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> |
1da177e4 LT |
49 | |
50 | #include <linux/nfs_fs.h> | |
51 | #include <linux/nfs_page.h> | |
52 | #include <linux/sunrpc/clnt.h> | |
53 | ||
1da177e4 | 54 | #include <asm/uaccess.h> |
60063497 | 55 | #include <linux/atomic.h> |
1da177e4 | 56 | |
8d5658c9 | 57 | #include "internal.h" |
91d5b470 CL |
58 | #include "iostat.h" |
59 | ||
1da177e4 | 60 | #define NFSDBG_FACILITY NFSDBG_VFS |
1da177e4 | 61 | |
e18b890b | 62 | static struct kmem_cache *nfs_direct_cachep; |
1da177e4 LT |
63 | |
64 | /* | |
65 | * This represents a set of asynchronous requests that we're waiting on | |
66 | */ | |
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 */ |
15ce4a0c | 79 | ssize_t count, /* bytes actually processed */ |
1da177e4 | 80 | error; /* any reported error */ |
d72b7a6b | 81 | struct completion completion; /* wait for i/o completion */ |
fad61490 TM |
82 | |
83 | /* commit state */ | |
607f31e8 | 84 | struct list_head rewrite_list; /* saved nfs_write_data structs */ |
fad61490 TM |
85 | struct nfs_write_data * commit_data; /* special write_data for commits */ |
86 | int flags; | |
87 | #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */ | |
88 | #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */ | |
89 | struct nfs_writeverf verf; /* unstable write verifier */ | |
1da177e4 LT |
90 | }; |
91 | ||
fad61490 | 92 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode); |
607f31e8 TM |
93 | static const struct rpc_call_ops nfs_write_direct_ops; |
94 | ||
95 | static inline void get_dreq(struct nfs_direct_req *dreq) | |
96 | { | |
97 | atomic_inc(&dreq->io_count); | |
98 | } | |
99 | ||
100 | static inline int put_dreq(struct nfs_direct_req *dreq) | |
101 | { | |
102 | return atomic_dec_and_test(&dreq->io_count); | |
103 | } | |
104 | ||
1da177e4 | 105 | /** |
b8a32e2b CL |
106 | * nfs_direct_IO - NFS address space operation for direct I/O |
107 | * @rw: direction (read or write) | |
108 | * @iocb: target I/O control block | |
109 | * @iov: array of vectors that define I/O buffer | |
110 | * @pos: offset in file to begin the operation | |
111 | * @nr_segs: size of iovec array | |
112 | * | |
113 | * The presence of this routine in the address space ops vector means | |
114 | * the NFS client supports direct I/O. However, we shunt off direct | |
115 | * read and write requests before the VFS gets them, so this method | |
116 | * should never be called. | |
1da177e4 | 117 | */ |
b8a32e2b CL |
118 | ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs) |
119 | { | |
b8a32e2b | 120 | dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n", |
01cce933 | 121 | iocb->ki_filp->f_path.dentry->d_name.name, |
e99170ff | 122 | (long long) pos, nr_segs); |
b8a32e2b CL |
123 | |
124 | return -EINVAL; | |
125 | } | |
126 | ||
d4a8f367 | 127 | static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count) |
6b45d858 | 128 | { |
d4a8f367 | 129 | unsigned int npages; |
749e146e | 130 | unsigned int i; |
d4a8f367 TM |
131 | |
132 | if (count == 0) | |
133 | return; | |
134 | pages += (pgbase >> PAGE_SHIFT); | |
135 | npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
6b45d858 TM |
136 | for (i = 0; i < npages; i++) { |
137 | struct page *page = pages[i]; | |
607f31e8 | 138 | if (!PageCompound(page)) |
d4a8f367 | 139 | set_page_dirty(page); |
6b45d858 | 140 | } |
9c93ab7d CL |
141 | } |
142 | ||
749e146e | 143 | static void nfs_direct_release_pages(struct page **pages, unsigned int npages) |
9c93ab7d | 144 | { |
749e146e | 145 | unsigned int i; |
607f31e8 TM |
146 | for (i = 0; i < npages; i++) |
147 | page_cache_release(pages[i]); | |
6b45d858 TM |
148 | } |
149 | ||
93619e59 | 150 | static inline struct nfs_direct_req *nfs_direct_req_alloc(void) |
1da177e4 | 151 | { |
93619e59 CL |
152 | struct nfs_direct_req *dreq; |
153 | ||
e94b1766 | 154 | dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL); |
93619e59 CL |
155 | if (!dreq) |
156 | return NULL; | |
157 | ||
158 | kref_init(&dreq->kref); | |
607f31e8 | 159 | kref_get(&dreq->kref); |
d72b7a6b | 160 | init_completion(&dreq->completion); |
fad61490 | 161 | INIT_LIST_HEAD(&dreq->rewrite_list); |
93619e59 | 162 | dreq->iocb = NULL; |
a8881f5a | 163 | dreq->ctx = NULL; |
f11ac8db | 164 | dreq->l_ctx = NULL; |
15ce4a0c | 165 | spin_lock_init(&dreq->lock); |
607f31e8 | 166 | atomic_set(&dreq->io_count, 0); |
15ce4a0c CL |
167 | dreq->count = 0; |
168 | dreq->error = 0; | |
fad61490 | 169 | dreq->flags = 0; |
93619e59 CL |
170 | |
171 | return dreq; | |
1da177e4 LT |
172 | } |
173 | ||
b4946ffb | 174 | static void nfs_direct_req_free(struct kref *kref) |
1da177e4 LT |
175 | { |
176 | struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); | |
a8881f5a | 177 | |
f11ac8db TM |
178 | if (dreq->l_ctx != NULL) |
