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> | |
47 | ||
48 | #include <linux/nfs_fs.h> | |
49 | #include <linux/nfs_page.h> | |
50 | #include <linux/sunrpc/clnt.h> | |
51 | ||
52 | #include <asm/system.h> | |
53 | #include <asm/uaccess.h> | |
54 | #include <asm/atomic.h> | |
55 | ||
8d5658c9 | 56 | #include "internal.h" |
91d5b470 CL |
57 | #include "iostat.h" |
58 | ||
1da177e4 | 59 | #define NFSDBG_FACILITY NFSDBG_VFS |
1da177e4 | 60 | |
e18b890b | 61 | static struct kmem_cache *nfs_direct_cachep; |
1da177e4 LT |
62 | |
63 | /* | |
64 | * This represents a set of asynchronous requests that we're waiting on | |
65 | */ | |
66 | struct nfs_direct_req { | |
67 | struct kref kref; /* release manager */ | |
15ce4a0c CL |
68 | |
69 | /* I/O parameters */ | |
a8881f5a | 70 | struct nfs_open_context *ctx; /* file open context info */ |
99514f8f | 71 | struct kiocb * iocb; /* controlling i/o request */ |
88467055 | 72 | struct inode * inode; /* target file of i/o */ |
15ce4a0c CL |
73 | |
74 | /* completion state */ | |
607f31e8 | 75 | atomic_t io_count; /* i/os we're waiting for */ |
15ce4a0c | 76 | spinlock_t lock; /* protect completion state */ |
15ce4a0c | 77 | ssize_t count, /* bytes actually processed */ |
1da177e4 | 78 | error; /* any reported error */ |
d72b7a6b | 79 | struct completion completion; /* wait for i/o completion */ |
fad61490 TM |
80 | |
81 | /* commit state */ | |
607f31e8 | 82 | struct list_head rewrite_list; /* saved nfs_write_data structs */ |
fad61490 TM |
83 | struct nfs_write_data * commit_data; /* special write_data for commits */ |
84 | int flags; | |
85 | #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */ | |
86 | #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */ | |
87 | struct nfs_writeverf verf; /* unstable write verifier */ | |
1da177e4 LT |
88 | }; |
89 | ||
fad61490 | 90 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode); |
607f31e8 TM |
91 | static const struct rpc_call_ops nfs_write_direct_ops; |
92 | ||
93 | static inline void get_dreq(struct nfs_direct_req *dreq) | |
94 | { | |
95 | atomic_inc(&dreq->io_count); | |
96 | } | |
97 | ||
98 | static inline int put_dreq(struct nfs_direct_req *dreq) | |
99 | { | |
100 | return atomic_dec_and_test(&dreq->io_count); | |
101 | } | |
102 | ||
1da177e4 | 103 | /** |
b8a32e2b CL |
104 | * nfs_direct_IO - NFS address space operation for direct I/O |
105 | * @rw: direction (read or write) | |
106 | * @iocb: target I/O control block | |
107 | * @iov: array of vectors that define I/O buffer | |
108 | * @pos: offset in file to begin the operation | |
109 | * @nr_segs: size of iovec array | |
110 | * | |
111 | * The presence of this routine in the address space ops vector means | |
112 | * the NFS client supports direct I/O. However, we shunt off direct | |
113 | * read and write requests before the VFS gets them, so this method | |
114 | * should never be called. | |
1da177e4 | 115 | */ |
b8a32e2b CL |
116 | ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs) |
117 | { | |
b8a32e2b | 118 | dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n", |
01cce933 | 119 | iocb->ki_filp->f_path.dentry->d_name.name, |
e99170ff | 120 | (long long) pos, nr_segs); |
b8a32e2b CL |
121 | |
122 | return -EINVAL; | |
123 | } | |
124 | ||
749e146e | 125 | static void nfs_direct_dirty_pages(struct page **pages, unsigned int npages) |
6b45d858 | 126 | { |
749e146e | 127 | unsigned int i; |
6b45d858 TM |
128 | for (i = 0; i < npages; i++) { |
129 | struct page *page = pages[i]; | |
607f31e8 | 130 | if (!PageCompound(page)) |
6b45d858 | 131 | set_page_dirty_lock(page); |
6b45d858 | 132 | } |
9c93ab7d CL |
133 | } |
134 | ||
749e146e | 135 | static void nfs_direct_release_pages(struct page **pages, unsigned int npages) |
9c93ab7d | 136 | { |
749e146e | 137 | unsigned int i; |
607f31e8 TM |
138 | for (i = 0; i < npages; i++) |
139 | page_cache_release(pages[i]); | |
6b45d858 TM |
140 | } |
141 | ||
93619e59 | 142 | static inline struct nfs_direct_req *nfs_direct_req_alloc(void) |
1da177e4 | 143 | { |
93619e59 CL |
144 | struct nfs_direct_req *dreq; |
