]>
Commit | Line | Data |
---|---|---|
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 | ||
7c0f6ba6 | 55 | #include <linux/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 | 64 | static 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 |
69 | struct nfs_direct_mirror { |
70 | ssize_t count; | |
71 | }; | |
72 | ||
1da177e4 LT |
73 | struct 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 | ||
8acb2a93 | 89 | loff_t io_start; /* Start offset for I/O */ |
15ce4a0c | 90 | ssize_t count, /* bytes actually processed */ |
ed3743a6 | 91 | max_count, /* max expected count */ |
35754bc0 | 92 | bytes_left, /* bytes left to be sent */ |
1da177e4 | 93 | error; /* any reported error */ |
d72b7a6b | 94 | struct completion completion; /* wait for i/o completion */ |
fad61490 TM |
95 | |
96 | /* commit state */ | |
1763da12 FI |
97 | struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo */ |
98 | struct pnfs_ds_commit_info ds_cinfo; /* Storage for cinfo */ | |
99 | struct work_struct work; | |
fad61490 | 100 | int flags; |
999f3743 | 101 | /* for write */ |
fad61490 TM |
102 | #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */ |
103 | #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */ | |
999f3743 DK |
104 | /* for read */ |
105 | #define NFS_ODIRECT_SHOULD_DIRTY (3) /* dirty user-space page after read */ | |
fad61490 | 106 | struct nfs_writeverf verf; /* unstable write verifier */ |
1da177e4 LT |
107 | }; |
108 | ||
1763da12 FI |
109 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops; |
110 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops; | |
4d3b55d3 | 111 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq); |
1763da12 | 112 | static void nfs_direct_write_schedule_work(struct work_struct *work); |
607f31e8 TM |
113 | |
114 | static inline void get_dreq(struct nfs_direct_req *dreq) | |
115 | { | |
116 | atomic_inc(&dreq->io_count); | |
117 | } | |
118 | ||
119 | static inline int put_dreq(struct nfs_direct_req *dreq) | |
120 | { | |
121 | return atomic_dec_and_test(&dreq->io_count); | |
122 | } | |
123 | ||
0a00b77b | 124 | static void |
60e9e1ff TM |
125 | nfs_direct_handle_truncated(struct nfs_direct_req *dreq, |
126 | const struct nfs_pgio_header *hdr, | |
127 | ssize_t dreq_len) | |
0a00b77b | 128 | { |
60e9e1ff TM |
129 | struct nfs_direct_mirror *mirror = &dreq->mirrors[hdr->pgio_mirror_idx]; |
130 | ||
131 | if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) || | |
132 | test_bit(NFS_IOHDR_EOF, &hdr->flags))) | |
133 | return; | |
134 | if (dreq->max_count >= dreq_len) { | |
135 | dreq->max_count = dreq_len; | |
136 | if (dreq->count > dreq_len) | |
137 | dreq->count = dreq_len; | |
138 | ||
139 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) | |
140 | dreq->error = hdr->error; | |
141 | else /* Clear outstanding error if this is EOF */ | |
142 | dreq->error = 0; | |
143 | } | |
144 | if (mirror->count > dreq_len) | |
145 | mirror->count = dreq_len; | |
146 | } | |
0a00b77b | 147 | |
60e9e1ff TM |
148 | static void |
149 | nfs_direct_count_bytes(struct nfs_direct_req *dreq, | |
150 | const struct nfs_pgio_header *hdr) | |
151 | { | |
152 | struct nfs_direct_mirror *mirror = &dreq->mirrors[hdr->pgio_mirror_idx]; | |
153 | loff_t hdr_end = hdr->io_start + hdr->good_bytes; | |
154 | ssize_t dreq_len = 0; | |
ed3743a6 | 155 | |
60e9e1ff TM |
156 | if (hdr_end > dreq->io_start) |
157 | dreq_len = hdr_end - dreq->io_start; | |
0a00b77b | 158 | |
60e9e1ff | 159 | nfs_direct_handle_truncated(dreq, hdr, dreq_len); |
0a00b77b | 160 | |
60e9e1ff TM |
161 | if (dreq_len > dreq->max_count) |
162 | dreq_len = dreq->max_count; | |
163 | ||
164 | if (mirror->count < dreq_len) | |
165 | mirror->count = dreq_len; | |
166 | if (dreq->count < dreq_len) | |
167 | dreq->count = dreq_len; | |
0a00b77b WAA |
168 | } |
169 | ||
5002c586 WAA |
170 | /* |
171 | * nfs_direct_select_verf - select the right verifier | |
172 | * @dreq - direct request possibly spanning multiple servers | |
173 | * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs | |
6cccbb6f | 174 | * @commit_idx - commit bucket index for the DS |
5002c586 WAA |
175 | * |
176 | * returns the correct verifier to use given the role of the server | |
177 | */ | |
178 | static struct nfs_writeverf * | |
179 | nfs_direct_select_verf(struct nfs_direct_req *dreq, | |
180 | struct nfs_client *ds_clp, | |
6cccbb6f | 181 | int commit_idx) |
5002c586 WAA |
182 | { |
183 | struct nfs_writeverf *verfp = &dreq->verf; | |
184 | ||
185 | #ifdef CONFIG_NFS_V4_1 | |
834e465b KM |
186 | /* |
187 | * pNFS is in use, use the DS verf except commit_through_mds is set | |
188 | * for layout segment where nbuckets is zero. | |
189 | */ | |
190 | if (ds_clp && dreq->ds_cinfo.nbuckets > 0) { | |
6cccbb6f WAA |
191 | if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets) |
192 | verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf; | |
5002c586 WAA |
193 | else |
