]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame - fs/nfs/direct.c
Merge branch 'core-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[mirror_ubuntu-bionic-kernel.git] / fs / nfs / direct.c
CommitLineData
1da177e4
LT
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
88467055 10 * (multiple copies of the same instance running on separate hosts)
1da177e4 11 * implement their own cache coherency protocol that subsumes file
88467055
CL
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
1da177e4
LT
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
88467055 37 * 04 May 2005 support O_DIRECT with aio --cel
1da177e4
LT
38 *
39 */
40
1da177e4
LT
41#include <linux/errno.h>
42#include <linux/sched.h>
43#include <linux/kernel.h>
1da177e4
LT
44#include <linux/file.h>
45#include <linux/pagemap.h>
46#include <linux/kref.h>
5a0e3ad6 47#include <linux/slab.h>
7ec10f26 48#include <linux/task_io_accounting_ops.h>
6296556f 49#include <linux/module.h>
1da177e4
LT
50
51#include <linux/nfs_fs.h>
52#include <linux/nfs_page.h>
53#include <linux/sunrpc/clnt.h>
54
1da177e4 55#include <asm/uaccess.h>
60063497 56#include <linux/atomic.h>
1da177e4 57
8d5658c9 58#include "internal.h"
91d5b470 59#include "iostat.h"
1763da12 60#include "pnfs.h"
91d5b470 61
1da177e4 62#define NFSDBG_FACILITY NFSDBG_VFS
1da177e4 63
e18b890b 64static struct kmem_cache *nfs_direct_cachep;
1da177e4
LT
65
66/*
67 * This represents a set of asynchronous requests that we're waiting on
68 */
0a00b77b
WAA
69struct nfs_direct_mirror {
70 ssize_t count;
71};
72
1da177e4
LT
73struct nfs_direct_req {
74 struct kref kref; /* release manager */
15ce4a0c
CL
75
76 /* I/O parameters */
a8881f5a 77 struct nfs_open_context *ctx; /* file open context info */
f11ac8db 78 struct nfs_lock_context *l_ctx; /* Lock context info */
99514f8f 79 struct kiocb * iocb; /* controlling i/o request */
88467055 80 struct inode * inode; /* target file of i/o */
15ce4a0c
CL
81
82 /* completion state */
607f31e8 83 atomic_t io_count; /* i/os we're waiting for */
15ce4a0c 84 spinlock_t lock; /* protect completion state */
0a00b77b
WAA
85
86 struct nfs_direct_mirror mirrors[NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX];
87 int mirror_count;
88
15ce4a0c 89 ssize_t count, /* bytes actually processed */
35754bc0 90 bytes_left, /* bytes left to be sent */
5fadeb47 91 io_start, /* start of IO */
1da177e4 92 error; /* any reported error */
d72b7a6b 93 struct completion completion; /* wait for i/o completion */
fad61490
TM
94
95 /* commit state */
1763da12
FI
96 struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo */
97 struct pnfs_ds_commit_info ds_cinfo; /* Storage for cinfo */
98 struct work_struct work;
fad61490
TM
99 int flags;
100#define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
101#define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
102 struct nfs_writeverf verf; /* unstable write verifier */
1da177e4
LT
103};
104
1763da12
FI
105static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
106static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
fad61490 107static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
1763da12 108static void nfs_direct_write_schedule_work(struct work_struct *work);
607f31e8
TM
109
110static inline void get_dreq(struct nfs_direct_req *dreq)
111{
112 atomic_inc(&dreq->io_count);
113}
114
115static inline int put_dreq(struct nfs_direct_req *dreq)
116{
117 return atomic_dec_and_test(&dreq->io_count);
118}
119
0a00b77b
WAA
120static void
121nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
122{
123 int i;
124 ssize_t count;
125
1ccbad9f
PT
126 if (dreq->mirror_count == 1) {
127 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
128 dreq->count += hdr->good_bytes;
129 } else {
130 /* mirrored writes */
131 count = dreq->mirrors[hdr->pgio_mirror_idx].count;
132 if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
133 count = hdr->io_start + hdr->good_bytes - dreq->io_start;
134 dreq->mirrors[hdr->pgio_mirror_idx].count = count;
135 }
136 /* update the dreq->count by finding the minimum agreed count from all
137 * mirrors */
138 count = dreq->mirrors[0].count;
0a00b77b 139
1ccbad9f
PT
140 for (i = 1; i < dreq->mirror_count; i++)
141 count = min(count, dreq->mirrors[i].count);
0a00b77b 142
1ccbad9f 143 dreq->count = count;
5fadeb47 144 }
0a00b77b
WAA
145}
146
5002c586
WAA
147/*
148 * nfs_direct_select_verf - select the right verifier
149 * @dreq - direct request possibly spanning multiple servers
150 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
6cccbb6f 151 * @commit_idx - commit bucket index for the DS
5002c586
WAA
152 *
153 * returns the correct verifier to use given the role of the server
154 */
155static struct nfs_writeverf *
156nfs_direct_select_verf(struct nfs_direct_req *dreq,
157 struct nfs_client *ds_clp,
6cccbb6f 158 int commit_idx)
5002c586
WAA
159{
160 struct nfs_writeverf *verfp = &dreq->verf;
161
162#ifdef CONFIG_NFS_V4_1
834e465b
KM
163 /*
164 * pNFS is in use, use the DS verf except commit_through_mds is set
165 * for layout segment where nbuckets is zero.