145 | ||
e94b1766 | 146 | dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL); |
93619e59 CL |
147 | if (!dreq) |
148 | return NULL; | |
149 | ||
150 | kref_init(&dreq->kref); | |
607f31e8 | 151 | kref_get(&dreq->kref); |
d72b7a6b | 152 | init_completion(&dreq->completion); |
fad61490 | 153 | INIT_LIST_HEAD(&dreq->rewrite_list); |
93619e59 | 154 | dreq->iocb = NULL; |
a8881f5a | 155 | dreq->ctx = NULL; |
15ce4a0c | 156 | spin_lock_init(&dreq->lock); |
607f31e8 | 157 | atomic_set(&dreq->io_count, 0); |
15ce4a0c CL |
158 | dreq->count = 0; |
159 | dreq->error = 0; | |
fad61490 | 160 | dreq->flags = 0; |
93619e59 CL |
161 | |
162 | return dreq; | |
1da177e4 LT |
163 | } |
164 | ||
1da177e4 LT |
165 | static void nfs_direct_req_release(struct kref *kref) |
166 | { | |
167 | struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); | |
a8881f5a TM |
168 | |
169 | if (dreq->ctx != NULL) | |
170 | put_nfs_open_context(dreq->ctx); | |
1da177e4 LT |
171 | kmem_cache_free(nfs_direct_cachep, dreq); |
172 | } | |
173 | ||
bc0fb201 CL |
174 | /* |
175 | * Collects and returns the final error value/byte-count. | |
176 | */ | |
177 | static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) | |
178 | { | |
15ce4a0c | 179 | ssize_t result = -EIOCBQUEUED; |
bc0fb201 CL |
180 | |
181 | /* Async requests don't wait here */ | |
182 | if (dreq->iocb) | |
183 | goto out; | |
184 | ||
d72b7a6b | 185 | result = wait_for_completion_interruptible(&dreq->completion); |
bc0fb201 CL |
186 | |
187 | if (!result) | |
15ce4a0c | 188 | result = dreq->error; |
bc0fb201 | 189 | if (!result) |
15ce4a0c | 190 | result = dreq->count; |
bc0fb201 CL |
191 | |
192 | out: | |
193 | kref_put(&dreq->kref, nfs_direct_req_release); | |
194 | return (ssize_t) result; | |
195 | } | |
196 | ||
63ab46ab | 197 | /* |
607f31e8 TM |
198 | * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust |
199 | * the iocb is still valid here if this is a synchronous request. | |
63ab46ab CL |
200 | */ |
201 | static void nfs_direct_complete(struct nfs_direct_req *dreq) | |
202 | { | |
63ab46ab | 203 | if (dreq->iocb) { |
15ce4a0c | 204 | long res = (long) dreq->error; |
63ab46ab | 205 | if (!res) |
15ce4a0c | 206 | res = (long) dreq->count; |
63ab46ab | 207 | aio_complete(dreq->iocb, res, 0); |
d72b7a6b TM |
208 | } |
209 | complete_all(&dreq->completion); | |
63ab46ab CL |
210 | |
211 | kref_put(&dreq->kref, nfs_direct_req_release); | |
212 | } | |
213 | ||
06cf6f2e | 214 | /* |
607f31e8 TM |
215 | * We must hold a reference to all the pages in this direct read request |
216 | * until the RPCs complete. This could be long *after* we are woken up in | |
217 | * nfs_direct_wait (for instance, if someone hits ^C on a slow server). | |
06cf6f2e | 218 | */ |
ec06c096 | 219 | static void nfs_direct_read_result(struct rpc_task *task, void *calldata) |
1da177e4 | 220 | { |
ec06c096 | 221 | struct nfs_read_data *data = calldata; |
1da177e4 LT |
222 | struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req; |
223 | ||
ec06c096 TM |
224 | if (nfs_readpage_result(task, data) != 0) |
225 | return; | |
15ce4a0c | 226 | |
607f31e8 TM |
227 | nfs_direct_dirty_pages(data->pagevec, data->npages); |
228 | nfs_direct_release_pages(data->pagevec, data->npages); | |
229 | ||
15ce4a0c CL |
230 | spin_lock(&dreq->lock); |
231 | ||
ec06c096 | 232 | if (likely(task->tk_status >= 0)) |
15ce4a0c | 233 | dreq->count += data->res.count; |
1da177e4 | 234 | else |
15ce4a0c | 235 | dreq->error = task->tk_status; |
1da177e4 | 236 | |
ccf01ef7 | 237 | spin_unlock(&dreq->lock); |
607f31e8 TM |
238 | |
239 | if (put_dreq(dreq)) | |
240 | nfs_direct_complete(dreq); | |
1da177e4 LT |
241 | } |
242 | ||
ec06c096 TM |
243 | static const struct rpc_call_ops nfs_read_direct_ops = { |
244 | .rpc_call_done = nfs_direct_read_result, | |
245 | .rpc_release = nfs_readdata_release, | |
246 | }; | |
247 | ||
d4cc948b | 248 | /* |
607f31e8 TM |
249 | * For each rsize'd chunk of the user's buffer, dispatch an NFS READ |
250 | * operation. If nfs_readdata_alloc() or get_user_pages() fails, | |
