]>
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 | 58 | #include "iostat.h" |
1763da12 | 59 | #include "pnfs.h" |
91d5b470 | 60 | |
1da177e4 | 61 | #define NFSDBG_FACILITY NFSDBG_VFS |
1da177e4 | 62 | |
e18b890b | 63 | static struct kmem_cache *nfs_direct_cachep; |
1da177e4 LT |
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 */ | |
a8881f5a | 72 | struct nfs_open_context *ctx; /* file open context info */ |
f11ac8db | 73 | struct nfs_lock_context *l_ctx; /* Lock context info */ |
99514f8f | 74 | struct kiocb * iocb; /* controlling i/o request */ |
88467055 | 75 | struct inode * inode; /* target file of i/o */ |
15ce4a0c CL |
76 | |
77 | /* completion state */ | |
607f31e8 | 78 | atomic_t io_count; /* i/os we're waiting for */ |
15ce4a0c | 79 | spinlock_t lock; /* protect completion state */ |
15ce4a0c | 80 | ssize_t count, /* bytes actually processed */ |
1da177e4 | 81 | error; /* any reported error */ |
d72b7a6b | 82 | struct completion completion; /* wait for i/o completion */ |
fad61490 TM |
83 | |
84 | /* commit state */ | |
1763da12 FI |
85 | struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo */ |
86 | struct pnfs_ds_commit_info ds_cinfo; /* Storage for cinfo */ | |
87 | struct work_struct work; | |
fad61490 TM |
88 | int flags; |
89 | #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */ | |
90 | #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */ | |
91 | struct nfs_writeverf verf; /* unstable write verifier */ | |
1da177e4 LT |
92 | }; |
93 | ||
1763da12 FI |
94 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops; |
95 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops; | |
fad61490 | 96 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode); |
1763da12 | 97 | static void nfs_direct_write_schedule_work(struct work_struct *work); |
607f31e8 TM |
98 | |
99 | static inline void get_dreq(struct nfs_direct_req *dreq) | |
100 | { | |
101 | atomic_inc(&dreq->io_count); | |
102 | } | |
103 | ||
104 | static inline int put_dreq(struct nfs_direct_req *dreq) | |
105 | { | |
106 | return atomic_dec_and_test(&dreq->io_count); | |
107 | } | |
108 | ||
1da177e4 | 109 | /** |
b8a32e2b CL |
110 | * nfs_direct_IO - NFS address space operation for direct I/O |
111 | * @rw: direction (read or write) | |
112 | * @iocb: target I/O control block | |
113 | * @iov: array of vectors that define I/O buffer | |
114 | * @pos: offset in file to begin the operation | |
115 | * @nr_segs: size of iovec array | |
116 | * | |
117 | * The presence of this routine in the address space ops vector means | |
118 | * the NFS client supports direct I/O. However, we shunt off direct | |
119 | * read and write requests before the VFS gets them, so this method | |
120 | * should never be called. | |
1da177e4 | 121 | */ |
b8a32e2b CL |
122 | ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs) |
123 | { | |
b8a32e2b | 124 | dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n", |
01cce933 | 125 | iocb->ki_filp->f_path.dentry->d_name.name, |
e99170ff | 126 | (long long) pos, nr_segs); |
b8a32e2b CL |
127 | |
128 | return -EINVAL; | |
129 | } | |
130 | ||
749e146e | 131 | static void nfs_direct_release_pages(struct page **pages, unsigned int npages) |
9c93ab7d | 132 | { |
749e146e | 133 | unsigned int i; |
607f31e8 TM |
134 | for (i = 0; i < npages; i++) |
135 | page_cache_release(pages[i]); | |
6b45d858 TM |
136 | } |
137 | ||
1763da12 FI |
138 | void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo, |
139 | struct nfs_direct_req *dreq) | |
140 | { | |
141 | cinfo->lock = &dreq->lock; | |
142 | cinfo->mds = &dreq->mds_cinfo; | |
143 | cinfo->ds = &dreq->ds_cinfo; | |
144 | cinfo->dreq = dreq; | |
145 | cinfo->completion_ops = &nfs_direct_commit_completion_ops; | |
146 | } | |
147 | ||
93619e59 | 148 | static inline struct nfs_direct_req *nfs_direct_req_alloc(void) |
1da177e4 | 149 | { |
93619e59 CL |
150 | struct nfs_direct_req *dreq; |
151 | ||
e94b1766 | 152 | dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL); |
93619e59 CL |
153 | if (!dreq) |
154 | return NULL; | |
155 | ||
156 | kref_init(&dreq->kref); | |
607f31e8 | 157 | kref_get(&dreq->kref); |
d72b7a6b | 158 | init_completion(&dreq->completion); |
1763da12 FI |
159 | dreq->mds_cinfo.ncommit = 0; |
160 | atomic_set(&dreq->mds_cinfo.rpcs_out, 0); | |
161 | INIT_LIST_HEAD(&dreq->mds_cinfo.list); | |
162 | INIT_WORK(&dreq->work, nfs_direct_write_schedule_work); | |
163 | memset(&dreq->ds_cinfo, 0, sizeof(dreq->ds_cinfo)); | |
93619e59 | 164 | dreq->iocb = NULL; |
a8881f5a | 165 | dreq->ctx = NULL; |
f11ac8db | 166 | dreq->l_ctx = NULL; |