166 */
167 if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
6cccbb6f
WAA
168 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
169 verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
5002c586
WAA
170 else
171 WARN_ON_ONCE(1);
172 }
173#endif
174 return verfp;
175}
176
177
178/*
179 * nfs_direct_set_hdr_verf - set the write/commit verifier
180 * @dreq - direct request possibly spanning multiple servers
181 * @hdr - pageio header to validate against previously seen verfs
182 *
183 * Set the server's (MDS or DS) "seen" verifier
184 */
185static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
186 struct nfs_pgio_header *hdr)
187{
188 struct nfs_writeverf *verfp;
189
6cccbb6f 190 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
5002c586
WAA
191 WARN_ON_ONCE(verfp->committed >= 0);
192 memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
193 WARN_ON_ONCE(verfp->committed < 0);
194}
195
196/*
197 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
198 * @dreq - direct request possibly spanning multiple servers
199 * @hdr - pageio header to validate against previously seen verf
200 *
201 * set the server's "seen" verf if not initialized.
202 * returns result of comparison between @hdr->verf and the "seen"
203 * verf of the server used by @hdr (DS or MDS)
204 */
205static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
206 struct nfs_pgio_header *hdr)
207{
208 struct nfs_writeverf *verfp;
209
6cccbb6f 210 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
5002c586
WAA
211 if (verfp->committed < 0) {
212 nfs_direct_set_hdr_verf(dreq, hdr);
213 return 0;
214 }
215 return memcmp(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
216}
217
5002c586
WAA
218/*
219 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
220 * @dreq - direct request possibly spanning multiple servers
221 * @data - commit data to validate against previously seen verf
222 *
223 * returns result of comparison between @data->verf and the verf of
224 * the server used by @data (DS or MDS)
225 */
226static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
227 struct nfs_commit_data *data)
228{
229 struct nfs_writeverf *verfp;
230
231 verfp = nfs_direct_select_verf(dreq, data->ds_clp,
232 data->ds_commit_index);
80c76fe3
WAA
233
234 /* verifier not set so always fail */
235 if (verfp->committed < 0)
236 return 1;
237
5002c586
WAA
238 return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf));
239}
5002c586 240
1da177e4 241/**
b8a32e2b 242 * nfs_direct_IO - NFS address space operation for direct I/O
b8a32e2b
CL
243 * @iocb: target I/O control block
244 * @iov: array of vectors that define I/O buffer
245 * @pos: offset in file to begin the operation
246 * @nr_segs: size of iovec array
247 *
248 * The presence of this routine in the address space ops vector means
a564b8f0
MG
249 * the NFS client supports direct I/O. However, for most direct IO, we
250 * shunt off direct read and write requests before the VFS gets them,
251 * so this method is only ever called for swap.
1da177e4 252 */
22c6186e 253ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter, loff_t pos)
b8a32e2b 254{
ee8a1a8b
PT
255 struct inode *inode = iocb->ki_filp->f_mapping->host;
256
257 /* we only support swap file calling nfs_direct_IO */
258 if (!IS_SWAPFILE(inode))
259 return 0;
260
66ee59af 261 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
a564b8f0 262
6f673763 263 if (iov_iter_rw(iter) == READ)
e19a8a0a 264 return nfs_file_direct_read(iocb, iter, pos);
65a4a1ca 265 return nfs_file_direct_write(iocb, iter);
b8a32e2b
CL
266}
267
749e146e 268static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
9c93ab7d 269{
749e146e 270 unsigned int i;
607f31e8 271 for (i = 0; i < npages; i++)
09cbfeaf 272 put_page(pages[i]);
6b45d858
TM
273}
274
1763da12
FI
275void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
276 struct nfs_direct_req *dreq)
277{
f4086a3d 278 cinfo->lock = &dreq->inode->i_lock;
1763da12
FI
279 cinfo->mds = &dreq->mds_cinfo;
280 cinfo->ds = &dreq->ds_cinfo;
281 cinfo->dreq = dreq;
282 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
283}
284
0a00b77b
WAA
285static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
286 struct nfs_pageio_descriptor *pgio,
287 struct nfs_page *req)
288{
289 int mirror_count = 1;
290
291 if (pgio->pg_ops->pg_get_mirror_count)
292 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
293
294 dreq->mirror_count = mirror_count;
295}
296
93619e59 297static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
1da177e4 298{
93619e59
CL
299 struct nfs_direct_req *dreq;
300
292f3eee 301 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
93619e59
CL
302 if (!dreq)
303 return NULL;
304
305 kref_init(&dreq->kref);
607f31e8 306 kref_get(&dreq->kref);
d72b7a6b 307 init_completion(&dreq->completion);
1763da12 308 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
5002c586 309 dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */
1763da12 310 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
0a00b77b 311 dreq->mirror_count = 1;
15ce4a0c 312 spin_lock_init(&dreq->lock);
93619e59
CL
313
314 return dreq;
1da177e4
LT
315}
316
b4946ffb 317static void nfs_direct_req_free(struct kref *kref)
1da177e4
LT
318{
319 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
a8881f5a 320
8c393f9a 321 nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
f11ac8db
TM
322 if (dreq->l_ctx != NULL)
323 nfs_put_lock_context(dreq->l_ctx);
a8881f5a
TM
324 if (dreq->ctx != NULL)
325 put_nfs_open_context(dreq->ctx);
1da177e4
LT
326 kmem_cache_free(nfs_direct_cachep, dreq);
327}
328
b4946ffb
TM
329static void nfs_direct_req_release(struct nfs_direct_req *dreq)
330{
331 kref_put(&dreq->kref, nfs_direct_req_free);
332}
333
6296556f
PT
334ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
335{
336 return dreq->bytes_left;
337}
338EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
339
bc0fb201
CL
340/*
341 * Collects and returns the final error value/byte-count.
