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 | ||
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 | 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 | ||
15ce4a0c | 89 | ssize_t count, /* bytes actually processed */ |
ed3743a6 | 90 | max_count, /* max expected count */ |
35754bc0 | 91 | bytes_left, /* bytes left to be sent */ |
5fadeb47 | 92 | io_start, /* start of IO */ |
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 TM |
100 | int flags; |
101 | #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */ | |
102 | #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */ | |
103 | struct nfs_writeverf verf; /* unstable write verifier */ | |
1da177e4 LT |
104 | }; |
105 | ||
1763da12 FI |
106 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops; |
107 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops; | |
fad61490 | 108 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode); |
1763da12 | 109 | static void nfs_direct_write_schedule_work(struct work_struct *work); |
607f31e8 TM |
110 | |
111 | static inline void get_dreq(struct nfs_direct_req *dreq) | |
112 | { | |
113 | atomic_inc(&dreq->io_count); | |
114 | } | |
115 | ||
116 | static inline int put_dreq(struct nfs_direct_req *dreq) | |
117 | { | |
118 | return atomic_dec_and_test(&dreq->io_count); | |
119 | } | |
120 | ||
0a00b77b WAA |
121 | static void |
122 | nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr) | |
123 | { | |
124 | int i; | |
125 | ssize_t count; | |
126 | ||
ed3743a6 WAA |
127 | WARN_ON_ONCE(dreq->count >= dreq->max_count); |
128 | ||
1ccbad9f PT |
129 | if (dreq->mirror_count == 1) { |
130 | dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes; | |
131 | dreq->count += hdr->good_bytes; | |
132 | } else { | |
133 | /* mirrored writes */ | |
134 | count = dreq->mirrors[hdr->pgio_mirror_idx].count; | |
135 | if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) { | |
136 | count = hdr->io_start + hdr->good_bytes - dreq->io_start; | |
137 | dreq->mirrors[hdr->pgio_mirror_idx].count = count; | |
138 | } | |
139 | /* update the dreq->count by finding the minimum agreed count from all | |
140 | * mirrors */ | |
141 | count = dreq->mirrors[0].count; | |
0a00b77b | 142 | |
1ccbad9f PT |
143 | for (i = 1; i < dreq->mirror_count; i++) |
144 | count = min(count, dreq->mirrors[i].count); | |
0a00b77b | 145 | |
1ccbad9f | 146 | dreq->count = count; |
5fadeb47 | 147 | } |
0a00b77b WAA |
148 | } |
149 | ||
5002c586 WAA |
150 | /* |
151 | * nfs_direct_select_verf - select the right verifier | |
152 | * @dreq - direct request possibly spanning multiple servers | |
153 | * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs | |
6cccbb6f | 154 | * @commit_idx - commit bucket index for the DS |
5002c586 WAA |
155 | * |
156 | * returns the correct verifier to use given the role of the server | |
157 | */ | |
158 | static struct nfs_writeverf * | |
159 | nfs_direct_select_verf(struct nfs_direct_req *dreq, | |
160 | struct nfs_client *ds_clp, | |
6cccbb6f | 161 | int commit_idx) |
5002c586 WAA |
162 | { |
163 | struct nfs_writeverf *verfp = &dreq->verf; | |
164 | ||
165 | #ifdef CONFIG_NFS_V4_1 | |
834e465b KM |
166 | /* |
167 | * pNFS is in use, use the DS verf except commit_through_mds is set | |
168 | * for layout segment where nbuckets is zero. | |
169 | */ | |
170 | if (ds_clp && dreq->ds_cinfo.nbuckets > 0) { | |
6cccbb6f WAA |
171 | if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets) |
172 | verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf; | |
5002c586 WAA |
173 | else |
174 | WARN_ON_ONCE(1); | |
175 | } | |
176 | #endif | |
177 | return verfp; | |
178 | } | |
179 | ||
180 | ||
181 | /* | |
182 | * nfs_direct_set_hdr_verf - set the write/commit verifier | |
183 | * @dreq - direct request possibly spanning multiple servers | |
184 | * @hdr - pageio header to validate against previously seen verfs | |
185 | * | |
186 | * Set the server's (MDS or DS) "seen" verifier | |
187 | */ | |
188 | static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq, | |
189 | struct nfs_pgio_header *hdr) | |
190 | { | |
191 | struct nfs_writeverf *verfp; | |
192 | ||
6cccbb6f | 193 | verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx); |
5002c586 WAA |
194 | WARN_ON_ONCE(verfp->committed >= 0); |
195 | memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf)); | |
196 | WARN_ON_ONCE(verfp->committed < 0); | |
197 | } | |
198 | ||
199 | /* | |
200 | * nfs_direct_cmp_hdr_verf - compare verifier for pgio header | |
201 | * @dreq - direct request possibly spanning multiple servers | |
202 | * @hdr - pageio header to validate against previously seen verf | |
203 | * | |
204 | * set the server's "seen" verf if not initialized. | |
205 | * returns result of comparison between @hdr->verf and the "seen" | |
206 | * verf of the server used by @hdr (DS or MDS) | |
207 | */ | |
208 | static int nfs_direct_set_or_cmp_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 | if (verfp->committed < 0) { |