15ce4a0c | 167 | spin_lock_init(&dreq->lock); |
607f31e8 | 168 | atomic_set(&dreq->io_count, 0); |
15ce4a0c CL |
169 | dreq->count = 0; |
170 | dreq->error = 0; | |
fad61490 | 171 | dreq->flags = 0; |
93619e59 CL |
172 | |
173 | return dreq; | |
1da177e4 LT |
174 | } |
175 | ||
b4946ffb | 176 | static void nfs_direct_req_free(struct kref *kref) |
1da177e4 LT |
177 | { |
178 | struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); | |
a8881f5a | 179 | |
f11ac8db TM |
180 | if (dreq->l_ctx != NULL) |
181 | nfs_put_lock_context(dreq->l_ctx); | |
a8881f5a TM |
182 | if (dreq->ctx != NULL) |
183 | put_nfs_open_context(dreq->ctx); | |
1da177e4 LT |
184 | kmem_cache_free(nfs_direct_cachep, dreq); |
185 | } | |
186 | ||
b4946ffb TM |
187 | static void nfs_direct_req_release(struct nfs_direct_req *dreq) |
188 | { | |
189 | kref_put(&dreq->kref, nfs_direct_req_free); | |
190 | } | |
191 | ||
bc0fb201 CL |
192 | /* |
193 | * Collects and returns the final error value/byte-count. | |
194 | */ | |
195 | static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) | |
196 | { | |
15ce4a0c | 197 | ssize_t result = -EIOCBQUEUED; |
bc0fb201 CL |
198 | |
199 | /* Async requests don't wait here */ | |
200 | if (dreq->iocb) | |
201 | goto out; | |
202 | ||
150030b7 | 203 | result = wait_for_completion_killable(&dreq->completion); |
bc0fb201 CL |
204 | |
205 | if (!result) | |
15ce4a0c | 206 | result = dreq->error; |
bc0fb201 | 207 | if (!result) |
15ce4a0c | 208 | result = dreq->count; |
bc0fb201 CL |
209 | |
210 | out: | |
bc0fb201 CL |
211 | return (ssize_t) result; |
212 | } | |
213 | ||
63ab46ab | 214 | /* |
607f31e8 TM |
215 | * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust |
216 | * the iocb is still valid here if this is a synchronous request. | |
63ab46ab CL |
217 | */ |
218 | static void nfs_direct_complete(struct nfs_direct_req *dreq) | |
219 | { | |
63ab46ab | 220 | if (dreq->iocb) { |
15ce4a0c | 221 | long res = (long) dreq->error; |
63ab46ab | 222 | if (!res) |
15ce4a0c | 223 | res = (long) dreq->count; |
63ab46ab | 224 | aio_complete(dreq->iocb, res, 0); |
d72b7a6b TM |
225 | } |
226 | complete_all(&dreq->completion); | |
63ab46ab | 227 | |
b4946ffb | 228 | nfs_direct_req_release(dreq); |
63ab46ab CL |
229 | } |
230 | ||
584aa810 | 231 | void nfs_direct_readpage_release(struct nfs_page *req) |
1da177e4 | 232 | { |
584aa810 FI |
233 | dprintk("NFS: direct read done (%s/%lld %d@%lld)\n", |
234 | req->wb_context->dentry->d_inode->i_sb->s_id, | |
235 | (long long)NFS_FILEID(req->wb_context->dentry->d_inode), | |
236 | req->wb_bytes, | |
237 | (long long)req_offset(req)); | |
238 | nfs_release_request(req); | |
fdd1e74c TM |
239 | } |
240 | ||
584aa810 | 241 | static void nfs_direct_read_completion(struct nfs_pgio_header *hdr) |
fdd1e74c | 242 | { |
584aa810 FI |
243 | unsigned long bytes = 0; |
244 | struct nfs_direct_req *dreq = hdr->dreq; | |
fdd1e74c | 245 | |
584aa810 FI |
246 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) |
247 | goto out_put; | |
15ce4a0c CL |
248 | |
249 | spin_lock(&dreq->lock); | |
584aa810 FI |
250 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0)) |
251 | dreq->error = hdr->error; | |
252 | else | |
253 | dreq->count += hdr->good_bytes; | |
254 | spin_unlock(&dreq->lock); | |
255 | ||
256 | if (!test_bit(NFS_IOHDR_ERROR, &hdr->flags)) { | |
257 | while (!list_empty(&hdr->pages)) { | |
258 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
259 | struct page *page = req->wb_page; | |
260 | ||
261 | if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) { | |
262 | if (bytes > hdr->good_bytes) | |
263 | zero_user(page, 0, PAGE_SIZE); | |
264 | else if (hdr->good_bytes - bytes < PAGE_SIZE) | |
265 | zero_user_segment(page, | |
266 | hdr->good_bytes & ~PAGE_MASK, | |
267 | PAGE_SIZE); | |
268 | } | |
269 | bytes += req->wb_bytes; | |
270 | nfs_list_remove_request(req); | |
271 | nfs_direct_readpage_release(req); | |
272 | if (!PageCompound(page)) | |
273 | set_page_dirty(page); | |
274 | page_cache_release(page); | |
275 | } | |
d4a8f367 | 276 | } else { |
584aa810 FI |
277 | while (!list_empty(&hdr->pages)) { |
278 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
279 | ||
280 | if (bytes < hdr->good_bytes) | |
281 | if (!PageCompound(req->wb_page)) | |
282 | set_page_dirty(req->wb_page); | |
283 | bytes += req->wb_bytes; | |
284 | page_cache_release(req->wb_page); | |
285 | nfs_list_remove_request(req); | |
286 | nfs_direct_readpage_release(req); | |
287 | } | |
d4a8f367 | 288 | } |
584aa810 | 289 | out_put: |
607f31e8 TM |
