]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/fs/nfs/direct.c | |
3 | * | |
4 | * Copyright (C) 2003 by Chuck Lever <cel@netapp.com> | |
5 | * | |
6 | * High-performance uncached I/O for the Linux NFS client | |
7 | * | |
8 | * There are important applications whose performance or correctness | |
9 | * depends on uncached access to file data. Database clusters | |
88467055 | 10 | * (multiple copies of the same instance running on separate hosts) |
1da177e4 | 11 | * implement their own cache coherency protocol that subsumes file |
88467055 CL |
12 | * system cache protocols. Applications that process datasets |
13 | * considerably larger than the client's memory do not always benefit | |
14 | * from a local cache. A streaming video server, for instance, has no | |
1da177e4 LT |
15 | * need to cache the contents of a file. |
16 | * | |
17 | * When an application requests uncached I/O, all read and write requests | |
18 | * are made directly to the server; data stored or fetched via these | |
19 | * requests is not cached in the Linux page cache. The client does not | |
20 | * correct unaligned requests from applications. All requested bytes are | |
21 | * held on permanent storage before a direct write system call returns to | |
22 | * an application. | |
23 | * | |
24 | * Solaris implements an uncached I/O facility called directio() that | |
25 | * is used for backups and sequential I/O to very large files. Solaris | |
26 | * also supports uncaching whole NFS partitions with "-o forcedirectio," | |
27 | * an undocumented mount option. | |
28 | * | |
29 | * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with | |
30 | * help from Andrew Morton. | |
31 | * | |
32 | * 18 Dec 2001 Initial implementation for 2.4 --cel | |
33 | * 08 Jul 2002 Version for 2.4.19, with bug fixes --trondmy | |
34 | * 08 Jun 2003 Port to 2.5 APIs --cel | |
35 | * 31 Mar 2004 Handle direct I/O without VFS support --cel | |
36 | * 15 Sep 2004 Parallel async reads --cel | |
88467055 | 37 | * 04 May 2005 support O_DIRECT with aio --cel |
1da177e4 LT |
38 | * |
39 | */ | |
40 | ||
1da177e4 LT |
41 | #include <linux/errno.h> |
42 | #include <linux/sched.h> | |
43 | #include <linux/kernel.h> | |
1da177e4 LT |
44 | #include <linux/file.h> |
45 | #include <linux/pagemap.h> | |
46 | #include <linux/kref.h> | |
5a0e3ad6 | 47 | #include <linux/slab.h> |
7ec10f26 | 48 | #include <linux/task_io_accounting_ops.h> |
6296556f | 49 | #include <linux/module.h> |
1da177e4 LT |
50 | |
51 | #include <linux/nfs_fs.h> | |
52 | #include <linux/nfs_page.h> | |
53 | #include <linux/sunrpc/clnt.h> | |
54 | ||
1da177e4 | 55 | #include <asm/uaccess.h> |
60063497 | 56 | #include <linux/atomic.h> |
1da177e4 | 57 | |
8d5658c9 | 58 | #include "internal.h" |
91d5b470 | 59 | #include "iostat.h" |
1763da12 | 60 | #include "pnfs.h" |
91d5b470 | 61 | |
1da177e4 | 62 | #define NFSDBG_FACILITY NFSDBG_VFS |
1da177e4 | 63 | |
e18b890b | 64 | static struct kmem_cache *nfs_direct_cachep; |
1da177e4 LT |
65 | |
66 | /* | |
67 | * This represents a set of asynchronous requests that we're waiting on | |
68 | */ | |
69 | struct nfs_direct_req { | |
70 | struct kref kref; /* release manager */ | |
15ce4a0c CL |
71 | |
72 | /* I/O parameters */ | |
a8881f5a | 73 | struct nfs_open_context *ctx; /* file open context info */ |
f11ac8db | 74 | struct nfs_lock_context *l_ctx; /* Lock context info */ |
99514f8f | 75 | struct kiocb * iocb; /* controlling i/o request */ |
88467055 | 76 | struct inode * inode; /* target file of i/o */ |
15ce4a0c CL |
77 | |
78 | /* completion state */ | |
607f31e8 | 79 | atomic_t io_count; /* i/os we're waiting for */ |
15ce4a0c | 80 | spinlock_t lock; /* protect completion state */ |
15ce4a0c | 81 | ssize_t count, /* bytes actually processed */ |
35754bc0 | 82 | bytes_left, /* bytes left to be sent */ |