251 | * bail and stop sending more reads. Read length accounting is | |
252 | * handled automatically by nfs_direct_read_result(). Otherwise, if | |
253 | * no requests have been sent, just return an error. | |
1da177e4 | 254 | */ |
607f31e8 | 255 | static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos) |
1da177e4 | 256 | { |
a8881f5a TM |
257 | struct nfs_open_context *ctx = dreq->ctx; |
258 | struct inode *inode = ctx->dentry->d_inode; | |
5dd602f2 | 259 | size_t rsize = NFS_SERVER(inode)->rsize; |
607f31e8 TM |
260 | unsigned int pgbase; |
261 | int result; | |
262 | ssize_t started = 0; | |
263 | ||
264 | get_dreq(dreq); | |
1da177e4 | 265 | |
1da177e4 | 266 | do { |
82b145c5 | 267 | struct nfs_read_data *data; |
5dd602f2 | 268 | size_t bytes; |
1da177e4 | 269 | |
e9f7bee1 TM |
270 | pgbase = user_addr & ~PAGE_MASK; |
271 | bytes = min(rsize,count); | |
272 | ||
607f31e8 | 273 | result = -ENOMEM; |
8d5658c9 | 274 | data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes)); |
607f31e8 TM |
275 | if (unlikely(!data)) |
276 | break; | |
277 | ||
607f31e8 TM |
278 | down_read(¤t->mm->mmap_sem); |
279 | result = get_user_pages(current, current->mm, user_addr, | |
280 | data->npages, 1, 0, data->pagevec, NULL); | |
281 | up_read(¤t->mm->mmap_sem); | |
749e146e CL |
282 | if (result < 0) { |
283 | nfs_readdata_release(data); | |
284 | break; | |
285 | } | |
286 | if ((unsigned)result < data->npages) { | |
287 | nfs_direct_release_pages(data->pagevec, result); | |
607f31e8 TM |
288 | nfs_readdata_release(data); |
289 | break; | |
290 | } | |
291 | ||
292 | get_dreq(dreq); | |
82b145c5 | 293 | |
607f31e8 | 294 | data->req = (struct nfs_page *) dreq; |
1da177e4 LT |
295 | data->inode = inode; |
296 | data->cred = ctx->cred; | |
297 | data->args.fh = NFS_FH(inode); | |
298 | data->args.context = ctx; | |
88467055 | 299 | data->args.offset = pos; |
1da177e4 | 300 | data->args.pgbase = pgbase; |
607f31e8 | 301 | data->args.pages = data->pagevec; |
1da177e4 LT |
302 | data->args.count = bytes; |
303 | data->res.fattr = &data->fattr; | |
304 | data->res.eof = 0; | |
305 | data->res.count = bytes; | |
306 | ||
ec06c096 TM |
307 | rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC, |
308 | &nfs_read_direct_ops, data); | |
1da177e4 LT |
309 | NFS_PROTO(inode)->read_setup(data); |
310 | ||
311 | data->task.tk_cookie = (unsigned long) inode; | |
1da177e4 | 312 | |
1da177e4 | 313 | rpc_execute(&data->task); |
1da177e4 | 314 | |
a3f565b1 CL |
315 | dprintk("NFS: %5u initiated direct read call " |
316 | "(req %s/%Ld, %zu bytes @ offset %Lu)\n", | |
1da177e4 LT |
317 | data->task.tk_pid, |
318 | inode->i_sb->s_id, | |
319 | (long long)NFS_FILEID(inode), | |
320 | bytes, | |
321 | (unsigned long long)data->args.offset); | |
322 | ||
607f31e8 TM |
323 | started += bytes; |
324 | user_addr += bytes; | |
88467055 | 325 | pos += bytes; |
e9f7bee1 | 326 | /* FIXME: Remove this unnecessary math from final patch */ |
1da177e4 | 327 | pgbase += bytes; |
1da177e4 | 328 | pgbase &= ~PAGE_MASK; |
e9f7bee1 | 329 | BUG_ON(pgbase != (user_addr & ~PAGE_MASK)); |
1da177e4 LT |
330 | |
331 | count -= bytes; | |
332 | } while (count != 0); | |
607f31e8 TM |
333 | |
334 | if (put_dreq(dreq)) | |
335 | nfs_direct_complete(dreq); | |
336 | ||
337 | if (started) | |
338 | return 0; | |
339 | return result < 0 ? (ssize_t) result : -EFAULT; | |
1da177e4 LT |
340 | } |
341 | ||
607f31e8 | 342 | static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos) |
1da177e4 | 343 | { |
607f31e8 | 344 | ssize_t result = 0; |
1da177e4 | 345 | sigset_t oldset; |
99514f8f | 346 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
1da177e4 LT |
347 | struct rpc_clnt *clnt = NFS_CLIENT(inode); |
348 | struct nfs_direct_req *dreq; | |
349 | ||
607f31e8 | 350 | dreq = nfs_direct_req_alloc(); |
1da177e4 LT |
351 | if (!dreq) |
352 | return -ENOMEM; | |
353 | ||
91d5b470 | 354 | dreq->inode = inode; |
a8881f5a | 355 | dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data); |
487b8372 CL |
356 | if (!is_sync_kiocb(iocb)) |
357 | dreq->iocb = iocb; | |
1da177e4 | 358 | |