342 */
343static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
344{
15ce4a0c 345 ssize_t result = -EIOCBQUEUED;
bc0fb201
CL
346
347 /* Async requests don't wait here */
348 if (dreq->iocb)
349 goto out;
350
150030b7 351 result = wait_for_completion_killable(&dreq->completion);
bc0fb201
CL
352
353 if (!result)
15ce4a0c 354 result = dreq->error;
bc0fb201 355 if (!result)
15ce4a0c 356 result = dreq->count;
bc0fb201
CL
357
358out:
bc0fb201
CL
359 return (ssize_t) result;
360}
361
63ab46ab 362/*
607f31e8
TM
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.
63ab46ab 365 */
9811cd57 366static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
63ab46ab 367{
9811cd57
CH
368 struct inode *inode = dreq->inode;
369
2a009ec9 370 if (dreq->iocb && write) {
9811cd57 371 loff_t pos = dreq->iocb->ki_pos + dreq->count;
2a009ec9
CH
372
373 spin_lock(&inode->i_lock);
374 if (i_size_read(inode) < pos)
375 i_size_write(inode, pos);
376 spin_unlock(&inode->i_lock);
377 }
378
1f90ee27 379 if (write)
2a009ec9 380 nfs_zap_mapping(inode, inode->i_mapping);
1f90ee27 381
fe0f07d0 382 inode_dio_end(inode);
2a009ec9
CH
383
384 if (dreq->iocb) {
15ce4a0c 385 long res = (long) dreq->error;
63ab46ab 386 if (!res)
15ce4a0c 387 res = (long) dreq->count;
04b2fa9f 388 dreq->iocb->ki_complete(dreq->iocb, res, 0);
d72b7a6b 389 }
2a009ec9 390
d72b7a6b 391 complete_all(&dreq->completion);
63ab46ab 392
b4946ffb 393 nfs_direct_req_release(dreq);
63ab46ab
CL
394}
395
1385b811 396static void nfs_direct_readpage_release(struct nfs_page *req)
1da177e4 397{
1e8968c5 398 dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
2b0143b5
DH
399 d_inode(req->wb_context->dentry)->i_sb->s_id,
400 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
584aa810
FI
401 req->wb_bytes,
402 (long long)req_offset(req));
403 nfs_release_request(req);
fdd1e74c
TM
404}
405
584aa810 406static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
fdd1e74c 407{
584aa810
FI
408 unsigned long bytes = 0;
409 struct nfs_direct_req *dreq = hdr->dreq;
fdd1e74c 410
584aa810
FI
411 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
412 goto out_put;
15ce4a0c
CL
413
414 spin_lock(&dreq->lock);
584aa810
FI
415 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
416 dreq->error = hdr->error;
417 else
0a00b77b
WAA
418 nfs_direct_good_bytes(dreq, hdr);
419
584aa810
FI
420 spin_unlock(&dreq->lock);
421
4bd8b010
TM
422 while (!list_empty(&hdr->pages)) {
423 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
424 struct page *page = req->wb_page;
425
be7e9858
JL
426 if (!PageCompound(page) && bytes < hdr->good_bytes)
427 set_page_dirty(page);
4bd8b010
TM
428 bytes += req->wb_bytes;
429 nfs_list_remove_request(req);
430 nfs_direct_readpage_release(req);
d4a8f367 431 }
584aa810 432out_put:
607f31e8 433 if (put_dreq(dreq))
9811cd57 434 nfs_direct_complete(dreq, false);
584aa810 435 hdr->release(hdr);
1da177e4
LT
436}
437
3e9e0ca3 438static void nfs_read_sync_pgio_error(struct list_head *head)
cd841605 439{
584aa810 440 struct nfs_page *req;
cd841605 441
584aa810
FI
442 while (!list_empty(head)) {
443 req = nfs_list_entry(head->next);
444 nfs_list_remove_request(req);
445 nfs_release_request(req);
446 }
cd841605
FI
447}
448
584aa810
FI
449static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
450{
451 get_dreq(hdr->dreq);
452}
453
454static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
3e9e0ca3 455 .error_cleanup = nfs_read_sync_pgio_error,
584aa810
FI
456 .init_hdr = nfs_direct_pgio_init,
457 .completion = nfs_direct_read_completion,
458};
459
d4cc948b 460/*
607f31e8
TM
461 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
462 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
463 * bail and stop sending more reads. Read length accounting is
464 * handled automatically by nfs_direct_read_result(). Otherwise, if
465 * no requests have been sent, just return an error.