290 | if (put_dreq(dreq)) |
291 | nfs_direct_complete(dreq); | |
584aa810 | 292 | hdr->release(hdr); |
1da177e4 LT |
293 | } |
294 | ||
584aa810 | 295 | static void nfs_sync_pgio_error(struct list_head *head) |
cd841605 | 296 | { |
584aa810 | 297 | struct nfs_page *req; |
cd841605 | 298 | |
584aa810 FI |
299 | while (!list_empty(head)) { |
300 | req = nfs_list_entry(head->next); | |
301 | nfs_list_remove_request(req); | |
302 | nfs_release_request(req); | |
303 | } | |
cd841605 FI |
304 | } |
305 | ||
584aa810 FI |
306 | static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr) |
307 | { | |
308 | get_dreq(hdr->dreq); | |
309 | } | |
310 | ||
311 | static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = { | |
312 | .error_cleanup = nfs_sync_pgio_error, | |
313 | .init_hdr = nfs_direct_pgio_init, | |
314 | .completion = nfs_direct_read_completion, | |
315 | }; | |
316 | ||
d4cc948b | 317 | /* |
607f31e8 TM |
318 | * For each rsize'd chunk of the user's buffer, dispatch an NFS READ |
319 | * operation. If nfs_readdata_alloc() or get_user_pages() fails, | |
320 | * bail and stop sending more reads. Read length accounting is | |
321 | * handled automatically by nfs_direct_read_result(). Otherwise, if | |
322 | * no requests have been sent, just return an error. | |
1da177e4 | 323 | */ |
584aa810 | 324 | static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc, |
02fe4946 CL |
325 | const struct iovec *iov, |
326 | loff_t pos) | |
1da177e4 | 327 | { |
584aa810 | 328 | struct nfs_direct_req *dreq = desc->pg_dreq; |
a8881f5a | 329 | struct nfs_open_context *ctx = dreq->ctx; |
3d4ff43d | 330 | struct inode *inode = ctx->dentry->d_inode; |
02fe4946 CL |
331 | unsigned long user_addr = (unsigned long)iov->iov_base; |
332 | size_t count = iov->iov_len; | |
5dd602f2 | 333 | size_t rsize = NFS_SERVER(inode)->rsize; |
607f31e8 TM |
334 | unsigned int pgbase; |
335 | int result; | |
336 | ssize_t started = 0; | |
584aa810 FI |
337 | struct page **pagevec = NULL; |
338 | unsigned int npages; | |
607f31e8 | 339 | |
1da177e4 | 340 | do { |
5dd602f2 | 341 | size_t bytes; |
584aa810 | 342 | int i; |
1da177e4 | 343 | |
e9f7bee1 | 344 | pgbase = user_addr & ~PAGE_MASK; |
584aa810 | 345 | bytes = min(max(rsize, PAGE_SIZE), count); |
e9f7bee1 | 346 | |
607f31e8 | 347 | result = -ENOMEM; |
584aa810 FI |
348 | npages = nfs_page_array_len(pgbase, bytes); |
349 | if (!pagevec) | |
350 | pagevec = kmalloc(npages * sizeof(struct page *), | |
351 | GFP_KERNEL); | |
352 | if (!pagevec) | |
4db6e0b7 | 353 | break; |
607f31e8 TM |
354 | down_read(¤t->mm->mmap_sem); |
355 | result = get_user_pages(current, current->mm, user_addr, | |
584aa810 | 356 | npages, 1, 0, pagevec, NULL); |
607f31e8 | 357 | up_read(¤t->mm->mmap_sem); |
584aa810 | 358 | if (result < 0) |
749e146e | 359 | break; |
584aa810 | 360 | if ((unsigned)result < npages) { |
d9df8d6b TM |
361 | bytes = result * PAGE_SIZE; |
362 | if (bytes <= pgbase) { | |
584aa810 | 363 | nfs_direct_release_pages(pagevec, result); |
d9df8d6b TM |
364 | break; |
365 | } | |
366 | bytes -= pgbase; | |
584aa810 | 367 | npages = result; |
607f31e8 TM |
368 | } |
369 | ||
584aa810 FI |
370 | for (i = 0; i < npages; i++) { |
371 | struct nfs_page *req; | |
372 | unsigned int req_len = min(bytes, PAGE_SIZE - pgbase); | |
373 | /* XXX do we need to do the eof zeroing found in async_filler? */ | |
374 | req = nfs_create_request(dreq->ctx, dreq->inode, | |
375 | pagevec[i], | |
376 | pgbase, req_len); | |
377 | if (IS_ERR(req)) { | |
378 | nfs_direct_release_pages(pagevec + i, | |
379 | npages - i); | |
380 | result = PTR_ERR(req); | |
381 | break; | |
382 | } | |
383 | req->wb_index = pos >> PAGE_SHIFT; | |
384 | req->wb_offset = pos & ~PAGE_MASK; | |
385 | if (!nfs_pageio_add_request(desc, req)) { | |
386 | result = desc->pg_error; | |
387 | nfs_release_request(req); | |
388 | nfs_direct_release_pages(pagevec + i, | |
389 | npages - i); | |
390 | break; | |
391 | } | |
392 | pgbase = 0; | |
393 | bytes -= req_len; | |
394 | started += req_len; | |
395 | user_addr += req_len; | |
396 | pos += req_len; | |
397 | count -= req_len; | |
398 | } | |
71e8cc00 | 399 | } while (count != 0 && result >= 0); |
607f31e8 | 400 | |
584aa810 FI |
401 | kfree(pagevec); |
402 | ||
607f31e8 | 403 | if (started) |
c216fd70 | 404 | return started; |
607f31e8 | 405 | return result < 0 ? (ssize_t) result : -EFAULT; |
1da177e4 LT |
406 | } |
407 | ||
19f73787 CL |
408 | static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq, |
409 | const struct iovec *iov, | |
410 | unsigned long nr_segs, | |
411 | loff_t pos) | |