91d5b470 | 359 | nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count); |
1da177e4 | 360 | rpc_clnt_sigmask(clnt, &oldset); |
607f31e8 TM |
361 | result = nfs_direct_read_schedule(dreq, user_addr, count, pos); |
362 | if (!result) | |
363 | result = nfs_direct_wait(dreq); | |
1da177e4 LT |
364 | rpc_clnt_sigunmask(clnt, &oldset); |
365 | ||
366 | return result; | |
367 | } | |
368 | ||
fad61490 | 369 | static void nfs_direct_free_writedata(struct nfs_direct_req *dreq) |
1da177e4 | 370 | { |
607f31e8 TM |
371 | while (!list_empty(&dreq->rewrite_list)) { |
372 | struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages); | |
fad61490 | 373 | list_del(&data->pages); |
607f31e8 | 374 | nfs_direct_release_pages(data->pagevec, data->npages); |
fad61490 TM |
375 | nfs_writedata_release(data); |
376 | } | |
377 | } | |
1da177e4 | 378 | |
fad61490 TM |
379 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) |
380 | static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) | |
381 | { | |
607f31e8 TM |
382 | struct inode *inode = dreq->inode; |
383 | struct list_head *p; | |
384 | struct nfs_write_data *data; | |
1da177e4 | 385 | |
fad61490 | 386 | dreq->count = 0; |
607f31e8 TM |
387 | get_dreq(dreq); |
388 | ||
389 | list_for_each(p, &dreq->rewrite_list) { | |
390 | data = list_entry(p, struct nfs_write_data, pages); | |
391 | ||
392 | get_dreq(dreq); | |
393 | ||
394 | /* | |
395 | * Reset data->res. | |
396 | */ | |
397 | nfs_fattr_init(&data->fattr); | |
398 | data->res.count = data->args.count; | |
399 | memset(&data->verf, 0, sizeof(data->verf)); | |
400 | ||
401 | /* | |
402 | * Reuse data->task; data->args should not have changed | |
403 | * since the original request was sent. | |
404 | */ | |
405 | rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC, | |
406 | &nfs_write_direct_ops, data); | |
407 | NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE); | |
408 | ||
409 | data->task.tk_priority = RPC_PRIORITY_NORMAL; | |
410 | data->task.tk_cookie = (unsigned long) inode; | |
411 | ||
412 | /* | |
413 | * We're called via an RPC callback, so BKL is already held. | |
414 | */ | |
415 | rpc_execute(&data->task); | |
416 | ||
417 | dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n", | |
418 | data->task.tk_pid, | |
419 | inode->i_sb->s_id, | |
420 | (long long)NFS_FILEID(inode), | |
421 | data->args.count, | |
422 | (unsigned long long)data->args.offset); | |
423 | } | |
fedb595c | 424 | |
607f31e8 TM |
425 | if (put_dreq(dreq)) |
426 | nfs_direct_write_complete(dreq, inode); | |
fad61490 TM |
427 | } |
428 | ||
429 | static void nfs_direct_commit_result(struct rpc_task *task, void *calldata) | |
430 | { | |
431 | struct nfs_write_data *data = calldata; | |
432 | struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req; | |
433 | ||
434 | /* Call the NFS version-specific code */ | |
435 | if (NFS_PROTO(data->inode)->commit_done(task, data) != 0) | |
436 | return; | |
437 | if (unlikely(task->tk_status < 0)) { | |
60fa3f76 TM |
438 | dprintk("NFS: %5u commit failed with error %d.\n", |
439 | task->tk_pid, task->tk_status); | |
fad61490 | 440 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
60fa3f76 | 441 | } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) { |
fad61490 TM |
442 | dprintk("NFS: %5u commit verify failed\n", task->tk_pid); |
443 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; | |
1da177e4 LT |
444 | } |
445 | ||
fad61490 TM |
446 | dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status); |
447 | nfs_direct_write_complete(dreq, data->inode); | |
1da177e4 LT |
448 | } |
449 | ||
fad61490 TM |
450 | static const struct rpc_call_ops nfs_commit_direct_ops = { |
451 | .rpc_call_done = nfs_direct_commit_result, | |
452 | .rpc_release = nfs_commit_release, | |
453 | }; | |
454 | ||
455 | static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) | |
1da177e4 | 456 | { |
fad61490 | 457 | struct nfs_write_data *data = dreq->commit_data; |
1da177e4 | 458 | |
fad61490 | 459 | data->inode = dreq->inode; |
a8881f5a | 460 | data->cred = dreq->ctx->cred; |
1da177e4 | 461 | |
fad61490 | 462 | data->args.fh = NFS_FH(data->inode); |
