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 | ||
41 | #include <linux/config.h> | |
42 | #include <linux/errno.h> | |
43 | #include <linux/sched.h> | |
44 | #include <linux/kernel.h> | |
45 | #include <linux/smp_lock.h> | |
46 | #include <linux/file.h> | |
47 | #include <linux/pagemap.h> | |
48 | #include <linux/kref.h> | |
49 | ||
50 | #include <linux/nfs_fs.h> | |
51 | #include <linux/nfs_page.h> | |
52 | #include <linux/sunrpc/clnt.h> | |
53 | ||
54 | #include <asm/system.h> | |
55 | #include <asm/uaccess.h> | |
56 | #include <asm/atomic.h> | |
57 | ||
91d5b470 CL |
58 | #include "iostat.h" |
59 | ||
1da177e4 | 60 | #define NFSDBG_FACILITY NFSDBG_VFS |
1da177e4 | 61 | |
143f412e | 62 | static void nfs_free_user_pages(struct page **pages, int npages, int do_dirty); |
1da177e4 LT |
63 | static kmem_cache_t *nfs_direct_cachep; |
64 | ||
65 | /* | |
66 | * This represents a set of asynchronous requests that we're waiting on | |
67 | */ | |
68 | struct nfs_direct_req { | |
69 | struct kref kref; /* release manager */ | |
15ce4a0c CL |
70 | |
71 | /* I/O parameters */ | |
88467055 | 72 | struct list_head list; /* nfs_read/write_data structs */ |
99514f8f CL |
73 | struct file * filp; /* file descriptor */ |
74 | struct kiocb * iocb; /* controlling i/o request */ | |
1da177e4 | 75 | wait_queue_head_t wait; /* wait for i/o completion */ |
88467055 | 76 | struct inode * inode; /* target file of i/o */ |
1da177e4 LT |
77 | struct page ** pages; /* pages in our buffer */ |
78 | unsigned int npages; /* count of pages */ | |
15ce4a0c CL |
79 | |
80 | /* completion state */ | |
81 | spinlock_t lock; /* protect completion state */ | |
82 | int outstanding; /* i/os we're waiting for */ | |
83 | ssize_t count, /* bytes actually processed */ | |
1da177e4 LT |
84 | error; /* any reported error */ |
85 | }; | |
86 | ||
b8a32e2b CL |
87 | /** |
88 | * nfs_direct_IO - NFS address space operation for direct I/O | |
89 | * @rw: direction (read or write) | |
90 | * @iocb: target I/O control block | |
91 | * @iov: array of vectors that define I/O buffer | |
92 | * @pos: offset in file to begin the operation | |
93 | * @nr_segs: size of iovec array | |
94 | * | |
95 | * The presence of this routine in the address space ops vector means | |
96 | * the NFS client supports direct I/O. However, we shunt off direct | |
97 | * read and write requests before the VFS gets them, so this method | |
98 | * should never be called. | |
99 | */ | |
100 | ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs) | |
101 | { | |
102 | struct dentry *dentry = iocb->ki_filp->f_dentry; | |
103 | ||
104 | dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n", | |
105 | dentry->d_name.name, (long long) pos, nr_segs); | |
106 | ||
107 | return -EINVAL; | |
108 | } | |
109 | ||
d4cc948b | 110 | static inline int nfs_get_user_pages(int rw, unsigned long user_addr, size_t size, struct page ***pages) |
1da177e4 LT |
111 | { |
112 | int result = -ENOMEM; | |
113 | unsigned long page_count; | |
114 | size_t array_size; | |
115 | ||
1da177e4 LT |
116 | page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT; |
117 | page_count -= user_addr >> PAGE_SHIFT; | |
118 | ||
119 | array_size = (page_count * sizeof(struct page *)); | |
120 | *pages = kmalloc(array_size, GFP_KERNEL); | |
121 | if (*pages) { | |
122 | down_read(¤t->mm->mmap_sem); | |