1da177e4 | 83 | error; /* any reported error */ |
d72b7a6b | 84 | struct completion completion; /* wait for i/o completion */ |
fad61490 TM |
85 | |
86 | /* commit state */ | |
1763da12 FI |
87 | struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo */ |
88 | struct pnfs_ds_commit_info ds_cinfo; /* Storage for cinfo */ | |
89 | struct work_struct work; | |
fad61490 TM |
90 | int flags; |
91 | #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */ | |
92 | #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */ | |
93 | struct nfs_writeverf verf; /* unstable write verifier */ | |
1da177e4 LT |
94 | }; |
95 | ||
1763da12 FI |
96 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops; |
97 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops; | |
fad61490 | 98 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode); |
1763da12 | 99 | static void nfs_direct_write_schedule_work(struct work_struct *work); |
607f31e8 TM |
100 | |
101 | static inline void get_dreq(struct nfs_direct_req *dreq) | |
102 | { | |
103 | atomic_inc(&dreq->io_count); | |
104 | } | |
105 | ||
106 | static inline int put_dreq(struct nfs_direct_req *dreq) | |
107 | { | |
108 | return atomic_dec_and_test(&dreq->io_count); | |
109 | } | |
110 | ||
1da177e4 | 111 | /** |
b8a32e2b CL |
112 | * nfs_direct_IO - NFS address space operation for direct I/O |
113 | * @rw: direction (read or write) | |
114 | * @iocb: target I/O control block | |
115 | * @iov: array of vectors that define I/O buffer | |
116 | * @pos: offset in file to begin the operation | |
117 | * @nr_segs: size of iovec array | |
118 | * | |
119 | * The presence of this routine in the address space ops vector means | |
a564b8f0 MG |
120 | * the NFS client supports direct I/O. However, for most direct IO, we |
121 | * shunt off direct read and write requests before the VFS gets them, | |
122 | * so this method is only ever called for swap. | |
1da177e4 | 123 | */ |
b8a32e2b CL |
124 | ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs) |
125 | { | |
a564b8f0 | 126 | #ifndef CONFIG_NFS_SWAP |
6de1472f AV |
127 | dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n", |
128 | iocb->ki_filp, (long long) pos, nr_segs); | |
b8a32e2b CL |
129 | |
130 | return -EINVAL; | |
a564b8f0 | 131 | #else |
a564b8f0 MG |
132 | VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE); |
133 | ||
134 | if (rw == READ || rw == KERNEL_READ) | |
135 | return nfs_file_direct_read(iocb, iov, nr_segs, pos, | |
136 | rw == READ ? true : false); | |
137 | return nfs_file_direct_write(iocb, iov, nr_segs, pos, | |
138 | rw == WRITE ? true : false); | |
139 | #endif /* CONFIG_NFS_SWAP */ | |
b8a32e2b CL |
140 | } |
141 | ||
749e146e | 142 | static void nfs_direct_release_pages(struct page **pages, unsigned int npages) |
9c93ab7d | 143 | { |
749e146e | 144 | unsigned int i; |
607f31e8 TM |
145 | for (i = 0; i < npages; i++) |
146 | page_cache_release(pages[i]); | |
6b45d858 TM |
147 | } |
148 | ||
1763da12 FI |
149 | void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo, |
150 | struct nfs_direct_req *dreq) | |
151 | { | |
152 | cinfo->lock = &dreq->lock; | |
153 | cinfo->mds = &dreq->mds_cinfo; | |
154 | cinfo->ds = &dreq->ds_cinfo; | |
155 | cinfo->dreq = dreq; | |
156 | cinfo->completion_ops = &nfs_direct_commit_completion_ops; | |
157 | } | |
158 | ||
93619e59 | 159 | static inline struct nfs_direct_req *nfs_direct_req_alloc(void) |
1da177e4 | 160 | { |
93619e59 CL |
161 | struct nfs_direct_req *dreq; |
162 | ||
292f3eee | 163 | dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL); |
93619e59 CL |
164 | if (!dreq) |
165 | return NULL; | |
166 | ||
167 | kref_init(&dreq->kref); | |
607f31e8 | 168 | kref_get(&dreq->kref); |
d72b7a6b | 169 | init_completion(&dreq->completion); |
1763da12 FI |
170 | INIT_LIST_HEAD(&dreq->mds_cinfo.list); |
171 | INIT_WORK(&dreq->work, nfs_direct_write_schedule_work); | |
