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 | */ | |
69 | struct nfs_direct_req { | |
70 | struct kref kref; /* release manager */ | |
15ce4a0c CL |
71 | |
72 | /* I/O parameters */ | |
a8881f5a | 73 | struct nfs_open_context *ctx; /* file open context info */ |
f11ac8db | 74 | struct nfs_lock_context *l_ctx; /* Lock context info */ |
99514f8f | 75 | struct kiocb * iocb; /* controlling i/o request */ |
88467055 | 76 | struct inode * inode; /* target file of i/o */ |
15ce4a0c CL |
77 | |
78 | /* completion state */ | |
607f31e8 | 79 | atomic_t io_count; /* i/os we're waiting for */ |
15ce4a0c | 80 | spinlock_t lock; /* protect completion state */ |
15ce4a0c | 81 | ssize_t count, /* bytes actually processed */ |
35754bc0 | 82 | bytes_left, /* bytes left to be sent */ |
1da177e4 | 83 | error; /* any reported error */ |
d72b7a6b | 84 | struct completion completion; /* wait for i/o completion */ |
fad61490 TM |
85 | |
86 | /* commit state */ | |
1763da12 FI |
87 | struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo */ |
88 | struct pnfs_ds_commit_info ds_cinfo; /* Storage for cinfo */ | |
89 | struct work_struct work; | |
fad61490 TM |
90 | int flags; |
91 | #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */ | |
92 | #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */ | |
93 | struct nfs_writeverf verf; /* unstable write verifier */ | |
1da177e4 LT |
94 | }; |
95 | ||
1763da12 FI |
96 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops; |
97 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops; | |
fad61490 | 98 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode); |
1763da12 | 99 | static void nfs_direct_write_schedule_work(struct work_struct *work); |
607f31e8 TM |
100 | |
101 | static inline void get_dreq(struct nfs_direct_req *dreq) | |
102 | { | |
103 | atomic_inc(&dreq->io_count); | |
104 | } | |
105 | ||
106 | static inline int put_dreq(struct nfs_direct_req *dreq) | |
107 | { | |
108 | return atomic_dec_and_test(&dreq->io_count); | |
109 | } | |
110 | ||
5002c586 WAA |
111 | /* |
112 | * nfs_direct_select_verf - select the right verifier | |
113 | * @dreq - direct request possibly spanning multiple servers | |
114 | * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs | |
115 | * @ds_idx - index of data server in data server list, only valid if ds_clp set | |
116 | * | |
117 | * returns the correct verifier to use given the role of the server | |
118 | */ | |
119 | static struct nfs_writeverf * | |
120 | nfs_direct_select_verf(struct nfs_direct_req *dreq, | |
121 | struct nfs_client *ds_clp, | |
122 | int ds_idx) | |
123 | { | |
124 | struct nfs_writeverf *verfp = &dreq->verf; | |
125 | ||
126 | #ifdef CONFIG_NFS_V4_1 | |
127 | if (ds_clp) { | |
128 | /* pNFS is in use, use the DS verf */ | |
129 | if (ds_idx >= 0 && ds_idx < dreq->ds_cinfo.nbuckets) | |
130 | verfp = &dreq->ds_cinfo.buckets[ds_idx].direct_verf; | |
131 | else | |
132 | WARN_ON_ONCE(1); | |
133 | } | |
134 | #endif | |
135 | return verfp; | |
136 | } | |
137 | ||
138 | ||
139 | /* | |
140 | * nfs_direct_set_hdr_verf - set the write/commit verifier | |
141 | * @dreq - direct request possibly spanning multiple servers | |
142 | * @hdr - pageio header to validate against previously seen verfs | |
143 | * | |
144 | * Set the server's (MDS or DS) "seen" verifier | |
145 | */ | |
146 | static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq, | |
147 | struct nfs_pgio_header *hdr) | |
148 | { | |
149 | struct nfs_writeverf *verfp; | |
150 | ||
151 | verfp = nfs_direct_select_verf(dreq, hdr->data->ds_clp, | |
152 | hdr->data->ds_idx); | |
153 | WARN_ON_ONCE(verfp->committed >= 0); | |
154 | memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf)); | |
155 | WARN_ON_ONCE(verfp->committed < 0); | |
156 | } | |
157 | ||
158 | /* | |
159 | * nfs_direct_cmp_hdr_verf - compare verifier for pgio header | |
160 | * @dreq - direct request possibly spanning multiple servers | |
161 | * @hdr - pageio header to validate against previously seen verf | |
162 | * | |
163 | * set the server's "seen" verf if not initialized. | |
164 | * returns result of comparison between @hdr->verf and the "seen" | |
165 | * verf of the server used by @hdr (DS or MDS) | |
166 | */ | |
167 | static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq, | |
168 | struct nfs_pgio_header *hdr) | |
169 | { | |
170 | struct nfs_writeverf *verfp; | |
171 | ||
172 | verfp = nfs_direct_select_verf(dreq, hdr->data->ds_clp, | |
173 | hdr->data->ds_idx); | |
174 | if (verfp->committed < 0) { | |
175 | nfs_direct_set_hdr_verf(dreq, hdr); | |
176 | return 0; | |
177 | } | |
178 | return memcmp(verfp, &hdr->verf, sizeof(struct nfs_writeverf)); | |