123 | result = get_user_pages(current, current->mm, user_addr, | |
124 | page_count, (rw == READ), 0, | |
125 | *pages, NULL); | |
126 | up_read(¤t->mm->mmap_sem); | |
143f412e TM |
127 | /* |
128 | * If we got fewer pages than expected from get_user_pages(), | |
129 | * the user buffer runs off the end of a mapping; return EFAULT. | |
130 | */ | |
131 | if (result >= 0 && result < page_count) { | |
132 | nfs_free_user_pages(*pages, result, 0); | |
133 | *pages = NULL; | |
134 | result = -EFAULT; | |
135 | } | |
1da177e4 LT |
136 | } |
137 | return result; | |
138 | } | |
139 | ||
d4cc948b | 140 | static void nfs_free_user_pages(struct page **pages, int npages, int do_dirty) |
1da177e4 LT |
141 | { |
142 | int i; | |
143 | for (i = 0; i < npages; i++) { | |
566dd606 TM |
144 | struct page *page = pages[i]; |
145 | if (do_dirty && !PageCompound(page)) | |
146 | set_page_dirty_lock(page); | |
147 | page_cache_release(page); | |
1da177e4 LT |
148 | } |
149 | kfree(pages); | |
150 | } | |
151 | ||
93619e59 CL |
152 | static inline struct nfs_direct_req *nfs_direct_req_alloc(void) |
153 | { | |
154 | struct nfs_direct_req *dreq; | |
155 | ||
156 | dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL); | |
157 | if (!dreq) | |
158 | return NULL; | |
159 | ||
160 | kref_init(&dreq->kref); | |
161 | init_waitqueue_head(&dreq->wait); | |
162 | INIT_LIST_HEAD(&dreq->list); | |
163 | dreq->iocb = NULL; | |
15ce4a0c CL |
164 | spin_lock_init(&dreq->lock); |
165 | dreq->outstanding = 0; | |
166 | dreq->count = 0; | |
167 | dreq->error = 0; | |
93619e59 CL |
168 | |
169 | return dreq; | |
170 | } | |
171 | ||
1da177e4 LT |
172 | static void nfs_direct_req_release(struct kref *kref) |
173 | { | |
174 | struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); | |
175 | kmem_cache_free(nfs_direct_cachep, dreq); | |
176 | } | |
177 | ||
bc0fb201 CL |
178 | /* |
179 | * Collects and returns the final error value/byte-count. | |
180 | */ | |
181 | static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) | |
182 | { | |
15ce4a0c | 183 | ssize_t result = -EIOCBQUEUED; |
bc0fb201 CL |
184 | |
185 | /* Async requests don't wait here */ | |
186 | if (dreq->iocb) | |
187 | goto out; | |
188 | ||
15ce4a0c | 189 | result = wait_event_interruptible(dreq->wait, (dreq->outstanding == 0)); |
bc0fb201 CL |
190 | |
191 | if (!result) | |
15ce4a0c | 192 | result = dreq->error; |
bc0fb201 | 193 | if (!result) |
15ce4a0c | 194 | result = dreq->count; |
bc0fb201 CL |
195 | |
196 | out: | |
197 | kref_put(&dreq->kref, nfs_direct_req_release); | |
198 | return (ssize_t) result; | |
199 | } | |
200 | ||
63ab46ab CL |
201 | /* |
202 | * We must hold a reference to all the pages in this direct read request | |
203 | * until the RPCs complete. This could be long *after* we are woken up in | |
204 | * nfs_direct_wait (for instance, if someone hits ^C on a slow server). | |
205 | * | |
206 | * In addition, synchronous I/O uses a stack-allocated iocb. Thus we | |
207 | * can't trust the iocb is still valid here if this is a synchronous | |
208 | * request. If the waiter is woken prematurely, the iocb is long gone. | |
209 | */ | |
210 | static void nfs_direct_complete(struct nfs_direct_req *dreq) | |
211 | { | |
212 | nfs_free_user_pages(dreq->pages, dreq->npages, 1); | |
213 | ||
214 | if (dreq->iocb) { | |
15ce4a0c | 215 | long res = (long) dreq->error; |