15ce4a0c | 172 | spin_lock_init(&dreq->lock); |
93619e59 CL |
173 | |
174 | return dreq; | |
1da177e4 LT |
175 | } |
176 | ||
b4946ffb | 177 | static void nfs_direct_req_free(struct kref *kref) |
1da177e4 LT |
178 | { |
179 | struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); | |
a8881f5a | 180 | |
f11ac8db TM |
181 | if (dreq->l_ctx != NULL) |
182 | nfs_put_lock_context(dreq->l_ctx); | |
a8881f5a TM |
183 | if (dreq->ctx != NULL) |
184 | put_nfs_open_context(dreq->ctx); | |
1da177e4 LT |
185 | kmem_cache_free(nfs_direct_cachep, dreq); |
186 | } | |
187 | ||
b4946ffb TM |
188 | static void nfs_direct_req_release(struct nfs_direct_req *dreq) |
189 | { | |
190 | kref_put(&dreq->kref, nfs_direct_req_free); | |
191 | } | |
192 | ||
6296556f PT |
193 | ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq) |
194 | { | |
195 | return dreq->bytes_left; | |
196 | } | |
197 | EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left); | |
198 | ||
bc0fb201 CL |
199 | /* |
200 | * Collects and returns the final error value/byte-count. | |
201 | */ | |
202 | static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) | |
203 | { | |
15ce4a0c | 204 | ssize_t result = -EIOCBQUEUED; |
bc0fb201 CL |
205 | |
206 | /* Async requests don't wait here */ | |
207 | if (dreq->iocb) | |
208 | goto out; | |
209 | ||
150030b7 | 210 | result = wait_for_completion_killable(&dreq->completion); |
bc0fb201 CL |
211 | |
212 | if (!result) | |
15ce4a0c | 213 | result = dreq->error; |
bc0fb201 | 214 | if (!result) |
15ce4a0c | 215 | result = dreq->count; |
bc0fb201 CL |
216 | |
217 | out: | |
bc0fb201 CL |
218 | return (ssize_t) result; |
219 | } | |
220 | ||
63ab46ab | 221 | /* |
607f31e8 TM |
222 | * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust |
223 | * the iocb is still valid here if this is a synchronous request. | |
63ab46ab | 224 | */ |
9811cd57 | 225 | static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write) |
63ab46ab | 226 | { |
9811cd57 CH |
227 | struct inode *inode = dreq->inode; |
228 | ||
2a009ec9 | 229 | if (dreq->iocb && write) { |
9811cd57 | 230 | loff_t pos = dreq->iocb->ki_pos + dreq->count; |
2a009ec9 CH |
231 | |
232 | spin_lock(&inode->i_lock); | |
233 | if (i_size_read(inode) < pos) | |
234 | i_size_write(inode, pos); | |
235 | spin_unlock(&inode->i_lock); | |
236 | } | |
237 | ||
1f90ee27 | 238 | if (write) |
2a009ec9 | 239 | nfs_zap_mapping(inode, inode->i_mapping); |
1f90ee27 CH |
240 | |
241 | inode_dio_done(inode); | |
2a009ec9 CH |
242 | |
243 | if (dreq->iocb) { | |
15ce4a0c | 244 | long res = (long) dreq->error; |
63ab46ab | 245 | if (!res) |
15ce4a0c | 246 | res = (long) dreq->count; |
63ab46ab | 247 | aio_complete(dreq->iocb, res, 0); |
d72b7a6b | 248 | } |
2a009ec9 | 249 | |
d72b7a6b | 250 | complete_all(&dreq->completion); |
63ab46ab | 251 | |
b4946ffb | 252 | nfs_direct_req_release(dreq); |
63ab46ab CL |
253 | } |
254 | ||
1385b811 | 255 | static void nfs_direct_readpage_release(struct nfs_page *req) |
1da177e4 | 256 | { |
1e8968c5 | 257 | dprintk("NFS: direct read done (%s/%llu %d@%lld)\n", |
584aa810 | 258 | req->wb_context->dentry->d_inode->i_sb->s_id, |
1e8968c5 | 259 | (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode), |
584aa810 FI |
260 | req->wb_bytes, |
261 | (long long)req_offset(req)); | |
262 | nfs_release_request(req); | |
fdd1e74c TM |
263 | } |
264 | ||
584aa810 | 265 | static void nfs_direct_read_completion(struct nfs_pgio_header *hdr) |
fdd1e74c | 266 | { |
584aa810 FI |
267 | unsigned long bytes = 0; |
268 | struct nfs_direct_req *dreq = hdr->dreq; | |
fdd1e74c | 269 | |
584aa810 FI |
270 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) |
271 | goto out_put; | |
15ce4a0c CL |
272 | |
273 | spin_lock(&dreq->lock); | |
584aa810 FI |
274 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0)) |