412 | { | |
584aa810 | 413 | struct nfs_pageio_descriptor desc; |
19f73787 CL |
414 | ssize_t result = -EINVAL; |
415 | size_t requested_bytes = 0; | |
416 | unsigned long seg; | |
417 | ||
584aa810 FI |
418 | nfs_pageio_init_read(&desc, dreq->inode, |
419 | &nfs_direct_read_completion_ops); | |
19f73787 | 420 | get_dreq(dreq); |
584aa810 | 421 | desc.pg_dreq = dreq; |
19f73787 CL |
422 | |
423 | for (seg = 0; seg < nr_segs; seg++) { | |
424 | const struct iovec *vec = &iov[seg]; | |
584aa810 | 425 | result = nfs_direct_read_schedule_segment(&desc, vec, pos); |
19f73787 CL |
426 | if (result < 0) |
427 | break; | |
428 | requested_bytes += result; | |
429 | if ((size_t)result < vec->iov_len) | |
430 | break; | |
431 | pos += vec->iov_len; | |
432 | } | |
433 | ||
584aa810 FI |
434 | nfs_pageio_complete(&desc); |
435 | ||
839f7ad6 CL |
436 | /* |
437 | * If no bytes were started, return the error, and let the | |
438 | * generic layer handle the completion. | |
439 | */ | |
440 | if (requested_bytes == 0) { | |
441 | nfs_direct_req_release(dreq); | |
442 | return result < 0 ? result : -EIO; | |
443 | } | |
444 | ||
19f73787 CL |
445 | if (put_dreq(dreq)) |
446 | nfs_direct_complete(dreq); | |
839f7ad6 | 447 | return 0; |
19f73787 CL |
448 | } |
449 | ||
c216fd70 CL |
450 | static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov, |
451 | unsigned long nr_segs, loff_t pos) | |
1da177e4 | 452 | { |
f11ac8db | 453 | ssize_t result = -ENOMEM; |
99514f8f | 454 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
1da177e4 LT |
455 | struct nfs_direct_req *dreq; |
456 | ||
607f31e8 | 457 | dreq = nfs_direct_req_alloc(); |
f11ac8db TM |
458 | if (dreq == NULL) |
459 | goto out; | |
1da177e4 | 460 | |
91d5b470 | 461 | dreq->inode = inode; |
cd3758e3 | 462 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
f11ac8db TM |
463 | dreq->l_ctx = nfs_get_lock_context(dreq->ctx); |
464 | if (dreq->l_ctx == NULL) | |
465 | goto out_release; | |
487b8372 CL |
466 | if (!is_sync_kiocb(iocb)) |
467 | dreq->iocb = iocb; | |
1da177e4 | 468 | |
c216fd70 | 469 | result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos); |
607f31e8 TM |
470 | if (!result) |
471 | result = nfs_direct_wait(dreq); | |
f11ac8db | 472 | out_release: |
b4946ffb | 473 | nfs_direct_req_release(dreq); |
f11ac8db | 474 | out: |
1da177e4 LT |
475 | return result; |
476 | } | |
477 | ||
fad61490 TM |
478 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) |
479 | static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) | |
480 | { | |
1763da12 FI |
481 | struct nfs_pageio_descriptor desc; |
482 | struct nfs_page *req, *tmp; | |
483 | LIST_HEAD(reqs); | |
484 | struct nfs_commit_info cinfo; | |
485 | LIST_HEAD(failed); | |
486 | ||
487 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
488 | pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo); | |
489 | spin_lock(cinfo.lock); | |
490 | nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0); | |
491 | spin_unlock(cinfo.lock); | |
1da177e4 | 492 | |
fad61490 | 493 | dreq->count = 0; |
607f31e8 TM |
494 | get_dreq(dreq); |
495 | ||
1763da12 FI |
496 | nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, |
497 | &nfs_direct_write_completion_ops); | |
498 | desc.pg_dreq = dreq; | |
fedb595c | 499 | |
1763da12 FI |
500 | list_for_each_entry_safe(req, tmp, &reqs, wb_list) { |
501 | if (!nfs_pageio_add_request(&desc, req)) { | |
502 | nfs_list_add_request(req, &failed); | |
503 | spin_lock(cinfo.lock); | |
504 | dreq->flags = 0; | |
505 | dreq->error = -EIO; | |
506 | spin_unlock(cinfo.lock); | |
507 | } | |
508 | } | |
509 | nfs_pageio_complete(&desc); | |
fad61490 | 510 | |
1763da12 FI |
511 | while (!list_empty(&failed)) { |
512 | page_cache_release(req->wb_page); | |
513 | nfs_release_request(req); | |
514 | nfs_unlock_request(req); | |
515 | } | |
fad61490 | 516 | |
1763da12 FI |
517 | if (put_dreq(dreq)) |
518 | nfs_direct_write_complete(dreq, dreq->inode); | |
c9d8f89d TM |
519 | } |
520 | ||
1763da12 | 521 | static void nfs_direct_commit_complete(struct nfs_commit_data *data) |
c9d8f89d | 522 | { |
0b7c0153 | 523 | struct nfs_direct_req *dreq = data->dreq; |
1763da12 FI |
524 | struct nfs_commit_info cinfo; |
525 | struct nfs_page *req; | |
c9d8f89d TM |
526 | int status = data->task.tk_status; |
527 | ||
1763da12 | 528 | nfs_init_cinfo_from_dreq(&cinfo, dreq); |
c9d8f89d | 529 | if (status < 0) { |
60fa3f76 | 530 | dprintk("NFS: %5u commit failed with error %d.\n", |
1763da12 | 531 | data->task.tk_pid, status); |
fad61490 | 532 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