179 | nfs_put_lock_context(dreq->l_ctx); | |
a8881f5a TM |
180 | if (dreq->ctx != NULL) |
181 | put_nfs_open_context(dreq->ctx); | |
1da177e4 LT |
182 | kmem_cache_free(nfs_direct_cachep, dreq); |
183 | } | |
184 | ||
b4946ffb TM |
185 | static void nfs_direct_req_release(struct nfs_direct_req *dreq) |
186 | { | |
187 | kref_put(&dreq->kref, nfs_direct_req_free); | |
188 | } | |
189 | ||
bc0fb201 CL |
190 | /* |
191 | * Collects and returns the final error value/byte-count. | |
192 | */ | |
193 | static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) | |
194 | { | |
15ce4a0c | 195 | ssize_t result = -EIOCBQUEUED; |
bc0fb201 CL |
196 | |
197 | /* Async requests don't wait here */ | |
198 | if (dreq->iocb) | |
199 | goto out; | |
200 | ||
150030b7 | 201 | result = wait_for_completion_killable(&dreq->completion); |
bc0fb201 CL |
202 | |
203 | if (!result) | |
15ce4a0c | 204 | result = dreq->error; |
bc0fb201 | 205 | if (!result) |
15ce4a0c | 206 | result = dreq->count; |
bc0fb201 CL |
207 | |
208 | out: | |
bc0fb201 CL |
209 | return (ssize_t) result; |
210 | } | |
211 | ||
63ab46ab | 212 | /* |
607f31e8 TM |
213 | * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust |
214 | * the iocb is still valid here if this is a synchronous request. | |
63ab46ab CL |
215 | */ |
216 | static void nfs_direct_complete(struct nfs_direct_req *dreq) | |
217 | { | |
63ab46ab | 218 | if (dreq->iocb) { |
15ce4a0c | 219 | long res = (long) dreq->error; |
63ab46ab | 220 | if (!res) |
15ce4a0c | 221 | res = (long) dreq->count; |
63ab46ab | 222 | aio_complete(dreq->iocb, res, 0); |
d72b7a6b TM |
223 | } |
224 | complete_all(&dreq->completion); | |
63ab46ab | 225 | |
b4946ffb | 226 | nfs_direct_req_release(dreq); |
63ab46ab CL |
227 | } |
228 | ||
06cf6f2e | 229 | /* |
607f31e8 TM |
230 | * We must hold a reference to all the pages in this direct read request |
231 | * until the RPCs complete. This could be long *after* we are woken up in | |
232 | * nfs_direct_wait (for instance, if someone hits ^C on a slow server). | |
06cf6f2e | 233 | */ |
ec06c096 | 234 | static void nfs_direct_read_result(struct rpc_task *task, void *calldata) |
1da177e4 | 235 | { |
ec06c096 | 236 | struct nfs_read_data *data = calldata; |
1da177e4 | 237 | |
fdd1e74c TM |
238 | nfs_readpage_result(task, data); |
239 | } | |
240 | ||
241 | static void nfs_direct_read_release(void *calldata) | |
242 | { | |
243 | ||
244 | struct nfs_read_data *data = calldata; | |
245 | struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req; | |
246 | int status = data->task.tk_status; | |
15ce4a0c CL |
247 | |
248 | spin_lock(&dreq->lock); | |
fdd1e74c TM |
249 | if (unlikely(status < 0)) { |
250 | dreq->error = status; | |
d4a8f367 TM |
251 | spin_unlock(&dreq->lock); |
252 | } else { | |
253 | dreq->count += data->res.count; | |
254 | spin_unlock(&dreq->lock); | |
255 | nfs_direct_dirty_pages(data->pagevec, | |
256 | data->args.pgbase, | |
257 | data->res.count); | |
258 | } | |
259 | nfs_direct_release_pages(data->pagevec, data->npages); | |
607f31e8 TM |
260 | |
261 | if (put_dreq(dreq)) | |
262 | nfs_direct_complete(dreq); | |
1ae88b2e | 263 | nfs_readdata_free(data); |
1da177e4 LT |
264 | } |
265 | ||
ec06c096 | 266 | static const struct rpc_call_ops nfs_read_direct_ops = { |
f11c88af | 267 | .rpc_call_prepare = nfs_read_prepare, |
ec06c096 | 268 | .rpc_call_done = nfs_direct_read_result, |
fdd1e74c | 269 | .rpc_release = nfs_direct_read_release, |
ec06c096 TM |
270 | }; |
271 | ||
d4cc948b | 272 | /* |
607f31e8 TM |
273 | * For each rsize'd chunk of the user's buffer, dispatch an NFS READ |
274 | * operation. If nfs_readdata_alloc() or get_user_pages() fails, | |
275 | * bail and stop sending more reads. Read length accounting is | |
276 | * handled automatically by nfs_direct_read_result(). Otherwise, if | |
277 | * no requests have been sent, just return an error. | |
1da177e4 | 278 | */ |
02fe4946 CL |
279 | static ssize_t nfs_direct_read_schedule_segment(struct nfs_direct_req *dreq, |
280 | const struct iovec *iov, | |
281 | loff_t pos) | |
1da177e4 | 282 | { |
a8881f5a | 283 | struct nfs_open_context *ctx = dreq->ctx; |
3d4ff43d | 284 | struct inode *inode = ctx->dentry->d_inode; |
02fe4946 CL |
285 | unsigned long user_addr = (unsigned long)iov->iov_base; |
286 | size_t count = iov->iov_len; | |
5dd602f2 | 287 | size_t rsize = NFS_SERVER(inode)->rsize; |
07737691 | 288 | struct rpc_task *task; |
bdc7f021 TM |
289 | struct rpc_message msg = { |
290 | .rpc_cred = ctx->cred, | |
291 | }; | |
84115e1c TM |
292 | struct rpc_task_setup task_setup_data = { |
293 | .rpc_client = NFS_CLIENT(inode), | |
bdc7f021 | 294 | .rpc_message = &msg, |
84115e1c | 295 | .callback_ops = &nfs_read_direct_ops, |
101070ca | 296 | .workqueue = nfsiod_workqueue, |
84115e1c TM |
297 | .flags = RPC_TASK_ASYNC, |