194 | WARN_ON_ONCE(1); | |
195 | } | |
196 | #endif | |
197 | return verfp; | |
198 | } | |
199 | ||
200 | ||
201 | /* | |
202 | * nfs_direct_set_hdr_verf - set the write/commit verifier | |
203 | * @dreq - direct request possibly spanning multiple servers | |
204 | * @hdr - pageio header to validate against previously seen verfs | |
205 | * | |
206 | * Set the server's (MDS or DS) "seen" verifier | |
207 | */ | |
208 | static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq, | |
209 | struct nfs_pgio_header *hdr) | |
210 | { | |
211 | struct nfs_writeverf *verfp; | |
212 | ||
6cccbb6f | 213 | verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx); |
5002c586 WAA |
214 | WARN_ON_ONCE(verfp->committed >= 0); |
215 | memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf)); | |
216 | WARN_ON_ONCE(verfp->committed < 0); | |
217 | } | |
218 | ||
8fc3c386 TM |
219 | static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1, |
220 | const struct nfs_writeverf *v2) | |
221 | { | |
222 | return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier); | |
223 | } | |
224 | ||
5002c586 WAA |
225 | /* |
226 | * nfs_direct_cmp_hdr_verf - compare verifier for pgio header | |
227 | * @dreq - direct request possibly spanning multiple servers | |
228 | * @hdr - pageio header to validate against previously seen verf | |
229 | * | |
230 | * set the server's "seen" verf if not initialized. | |
231 | * returns result of comparison between @hdr->verf and the "seen" | |
232 | * verf of the server used by @hdr (DS or MDS) | |
233 | */ | |
234 | static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq, | |
235 | struct nfs_pgio_header *hdr) | |
236 | { | |
237 | struct nfs_writeverf *verfp; | |
238 | ||
6cccbb6f | 239 | verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx); |
5002c586 WAA |
240 | if (verfp->committed < 0) { |
241 | nfs_direct_set_hdr_verf(dreq, hdr); | |
242 | return 0; | |
243 | } | |
8fc3c386 | 244 | return nfs_direct_cmp_verf(verfp, &hdr->verf); |
5002c586 WAA |
245 | } |
246 | ||
5002c586 WAA |
247 | /* |
248 | * nfs_direct_cmp_commit_data_verf - compare verifier for commit data | |
249 | * @dreq - direct request possibly spanning multiple servers | |
250 | * @data - commit data to validate against previously seen verf | |
251 | * | |
252 | * returns result of comparison between @data->verf and the verf of | |
253 | * the server used by @data (DS or MDS) | |
254 | */ | |
255 | static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq, | |
256 | struct nfs_commit_data *data) | |
257 | { | |
258 | struct nfs_writeverf *verfp; | |
259 | ||
260 | verfp = nfs_direct_select_verf(dreq, data->ds_clp, | |
261 | data->ds_commit_index); | |
80c76fe3 WAA |
262 | |
263 | /* verifier not set so always fail */ | |
443e7db4 | 264 | if (verfp->committed < 0 || data->res.verf->committed <= NFS_UNSTABLE) |
80c76fe3 WAA |
265 | return 1; |
266 | ||
443e7db4 | 267 | return nfs_direct_cmp_verf(verfp, data->res.verf); |
5002c586 | 268 | } |
5002c586 | 269 | |
1da177e4 | 270 | /** |
b8a32e2b | 271 | * nfs_direct_IO - NFS address space operation for direct I/O |
b8a32e2b | 272 | * @iocb: target I/O control block |
90090ae6 | 273 | * @iter: I/O buffer |
b8a32e2b CL |
274 | * |
275 | * The presence of this routine in the address space ops vector means | |
a564b8f0 MG |
276 | * the NFS client supports direct I/O. However, for most direct IO, we |
277 | * shunt off direct read and write requests before the VFS gets them, | |
278 | * so this method is only ever called for swap. | |
1da177e4 | 279 | */ |
c8b8e32d | 280 | ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter) |
b8a32e2b | 281 | { |
ee8a1a8b PT |
282 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
283 | ||
284 | /* we only support swap file calling nfs_direct_IO */ | |
285 | if (!IS_SWAPFILE(inode)) | |
286 | return 0; | |
287 | ||
66ee59af | 288 | VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE); |
a564b8f0 | 289 | |
6f673763 | 290 | if (iov_iter_rw(iter) == READ) |
c8b8e32d | 291 | return nfs_file_direct_read(iocb, iter); |
65a4a1ca | 292 | return nfs_file_direct_write(iocb, iter); |
b8a32e2b CL |
293 | } |
294 | ||
749e146e | 295 | static void nfs_direct_release_pages(struct page **pages, unsigned int npages) |
9c93ab7d | 296 | { |
749e146e | 297 | unsigned int i; |
607f31e8 | 298 | for (i = 0; i < npages; i++) |
09cbfeaf | 299 | put_page(pages[i]); |
6b45d858 TM |
300 | } |
301 | ||
1763da12 FI |
302 | void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo, |
303 | struct nfs_direct_req *dreq) | |
304 | { | |
fe238e60 | 305 | cinfo->inode = dreq->inode; |
1763da12 FI |
306 | cinfo->mds = &dreq->mds_cinfo; |
307 | cinfo->ds = &dreq->ds_cinfo; | |
308 | cinfo->dreq = dreq; | |
309 | cinfo->completion_ops = &nfs_direct_commit_completion_ops; | |
310 | } | |
311 | ||
0a00b77b WAA |
312 | static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq, |
313 | struct nfs_pageio_descriptor *pgio, | |