215 | nfs_direct_set_hdr_verf(dreq, hdr); | |
216 | return 0; | |
217 | } | |
218 | return memcmp(verfp, &hdr->verf, sizeof(struct nfs_writeverf)); | |
219 | } | |
220 | ||
5002c586 WAA |
221 | /* |
222 | * nfs_direct_cmp_commit_data_verf - compare verifier for commit data | |
223 | * @dreq - direct request possibly spanning multiple servers | |
224 | * @data - commit data to validate against previously seen verf | |
225 | * | |
226 | * returns result of comparison between @data->verf and the verf of | |
227 | * the server used by @data (DS or MDS) | |
228 | */ | |
229 | static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq, | |
230 | struct nfs_commit_data *data) | |
231 | { | |
232 | struct nfs_writeverf *verfp; | |
233 | ||
234 | verfp = nfs_direct_select_verf(dreq, data->ds_clp, | |
235 | data->ds_commit_index); | |
80c76fe3 WAA |
236 | |
237 | /* verifier not set so always fail */ | |
238 | if (verfp->committed < 0) | |
239 | return 1; | |
240 | ||
5002c586 WAA |
241 | return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf)); |
242 | } | |
5002c586 | 243 | |
1da177e4 | 244 | /** |
b8a32e2b | 245 | * nfs_direct_IO - NFS address space operation for direct I/O |
b8a32e2b CL |
246 | * @iocb: target I/O control block |
247 | * @iov: array of vectors that define I/O buffer | |
248 | * @pos: offset in file to begin the operation | |
249 | * @nr_segs: size of iovec array | |
250 | * | |
251 | * The presence of this routine in the address space ops vector means | |
a564b8f0 MG |
252 | * the NFS client supports direct I/O. However, for most direct IO, we |
253 | * shunt off direct read and write requests before the VFS gets them, | |
254 | * so this method is only ever called for swap. | |
1da177e4 | 255 | */ |
c8b8e32d | 256 | ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter) |
b8a32e2b | 257 | { |
ee8a1a8b PT |
258 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
259 | ||
260 | /* we only support swap file calling nfs_direct_IO */ | |
261 | if (!IS_SWAPFILE(inode)) | |
262 | return 0; | |
263 | ||
66ee59af | 264 | VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE); |
a564b8f0 | 265 | |
6f673763 | 266 | if (iov_iter_rw(iter) == READ) |
c8b8e32d | 267 | return nfs_file_direct_read(iocb, iter); |
65a4a1ca | 268 | return nfs_file_direct_write(iocb, iter); |
b8a32e2b CL |
269 | } |
270 | ||
749e146e | 271 | static void nfs_direct_release_pages(struct page **pages, unsigned int npages) |
9c93ab7d | 272 | { |
749e146e | 273 | unsigned int i; |
607f31e8 | 274 | for (i = 0; i < npages; i++) |
09cbfeaf | 275 | put_page(pages[i]); |
6b45d858 TM |
276 | } |
277 | ||
1763da12 FI |
278 | void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo, |
279 | struct nfs_direct_req *dreq) | |
280 | { | |
fe238e60 | 281 | cinfo->inode = dreq->inode; |
1763da12 FI |
282 | cinfo->mds = &dreq->mds_cinfo; |
283 | cinfo->ds = &dreq->ds_cinfo; | |
284 | cinfo->dreq = dreq; | |
285 | cinfo->completion_ops = &nfs_direct_commit_completion_ops; | |
286 | } | |
287 | ||
0a00b77b WAA |
288 | static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq, |
289 | struct nfs_pageio_descriptor *pgio, | |
290 | struct nfs_page *req) | |
291 | { | |
292 | int mirror_count = 1; | |
293 | ||
294 | if (pgio->pg_ops->pg_get_mirror_count) | |
295 | mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req); | |
296 | ||
297 | dreq->mirror_count = mirror_count; | |
298 | } | |
299 | ||
93619e59 | 300 | static inline struct nfs_direct_req *nfs_direct_req_alloc(void) |
1da177e4 | 301 | { |
93619e59 CL |
302 | struct nfs_direct_req *dreq; |
303 | ||
292f3eee | 304 | dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL); |
93619e59 CL |
305 | if (!dreq) |
306 | return NULL; | |
307 | ||
308 | kref_init(&dreq->kref); | |
607f31e8 | 309 | kref_get(&dreq->kref); |
d72b7a6b | 310 | init_completion(&dreq->completion); |
1763da12 | 311 | INIT_LIST_HEAD(&dreq->mds_cinfo.list); |
5002c586 | 312 | dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */ |
1763da12 | 313 | INIT_WORK(&dreq->work, nfs_direct_write_schedule_work); |
0a00b77b | 314 | dreq->mirror_count = 1; |
15ce4a0c | 315 | spin_lock_init(&dreq->lock); |
93619e59 CL |
316 | |
317 | return dreq; | |
1da177e4 LT |
318 | } |
319 | ||
b4946ffb | 320 | static void nfs_direct_req_free(struct kref *kref) |
1da177e4 LT |
321 | { |
322 | struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); | |
a8881f5a | 323 | |
8c393f9a | 324 | nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo); |
f11ac8db TM |
325 | if (dreq->l_ctx != NULL) |
326 | nfs_put_lock_context(dreq->l_ctx); | |
a8881f5a TM |
327 | if (dreq->ctx != NULL) |
328 | put_nfs_open_context(dreq->ctx); | |
1da177e4 LT |
329 | kmem_cache_free(nfs_direct_cachep, dreq); |
330 | } | |
331 | ||
b4946ffb TM |
332 | static void nfs_direct_req_release(struct nfs_direct_req *dreq) |
333 | { | |
334 | kref_put(&dreq->kref, nfs_direct_req_free); | |
335 | } | |
336 | ||
6296556f PT |