1da177e4 466 */
1da177e4 467
91f79c43
AV
468static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
469 struct iov_iter *iter,
470 loff_t pos)
471{
472 struct nfs_pageio_descriptor desc;
473 struct inode *inode = dreq->inode;
474 ssize_t result = -EINVAL;
475 size_t requested_bytes = 0;
476 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
e9f7bee1 477
16b90578 478 nfs_pageio_init_read(&desc, dreq->inode, false,
91f79c43
AV
479 &nfs_direct_read_completion_ops);
480 get_dreq(dreq);
481 desc.pg_dreq = dreq;
fe0f07d0 482 inode_dio_begin(inode);
a564b8f0 483
91f79c43
AV
484 while (iov_iter_count(iter)) {
485 struct page **pagevec;
486 size_t bytes;
487 size_t pgbase;
488 unsigned npages, i;
607f31e8 489
91f79c43
AV
490 result = iov_iter_get_pages_alloc(iter, &pagevec,
491 rsize, &pgbase);
492 if (result < 0)
493 break;
494
495 bytes = result;
496 iov_iter_advance(iter, bytes);
497 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
584aa810
FI
498 for (i = 0; i < npages; i++) {
499 struct nfs_page *req;
bf5fc402 500 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
584aa810 501 /* XXX do we need to do the eof zeroing found in async_filler? */
2bfc6e56 502 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
584aa810
FI
503 pgbase, req_len);
504 if (IS_ERR(req)) {
584aa810
FI
505 result = PTR_ERR(req);
506 break;
507 }
508 req->wb_index = pos >> PAGE_SHIFT;
509 req->wb_offset = pos & ~PAGE_MASK;
91f79c43
AV
510 if (!nfs_pageio_add_request(&desc, req)) {
511 result = desc.pg_error;
584aa810 512 nfs_release_request(req);
584aa810
FI
513 break;
514 }
515 pgbase = 0;
516 bytes -= req_len;
91f79c43 517 requested_bytes += req_len;
584aa810 518 pos += req_len;
35754bc0 519 dreq->bytes_left -= req_len;
584aa810 520 }
6d74743b 521 nfs_direct_release_pages(pagevec, npages);
91f79c43 522 kvfree(pagevec);
19f73787
CL
523 if (result < 0)
524 break;
19f73787
CL
525 }
526
584aa810
FI
527 nfs_pageio_complete(&desc);
528
839f7ad6
CL
529 /*
530 * If no bytes were started, return the error, and let the
531 * generic layer handle the completion.
532 */
533 if (requested_bytes == 0) {
fe0f07d0 534 inode_dio_end(inode);
839f7ad6
CL
535 nfs_direct_req_release(dreq);
536 return result < 0 ? result : -EIO;
537 }
538
19f73787 539 if (put_dreq(dreq))
9811cd57 540 nfs_direct_complete(dreq, false);
839f7ad6 541 return 0;
19f73787
CL
542}
543
14a3ec79
CH
544/**
545 * nfs_file_direct_read - file direct read operation for NFS files
546 * @iocb: target I/O control block
619d30b4 547 * @iter: vector of user buffers into which to read data
14a3ec79
CH
548 * @pos: byte offset in file where reading starts
549 *
550 * We use this function for direct reads instead of calling
551 * generic_file_aio_read() in order to avoid gfar's check to see if
552 * the request starts before the end of the file. For that check
553 * to work, we must generate a GETATTR before each direct read, and
554 * even then there is a window between the GETATTR and the subsequent
555 * READ where the file size could change. Our preference is simply
556 * to do all reads the application wants, and the server will take
557 * care of managing the end of file boundary.
558 *
559 * This function also eliminates unnecessarily updating the file's
560 * atime locally, as the NFS server sets the file's atime, and this
561 * client must read the updated atime from the server back into its
562 * cache.