314 | struct nfs_page *req) | |
315 | { | |
316 | int mirror_count = 1; | |
317 | ||
318 | if (pgio->pg_ops->pg_get_mirror_count) | |
319 | mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req); | |
320 | ||
321 | dreq->mirror_count = mirror_count; | |
322 | } | |
323 | ||
93619e59 | 324 | static inline struct nfs_direct_req *nfs_direct_req_alloc(void) |
1da177e4 | 325 | { |
93619e59 CL |
326 | struct nfs_direct_req *dreq; |
327 | ||
292f3eee | 328 | dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL); |
93619e59 CL |
329 | if (!dreq) |
330 | return NULL; | |
331 | ||
332 | kref_init(&dreq->kref); | |
607f31e8 | 333 | kref_get(&dreq->kref); |
d72b7a6b | 334 | init_completion(&dreq->completion); |
1763da12 | 335 | INIT_LIST_HEAD(&dreq->mds_cinfo.list); |
5002c586 | 336 | dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */ |
1763da12 | 337 | INIT_WORK(&dreq->work, nfs_direct_write_schedule_work); |
0a00b77b | 338 | dreq->mirror_count = 1; |
15ce4a0c | 339 | spin_lock_init(&dreq->lock); |
93619e59 CL |
340 | |
341 | return dreq; | |
1da177e4 LT |
342 | } |
343 | ||
b4946ffb | 344 | static void nfs_direct_req_free(struct kref *kref) |
1da177e4 LT |
345 | { |
346 | struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); | |
a8881f5a | 347 | |
8c393f9a | 348 | nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo); |
f11ac8db TM |
349 | if (dreq->l_ctx != NULL) |
350 | nfs_put_lock_context(dreq->l_ctx); | |
a8881f5a TM |
351 | if (dreq->ctx != NULL) |
352 | put_nfs_open_context(dreq->ctx); | |
1da177e4 LT |
353 | kmem_cache_free(nfs_direct_cachep, dreq); |
354 | } | |
355 | ||
b4946ffb TM |
356 | static void nfs_direct_req_release(struct nfs_direct_req *dreq) |
357 | { | |
358 | kref_put(&dreq->kref, nfs_direct_req_free); | |
359 | } | |
360 | ||
6296556f PT |
361 | ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq) |
362 | { | |
363 | return dreq->bytes_left; | |
364 | } | |
365 | EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left); | |
366 | ||
bc0fb201 CL |
367 | /* |
368 | * Collects and returns the final error value/byte-count. | |
369 | */ | |
370 | static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) | |
371 | { | |
15ce4a0c | 372 | ssize_t result = -EIOCBQUEUED; |
bc0fb201 CL |
373 | |
374 | /* Async requests don't wait here */ | |
375 | if (dreq->iocb) | |
376 | goto out; | |
377 | ||
150030b7 | 378 | result = wait_for_completion_killable(&dreq->completion); |
bc0fb201 | 379 | |
d2a7de0b TM |
380 | if (!result) { |
381 | result = dreq->count; | |
382 | WARN_ON_ONCE(dreq->count < 0); | |
383 | } | |
bc0fb201 | 384 | if (!result) |
15ce4a0c | 385 | result = dreq->error; |
bc0fb201 CL |
386 | |
387 | out: | |
bc0fb201 CL |
388 | return (ssize_t) result; |
389 | } | |
390 | ||
63ab46ab | 391 | /* |
607f31e8 TM |
392 | * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust |
393 | * the iocb is still valid here if this is a synchronous request. | |
63ab46ab | 394 | */ |
f7b5c340 | 395 | static void nfs_direct_complete(struct nfs_direct_req *dreq) |
63ab46ab | 396 | { |
9811cd57 CH |
397 | struct inode *inode = dreq->inode; |
398 | ||
fe0f07d0 | 399 | inode_dio_end(inode); |
2a009ec9 CH |
400 | |
401 | if (dreq->iocb) { | |
15ce4a0c | 402 | long res = (long) dreq->error; |
d2a7de0b | 403 | if (dreq->count != 0) { |
15ce4a0c | 404 | res = (long) dreq->count; |
d2a7de0b TM |
405 | WARN_ON_ONCE(dreq->count < 0); |
406 | } | |
04b2fa9f | 407 | dreq->iocb->ki_complete(dreq->iocb, res, 0); |
d72b7a6b | 408 | } |
2a009ec9 | 409 | |
024de8f1 | 410 | complete(&dreq->completion); |
63ab46ab | 411 | |
b4946ffb | 412 | nfs_direct_req_release(dreq); |
63ab46ab CL |
413 | } |
414 | ||
584aa810 | 415 | static void nfs_direct_read_completion(struct nfs_pgio_header *hdr) |
fdd1e74c | 416 | { |
584aa810 FI |
417 | unsigned long bytes = 0; |
418 | struct nfs_direct_req *dreq = hdr->dreq; | |
fdd1e74c | 419 | |
15ce4a0c | 420 | spin_lock(&dreq->lock); |
a206ab6f TM |
421 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) { |
422 | spin_unlock(&dreq->lock); | |
423 | goto out_put; | |
424 | } | |
425 | ||
60e9e1ff | 426 | nfs_direct_count_bytes(dreq, hdr); |
584aa810 FI |
427 | spin_unlock(&dreq->lock); |
428 | ||
4bd8b010 TM |
429 | while (!list_empty(&hdr->pages)) { |
430 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
431 | struct page *page = req->wb_page; | |
432 | ||
999f3743 DK |
433 | if (!PageCompound(page) && bytes < hdr->good_bytes && |
434 | (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY)) | |
be7e9858 | 435 | set_page_dirty(page); |
4bd8b010 TM |
436 | bytes += req->wb_bytes; |
437 | nfs_list_remove_request(req); | |
beeb5338 | 438 | nfs_release_request(req); |
d4a8f367 | 439 | } |
584aa810 | 440 | out_put: |
607f31e8 | 441 | if (put_dreq(dreq)) |
f7b5c340 | 442 | nfs_direct_complete(dreq); |
584aa810 | 443 | hdr->release(hdr); |
1da177e4 LT |
444 | } |
445 | ||