179 | } | |
180 | ||
181 | #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) | |
182 | /* | |
183 | * nfs_direct_cmp_commit_data_verf - compare verifier for commit data | |
184 | * @dreq - direct request possibly spanning multiple servers | |
185 | * @data - commit data to validate against previously seen verf | |
186 | * | |
187 | * returns result of comparison between @data->verf and the verf of | |
188 | * the server used by @data (DS or MDS) | |
189 | */ | |
190 | static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq, | |
191 | struct nfs_commit_data *data) | |
192 | { | |
193 | struct nfs_writeverf *verfp; | |
194 | ||
195 | verfp = nfs_direct_select_verf(dreq, data->ds_clp, | |
196 | data->ds_commit_index); | |
197 | WARN_ON_ONCE(verfp->committed < 0); | |
198 | return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf)); | |
199 | } | |
200 | #endif | |
201 | ||
1da177e4 | 202 | /** |
b8a32e2b CL |
203 | * nfs_direct_IO - NFS address space operation for direct I/O |
204 | * @rw: direction (read or write) | |
205 | * @iocb: target I/O control block | |
206 | * @iov: array of vectors that define I/O buffer | |
207 | * @pos: offset in file to begin the operation | |
208 | * @nr_segs: size of iovec array | |
209 | * | |
210 | * The presence of this routine in the address space ops vector means | |
a564b8f0 MG |
211 | * the NFS client supports direct I/O. However, for most direct IO, we |
212 | * shunt off direct read and write requests before the VFS gets them, | |
213 | * so this method is only ever called for swap. | |
1da177e4 | 214 | */ |
d8d3d94b | 215 | ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, struct iov_iter *iter, loff_t pos) |
b8a32e2b | 216 | { |
a564b8f0 | 217 | #ifndef CONFIG_NFS_SWAP |
6de1472f | 218 | dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n", |
d8d3d94b | 219 | iocb->ki_filp, (long long) pos, iter->nr_segs); |
b8a32e2b CL |
220 | |
221 | return -EINVAL; | |
a564b8f0 | 222 | #else |
a564b8f0 MG |
223 | VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE); |
224 | ||
225 | if (rw == READ || rw == KERNEL_READ) | |
619d30b4 | 226 | return nfs_file_direct_read(iocb, iter, pos, |
a564b8f0 | 227 | rw == READ ? true : false); |
619d30b4 | 228 | return nfs_file_direct_write(iocb, iter, pos, |
a564b8f0 MG |
229 | rw == WRITE ? true : false); |
230 | #endif /* CONFIG_NFS_SWAP */ | |
b8a32e2b CL |
231 | } |
232 | ||
749e146e | 233 | static void nfs_direct_release_pages(struct page **pages, unsigned int npages) |
9c93ab7d | 234 | { |
749e146e | 235 | unsigned int i; |
607f31e8 TM |
236 | for (i = 0; i < npages; i++) |
237 | page_cache_release(pages[i]); | |
6b45d858 TM |
238 | } |
239 | ||
1763da12 FI |
240 | void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo, |
241 | struct nfs_direct_req *dreq) | |
242 | { | |
243 | cinfo->lock = &dreq->lock; | |
244 | cinfo->mds = &dreq->mds_cinfo; | |
245 | cinfo->ds = &dreq->ds_cinfo; | |
246 | cinfo->dreq = dreq; | |
247 | cinfo->completion_ops = &nfs_direct_commit_completion_ops; | |
248 | } | |
249 | ||
93619e59 | 250 | static inline struct nfs_direct_req *nfs_direct_req_alloc(void) |
1da177e4 | 251 | { |
93619e59 CL |
252 | struct nfs_direct_req *dreq; |
253 | ||
292f3eee | 254 | dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL); |
93619e59 CL |
255 | if (!dreq) |
256 | return NULL; | |
257 | ||
258 | kref_init(&dreq->kref); | |
607f31e8 | 259 | kref_get(&dreq->kref); |
d72b7a6b | 260 | init_completion(&dreq->completion); |
1763da12 | 261 | INIT_LIST_HEAD(&dreq->mds_cinfo.list); |
5002c586 | 262 | dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */ |
1763da12 | 263 | INIT_WORK(&dreq->work, nfs_direct_write_schedule_work); |
15ce4a0c | 264 | spin_lock_init(&dreq->lock); |
93619e59 CL |
265 | |
266 | return dreq; | |
1da177e4 LT |
267 | } |
268 | ||
b4946ffb | 269 | static void nfs_direct_req_free(struct kref *kref) |
1da177e4 LT |
270 | { |
271 | struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); | |
a8881f5a | 272 | |
f11ac8db TM |
273 | if (dreq->l_ctx != NULL) |
274 | nfs_put_lock_context(dreq->l_ctx); | |
a8881f5a TM |
275 | if (dreq->ctx != NULL) |
276 | put_nfs_open_context(dreq->ctx); | |
1da177e4 LT |
277 | kmem_cache_free(nfs_direct_cachep, dreq); |
278 | } | |
279 | ||
b4946ffb TM |
280 | static void nfs_direct_req_release(struct nfs_direct_req *dreq) |
281 | { | |
282 | kref_put(&dreq->kref, nfs_direct_req_free); | |
283 | } | |
284 | ||
6296556f PT |
285 | ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq) |
286 | { | |
287 | return dreq->bytes_left; | |
288 | } | |
289 | EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left); | |
290 | ||
bc0fb201 CL |
291 | /* |