563 */
619d30b4 564ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
e19a8a0a 565 loff_t pos)
1da177e4 566{
14a3ec79
CH
567 struct file *file = iocb->ki_filp;
568 struct address_space *mapping = file->f_mapping;
569 struct inode *inode = mapping->host;
1da177e4 570 struct nfs_direct_req *dreq;
b3c54de6 571 struct nfs_lock_context *l_ctx;
14a3ec79 572 ssize_t result = -EINVAL;
a6cbcd4a 573 size_t count = iov_iter_count(iter);
14a3ec79
CH
574 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
575
576 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
577 file, count, (long long) pos);
578
579 result = 0;
580 if (!count)
581 goto out;
582
5955102c 583 inode_lock(inode);
14a3ec79
CH
584 result = nfs_sync_mapping(mapping);
585 if (result)
d0b9875d 586 goto out_unlock;
1da177e4 587
14a3ec79
CH
588 task_io_account_read(count);
589
590 result = -ENOMEM;
607f31e8 591 dreq = nfs_direct_req_alloc();
f11ac8db 592 if (dreq == NULL)
d0b9875d 593 goto out_unlock;
1da177e4 594
91d5b470 595 dreq->inode = inode;
619d30b4 596 dreq->bytes_left = count;
5fadeb47 597 dreq->io_start = pos;
cd3758e3 598 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
b3c54de6
TM
599 l_ctx = nfs_get_lock_context(dreq->ctx);
600 if (IS_ERR(l_ctx)) {
601 result = PTR_ERR(l_ctx);
f11ac8db 602 goto out_release;
b3c54de6
TM
603 }
604 dreq->l_ctx = l_ctx;
487b8372
CL
605 if (!is_sync_kiocb(iocb))
606 dreq->iocb = iocb;
1da177e4 607
619d30b4 608 NFS_I(inode)->read_io += count;
91f79c43 609 result = nfs_direct_read_schedule_iovec(dreq, iter, pos);
d0b9875d 610
5955102c 611 inode_unlock(inode);
d0b9875d 612
14a3ec79 613 if (!result) {
607f31e8 614 result = nfs_direct_wait(dreq);
14a3ec79
CH
615 if (result > 0)
616 iocb->ki_pos = pos + result;
617 }
d0b9875d
CH
618
619 nfs_direct_req_release(dreq);
620 return result;
621
f11ac8db 622out_release:
b4946ffb 623 nfs_direct_req_release(dreq);
d0b9875d 624out_unlock:
5955102c 625 inode_unlock(inode);
f11ac8db 626out:
1da177e4
LT
627 return result;
628}
629
085d1e33
TH
630static void
631nfs_direct_write_scan_commit_list(struct inode *inode,
632 struct list_head *list,
633 struct nfs_commit_info *cinfo)
634{
635 spin_lock(cinfo->lock);
636#ifdef CONFIG_NFS_V4_1
637 if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
638 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
639#endif
640 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
641 spin_unlock(cinfo->lock);
642}
643
fad61490
TM
644static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
645{
1763da12
FI
646 struct nfs_pageio_descriptor desc;
647 struct nfs_page *req, *tmp;
648 LIST_HEAD(reqs);
649 struct nfs_commit_info cinfo;
650 LIST_HEAD(failed);
0a00b77b 651 int i;
1763da12
FI
652
653 nfs_init_cinfo_from_dreq(&cinfo, dreq);
085d1e33 654 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
1da177e4 655
fad61490 656 dreq->count = 0;
0a00b77b
WAA
657 for (i = 0; i < dreq->mirror_count; i++)
658 dreq->mirrors[i].count = 0;
607f31e8
TM
659 get_dreq(dreq);
660
a20c93e3 661 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
1763da12
FI
662 &nfs_direct_write_completion_ops);
663 desc.pg_dreq = dreq;
fedb595c 664
0a00b77b
WAA
665 req = nfs_list_entry(reqs.next);
666 nfs_direct_setup_mirroring(dreq, &desc, req);
d600ad1f
PT
667 if (desc.pg_error < 0) {
668 list_splice_init(&reqs, &failed);
669 goto out_failed;
670 }
0a00b77b 671
1763da12
FI
672 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
673 if (!nfs_pageio_add_request(&desc, req)) {
4035c248 674 nfs_list_remove_request(req);
1763da12
FI
675 nfs_list_add_request(req, &failed);
676 spin_lock(cinfo.lock);
677 dreq->flags = 0;
d600ad1f
PT
678 if (desc.pg_error < 0)
679 dreq->error = desc.pg_error;
680 else
681 dreq->error = -EIO;
1763da12
FI
682 spin_unlock(cinfo.lock);
683 }
5a695da2 684 nfs_release_request(req);
1763da12
FI
685 }
686 nfs_pageio_complete(&desc);
fad61490 687
d600ad1f 688out_failed:
4035c248
TM
689 while (!list_empty(&failed)) {
690 req = nfs_list_entry(failed.next);
691 nfs_list_remove_request(req);
1d1afcbc 692 nfs_unlock_and_release_request(req);
4035c248 693 }
fad61490 694
1763da12
FI
695 if (put_dreq(dreq))
696 nfs_direct_write_complete(dreq, dreq->inode);
c9d8f89d
TM
697}
698
1763da12 699static void nfs_direct_commit_complete(struct nfs_commit_data *data)
c9d8f89d 700{
0b7c0153 701 struct nfs_direct_req *dreq = data->dreq;
1763da12
FI
702 struct nfs_commit_info cinfo;
703 struct nfs_page *req;
c9d8f89d
TM
704 int status = data->task.tk_status;
705
1763da12 706 nfs_init_cinfo_from_dreq(&cinfo, dreq);
c9d8f89d 707 if (status < 0) {
60fa3f76 708 dprintk("NFS: %5u commit failed with error %d.\n",
1763da12 709 data->task.tk_pid, status);
fad61490 710 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
5002c586 711 } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
c9d8f89d 712 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
fad61490 713 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
1da177e4
LT
714 }
715
c9d8f89d 716 dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
1763da12
FI
717 while (!list_empty(&data->pages)) {
718 req = nfs_list_entry(data->pages.next);