298 | }; | |
607f31e8 TM |
299 | unsigned int pgbase; |
300 | int result; | |
301 | ssize_t started = 0; | |
302 | ||
1da177e4 | 303 | do { |
82b145c5 | 304 | struct nfs_read_data *data; |
5dd602f2 | 305 | size_t bytes; |
1da177e4 | 306 | |
e9f7bee1 TM |
307 | pgbase = user_addr & ~PAGE_MASK; |
308 | bytes = min(rsize,count); | |
309 | ||
607f31e8 | 310 | result = -ENOMEM; |
8d5658c9 | 311 | data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes)); |
607f31e8 TM |
312 | if (unlikely(!data)) |
313 | break; | |
314 | ||
607f31e8 TM |
315 | down_read(¤t->mm->mmap_sem); |
316 | result = get_user_pages(current, current->mm, user_addr, | |
317 | data->npages, 1, 0, data->pagevec, NULL); | |
318 | up_read(¤t->mm->mmap_sem); | |
749e146e | 319 | if (result < 0) { |
1ae88b2e | 320 | nfs_readdata_free(data); |
749e146e CL |
321 | break; |
322 | } | |
323 | if ((unsigned)result < data->npages) { | |
d9df8d6b TM |
324 | bytes = result * PAGE_SIZE; |
325 | if (bytes <= pgbase) { | |
326 | nfs_direct_release_pages(data->pagevec, result); | |
1ae88b2e | 327 | nfs_readdata_free(data); |
d9df8d6b TM |
328 | break; |
329 | } | |
330 | bytes -= pgbase; | |
331 | data->npages = result; | |
607f31e8 TM |
332 | } |
333 | ||
334 | get_dreq(dreq); | |
82b145c5 | 335 | |
607f31e8 | 336 | data->req = (struct nfs_page *) dreq; |
1da177e4 | 337 | data->inode = inode; |
bdc7f021 | 338 | data->cred = msg.rpc_cred; |
1da177e4 | 339 | data->args.fh = NFS_FH(inode); |
1ae88b2e | 340 | data->args.context = ctx; |
f11ac8db | 341 | data->args.lock_context = dreq->l_ctx; |
88467055 | 342 | data->args.offset = pos; |
1da177e4 | 343 | data->args.pgbase = pgbase; |
607f31e8 | 344 | data->args.pages = data->pagevec; |
1da177e4 LT |
345 | data->args.count = bytes; |
346 | data->res.fattr = &data->fattr; | |
347 | data->res.eof = 0; | |
348 | data->res.count = bytes; | |
65d26953 | 349 | nfs_fattr_init(&data->fattr); |
bdc7f021 TM |
350 | msg.rpc_argp = &data->args; |
351 | msg.rpc_resp = &data->res; | |
1da177e4 | 352 | |
07737691 | 353 | task_setup_data.task = &data->task; |
84115e1c | 354 | task_setup_data.callback_data = data; |
bdc7f021 | 355 | NFS_PROTO(inode)->read_setup(data, &msg); |
1da177e4 | 356 | |
07737691 | 357 | task = rpc_run_task(&task_setup_data); |
dbae4c73 TM |
358 | if (IS_ERR(task)) |
359 | break; | |
360 | rpc_put_task(task); | |
1da177e4 | 361 | |
a3f565b1 CL |
362 | dprintk("NFS: %5u initiated direct read call " |
363 | "(req %s/%Ld, %zu bytes @ offset %Lu)\n", | |
1da177e4 LT |
364 | data->task.tk_pid, |
365 | inode->i_sb->s_id, | |
366 | (long long)NFS_FILEID(inode), | |
367 | bytes, | |
368 | (unsigned long long)data->args.offset); | |
369 | ||
607f31e8 TM |
370 | started += bytes; |
371 | user_addr += bytes; | |
88467055 | 372 | pos += bytes; |
e9f7bee1 | 373 | /* FIXME: Remove this unnecessary math from final patch */ |
1da177e4 | 374 | pgbase += bytes; |
1da177e4 | 375 | pgbase &= ~PAGE_MASK; |
e9f7bee1 | 376 | BUG_ON(pgbase != (user_addr & ~PAGE_MASK)); |
1da177e4 LT |
377 | |
378 | count -= bytes; | |
379 | } while (count != 0); | |
607f31e8 | 380 | |
607f31e8 | 381 | if (started) |
c216fd70 | 382 | return started; |
607f31e8 | 383 | return result < 0 ? (ssize_t) result : -EFAULT; |
1da177e4 LT |
384 | } |
385 | ||
19f73787 CL |
386 | static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq, |
387 | const struct iovec *iov, | |
388 | unsigned long nr_segs, | |
389 | loff_t pos) | |
390 | { | |
391 | ssize_t result = -EINVAL; | |
392 | size_t requested_bytes = 0; | |
393 | unsigned long seg; | |
394 | ||
395 | get_dreq(dreq); | |
396 | ||
397 | for (seg = 0; seg < nr_segs; seg++) { | |
398 | const struct iovec *vec = &iov[seg]; | |
02fe4946 | 399 | result = nfs_direct_read_schedule_segment(dreq, vec, pos); |
19f73787 CL |
400 | if (result < 0) |
401 | break; | |
402 | requested_bytes += result; | |
403 | if ((size_t)result < vec->iov_len) | |
404 | break; | |
405 | pos += vec->iov_len; | |
406 | } | |
407 | ||
839f7ad6 CL |
408 | /* |
409 | * If no bytes were started, return the error, and let the | |
410 | * generic layer handle the completion. | |
411 | */ | |
412 | if (requested_bytes == 0) { | |
413 | nfs_direct_req_release(dreq); | |
414 | return result < 0 ? result : -EIO; | |
415 | } | |
416 | ||
19f73787 CL |
417 | if (put_dreq(dreq)) |
418 | nfs_direct_complete(dreq); | |
839f7ad6 | 419 | return 0; |
19f73787 CL |
420 | } |
421 | ||
c216fd70 CL |
422 | static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov, |
423 | unsigned long nr_segs, loff_t pos) | |
1da177e4 | 424 | { |
f11ac8db | 425 | ssize_t result = -ENOMEM; |
99514f8f | 426 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
1da177e4 LT |
427 | struct nfs_direct_req *dreq; |
428 | ||
607f31e8 | 429 | dreq = nfs_direct_req_alloc(); |
f11ac8db TM |