292 | * Collects and returns the final error value/byte-count. | |
293 | */ | |
294 | static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) | |
295 | { | |
15ce4a0c | 296 | ssize_t result = -EIOCBQUEUED; |
bc0fb201 CL |
297 | |
298 | /* Async requests don't wait here */ | |
299 | if (dreq->iocb) | |
300 | goto out; | |
301 | ||
150030b7 | 302 | result = wait_for_completion_killable(&dreq->completion); |
bc0fb201 CL |
303 | |
304 | if (!result) | |
15ce4a0c | 305 | result = dreq->error; |
bc0fb201 | 306 | if (!result) |
15ce4a0c | 307 | result = dreq->count; |
bc0fb201 CL |
308 | |
309 | out: | |
bc0fb201 CL |
310 | return (ssize_t) result; |
311 | } | |
312 | ||
63ab46ab | 313 | /* |
607f31e8 TM |
314 | * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust |
315 | * the iocb is still valid here if this is a synchronous request. | |
63ab46ab | 316 | */ |
9811cd57 | 317 | static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write) |
63ab46ab | 318 | { |
9811cd57 CH |
319 | struct inode *inode = dreq->inode; |
320 | ||
2a009ec9 | 321 | if (dreq->iocb && write) { |
9811cd57 | 322 | loff_t pos = dreq->iocb->ki_pos + dreq->count; |
2a009ec9 CH |
323 | |
324 | spin_lock(&inode->i_lock); | |
325 | if (i_size_read(inode) < pos) | |
326 | i_size_write(inode, pos); | |
327 | spin_unlock(&inode->i_lock); | |
328 | } | |
329 | ||
1f90ee27 | 330 | if (write) |
2a009ec9 | 331 | nfs_zap_mapping(inode, inode->i_mapping); |
1f90ee27 CH |
332 | |
333 | inode_dio_done(inode); | |
2a009ec9 CH |
334 | |
335 | if (dreq->iocb) { | |
15ce4a0c | 336 | long res = (long) dreq->error; |
63ab46ab | 337 | if (!res) |
15ce4a0c | 338 | res = (long) dreq->count; |
63ab46ab | 339 | aio_complete(dreq->iocb, res, 0); |
d72b7a6b | 340 | } |
2a009ec9 | 341 | |
d72b7a6b | 342 | complete_all(&dreq->completion); |
63ab46ab | 343 | |
b4946ffb | 344 | nfs_direct_req_release(dreq); |
63ab46ab CL |
345 | } |
346 | ||
1385b811 | 347 | static void nfs_direct_readpage_release(struct nfs_page *req) |
1da177e4 | 348 | { |
1e8968c5 | 349 | dprintk("NFS: direct read done (%s/%llu %d@%lld)\n", |
584aa810 | 350 | req->wb_context->dentry->d_inode->i_sb->s_id, |
1e8968c5 | 351 | (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode), |
584aa810 FI |
352 | req->wb_bytes, |
353 | (long long)req_offset(req)); | |
354 | nfs_release_request(req); | |
fdd1e74c TM |
355 | } |
356 | ||
584aa810 | 357 | static void nfs_direct_read_completion(struct nfs_pgio_header *hdr) |
fdd1e74c | 358 | { |
584aa810 FI |
359 | unsigned long bytes = 0; |
360 | struct nfs_direct_req *dreq = hdr->dreq; | |
fdd1e74c | 361 | |
584aa810 FI |
362 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) |
363 | goto out_put; | |
15ce4a0c CL |
364 | |
365 | spin_lock(&dreq->lock); | |
584aa810 FI |
366 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0)) |
367 | dreq->error = hdr->error; | |
368 | else | |
369 | dreq->count += hdr->good_bytes; | |
370 | spin_unlock(&dreq->lock); | |
371 | ||
4bd8b010 TM |
372 | while (!list_empty(&hdr->pages)) { |
373 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
374 | struct page *page = req->wb_page; | |
375 | ||
be7e9858 JL |
376 | if (!PageCompound(page) && bytes < hdr->good_bytes) |
377 | set_page_dirty(page); | |
4bd8b010 TM |
378 | bytes += req->wb_bytes; |
379 | nfs_list_remove_request(req); | |
380 | nfs_direct_readpage_release(req); | |
d4a8f367 | 381 | } |
584aa810 | 382 | out_put: |
607f31e8 | 383 | if (put_dreq(dreq)) |
9811cd57 | 384 | nfs_direct_complete(dreq, false); |
584aa810 | 385 | hdr->release(hdr); |
1da177e4 LT |
386 | } |
387 | ||
3e9e0ca3 | 388 | static void nfs_read_sync_pgio_error(struct list_head *head) |
cd841605 | 389 | { |
584aa810 | 390 | struct nfs_page *req; |
cd841605 | 391 | |
584aa810 FI |
392 | while (!list_empty(head)) { |
393 | req = nfs_list_entry(head->next); | |
394 | nfs_list_remove_request(req); | |
395 | nfs_release_request(req); | |
396 | } | |
cd841605 FI |
397 | } |
398 | ||
584aa810 FI |
399 | static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr) |
400 | { | |
401 | get_dreq(hdr->dreq); | |
402 | } | |
403 | ||
404 | static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = { | |
3e9e0ca3 | 405 | .error_cleanup = nfs_read_sync_pgio_error, |
584aa810 FI |
406 | .init_hdr = nfs_direct_pgio_init, |
407 | .completion = nfs_direct_read_completion, | |
408 | }; | |
409 | ||
d4cc948b | 410 | /* |
607f31e8 TM |
411 | * For each rsize'd chunk of the user's buffer, dispatch an NFS READ |
412 | * operation. If nfs_readdata_alloc() or get_user_pages() fails, | |
413 | * bail and stop sending more reads. Read length accounting is | |
414 | * handled automatically by nfs_direct_read_result(). Otherwise, if | |