337 | ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq) |
338 | { | |
339 | return dreq->bytes_left; | |
340 | } | |
341 | EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left); | |
342 | ||
bc0fb201 CL |
343 | /* |
344 | * Collects and returns the final error value/byte-count. | |
345 | */ | |
346 | static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) | |
347 | { | |
15ce4a0c | 348 | ssize_t result = -EIOCBQUEUED; |
bc0fb201 CL |
349 | |
350 | /* Async requests don't wait here */ | |
351 | if (dreq->iocb) | |
352 | goto out; | |
353 | ||
150030b7 | 354 | result = wait_for_completion_killable(&dreq->completion); |
bc0fb201 | 355 | |
d2a7de0b TM |
356 | if (!result) { |
357 | result = dreq->count; | |
358 | WARN_ON_ONCE(dreq->count < 0); | |
359 | } | |
bc0fb201 | 360 | if (!result) |
15ce4a0c | 361 | result = dreq->error; |
bc0fb201 CL |
362 | |
363 | out: | |
bc0fb201 CL |
364 | return (ssize_t) result; |
365 | } | |
366 | ||
63ab46ab | 367 | /* |
607f31e8 TM |
368 | * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust |
369 | * the iocb is still valid here if this is a synchronous request. | |
63ab46ab | 370 | */ |
9811cd57 | 371 | static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write) |
63ab46ab | 372 | { |
9811cd57 CH |
373 | struct inode *inode = dreq->inode; |
374 | ||
2a009ec9 | 375 | if (dreq->iocb && write) { |
9811cd57 | 376 | loff_t pos = dreq->iocb->ki_pos + dreq->count; |
2a009ec9 CH |
377 | |
378 | spin_lock(&inode->i_lock); | |
379 | if (i_size_read(inode) < pos) | |
380 | i_size_write(inode, pos); | |
381 | spin_unlock(&inode->i_lock); | |
382 | } | |
383 | ||
1f90ee27 | 384 | if (write) |
2a009ec9 | 385 | nfs_zap_mapping(inode, inode->i_mapping); |
1f90ee27 | 386 | |
fe0f07d0 | 387 | inode_dio_end(inode); |
2a009ec9 CH |
388 | |
389 | if (dreq->iocb) { | |
15ce4a0c | 390 | long res = (long) dreq->error; |
d2a7de0b | 391 | if (dreq->count != 0) { |
15ce4a0c | 392 | res = (long) dreq->count; |
d2a7de0b TM |
393 | WARN_ON_ONCE(dreq->count < 0); |
394 | } | |
04b2fa9f | 395 | dreq->iocb->ki_complete(dreq->iocb, res, 0); |
d72b7a6b | 396 | } |
2a009ec9 | 397 | |
d72b7a6b | 398 | complete_all(&dreq->completion); |
63ab46ab | 399 | |
b4946ffb | 400 | nfs_direct_req_release(dreq); |
63ab46ab CL |
401 | } |
402 | ||
1385b811 | 403 | static void nfs_direct_readpage_release(struct nfs_page *req) |
1da177e4 | 404 | { |
1e8968c5 | 405 | dprintk("NFS: direct read done (%s/%llu %d@%lld)\n", |
fc64005c | 406 | req->wb_context->dentry->d_sb->s_id, |
2b0143b5 | 407 | (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)), |
584aa810 FI |
408 | req->wb_bytes, |
409 | (long long)req_offset(req)); | |
410 | nfs_release_request(req); | |
fdd1e74c TM |
411 | } |
412 | ||
584aa810 | 413 | static void nfs_direct_read_completion(struct nfs_pgio_header *hdr) |
fdd1e74c | 414 | { |
584aa810 FI |
415 | unsigned long bytes = 0; |
416 | struct nfs_direct_req *dreq = hdr->dreq; | |
fdd1e74c | 417 | |
584aa810 FI |
418 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) |
419 | goto out_put; | |
15ce4a0c CL |
420 | |
421 | spin_lock(&dreq->lock); | |
584aa810 FI |
422 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0)) |
423 | dreq->error = hdr->error; | |
424 | else | |
0a00b77b WAA |
425 | nfs_direct_good_bytes(dreq, hdr); |
426 | ||
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 | ||
be7e9858 JL |
433 | if (!PageCompound(page) && bytes < hdr->good_bytes) |
434 | set_page_dirty(page); | |
4bd8b010 TM |
435 | bytes += req->wb_bytes; |
436 | nfs_list_remove_request(req); | |
437 | nfs_direct_readpage_release(req); | |
d4a8f367 | 438 | } |
584aa810 | 439 | out_put: |
607f31e8 | 440 | if (put_dreq(dreq)) |
9811cd57 | 441 | nfs_direct_complete(dreq, false); |
584aa810 | 442 | hdr->release(hdr); |
1da177e4 LT |
443 | } |
444 | ||
3e9e0ca3 | 445 | static void nfs_read_sync_pgio_error(struct list_head *head) |
cd841605 | 446 | { |
584aa810 | 447 | struct nfs_page *req; |
cd841605 | 448 | |
584aa810 FI |
449 | while (!list_empty(head)) { |
450 | req = nfs_list_entry(head->next); | |
451 | nfs_list_remove_request(req); | |
452 | nfs_release_request(req); | |
453 | } | |
cd841605 FI |
454 | } |
455 | ||
584aa810 FI |
456 | static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr) |
457 | { | |
458 | get_dreq(hdr->dreq); | |
459 | } | |
460 | ||
461 | static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = { | |
3e9e0ca3 | 462 | .error_cleanup = nfs_read_sync_pgio_error, |
584aa810 FI |
463 | .init_hdr = nfs_direct_pgio_init, |
464 | .completion = nfs_direct_read_completion, | |
465 | }; | |
466 | ||
d4cc948b | 467 | /* |
607f31e8 TM |
468 | * For each rsize'd chunk of the user's buffer, dispatch an NFS READ |
469 | * operation. If nfs_readdata_alloc() or get_user_pages() fails, | |
470 | * bail and stop sending more reads. Read length accounting is | |
471 | * handled automatically by nfs_direct_read_result(). Otherwise, if | |