63ab46ab | 216 | if (!res) |
15ce4a0c | 217 | res = (long) dreq->count; |
63ab46ab CL |
218 | aio_complete(dreq->iocb, res, 0); |
219 | } else | |
220 | wake_up(&dreq->wait); | |
221 | ||
a37ec012 | 222 | iput(dreq->inode); |
63ab46ab CL |
223 | kref_put(&dreq->kref, nfs_direct_req_release); |
224 | } | |
225 | ||
d4cc948b | 226 | /* |
1da177e4 LT |
227 | * Note we also set the number of requests we have in the dreq when we are |
228 | * done. This prevents races with I/O completion so we will always wait | |
229 | * until all requests have been dispatched and completed. | |
230 | */ | |
5dd602f2 | 231 | static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize) |
1da177e4 LT |
232 | { |
233 | struct list_head *list; | |
234 | struct nfs_direct_req *dreq; | |
40859d7e | 235 | unsigned int rpages = (rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT; |
1da177e4 | 236 | |
93619e59 | 237 | dreq = nfs_direct_req_alloc(); |
1da177e4 LT |
238 | if (!dreq) |
239 | return NULL; | |
240 | ||
1da177e4 LT |
241 | list = &dreq->list; |
242 | for(;;) { | |
40859d7e | 243 | struct nfs_read_data *data = nfs_readdata_alloc(rpages); |
1da177e4 LT |
244 | |
245 | if (unlikely(!data)) { | |
246 | while (!list_empty(list)) { | |
247 | data = list_entry(list->next, | |
248 | struct nfs_read_data, pages); | |
249 | list_del(&data->pages); | |
250 | nfs_readdata_free(data); | |
251 | } | |
252 | kref_put(&dreq->kref, nfs_direct_req_release); | |
253 | return NULL; | |
254 | } | |
255 | ||
256 | INIT_LIST_HEAD(&data->pages); | |
257 | list_add(&data->pages, list); | |
258 | ||
259 | data->req = (struct nfs_page *) dreq; | |
15ce4a0c | 260 | dreq->outstanding++; |
1da177e4 LT |
261 | if (nbytes <= rsize) |
262 | break; | |
263 | nbytes -= rsize; | |
264 | } | |
265 | kref_get(&dreq->kref); | |
1da177e4 LT |
266 | return dreq; |
267 | } | |
268 | ||
ec06c096 | 269 | static void nfs_direct_read_result(struct rpc_task *task, void *calldata) |
1da177e4 | 270 | { |
ec06c096 | 271 | struct nfs_read_data *data = calldata; |
1da177e4 LT |
272 | struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req; |
273 | ||
ec06c096 TM |
274 | if (nfs_readpage_result(task, data) != 0) |
275 | return; | |
15ce4a0c CL |
276 | |
277 | spin_lock(&dreq->lock); | |
278 | ||
ec06c096 | 279 | if (likely(task->tk_status >= 0)) |
15ce4a0c | 280 | dreq->count += data->res.count; |
1da177e4 | 281 | else |
15ce4a0c CL |
282 | dreq->error = task->tk_status; |
283 | ||
284 | if (--dreq->outstanding) { | |
285 | spin_unlock(&dreq->lock); | |
286 | return; | |
287 | } | |
1da177e4 | 288 | |
15ce4a0c CL |
289 | spin_unlock(&dreq->lock); |
290 | nfs_direct_complete(dreq); | |
1da177e4 LT |
291 | } |
292 | ||
ec06c096 TM |
293 | static const struct rpc_call_ops nfs_read_direct_ops = { |
294 | .rpc_call_done = nfs_direct_read_result, | |
295 | .rpc_release = nfs_readdata_release, | |
296 | }; | |
297 | ||
d4cc948b | 298 | /* |
1da177e4 LT |
299 | * For each nfs_read_data struct that was allocated on the list, dispatch |
300 | * an NFS READ operation | |
301 | */ | |
88467055 | 302 | static void nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos) |
1da177e4 | 303 | { |
99514f8f CL |
304 | struct file *file = dreq->filp; |
305 | struct inode *inode = file->f_mapping->host; | |
306 | struct nfs_open_context *ctx = (struct nfs_open_context *) | |