415 | * no requests have been sent, just return an error. | |
1da177e4 | 416 | */ |
1da177e4 | 417 | |
91f79c43 AV |
418 | static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq, |
419 | struct iov_iter *iter, | |
420 | loff_t pos) | |
421 | { | |
422 | struct nfs_pageio_descriptor desc; | |
423 | struct inode *inode = dreq->inode; | |
424 | ssize_t result = -EINVAL; | |
425 | size_t requested_bytes = 0; | |
426 | size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE); | |
e9f7bee1 | 427 | |
16b90578 | 428 | nfs_pageio_init_read(&desc, dreq->inode, false, |
91f79c43 AV |
429 | &nfs_direct_read_completion_ops); |
430 | get_dreq(dreq); | |
431 | desc.pg_dreq = dreq; | |
432 | atomic_inc(&inode->i_dio_count); | |
a564b8f0 | 433 | |
91f79c43 AV |
434 | while (iov_iter_count(iter)) { |
435 | struct page **pagevec; | |
436 | size_t bytes; | |
437 | size_t pgbase; | |
438 | unsigned npages, i; | |
607f31e8 | 439 | |
91f79c43 AV |
440 | result = iov_iter_get_pages_alloc(iter, &pagevec, |
441 | rsize, &pgbase); | |
442 | if (result < 0) | |
443 | break; | |
444 | ||
445 | bytes = result; | |
446 | iov_iter_advance(iter, bytes); | |
447 | npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; | |
584aa810 FI |
448 | for (i = 0; i < npages; i++) { |
449 | struct nfs_page *req; | |
bf5fc402 | 450 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); |
584aa810 | 451 | /* XXX do we need to do the eof zeroing found in async_filler? */ |
2bfc6e56 | 452 | req = nfs_create_request(dreq->ctx, pagevec[i], NULL, |
584aa810 FI |
453 | pgbase, req_len); |
454 | if (IS_ERR(req)) { | |
584aa810 FI |
455 | result = PTR_ERR(req); |
456 | break; | |
457 | } | |
458 | req->wb_index = pos >> PAGE_SHIFT; | |
459 | req->wb_offset = pos & ~PAGE_MASK; | |
91f79c43 AV |
460 | if (!nfs_pageio_add_request(&desc, req)) { |
461 | result = desc.pg_error; | |
584aa810 | 462 | nfs_release_request(req); |
584aa810 FI |
463 | break; |
464 | } | |
465 | pgbase = 0; | |
466 | bytes -= req_len; | |
91f79c43 | 467 | requested_bytes += req_len; |
584aa810 | 468 | pos += req_len; |
35754bc0 | 469 | dreq->bytes_left -= req_len; |
584aa810 | 470 | } |
6d74743b | 471 | nfs_direct_release_pages(pagevec, npages); |
91f79c43 | 472 | kvfree(pagevec); |
19f73787 CL |
473 | if (result < 0) |
474 | break; | |
19f73787 CL |
475 | } |
476 | ||
584aa810 FI |
477 | nfs_pageio_complete(&desc); |
478 | ||
839f7ad6 CL |
479 | /* |
480 | * If no bytes were started, return the error, and let the | |
481 | * generic layer handle the completion. | |
482 | */ | |
483 | if (requested_bytes == 0) { | |
1f90ee27 | 484 | inode_dio_done(inode); |
839f7ad6 CL |
485 | nfs_direct_req_release(dreq); |
486 | return result < 0 ? result : -EIO; | |
487 | } | |
488 | ||
19f73787 | 489 | if (put_dreq(dreq)) |
9811cd57 | 490 | nfs_direct_complete(dreq, false); |
839f7ad6 | 491 | return 0; |
19f73787 CL |
492 | } |
493 | ||
14a3ec79 CH |
494 | /** |
495 | * nfs_file_direct_read - file direct read operation for NFS files | |
496 | * @iocb: target I/O control block | |
619d30b4 | 497 | * @iter: vector of user buffers into which to read data |
14a3ec79 CH |
498 | * @pos: byte offset in file where reading starts |
499 | * | |
500 | * We use this function for direct reads instead of calling | |
501 | * generic_file_aio_read() in order to avoid gfar's check to see if | |
502 | * the request starts before the end of the file. For that check | |
503 | * to work, we must generate a GETATTR before each direct read, and | |
504 | * even then there is a window between the GETATTR and the subsequent | |
505 | * READ where the file size could change. Our preference is simply | |
506 | * to do all reads the application wants, and the server will take | |
507 | * care of managing the end of file boundary. | |
508 | * | |
509 | * This function also eliminates unnecessarily updating the file's | |
510 | * atime locally, as the NFS server sets the file's atime, and this | |
511 | * client must read the updated atime from the server back into its | |
512 | * cache. | |
513 | */ | |
619d30b4 AV |
514 | ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter, |
515 | loff_t pos, bool uio) | |
1da177e4 | 516 | { |
14a3ec79 CH |
517 | struct file *file = iocb->ki_filp; |
518 | struct address_space *mapping = file->f_mapping; | |
519 | struct inode *inode = mapping->host; | |
1da177e4 | 520 | struct nfs_direct_req *dreq; |
b3c54de6 | 521 | struct nfs_lock_context *l_ctx; |
14a3ec79 | 522 | ssize_t result = -EINVAL; |
a6cbcd4a | 523 | size_t count = iov_iter_count(iter); |
14a3ec79 CH |
524 | nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); |
525 | ||