472 | * no requests have been sent, just return an error. | |
1da177e4 | 473 | */ |
1da177e4 | 474 | |
91f79c43 AV |
475 | static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq, |
476 | struct iov_iter *iter, | |
477 | loff_t pos) | |
478 | { | |
479 | struct nfs_pageio_descriptor desc; | |
480 | struct inode *inode = dreq->inode; | |
481 | ssize_t result = -EINVAL; | |
482 | size_t requested_bytes = 0; | |
483 | size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE); | |
e9f7bee1 | 484 | |
16b90578 | 485 | nfs_pageio_init_read(&desc, dreq->inode, false, |
91f79c43 AV |
486 | &nfs_direct_read_completion_ops); |
487 | get_dreq(dreq); | |
488 | desc.pg_dreq = dreq; | |
fe0f07d0 | 489 | inode_dio_begin(inode); |
a564b8f0 | 490 | |
91f79c43 AV |
491 | while (iov_iter_count(iter)) { |
492 | struct page **pagevec; | |
493 | size_t bytes; | |
494 | size_t pgbase; | |
495 | unsigned npages, i; | |
607f31e8 | 496 | |
91f79c43 AV |
497 | result = iov_iter_get_pages_alloc(iter, &pagevec, |
498 | rsize, &pgbase); | |
499 | if (result < 0) | |
500 | break; | |
501 | ||
502 | bytes = result; | |
503 | iov_iter_advance(iter, bytes); | |
504 | npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; | |
584aa810 FI |
505 | for (i = 0; i < npages; i++) { |
506 | struct nfs_page *req; | |
bf5fc402 | 507 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); |
584aa810 | 508 | /* XXX do we need to do the eof zeroing found in async_filler? */ |
2bfc6e56 | 509 | req = nfs_create_request(dreq->ctx, pagevec[i], NULL, |
584aa810 FI |
510 | pgbase, req_len); |
511 | if (IS_ERR(req)) { | |
584aa810 FI |
512 | result = PTR_ERR(req); |
513 | break; | |
514 | } | |
515 | req->wb_index = pos >> PAGE_SHIFT; | |
516 | req->wb_offset = pos & ~PAGE_MASK; | |
91f79c43 AV |
517 | if (!nfs_pageio_add_request(&desc, req)) { |
518 | result = desc.pg_error; | |
584aa810 | 519 | nfs_release_request(req); |
584aa810 FI |
520 | break; |
521 | } | |
522 | pgbase = 0; | |
523 | bytes -= req_len; | |
91f79c43 | 524 | requested_bytes += req_len; |
584aa810 | 525 | pos += req_len; |
35754bc0 | 526 | dreq->bytes_left -= req_len; |
584aa810 | 527 | } |
6d74743b | 528 | nfs_direct_release_pages(pagevec, npages); |
91f79c43 | 529 | kvfree(pagevec); |
19f73787 CL |
530 | if (result < 0) |
531 | break; | |
19f73787 CL |
532 | } |
533 | ||
584aa810 FI |
534 | nfs_pageio_complete(&desc); |
535 | ||
839f7ad6 CL |
536 | /* |
537 | * If no bytes were started, return the error, and let the | |
538 | * generic layer handle the completion. | |
539 | */ | |
540 | if (requested_bytes == 0) { | |
fe0f07d0 | 541 | inode_dio_end(inode); |
839f7ad6 CL |
542 | nfs_direct_req_release(dreq); |
543 | return result < 0 ? result : -EIO; | |
544 | } | |
545 | ||
19f73787 | 546 | if (put_dreq(dreq)) |
9811cd57 | 547 | nfs_direct_complete(dreq, false); |
839f7ad6 | 548 | return 0; |
19f73787 CL |
549 | } |
550 | ||
14a3ec79 CH |
551 | /** |
552 | * nfs_file_direct_read - file direct read operation for NFS files | |
553 | * @iocb: target I/O control block | |
619d30b4 | 554 | * @iter: vector of user buffers into which to read data |
14a3ec79 CH |
555 | * |
556 | * We use this function for direct reads instead of calling | |
557 | * generic_file_aio_read() in order to avoid gfar's check to see if | |
558 | * the request starts before the end of the file. For that check | |
559 | * to work, we must generate a GETATTR before each direct read, and | |
560 | * even then there is a window between the GETATTR and the subsequent | |
561 | * READ where the file size could change. Our preference is simply | |
562 | * to do all reads the application wants, and the server will take | |
563 | * care of managing the end of file boundary. | |
564 | * | |
565 | * This function also eliminates unnecessarily updating the file's | |
566 | * atime locally, as the NFS server sets the file's atime, and this | |
567 | * client must read the updated atime from the server back into its | |
568 | * cache. | |
569 | */ | |
c8b8e32d | 570 | ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter) |
1da177e4 | 571 | { |
14a3ec79 CH |
572 | struct file *file = iocb->ki_filp; |
573 | struct address_space *mapping = file->f_mapping; | |
574 | struct inode *inode = mapping->host; | |
1da177e4 | 575 | struct nfs_direct_req *dreq; |
b3c54de6 | 576 | struct nfs_lock_context *l_ctx; |
14a3ec79 | 577 | ssize_t result = -EINVAL; |
a6cbcd4a | 578 | size_t count = iov_iter_count(iter); |
14a3ec79 CH |
579 | nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); |
580 | ||
581 | dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n", | |
c8b8e32d | 582 | file, count, (long long) iocb->ki_pos); |
14a3ec79 CH |
583 | |
584 | result = 0; | |
585 | if (!count) | |
586 | goto out; | |
587 | ||
5955102c | 588 | inode_lock(inode); |
14a3ec79 CH |
589 | result = nfs_sync_mapping(mapping); |
590 | if (result) | |
d0b9875d | 591 | goto out_unlock; |
1da177e4 | 592 | |
14a3ec79 CH |
593 | task_io_account_read(count); |