719 nfs_list_remove_request(req);
720 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
721 /* Note the rewrite will go through mds */
b57ff130 722 nfs_mark_request_commit(req, NULL, &cinfo, 0);
906369e4
FI
723 } else
724 nfs_release_request(req);
1d1afcbc 725 nfs_unlock_and_release_request(req);
1763da12
FI
726 }
727
728 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
729 nfs_direct_write_complete(dreq, data->inode);
1da177e4
LT
730}
731
b20135d0
TM
732static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
733 struct nfs_page *req)
1763da12 734{
b20135d0
TM
735 struct nfs_direct_req *dreq = cinfo->dreq;
736
737 spin_lock(&dreq->lock);
738 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
739 spin_unlock(&dreq->lock);
740 nfs_mark_request_commit(req, NULL, cinfo, 0);
1763da12
FI
741}
742
743static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
744 .completion = nfs_direct_commit_complete,
b20135d0 745 .resched_write = nfs_direct_resched_write,
fad61490
TM
746};
747
748static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
1da177e4 749{
1763da12
FI
750 int res;
751 struct nfs_commit_info cinfo;
752 LIST_HEAD(mds_list);
753
754 nfs_init_cinfo_from_dreq(&cinfo, dreq);
755 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
756 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
757 if (res < 0) /* res == -ENOMEM */
758 nfs_direct_write_reschedule(dreq);
fad61490 759}
1da177e4 760
1763da12 761static void nfs_direct_write_schedule_work(struct work_struct *work)
fad61490 762{
1763da12 763 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
fad61490 764 int flags = dreq->flags;
1da177e4 765
fad61490
TM
766 dreq->flags = 0;
767 switch (flags) {
768 case NFS_ODIRECT_DO_COMMIT:
769 nfs_direct_commit_schedule(dreq);
1da177e4 770 break;
fad61490
TM
771 case NFS_ODIRECT_RESCHED_WRITES:
772 nfs_direct_write_reschedule(dreq);
773 break;
774 default:
9811cd57 775 nfs_direct_complete(dreq, true);
fad61490
TM
776 }
777}
1da177e4 778
1763da12 779static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
fad61490 780{
1763da12 781 schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
fad61490 782}
1763da12 783
1763da12
FI
784static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
785{
786 struct nfs_direct_req *dreq = hdr->dreq;
787 struct nfs_commit_info cinfo;
c65e6254 788 bool request_commit = false;
1763da12
FI
789 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
790
791 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
792 goto out_put;
793
794 nfs_init_cinfo_from_dreq(&cinfo, dreq);
795
796 spin_lock(&dreq->lock);
797
798 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
799 dreq->flags = 0;
800 dreq->error = hdr->error;
801 }
c65e6254 802 if (dreq->error == 0) {
0a00b77b 803 nfs_direct_good_bytes(dreq, hdr);
c65e6254 804 if (nfs_write_need_commit(hdr)) {
1763da12 805 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
c65e6254 806 request_commit = true;
1763da12 807 else if (dreq->flags == 0) {
5002c586 808 nfs_direct_set_hdr_verf(dreq, hdr);
c65e6254 809 request_commit = true;
1763da12
FI
810 dreq->flags = NFS_ODIRECT_DO_COMMIT;
811 } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
c65e6254
WAA
812 request_commit = true;
813 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
5002c586
WAA
814 dreq->flags =
815 NFS_ODIRECT_RESCHED_WRITES;
1763da12
FI
816 }
817 }
818 }
819 spin_unlock(&dreq->lock);
820
821 while (!list_empty(&hdr->pages)) {
2bfc6e56 822
1763da12
FI
823 req = nfs_list_entry(hdr->pages.next);
824 nfs_list_remove_request(req);
c65e6254 825 if (request_commit) {
04277086 826 kref_get(&req->wb_kref);
b57ff130
WAA
827 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
828 hdr->ds_commit_idx);
1763da12 829 }
1d1afcbc 830 nfs_unlock_and_release_request(req);
1763da12
FI
831 }
832
833out_put:
834 if (put_dreq(dreq))
835 nfs_direct_write_complete(dreq, hdr->inode);
836 hdr->release(hdr);
837}
838
3e9e0ca3
TM
839static void nfs_write_sync_pgio_error(struct list_head *head)
840{
841 struct nfs_page *req;
842
843 while (!list_empty(head)) {
844 req = nfs_list_entry(head->next);
845 nfs_list_remove_request(req);
1d1afcbc 846 nfs_unlock_and_release_request(req);
3e9e0ca3
TM
847 }
848}
849
dc602dd7
TM
850static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
851{
852 struct nfs_direct_req *dreq = hdr->dreq;
853
854 spin_lock(&dreq->lock);
855 if (dreq->error == 0) {
856 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
857 /* fake unstable write to let common nfs resend pages */
858 hdr->verf.committed = NFS_UNSTABLE;
859 hdr->good_bytes = hdr->args.count;
860 }
861 spin_unlock(&dreq->lock);
862}
863
1763da12 864static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
3e9e0ca3 865 .error_cleanup = nfs_write_sync_pgio_error,
1763da12
FI
866 .init_hdr = nfs_direct_pgio_init,
867 .completion = nfs_direct_write_completion,
dc602dd7 868 .reschedule_io = nfs_direct_write_reschedule_io,
1763da12
FI
869};
870
91f79c43
AV
871
872/*
873 * NB: Return the value of the first error return code. Subsequent
874 * errors after the first one are ignored.