60fa3f76 | 533 | } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) { |
c9d8f89d | 534 | dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid); |
fad61490 | 535 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
1da177e4 LT |
536 | } |
537 | ||
c9d8f89d | 538 | dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status); |
1763da12 FI |
539 | while (!list_empty(&data->pages)) { |
540 | req = nfs_list_entry(data->pages.next); | |
541 | nfs_list_remove_request(req); | |
542 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) { | |
543 | /* Note the rewrite will go through mds */ | |
544 | nfs_mark_request_commit(req, NULL, &cinfo); | |
545 | } else { | |
546 | page_cache_release(req->wb_page); | |
547 | nfs_release_request(req); | |
548 | } | |
549 | nfs_unlock_request(req); | |
550 | } | |
551 | ||
552 | if (atomic_dec_and_test(&cinfo.mds->rpcs_out)) | |
553 | nfs_direct_write_complete(dreq, data->inode); | |
1da177e4 LT |
554 | } |
555 | ||
1763da12 FI |
556 | static void nfs_direct_error_cleanup(struct nfs_inode *nfsi) |
557 | { | |
558 | /* There is no lock to clear */ | |
559 | } | |
560 | ||
561 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = { | |
562 | .completion = nfs_direct_commit_complete, | |
563 | .error_cleanup = nfs_direct_error_cleanup, | |
fad61490 TM |
564 | }; |
565 | ||
566 | static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) | |
1da177e4 | 567 | { |
1763da12 FI |
568 | int res; |
569 | struct nfs_commit_info cinfo; | |
570 | LIST_HEAD(mds_list); | |
571 | ||
572 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
573 | nfs_scan_commit(dreq->inode, &mds_list, &cinfo); | |
574 | res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo); | |
575 | if (res < 0) /* res == -ENOMEM */ | |
576 | nfs_direct_write_reschedule(dreq); | |
fad61490 | 577 | } |
1da177e4 | 578 | |
1763da12 | 579 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
fad61490 | 580 | { |
1763da12 | 581 | struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work); |
fad61490 | 582 | int flags = dreq->flags; |
1da177e4 | 583 | |
fad61490 TM |
584 | dreq->flags = 0; |
585 | switch (flags) { | |
586 | case NFS_ODIRECT_DO_COMMIT: | |
587 | nfs_direct_commit_schedule(dreq); | |
1da177e4 | 588 | break; |
fad61490 TM |
589 | case NFS_ODIRECT_RESCHED_WRITES: |
590 | nfs_direct_write_reschedule(dreq); | |
591 | break; | |
592 | default: | |
1763da12 | 593 | nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping); |
fad61490 TM |
594 | nfs_direct_complete(dreq); |
595 | } | |
596 | } | |
1da177e4 | 597 | |
1763da12 | 598 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
fad61490 | 599 | { |
1763da12 | 600 | schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */ |
fad61490 | 601 | } |
1763da12 | 602 | |
fad61490 | 603 | #else |
24fc9211 BS |
604 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
605 | { | |
606 | } | |
1da177e4 | 607 | |
fad61490 TM |
608 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
609 | { | |
cd9ae2b6 | 610 | nfs_zap_mapping(inode, inode->i_mapping); |
fad61490 TM |
611 | nfs_direct_complete(dreq); |
612 | } | |
613 | #endif | |
1da177e4 | 614 | |
c9d8f89d TM |
615 | /* |
616 | * NB: Return the value of the first error return code. Subsequent | |
617 | * errors after the first one are ignored. | |
618 | */ | |
462d5b32 | 619 | /* |
607f31e8 TM |
620 | * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE |
621 | * operation. If nfs_writedata_alloc() or get_user_pages() fails, | |
622 | * bail and stop sending more writes. Write length accounting is | |
623 | * handled automatically by nfs_direct_write_result(). Otherwise, if | |
624 | * no requests have been sent, just return an error. | |
462d5b32 | 625 | */ |
1763da12 | 626 | static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc, |
02fe4946 | 627 | const struct iovec *iov, |
1763da12 | 628 | loff_t pos) |
462d5b32 | 629 | { |
1763da12 | 630 | struct nfs_direct_req *dreq = desc->pg_dreq; |
a8881f5a | 631 | struct nfs_open_context *ctx = dreq->ctx; |
3d4ff43d | 632 | struct inode *inode = ctx->dentry->d_inode; |
02fe4946 CL |
633 | unsigned long user_addr = (unsigned long)iov->iov_base; |
634 | size_t count = iov->iov_len; | |
462d5b32 | 635 | size_t wsize = NFS_SERVER(inode)->wsize; |
607f31e8 TM |
636 | unsigned int pgbase; |
637 | int result; | |
638 | ssize_t started = 0; | |
1763da12 FI |
639 | struct page **pagevec = NULL; |
640 | unsigned int npages; | |
82b145c5 | 641 | |
1da177e4 | 642 | do { |
462d5b32 | 643 | size_t bytes; |
1763da12 | 644 | int i; |
462d5b32 | 645 | |
