Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 LT |
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 | |
88467055 | 11 | * (multiple copies of the same instance running on separate hosts) |
1da177e4 | 12 | * implement their own cache coherency protocol that subsumes file |
88467055 CL |
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 | |
1da177e4 LT |
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 | |
88467055 | 38 | * 04 May 2005 support O_DIRECT with aio --cel |
1da177e4 LT |
39 | * |
40 | */ | |
41 | ||
1da177e4 LT |
42 | #include <linux/errno.h> |
43 | #include <linux/sched.h> | |
44 | #include <linux/kernel.h> | |
1da177e4 LT |
45 | #include <linux/file.h> |
46 | #include <linux/pagemap.h> | |
47 | #include <linux/kref.h> | |
5a0e3ad6 | 48 | #include <linux/slab.h> |
7ec10f26 | 49 | #include <linux/task_io_accounting_ops.h> |
6296556f | 50 | #include <linux/module.h> |
1da177e4 LT |
51 | |
52 | #include <linux/nfs_fs.h> | |
53 | #include <linux/nfs_page.h> | |
54 | #include <linux/sunrpc/clnt.h> | |
55 | ||
7c0f6ba6 | 56 | #include <linux/uaccess.h> |
60063497 | 57 | #include <linux/atomic.h> |
1da177e4 | 58 | |
8d5658c9 | 59 | #include "internal.h" |
91d5b470 | 60 | #include "iostat.h" |
1763da12 | 61 | #include "pnfs.h" |
91d5b470 | 62 | |
1da177e4 | 63 | #define NFSDBG_FACILITY NFSDBG_VFS |
1da177e4 | 64 | |
e18b890b | 65 | static struct kmem_cache *nfs_direct_cachep; |
1da177e4 | 66 | |
1da177e4 LT |
67 | struct nfs_direct_req { |
68 | struct kref kref; /* release manager */ | |
15ce4a0c CL |
69 | |
70 | /* I/O parameters */ | |
a8881f5a | 71 | struct nfs_open_context *ctx; /* file open context info */ |
f11ac8db | 72 | struct nfs_lock_context *l_ctx; /* Lock context info */ |
99514f8f | 73 | struct kiocb * iocb; /* controlling i/o request */ |
88467055 | 74 | struct inode * inode; /* target file of i/o */ |
15ce4a0c CL |
75 | |
76 | /* completion state */ | |
607f31e8 | 77 | atomic_t io_count; /* i/os we're waiting for */ |
15ce4a0c | 78 | spinlock_t lock; /* protect completion state */ |
0a00b77b | 79 | |
d9ee6553 | 80 | loff_t io_start; /* Start offset for I/O */ |
15ce4a0c | 81 | ssize_t count, /* bytes actually processed */ |
ed3743a6 | 82 | max_count, /* max expected count */ |
35754bc0 | 83 | bytes_left, /* bytes left to be sent */ |
1da177e4 | 84 | error; /* any reported error */ |
d72b7a6b | 85 | struct completion completion; /* wait for i/o completion */ |
fad61490 TM |
86 | |
87 | /* commit state */ | |
1763da12 FI |
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; | |
fad61490 | 91 | int flags; |
ad3cba22 | 92 | /* for write */ |
fad61490 TM |
93 | #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */ |
94 | #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */ | |
ad3cba22 DK |
95 | /* for read */ |
96 | #define NFS_ODIRECT_SHOULD_DIRTY (3) /* dirty user-space page after read */ | |
fad61490 | 97 | struct nfs_writeverf verf; /* unstable write verifier */ |
1da177e4 LT |
98 | }; |
99 | ||
1763da12 FI |
100 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops; |
101 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops; | |
4d3b55d3 | 102 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq); |
1763da12 | 103 | static void nfs_direct_write_schedule_work(struct work_struct *work); |
607f31e8 TM |
104 | |
105 | static inline void get_dreq(struct nfs_direct_req *dreq) | |
106 | { | |
107 | atomic_inc(&dreq->io_count); | |
108 | } | |
109 | ||
110 | static inline int put_dreq(struct nfs_direct_req *dreq) | |
111 | { | |
112 | return atomic_dec_and_test(&dreq->io_count); | |
113 | } | |
114 | ||
0a00b77b | 115 | static void |
031d73ed TM |
116 | nfs_direct_handle_truncated(struct nfs_direct_req *dreq, |
117 | const struct nfs_pgio_header *hdr, | |
118 | ssize_t dreq_len) | |
0a00b77b | 119 | { |
031d73ed TM |
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 | } | |
031d73ed | 133 | } |
0a00b77b | 134 | |
031d73ed TM |