430 | if (dreq == NULL) |
431 | goto out; | |
1da177e4 | 432 | |
91d5b470 | 433 | dreq->inode = inode; |
cd3758e3 | 434 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
f11ac8db TM |
435 | dreq->l_ctx = nfs_get_lock_context(dreq->ctx); |
436 | if (dreq->l_ctx == NULL) | |
437 | goto out_release; | |
487b8372 CL |
438 | if (!is_sync_kiocb(iocb)) |
439 | dreq->iocb = iocb; | |
1da177e4 | 440 | |
c216fd70 | 441 | result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos); |
607f31e8 TM |
442 | if (!result) |
443 | result = nfs_direct_wait(dreq); | |
f11ac8db | 444 | out_release: |
b4946ffb | 445 | nfs_direct_req_release(dreq); |
f11ac8db | 446 | out: |
1da177e4 LT |
447 | return result; |
448 | } | |
449 | ||
fad61490 | 450 | static void nfs_direct_free_writedata(struct nfs_direct_req *dreq) |
1da177e4 | 451 | { |
607f31e8 TM |
452 | while (!list_empty(&dreq->rewrite_list)) { |
453 | struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages); | |
fad61490 | 454 | list_del(&data->pages); |
607f31e8 | 455 | nfs_direct_release_pages(data->pagevec, data->npages); |
1ae88b2e | 456 | nfs_writedata_free(data); |
fad61490 TM |
457 | } |
458 | } | |
1da177e4 | 459 | |
fad61490 TM |
460 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) |
461 | static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) | |
462 | { | |
607f31e8 TM |
463 | struct inode *inode = dreq->inode; |
464 | struct list_head *p; | |
465 | struct nfs_write_data *data; | |
07737691 | 466 | struct rpc_task *task; |
bdc7f021 TM |
467 | struct rpc_message msg = { |
468 | .rpc_cred = dreq->ctx->cred, | |
469 | }; | |
84115e1c TM |
470 | struct rpc_task_setup task_setup_data = { |
471 | .rpc_client = NFS_CLIENT(inode), | |
a8b40bc7 | 472 | .rpc_message = &msg, |
84115e1c | 473 | .callback_ops = &nfs_write_direct_ops, |
101070ca | 474 | .workqueue = nfsiod_workqueue, |
84115e1c TM |
475 | .flags = RPC_TASK_ASYNC, |
476 | }; | |
1da177e4 | 477 | |
fad61490 | 478 | dreq->count = 0; |
607f31e8 TM |
479 | get_dreq(dreq); |
480 | ||
481 | list_for_each(p, &dreq->rewrite_list) { | |
482 | data = list_entry(p, struct nfs_write_data, pages); | |
483 | ||
484 | get_dreq(dreq); | |
485 | ||
bdc7f021 TM |
486 | /* Use stable writes */ |
487 | data->args.stable = NFS_FILE_SYNC; | |
488 | ||
607f31e8 TM |
489 | /* |
490 | * Reset data->res. | |
491 | */ | |
492 | nfs_fattr_init(&data->fattr); | |
493 | data->res.count = data->args.count; | |
494 | memset(&data->verf, 0, sizeof(data->verf)); | |
495 | ||
496 | /* | |
497 | * Reuse data->task; data->args should not have changed | |
498 | * since the original request was sent. | |
499 | */ | |
07737691 | 500 | task_setup_data.task = &data->task; |
84115e1c | 501 | task_setup_data.callback_data = data; |
bdc7f021 TM |
502 | msg.rpc_argp = &data->args; |
503 | msg.rpc_resp = &data->res; | |
504 | NFS_PROTO(inode)->write_setup(data, &msg); | |
607f31e8 | 505 | |
607f31e8 TM |
506 | /* |
507 | * We're called via an RPC callback, so BKL is already held. | |
508 | */ | |
07737691 TM |
509 | task = rpc_run_task(&task_setup_data); |
510 | if (!IS_ERR(task)) | |
511 | rpc_put_task(task); | |
607f31e8 TM |
512 | |
513 | dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n", | |
514 | data->task.tk_pid, | |
515 | inode->i_sb->s_id, | |
516 | (long long)NFS_FILEID(inode), | |
517 | data->args.count, | |
518 | (unsigned long long)data->args.offset); | |
519 | } | |
fedb595c | 520 | |
607f31e8 TM |
521 | if (put_dreq(dreq)) |
522 | nfs_direct_write_complete(dreq, inode); | |
fad61490 TM |
523 | } |
524 | ||
525 | static void nfs_direct_commit_result(struct rpc_task *task, void *calldata) | |
526 | { | |
527 | struct nfs_write_data *data = calldata; | |
fad61490 TM |
528 | |
529 | /* Call the NFS version-specific code */ | |
c9d8f89d TM |
530 | NFS_PROTO(data->inode)->commit_done(task, data); |
531 | } | |
532 | ||
533 | static void nfs_direct_commit_release(void *calldata) | |
534 | { | |
535 | struct nfs_write_data *data = calldata; | |
536 | struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req; | |
537 | int status = data->task.tk_status; | |
538 | ||
539 | if (status < 0) { | |
60fa3f76 | 540 | dprintk("NFS: %5u commit failed with error %d.\n", |
c9d8f89d | 541 | data->task.tk_pid, status); |
fad61490 | 542 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
60fa3f76 | 543 | } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) { |
c9d8f89d | 544 | dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid); |
fad61490 | 545 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
1da177e4 LT |
546 | } |
547 | ||
c9d8f89d | 548 | dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status); |
fad61490 | 549 | nfs_direct_write_complete(dreq, data->inode); |
1ae88b2e | 550 | nfs_commit_free(data); |
1da177e4 LT |
551 | } |
552 | ||