e9f7bee1 | 646 | pgbase = user_addr & ~PAGE_MASK; |
1763da12 | 647 | bytes = min(max(wsize, PAGE_SIZE), count); |
e9f7bee1 | 648 | |
607f31e8 | 649 | result = -ENOMEM; |
1763da12 FI |
650 | npages = nfs_page_array_len(pgbase, bytes); |
651 | if (!pagevec) | |
652 | pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL); | |
653 | if (!pagevec) | |
607f31e8 TM |
654 | break; |
655 | ||
607f31e8 TM |
656 | down_read(¤t->mm->mmap_sem); |
657 | result = get_user_pages(current, current->mm, user_addr, | |
1763da12 | 658 | npages, 0, 0, pagevec, NULL); |
607f31e8 | 659 | up_read(¤t->mm->mmap_sem); |
1763da12 | 660 | if (result < 0) |
749e146e | 661 | break; |
1763da12 FI |
662 | |
663 | if ((unsigned)result < npages) { | |
d9df8d6b TM |
664 | bytes = result * PAGE_SIZE; |
665 | if (bytes <= pgbase) { | |
1763da12 | 666 | nfs_direct_release_pages(pagevec, result); |
d9df8d6b TM |
667 | break; |
668 | } | |
669 | bytes -= pgbase; | |
1763da12 | 670 | npages = result; |
607f31e8 TM |
671 | } |
672 | ||
1763da12 FI |
673 | for (i = 0; i < npages; i++) { |
674 | struct nfs_page *req; | |
675 | unsigned int req_len = min(bytes, PAGE_SIZE - pgbase); | |
1da177e4 | 676 | |
1763da12 FI |
677 | req = nfs_create_request(dreq->ctx, dreq->inode, |
678 | pagevec[i], | |
679 | pgbase, req_len); | |
680 | if (IS_ERR(req)) { | |
681 | nfs_direct_release_pages(pagevec + i, | |
682 | npages - i); | |
683 | result = PTR_ERR(req); | |
684 | break; | |
685 | } | |
686 | nfs_lock_request(req); | |
687 | req->wb_index = pos >> PAGE_SHIFT; | |
688 | req->wb_offset = pos & ~PAGE_MASK; | |
689 | if (!nfs_pageio_add_request(desc, req)) { | |
690 | result = desc->pg_error; | |
691 | nfs_unlock_request(req); | |
692 | nfs_release_request(req); | |
693 | nfs_direct_release_pages(pagevec + i, | |
694 | npages - i); | |
71e8cc00 | 695 | break; |
1763da12 FI |
696 | } |
697 | pgbase = 0; | |
698 | bytes -= req_len; | |
699 | started += req_len; | |
700 | user_addr += req_len; | |
701 | pos += req_len; | |
702 | count -= req_len; | |
703 | } | |
71e8cc00 | 704 | } while (count != 0 && result >= 0); |
607f31e8 | 705 | |
1763da12 FI |
706 | kfree(pagevec); |
707 | ||
607f31e8 | 708 | if (started) |
c216fd70 | 709 | return started; |
607f31e8 | 710 | return result < 0 ? (ssize_t) result : -EFAULT; |
462d5b32 | 711 | } |
1da177e4 | 712 | |
1763da12 FI |
713 | static void nfs_direct_write_completion(struct nfs_pgio_header *hdr) |
714 | { | |
715 | struct nfs_direct_req *dreq = hdr->dreq; | |
716 | struct nfs_commit_info cinfo; | |
717 | int bit = -1; | |
718 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
719 | ||
720 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) | |
721 | goto out_put; | |
722 | ||
723 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
724 | ||
725 | spin_lock(&dreq->lock); | |
726 | ||
727 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) { | |
728 | dreq->flags = 0; | |
729 | dreq->error = hdr->error; | |
730 | } | |
731 | if (dreq->error != 0) | |
732 | bit = NFS_IOHDR_ERROR; | |
733 | else { | |
734 | dreq->count += hdr->good_bytes; | |
735 | if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) { | |
736 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; | |
737 | bit = NFS_IOHDR_NEED_RESCHED; | |
738 | } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) { | |
739 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) | |
740 | bit = NFS_IOHDR_NEED_RESCHED; | |
741 | else if (dreq->flags == 0) { | |
742 | memcpy(&dreq->verf, &req->wb_verf, | |
743 | sizeof(dreq->verf)); | |
744 | bit = NFS_IOHDR_NEED_COMMIT; | |
745 | dreq->flags = NFS_ODIRECT_DO_COMMIT; | |
746 | } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) { | |
747 | if (memcmp(&dreq->verf, &req->wb_verf, sizeof(dreq->verf))) { | |
748 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; | |
749 | bit = NFS_IOHDR_NEED_RESCHED; | |
750 | } else | |
751 | bit = NFS_IOHDR_NEED_COMMIT; | |
752 | } | |
753 | } | |
754 | } | |
755 | spin_unlock(&dreq->lock); | |
756 | ||
757 | while (!list_empty(&hdr->pages)) { | |
758 | req = nfs_list_entry(hdr->pages.next); | |
759 | nfs_list_remove_request(req); | |
760 | switch (bit) { | |
761 | case NFS_IOHDR_NEED_RESCHED: | |
762 | case NFS_IOHDR_NEED_COMMIT: | |
763 | nfs_mark_request_commit(req, hdr->lseg, &cinfo); | |
764 | break; | |
765 | default: | |
766 | page_cache_release(req->wb_page); | |
767 | nfs_release_request(req); | |
768 | } | |
769 | nfs_unlock_request(req); | |
770 | } | |
771 | ||
772 | out_put: | |
773 | if (put_dreq(dreq)) | |
774 | nfs_direct_write_complete(dreq, hdr->inode); | |
