pgbase) {
760 nfs_direct_release_pages(data->pagevec, result);
761 nfs_writedata_free(data);
762 break;
763 }
764 bytes -= pgbase;
765 data->npages = result;
766 }
767
768 get_dreq(dreq);
769
770 list_move_tail(&data->pages, &dreq->rewrite_list);
771
772 data->req = (struct nfs_page *) dreq;
773 data->inode = inode;
774 data->cred = msg.rpc_cred;
775 data->args.fh = NFS_FH(inode);
776 data->args.context = ctx;
777 data->args.lock_context = dreq->l_ctx;
778 data->args.offset = pos;
779 data->args.pgbase = pgbase;
780 data->args.pages = data->pagevec;
781 data->args.count = bytes;
782 data->args.stable = sync;
783 data->res.fattr = &data->fattr;
784 data->res.count = bytes;
785 data->res.verf = &data->verf;
786 nfs_fattr_init(&data->fattr);
787
788 task_setup_data.task = &data->task;
789 task_setup_data.callback_data = data;
790 msg.rpc_argp = &data->args;
791 msg.rpc_resp = &data->res;
792 NFS_PROTO(inode)->write_setup(data, &msg);
793
794 task = rpc_run_task(&task_setup_data);
795 if (IS_ERR(task))
796 break;
797 rpc_put_task(task);
798
799 dprintk("NFS: %5u initiated direct write call "
800 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
801 data->task.tk_pid,
802 inode->i_sb->s_id,
803 (long long)NFS_FILEID(inode),
804 bytes,
805 (unsigned long long)data->args.offset);
806
807 started += bytes;
808 user_addr += bytes;
809 pos += bytes;
810
811 /* FIXME: Remove this useless math from the final patch */
812 pgbase += bytes;
813 pgbase &= ~PAGE_MASK;
814 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
815
816 count -= bytes;
817 } while (count != 0);
818
819 if (started)
820 return started;
821 return result < 0 ? (ssize_t) result : -EFAULT;
822}
823
824static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
825 const struct iovec *iov,
826 unsigned long nr_segs,
827 loff_t pos, int sync)
828{
829 ssize_t result = 0;
830 size_t requested_bytes = 0;
831 unsigned long seg;
832
833 get_dreq(dreq);
834
835 for (seg = 0; seg < nr_segs; seg++) {
836 const struct iovec *vec = &iov[seg];
837 result = nfs_direct_write_schedule_segment(dreq, vec,
838 pos, sync);
839 if (result < 0)
840 break;
841 requested_bytes += result;
842 if ((size_t)result < vec->iov_len)
843 break;
844 pos += vec->iov_len;
845 }
846
847 /*
848 * If no bytes were started, return the error, and let the
849 * generic layer handle the completion.
850 */
851 if (requested_bytes == 0) {
852 nfs_direct_req_release(dreq);
853 return result < 0 ? result : -EIO;
854 }
855
856 if (put_dreq(dreq))
857 nfs_direct_write_complete(dreq, dreq->inode);
858 return 0;
859}
860
861static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
862 unsigned long nr_segs, loff_t pos,
863 size_t count)
864{
865 ssize_t result = -ENOMEM;
866 struct inode *inode = iocb->ki_filp->f_mapping->host;
867 struct nfs_direct_req *dreq;
868 size_t wsize = NFS_SERVER(inode)->wsize;
869 int sync = NFS_UNSTABLE;
870
871 dreq = nfs_direct_req_alloc();
872 if (!dreq)
873 goto out;
874 nfs_alloc_commit_data(dreq);
875
876 if (dreq->commit_data == NULL || count <= wsize)
877 sync = NFS_FILE_SYNC;
878
879 dreq->inode = inode;
880 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
881 dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
882 if (dreq->l_ctx == NULL)
883 goto out_release;
884 if (!is_sync_kiocb(iocb))
885 dreq->iocb = iocb;
886
887 result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
888 if (!result)
889 result = nfs_direct_wait(dreq);
890out_release:
891 nfs_direct_req_release(dreq);
892out:
893 return result;
894}
895
896/**
897 * nfs_file_direct_read - file direct read operation for NFS files
898 * @iocb: target I/O control block
899 * @iov: vector of user buffers into which to read data
900 * @nr_segs: size of iov vector
901 * @pos: byte offset in file where reading starts
902 *
903 * We use this function for direct reads instead of calling
904 * generic_file_aio_read() in order to avoid gfar's check to see if
905 * the request starts before the end of the file. For that check
906 * to work, we must generate a GETATTR before each direct read, and
907 * even then there is a window between the GETATTR and the subsequent
908 * READ where the file size could change. Our preference is simply
909 * to do all reads the application wants, and the server will take
910 * care of managing the end of file boundary.