c81ad389 | 446 | static void nfs_read_sync_pgio_error(struct list_head *head, int error) |
cd841605 | 447 | { |
584aa810 | 448 | struct nfs_page *req; |
cd841605 | 449 | |
584aa810 FI |
450 | while (!list_empty(head)) { |
451 | req = nfs_list_entry(head->next); | |
452 | nfs_list_remove_request(req); | |
453 | nfs_release_request(req); | |
454 | } | |
cd841605 FI |
455 | } |
456 | ||
584aa810 FI |
457 | static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr) |
458 | { | |
459 | get_dreq(hdr->dreq); | |
460 | } | |
461 | ||
462 | static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = { | |
3e9e0ca3 | 463 | .error_cleanup = nfs_read_sync_pgio_error, |
584aa810 FI |
464 | .init_hdr = nfs_direct_pgio_init, |
465 | .completion = nfs_direct_read_completion, | |
466 | }; | |
467 | ||
d4cc948b | 468 | /* |
607f31e8 TM |
469 | * For each rsize'd chunk of the user's buffer, dispatch an NFS READ |
470 | * operation. If nfs_readdata_alloc() or get_user_pages() fails, | |
471 | * bail and stop sending more reads. Read length accounting is | |
472 | * handled automatically by nfs_direct_read_result(). Otherwise, if | |
473 | * no requests have been sent, just return an error. | |
1da177e4 | 474 | */ |
1da177e4 | 475 | |
91f79c43 AV |
476 | static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq, |
477 | struct iov_iter *iter, | |
478 | loff_t pos) | |
479 | { | |
480 | struct nfs_pageio_descriptor desc; | |
481 | struct inode *inode = dreq->inode; | |
482 | ssize_t result = -EINVAL; | |
483 | size_t requested_bytes = 0; | |
484 | size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE); | |
e9f7bee1 | 485 | |
16b90578 | 486 | nfs_pageio_init_read(&desc, dreq->inode, false, |
91f79c43 AV |
487 | &nfs_direct_read_completion_ops); |
488 | get_dreq(dreq); | |
489 | desc.pg_dreq = dreq; | |
fe0f07d0 | 490 | inode_dio_begin(inode); |
a564b8f0 | 491 | |
91f79c43 AV |
492 | while (iov_iter_count(iter)) { |
493 | struct page **pagevec; | |
494 | size_t bytes; | |
495 | size_t pgbase; | |
496 | unsigned npages, i; | |
607f31e8 | 497 | |
91f79c43 AV |
498 | result = iov_iter_get_pages_alloc(iter, &pagevec, |
499 | rsize, &pgbase); | |
500 | if (result < 0) | |
501 | break; | |
502 | ||
503 | bytes = result; | |
504 | iov_iter_advance(iter, bytes); | |
505 | npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; | |
584aa810 FI |
506 | for (i = 0; i < npages; i++) { |
507 | struct nfs_page *req; | |
bf5fc402 | 508 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); |
584aa810 | 509 | /* XXX do we need to do the eof zeroing found in async_filler? */ |
2bfc6e56 | 510 | req = nfs_create_request(dreq->ctx, pagevec[i], NULL, |
584aa810 FI |
511 | pgbase, req_len); |
512 | if (IS_ERR(req)) { | |
584aa810 FI |
513 | result = PTR_ERR(req); |
514 | break; | |
515 | } | |
516 | req->wb_index = pos >> PAGE_SHIFT; | |
517 | req->wb_offset = pos & ~PAGE_MASK; | |
91f79c43 AV |
518 | if (!nfs_pageio_add_request(&desc, req)) { |
519 | result = desc.pg_error; | |
584aa810 | 520 | nfs_release_request(req); |
584aa810 FI |
521 | break; |
522 | } | |
523 | pgbase = 0; | |
524 | bytes -= req_len; | |
91f79c43 | 525 | requested_bytes += req_len; |
584aa810 | 526 | pos += req_len; |
35754bc0 | 527 | dreq->bytes_left -= req_len; |
584aa810 | 528 | } |
6d74743b | 529 | nfs_direct_release_pages(pagevec, npages); |
91f79c43 | 530 | kvfree(pagevec); |
19f73787 CL |
531 | if (result < 0) |
532 | break; | |
19f73787 CL |
533 | } |
534 | ||
584aa810 FI |
535 | nfs_pageio_complete(&desc); |
536 | ||
839f7ad6 CL |
537 | /* |
538 | * If no bytes were started, return the error, and let the | |
539 | * generic layer handle the completion. | |
540 | */ | |
541 | if (requested_bytes == 0) { | |
fe0f07d0 | 542 | inode_dio_end(inode); |
839f7ad6 CL |
543 | nfs_direct_req_release(dreq); |
544 | return result < 0 ? result : -EIO; | |
545 | } | |
546 | ||
19f73787 | 547 | if (put_dreq(dreq)) |
f7b5c340 | 548 | nfs_direct_complete(dreq); |
85128b2b | 549 | return requested_bytes; |
19f73787 CL |
550 | } |
551 | ||
14a3ec79 CH |
552 | /** |
553 | * nfs_file_direct_read - file direct read operation for NFS files | |
554 | * @iocb: target I/O control block | |
619d30b4 | 555 | * @iter: vector of user buffers into which to read data |
14a3ec79 CH |
556 | * |
557 | * We use this function for direct reads instead of calling | |
558 | * generic_file_aio_read() in order to avoid gfar's check to see if | |
559 | * the request starts before the end of the file. For that check | |
560 | * to work, we must generate a GETATTR before each direct read, and | |
561 | * even then there is a window between the GETATTR and the subsequent | |
562 | * READ where the file size could change. Our preference is simply | |
563 | * to do all reads the application wants, and the server will take | |
564 | * care of managing the end of file boundary. | |
565 | * | |
566 | * This function also eliminates unnecessarily updating the file's | |