594 | ||
595 | result = -ENOMEM; | |
607f31e8 | 596 | dreq = nfs_direct_req_alloc(); |
f11ac8db | 597 | if (dreq == NULL) |
d0b9875d | 598 | goto out_unlock; |
1da177e4 | 599 | |
91d5b470 | 600 | dreq->inode = inode; |
ed3743a6 | 601 | dreq->bytes_left = dreq->max_count = count; |
c8b8e32d | 602 | dreq->io_start = iocb->ki_pos; |
cd3758e3 | 603 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
b3c54de6 TM |
604 | l_ctx = nfs_get_lock_context(dreq->ctx); |
605 | if (IS_ERR(l_ctx)) { | |
606 | result = PTR_ERR(l_ctx); | |
f11ac8db | 607 | goto out_release; |
b3c54de6 TM |
608 | } |
609 | dreq->l_ctx = l_ctx; | |
487b8372 CL |
610 | if (!is_sync_kiocb(iocb)) |
611 | dreq->iocb = iocb; | |
1da177e4 | 612 | |
619d30b4 | 613 | NFS_I(inode)->read_io += count; |
c8b8e32d | 614 | result = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos); |
d0b9875d | 615 | |
5955102c | 616 | inode_unlock(inode); |
d0b9875d | 617 | |
14a3ec79 | 618 | if (!result) { |
607f31e8 | 619 | result = nfs_direct_wait(dreq); |
14a3ec79 | 620 | if (result > 0) |
c8b8e32d | 621 | iocb->ki_pos += result; |
14a3ec79 | 622 | } |
d0b9875d CH |
623 | |
624 | nfs_direct_req_release(dreq); | |
625 | return result; | |
626 | ||
f11ac8db | 627 | out_release: |
b4946ffb | 628 | nfs_direct_req_release(dreq); |
d0b9875d | 629 | out_unlock: |
5955102c | 630 | inode_unlock(inode); |
f11ac8db | 631 | out: |
1da177e4 LT |
632 | return result; |
633 | } | |
634 | ||
085d1e33 TH |
635 | static void |
636 | nfs_direct_write_scan_commit_list(struct inode *inode, | |
637 | struct list_head *list, | |
638 | struct nfs_commit_info *cinfo) | |
639 | { | |
fe238e60 | 640 | spin_lock(&cinfo->inode->i_lock); |
085d1e33 TH |
641 | #ifdef CONFIG_NFS_V4_1 |
642 | if (cinfo->ds != NULL && cinfo->ds->nwritten != 0) | |
643 | NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo); | |
644 | #endif | |
645 | nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0); | |
fe238e60 | 646 | spin_unlock(&cinfo->inode->i_lock); |
085d1e33 TH |
647 | } |
648 | ||
fad61490 TM |
649 | static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) |
650 | { | |
1763da12 FI |
651 | struct nfs_pageio_descriptor desc; |
652 | struct nfs_page *req, *tmp; | |
653 | LIST_HEAD(reqs); | |
654 | struct nfs_commit_info cinfo; | |
655 | LIST_HEAD(failed); | |
0a00b77b | 656 | int i; |
1763da12 FI |
657 | |
658 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
085d1e33 | 659 | nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo); |
1da177e4 | 660 | |
fad61490 | 661 | dreq->count = 0; |
0a00b77b WAA |
662 | for (i = 0; i < dreq->mirror_count; i++) |
663 | dreq->mirrors[i].count = 0; | |
607f31e8 TM |
664 | get_dreq(dreq); |
665 | ||
a20c93e3 | 666 | nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false, |
1763da12 FI |
667 | &nfs_direct_write_completion_ops); |
668 | desc.pg_dreq = dreq; | |
fedb595c | 669 | |
0a00b77b WAA |
670 | req = nfs_list_entry(reqs.next); |
671 | nfs_direct_setup_mirroring(dreq, &desc, req); | |
d600ad1f PT |
672 | if (desc.pg_error < 0) { |
673 | list_splice_init(&reqs, &failed); | |
674 | goto out_failed; | |
675 | } | |
0a00b77b | 676 | |
1763da12 FI |
677 | list_for_each_entry_safe(req, tmp, &reqs, wb_list) { |
678 | if (!nfs_pageio_add_request(&desc, req)) { | |
4035c248 | 679 | nfs_list_remove_request(req); |
1763da12 | 680 | nfs_list_add_request(req, &failed); |
fe238e60 | 681 | spin_lock(&cinfo.inode->i_lock); |
1763da12 | 682 | dreq->flags = 0; |
d600ad1f PT |
683 | if (desc.pg_error < 0) |
684 | dreq->error = desc.pg_error; | |
685 | else | |
686 | dreq->error = -EIO; | |
fe238e60 | 687 | spin_unlock(&cinfo.inode->i_lock); |
1763da12 | 688 | } |
5a695da2 | 689 | nfs_release_request(req); |
1763da12 FI |
690 | } |
691 | nfs_pageio_complete(&desc); | |
fad61490 | 692 | |
d600ad1f | 693 | out_failed: |
4035c248 TM |
694 | while (!list_empty(&failed)) { |
695 | req = nfs_list_entry(failed.next); | |
696 | nfs_list_remove_request(req); | |
1d1afcbc | 697 | nfs_unlock_and_release_request(req); |
4035c248 | 698 | } |
fad61490 | 699 | |
1763da12 FI |
700 | if (put_dreq(dreq)) |
701 | nfs_direct_write_complete(dreq, dreq->inode); | |
c9d8f89d TM |
702 | } |
703 | ||
1763da12 | 704 | static void nfs_direct_commit_complete(struct nfs_commit_data *data) |
c9d8f89d | 705 | { |
0b7c0153 | 706 | struct nfs_direct_req *dreq = data->dreq; |
1763da12 FI |
707 | struct nfs_commit_info cinfo; |
708 | struct nfs_page *req; | |
c9d8f89d TM |
709 | int status = data->task.tk_status; |
710 | ||
1763da12 | 711 | nfs_init_cinfo_from_dreq(&cinfo, dreq); |
c9d8f89d | 712 | if (status < 0) { |
60fa3f76 | 713 | dprintk("NFS: %5u commit failed with error %d.\n", |
1763da12 | 714 | data->task.tk_pid, status); |
fad61490 | 715 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
5002c586 | 716 | } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) { |
c9d8f89d | 717 | dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid); |