775 | hdr->release(hdr); | |
776 | } | |
777 | ||
778 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = { | |
779 | .error_cleanup = nfs_sync_pgio_error, | |
780 | .init_hdr = nfs_direct_pgio_init, | |
781 | .completion = nfs_direct_write_completion, | |
782 | }; | |
783 | ||
19f73787 CL |
784 | static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, |
785 | const struct iovec *iov, | |
786 | unsigned long nr_segs, | |
1763da12 | 787 | loff_t pos) |
19f73787 | 788 | { |
1763da12 | 789 | struct nfs_pageio_descriptor desc; |
19f73787 CL |
790 | ssize_t result = 0; |
791 | size_t requested_bytes = 0; | |
792 | unsigned long seg; | |
793 | ||
1763da12 FI |
794 | nfs_pageio_init_write(&desc, dreq->inode, FLUSH_COND_STABLE, |
795 | &nfs_direct_write_completion_ops); | |
796 | desc.pg_dreq = dreq; | |
19f73787 CL |
797 | get_dreq(dreq); |
798 | ||
799 | for (seg = 0; seg < nr_segs; seg++) { | |
800 | const struct iovec *vec = &iov[seg]; | |
1763da12 | 801 | result = nfs_direct_write_schedule_segment(&desc, vec, pos); |
19f73787 CL |
802 | if (result < 0) |
803 | break; | |
804 | requested_bytes += result; | |
805 | if ((size_t)result < vec->iov_len) | |
806 | break; | |
807 | pos += vec->iov_len; | |
808 | } | |
1763da12 | 809 | nfs_pageio_complete(&desc); |
19f73787 | 810 | |
839f7ad6 CL |
811 | /* |
812 | * If no bytes were started, return the error, and let the | |
813 | * generic layer handle the completion. | |
814 | */ | |
815 | if (requested_bytes == 0) { | |
816 | nfs_direct_req_release(dreq); | |
817 | return result < 0 ? result : -EIO; | |
818 | } | |
819 | ||
19f73787 CL |
820 | if (put_dreq(dreq)) |
821 | nfs_direct_write_complete(dreq, dreq->inode); | |
839f7ad6 | 822 | return 0; |
19f73787 CL |
823 | } |
824 | ||
c216fd70 CL |
825 | static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov, |
826 | unsigned long nr_segs, loff_t pos, | |
827 | size_t count) | |
462d5b32 | 828 | { |
f11ac8db | 829 | ssize_t result = -ENOMEM; |
c89f2ee5 | 830 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
462d5b32 | 831 | struct nfs_direct_req *dreq; |
1da177e4 | 832 | |
607f31e8 | 833 | dreq = nfs_direct_req_alloc(); |
462d5b32 | 834 | if (!dreq) |
f11ac8db | 835 | goto out; |
1da177e4 | 836 | |
c89f2ee5 | 837 | dreq->inode = inode; |
cd3758e3 | 838 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
f11ac8db | 839 | dreq->l_ctx = nfs_get_lock_context(dreq->ctx); |
568a810d | 840 | if (dreq->l_ctx == NULL) |
f11ac8db | 841 | goto out_release; |
c89f2ee5 CL |
842 | if (!is_sync_kiocb(iocb)) |
843 | dreq->iocb = iocb; | |
1da177e4 | 844 | |
1763da12 | 845 | result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos); |
607f31e8 TM |
846 | if (!result) |
847 | result = nfs_direct_wait(dreq); | |
f11ac8db | 848 | out_release: |
b4946ffb | 849 | nfs_direct_req_release(dreq); |
f11ac8db | 850 | out: |
1da177e4 LT |
851 | return result; |
852 | } | |
853 | ||
854 | /** | |
855 | * nfs_file_direct_read - file direct read operation for NFS files | |
856 | * @iocb: target I/O control block | |
027445c3 BP |
857 | * @iov: vector of user buffers into which to read data |
858 | * @nr_segs: size of iov vector | |
88467055 | 859 | * @pos: byte offset in file where reading starts |
1da177e4 LT |
860 | * |
861 | * We use this function for direct reads instead of calling | |
862 | * generic_file_aio_read() in order to avoid gfar's check to see if | |
863 | * the request starts before the end of the file. For that check | |
864 | * to work, we must generate a GETATTR before each direct read, and | |
865 | * even then there is a window between the GETATTR and the subsequent | |
88467055 | 866 | * READ where the file size could change. Our preference is simply |
1da177e4 LT |
867 | * to do all reads the application wants, and the server will take |
868 | * care of managing the end of file boundary. | |
88467055 | 869 | * |
1da177e4 LT |
870 | * This function also eliminates unnecessarily updating the file's |
871 | * atime locally, as the NFS server sets the file's atime, and this | |
872 | * client must read the updated atime from the server back into its | |
873 | * cache. | |
874 | */ | |
027445c3 BP |
875 | ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov, |
876 | unsigned long nr_segs, loff_t pos) | |
1da177e4 LT |
877 | { |
878 | ssize_t retval = -EINVAL; | |
1da177e4 | 879 | struct file *file = iocb->ki_filp; |
1da177e4 | 880 | struct address_space *mapping = file->f_mapping; |
c216fd70 CL |
881 | size_t count; |
882 | ||
883 | count = iov_length(iov, nr_segs); | |