567 | * atime locally, as the NFS server sets the file's atime, and this | |
568 | * client must read the updated atime from the server back into its | |
569 | * cache. | |
570 | */ | |
c8b8e32d | 571 | ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter) |
1da177e4 | 572 | { |
14a3ec79 CH |
573 | struct file *file = iocb->ki_filp; |
574 | struct address_space *mapping = file->f_mapping; | |
575 | struct inode *inode = mapping->host; | |
1da177e4 | 576 | struct nfs_direct_req *dreq; |
b3c54de6 | 577 | struct nfs_lock_context *l_ctx; |
85128b2b | 578 | ssize_t result = -EINVAL, requested; |
a6cbcd4a | 579 | size_t count = iov_iter_count(iter); |
14a3ec79 CH |
580 | nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); |
581 | ||
582 | dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n", | |
c8b8e32d | 583 | file, count, (long long) iocb->ki_pos); |
14a3ec79 CH |
584 | |
585 | result = 0; | |
586 | if (!count) | |
587 | goto out; | |
588 | ||
14a3ec79 CH |
589 | task_io_account_read(count); |
590 | ||
591 | result = -ENOMEM; | |
607f31e8 | 592 | dreq = nfs_direct_req_alloc(); |
f11ac8db | 593 | if (dreq == NULL) |
a5864c99 | 594 | goto out; |
1da177e4 | 595 | |
91d5b470 | 596 | dreq->inode = inode; |
ed3743a6 | 597 | dreq->bytes_left = dreq->max_count = count; |
c8b8e32d | 598 | dreq->io_start = iocb->ki_pos; |
cd3758e3 | 599 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
b3c54de6 TM |
600 | l_ctx = nfs_get_lock_context(dreq->ctx); |
601 | if (IS_ERR(l_ctx)) { | |
602 | result = PTR_ERR(l_ctx); | |
f11ac8db | 603 | goto out_release; |
b3c54de6 TM |
604 | } |
605 | dreq->l_ctx = l_ctx; | |
487b8372 CL |
606 | if (!is_sync_kiocb(iocb)) |
607 | dreq->iocb = iocb; | |
1da177e4 | 608 | |
999f3743 DK |
609 | if (iter_is_iovec(iter)) |
610 | dreq->flags = NFS_ODIRECT_SHOULD_DIRTY; | |
611 | ||
a5864c99 TM |
612 | nfs_start_io_direct(inode); |
613 | ||
619d30b4 | 614 | NFS_I(inode)->read_io += count; |
85128b2b | 615 | requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos); |
d0b9875d | 616 | |
a5864c99 | 617 | nfs_end_io_direct(inode); |
d0b9875d | 618 | |
85128b2b | 619 | if (requested > 0) { |
607f31e8 | 620 | result = nfs_direct_wait(dreq); |
85128b2b AV |
621 | if (result > 0) { |
622 | requested -= result; | |
c8b8e32d | 623 | iocb->ki_pos += result; |
85128b2b AV |
624 | } |
625 | iov_iter_revert(iter, requested); | |
626 | } else { | |
627 | result = requested; | |
14a3ec79 | 628 | } |
d0b9875d | 629 | |
f11ac8db | 630 | out_release: |
b4946ffb | 631 | nfs_direct_req_release(dreq); |
f11ac8db | 632 | out: |
1da177e4 LT |
633 | return result; |
634 | } | |
635 | ||
085d1e33 TH |
636 | static void |
637 | nfs_direct_write_scan_commit_list(struct inode *inode, | |
638 | struct list_head *list, | |
639 | struct nfs_commit_info *cinfo) | |
640 | { | |
e824f99a | 641 | mutex_lock(&NFS_I(cinfo->inode)->commit_mutex); |
085d1e33 TH |
642 | #ifdef CONFIG_NFS_V4_1 |
643 | if (cinfo->ds != NULL && cinfo->ds->nwritten != 0) | |
644 | NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo); | |
645 | #endif | |
646 | nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0); | |
e824f99a | 647 | mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex); |
085d1e33 TH |
648 | } |
649 | ||
fad61490 TM |
650 | static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) |
651 | { | |
1763da12 FI |
652 | struct nfs_pageio_descriptor desc; |
653 | struct nfs_page *req, *tmp; | |
654 | LIST_HEAD(reqs); | |
655 | struct nfs_commit_info cinfo; | |
656 | LIST_HEAD(failed); | |
0a00b77b | 657 | int i; |
1763da12 FI |
658 | |
659 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
085d1e33 | 660 | nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo); |
1da177e4 | 661 | |
fad61490 | 662 | dreq->count = 0; |
60e9e1ff TM |
663 | dreq->max_count = 0; |
664 | list_for_each_entry(req, &reqs, wb_list) | |
665 | dreq->max_count += req->wb_bytes; | |
a5314a74 TM |
666 | dreq->verf.committed = NFS_INVALID_STABLE_HOW; |
667 | nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo); | |
0a00b77b WAA |
668 | for (i = 0; i < dreq->mirror_count; i++) |
669 | dreq->mirrors[i].count = 0; | |
607f31e8 TM |
670 | get_dreq(dreq); |
671 | ||
a20c93e3 | 672 | nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false, |
1763da12 FI |
673 | &nfs_direct_write_completion_ops); |
674 | desc.pg_dreq = dreq; | |
fedb595c | 675 | |
0a00b77b WAA |
676 | req = nfs_list_entry(reqs.next); |
677 | nfs_direct_setup_mirroring(dreq, &desc, req); | |
d600ad1f PT |
678 | if (desc.pg_error < 0) { |
679 | list_splice_init(&reqs, &failed); | |
680 | goto out_failed; | |
681 | } | |
0a00b77b | 682 | |
1763da12 FI |
683 | list_for_each_entry_safe(req, tmp, &reqs, wb_list) { |
684 | if (!nfs_pageio_add_request(&desc, req)) { | |
f8e8a134 | 685 | nfs_list_move_request(req, &failed); |
fe238e60 | 686 | spin_lock(&cinfo.inode->i_lock); |
1763da12 | 687 | dreq->flags = 0; |
d600ad1f PT |