607f31e8 TM |
463 | data->args.offset = 0; |
464 | data->args.count = 0; | |
fad61490 TM |
465 | data->res.count = 0; |
466 | data->res.fattr = &data->fattr; | |
467 | data->res.verf = &data->verf; | |
1da177e4 | 468 | |
fad61490 TM |
469 | rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC, |
470 | &nfs_commit_direct_ops, data); | |
471 | NFS_PROTO(data->inode)->commit_setup(data, 0); | |
1da177e4 | 472 | |
fad61490 TM |
473 | data->task.tk_priority = RPC_PRIORITY_NORMAL; |
474 | data->task.tk_cookie = (unsigned long)data->inode; | |
475 | /* Note: task.tk_ops->rpc_release will free dreq->commit_data */ | |
476 | dreq->commit_data = NULL; | |
1da177e4 | 477 | |
e99170ff | 478 | dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid); |
1da177e4 | 479 | |
fad61490 | 480 | rpc_execute(&data->task); |
fad61490 | 481 | } |
1da177e4 | 482 | |
fad61490 TM |
483 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
484 | { | |
485 | int flags = dreq->flags; | |
1da177e4 | 486 | |
fad61490 TM |
487 | dreq->flags = 0; |
488 | switch (flags) { | |
489 | case NFS_ODIRECT_DO_COMMIT: | |
490 | nfs_direct_commit_schedule(dreq); | |
1da177e4 | 491 | break; |
fad61490 TM |
492 | case NFS_ODIRECT_RESCHED_WRITES: |
493 | nfs_direct_write_reschedule(dreq); | |
494 | break; | |
495 | default: | |
496 | nfs_end_data_update(inode); | |
497 | if (dreq->commit_data != NULL) | |
498 | nfs_commit_free(dreq->commit_data); | |
499 | nfs_direct_free_writedata(dreq); | |
cd9ae2b6 | 500 | nfs_zap_mapping(inode, inode->i_mapping); |
fad61490 TM |
501 | nfs_direct_complete(dreq); |
502 | } | |
503 | } | |
1da177e4 | 504 | |
fad61490 TM |
505 | static void nfs_alloc_commit_data(struct nfs_direct_req *dreq) |
506 | { | |
e9f7bee1 | 507 | dreq->commit_data = nfs_commit_alloc(); |
fad61490 TM |
508 | if (dreq->commit_data != NULL) |
509 | dreq->commit_data->req = (struct nfs_page *) dreq; | |
510 | } | |
511 | #else | |
512 | static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq) | |
513 | { | |
514 | dreq->commit_data = NULL; | |
515 | } | |
1da177e4 | 516 | |
fad61490 TM |
517 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
518 | { | |
519 | nfs_end_data_update(inode); | |
520 | nfs_direct_free_writedata(dreq); | |
cd9ae2b6 | 521 | nfs_zap_mapping(inode, inode->i_mapping); |
fad61490 TM |
522 | nfs_direct_complete(dreq); |
523 | } | |
524 | #endif | |
1da177e4 | 525 | |
462d5b32 | 526 | static void nfs_direct_write_result(struct rpc_task *task, void *calldata) |
1da177e4 | 527 | { |
462d5b32 CL |
528 | struct nfs_write_data *data = calldata; |
529 | struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req; | |
530 | int status = task->tk_status; | |
531 | ||
532 | if (nfs_writeback_done(task, data) != 0) | |
533 | return; | |
462d5b32 | 534 | |
15ce4a0c | 535 | spin_lock(&dreq->lock); |
1da177e4 | 536 | |
60fa3f76 TM |
537 | if (unlikely(dreq->error != 0)) |
538 | goto out_unlock; | |
eda3cef8 | 539 | if (unlikely(status < 0)) { |
60fa3f76 TM |
540 | /* An error has occured, so we should not commit */ |
541 | dreq->flags = 0; | |
eda3cef8 | 542 | dreq->error = status; |
eda3cef8 TM |
543 | } |
544 | ||
545 | dreq->count += data->res.count; | |
1da177e4 | 546 | |
fad61490 TM |
547 | if (data->res.verf->committed != NFS_FILE_SYNC) { |
548 | switch (dreq->flags) { | |
549 | case 0: | |
550 | memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf)); | |
551 | dreq->flags = NFS_ODIRECT_DO_COMMIT; | |
1da177e4 | 552 | break; |
fad61490 TM |
553 | case NFS_ODIRECT_DO_COMMIT: |
554 | if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) { | |
555 | dprintk("NFS: %5u write verify failed\n", task->tk_pid); | |
556 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; | |
557 | } | |
1da177e4 | 558 | } |
1da177e4 | 559 | } |
eda3cef8 | 560 | out_unlock: |
fad61490 | 561 | spin_unlock(&dreq->lock); |
1da177e4 LT |
562 | } |
563 | ||
fad61490 TM |
564 | /* |
565 | * NB: Return the value of the first error return code. Subsequent | |
566 | * errors after the first one are ignored. | |
1da177e4 | 567 | */ |