135 | static void |
136 | nfs_direct_count_bytes(struct nfs_direct_req *dreq, | |
137 | const struct nfs_pgio_header *hdr) | |
138 | { | |
031d73ed TM |
139 | loff_t hdr_end = hdr->io_start + hdr->good_bytes; |
140 | ssize_t dreq_len = 0; | |
ed3743a6 | 141 | |
031d73ed TM |
142 | if (hdr_end > dreq->io_start) |
143 | dreq_len = hdr_end - dreq->io_start; | |
0a00b77b | 144 | |
031d73ed | 145 | nfs_direct_handle_truncated(dreq, hdr, dreq_len); |
0a00b77b | 146 | |
031d73ed TM |
147 | if (dreq_len > dreq->max_count) |
148 | dreq_len = dreq->max_count; | |
149 | ||
031d73ed TM |
150 | if (dreq->count < dreq_len) |
151 | dreq->count = dreq_len; | |
0a00b77b WAA |
152 | } |
153 | ||
5002c586 WAA |
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 | |
6cccbb6f | 158 | * @commit_idx - commit bucket index for the DS |
5002c586 WAA |
159 | * |
160 | * returns the correct verifier to use given the role of the server | |
161 | */ | |
162 | static struct nfs_writeverf * | |
163 | nfs_direct_select_verf(struct nfs_direct_req *dreq, | |
164 | struct nfs_client *ds_clp, | |
6cccbb6f | 165 | int commit_idx) |
5002c586 WAA |
166 | { |
167 | struct nfs_writeverf *verfp = &dreq->verf; | |
168 | ||
169 | #ifdef CONFIG_NFS_V4_1 | |
834e465b KM |
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) { | |
6cccbb6f WAA |
175 | if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets) |
176 | verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf; | |
5002c586 WAA |
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 | */ | |
192 | static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq, | |
193 | struct nfs_pgio_header *hdr) | |
194 | { | |
195 | struct nfs_writeverf *verfp; | |
196 | ||
6cccbb6f | 197 | verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx); |
5002c586 WAA |
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 | ||
8fc3c386 TM |
203 | static 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 | ||
5002c586 WAA |
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 | */ | |
218 | static 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 | ||
6cccbb6f | 223 | verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx); |
5002c586 WAA |
224 | if (verfp->committed < 0) { |
225 | nfs_direct_set_hdr_verf(dreq, hdr); | |
226 | return 0; | |
227 | } | |
8fc3c386 | 228 | return nfs_direct_cmp_verf(verfp, &hdr->verf); |
5002c586 WAA |
229 | } |
230 | ||
5002c586 WAA |
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 | */ | |
239 | static 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); | |
80c76fe3 WAA |
246 | |
247 | /* verifier not set so always fail */ | |
248 | if (verfp->committed < 0) | |
249 | return 1; | |
250 | ||
8fc3c386 | 251 | return nfs_direct_cmp_verf(verfp, &data->verf); |
5002c586 | 252 | } |
5002c586 | 253 | |
1da177e4 | 254 | /** |
b8a32e2b | 255 | * nfs_direct_IO - NFS address space operation for direct I/O |
b8a32e2b | 256 | * @iocb: target I/O control block |
90090ae6 | 257 | * @iter: I/O buffer |
b8a32e2b CL |
258 | * |
259 | * The presence of this routine in the address space ops vector means | |
a564b8f0 MG |
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. | |
1da177e4 | 263 | */ |
c8b8e32d | 264 | ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter) |
b8a32e2b | 265 | { |
ee8a1a8b PT |
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 | ||
66ee59af | 272 | VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE); |
a564b8f0 | 273 | |
6f673763 | 274 | if (iov_iter_rw(iter) == READ) |
c8b8e32d | 275 | return nfs_file_direct_read(iocb, iter); |
65a4a1ca | 276 | return nfs_file_direct_write(iocb, iter); |
b8a32e2b CL |
277 | } |
278 | ||
749e146e | 279 | static void nfs_direct_release_pages(struct page **pages, unsigned int npages) |
9c93ab7d | 280 | { |
749e146e | 281 | unsigned int i; |
607f31e8 | 282 | for (i = 0; i < npages; i++) |
09cbfeaf | 283 | put_page(pages[i]); |
6b45d858 TM |
284 | } |
285 | ||
1763da12 FI |