fad61490 | 553 | static const struct rpc_call_ops nfs_commit_direct_ops = { |
21d9a851 | 554 | .rpc_call_prepare = nfs_write_prepare, |
fad61490 | 555 | .rpc_call_done = nfs_direct_commit_result, |
c9d8f89d | 556 | .rpc_release = nfs_direct_commit_release, |
fad61490 TM |
557 | }; |
558 | ||
559 | static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) | |
1da177e4 | 560 | { |
fad61490 | 561 | struct nfs_write_data *data = dreq->commit_data; |
07737691 | 562 | struct rpc_task *task; |
bdc7f021 TM |
563 | struct rpc_message msg = { |
564 | .rpc_argp = &data->args, | |
565 | .rpc_resp = &data->res, | |
566 | .rpc_cred = dreq->ctx->cred, | |
567 | }; | |
84115e1c | 568 | struct rpc_task_setup task_setup_data = { |
07737691 | 569 | .task = &data->task, |
84115e1c | 570 | .rpc_client = NFS_CLIENT(dreq->inode), |
bdc7f021 | 571 | .rpc_message = &msg, |
84115e1c TM |
572 | .callback_ops = &nfs_commit_direct_ops, |
573 | .callback_data = data, | |
101070ca | 574 | .workqueue = nfsiod_workqueue, |
84115e1c TM |
575 | .flags = RPC_TASK_ASYNC, |
576 | }; | |
1da177e4 | 577 | |
fad61490 | 578 | data->inode = dreq->inode; |
bdc7f021 | 579 | data->cred = msg.rpc_cred; |
1da177e4 | 580 | |
fad61490 | 581 | data->args.fh = NFS_FH(data->inode); |
607f31e8 TM |
582 | data->args.offset = 0; |
583 | data->args.count = 0; | |
1ae88b2e | 584 | data->args.context = dreq->ctx; |
f11ac8db | 585 | data->args.lock_context = dreq->l_ctx; |
fad61490 TM |
586 | data->res.count = 0; |
587 | data->res.fattr = &data->fattr; | |
588 | data->res.verf = &data->verf; | |
65d26953 | 589 | nfs_fattr_init(&data->fattr); |
1da177e4 | 590 | |
bdc7f021 | 591 | NFS_PROTO(data->inode)->commit_setup(data, &msg); |
1da177e4 | 592 | |
fad61490 TM |
593 | /* Note: task.tk_ops->rpc_release will free dreq->commit_data */ |
594 | dreq->commit_data = NULL; | |
1da177e4 | 595 | |
e99170ff | 596 | dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid); |
1da177e4 | 597 | |
07737691 TM |
598 | task = rpc_run_task(&task_setup_data); |
599 | if (!IS_ERR(task)) | |
600 | rpc_put_task(task); | |
fad61490 | 601 | } |
1da177e4 | 602 | |
fad61490 TM |
603 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
604 | { | |
605 | int flags = dreq->flags; | |
1da177e4 | 606 | |
fad61490 TM |
607 | dreq->flags = 0; |
608 | switch (flags) { | |
609 | case NFS_ODIRECT_DO_COMMIT: | |
610 | nfs_direct_commit_schedule(dreq); | |
1da177e4 | 611 | break; |
fad61490 TM |
612 | case NFS_ODIRECT_RESCHED_WRITES: |
613 | nfs_direct_write_reschedule(dreq); | |
614 | break; | |
615 | default: | |
fad61490 TM |
616 | if (dreq->commit_data != NULL) |
617 | nfs_commit_free(dreq->commit_data); | |
618 | nfs_direct_free_writedata(dreq); | |
cd9ae2b6 | 619 | nfs_zap_mapping(inode, inode->i_mapping); |
fad61490 TM |
620 | nfs_direct_complete(dreq); |
621 | } | |
622 | } | |
1da177e4 | 623 | |
fad61490 TM |
624 | static void nfs_alloc_commit_data(struct nfs_direct_req *dreq) |
625 | { | |
c9d8f89d | 626 | dreq->commit_data = nfs_commitdata_alloc(); |
fad61490 TM |
627 | if (dreq->commit_data != NULL) |
628 | dreq->commit_data->req = (struct nfs_page *) dreq; | |
629 | } | |
630 | #else | |
631 | static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq) | |
632 | { | |
633 | dreq->commit_data = NULL; | |
634 | } | |
1da177e4 | 635 | |
fad61490 TM |
636 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
637 | { | |
fad61490 | 638 | nfs_direct_free_writedata(dreq); |
cd9ae2b6 | 639 | nfs_zap_mapping(inode, inode->i_mapping); |
fad61490 TM |
640 | nfs_direct_complete(dreq); |
641 | } | |
642 | #endif | |
1da177e4 | 643 | |
462d5b32 | 644 | static void nfs_direct_write_result(struct rpc_task *task, void *calldata) |
1da177e4 | 645 | { |
462d5b32 | 646 | struct nfs_write_data *data = calldata; |
462d5b32 | 647 | |
83762c56 | 648 | nfs_writeback_done(task, data); |
c9d8f89d TM |
649 | } |
650 | ||
651 | /* | |
652 | * NB: Return the value of the first error return code. Subsequent | |
653 | * errors after the first one are ignored. | |
654 | */ | |
655 | static void nfs_direct_write_release(void *calldata) | |
656 | { | |
657 | struct nfs_write_data *data = calldata; | |
658 | struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req; | |
659 | int status = data->task.tk_status; | |
462d5b32 | 660 | |
15ce4a0c | 661 | spin_lock(&dreq->lock); |
1da177e4 | 662 | |
eda3cef8 | 663 | if (unlikely(status < 0)) { |
432409ee | 664 | /* An error has occurred, so we should not commit */ |
60fa3f76 | 665 | dreq->flags = 0; |
eda3cef8 | 666 | dreq->error = status; |
eda3cef8 | 667 | } |
432409ee NB |
668 | if (unlikely(dreq->error != 0)) |
669 | goto out_unlock; | |
eda3cef8 TM |
670 | |
671 | dreq->count += data->res.count; | |
1da177e4 | 672 | |
fad61490 TM |
673 | if (data->res.verf->committed != NFS_FILE_SYNC) { |