307 | file->private_data; | |
1da177e4 LT |
308 | struct list_head *list = &dreq->list; |
309 | struct page **pages = dreq->pages; | |
5dd602f2 | 310 | size_t rsize = NFS_SERVER(inode)->rsize; |
1da177e4 | 311 | unsigned int curpage, pgbase; |
1da177e4 LT |
312 | |
313 | curpage = 0; | |
314 | pgbase = user_addr & ~PAGE_MASK; | |
315 | do { | |
316 | struct nfs_read_data *data; | |
5dd602f2 | 317 | size_t bytes; |
1da177e4 LT |
318 | |
319 | bytes = rsize; | |
320 | if (count < rsize) | |
321 | bytes = count; | |
322 | ||
323 | data = list_entry(list->next, struct nfs_read_data, pages); | |
324 | list_del_init(&data->pages); | |
325 | ||
326 | data->inode = inode; | |
327 | data->cred = ctx->cred; | |
328 | data->args.fh = NFS_FH(inode); | |
329 | data->args.context = ctx; | |
88467055 | 330 | data->args.offset = pos; |
1da177e4 LT |
331 | data->args.pgbase = pgbase; |
332 | data->args.pages = &pages[curpage]; | |
333 | data->args.count = bytes; | |
334 | data->res.fattr = &data->fattr; | |
335 | data->res.eof = 0; | |
336 | data->res.count = bytes; | |
337 | ||
ec06c096 TM |
338 | rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC, |
339 | &nfs_read_direct_ops, data); | |
1da177e4 LT |
340 | NFS_PROTO(inode)->read_setup(data); |
341 | ||
342 | data->task.tk_cookie = (unsigned long) inode; | |
1da177e4 LT |
343 | |
344 | lock_kernel(); | |
345 | rpc_execute(&data->task); | |
346 | unlock_kernel(); | |
347 | ||
348 | dfprintk(VFS, "NFS: %4d initiated direct read call (req %s/%Ld, %u bytes @ offset %Lu)\n", | |
349 | data->task.tk_pid, | |
350 | inode->i_sb->s_id, | |
351 | (long long)NFS_FILEID(inode), | |
352 | bytes, | |
353 | (unsigned long long)data->args.offset); | |
354 | ||
88467055 | 355 | pos += bytes; |
1da177e4 LT |
356 | pgbase += bytes; |
357 | curpage += pgbase >> PAGE_SHIFT; | |
358 | pgbase &= ~PAGE_MASK; | |
359 | ||
360 | count -= bytes; | |
361 | } while (count != 0); | |
362 | } | |
363 | ||
88467055 | 364 | static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos, struct page **pages, unsigned int nr_pages) |
1da177e4 LT |
365 | { |
366 | ssize_t result; | |
367 | sigset_t oldset; | |
99514f8f | 368 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
1da177e4 LT |
369 | struct rpc_clnt *clnt = NFS_CLIENT(inode); |
370 | struct nfs_direct_req *dreq; | |
371 | ||
372 | dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize); | |
373 | if (!dreq) | |
374 | return -ENOMEM; | |
375 | ||
376 | dreq->pages = pages; | |
377 | dreq->npages = nr_pages; | |
a37ec012 | 378 | igrab(inode); |
91d5b470 | 379 | dreq->inode = inode; |
99514f8f | 380 | dreq->filp = iocb->ki_filp; |
487b8372 CL |
381 | if (!is_sync_kiocb(iocb)) |
382 | dreq->iocb = iocb; | |
1da177e4 | 383 | |
91d5b470 | 384 | nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count); |
1da177e4 | 385 | rpc_clnt_sigmask(clnt, &oldset); |
88467055 | 386 | nfs_direct_read_schedule(dreq, user_addr, count, pos); |
bc0fb201 | 387 | result = nfs_direct_wait(dreq); |
1da177e4 LT |
388 | rpc_clnt_sigunmask(clnt, &oldset); |
389 | ||
390 | return result; | |
391 | } | |
392 | ||
462d5b32 | 393 | static struct nfs_direct_req *nfs_direct_write_alloc(size_t nbytes, size_t wsize) |
1da177e4 | 394 | { |
462d5b32 CL |
395 | struct list_head *list; |
396 | struct nfs_direct_req *dreq; | |
462d5b32 | 397 | unsigned int wpages = (wsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT; |