911 *
912 * This function also eliminates unnecessarily updating the file's
913 * atime locally, as the NFS server sets the file's atime, and this
914 * client must read the updated atime from the server back into its
915 * cache.
916 */
917ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
918 unsigned long nr_segs, loff_t pos)
919{
920 ssize_t retval = -EINVAL;
921 struct file *file = iocb->ki_filp;
922 struct address_space *mapping = file->f_mapping;
923 size_t count;
924
925 count = iov_length(iov, nr_segs);
926 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
927
928 dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
929 file->f_path.dentry->d_parent->d_name.name,
930 file->f_path.dentry->d_name.name,
931 count, (long long) pos);
932
933 retval = 0;
934 if (!count)
935 goto out;
936
937 retval = nfs_sync_mapping(mapping);
938 if (retval)
939 goto out;
940
941 task_io_account_read(count);
942
943 retval = nfs_direct_read(iocb, iov, nr_segs, pos);
944 if (retval > 0)
945 iocb->ki_pos = pos + retval;
946
947out:
948 return retval;
949}
950
951/**
952 * nfs_file_direct_write - file direct write operation for NFS files
953 * @iocb: target I/O control block
954 * @iov: vector of user buffers from which to write data
955 * @nr_segs: size of iov vector
956 * @pos: byte offset in file where writing starts
957 *
958 * We use this function for direct writes instead of calling
959 * generic_file_aio_write() in order to avoid taking the inode
960 * semaphore and updating the i_size. The NFS server will set
961 * the new i_size and this client must read the updated size
962 * back into its cache. We let the server do generic write
963 * parameter checking and report problems.
964 *
965 * We eliminate local atime updates, see direct read above.
966 *
967 * We avoid unnecessary page cache invalidations for normal cached
968 * readers of this file.
969 *
970 * Note that O_APPEND is not supported for NFS direct writes, as there
971 * is no atomic O_APPEND write facility in the NFS protocol.
972 */
973ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
974 unsigned long nr_segs, loff_t pos)
975{
976 ssize_t retval = -EINVAL;
977 struct file *file = iocb->ki_filp;
978 struct address_space *mapping = file->f_mapping;
979 size_t count;
980
981 count = iov_length(iov, nr_segs);
982 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
983
984 dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
985 file->f_path.dentry->d_parent->d_name.name,
986 file->f_path.dentry->d_name.name,
987 count, (long long) pos);
988
989 retval = generic_write_checks(file, &pos, &count, 0);
990 if (retval)
991 goto out;
992
993 retval = -EINVAL;
994 if ((ssize_t) count < 0)
995 goto out;
996 retval = 0;
997 if (!count)
998 goto out;
999
1000 retval = nfs_sync_mapping(mapping);
1001 if (retval)
1002 goto out;
1003
1004 task_io_account_write(count);
1005
1006 retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
1007
1008 if (retval > 0)
1009 iocb->ki_pos = pos + retval;
1010
1011out:
1012 return retval;
1013}
1014
1015/**
1016 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1017 *
1018 */
1019int __init nfs_init_directcache(void)
1020{
1021 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1022 sizeof(struct nfs_direct_req),
1023 0, (SLAB_RECLAIM_ACCOUNT|
1024 SLAB_MEM_SPREAD),
1025 NULL);
1026 if (nfs_direct_cachep == NULL)
1027 return -ENOMEM;
1028
1029 return 0;
1030}
1031
1032/**
1033 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1034 *
1035 */
1036void nfs_destroy_directcache(void)
1037{
1038 kmem_cache_destroy(nfs_direct_cachep);
1039}