526 | dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n", | |
527 | file, count, (long long) pos); | |
528 | ||
529 | result = 0; | |
530 | if (!count) | |
531 | goto out; | |
532 | ||
d0b9875d | 533 | mutex_lock(&inode->i_mutex); |
14a3ec79 CH |
534 | result = nfs_sync_mapping(mapping); |
535 | if (result) | |
d0b9875d | 536 | goto out_unlock; |
1da177e4 | 537 | |
14a3ec79 CH |
538 | task_io_account_read(count); |
539 | ||
540 | result = -ENOMEM; | |
607f31e8 | 541 | dreq = nfs_direct_req_alloc(); |
f11ac8db | 542 | if (dreq == NULL) |
d0b9875d | 543 | goto out_unlock; |
1da177e4 | 544 | |
91d5b470 | 545 | dreq->inode = inode; |
619d30b4 | 546 | dreq->bytes_left = count; |
cd3758e3 | 547 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
b3c54de6 TM |
548 | l_ctx = nfs_get_lock_context(dreq->ctx); |
549 | if (IS_ERR(l_ctx)) { | |
550 | result = PTR_ERR(l_ctx); | |
f11ac8db | 551 | goto out_release; |
b3c54de6 TM |
552 | } |
553 | dreq->l_ctx = l_ctx; | |
487b8372 CL |
554 | if (!is_sync_kiocb(iocb)) |
555 | dreq->iocb = iocb; | |
1da177e4 | 556 | |
619d30b4 | 557 | NFS_I(inode)->read_io += count; |
91f79c43 | 558 | result = nfs_direct_read_schedule_iovec(dreq, iter, pos); |
d0b9875d CH |
559 | |
560 | mutex_unlock(&inode->i_mutex); | |
561 | ||
14a3ec79 | 562 | if (!result) { |
607f31e8 | 563 | result = nfs_direct_wait(dreq); |
14a3ec79 CH |
564 | if (result > 0) |
565 | iocb->ki_pos = pos + result; | |
566 | } | |
d0b9875d CH |
567 | |
568 | nfs_direct_req_release(dreq); | |
569 | return result; | |
570 | ||
f11ac8db | 571 | out_release: |
b4946ffb | 572 | nfs_direct_req_release(dreq); |
d0b9875d CH |
573 | out_unlock: |
574 | mutex_unlock(&inode->i_mutex); | |
f11ac8db | 575 | out: |
1da177e4 LT |
576 | return result; |
577 | } | |
578 | ||
89d77c8f | 579 | #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) |
fad61490 TM |
580 | static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) |
581 | { | |
1763da12 FI |
582 | struct nfs_pageio_descriptor desc; |
583 | struct nfs_page *req, *tmp; | |
584 | LIST_HEAD(reqs); | |
585 | struct nfs_commit_info cinfo; | |
586 | LIST_HEAD(failed); | |
587 | ||
588 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
589 | pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo); | |
590 | spin_lock(cinfo.lock); | |
591 | nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0); | |
592 | spin_unlock(cinfo.lock); | |
1da177e4 | 593 | |
fad61490 | 594 | dreq->count = 0; |
607f31e8 TM |
595 | get_dreq(dreq); |
596 | ||
a20c93e3 | 597 | nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false, |
1763da12 FI |
598 | &nfs_direct_write_completion_ops); |
599 | desc.pg_dreq = dreq; | |
fedb595c | 600 | |
1763da12 FI |
601 | list_for_each_entry_safe(req, tmp, &reqs, wb_list) { |
602 | if (!nfs_pageio_add_request(&desc, req)) { | |
4035c248 | 603 | nfs_list_remove_request(req); |
1763da12 FI |
604 | nfs_list_add_request(req, &failed); |
605 | spin_lock(cinfo.lock); | |
606 | dreq->flags = 0; | |
607 | dreq->error = -EIO; | |
608 | spin_unlock(cinfo.lock); | |
609 | } | |
5a695da2 | 610 | nfs_release_request(req); |
1763da12 FI |
611 | } |
612 | nfs_pageio_complete(&desc); | |
fad61490 | 613 | |
4035c248 TM |
614 | while (!list_empty(&failed)) { |
615 | req = nfs_list_entry(failed.next); | |
616 | nfs_list_remove_request(req); | |
1d1afcbc | 617 | nfs_unlock_and_release_request(req); |
4035c248 | 618 | } |
fad61490 | 619 | |
1763da12 FI |
620 | if (put_dreq(dreq)) |
621 | nfs_direct_write_complete(dreq, dreq->inode); | |
c9d8f89d TM |
622 | } |
623 | ||
1763da12 | 624 | static void nfs_direct_commit_complete(struct nfs_commit_data *data) |
c9d8f89d | 625 | { |
0b7c0153 | 626 | struct nfs_direct_req *dreq = data->dreq; |
1763da12 FI |
627 | struct nfs_commit_info cinfo; |
628 | struct nfs_page *req; | |
c9d8f89d TM |
629 | int status = data->task.tk_status; |
630 | ||
1763da12 | 631 | nfs_init_cinfo_from_dreq(&cinfo, dreq); |
c9d8f89d | 632 | if (status < 0) { |
60fa3f76 | 633 | dprintk("NFS: %5u commit failed with error %d.\n", |
1763da12 | 634 | data->task.tk_pid, status); |
fad61490 | 635 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
5002c586 | 636 | } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) { |
c9d8f89d | 637 | dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid); |
fad61490 | 638 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
1da177e4 LT |
639 | } |
640 | ||
c9d8f89d | 641 | dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status); |
1763da12 FI |
642 | while (!list_empty(&data->pages)) { |
643 | req = nfs_list_entry(data->pages.next); | |
644 | nfs_list_remove_request(req); | |
645 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) { | |