286 | void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo, |
287 | struct nfs_direct_req *dreq) | |
288 | { | |
fe238e60 | 289 | cinfo->inode = dreq->inode; |
1763da12 FI |
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 | ||
93619e59 | 296 | static inline struct nfs_direct_req *nfs_direct_req_alloc(void) |
1da177e4 | 297 | { |
93619e59 CL |
298 | struct nfs_direct_req *dreq; |
299 | ||
292f3eee | 300 | dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL); |
93619e59 CL |
301 | if (!dreq) |
302 | return NULL; | |
303 | ||
304 | kref_init(&dreq->kref); | |
607f31e8 | 305 | kref_get(&dreq->kref); |
d72b7a6b | 306 | init_completion(&dreq->completion); |
1763da12 | 307 | INIT_LIST_HEAD(&dreq->mds_cinfo.list); |
5002c586 | 308 | dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */ |
1763da12 | 309 | INIT_WORK(&dreq->work, nfs_direct_write_schedule_work); |
15ce4a0c | 310 | spin_lock_init(&dreq->lock); |
93619e59 CL |
311 | |
312 | return dreq; | |
1da177e4 LT |
313 | } |
314 | ||
b4946ffb | 315 | static void nfs_direct_req_free(struct kref *kref) |
1da177e4 LT |
316 | { |
317 | struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); | |
a8881f5a | 318 | |
8c393f9a | 319 | nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo); |
f11ac8db TM |
320 | if (dreq->l_ctx != NULL) |
321 | nfs_put_lock_context(dreq->l_ctx); | |
a8881f5a TM |
322 | if (dreq->ctx != NULL) |
323 | put_nfs_open_context(dreq->ctx); | |
1da177e4 LT |
324 | kmem_cache_free(nfs_direct_cachep, dreq); |
325 | } | |
326 | ||
b4946ffb TM |
327 | static void nfs_direct_req_release(struct nfs_direct_req *dreq) |
328 | { | |
329 | kref_put(&dreq->kref, nfs_direct_req_free); | |
330 | } | |
331 | ||
6296556f PT |
332 | ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq) |
333 | { | |
334 | return dreq->bytes_left; | |
335 | } | |
336 | EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left); | |
337 | ||
bc0fb201 CL |
338 | /* |
339 | * Collects and returns the final error value/byte-count. | |
340 | */ | |
341 | static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) | |
342 | { | |
15ce4a0c | 343 | ssize_t result = -EIOCBQUEUED; |
bc0fb201 CL |
344 | |
345 | /* Async requests don't wait here */ | |
346 | if (dreq->iocb) | |
347 | goto out; | |
348 | ||
150030b7 | 349 | result = wait_for_completion_killable(&dreq->completion); |
bc0fb201 | 350 | |
d2a7de0b TM |
351 | if (!result) { |
352 | result = dreq->count; | |
353 | WARN_ON_ONCE(dreq->count < 0); | |
354 | } | |
bc0fb201 | 355 | if (!result) |
15ce4a0c | 356 | result = dreq->error; |
bc0fb201 CL |
357 | |
358 | out: | |
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 | */ |
f7b5c340 | 366 | static void nfs_direct_complete(struct nfs_direct_req *dreq) |
63ab46ab | 367 | { |
9811cd57 CH |
368 | struct inode *inode = dreq->inode; |
369 | ||
fe0f07d0 | 370 | inode_dio_end(inode); |
2a009ec9 CH |
371 | |
372 | if (dreq->iocb) { | |
15ce4a0c | 373 | long res = (long) dreq->error; |
d2a7de0b | 374 | if (dreq->count != 0) { |
15ce4a0c | 375 | res = (long) dreq->count; |
d2a7de0b TM |
376 | WARN_ON_ONCE(dreq->count < 0); |
377 | } | |
04b2fa9f | 378 | dreq->iocb->ki_complete(dreq->iocb, res, 0); |
d72b7a6b | 379 | } |
2a009ec9 | 380 | |
024de8f1 | 381 | complete(&dreq->completion); |
63ab46ab | 382 | |
b4946ffb | 383 | nfs_direct_req_release(dreq); |
63ab46ab CL |
384 | } |
385 | ||
584aa810 | 386 | static void nfs_direct_read_completion(struct nfs_pgio_header *hdr) |
fdd1e74c | 387 | { |
584aa810 FI |
388 | unsigned long bytes = 0; |
389 | struct nfs_direct_req *dreq = hdr->dreq; | |
fdd1e74c | 390 | |
15ce4a0c | 391 | spin_lock(&dreq->lock); |
eb2c50da TM |
392 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) { |
393 | spin_unlock(&dreq->lock); | |
394 | goto out_put; | |
395 | } | |
396 | ||
031d73ed | 397 | nfs_direct_count_bytes(dreq, hdr); |
584aa810 FI |
398 | spin_unlock(&dreq->lock); |
399 | ||
4bd8b010 TM |