875 */
876/*
877 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
878 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
879 * bail and stop sending more writes. Write length accounting is
880 * handled automatically by nfs_direct_write_result(). Otherwise, if
881 * no requests have been sent, just return an error.
882 */
19f73787 883static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
619d30b4 884 struct iov_iter *iter,
91f79c43 885 loff_t pos)
19f73787 886{
1763da12 887 struct nfs_pageio_descriptor desc;
1d59d61f 888 struct inode *inode = dreq->inode;
19f73787
CL
889 ssize_t result = 0;
890 size_t requested_bytes = 0;
91f79c43 891 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
19f73787 892
a20c93e3 893 nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
1763da12
FI
894 &nfs_direct_write_completion_ops);
895 desc.pg_dreq = dreq;
19f73787 896 get_dreq(dreq);
fe0f07d0 897 inode_dio_begin(inode);
19f73787 898
91f79c43
AV
899 NFS_I(inode)->write_io += iov_iter_count(iter);
900 while (iov_iter_count(iter)) {
901 struct page **pagevec;
902 size_t bytes;
903 size_t pgbase;
904 unsigned npages, i;
905
906 result = iov_iter_get_pages_alloc(iter, &pagevec,
907 wsize, &pgbase);
19f73787
CL
908 if (result < 0)
909 break;
91f79c43
AV
910
911 bytes = result;
912 iov_iter_advance(iter, bytes);
913 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
914 for (i = 0; i < npages; i++) {
915 struct nfs_page *req;
916 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
917
16b90578 918 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
91f79c43
AV
919 pgbase, req_len);
920 if (IS_ERR(req)) {
921 result = PTR_ERR(req);
922 break;
923 }
0a00b77b
WAA
924
925 nfs_direct_setup_mirroring(dreq, &desc, req);
d600ad1f
PT
926 if (desc.pg_error < 0) {
927 nfs_free_request(req);
928 result = desc.pg_error;
929 break;
930 }
0a00b77b 931
91f79c43
AV
932 nfs_lock_request(req);
933 req->wb_index = pos >> PAGE_SHIFT;
934 req->wb_offset = pos & ~PAGE_MASK;
935 if (!nfs_pageio_add_request(&desc, req)) {
936 result = desc.pg_error;
937 nfs_unlock_and_release_request(req);
938 break;
939 }
940 pgbase = 0;
941 bytes -= req_len;
942 requested_bytes += req_len;
943 pos += req_len;
944 dreq->bytes_left -= req_len;
945 }
946 nfs_direct_release_pages(pagevec, npages);
947 kvfree(pagevec);
948 if (result < 0)
19f73787 949 break;
19f73787 950 }
1763da12 951 nfs_pageio_complete(&desc);
19f73787 952
839f7ad6
CL
953 /*
954 * If no bytes were started, return the error, and let the
955 * generic layer handle the completion.
956 */
957 if (requested_bytes == 0) {
fe0f07d0 958 inode_dio_end(inode);
839f7ad6
CL
959 nfs_direct_req_release(dreq);
960 return result < 0 ? result : -EIO;
961 }
962
19f73787
CL
963 if (put_dreq(dreq))
964 nfs_direct_write_complete(dreq, dreq->inode);
839f7ad6 965 return 0;
19f73787
CL
966}
967
1da177e4
LT
968/**
969 * nfs_file_direct_write - file direct write operation for NFS files
970 * @iocb: target I/O control block
619d30b4 971 * @iter: vector of user buffers from which to write data
88467055 972 * @pos: byte offset in file where writing starts
1da177e4
LT
973 *
974 * We use this function for direct writes instead of calling
975 * generic_file_aio_write() in order to avoid taking the inode
976 * semaphore and updating the i_size. The NFS server will set
977 * the new i_size and this client must read the updated size
978 * back into its cache. We let the server do generic write
979 * parameter checking and report problems.