fad61490 | 568 | static void nfs_direct_write_release(void *calldata) |
1da177e4 | 569 | { |
fad61490 TM |
570 | struct nfs_write_data *data = calldata; |
571 | struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req; | |
1da177e4 | 572 | |
607f31e8 TM |
573 | if (put_dreq(dreq)) |
574 | nfs_direct_write_complete(dreq, data->inode); | |
462d5b32 CL |
575 | } |
576 | ||
577 | static const struct rpc_call_ops nfs_write_direct_ops = { | |
578 | .rpc_call_done = nfs_direct_write_result, | |
fad61490 | 579 | .rpc_release = nfs_direct_write_release, |
462d5b32 CL |
580 | }; |
581 | ||
582 | /* | |
607f31e8 TM |
583 | * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE |
584 | * operation. If nfs_writedata_alloc() or get_user_pages() fails, | |
585 | * bail and stop sending more writes. Write length accounting is | |
586 | * handled automatically by nfs_direct_write_result(). Otherwise, if | |
587 | * no requests have been sent, just return an error. | |
462d5b32 | 588 | */ |
607f31e8 | 589 | static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync) |
462d5b32 | 590 | { |
a8881f5a TM |
591 | struct nfs_open_context *ctx = dreq->ctx; |
592 | struct inode *inode = ctx->dentry->d_inode; | |
462d5b32 | 593 | size_t wsize = NFS_SERVER(inode)->wsize; |
607f31e8 TM |
594 | unsigned int pgbase; |
595 | int result; | |
596 | ssize_t started = 0; | |
82b145c5 | 597 | |
607f31e8 TM |
598 | get_dreq(dreq); |
599 | ||
1da177e4 | 600 | do { |
82b145c5 | 601 | struct nfs_write_data *data; |
462d5b32 CL |
602 | size_t bytes; |
603 | ||
e9f7bee1 TM |
604 | pgbase = user_addr & ~PAGE_MASK; |
605 | bytes = min(wsize,count); | |
606 | ||
607f31e8 | 607 | result = -ENOMEM; |
8d5658c9 | 608 | data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes)); |
607f31e8 TM |
609 | if (unlikely(!data)) |
610 | break; | |
611 | ||
607f31e8 TM |
612 | down_read(¤t->mm->mmap_sem); |
613 | result = get_user_pages(current, current->mm, user_addr, | |
614 | data->npages, 0, 0, data->pagevec, NULL); | |
615 | up_read(¤t->mm->mmap_sem); | |
749e146e CL |
616 | if (result < 0) { |
617 | nfs_writedata_release(data); | |
618 | break; | |
619 | } | |
620 | if ((unsigned)result < data->npages) { | |
621 | nfs_direct_release_pages(data->pagevec, result); | |
607f31e8 TM |
622 | nfs_writedata_release(data); |
623 | break; | |
624 | } | |
625 | ||
626 | get_dreq(dreq); | |
627 | ||
fad61490 | 628 | list_move_tail(&data->pages, &dreq->rewrite_list); |
462d5b32 | 629 | |
607f31e8 | 630 | data->req = (struct nfs_page *) dreq; |
462d5b32 CL |
631 | data->inode = inode; |
632 | data->cred = ctx->cred; | |
633 | data->args.fh = NFS_FH(inode); | |
634 | data->args.context = ctx; | |
88467055 | 635 | data->args.offset = pos; |
462d5b32 | 636 | data->args.pgbase = pgbase; |
607f31e8 | 637 | data->args.pages = data->pagevec; |
462d5b32 CL |
638 | data->args.count = bytes; |
639 | data->res.fattr = &data->fattr; | |
640 | data->res.count = bytes; | |
47989d74 | 641 | data->res.verf = &data->verf; |
462d5b32 CL |
642 | |
643 | rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC, | |
644 | &nfs_write_direct_ops, data); | |
fad61490 | 645 | NFS_PROTO(inode)->write_setup(data, sync); |
1da177e4 | 646 | |
462d5b32 CL |
647 | data->task.tk_priority = RPC_PRIORITY_NORMAL; |
648 | data->task.tk_cookie = (unsigned long) inode; | |
1da177e4 | 649 | |
462d5b32 | 650 | rpc_execute(&data->task); |
1da177e4 | 651 | |
a3f565b1 CL |
652 | dprintk("NFS: %5u initiated direct write call " |
653 | "(req %s/%Ld, %zu bytes @ offset %Lu)\n", | |
462d5b32 CL |
654 | data->task.tk_pid, |
655 | inode->i_sb->s_id, | |
656 | (long long)NFS_FILEID(inode), | |
657 | bytes, | |
658 | (unsigned long long)data->args.offset); | |
1da177e4 | 659 | |
607f31e8 TM |
660 | started += bytes; |
661 | user_addr += bytes; | |
88467055 | 662 | pos += bytes; |
e9f7bee1 TM |
663 | |
664 | /* FIXME: Remove this useless math from the final patch */ | |
462d5b32 | 665 | pgbase += bytes; |
462d5b32 | 666 | pgbase &= ~PAGE_MASK; |
e9f7bee1 | 667 | BUG_ON(pgbase != (user_addr & ~PAGE_MASK)); |
1da177e4 | 668 | |
462d5b32 CL |