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 | ||
ad3cba22 DK |
404 | if (!PageCompound(page) && bytes < hdr->good_bytes && |
405 | (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY)) | |
be7e9858 | 406 | set_page_dirty(page); |
4bd8b010 TM |
407 | bytes += req->wb_bytes; |
408 | nfs_list_remove_request(req); | |
beeb5338 | 409 | nfs_release_request(req); |
d4a8f367 | 410 | } |
584aa810 | 411 | out_put: |
607f31e8 | 412 | if (put_dreq(dreq)) |
f7b5c340 | 413 | nfs_direct_complete(dreq); |
584aa810 | 414 | hdr->release(hdr); |
1da177e4 LT |
415 | } |
416 | ||
df3accb8 | 417 | static void nfs_read_sync_pgio_error(struct list_head *head, int error) |
cd841605 | 418 | { |
584aa810 | 419 | struct nfs_page *req; |
cd841605 | 420 | |
584aa810 FI |
421 | while (!list_empty(head)) { |
422 | req = nfs_list_entry(head->next); | |
423 | nfs_list_remove_request(req); | |
424 | nfs_release_request(req); | |
425 | } | |
cd841605 FI |
426 | } |
427 | ||
584aa810 FI |
428 | static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr) |
429 | { | |
430 | get_dreq(hdr->dreq); | |
431 | } | |
432 | ||
433 | static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = { | |
3e9e0ca3 | 434 | .error_cleanup = nfs_read_sync_pgio_error, |
584aa810 FI |
435 | .init_hdr = nfs_direct_pgio_init, |
436 | .completion = nfs_direct_read_completion, | |
437 | }; | |
438 | ||
d4cc948b | 439 | /* |
607f31e8 TM |
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. | |
1da177e4 | 445 | */ |
1da177e4 | 446 | |
91f79c43 AV |
447 | static 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); | |
e9f7bee1 | 456 | |
16b90578 | 457 | nfs_pageio_init_read(&desc, dreq->inode, false, |
91f79c43 AV |
458 | &nfs_direct_read_completion_ops); |
459 | get_dreq(dreq); | |
460 | desc.pg_dreq = dreq; | |
fe0f07d0 | 461 | inode_dio_begin(inode); |
a564b8f0 | 462 | |
91f79c43 AV |
463 | while (iov_iter_count(iter)) { |
464 | struct page **pagevec; | |
465 | size_t bytes; | |
466 | size_t pgbase; | |
467 | unsigned npages, i; | |
607f31e8 | 468 | |
91f79c43 AV |
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; | |
584aa810 FI |
477 | for (i = 0; i < npages; i++) { |
478 | struct nfs_page *req; | |
bf5fc402 | 479 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); |
584aa810 | 480 | /* XXX do we need to do the eof zeroing found in async_filler? */ |
28b1d3f5 | 481 | req = nfs_create_request(dreq->ctx, pagevec[i], |
584aa810 FI |
482 | pgbase, req_len); |
483 | if (IS_ERR(req)) { | |
584aa810 FI |
484 | result = PTR_ERR(req); |
485 | break; | |
486 | } | |
487 | req->wb_index = pos >> PAGE_SHIFT; | |
488 | req->wb_offset = pos & ~PAGE_MASK; | |
91f79c43 AV |
489 | if (!nfs_pageio_add_request(&desc, req)) { |
490 | result = desc.pg_error; | |
584aa810 | 491 | nfs_release_request(req); |
584aa810 FI |
492 | break; |
493 | } | |
494 | pgbase = 0; | |
495 | bytes -= req_len; | |
91f79c43 | 496 | requested_bytes += req_len; |
584aa810 | 497 | pos += req_len; |
35754bc0 | 498 | dreq->bytes_left -= req_len; |
584aa810 | 499 | } |
6d74743b | 500 | nfs_direct_release_pages(pagevec, npages); |
91f79c43 | 501 | kvfree(pagevec); |
19f73787 CL |
502 | if (result < 0) |
503 | break; | |
19f73787 CL |
504 | } |
505 | ||
584aa810 FI |
506 | nfs_pageio_complete(&desc); |
507 | ||
839f7ad6 CL |
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) { | |
fe0f07d0 | 513 | inode_dio_end(inode); |
839f7ad6 CL |
514 | nfs_direct_req_release(dreq); |
515 | return result < 0 ? result : -EIO; | |
516 | } | |
517 | ||
19f73787 | 518 | if (put_dreq(dreq)) |
f7b5c340 | 519 | nfs_direct_complete(dreq); |
85128b2b | 520 | return requested_bytes; |
19f73787 CL |
521 | } |
522 | ||
14a3ec79 CH |
523 | /** |
524 | * nfs_file_direct_read - file direct read operation for NFS files | |
525 | * @iocb: target I/O control block | |
619d30b4 | 526 | * @iter: vector of user buffers into which to read data |
14a3ec79 CH |
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 | */ | |