646 | /* Note the rewrite will go through mds */ | |
647 | nfs_mark_request_commit(req, NULL, &cinfo); | |
906369e4 FI |
648 | } else |
649 | nfs_release_request(req); | |
1d1afcbc | 650 | nfs_unlock_and_release_request(req); |
1763da12 FI |
651 | } |
652 | ||
653 | if (atomic_dec_and_test(&cinfo.mds->rpcs_out)) | |
654 | nfs_direct_write_complete(dreq, data->inode); | |
1da177e4 LT |
655 | } |
656 | ||
1763da12 FI |
657 | static void nfs_direct_error_cleanup(struct nfs_inode *nfsi) |
658 | { | |
659 | /* There is no lock to clear */ | |
660 | } | |
661 | ||
662 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = { | |
663 | .completion = nfs_direct_commit_complete, | |
664 | .error_cleanup = nfs_direct_error_cleanup, | |
fad61490 TM |
665 | }; |
666 | ||
667 | static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) | |
1da177e4 | 668 | { |
1763da12 FI |
669 | int res; |
670 | struct nfs_commit_info cinfo; | |
671 | LIST_HEAD(mds_list); | |
672 | ||
673 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
674 | nfs_scan_commit(dreq->inode, &mds_list, &cinfo); | |
675 | res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo); | |
676 | if (res < 0) /* res == -ENOMEM */ | |
677 | nfs_direct_write_reschedule(dreq); | |
fad61490 | 678 | } |
1da177e4 | 679 | |
1763da12 | 680 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
fad61490 | 681 | { |
1763da12 | 682 | struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work); |
fad61490 | 683 | int flags = dreq->flags; |
1da177e4 | 684 | |
fad61490 TM |
685 | dreq->flags = 0; |
686 | switch (flags) { | |
687 | case NFS_ODIRECT_DO_COMMIT: | |
688 | nfs_direct_commit_schedule(dreq); | |
1da177e4 | 689 | break; |
fad61490 TM |
690 | case NFS_ODIRECT_RESCHED_WRITES: |
691 | nfs_direct_write_reschedule(dreq); | |
692 | break; | |
693 | default: | |
9811cd57 | 694 | nfs_direct_complete(dreq, true); |
fad61490 TM |
695 | } |
696 | } | |
1da177e4 | 697 | |
1763da12 | 698 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
fad61490 | 699 | { |
1763da12 | 700 | schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */ |
fad61490 | 701 | } |
1763da12 | 702 | |
fad61490 | 703 | #else |
24fc9211 BS |
704 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
705 | { | |
706 | } | |
1da177e4 | 707 | |
fad61490 TM |
708 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
709 | { | |
9811cd57 | 710 | nfs_direct_complete(dreq, true); |
fad61490 TM |
711 | } |
712 | #endif | |
1da177e4 | 713 | |
1763da12 FI |
714 | static void nfs_direct_write_completion(struct nfs_pgio_header *hdr) |
715 | { | |
716 | struct nfs_direct_req *dreq = hdr->dreq; | |
717 | struct nfs_commit_info cinfo; | |
718 | int bit = -1; | |
719 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
720 | ||
721 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) | |
722 | goto out_put; | |
723 | ||
724 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
725 | ||
726 | spin_lock(&dreq->lock); | |
727 | ||
728 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) { | |
729 | dreq->flags = 0; | |
730 | dreq->error = hdr->error; | |
731 | } | |
732 | if (dreq->error != 0) | |
733 | bit = NFS_IOHDR_ERROR; | |
734 | else { | |
735 | dreq->count += hdr->good_bytes; | |
736 | if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) { | |
737 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; | |
738 | bit = NFS_IOHDR_NEED_RESCHED; | |
739 | } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) { | |
740 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) | |
741 | bit = NFS_IOHDR_NEED_RESCHED; | |
742 | else if (dreq->flags == 0) { | |
5002c586 | 743 | nfs_direct_set_hdr_verf(dreq, hdr); |
1763da12 FI |
744 | bit = NFS_IOHDR_NEED_COMMIT; |
745 | dreq->flags = NFS_ODIRECT_DO_COMMIT; | |
746 | } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) { | |
5002c586 WAA |
747 | if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr)) { |
748 | dreq->flags = | |
749 | NFS_ODIRECT_RESCHED_WRITES; | |
1763da12 FI |
750 | bit = NFS_IOHDR_NEED_RESCHED; |
751 | } else | |
752 | bit = NFS_IOHDR_NEED_COMMIT; | |
753 | } | |
754 | } | |
755 | } | |
756 | spin_unlock(&dreq->lock); | |
757 | ||
758 | while (!list_empty(&hdr->pages)) { | |
2bfc6e56 WAA |
759 | bool do_destroy = true; |
760 | ||
1763da12 FI |
761 | req = nfs_list_entry(hdr->pages.next); |
762 | nfs_list_remove_request(req); | |
763 | switch (bit) { | |
764 | case NFS_IOHDR_NEED_RESCHED: | |
765 | case NFS_IOHDR_NEED_COMMIT: | |
04277086 | 766 | kref_get(&req->wb_kref); |
1763da12 | 767 | nfs_mark_request_commit(req, hdr->lseg, &cinfo); |
2bfc6e56 | 768 | do_destroy = false; |
1763da12 | 769 | } |
1d1afcbc | 770 | nfs_unlock_and_release_request(req); |