669 | count -= bytes; |
670 | } while (count != 0); | |
607f31e8 TM |
671 | |
672 | if (put_dreq(dreq)) | |
673 | nfs_direct_write_complete(dreq, inode); | |
674 | ||
675 | if (started) | |
676 | return 0; | |
677 | return result < 0 ? (ssize_t) result : -EFAULT; | |
462d5b32 | 678 | } |
1da177e4 | 679 | |
607f31e8 | 680 | static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos) |
462d5b32 | 681 | { |
607f31e8 | 682 | ssize_t result = 0; |
462d5b32 | 683 | sigset_t oldset; |
c89f2ee5 | 684 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
462d5b32 CL |
685 | struct rpc_clnt *clnt = NFS_CLIENT(inode); |
686 | struct nfs_direct_req *dreq; | |
fad61490 TM |
687 | size_t wsize = NFS_SERVER(inode)->wsize; |
688 | int sync = 0; | |
1da177e4 | 689 | |
607f31e8 | 690 | dreq = nfs_direct_req_alloc(); |
462d5b32 CL |
691 | if (!dreq) |
692 | return -ENOMEM; | |
607f31e8 TM |
693 | nfs_alloc_commit_data(dreq); |
694 | ||
fad61490 TM |
695 | if (dreq->commit_data == NULL || count < wsize) |
696 | sync = FLUSH_STABLE; | |
1da177e4 | 697 | |
c89f2ee5 | 698 | dreq->inode = inode; |
a8881f5a | 699 | dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data); |
c89f2ee5 CL |
700 | if (!is_sync_kiocb(iocb)) |
701 | dreq->iocb = iocb; | |
1da177e4 | 702 | |
47989d74 CL |
703 | nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count); |
704 | ||
462d5b32 | 705 | nfs_begin_data_update(inode); |
1da177e4 | 706 | |
462d5b32 | 707 | rpc_clnt_sigmask(clnt, &oldset); |
607f31e8 TM |
708 | result = nfs_direct_write_schedule(dreq, user_addr, count, pos, sync); |
709 | if (!result) | |
710 | result = nfs_direct_wait(dreq); | |
462d5b32 | 711 | rpc_clnt_sigunmask(clnt, &oldset); |
1da177e4 | 712 | |
1da177e4 LT |
713 | return result; |
714 | } | |
715 | ||
716 | /** | |
717 | * nfs_file_direct_read - file direct read operation for NFS files | |
718 | * @iocb: target I/O control block | |
027445c3 BP |
719 | * @iov: vector of user buffers into which to read data |
720 | * @nr_segs: size of iov vector | |
88467055 | 721 | * @pos: byte offset in file where reading starts |
1da177e4 LT |
722 | * |
723 | * We use this function for direct reads instead of calling | |
724 | * generic_file_aio_read() in order to avoid gfar's check to see if | |
725 | * the request starts before the end of the file. For that check | |
726 | * to work, we must generate a GETATTR before each direct read, and | |
727 | * even then there is a window between the GETATTR and the subsequent | |
88467055 | 728 | * READ where the file size could change. Our preference is simply |
1da177e4 LT |
729 | * to do all reads the application wants, and the server will take |
730 | * care of managing the end of file boundary. | |
88467055 | 731 | * |
1da177e4 LT |
732 | * This function also eliminates unnecessarily updating the file's |
733 | * atime locally, as the NFS server sets the file's atime, and this | |
734 | * client must read the updated atime from the server back into its | |
735 | * cache. | |
736 | */ | |
027445c3 BP |
737 | ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov, |
738 | unsigned long nr_segs, loff_t pos) | |
1da177e4 LT |
739 | { |
740 | ssize_t retval = -EINVAL; | |
1da177e4 | 741 | struct file *file = iocb->ki_filp; |
1da177e4 | 742 | struct address_space *mapping = file->f_mapping; |
027445c3 BP |
743 | /* XXX: temporary */ |
744 | const char __user *buf = iov[0].iov_base; | |
745 | size_t count = iov[0].iov_len; | |
1da177e4 | 746 | |
ce1a8e67 | 747 | dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n", |
01cce933 JJS |
748 | file->f_path.dentry->d_parent->d_name.name, |
749 | file->f_path.dentry->d_name.name, | |
ce1a8e67 | 750 | (unsigned long) count, (long long) pos); |
1da177e4 | 751 | |
027445c3 BP |
752 | if (nr_segs != 1) |
753 | return -EINVAL; | |
754 | ||
1da177e4 LT |
755 | if (count < 0) |
756 | goto out; | |
757 | retval = -EFAULT; | |
0cdd80d0 | 758 | if (!access_ok(VERIFY_WRITE, buf, count)) |
1da177e4 LT |
759 | goto out; |
760 | retval = 0; | |
761 | if (!count) | |
762 | goto out; | |
763 | ||
29884df0 TM |
764 | retval = nfs_sync_mapping(mapping); |