c8b8e32d | 542 | ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter) |
1da177e4 | 543 | { |
14a3ec79 CH |
544 | struct file *file = iocb->ki_filp; |
545 | struct address_space *mapping = file->f_mapping; | |
546 | struct inode *inode = mapping->host; | |
1da177e4 | 547 | struct nfs_direct_req *dreq; |
b3c54de6 | 548 | struct nfs_lock_context *l_ctx; |
85128b2b | 549 | ssize_t result = -EINVAL, requested; |
a6cbcd4a | 550 | size_t count = iov_iter_count(iter); |
14a3ec79 CH |
551 | nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); |
552 | ||
553 | dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n", | |
c8b8e32d | 554 | file, count, (long long) iocb->ki_pos); |
14a3ec79 CH |
555 | |
556 | result = 0; | |
557 | if (!count) | |
558 | goto out; | |
559 | ||
14a3ec79 CH |
560 | task_io_account_read(count); |
561 | ||
562 | result = -ENOMEM; | |
607f31e8 | 563 | dreq = nfs_direct_req_alloc(); |
f11ac8db | 564 | if (dreq == NULL) |
a5864c99 | 565 | goto out; |
1da177e4 | 566 | |
91d5b470 | 567 | dreq->inode = inode; |
ed3743a6 | 568 | dreq->bytes_left = dreq->max_count = count; |
c8b8e32d | 569 | dreq->io_start = iocb->ki_pos; |
cd3758e3 | 570 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
b3c54de6 TM |
571 | l_ctx = nfs_get_lock_context(dreq->ctx); |
572 | if (IS_ERR(l_ctx)) { | |
573 | result = PTR_ERR(l_ctx); | |
f11ac8db | 574 | goto out_release; |
b3c54de6 TM |
575 | } |
576 | dreq->l_ctx = l_ctx; | |
487b8372 CL |
577 | if (!is_sync_kiocb(iocb)) |
578 | dreq->iocb = iocb; | |
1da177e4 | 579 | |
ad3cba22 DK |
580 | if (iter_is_iovec(iter)) |
581 | dreq->flags = NFS_ODIRECT_SHOULD_DIRTY; | |
582 | ||
a5864c99 TM |
583 | nfs_start_io_direct(inode); |
584 | ||
619d30b4 | 585 | NFS_I(inode)->read_io += count; |
85128b2b | 586 | requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos); |
d0b9875d | 587 | |
a5864c99 | 588 | nfs_end_io_direct(inode); |
d0b9875d | 589 | |
85128b2b | 590 | if (requested > 0) { |
607f31e8 | 591 | result = nfs_direct_wait(dreq); |
85128b2b AV |
592 | if (result > 0) { |
593 | requested -= result; | |
c8b8e32d | 594 | iocb->ki_pos += result; |
85128b2b AV |
595 | } |
596 | iov_iter_revert(iter, requested); | |
597 | } else { | |
598 | result = requested; | |
14a3ec79 | 599 | } |
d0b9875d | 600 | |
f11ac8db | 601 | out_release: |
b4946ffb | 602 | nfs_direct_req_release(dreq); |
f11ac8db | 603 | out: |
1da177e4 LT |
604 | return result; |
605 | } | |
606 | ||
085d1e33 TH |
607 | static void |
608 | nfs_direct_write_scan_commit_list(struct inode *inode, | |
609 | struct list_head *list, | |
610 | struct nfs_commit_info *cinfo) | |
611 | { | |
e824f99a | 612 | mutex_lock(&NFS_I(cinfo->inode)->commit_mutex); |
085d1e33 TH |
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); | |
e824f99a | 618 | mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex); |
085d1e33 TH |
619 | } |
620 | ||
fad61490 TM |
621 | static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) |
622 | { | |
1763da12 FI |
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); | |
085d1e33 | 630 | nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo); |
1da177e4 | 631 | |
fad61490 | 632 | dreq->count = 0; |
031d73ed TM |
633 | dreq->max_count = 0; |
634 | list_for_each_entry(req, &reqs, wb_list) | |
635 | dreq->max_count += req->wb_bytes; | |
a5314a74 TM |
636 | dreq->verf.committed = NFS_INVALID_STABLE_HOW; |
637 | nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo); | |
607f31e8 TM |
638 | get_dreq(dreq); |
639 | ||
a20c93e3 | 640 | nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false, |
1763da12 FI |
641 | &nfs_direct_write_completion_ops); |
642 | desc.pg_dreq = dreq; | |
fedb595c | 643 | |
1763da12 | 644 | list_for_each_entry_safe(req, tmp, &reqs, wb_list) { |
33344e0f TM |
645 | /* Bump the transmission count */ |
646 | req->wb_nio++; | |
1763da12 | 647 | if (!nfs_pageio_add_request(&desc, req)) { |