674 | switch (dreq->flags) { | |
675 | case 0: | |
676 | memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf)); | |
677 | dreq->flags = NFS_ODIRECT_DO_COMMIT; | |
1da177e4 | 678 | break; |
fad61490 TM |
679 | case NFS_ODIRECT_DO_COMMIT: |
680 | if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) { | |
c9d8f89d | 681 | dprintk("NFS: %5u write verify failed\n", data->task.tk_pid); |
fad61490 TM |
682 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
683 | } | |
1da177e4 | 684 | } |
1da177e4 | 685 | } |
eda3cef8 | 686 | out_unlock: |
fad61490 | 687 | spin_unlock(&dreq->lock); |
1da177e4 | 688 | |
607f31e8 TM |
689 | if (put_dreq(dreq)) |
690 | nfs_direct_write_complete(dreq, data->inode); | |
462d5b32 CL |
691 | } |
692 | ||
693 | static const struct rpc_call_ops nfs_write_direct_ops = { | |
def6ed7e | 694 | .rpc_call_prepare = nfs_write_prepare, |
462d5b32 | 695 | .rpc_call_done = nfs_direct_write_result, |
fad61490 | 696 | .rpc_release = nfs_direct_write_release, |
462d5b32 CL |
697 | }; |
698 | ||
699 | /* | |
607f31e8 TM |
700 | * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE |
701 | * operation. If nfs_writedata_alloc() or get_user_pages() fails, | |
702 | * bail and stop sending more writes. Write length accounting is | |
703 | * handled automatically by nfs_direct_write_result(). Otherwise, if | |
704 | * no requests have been sent, just return an error. | |
462d5b32 | 705 | */ |
02fe4946 CL |
706 | static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq, |
707 | const struct iovec *iov, | |
708 | loff_t pos, int sync) | |
462d5b32 | 709 | { |
a8881f5a | 710 | struct nfs_open_context *ctx = dreq->ctx; |
3d4ff43d | 711 | struct inode *inode = ctx->dentry->d_inode; |
02fe4946 CL |
712 | unsigned long user_addr = (unsigned long)iov->iov_base; |
713 | size_t count = iov->iov_len; | |
07737691 | 714 | struct rpc_task *task; |
bdc7f021 TM |
715 | struct rpc_message msg = { |
716 | .rpc_cred = ctx->cred, | |
717 | }; | |
84115e1c TM |
718 | struct rpc_task_setup task_setup_data = { |
719 | .rpc_client = NFS_CLIENT(inode), | |
bdc7f021 | 720 | .rpc_message = &msg, |
84115e1c | 721 | .callback_ops = &nfs_write_direct_ops, |
101070ca | 722 | .workqueue = nfsiod_workqueue, |
84115e1c TM |
723 | .flags = RPC_TASK_ASYNC, |
724 | }; | |
462d5b32 | 725 | size_t wsize = NFS_SERVER(inode)->wsize; |
607f31e8 TM |
726 | unsigned int pgbase; |
727 | int result; | |
728 | ssize_t started = 0; | |
82b145c5 | 729 | |
1da177e4 | 730 | do { |
82b145c5 | 731 | struct nfs_write_data *data; |
462d5b32 CL |
732 | size_t bytes; |
733 | ||
e9f7bee1 TM |
734 | pgbase = user_addr & ~PAGE_MASK; |
735 | bytes = min(wsize,count); | |
736 | ||
607f31e8 | 737 | result = -ENOMEM; |
8d5658c9 | 738 | data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes)); |
607f31e8 TM |
739 | if (unlikely(!data)) |
740 | break; | |
741 | ||
607f31e8 TM |
742 | down_read(¤t->mm->mmap_sem); |
743 | result = get_user_pages(current, current->mm, user_addr, | |
744 | data->npages, 0, 0, data->pagevec, NULL); | |
745 | up_read(¤t->mm->mmap_sem); | |
749e146e | 746 | if (result < 0) { |
1ae88b2e | 747 | nfs_writedata_free(data); |
749e146e CL |
748 | break; |
749 | } | |
750 | if ((unsigned)result < data->npages) { | |
d9df8d6b TM |
751 | bytes = result * PAGE_SIZE; |
752 | if (bytes <= pgbase) { | |
753 | nfs_direct_release_pages(data->pagevec, result); | |
1ae88b2e | 754 | nfs_writedata_free(data); |
d9df8d6b TM |
755 | break; |
756 | } | |
757 | bytes -= pgbase; | |
758 | data->npages = result; | |
607f31e8 TM |
759 | } |
760 | ||
761 | get_dreq(dreq); | |
762 | ||
fad61490 | 763 | list_move_tail(&data->pages, &dreq->rewrite_list); |
462d5b32 | 764 | |
607f31e8 | 765 | data->req = (struct nfs_page *) dreq; |
462d5b32 | 766 | data->inode = inode; |
bdc7f021 | 767 | data->cred = msg.rpc_cred; |
462d5b32 | 768 | data->args.fh = NFS_FH(inode); |
1ae88b2e | 769 | data->args.context = ctx; |
f11ac8db | 770 | data->args.lock_context = dreq->l_ctx; |
88467055 | 771 | data->args.offset = pos; |
462d5b32 | 772 | data->args.pgbase = pgbase; |
607f31e8 | 773 | data->args.pages = data->pagevec; |
462d5b32 | 774 | data->args.count = bytes; |
bdc7f021 | 775 | data->args.stable = sync; |
462d5b32 CL |
776 | data->res.fattr = &data->fattr; |
777 | data->res.count = bytes; | |
47989d74 | 778 | data->res.verf = &data->verf; |
65d26953 | 779 | nfs_fattr_init(&data->fattr); |
462d5b32 | 780 | |
07737691 | 781 | task_setup_data.task = &data->task; |
84115e1c | 782 | task_setup_data.callback_data = data; |
bdc7f021 TM |
783 | msg.rpc_argp = &data->args; |
784 | msg.rpc_resp = &data->res; | |
785 | NFS_PROTO(inode)->write_setup(data, &msg); | |
1da177e4 | 786 | |
07737691 | 787 | task = rpc_run_task(&task_setup_data); |
dbae4c73 TM |
788 | if (IS_ERR(task)) |
789 | break; | |