fad61490 | 718 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
1da177e4 LT |
719 | } |
720 | ||
c9d8f89d | 721 | dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status); |
1763da12 FI |
722 | while (!list_empty(&data->pages)) { |
723 | req = nfs_list_entry(data->pages.next); | |
724 | nfs_list_remove_request(req); | |
725 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) { | |
726 | /* Note the rewrite will go through mds */ | |
b57ff130 | 727 | nfs_mark_request_commit(req, NULL, &cinfo, 0); |
906369e4 FI |
728 | } else |
729 | nfs_release_request(req); | |
1d1afcbc | 730 | nfs_unlock_and_release_request(req); |
1763da12 FI |
731 | } |
732 | ||
733 | if (atomic_dec_and_test(&cinfo.mds->rpcs_out)) | |
734 | nfs_direct_write_complete(dreq, data->inode); | |
1da177e4 LT |
735 | } |
736 | ||
b20135d0 TM |
737 | static void nfs_direct_resched_write(struct nfs_commit_info *cinfo, |
738 | struct nfs_page *req) | |
1763da12 | 739 | { |
b20135d0 TM |
740 | struct nfs_direct_req *dreq = cinfo->dreq; |
741 | ||
742 | spin_lock(&dreq->lock); | |
743 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; | |
744 | spin_unlock(&dreq->lock); | |
745 | nfs_mark_request_commit(req, NULL, cinfo, 0); | |
1763da12 FI |
746 | } |
747 | ||
748 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = { | |
749 | .completion = nfs_direct_commit_complete, | |
b20135d0 | 750 | .resched_write = nfs_direct_resched_write, |
fad61490 TM |
751 | }; |
752 | ||
753 | static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) | |
1da177e4 | 754 | { |
1763da12 FI |
755 | int res; |
756 | struct nfs_commit_info cinfo; | |
757 | LIST_HEAD(mds_list); | |
758 | ||
759 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
760 | nfs_scan_commit(dreq->inode, &mds_list, &cinfo); | |
761 | res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo); | |
762 | if (res < 0) /* res == -ENOMEM */ | |
763 | nfs_direct_write_reschedule(dreq); | |
fad61490 | 764 | } |
1da177e4 | 765 | |
1763da12 | 766 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
fad61490 | 767 | { |
1763da12 | 768 | struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work); |
fad61490 | 769 | int flags = dreq->flags; |
1da177e4 | 770 | |
fad61490 TM |
771 | dreq->flags = 0; |
772 | switch (flags) { | |
773 | case NFS_ODIRECT_DO_COMMIT: | |
774 | nfs_direct_commit_schedule(dreq); | |
1da177e4 | 775 | break; |
fad61490 TM |
776 | case NFS_ODIRECT_RESCHED_WRITES: |
777 | nfs_direct_write_reschedule(dreq); | |
778 | break; | |
779 | default: | |
9811cd57 | 780 | nfs_direct_complete(dreq, true); |
fad61490 TM |
781 | } |
782 | } | |
1da177e4 | 783 | |
1763da12 | 784 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
fad61490 | 785 | { |
1763da12 | 786 | schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */ |
fad61490 | 787 | } |
1763da12 | 788 | |
1763da12 FI |
789 | static void nfs_direct_write_completion(struct nfs_pgio_header *hdr) |
790 | { | |
791 | struct nfs_direct_req *dreq = hdr->dreq; | |
792 | struct nfs_commit_info cinfo; | |
c65e6254 | 793 | bool request_commit = false; |
1763da12 FI |
794 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); |
795 | ||
796 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) | |
797 | goto out_put; | |
798 | ||
799 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
800 | ||
801 | spin_lock(&dreq->lock); | |
802 | ||
803 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) { | |
804 | dreq->flags = 0; | |
805 | dreq->error = hdr->error; | |
806 | } | |
c65e6254 | 807 | if (dreq->error == 0) { |
0a00b77b | 808 | nfs_direct_good_bytes(dreq, hdr); |
c65e6254 | 809 | if (nfs_write_need_commit(hdr)) { |
1763da12 | 810 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) |
c65e6254 | 811 | request_commit = true; |
1763da12 | 812 | else if (dreq->flags == 0) { |
5002c586 | 813 | nfs_direct_set_hdr_verf(dreq, hdr); |
c65e6254 | 814 | request_commit = true; |
1763da12 FI |
815 | dreq->flags = NFS_ODIRECT_DO_COMMIT; |
816 | } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) { | |
c65e6254 WAA |
817 | request_commit = true; |
818 | if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr)) | |
5002c586 WAA |
819 | dreq->flags = |
820 | NFS_ODIRECT_RESCHED_WRITES; | |
1763da12 FI |
821 | } |
822 | } | |
823 | } | |
824 | spin_unlock(&dreq->lock); | |
825 | ||
826 | while (!list_empty(&hdr->pages)) { | |
2bfc6e56 | 827 | |
1763da12 FI |
828 | req = nfs_list_entry(hdr->pages.next); |
829 | nfs_list_remove_request(req); | |
c65e6254 | 830 | if (request_commit) { |
04277086 | 831 | kref_get(&req->wb_kref); |
b57ff130 WAA |
832 | nfs_mark_request_commit(req, hdr->lseg, &cinfo, |
833 | hdr->ds_commit_idx); | |
1763da12 | 834 | } |
1d1afcbc | 835 | nfs_unlock_and_release_request(req); |
1763da12 FI |
836 | } |
837 | ||
838 | out_put: | |
839 | if (put_dreq(dreq)) | |
840 | nfs_direct_write_complete(dreq, hdr->inode); | |
841 | hdr->release(hdr); | |
842 | } | |
843 | ||
3e9e0ca3 TM |