078b5fd9 | 648 | nfs_list_move_request(req, &failed); |
fe238e60 | 649 | spin_lock(&cinfo.inode->i_lock); |
1763da12 | 650 | dreq->flags = 0; |
d600ad1f PT |
651 | if (desc.pg_error < 0) |
652 | dreq->error = desc.pg_error; | |
653 | else | |
654 | dreq->error = -EIO; | |
fe238e60 | 655 | spin_unlock(&cinfo.inode->i_lock); |
1763da12 | 656 | } |
5a695da2 | 657 | nfs_release_request(req); |
1763da12 FI |
658 | } |
659 | nfs_pageio_complete(&desc); | |
fad61490 | 660 | |
4035c248 TM |
661 | while (!list_empty(&failed)) { |
662 | req = nfs_list_entry(failed.next); | |
663 | nfs_list_remove_request(req); | |
1d1afcbc | 664 | nfs_unlock_and_release_request(req); |
4035c248 | 665 | } |
fad61490 | 666 | |
1763da12 | 667 | if (put_dreq(dreq)) |
4d3b55d3 | 668 | nfs_direct_write_complete(dreq); |
c9d8f89d TM |
669 | } |
670 | ||
1763da12 | 671 | static void nfs_direct_commit_complete(struct nfs_commit_data *data) |
c9d8f89d | 672 | { |
0b7c0153 | 673 | struct nfs_direct_req *dreq = data->dreq; |
1763da12 FI |
674 | struct nfs_commit_info cinfo; |
675 | struct nfs_page *req; | |
c9d8f89d TM |
676 | int status = data->task.tk_status; |
677 | ||
1763da12 | 678 | nfs_init_cinfo_from_dreq(&cinfo, dreq); |
fe4f844d | 679 | if (status < 0 || nfs_direct_cmp_commit_data_verf(dreq, data)) |
fad61490 | 680 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
1da177e4 | 681 | |
1763da12 FI |
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) { | |
33344e0f TM |
686 | /* |
687 | * Despite the reboot, the write was successful, | |
688 | * so reset wb_nio. | |
689 | */ | |
690 | req->wb_nio = 0; | |
1763da12 | 691 | /* Note the rewrite will go through mds */ |
b57ff130 | 692 | nfs_mark_request_commit(req, NULL, &cinfo, 0); |
906369e4 FI |
693 | } else |
694 | nfs_release_request(req); | |
1d1afcbc | 695 | nfs_unlock_and_release_request(req); |
1763da12 FI |
696 | } |
697 | ||
698 | if (atomic_dec_and_test(&cinfo.mds->rpcs_out)) | |
4d3b55d3 | 699 | nfs_direct_write_complete(dreq); |
1da177e4 LT |
700 | } |
701 | ||
b20135d0 TM |
702 | static void nfs_direct_resched_write(struct nfs_commit_info *cinfo, |
703 | struct nfs_page *req) | |
1763da12 | 704 | { |
b20135d0 TM |
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); | |
1763da12 FI |
711 | } |
712 | ||
713 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = { | |
714 | .completion = nfs_direct_commit_complete, | |
b20135d0 | 715 | .resched_write = nfs_direct_resched_write, |
fad61490 TM |
716 | }; |
717 | ||
718 | static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) | |
1da177e4 | 719 | { |
1763da12 FI |
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); | |
fad61490 | 729 | } |
1da177e4 | 730 | |
1763da12 | 731 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
fad61490 | 732 | { |
1763da12 | 733 | struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work); |
fad61490 | 734 | int flags = dreq->flags; |
1da177e4 | 735 | |
fad61490 TM |
736 | dreq->flags = 0; |
737 | switch (flags) { | |
738 | case NFS_ODIRECT_DO_COMMIT: | |
739 | nfs_direct_commit_schedule(dreq); | |
1da177e4 | 740 | break; |
fad61490 TM |
741 | case NFS_ODIRECT_RESCHED_WRITES: |
742 | nfs_direct_write_reschedule(dreq); | |
743 | break; | |
744 | default: | |
f7b5c340 TM |
745 | nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping); |
746 | nfs_direct_complete(dreq); | |
fad61490 TM |
747 | } |
748 | } | |
1da177e4 | 749 | |
4d3b55d3 | 750 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq) |
fad61490 | 751 | { |
46483c2e | 752 | queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */ |
fad61490 | 753 | } |
1763da12 | 754 | |
1763da12 FI |
755 | static 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; | |
c65e6254 | 759 | bool request_commit = false; |
1763da12 FI |
760 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); |
761 | ||
1763da12 FI |
762 | nfs_init_cinfo_from_dreq(&cinfo, dreq); |
763 | ||
764 | spin_lock(&dreq->lock); | |
eb2c50da TM |