790 | rpc_put_task(task); | |
1da177e4 | 791 | |
a3f565b1 CL |
792 | dprintk("NFS: %5u initiated direct write call " |
793 | "(req %s/%Ld, %zu bytes @ offset %Lu)\n", | |
462d5b32 CL |
794 | data->task.tk_pid, |
795 | inode->i_sb->s_id, | |
796 | (long long)NFS_FILEID(inode), | |
797 | bytes, | |
798 | (unsigned long long)data->args.offset); | |
1da177e4 | 799 | |
607f31e8 TM |
800 | started += bytes; |
801 | user_addr += bytes; | |
88467055 | 802 | pos += bytes; |
e9f7bee1 TM |
803 | |
804 | /* FIXME: Remove this useless math from the final patch */ | |
462d5b32 | 805 | pgbase += bytes; |
462d5b32 | 806 | pgbase &= ~PAGE_MASK; |
e9f7bee1 | 807 | BUG_ON(pgbase != (user_addr & ~PAGE_MASK)); |
1da177e4 | 808 | |
462d5b32 CL |
809 | count -= bytes; |
810 | } while (count != 0); | |
607f31e8 | 811 | |
607f31e8 | 812 | if (started) |
c216fd70 | 813 | return started; |
607f31e8 | 814 | return result < 0 ? (ssize_t) result : -EFAULT; |
462d5b32 | 815 | } |
1da177e4 | 816 | |
19f73787 CL |
817 | static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, |
818 | const struct iovec *iov, | |
819 | unsigned long nr_segs, | |
820 | loff_t pos, int sync) | |
821 | { | |
822 | ssize_t result = 0; | |
823 | size_t requested_bytes = 0; | |
824 | unsigned long seg; | |
825 | ||
826 | get_dreq(dreq); | |
827 | ||
828 | for (seg = 0; seg < nr_segs; seg++) { | |
829 | const struct iovec *vec = &iov[seg]; | |
02fe4946 CL |
830 | result = nfs_direct_write_schedule_segment(dreq, vec, |
831 | pos, sync); | |
19f73787 CL |
832 | if (result < 0) |
833 | break; | |
834 | requested_bytes += result; | |
835 | if ((size_t)result < vec->iov_len) | |
836 | break; | |
837 | pos += vec->iov_len; | |
838 | } | |
839 | ||
839f7ad6 CL |
840 | /* |
841 | * If no bytes were started, return the error, and let the | |
842 | * generic layer handle the completion. | |
843 | */ | |
844 | if (requested_bytes == 0) { | |
845 | nfs_direct_req_release(dreq); | |
846 | return result < 0 ? result : -EIO; | |
847 | } | |
848 | ||
19f73787 CL |
849 | if (put_dreq(dreq)) |
850 | nfs_direct_write_complete(dreq, dreq->inode); | |
839f7ad6 | 851 | return 0; |
19f73787 CL |
852 | } |
853 | ||
c216fd70 CL |
854 | static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov, |
855 | unsigned long nr_segs, loff_t pos, | |
856 | size_t count) | |
462d5b32 | 857 | { |
f11ac8db | 858 | ssize_t result = -ENOMEM; |
c89f2ee5 | 859 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
462d5b32 | 860 | struct nfs_direct_req *dreq; |
fad61490 | 861 | size_t wsize = NFS_SERVER(inode)->wsize; |
bdc7f021 | 862 | int sync = NFS_UNSTABLE; |
1da177e4 | 863 | |
607f31e8 | 864 | dreq = nfs_direct_req_alloc(); |
462d5b32 | 865 | if (!dreq) |
f11ac8db | 866 | goto out; |
607f31e8 TM |
867 | nfs_alloc_commit_data(dreq); |
868 | ||
b47d19de | 869 | if (dreq->commit_data == NULL || count <= wsize) |
bdc7f021 | 870 | sync = NFS_FILE_SYNC; |
1da177e4 | 871 | |
c89f2ee5 | 872 | dreq->inode = inode; |
cd3758e3 | 873 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
f11ac8db | 874 | dreq->l_ctx = nfs_get_lock_context(dreq->ctx); |
568a810d | 875 | if (dreq->l_ctx == NULL) |
f11ac8db | 876 | goto out_release; |
c89f2ee5 CL |
877 | if (!is_sync_kiocb(iocb)) |
878 | dreq->iocb = iocb; | |
1da177e4 | 879 | |
c216fd70 | 880 | result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync); |
607f31e8 TM |
881 | if (!result) |
882 | result = nfs_direct_wait(dreq); | |
f11ac8db | 883 | out_release: |
b4946ffb | 884 | nfs_direct_req_release(dreq); |
f11ac8db | 885 | out: |
1da177e4 LT |
886 | return result; |
887 | } | |
888 | ||
889 | /** | |
890 | * nfs_file_direct_read - file direct read operation for NFS files | |
891 | * @iocb: target I/O control block | |
027445c3 BP |
892 | * @iov: vector of user buffers into which to read data |
893 | * @nr_segs: size of iov vector | |
88467055 | 894 | * @pos: byte offset in file where reading starts |
1da177e4 LT |
895 | * |
896 | * We use this function for direct reads instead of calling | |
897 | * generic_file_aio_read() in order to avoid gfar's check to see if | |
898 | * the request starts before the end of the file. For that check | |
899 | * to work, we must generate a GETATTR before each direct read, and | |
900 | * even then there is a window between the GETATTR and the subsequent | |
88467055 | 901 | * READ where the file size could change. Our preference is simply |
1da177e4 LT |
902 | * to do all reads the application wants, and the server will take |
903 | * care of managing the end of file boundary. | |
88467055 | 904 | * |
1da177e4 LT |
905 | * This function also eliminates unnecessarily updating the file's |
906 | * atime locally, as the NFS server sets the file's atime, and this | |
907 | * client must read the updated atime from the server back into its | |