844 | static void nfs_write_sync_pgio_error(struct list_head *head) |
845 | { | |
846 | struct nfs_page *req; | |
847 | ||
848 | while (!list_empty(head)) { | |
849 | req = nfs_list_entry(head->next); | |
850 | nfs_list_remove_request(req); | |
1d1afcbc | 851 | nfs_unlock_and_release_request(req); |
3e9e0ca3 TM |
852 | } |
853 | } | |
854 | ||
dc602dd7 TM |
855 | static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr) |
856 | { | |
857 | struct nfs_direct_req *dreq = hdr->dreq; | |
858 | ||
859 | spin_lock(&dreq->lock); | |
860 | if (dreq->error == 0) { | |
861 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; | |
862 | /* fake unstable write to let common nfs resend pages */ | |
863 | hdr->verf.committed = NFS_UNSTABLE; | |
864 | hdr->good_bytes = hdr->args.count; | |
865 | } | |
866 | spin_unlock(&dreq->lock); | |
867 | } | |
868 | ||
1763da12 | 869 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = { |
3e9e0ca3 | 870 | .error_cleanup = nfs_write_sync_pgio_error, |
1763da12 FI |
871 | .init_hdr = nfs_direct_pgio_init, |
872 | .completion = nfs_direct_write_completion, | |
dc602dd7 | 873 | .reschedule_io = nfs_direct_write_reschedule_io, |
1763da12 FI |
874 | }; |
875 | ||
91f79c43 AV |
876 | |
877 | /* | |
878 | * NB: Return the value of the first error return code. Subsequent | |
879 | * errors after the first one are ignored. | |
880 | */ | |
881 | /* | |
882 | * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE | |
883 | * operation. If nfs_writedata_alloc() or get_user_pages() fails, | |
884 | * bail and stop sending more writes. Write length accounting is | |
885 | * handled automatically by nfs_direct_write_result(). Otherwise, if | |
886 | * no requests have been sent, just return an error. | |
887 | */ | |
19f73787 | 888 | static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, |
619d30b4 | 889 | struct iov_iter *iter, |
91f79c43 | 890 | loff_t pos) |
19f73787 | 891 | { |
1763da12 | 892 | struct nfs_pageio_descriptor desc; |
1d59d61f | 893 | struct inode *inode = dreq->inode; |
19f73787 CL |
894 | ssize_t result = 0; |
895 | size_t requested_bytes = 0; | |
91f79c43 | 896 | size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE); |
19f73787 | 897 | |
a20c93e3 | 898 | nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false, |
1763da12 FI |
899 | &nfs_direct_write_completion_ops); |
900 | desc.pg_dreq = dreq; | |
19f73787 | 901 | get_dreq(dreq); |
fe0f07d0 | 902 | inode_dio_begin(inode); |
19f73787 | 903 | |
91f79c43 AV |
904 | NFS_I(inode)->write_io += iov_iter_count(iter); |
905 | while (iov_iter_count(iter)) { | |
906 | struct page **pagevec; | |
907 | size_t bytes; | |
908 | size_t pgbase; | |
909 | unsigned npages, i; | |
910 | ||
911 | result = iov_iter_get_pages_alloc(iter, &pagevec, | |
912 | wsize, &pgbase); | |
19f73787 CL |
913 | if (result < 0) |
914 | break; | |
91f79c43 AV |
915 | |
916 | bytes = result; | |
917 | iov_iter_advance(iter, bytes); | |
918 | npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; | |
919 | for (i = 0; i < npages; i++) { | |
920 | struct nfs_page *req; | |
921 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); | |
922 | ||
16b90578 | 923 | req = nfs_create_request(dreq->ctx, pagevec[i], NULL, |
91f79c43 AV |
924 | pgbase, req_len); |
925 | if (IS_ERR(req)) { | |
926 | result = PTR_ERR(req); | |
927 | break; | |
928 | } | |
0a00b77b WAA |
929 | |
930 | nfs_direct_setup_mirroring(dreq, &desc, req); | |
d600ad1f PT |
931 | if (desc.pg_error < 0) { |
932 | nfs_free_request(req); | |
933 | result = desc.pg_error; | |
934 | break; | |
935 | } | |
0a00b77b | 936 | |
91f79c43 AV |
937 | nfs_lock_request(req); |
938 | req->wb_index = pos >> PAGE_SHIFT; | |
939 | req->wb_offset = pos & ~PAGE_MASK; | |
940 | if (!nfs_pageio_add_request(&desc, req)) { | |
941 | result = desc.pg_error; | |
942 | nfs_unlock_and_release_request(req); | |
943 | break; | |
944 | } | |
945 | pgbase = 0; | |
946 | bytes -= req_len; | |
947 | requested_bytes += req_len; | |
948 | pos += req_len; | |
949 | dreq->bytes_left -= req_len; | |
950 | } | |
951 | nfs_direct_release_pages(pagevec, npages); | |
952 | kvfree(pagevec); | |
953 | if (result < 0) | |
19f73787 | 954 | break; |
19f73787 | 955 | } |
1763da12 | 956 | nfs_pageio_complete(&desc); |
19f73787 | 957 | |
839f7ad6 CL |
958 | /* |
959 | * If no bytes were started, return the error, and let the | |
960 | * generic layer handle the completion. | |
961 | */ | |
962 | if (requested_bytes == 0) { | |
fe0f07d0 | 963 | inode_dio_end(inode); |
839f7ad6 CL |
964 | nfs_direct_req_release(dreq); |
965 | return result < 0 ? result : -EIO; | |
966 | } | |
967 | ||
19f73787 CL |
968 | if (put_dreq(dreq)) |
969 | nfs_direct_write_complete(dreq, dreq->inode); | |
839f7ad6 | 970 | return 0; |
19f73787 CL |
971 | } |
972 | ||
1da177e4 LT |
973 | /** |
974 | * nfs_file_direct_write - file direct write operation for NFS files | |
975 | * @iocb: target I/O control block | |
619d30b4 | 976 | * @iter: vector of user buffers from which to write data |