765 | if (retval) | |
766 | goto out; | |
1da177e4 | 767 | |
607f31e8 | 768 | retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos); |
1da177e4 | 769 | if (retval > 0) |
0cdd80d0 | 770 | iocb->ki_pos = pos + retval; |
1da177e4 LT |
771 | |
772 | out: | |
773 | return retval; | |
774 | } | |
775 | ||
776 | /** | |
777 | * nfs_file_direct_write - file direct write operation for NFS files | |
778 | * @iocb: target I/O control block | |
027445c3 BP |
779 | * @iov: vector of user buffers from which to write data |
780 | * @nr_segs: size of iov vector | |
88467055 | 781 | * @pos: byte offset in file where writing starts |
1da177e4 LT |
782 | * |
783 | * We use this function for direct writes instead of calling | |
784 | * generic_file_aio_write() in order to avoid taking the inode | |
785 | * semaphore and updating the i_size. The NFS server will set | |
786 | * the new i_size and this client must read the updated size | |
787 | * back into its cache. We let the server do generic write | |
788 | * parameter checking and report problems. | |
789 | * | |
790 | * We also avoid an unnecessary invocation of generic_osync_inode(), | |
791 | * as it is fairly meaningless to sync the metadata of an NFS file. | |
792 | * | |
793 | * We eliminate local atime updates, see direct read above. | |
794 | * | |
795 | * We avoid unnecessary page cache invalidations for normal cached | |
796 | * readers of this file. | |
797 | * | |
798 | * Note that O_APPEND is not supported for NFS direct writes, as there | |
799 | * is no atomic O_APPEND write facility in the NFS protocol. | |
800 | */ | |
027445c3 BP |
801 | ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov, |
802 | unsigned long nr_segs, loff_t pos) | |
1da177e4 | 803 | { |
ce1a8e67 | 804 | ssize_t retval; |
1da177e4 | 805 | struct file *file = iocb->ki_filp; |
1da177e4 | 806 | struct address_space *mapping = file->f_mapping; |
027445c3 BP |
807 | /* XXX: temporary */ |
808 | const char __user *buf = iov[0].iov_base; | |
809 | size_t count = iov[0].iov_len; | |
1da177e4 | 810 | |
a3f565b1 | 811 | dprintk("nfs: direct write(%s/%s, %lu@%Ld)\n", |
01cce933 JJS |
812 | file->f_path.dentry->d_parent->d_name.name, |
813 | file->f_path.dentry->d_name.name, | |
ce1a8e67 | 814 | (unsigned long) count, (long long) pos); |
1da177e4 | 815 | |
027445c3 BP |
816 | if (nr_segs != 1) |
817 | return -EINVAL; | |
818 | ||
ce1a8e67 CL |
819 | retval = generic_write_checks(file, &pos, &count, 0); |
820 | if (retval) | |
1da177e4 | 821 | goto out; |
ce1a8e67 CL |
822 | |
823 | retval = -EINVAL; | |
824 | if ((ssize_t) count < 0) | |
1da177e4 | 825 | goto out; |
1da177e4 LT |
826 | retval = 0; |
827 | if (!count) | |
828 | goto out; | |
ce1a8e67 CL |
829 | |
830 | retval = -EFAULT; | |
47989d74 | 831 | if (!access_ok(VERIFY_READ, buf, count)) |
ce1a8e67 | 832 | goto out; |
1da177e4 | 833 | |
29884df0 TM |
834 | retval = nfs_sync_mapping(mapping); |
835 | if (retval) | |
836 | goto out; | |
1da177e4 | 837 | |
607f31e8 | 838 | retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos); |
9eafa8cc | 839 | |
1da177e4 | 840 | if (retval > 0) |
ce1a8e67 | 841 | iocb->ki_pos = pos + retval; |
1da177e4 LT |
842 | |
843 | out: | |
844 | return retval; | |
845 | } | |
846 | ||
88467055 CL |
847 | /** |
848 | * nfs_init_directcache - create a slab cache for nfs_direct_req structures | |
849 | * | |
850 | */ | |
f7b422b1 | 851 | int __init nfs_init_directcache(void) |
1da177e4 LT |
852 | { |
853 | nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", | |
854 | sizeof(struct nfs_direct_req), | |
fffb60f9 PJ |
855 | 0, (SLAB_RECLAIM_ACCOUNT| |
856 | SLAB_MEM_SPREAD), | |
1da177e4 LT |
857 | NULL, NULL); |
858 | if (nfs_direct_cachep == NULL) | |
859 | return -ENOMEM; | |
860 | ||
861 | return 0; | |
862 | } | |
863 | ||
88467055 | 864 | /** |
f7b422b1 | 865 | * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures |
88467055 CL |
866 | * |
867 | */ | |
266bee88 | 868 | void nfs_destroy_directcache(void) |
1da177e4 | 869 | { |
1a1d92c1 | 870 | kmem_cache_destroy(nfs_direct_cachep); |
1da177e4 | 871 | } |