908 | * cache. | |
909 | */ | |
027445c3 BP |
910 | ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov, |
911 | unsigned long nr_segs, loff_t pos) | |
1da177e4 LT |
912 | { |
913 | ssize_t retval = -EINVAL; | |
1da177e4 | 914 | struct file *file = iocb->ki_filp; |
1da177e4 | 915 | struct address_space *mapping = file->f_mapping; |
c216fd70 CL |
916 | size_t count; |
917 | ||
918 | count = iov_length(iov, nr_segs); | |
919 | nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); | |
1da177e4 | 920 | |
6da24bc9 | 921 | dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n", |
01cce933 JJS |
922 | file->f_path.dentry->d_parent->d_name.name, |
923 | file->f_path.dentry->d_name.name, | |
c216fd70 | 924 | count, (long long) pos); |
1da177e4 | 925 | |
1da177e4 LT |
926 | retval = 0; |
927 | if (!count) | |
928 | goto out; | |
929 | ||
29884df0 TM |
930 | retval = nfs_sync_mapping(mapping); |
931 | if (retval) | |
932 | goto out; | |
1da177e4 | 933 | |
7ec10f26 KK |
934 | task_io_account_read(count); |
935 | ||
c216fd70 | 936 | retval = nfs_direct_read(iocb, iov, nr_segs, pos); |
1da177e4 | 937 | if (retval > 0) |
0cdd80d0 | 938 | iocb->ki_pos = pos + retval; |
1da177e4 LT |
939 | |
940 | out: | |
941 | return retval; | |
942 | } | |
943 | ||
944 | /** | |
945 | * nfs_file_direct_write - file direct write operation for NFS files | |
946 | * @iocb: target I/O control block | |
027445c3 BP |
947 | * @iov: vector of user buffers from which to write data |
948 | * @nr_segs: size of iov vector | |
88467055 | 949 | * @pos: byte offset in file where writing starts |
1da177e4 LT |
950 | * |
951 | * We use this function for direct writes instead of calling | |
952 | * generic_file_aio_write() in order to avoid taking the inode | |
953 | * semaphore and updating the i_size. The NFS server will set | |
954 | * the new i_size and this client must read the updated size | |
955 | * back into its cache. We let the server do generic write | |
956 | * parameter checking and report problems. | |
957 | * | |
1da177e4 LT |
958 | * We eliminate local atime updates, see direct read above. |
959 | * | |
960 | * We avoid unnecessary page cache invalidations for normal cached | |
961 | * readers of this file. | |
962 | * | |
963 | * Note that O_APPEND is not supported for NFS direct writes, as there | |
964 | * is no atomic O_APPEND write facility in the NFS protocol. | |
965 | */ | |
027445c3 BP |
966 | ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov, |
967 | unsigned long nr_segs, loff_t pos) | |
1da177e4 | 968 | { |
070ea602 | 969 | ssize_t retval = -EINVAL; |
1da177e4 | 970 | struct file *file = iocb->ki_filp; |
1da177e4 | 971 | struct address_space *mapping = file->f_mapping; |
c216fd70 | 972 | size_t count; |
1da177e4 | 973 | |
c216fd70 CL |
974 | count = iov_length(iov, nr_segs); |
975 | nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count); | |
976 | ||
6da24bc9 | 977 | dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n", |
01cce933 JJS |
978 | file->f_path.dentry->d_parent->d_name.name, |
979 | file->f_path.dentry->d_name.name, | |
c216fd70 | 980 | count, (long long) pos); |
027445c3 | 981 | |
ce1a8e67 CL |
982 | retval = generic_write_checks(file, &pos, &count, 0); |
983 | if (retval) | |
1da177e4 | 984 | goto out; |
ce1a8e67 CL |
985 | |
986 | retval = -EINVAL; | |
987 | if ((ssize_t) count < 0) | |
1da177e4 | 988 | goto out; |
1da177e4 LT |
989 | retval = 0; |
990 | if (!count) | |
991 | goto out; | |
ce1a8e67 | 992 | |
29884df0 TM |
993 | retval = nfs_sync_mapping(mapping); |
994 | if (retval) | |
995 | goto out; | |
1da177e4 | 996 | |
7ec10f26 KK |
997 | task_io_account_write(count); |
998 | ||
c216fd70 | 999 | retval = nfs_direct_write(iocb, iov, nr_segs, pos, count); |
9eafa8cc | 1000 | |
1da177e4 | 1001 | if (retval > 0) |
ce1a8e67 | 1002 | iocb->ki_pos = pos + retval; |
1da177e4 LT |
1003 | |
1004 | out: | |
1005 | return retval; | |
1006 | } | |
1007 | ||
88467055 CL |
1008 | /** |
1009 | * nfs_init_directcache - create a slab cache for nfs_direct_req structures | |
1010 | * | |
1011 | */ | |
f7b422b1 | 1012 | int __init nfs_init_directcache(void) |
1da177e4 LT |
1013 | { |
1014 | nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", | |
1015 | sizeof(struct nfs_direct_req), | |
fffb60f9 PJ |
1016 | 0, (SLAB_RECLAIM_ACCOUNT| |
1017 | SLAB_MEM_SPREAD), | |
20c2df83 | 1018 | NULL); |
1da177e4 LT |
1019 | if (nfs_direct_cachep == NULL) |
1020 | return -ENOMEM; | |
1021 | ||
1022 | return 0; | |
1023 | } | |
1024 | ||
88467055 | 1025 | /** |
f7b422b1 | 1026 | * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures |
88467055 CL |
1027 | * |
1028 | */ | |
266bee88 | 1029 | void nfs_destroy_directcache(void) |
1da177e4 | 1030 | { |
1a1d92c1 | 1031 | kmem_cache_destroy(nfs_direct_cachep); |
1da177e4 | 1032 | } |