1763da12 FI |
771 | } |
772 | ||
773 | out_put: | |
774 | if (put_dreq(dreq)) | |
775 | nfs_direct_write_complete(dreq, hdr->inode); | |
776 | hdr->release(hdr); | |
777 | } | |
778 | ||
3e9e0ca3 TM |
779 | static void nfs_write_sync_pgio_error(struct list_head *head) |
780 | { | |
781 | struct nfs_page *req; | |
782 | ||
783 | while (!list_empty(head)) { | |
784 | req = nfs_list_entry(head->next); | |
785 | nfs_list_remove_request(req); | |
1d1afcbc | 786 | nfs_unlock_and_release_request(req); |
3e9e0ca3 TM |
787 | } |
788 | } | |
789 | ||
1763da12 | 790 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = { |
3e9e0ca3 | 791 | .error_cleanup = nfs_write_sync_pgio_error, |
1763da12 FI |
792 | .init_hdr = nfs_direct_pgio_init, |
793 | .completion = nfs_direct_write_completion, | |
794 | }; | |
795 | ||
91f79c43 AV |
796 | |
797 | /* | |
798 | * NB: Return the value of the first error return code. Subsequent | |
799 | * errors after the first one are ignored. | |
800 | */ | |
801 | /* | |
802 | * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE | |
803 | * operation. If nfs_writedata_alloc() or get_user_pages() fails, | |
804 | * bail and stop sending more writes. Write length accounting is | |
805 | * handled automatically by nfs_direct_write_result(). Otherwise, if | |
806 | * no requests have been sent, just return an error. | |
807 | */ | |
19f73787 | 808 | static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, |
619d30b4 | 809 | struct iov_iter *iter, |
91f79c43 | 810 | loff_t pos) |
19f73787 | 811 | { |
1763da12 | 812 | struct nfs_pageio_descriptor desc; |
1d59d61f | 813 | struct inode *inode = dreq->inode; |
19f73787 CL |
814 | ssize_t result = 0; |
815 | size_t requested_bytes = 0; | |
91f79c43 | 816 | size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE); |
19f73787 | 817 | |
a20c93e3 | 818 | nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false, |
1763da12 FI |
819 | &nfs_direct_write_completion_ops); |
820 | desc.pg_dreq = dreq; | |
19f73787 | 821 | get_dreq(dreq); |
1d59d61f | 822 | atomic_inc(&inode->i_dio_count); |
19f73787 | 823 | |
91f79c43 AV |
824 | NFS_I(inode)->write_io += iov_iter_count(iter); |
825 | while (iov_iter_count(iter)) { | |
826 | struct page **pagevec; | |
827 | size_t bytes; | |
828 | size_t pgbase; | |
829 | unsigned npages, i; | |
830 | ||
831 | result = iov_iter_get_pages_alloc(iter, &pagevec, | |
832 | wsize, &pgbase); | |
19f73787 CL |
833 | if (result < 0) |
834 | break; | |
91f79c43 AV |
835 | |
836 | bytes = result; | |
837 | iov_iter_advance(iter, bytes); | |
838 | npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; | |
839 | for (i = 0; i < npages; i++) { | |
840 | struct nfs_page *req; | |
841 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); | |
842 | ||
16b90578 | 843 | req = nfs_create_request(dreq->ctx, pagevec[i], NULL, |
91f79c43 AV |
844 | pgbase, req_len); |
845 | if (IS_ERR(req)) { | |
846 | result = PTR_ERR(req); | |
847 | break; | |
848 | } | |
849 | nfs_lock_request(req); | |
850 | req->wb_index = pos >> PAGE_SHIFT; | |
851 | req->wb_offset = pos & ~PAGE_MASK; | |
852 | if (!nfs_pageio_add_request(&desc, req)) { | |
853 | result = desc.pg_error; | |
854 | nfs_unlock_and_release_request(req); | |
855 | break; | |
856 | } | |
857 | pgbase = 0; | |
858 | bytes -= req_len; | |
859 | requested_bytes += req_len; | |
860 | pos += req_len; | |
861 | dreq->bytes_left -= req_len; | |
862 | } | |
863 | nfs_direct_release_pages(pagevec, npages); | |
864 | kvfree(pagevec); | |
865 | if (result < 0) | |
19f73787 | 866 | break; |
19f73787 | 867 | } |
1763da12 | 868 | nfs_pageio_complete(&desc); |
19f73787 | 869 | |
839f7ad6 CL |
870 | /* |
871 | * If no bytes were started, return the error, and let the | |
872 | * generic layer handle the completion. | |
873 | */ | |
874 | if (requested_bytes == 0) { | |
1d59d61f | 875 | inode_dio_done(inode); |
839f7ad6 CL |
876 | nfs_direct_req_release(dreq); |
877 | return result < 0 ? result : -EIO; | |
878 | } | |
879 | ||
19f73787 CL |
880 | if (put_dreq(dreq)) |
881 | nfs_direct_write_complete(dreq, dreq->inode); | |
839f7ad6 | 882 | return 0; |
19f73787 CL |
883 | } |
884 | ||
1da177e4 LT |
885 | /** |
886 | * nfs_file_direct_write - file direct write operation for NFS files | |
887 | * @iocb: target I/O control block | |
619d30b4 | 888 | * @iter: vector of user buffers from which to write data |
88467055 | 889 | * @pos: byte offset in file where writing starts |
1da177e4 LT |
890 | * |
891 | * We use this function for direct writes instead of calling | |
892 | * generic_file_aio_write() in order to avoid taking the inode | |
893 | * semaphore and updating the i_size. The NFS server will set | |