1da177e4 LT |
977 | * |
978 | * We use this function for direct writes instead of calling | |
979 | * generic_file_aio_write() in order to avoid taking the inode | |
980 | * semaphore and updating the i_size. The NFS server will set | |
981 | * the new i_size and this client must read the updated size | |
982 | * back into its cache. We let the server do generic write | |
983 | * parameter checking and report problems. | |
984 | * | |
1da177e4 LT |
985 | * We eliminate local atime updates, see direct read above. |
986 | * | |
987 | * We avoid unnecessary page cache invalidations for normal cached | |
988 | * readers of this file. | |
989 | * | |
990 | * Note that O_APPEND is not supported for NFS direct writes, as there | |
991 | * is no atomic O_APPEND write facility in the NFS protocol. | |
992 | */ | |
65a4a1ca | 993 | ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter) |
1da177e4 | 994 | { |
22cd1bf1 | 995 | ssize_t result = -EINVAL; |
1da177e4 | 996 | struct file *file = iocb->ki_filp; |
1da177e4 | 997 | struct address_space *mapping = file->f_mapping; |
22cd1bf1 CH |
998 | struct inode *inode = mapping->host; |
999 | struct nfs_direct_req *dreq; | |
1000 | struct nfs_lock_context *l_ctx; | |
65a4a1ca | 1001 | loff_t pos, end; |
c216fd70 | 1002 | |
6de1472f | 1003 | dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n", |
3309dd04 | 1004 | file, iov_iter_count(iter), (long long) iocb->ki_pos); |
027445c3 | 1005 | |
3309dd04 AV |
1006 | nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, |
1007 | iov_iter_count(iter)); | |
1008 | ||
1009 | pos = iocb->ki_pos; | |
09cbfeaf | 1010 | end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT; |
ce1a8e67 | 1011 | |
5955102c | 1012 | inode_lock(inode); |
a9ab5e84 | 1013 | |
22cd1bf1 CH |
1014 | result = nfs_sync_mapping(mapping); |
1015 | if (result) | |
a9ab5e84 CH |
1016 | goto out_unlock; |
1017 | ||
1018 | if (mapping->nrpages) { | |
1019 | result = invalidate_inode_pages2_range(mapping, | |
09cbfeaf | 1020 | pos >> PAGE_SHIFT, end); |
a9ab5e84 CH |
1021 | if (result) |
1022 | goto out_unlock; | |
1023 | } | |
1da177e4 | 1024 | |
3309dd04 | 1025 | task_io_account_write(iov_iter_count(iter)); |
7ec10f26 | 1026 | |
22cd1bf1 CH |
1027 | result = -ENOMEM; |
1028 | dreq = nfs_direct_req_alloc(); | |
1029 | if (!dreq) | |
a9ab5e84 | 1030 | goto out_unlock; |
9eafa8cc | 1031 | |
22cd1bf1 | 1032 | dreq->inode = inode; |
ed3743a6 | 1033 | dreq->bytes_left = dreq->max_count = iov_iter_count(iter); |
5fadeb47 | 1034 | dreq->io_start = pos; |
22cd1bf1 CH |
1035 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
1036 | l_ctx = nfs_get_lock_context(dreq->ctx); | |
1037 | if (IS_ERR(l_ctx)) { | |
1038 | result = PTR_ERR(l_ctx); | |
1039 | goto out_release; | |
1040 | } | |
1041 | dreq->l_ctx = l_ctx; | |
1042 | if (!is_sync_kiocb(iocb)) | |
1043 | dreq->iocb = iocb; | |
1044 | ||
91f79c43 | 1045 | result = nfs_direct_write_schedule_iovec(dreq, iter, pos); |
a9ab5e84 CH |
1046 | |
1047 | if (mapping->nrpages) { | |
1048 | invalidate_inode_pages2_range(mapping, | |
09cbfeaf | 1049 | pos >> PAGE_SHIFT, end); |
a9ab5e84 CH |
1050 | } |
1051 | ||
5955102c | 1052 | inode_unlock(inode); |
a9ab5e84 | 1053 | |
22cd1bf1 CH |
1054 | if (!result) { |
1055 | result = nfs_direct_wait(dreq); | |
1056 | if (result > 0) { | |
1057 | struct inode *inode = mapping->host; | |
1058 | ||
1059 | iocb->ki_pos = pos + result; | |
1060 | spin_lock(&inode->i_lock); | |
1061 | if (i_size_read(inode) < iocb->ki_pos) | |
1062 | i_size_write(inode, iocb->ki_pos); | |
1063 | spin_unlock(&inode->i_lock); | |
e2592217 CH |
1064 | |
1065 | /* XXX: should check the generic_write_sync retval */ | |
1066 | generic_write_sync(iocb, result); | |
22cd1bf1 | 1067 | } |
1763da12 | 1068 | } |
a9ab5e84 CH |
1069 | nfs_direct_req_release(dreq); |
1070 | return result; | |
1071 | ||
22cd1bf1 CH |
1072 | out_release: |
1073 | nfs_direct_req_release(dreq); | |
a9ab5e84 | 1074 | out_unlock: |
5955102c | 1075 | inode_unlock(inode); |
22cd1bf1 | 1076 | return result; |
1da177e4 LT |
1077 | } |
1078 | ||
88467055 CL |
1079 | /** |
1080 | * nfs_init_directcache - create a slab cache for nfs_direct_req structures | |
1081 | * | |
1082 | */ | |
f7b422b1 | 1083 | int __init nfs_init_directcache(void) |
1da177e4 LT |
1084 | { |
1085 | nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", | |
1086 | sizeof(struct nfs_direct_req), | |
fffb60f9 PJ |
1087 | 0, (SLAB_RECLAIM_ACCOUNT| |
1088 | SLAB_MEM_SPREAD), | |
20c2df83 | 1089 | NULL); |
1da177e4 LT |
1090 | if (nfs_direct_cachep == NULL) |
1091 | return -ENOMEM; | |
1092 | ||
1093 | return 0; | |
1094 | } | |
1095 | ||
88467055 | 1096 | /** |
f7b422b1 | 1097 | * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures |
88467055 CL |
1098 | * |
1099 | */ | |
266bee88 | 1100 | void nfs_destroy_directcache(void) |
1da177e4 | 1101 | { |
1a1d92c1 | 1102 | kmem_cache_destroy(nfs_direct_cachep); |
1da177e4 | 1103 | } |