1da177e4 | 398 | |
462d5b32 CL |
399 | dreq = nfs_direct_req_alloc(); |
400 | if (!dreq) | |
401 | return NULL; | |
402 | ||
403 | list = &dreq->list; | |
404 | for(;;) { | |
405 | struct nfs_write_data *data = nfs_writedata_alloc(wpages); | |
406 | ||
407 | if (unlikely(!data)) { | |
408 | while (!list_empty(list)) { | |
409 | data = list_entry(list->next, | |
410 | struct nfs_write_data, pages); | |
411 | list_del(&data->pages); | |
412 | nfs_writedata_free(data); | |
413 | } | |
414 | kref_put(&dreq->kref, nfs_direct_req_release); | |
415 | return NULL; | |
416 | } | |
417 | ||
418 | INIT_LIST_HEAD(&data->pages); | |
419 | list_add(&data->pages, list); | |
420 | ||
421 | data->req = (struct nfs_page *) dreq; | |
15ce4a0c | 422 | dreq->outstanding++; |
462d5b32 CL |
423 | if (nbytes <= wsize) |
424 | break; | |
425 | nbytes -= wsize; | |
426 | } | |
427 | kref_get(&dreq->kref); | |
462d5b32 CL |
428 | return dreq; |
429 | } | |
430 | ||
15ce4a0c CL |
431 | /* |
432 | * NB: Return the value of the first error return code. Subsequent | |
433 | * errors after the first one are ignored. | |
434 | */ | |
462d5b32 CL |
435 | static void nfs_direct_write_result(struct rpc_task *task, void *calldata) |
436 | { | |
437 | struct nfs_write_data *data = calldata; | |
438 | struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req; | |
439 | int status = task->tk_status; | |
440 | ||
441 | if (nfs_writeback_done(task, data) != 0) | |
442 | return; | |
443 | /* If the server fell back to an UNSTABLE write, it's an error. */ | |
444 | if (unlikely(data->res.verf->committed != NFS_FILE_SYNC)) | |
445 | status = -EIO; | |
446 | ||
15ce4a0c CL |
447 | spin_lock(&dreq->lock); |
448 | ||
462d5b32 | 449 | if (likely(status >= 0)) |
15ce4a0c | 450 | dreq->count += data->res.count; |
462d5b32 | 451 | else |
15ce4a0c | 452 | dreq->error = status; |
462d5b32 | 453 | |
15ce4a0c CL |
454 | if (--dreq->outstanding) { |
455 | spin_unlock(&dreq->lock); | |
456 | return; | |
9eafa8cc | 457 | } |
15ce4a0c CL |
458 | |
459 | spin_unlock(&dreq->lock); | |
460 | ||
461 | nfs_end_data_update(data->inode); | |
462 | nfs_direct_complete(dreq); | |
462d5b32 CL |
463 | } |
464 | ||
465 | static const struct rpc_call_ops nfs_write_direct_ops = { | |
466 | .rpc_call_done = nfs_direct_write_result, | |
467 | .rpc_release = nfs_writedata_release, | |
468 | }; | |
469 | ||
470 | /* | |
471 | * For each nfs_write_data struct that was allocated on the list, dispatch | |
472 | * an NFS WRITE operation | |
473 | * | |
474 | * XXX: For now, support only FILE_SYNC writes. Later we may add | |
475 | * support for UNSTABLE + COMMIT. | |
476 | */ | |
88467055 | 477 | static void nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos) |
462d5b32 | 478 | { |
c89f2ee5 CL |
479 | struct file *file = dreq->filp; |
480 | struct inode *inode = file->f_mapping->host; | |
481 | struct nfs_open_context *ctx = (struct nfs_open_context *) | |
482 | file->private_data; | |
462d5b32 CL |
483 | struct list_head *list = &dreq->list; |
484 | struct page **pages = dreq->pages; | |
485 | size_t wsize = NFS_SERVER(inode)->wsize; | |
486 | unsigned int curpage, pgbase; | |
1da177e4 | 487 | |
1da177e4 | 488 | curpage = 0; |
462d5b32 | 489 | pgbase = user_addr & ~PAGE_MASK; |
1da177e4 | 490 | do { |
462d5b32 CL |
491 | struct nfs_write_data *data; |