894 | * the new i_size and this client must read the updated size | |
895 | * back into its cache. We let the server do generic write | |
896 | * parameter checking and report problems. | |
897 | * | |
1da177e4 LT |
898 | * We eliminate local atime updates, see direct read above. |
899 | * | |
900 | * We avoid unnecessary page cache invalidations for normal cached | |
901 | * readers of this file. | |
902 | * | |
903 | * Note that O_APPEND is not supported for NFS direct writes, as there | |
904 | * is no atomic O_APPEND write facility in the NFS protocol. | |
905 | */ | |
619d30b4 AV |
906 | ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter, |
907 | loff_t pos, bool uio) | |
1da177e4 | 908 | { |
22cd1bf1 | 909 | ssize_t result = -EINVAL; |
1da177e4 | 910 | struct file *file = iocb->ki_filp; |
1da177e4 | 911 | struct address_space *mapping = file->f_mapping; |
22cd1bf1 CH |
912 | struct inode *inode = mapping->host; |
913 | struct nfs_direct_req *dreq; | |
914 | struct nfs_lock_context *l_ctx; | |
a9ab5e84 | 915 | loff_t end; |
a6cbcd4a | 916 | size_t count = iov_iter_count(iter); |
a9ab5e84 CH |
917 | end = (pos + count - 1) >> PAGE_CACHE_SHIFT; |
918 | ||
c216fd70 CL |
919 | nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count); |
920 | ||
6de1472f AV |
921 | dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n", |
922 | file, count, (long long) pos); | |
027445c3 | 923 | |
22cd1bf1 CH |
924 | result = generic_write_checks(file, &pos, &count, 0); |
925 | if (result) | |
1da177e4 | 926 | goto out; |
ce1a8e67 | 927 | |
22cd1bf1 | 928 | result = -EINVAL; |
ce1a8e67 | 929 | if ((ssize_t) count < 0) |
1da177e4 | 930 | goto out; |
22cd1bf1 | 931 | result = 0; |
1da177e4 LT |
932 | if (!count) |
933 | goto out; | |
ce1a8e67 | 934 | |
a9ab5e84 CH |
935 | mutex_lock(&inode->i_mutex); |
936 | ||
22cd1bf1 CH |
937 | result = nfs_sync_mapping(mapping); |
938 | if (result) | |
a9ab5e84 CH |
939 | goto out_unlock; |
940 | ||
941 | if (mapping->nrpages) { | |
942 | result = invalidate_inode_pages2_range(mapping, | |
943 | pos >> PAGE_CACHE_SHIFT, end); | |
944 | if (result) | |
945 | goto out_unlock; | |
946 | } | |
1da177e4 | 947 | |
7ec10f26 KK |
948 | task_io_account_write(count); |
949 | ||
22cd1bf1 CH |
950 | result = -ENOMEM; |
951 | dreq = nfs_direct_req_alloc(); | |
952 | if (!dreq) | |
a9ab5e84 | 953 | goto out_unlock; |
9eafa8cc | 954 | |
22cd1bf1 CH |
955 | dreq->inode = inode; |
956 | dreq->bytes_left = count; | |
957 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); | |
958 | l_ctx = nfs_get_lock_context(dreq->ctx); | |
959 | if (IS_ERR(l_ctx)) { | |
960 | result = PTR_ERR(l_ctx); | |
961 | goto out_release; | |
962 | } | |
963 | dreq->l_ctx = l_ctx; | |
964 | if (!is_sync_kiocb(iocb)) | |
965 | dreq->iocb = iocb; | |
966 | ||
91f79c43 | 967 | result = nfs_direct_write_schedule_iovec(dreq, iter, pos); |
a9ab5e84 CH |
968 | |
969 | if (mapping->nrpages) { | |
970 | invalidate_inode_pages2_range(mapping, | |
971 | pos >> PAGE_CACHE_SHIFT, end); | |
972 | } | |
973 | ||
974 | mutex_unlock(&inode->i_mutex); | |
975 | ||
22cd1bf1 CH |
976 | if (!result) { |
977 | result = nfs_direct_wait(dreq); | |
978 | if (result > 0) { | |
979 | struct inode *inode = mapping->host; | |
980 | ||
981 | iocb->ki_pos = pos + result; | |
982 | spin_lock(&inode->i_lock); | |
983 | if (i_size_read(inode) < iocb->ki_pos) | |
984 | i_size_write(inode, iocb->ki_pos); | |
985 | spin_unlock(&inode->i_lock); | |
986 | } | |
1763da12 | 987 | } |
a9ab5e84 CH |
988 | nfs_direct_req_release(dreq); |
989 | return result; | |
990 | ||
22cd1bf1 CH |
991 | out_release: |
992 | nfs_direct_req_release(dreq); | |
a9ab5e84 CH |
993 | out_unlock: |
994 | mutex_unlock(&inode->i_mutex); | |
1da177e4 | 995 | out: |
22cd1bf1 | 996 | return result; |
1da177e4 LT |
997 | } |
998 | ||
88467055 CL |
999 | /** |
1000 | * nfs_init_directcache - create a slab cache for nfs_direct_req structures | |
1001 | * | |
1002 | */ | |
f7b422b1 | 1003 | int __init nfs_init_directcache(void) |
1da177e4 LT |
1004 | { |
1005 | nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", | |
1006 | sizeof(struct nfs_direct_req), | |
fffb60f9 PJ |
1007 | 0, (SLAB_RECLAIM_ACCOUNT| |
1008 | SLAB_MEM_SPREAD), | |
20c2df83 | 1009 | NULL); |
1da177e4 LT |
1010 | if (nfs_direct_cachep == NULL) |
1011 | return -ENOMEM; | |
1012 | ||
1013 | return 0; | |
1014 | } | |
1015 | ||
88467055 | 1016 | /** |
f7b422b1 | 1017 | * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures |
88467055 CL |
1018 | * |
1019 | */ | |
266bee88 | 1020 | void nfs_destroy_directcache(void) |
1da177e4 | 1021 | { |
1a1d92c1 | 1022 | kmem_cache_destroy(nfs_direct_cachep); |
1da177e4 | 1023 | } |