688 | if (desc.pg_error < 0) |
689 | dreq->error = desc.pg_error; | |
690 | else | |
691 | dreq->error = -EIO; | |
fe238e60 | 692 | spin_unlock(&cinfo.inode->i_lock); |
1763da12 | 693 | } |
5a695da2 | 694 | nfs_release_request(req); |
1763da12 FI |
695 | } |
696 | nfs_pageio_complete(&desc); | |
fad61490 | 697 | |
d600ad1f | 698 | out_failed: |
4035c248 TM |
699 | while (!list_empty(&failed)) { |
700 | req = nfs_list_entry(failed.next); | |
701 | nfs_list_remove_request(req); | |
1d1afcbc | 702 | nfs_unlock_and_release_request(req); |
4035c248 | 703 | } |
fad61490 | 704 | |
1763da12 | 705 | if (put_dreq(dreq)) |
4d3b55d3 | 706 | nfs_direct_write_complete(dreq); |
c9d8f89d TM |
707 | } |
708 | ||
1763da12 | 709 | static void nfs_direct_commit_complete(struct nfs_commit_data *data) |
c9d8f89d | 710 | { |
0b7c0153 | 711 | struct nfs_direct_req *dreq = data->dreq; |
1763da12 FI |
712 | struct nfs_commit_info cinfo; |
713 | struct nfs_page *req; | |
c9d8f89d TM |
714 | int status = data->task.tk_status; |
715 | ||
1763da12 | 716 | nfs_init_cinfo_from_dreq(&cinfo, dreq); |
fe4f844d | 717 | if (status < 0 || nfs_direct_cmp_commit_data_verf(dreq, data)) |
fad61490 | 718 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
1da177e4 | 719 | |
1763da12 FI |
720 | while (!list_empty(&data->pages)) { |
721 | req = nfs_list_entry(data->pages.next); | |
722 | nfs_list_remove_request(req); | |
723 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) { | |
724 | /* Note the rewrite will go through mds */ | |
b57ff130 | 725 | nfs_mark_request_commit(req, NULL, &cinfo, 0); |
906369e4 FI |
726 | } else |
727 | nfs_release_request(req); | |
1d1afcbc | 728 | nfs_unlock_and_release_request(req); |
1763da12 FI |
729 | } |
730 | ||
731 | if (atomic_dec_and_test(&cinfo.mds->rpcs_out)) | |
4d3b55d3 | 732 | nfs_direct_write_complete(dreq); |
1da177e4 LT |
733 | } |
734 | ||
b20135d0 TM |
735 | static void nfs_direct_resched_write(struct nfs_commit_info *cinfo, |
736 | struct nfs_page *req) | |
1763da12 | 737 | { |
b20135d0 TM |
738 | struct nfs_direct_req *dreq = cinfo->dreq; |
739 | ||
740 | spin_lock(&dreq->lock); | |
741 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; | |
742 | spin_unlock(&dreq->lock); | |
743 | nfs_mark_request_commit(req, NULL, cinfo, 0); | |
1763da12 FI |
744 | } |
745 | ||
746 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = { | |
747 | .completion = nfs_direct_commit_complete, | |
b20135d0 | 748 | .resched_write = nfs_direct_resched_write, |
fad61490 TM |
749 | }; |
750 | ||
751 | static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) | |
1da177e4 | 752 | { |
1763da12 FI |
753 | int res; |
754 | struct nfs_commit_info cinfo; | |
755 | LIST_HEAD(mds_list); | |
756 | ||
757 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
758 | nfs_scan_commit(dreq->inode, &mds_list, &cinfo); | |
759 | res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo); | |
760 | if (res < 0) /* res == -ENOMEM */ | |
761 | nfs_direct_write_reschedule(dreq); | |
fad61490 | 762 | } |
1da177e4 | 763 | |
1763da12 | 764 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
fad61490 | 765 | { |
1763da12 | 766 | struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work); |
fad61490 | 767 | int flags = dreq->flags; |
1da177e4 | 768 | |
fad61490 TM |
769 | dreq->flags = 0; |
770 | switch (flags) { | |
771 | case NFS_ODIRECT_DO_COMMIT: | |
772 | nfs_direct_commit_schedule(dreq); | |
1da177e4 | 773 | break; |
fad61490 TM |
774 | case NFS_ODIRECT_RESCHED_WRITES: |
775 | nfs_direct_write_reschedule(dreq); | |
776 | break; | |
777 | default: | |
f7b5c340 TM |
778 | nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping); |
779 | nfs_direct_complete(dreq); | |
fad61490 TM |
780 | } |
781 | } | |
1da177e4 | 782 | |
4d3b55d3 | 783 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq) |
fad61490 | 784 | { |
1763da12 | 785 | schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */ |
fad61490 | 786 | } |
1763da12 | 787 | |
1763da12 FI |
788 | static void nfs_direct_write_completion(struct nfs_pgio_header *hdr) |
789 | { | |
790 | struct nfs_direct_req *dreq = hdr->dreq; | |
791 | struct nfs_commit_info cinfo; | |
c65e6254 | 792 | bool request_commit = false; |
1763da12 FI |
793 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); |
794 | ||
1763da12 FI |
795 | nfs_init_cinfo_from_dreq(&cinfo, dreq); |
796 | ||
797 | spin_lock(&dreq->lock); | |
a206ab6f TM |
798 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) { |
799 | spin_unlock(&dreq->lock); | |
800 | goto out_put; | |
801 | } | |
802 | ||
60e9e1ff | 803 | nfs_direct_count_bytes(dreq, hdr); |
a206ab6f | 804 | if (hdr->good_bytes != 0) { |
c65e6254 | 805 | if (nfs_write_need_commit(hdr)) { |
1763da12 | 806 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) |
c65e6254 | 807 | request_commit = true; |
1763da12 | 808 | else if (dreq->flags == 0) { |
5002c586 | 809 | nfs_direct_set_hdr_verf(dreq, hdr); |