492 | size_t bytes; | |
493 | ||
494 | bytes = wsize; | |
495 | if (count < wsize) | |
496 | bytes = count; | |
497 | ||
498 | data = list_entry(list->next, struct nfs_write_data, pages); | |
499 | list_del_init(&data->pages); | |
500 | ||
501 | data->inode = inode; | |
502 | data->cred = ctx->cred; | |
503 | data->args.fh = NFS_FH(inode); | |
504 | data->args.context = ctx; | |
88467055 | 505 | data->args.offset = pos; |
462d5b32 CL |
506 | data->args.pgbase = pgbase; |
507 | data->args.pages = &pages[curpage]; | |
508 | data->args.count = bytes; | |
509 | data->res.fattr = &data->fattr; | |
510 | data->res.count = bytes; | |
47989d74 | 511 | data->res.verf = &data->verf; |
462d5b32 CL |
512 | |
513 | rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC, | |
514 | &nfs_write_direct_ops, data); | |
515 | NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE); | |
1da177e4 | 516 | |
462d5b32 CL |
517 | data->task.tk_priority = RPC_PRIORITY_NORMAL; |
518 | data->task.tk_cookie = (unsigned long) inode; | |
1da177e4 LT |
519 | |
520 | lock_kernel(); | |
462d5b32 | 521 | rpc_execute(&data->task); |
1da177e4 LT |
522 | unlock_kernel(); |
523 | ||
462d5b32 CL |
524 | dfprintk(VFS, "NFS: %4d initiated direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n", |
525 | data->task.tk_pid, | |
526 | inode->i_sb->s_id, | |
527 | (long long)NFS_FILEID(inode), | |
528 | bytes, | |
529 | (unsigned long long)data->args.offset); | |
1da177e4 | 530 | |
88467055 | 531 | pos += bytes; |
462d5b32 CL |
532 | pgbase += bytes; |
533 | curpage += pgbase >> PAGE_SHIFT; | |
534 | pgbase &= ~PAGE_MASK; | |
1da177e4 | 535 | |
462d5b32 CL |
536 | count -= bytes; |
537 | } while (count != 0); | |
538 | } | |
1da177e4 | 539 | |
88467055 | 540 | static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos, struct page **pages, int nr_pages) |
462d5b32 CL |
541 | { |
542 | ssize_t result; | |
543 | sigset_t oldset; | |
c89f2ee5 | 544 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
462d5b32 CL |
545 | struct rpc_clnt *clnt = NFS_CLIENT(inode); |
546 | struct nfs_direct_req *dreq; | |
1da177e4 | 547 | |
462d5b32 CL |
548 | dreq = nfs_direct_write_alloc(count, NFS_SERVER(inode)->wsize); |
549 | if (!dreq) | |
550 | return -ENOMEM; | |
1da177e4 | 551 | |
462d5b32 CL |
552 | dreq->pages = pages; |
553 | dreq->npages = nr_pages; | |
a37ec012 | 554 | igrab(inode); |
c89f2ee5 CL |
555 | dreq->inode = inode; |
556 | dreq->filp = iocb->ki_filp; | |
557 | if (!is_sync_kiocb(iocb)) | |
558 | dreq->iocb = iocb; | |
1da177e4 | 559 | |
47989d74 CL |
560 | nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count); |
561 | ||
462d5b32 | 562 | nfs_begin_data_update(inode); |
1da177e4 | 563 | |
462d5b32 | 564 | rpc_clnt_sigmask(clnt, &oldset); |
88467055 | 565 | nfs_direct_write_schedule(dreq, user_addr, count, pos); |
c89f2ee5 | 566 | result = nfs_direct_wait(dreq); |
462d5b32 | 567 | rpc_clnt_sigunmask(clnt, &oldset); |
1da177e4 | 568 | |
462d5b32 | 569 | return result; |
1da177e4 LT |
570 | } |
571 | ||
1da177e4 LT |
572 | /** |
573 | * nfs_file_direct_read - file direct read operation for NFS files | |
574 | * @iocb: target I/O control block | |
575 | * @buf: user's buffer into which to read data | |
88467055 CL |
576 | * @count: number of bytes to read |
577 | * @pos: byte offset in file where reading starts | |
1da177e4 LT |
578 | * |