884 | nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); | |
1da177e4 | 885 | |
6da24bc9 | 886 | dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n", |
01cce933 JJS |
887 | file->f_path.dentry->d_parent->d_name.name, |
888 | file->f_path.dentry->d_name.name, | |
c216fd70 | 889 | count, (long long) pos); |
1da177e4 | 890 | |
1da177e4 LT |
891 | retval = 0; |
892 | if (!count) | |
893 | goto out; | |
894 | ||
29884df0 TM |
895 | retval = nfs_sync_mapping(mapping); |
896 | if (retval) | |
897 | goto out; | |
1da177e4 | 898 | |
7ec10f26 KK |
899 | task_io_account_read(count); |
900 | ||
c216fd70 | 901 | retval = nfs_direct_read(iocb, iov, nr_segs, pos); |
1da177e4 | 902 | if (retval > 0) |
0cdd80d0 | 903 | iocb->ki_pos = pos + retval; |
1da177e4 LT |
904 | |
905 | out: | |
906 | return retval; | |
907 | } | |
908 | ||
909 | /** | |
910 | * nfs_file_direct_write - file direct write operation for NFS files | |
911 | * @iocb: target I/O control block | |
027445c3 BP |
912 | * @iov: vector of user buffers from which to write data |
913 | * @nr_segs: size of iov vector | |
88467055 | 914 | * @pos: byte offset in file where writing starts |
1da177e4 LT |
915 | * |
916 | * We use this function for direct writes instead of calling | |
917 | * generic_file_aio_write() in order to avoid taking the inode | |
918 | * semaphore and updating the i_size. The NFS server will set | |
919 | * the new i_size and this client must read the updated size | |
920 | * back into its cache. We let the server do generic write | |
921 | * parameter checking and report problems. | |
922 | * | |
1da177e4 LT |
923 | * We eliminate local atime updates, see direct read above. |
924 | * | |
925 | * We avoid unnecessary page cache invalidations for normal cached | |
926 | * readers of this file. | |
927 | * | |
928 | * Note that O_APPEND is not supported for NFS direct writes, as there | |
929 | * is no atomic O_APPEND write facility in the NFS protocol. | |
930 | */ | |
027445c3 BP |
931 | ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov, |
932 | unsigned long nr_segs, loff_t pos) | |
1da177e4 | 933 | { |
070ea602 | 934 | ssize_t retval = -EINVAL; |
1da177e4 | 935 | struct file *file = iocb->ki_filp; |
1da177e4 | 936 | struct address_space *mapping = file->f_mapping; |
c216fd70 | 937 | size_t count; |
1da177e4 | 938 | |
c216fd70 CL |
939 | count = iov_length(iov, nr_segs); |
940 | nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count); | |
941 | ||
6da24bc9 | 942 | dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n", |
01cce933 JJS |
943 | file->f_path.dentry->d_parent->d_name.name, |
944 | file->f_path.dentry->d_name.name, | |
c216fd70 | 945 | count, (long long) pos); |
027445c3 | 946 | |
ce1a8e67 CL |
947 | retval = generic_write_checks(file, &pos, &count, 0); |
948 | if (retval) | |
1da177e4 | 949 | goto out; |
ce1a8e67 CL |
950 | |
951 | retval = -EINVAL; | |
952 | if ((ssize_t) count < 0) | |
1da177e4 | 953 | goto out; |
1da177e4 LT |
954 | retval = 0; |
955 | if (!count) | |
956 | goto out; | |
ce1a8e67 | 957 | |
29884df0 TM |
958 | retval = nfs_sync_mapping(mapping); |
959 | if (retval) | |
960 | goto out; | |
1da177e4 | 961 | |
7ec10f26 KK |
962 | task_io_account_write(count); |
963 | ||
c216fd70 | 964 | retval = nfs_direct_write(iocb, iov, nr_segs, pos, count); |
1763da12 FI |
965 | if (retval > 0) { |
966 | struct inode *inode = mapping->host; | |
9eafa8cc | 967 | |
ce1a8e67 | 968 | iocb->ki_pos = pos + retval; |
1763da12 FI |
969 | spin_lock(&inode->i_lock); |
970 | if (i_size_read(inode) < iocb->ki_pos) | |
971 | i_size_write(inode, iocb->ki_pos); | |
972 | spin_unlock(&inode->i_lock); | |
973 | } | |
1da177e4 LT |
974 | out: |
975 | return retval; | |
976 | } | |
977 | ||
88467055 CL |
978 | /** |
979 | * nfs_init_directcache - create a slab cache for nfs_direct_req structures | |
980 | * | |
981 | */ | |
f7b422b1 | 982 | int __init nfs_init_directcache(void) |
1da177e4 LT |
983 | { |
984 | nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", | |
985 | sizeof(struct nfs_direct_req), | |
fffb60f9 PJ |
986 | 0, (SLAB_RECLAIM_ACCOUNT| |
987 | SLAB_MEM_SPREAD), | |
20c2df83 | 988 | NULL); |
1da177e4 LT |
989 | if (nfs_direct_cachep == NULL) |
990 | return -ENOMEM; | |
991 | ||
992 | return 0; | |
993 | } | |
994 | ||
88467055 | 995 | /** |
f7b422b1 | 996 | * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures |
88467055 CL |
997 | * |
998 | */ | |
266bee88 | 999 | void nfs_destroy_directcache(void) |
1da177e4 | 1000 | { |
1a1d92c1 | 1001 | kmem_cache_destroy(nfs_direct_cachep); |
1da177e4 | 1002 | } |