c65e6254 | 810 | request_commit = true; |
1763da12 FI |
811 | dreq->flags = NFS_ODIRECT_DO_COMMIT; |
812 | } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) { | |
c65e6254 WAA |
813 | request_commit = true; |
814 | if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr)) | |
5002c586 WAA |
815 | dreq->flags = |
816 | NFS_ODIRECT_RESCHED_WRITES; | |
1763da12 FI |
817 | } |
818 | } | |
819 | } | |
820 | spin_unlock(&dreq->lock); | |
821 | ||
822 | while (!list_empty(&hdr->pages)) { | |
2bfc6e56 | 823 | |
1763da12 FI |
824 | req = nfs_list_entry(hdr->pages.next); |
825 | nfs_list_remove_request(req); | |
c65e6254 | 826 | if (request_commit) { |
04277086 | 827 | kref_get(&req->wb_kref); |
b57ff130 WAA |
828 | nfs_mark_request_commit(req, hdr->lseg, &cinfo, |
829 | hdr->ds_commit_idx); | |
1763da12 | 830 | } |
1d1afcbc | 831 | nfs_unlock_and_release_request(req); |
1763da12 FI |
832 | } |
833 | ||
834 | out_put: | |
835 | if (put_dreq(dreq)) | |
4d3b55d3 | 836 | nfs_direct_write_complete(dreq); |
1763da12 FI |
837 | hdr->release(hdr); |
838 | } | |
839 | ||
c81ad389 | 840 | static void nfs_write_sync_pgio_error(struct list_head *head, int error) |
3e9e0ca3 TM |
841 | { |
842 | struct nfs_page *req; | |
843 | ||
844 | while (!list_empty(head)) { | |
845 | req = nfs_list_entry(head->next); | |
846 | nfs_list_remove_request(req); | |
1d1afcbc | 847 | nfs_unlock_and_release_request(req); |
3e9e0ca3 TM |
848 | } |
849 | } | |
850 | ||
dc602dd7 TM |
851 | static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr) |
852 | { | |
853 | struct nfs_direct_req *dreq = hdr->dreq; | |
854 | ||
855 | spin_lock(&dreq->lock); | |
856 | if (dreq->error == 0) { | |
857 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; | |
858 | /* fake unstable write to let common nfs resend pages */ | |
859 | hdr->verf.committed = NFS_UNSTABLE; | |
860 | hdr->good_bytes = hdr->args.count; | |
861 | } | |
862 | spin_unlock(&dreq->lock); | |
863 | } | |
864 | ||
1763da12 | 865 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = { |
3e9e0ca3 | 866 | .error_cleanup = nfs_write_sync_pgio_error, |
1763da12 FI |
867 | .init_hdr = nfs_direct_pgio_init, |
868 | .completion = nfs_direct_write_completion, | |
dc602dd7 | 869 | .reschedule_io = nfs_direct_write_reschedule_io, |
1763da12 FI |
870 | }; |
871 | ||
91f79c43 AV |
872 | |
873 | /* | |
874 | * NB: Return the value of the first error return code. Subsequent | |
875 | * errors after the first one are ignored. | |
876 | */ | |
877 | /* | |
878 | * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE | |
879 | * operation. If nfs_writedata_alloc() or get_user_pages() fails, | |
880 | * bail and stop sending more writes. Write length accounting is | |
881 | * handled automatically by nfs_direct_write_result(). Otherwise, if | |
882 | * no requests have been sent, just return an error. | |
883 | */ | |
19f73787 | 884 | static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, |
619d30b4 | 885 | struct iov_iter *iter, |
91f79c43 | 886 | loff_t pos) |
19f73787 | 887 | { |
1763da12 | 888 | struct nfs_pageio_descriptor desc; |
1d59d61f | 889 | struct inode *inode = dreq->inode; |
19f73787 CL |
890 | ssize_t result = 0; |
891 | size_t requested_bytes = 0; | |
91f79c43 | 892 | size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE); |
19f73787 | 893 | |
a20c93e3 | 894 | nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false, |
1763da12 FI |
895 | &nfs_direct_write_completion_ops); |
896 | desc.pg_dreq = dreq; | |
19f73787 | 897 | get_dreq(dreq); |
fe0f07d0 | 898 | inode_dio_begin(inode); |
19f73787 | 899 | |
91f79c43 AV |
900 | NFS_I(inode)->write_io += iov_iter_count(iter); |
901 | while (iov_iter_count(iter)) { | |
902 | struct page **pagevec; | |
903 | size_t bytes; | |
904 | size_t pgbase; | |
905 | unsigned npages, i; | |
906 | ||
907 | result = iov_iter_get_pages_alloc(iter, &pagevec, | |
908 | wsize, &pgbase); | |
19f73787 CL |
909 | if (result < 0) |
910 | break; | |
91f79c43 AV |
911 | |
912 | bytes = result; | |
913 | iov_iter_advance(iter, bytes); | |
914 | npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; | |
915 | for (i = 0; i < npages; i++) { | |
916 | struct nfs_page *req; | |
917 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); | |
918 | ||
16b90578 | 919 | req = nfs_create_request(dreq->ctx, pagevec[i], NULL, |
91f79c43 AV |
920 | pgbase, req_len); |
921 | if (IS_ERR(req)) { | |
922 | result = PTR_ERR(req); | |
923 | break; | |
924 | } | |
0a00b77b WAA |
925 | |
926 | nfs_direct_setup_mirroring(dreq, &desc, req); | |
d600ad1f PT |
927 | if (desc.pg_error < 0) { |
928 | nfs_free_request(req); | |
929 | result = desc.pg_error; | |
930 | break; | |
931 | } | |
0a00b77b | 932 | |
91f79c43 AV |
933 | nfs_lock_request(req); |
934 | req->wb_index = pos >> PAGE_SHIFT; | |
935 | req->wb_offset = pos & ~PAGE_MASK; | |
936 | if (!nfs_pageio_add_request(&desc, req)) { | |
937 | result = desc.pg_error; | |