579 | * We use this function for direct reads instead of calling | |
580 | * generic_file_aio_read() in order to avoid gfar's check to see if | |
581 | * the request starts before the end of the file. For that check | |
582 | * to work, we must generate a GETATTR before each direct read, and | |
583 | * even then there is a window between the GETATTR and the subsequent | |
88467055 | 584 | * READ where the file size could change. Our preference is simply |
1da177e4 LT |
585 | * to do all reads the application wants, and the server will take |
586 | * care of managing the end of file boundary. | |
88467055 | 587 | * |
1da177e4 LT |
588 | * This function also eliminates unnecessarily updating the file's |
589 | * atime locally, as the NFS server sets the file's atime, and this | |
590 | * client must read the updated atime from the server back into its | |
591 | * cache. | |
592 | */ | |
d4cc948b | 593 | ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos) |
1da177e4 LT |
594 | { |
595 | ssize_t retval = -EINVAL; | |
0cdd80d0 CL |
596 | int page_count; |
597 | struct page **pages; | |
1da177e4 | 598 | struct file *file = iocb->ki_filp; |
1da177e4 | 599 | struct address_space *mapping = file->f_mapping; |
1da177e4 | 600 | |
ce1a8e67 | 601 | dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n", |
0bbacc40 CL |
602 | file->f_dentry->d_parent->d_name.name, |
603 | file->f_dentry->d_name.name, | |
ce1a8e67 | 604 | (unsigned long) count, (long long) pos); |
1da177e4 | 605 | |
1da177e4 LT |
606 | if (count < 0) |
607 | goto out; | |
608 | retval = -EFAULT; | |
0cdd80d0 | 609 | if (!access_ok(VERIFY_WRITE, buf, count)) |
1da177e4 LT |
610 | goto out; |
611 | retval = 0; | |
612 | if (!count) | |
613 | goto out; | |
614 | ||
29884df0 TM |
615 | retval = nfs_sync_mapping(mapping); |
616 | if (retval) | |
617 | goto out; | |
1da177e4 | 618 | |
0cdd80d0 CL |
619 | page_count = nfs_get_user_pages(READ, (unsigned long) buf, |
620 | count, &pages); | |
621 | if (page_count < 0) { | |
622 | nfs_free_user_pages(pages, 0, 0); | |
623 | retval = page_count; | |
624 | goto out; | |
625 | } | |
626 | ||
99514f8f | 627 | retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos, |
0cdd80d0 | 628 | pages, page_count); |
1da177e4 | 629 | if (retval > 0) |
0cdd80d0 | 630 | iocb->ki_pos = pos + retval; |
1da177e4 LT |
631 | |
632 | out: | |
633 | return retval; | |
634 | } | |
635 | ||
636 | /** | |
637 | * nfs_file_direct_write - file direct write operation for NFS files | |
638 | * @iocb: target I/O control block | |
639 | * @buf: user's buffer from which to write data | |
88467055 CL |
640 | * @count: number of bytes to write |
641 | * @pos: byte offset in file where writing starts | |
1da177e4 LT |
642 | * |
643 | * We use this function for direct writes instead of calling | |
644 | * generic_file_aio_write() in order to avoid taking the inode | |
645 | * semaphore and updating the i_size. The NFS server will set | |
646 | * the new i_size and this client must read the updated size | |
647 | * back into its cache. We let the server do generic write | |
648 | * parameter checking and report problems. | |
649 | * | |
650 | * We also avoid an unnecessary invocation of generic_osync_inode(), | |
651 | * as it is fairly meaningless to sync the metadata of an NFS file. | |
652 | * | |
653 | * We eliminate local atime updates, see direct read above. | |