980 *
1da177e4
LT
981 * We eliminate local atime updates, see direct read above.
982 *
983 * We avoid unnecessary page cache invalidations for normal cached
984 * readers of this file.
985 *
986 * Note that O_APPEND is not supported for NFS direct writes, as there
987 * is no atomic O_APPEND write facility in the NFS protocol.
988 */
65a4a1ca 989ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
1da177e4 990{
22cd1bf1 991 ssize_t result = -EINVAL;
1da177e4 992 struct file *file = iocb->ki_filp;
1da177e4 993 struct address_space *mapping = file->f_mapping;
22cd1bf1
CH
994 struct inode *inode = mapping->host;
995 struct nfs_direct_req *dreq;
996 struct nfs_lock_context *l_ctx;
65a4a1ca 997 loff_t pos, end;
c216fd70 998
6de1472f 999 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
3309dd04 1000 file, iov_iter_count(iter), (long long) iocb->ki_pos);
027445c3 1001
3309dd04
AV
1002 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES,
1003 iov_iter_count(iter));
1004
1005 pos = iocb->ki_pos;
09cbfeaf 1006 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
ce1a8e67 1007
5955102c 1008 inode_lock(inode);
a9ab5e84 1009
22cd1bf1
CH
1010 result = nfs_sync_mapping(mapping);
1011 if (result)
a9ab5e84
CH
1012 goto out_unlock;
1013
1014 if (mapping->nrpages) {
1015 result = invalidate_inode_pages2_range(mapping,
09cbfeaf 1016 pos >> PAGE_SHIFT, end);
a9ab5e84
CH
1017 if (result)
1018 goto out_unlock;
1019 }
1da177e4 1020
3309dd04 1021 task_io_account_write(iov_iter_count(iter));
7ec10f26 1022
22cd1bf1
CH
1023 result = -ENOMEM;
1024 dreq = nfs_direct_req_alloc();
1025 if (!dreq)
a9ab5e84 1026 goto out_unlock;
9eafa8cc 1027
22cd1bf1 1028 dreq->inode = inode;
3309dd04 1029 dreq->bytes_left = iov_iter_count(iter);
5fadeb47 1030 dreq->io_start = pos;
22cd1bf1
CH
1031 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1032 l_ctx = nfs_get_lock_context(dreq->ctx);
1033 if (IS_ERR(l_ctx)) {
1034 result = PTR_ERR(l_ctx);
1035 goto out_release;
1036 }
1037 dreq->l_ctx = l_ctx;
1038 if (!is_sync_kiocb(iocb))
1039 dreq->iocb = iocb;
1040
91f79c43 1041 result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
a9ab5e84
CH
1042
1043 if (mapping->nrpages) {
1044 invalidate_inode_pages2_range(mapping,
09cbfeaf 1045 pos >> PAGE_SHIFT, end);
a9ab5e84
CH
1046 }
1047
5955102c 1048 inode_unlock(inode);
a9ab5e84 1049
22cd1bf1
CH
1050 if (!result) {
1051 result = nfs_direct_wait(dreq);
1052 if (result > 0) {
1053 struct inode *inode = mapping->host;
1054
1055 iocb->ki_pos = pos + result;
1056 spin_lock(&inode->i_lock);
1057 if (i_size_read(inode) < iocb->ki_pos)
1058 i_size_write(inode, iocb->ki_pos);
1059 spin_unlock(&inode->i_lock);
a0815d55 1060 generic_write_sync(file, pos, result);
22cd1bf1 1061 }
1763da12 1062 }
a9ab5e84
CH
1063 nfs_direct_req_release(dreq);
1064 return result;
1065
22cd1bf1
CH
1066out_release:
1067 nfs_direct_req_release(dreq);
a9ab5e84 1068out_unlock:
5955102c 1069 inode_unlock(inode);
22cd1bf1 1070 return result;
1da177e4
LT
1071}
1072
88467055
CL
1073/**
1074 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1075 *
1076 */
f7b422b1 1077int __init nfs_init_directcache(void)
1da177e4
LT
1078{
1079 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1080 sizeof(struct nfs_direct_req),
fffb60f9
PJ
1081 0, (SLAB_RECLAIM_ACCOUNT|
1082 SLAB_MEM_SPREAD),
20c2df83 1083 NULL);
1da177e4
LT
1084 if (nfs_direct_cachep == NULL)
1085 return -ENOMEM;
1086
1087 return 0;
1088}
1089
88467055 1090/**
f7b422b1 1091 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
88467055
CL
1092 *
1093 */
266bee88 1094void nfs_destroy_directcache(void)
1da177e4 1095{
1a1d92c1 1096 kmem_cache_destroy(nfs_direct_cachep);
1da177e4 1097}