765 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) { |
766 | spin_unlock(&dreq->lock); | |
767 | goto out_put; | |
768 | } | |
769 | ||
031d73ed | 770 | nfs_direct_count_bytes(dreq, hdr); |
eb2c50da | 771 | if (hdr->good_bytes != 0) { |
c65e6254 | 772 | if (nfs_write_need_commit(hdr)) { |
1763da12 | 773 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) |
c65e6254 | 774 | request_commit = true; |
1763da12 | 775 | else if (dreq->flags == 0) { |
5002c586 | 776 | nfs_direct_set_hdr_verf(dreq, hdr); |
c65e6254 | 777 | request_commit = true; |
1763da12 FI |
778 | dreq->flags = NFS_ODIRECT_DO_COMMIT; |
779 | } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) { | |
c65e6254 WAA |
780 | request_commit = true; |
781 | if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr)) | |
5002c586 WAA |
782 | dreq->flags = |
783 | NFS_ODIRECT_RESCHED_WRITES; | |
1763da12 FI |
784 | } |
785 | } | |
786 | } | |
787 | spin_unlock(&dreq->lock); | |
788 | ||
789 | while (!list_empty(&hdr->pages)) { | |
2bfc6e56 | 790 | |
1763da12 FI |
791 | req = nfs_list_entry(hdr->pages.next); |
792 | nfs_list_remove_request(req); | |
c65e6254 | 793 | if (request_commit) { |
04277086 | 794 | kref_get(&req->wb_kref); |
b57ff130 WAA |
795 | nfs_mark_request_commit(req, hdr->lseg, &cinfo, |
796 | hdr->ds_commit_idx); | |
1763da12 | 797 | } |
1d1afcbc | 798 | nfs_unlock_and_release_request(req); |
1763da12 FI |
799 | } |
800 | ||
801 | out_put: | |
802 | if (put_dreq(dreq)) | |
4d3b55d3 | 803 | nfs_direct_write_complete(dreq); |
1763da12 FI |
804 | hdr->release(hdr); |
805 | } | |
806 | ||
df3accb8 | 807 | static void nfs_write_sync_pgio_error(struct list_head *head, int error) |
3e9e0ca3 TM |
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); | |
1d1afcbc | 814 | nfs_unlock_and_release_request(req); |
3e9e0ca3 TM |
815 | } |
816 | } | |
817 | ||
dc602dd7 TM |
818 | static 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 | ||
1763da12 | 832 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = { |
3e9e0ca3 | 833 | .error_cleanup = nfs_write_sync_pgio_error, |
1763da12 FI |
834 | .init_hdr = nfs_direct_pgio_init, |
835 | .completion = nfs_direct_write_completion, | |
dc602dd7 | 836 | .reschedule_io = nfs_direct_write_reschedule_io, |
1763da12 FI |
837 | }; |
838 | ||
91f79c43 AV |
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 | */ | |
19f73787 | 851 | static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, |
619d30b4 | 852 | struct iov_iter *iter, |
91f79c43 | 853 | loff_t pos) |
19f73787 | 854 | { |
1763da12 | 855 | struct nfs_pageio_descriptor desc; |
1d59d61f | 856 | struct inode *inode = dreq->inode; |
19f73787 CL |
857 | ssize_t result = 0; |
858 | size_t requested_bytes = 0; | |
91f79c43 | 859 | size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE); |
19f73787 | 860 | |
a20c93e3 | 861 | nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false, |
1763da12 FI |
862 | &nfs_direct_write_completion_ops); |
863 | desc.pg_dreq = dreq; | |
19f73787 | 864 | get_dreq(dreq); |
fe0f07d0 | 865 | inode_dio_begin(inode); |
19f73787 | 866 | |
91f79c43 AV |
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); | |
19f73787 CL |
876 | if (result < 0) |
877 | break; | |
91f79c43 AV |
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 | ||
28b1d3f5 | 886 | req = nfs_create_request(dreq->ctx, pagevec[i], |
91f79c43 AV |
887 | pgbase, req_len); |
888 | if (IS_ERR(req)) { | |
889 | result = PTR_ERR(req); | |
890 | break; | |
891 | } | |
0a00b77b | 892 | |
d600ad1f PT |
893 | if (desc.pg_error < 0) { |
894 | nfs_free_request(req); | |
895 | result = desc.pg_error; | |
896 | break; | |
897 | } | |
0a00b77b | 898 | |
91f79c43 AV |
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) | |
19f73787 | 916 | break; |
19f73787 | 917 | } |
1763da12 | 918 | nfs_pageio_complete(&desc); |
19f73787 | 919 | |
839f7ad6 CL |
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) { | |