654 | * | |
655 | * We avoid unnecessary page cache invalidations for normal cached | |
656 | * readers of this file. | |
657 | * | |
658 | * Note that O_APPEND is not supported for NFS direct writes, as there | |
659 | * is no atomic O_APPEND write facility in the NFS protocol. | |
660 | */ | |
d4cc948b | 661 | ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos) |
1da177e4 | 662 | { |
ce1a8e67 | 663 | ssize_t retval; |
47989d74 CL |
664 | int page_count; |
665 | struct page **pages; | |
1da177e4 | 666 | struct file *file = iocb->ki_filp; |
1da177e4 | 667 | struct address_space *mapping = file->f_mapping; |
1da177e4 | 668 | |
ce1a8e67 | 669 | dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n", |
0bbacc40 | 670 | file->f_dentry->d_parent->d_name.name, |
ce1a8e67 CL |
671 | file->f_dentry->d_name.name, |
672 | (unsigned long) count, (long long) pos); | |
1da177e4 | 673 | |
ce1a8e67 CL |
674 | retval = generic_write_checks(file, &pos, &count, 0); |
675 | if (retval) | |
1da177e4 | 676 | goto out; |
ce1a8e67 CL |
677 | |
678 | retval = -EINVAL; | |
679 | if ((ssize_t) count < 0) | |
1da177e4 | 680 | goto out; |
1da177e4 LT |
681 | retval = 0; |
682 | if (!count) | |
683 | goto out; | |
ce1a8e67 CL |
684 | |
685 | retval = -EFAULT; | |
47989d74 | 686 | if (!access_ok(VERIFY_READ, buf, count)) |
ce1a8e67 | 687 | goto out; |
1da177e4 | 688 | |
29884df0 TM |
689 | retval = nfs_sync_mapping(mapping); |
690 | if (retval) | |
691 | goto out; | |
1da177e4 | 692 | |
47989d74 CL |
693 | page_count = nfs_get_user_pages(WRITE, (unsigned long) buf, |
694 | count, &pages); | |
695 | if (page_count < 0) { | |
696 | nfs_free_user_pages(pages, 0, 0); | |
697 | retval = page_count; | |
698 | goto out; | |
699 | } | |
700 | ||
c89f2ee5 | 701 | retval = nfs_direct_write(iocb, (unsigned long) buf, count, |
47989d74 | 702 | pos, pages, page_count); |
9eafa8cc CL |
703 | |
704 | /* | |
705 | * XXX: nfs_end_data_update() already ensures this file's | |
706 | * cached data is subsequently invalidated. Do we really | |
707 | * need to call invalidate_inode_pages2() again here? | |
708 | * | |
709 | * For aio writes, this invalidation will almost certainly | |
710 | * occur before the writes complete. Kind of racey. | |
711 | */ | |
1da177e4 LT |
712 | if (mapping->nrpages) |
713 | invalidate_inode_pages2(mapping); | |
9eafa8cc | 714 | |
1da177e4 | 715 | if (retval > 0) |
ce1a8e67 | 716 | iocb->ki_pos = pos + retval; |
1da177e4 LT |
717 | |
718 | out: | |
719 | return retval; | |
720 | } | |
721 | ||
88467055 CL |
722 | /** |
723 | * nfs_init_directcache - create a slab cache for nfs_direct_req structures | |
724 | * | |
725 | */ | |
1da177e4 LT |
726 | int nfs_init_directcache(void) |
727 | { | |
728 | nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", | |
729 | sizeof(struct nfs_direct_req), | |
730 | 0, SLAB_RECLAIM_ACCOUNT, | |
731 | NULL, NULL); | |
732 | if (nfs_direct_cachep == NULL) | |
733 | return -ENOMEM; | |
734 | ||
735 | return 0; | |
736 | } | |
737 | ||
88467055 CL |
738 | /** |
739 | * nfs_init_directcache - destroy the slab cache for nfs_direct_req structures | |
740 | * | |
741 | */ | |
1da177e4 LT |
742 | void nfs_destroy_directcache(void) |
743 | { | |
744 | if (kmem_cache_destroy(nfs_direct_cachep)) | |
745 | printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n"); | |
746 | } |