938 | nfs_unlock_and_release_request(req); | |
939 | break; | |
940 | } | |
941 | pgbase = 0; | |
942 | bytes -= req_len; | |
943 | requested_bytes += req_len; | |
944 | pos += req_len; | |
945 | dreq->bytes_left -= req_len; | |
946 | } | |
947 | nfs_direct_release_pages(pagevec, npages); | |
948 | kvfree(pagevec); | |
949 | if (result < 0) | |
19f73787 | 950 | break; |
19f73787 | 951 | } |
1763da12 | 952 | nfs_pageio_complete(&desc); |
19f73787 | 953 | |
839f7ad6 CL |
954 | /* |
955 | * If no bytes were started, return the error, and let the | |
956 | * generic layer handle the completion. | |
957 | */ | |
958 | if (requested_bytes == 0) { | |
fe0f07d0 | 959 | inode_dio_end(inode); |
839f7ad6 CL |
960 | nfs_direct_req_release(dreq); |
961 | return result < 0 ? result : -EIO; | |
962 | } | |
963 | ||
19f73787 | 964 | if (put_dreq(dreq)) |
4d3b55d3 | 965 | nfs_direct_write_complete(dreq); |
85128b2b | 966 | return requested_bytes; |
19f73787 CL |
967 | } |
968 | ||
1da177e4 LT |
969 | /** |
970 | * nfs_file_direct_write - file direct write operation for NFS files | |
971 | * @iocb: target I/O control block | |
619d30b4 | 972 | * @iter: vector of user buffers from which to write data |
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 | 989 | ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter) |
1da177e4 | 990 | { |
85128b2b | 991 | ssize_t result = -EINVAL, requested; |
89698b24 | 992 | size_t count; |
1da177e4 | 993 | struct file *file = iocb->ki_filp; |
1da177e4 | 994 | struct address_space *mapping = file->f_mapping; |
22cd1bf1 CH |
995 | struct inode *inode = mapping->host; |
996 | struct nfs_direct_req *dreq; | |
997 | struct nfs_lock_context *l_ctx; | |
65a4a1ca | 998 | loff_t pos, end; |
c216fd70 | 999 | |
6de1472f | 1000 | dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n", |
3309dd04 | 1001 | file, iov_iter_count(iter), (long long) iocb->ki_pos); |
027445c3 | 1002 | |
89698b24 TM |
1003 | result = generic_write_checks(iocb, iter); |
1004 | if (result <= 0) | |
1005 | return result; | |
1006 | count = result; | |
1007 | nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count); | |
3309dd04 AV |
1008 | |
1009 | pos = iocb->ki_pos; | |
09cbfeaf | 1010 | end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT; |
ce1a8e67 | 1011 | |
89698b24 | 1012 | task_io_account_write(count); |
7ec10f26 | 1013 | |
22cd1bf1 CH |
1014 | result = -ENOMEM; |
1015 | dreq = nfs_direct_req_alloc(); | |
1016 | if (!dreq) | |
a5864c99 | 1017 | goto out; |
9eafa8cc | 1018 | |
22cd1bf1 | 1019 | dreq->inode = inode; |
89698b24 | 1020 | dreq->bytes_left = dreq->max_count = count; |
5fadeb47 | 1021 | dreq->io_start = pos; |
22cd1bf1 CH |
1022 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
1023 | l_ctx = nfs_get_lock_context(dreq->ctx); | |
1024 | if (IS_ERR(l_ctx)) { | |
1025 | result = PTR_ERR(l_ctx); | |
1026 | goto out_release; | |
1027 | } | |
1028 | dreq->l_ctx = l_ctx; | |
1029 | if (!is_sync_kiocb(iocb)) | |
1030 | dreq->iocb = iocb; | |
1031 | ||
a5864c99 TM |
1032 | nfs_start_io_direct(inode); |
1033 | ||
85128b2b | 1034 | requested = nfs_direct_write_schedule_iovec(dreq, iter, pos); |
a9ab5e84 CH |
1035 | |
1036 | if (mapping->nrpages) { | |
1037 | invalidate_inode_pages2_range(mapping, | |
09cbfeaf | 1038 | pos >> PAGE_SHIFT, end); |
a9ab5e84 CH |
1039 | } |
1040 | ||
a5864c99 | 1041 | nfs_end_io_direct(inode); |
a9ab5e84 | 1042 | |
85128b2b | 1043 | if (requested > 0) { |
22cd1bf1 CH |
1044 | result = nfs_direct_wait(dreq); |
1045 | if (result > 0) { | |
85128b2b | 1046 | requested -= result; |
22cd1bf1 | 1047 | iocb->ki_pos = pos + result; |
e2592217 CH |
1048 | /* XXX: should check the generic_write_sync retval */ |
1049 | generic_write_sync(iocb, result); | |
22cd1bf1 | 1050 | } |
85128b2b AV |
1051 | iov_iter_revert(iter, requested); |
1052 | } else { | |
1053 | result = requested; | |
1763da12 | 1054 | } |
22cd1bf1 CH |
1055 | out_release: |
1056 | nfs_direct_req_release(dreq); | |
a5864c99 | 1057 | out: |
22cd1bf1 | 1058 | return result; |
1da177e4 LT |
1059 | } |
1060 | ||
88467055 CL |
1061 | /** |
1062 | * nfs_init_directcache - create a slab cache for nfs_direct_req structures | |
1063 | * | |
1064 | */ | |
f7b422b1 | 1065 | int __init nfs_init_directcache(void) |
1da177e4 LT |
1066 | { |
1067 | nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", | |
1068 | sizeof(struct nfs_direct_req), | |
fffb60f9 PJ |
1069 | 0, (SLAB_RECLAIM_ACCOUNT| |
1070 | SLAB_MEM_SPREAD), | |
20c2df83 | 1071 | NULL); |
1da177e4 LT |
1072 | if (nfs_direct_cachep == NULL) |
1073 | return -ENOMEM; | |
1074 | ||
1075 | return 0; | |
1076 | } | |
1077 | ||
88467055 | 1078 | /** |
f7b422b1 | 1079 | * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures |
88467055 CL |
1080 | * |
1081 | */ | |
266bee88 | 1082 | void nfs_destroy_directcache(void) |
1da177e4 | 1083 | { |
1a1d92c1 | 1084 | kmem_cache_destroy(nfs_direct_cachep); |
1da177e4 | 1085 | } |