fe0f07d0 | 925 | inode_dio_end(inode); |
839f7ad6 CL |
926 | nfs_direct_req_release(dreq); |
927 | return result < 0 ? result : -EIO; | |
928 | } | |
929 | ||
19f73787 | 930 | if (put_dreq(dreq)) |
4d3b55d3 | 931 | nfs_direct_write_complete(dreq); |
85128b2b | 932 | return requested_bytes; |
19f73787 CL |
933 | } |
934 | ||
1da177e4 LT |
935 | /** |
936 | * nfs_file_direct_write - file direct write operation for NFS files | |
937 | * @iocb: target I/O control block | |
619d30b4 | 938 | * @iter: vector of user buffers from which to write data |
1da177e4 LT |
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 | * | |
1da177e4 LT |
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 | */ | |
65a4a1ca | 955 | ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter) |
1da177e4 | 956 | { |
85128b2b | 957 | ssize_t result = -EINVAL, requested; |
89698b24 | 958 | size_t count; |
1da177e4 | 959 | struct file *file = iocb->ki_filp; |
1da177e4 | 960 | struct address_space *mapping = file->f_mapping; |
22cd1bf1 CH |
961 | struct inode *inode = mapping->host; |
962 | struct nfs_direct_req *dreq; | |
963 | struct nfs_lock_context *l_ctx; | |
65a4a1ca | 964 | loff_t pos, end; |
c216fd70 | 965 | |
6de1472f | 966 | dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n", |
3309dd04 | 967 | file, iov_iter_count(iter), (long long) iocb->ki_pos); |
027445c3 | 968 | |
89698b24 TM |
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); | |
3309dd04 AV |
974 | |
975 | pos = iocb->ki_pos; | |
09cbfeaf | 976 | end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT; |
ce1a8e67 | 977 | |
89698b24 | 978 | task_io_account_write(count); |
7ec10f26 | 979 | |
22cd1bf1 CH |
980 | result = -ENOMEM; |
981 | dreq = nfs_direct_req_alloc(); | |
982 | if (!dreq) | |
a5864c99 | 983 | goto out; |
9eafa8cc | 984 | |
22cd1bf1 | 985 | dreq->inode = inode; |
89698b24 | 986 | dreq->bytes_left = dreq->max_count = count; |
5fadeb47 | 987 | dreq->io_start = pos; |
22cd1bf1 CH |
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 | ||
a5864c99 TM |
998 | nfs_start_io_direct(inode); |
999 | ||
85128b2b | 1000 | requested = nfs_direct_write_schedule_iovec(dreq, iter, pos); |
a9ab5e84 CH |
1001 | |
1002 | if (mapping->nrpages) { | |
1003 | invalidate_inode_pages2_range(mapping, | |
09cbfeaf | 1004 | pos >> PAGE_SHIFT, end); |
a9ab5e84 CH |
1005 | } |
1006 | ||
a5864c99 | 1007 | nfs_end_io_direct(inode); |
a9ab5e84 | 1008 | |
85128b2b | 1009 | if (requested > 0) { |
22cd1bf1 CH |
1010 | result = nfs_direct_wait(dreq); |
1011 | if (result > 0) { | |
85128b2b | 1012 | requested -= result; |
22cd1bf1 | 1013 | iocb->ki_pos = pos + result; |
e2592217 CH |
1014 | /* XXX: should check the generic_write_sync retval */ |
1015 | generic_write_sync(iocb, result); | |
22cd1bf1 | 1016 | } |
85128b2b AV |
1017 | iov_iter_revert(iter, requested); |
1018 | } else { | |
1019 | result = requested; | |
1763da12 | 1020 | } |
22cd1bf1 CH |
1021 | out_release: |
1022 | nfs_direct_req_release(dreq); | |
a5864c99 | 1023 | out: |
22cd1bf1 | 1024 | return result; |
1da177e4 LT |
1025 | } |
1026 | ||
88467055 CL |
1027 | /** |
1028 | * nfs_init_directcache - create a slab cache for nfs_direct_req structures | |
1029 | * | |
1030 | */ | |
f7b422b1 | 1031 | int __init nfs_init_directcache(void) |
1da177e4 LT |
1032 | { |
1033 | nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", | |
1034 | sizeof(struct nfs_direct_req), | |
fffb60f9 PJ |
1035 | 0, (SLAB_RECLAIM_ACCOUNT| |
1036 | SLAB_MEM_SPREAD), | |
20c2df83 | 1037 | NULL); |
1da177e4 LT |
1038 | if (nfs_direct_cachep == NULL) |
1039 | return -ENOMEM; | |
1040 | ||
1041 | return 0; | |
1042 | } | |
1043 | ||
88467055 | 1044 | /** |
f7b422b1 | 1045 | * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures |
88467055 CL |
1046 | * |
1047 | */ | |
266bee88 | 1048 | void nfs_destroy_directcache(void) |
1da177e4 | 1049 | { |
1a1d92c1 | 1050 | kmem_cache_destroy(nfs_direct_cachep); |
1da177e4 | 1051 | } |