275 | dreq->error = hdr->error; | |
276 | else | |
277 | dreq->count += hdr->good_bytes; | |
278 | spin_unlock(&dreq->lock); | |
279 | ||
4bd8b010 TM |
280 | while (!list_empty(&hdr->pages)) { |
281 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
282 | struct page *page = req->wb_page; | |
283 | ||
be7e9858 JL |
284 | if (!PageCompound(page) && bytes < hdr->good_bytes) |
285 | set_page_dirty(page); | |
4bd8b010 TM |
286 | bytes += req->wb_bytes; |
287 | nfs_list_remove_request(req); | |
288 | nfs_direct_readpage_release(req); | |
d4a8f367 | 289 | } |
584aa810 | 290 | out_put: |
607f31e8 | 291 | if (put_dreq(dreq)) |
9811cd57 | 292 | nfs_direct_complete(dreq, false); |
584aa810 | 293 | hdr->release(hdr); |
1da177e4 LT |
294 | } |
295 | ||
3e9e0ca3 | 296 | static void nfs_read_sync_pgio_error(struct list_head *head) |
cd841605 | 297 | { |
584aa810 | 298 | struct nfs_page *req; |
cd841605 | 299 | |
584aa810 FI |
300 | while (!list_empty(head)) { |
301 | req = nfs_list_entry(head->next); | |
302 | nfs_list_remove_request(req); | |
303 | nfs_release_request(req); | |
304 | } | |
cd841605 FI |
305 | } |
306 | ||
584aa810 FI |
307 | static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr) |
308 | { | |
309 | get_dreq(hdr->dreq); | |
310 | } | |
311 | ||
312 | static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = { | |
3e9e0ca3 | 313 | .error_cleanup = nfs_read_sync_pgio_error, |
584aa810 FI |
314 | .init_hdr = nfs_direct_pgio_init, |
315 | .completion = nfs_direct_read_completion, | |
316 | }; | |
317 | ||
d4cc948b | 318 | /* |
607f31e8 TM |
319 | * For each rsize'd chunk of the user's buffer, dispatch an NFS READ |
320 | * operation. If nfs_readdata_alloc() or get_user_pages() fails, | |
321 | * bail and stop sending more reads. Read length accounting is | |
322 | * handled automatically by nfs_direct_read_result(). Otherwise, if | |
323 | * no requests have been sent, just return an error. | |
1da177e4 | 324 | */ |
584aa810 | 325 | static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc, |
02fe4946 | 326 | const struct iovec *iov, |
a564b8f0 | 327 | loff_t pos, bool uio) |
1da177e4 | 328 | { |
584aa810 | 329 | struct nfs_direct_req *dreq = desc->pg_dreq; |
a8881f5a | 330 | struct nfs_open_context *ctx = dreq->ctx; |
3d4ff43d | 331 | struct inode *inode = ctx->dentry->d_inode; |
02fe4946 CL |
332 | unsigned long user_addr = (unsigned long)iov->iov_base; |
333 | size_t count = iov->iov_len; | |
5dd602f2 | 334 | size_t rsize = NFS_SERVER(inode)->rsize; |
607f31e8 TM |
335 | unsigned int pgbase; |
336 | int result; | |
337 | ssize_t started = 0; | |
584aa810 FI |
338 | struct page **pagevec = NULL; |
339 | unsigned int npages; | |
607f31e8 | 340 | |
1da177e4 | 341 | do { |
5dd602f2 | 342 | size_t bytes; |
584aa810 | 343 | int i; |
1da177e4 | 344 | |
e9f7bee1 | 345 | pgbase = user_addr & ~PAGE_MASK; |
bf5fc402 | 346 | bytes = min(max_t(size_t, rsize, PAGE_SIZE), count); |
e9f7bee1 | 347 | |
607f31e8 | 348 | result = -ENOMEM; |
584aa810 FI |
349 | npages = nfs_page_array_len(pgbase, bytes); |
350 | if (!pagevec) | |
351 | pagevec = kmalloc(npages * sizeof(struct page *), | |
352 | GFP_KERNEL); | |
353 | if (!pagevec) | |
4db6e0b7 | 354 | break; |
a564b8f0 MG |
355 | if (uio) { |
356 | down_read(¤t->mm->mmap_sem); | |
357 | result = get_user_pages(current, current->mm, user_addr, | |
584aa810 | 358 | npages, 1, 0, pagevec, NULL); |
a564b8f0 MG |
359 | up_read(¤t->mm->mmap_sem); |
360 | if (result < 0) | |
361 | break; | |
362 | } else { | |
363 | WARN_ON(npages != 1); | |
364 | result = get_kernel_page(user_addr, 1, pagevec); | |
365 | if (WARN_ON(result != 1)) | |
366 | break; | |
367 | } | |
368 | ||
584aa810 | 369 | if ((unsigned)result < npages) { |
d9df8d6b TM |
370 | bytes = result * PAGE_SIZE; |
371 | if (bytes <= pgbase) { | |
584aa810 | 372 | nfs_direct_release_pages(pagevec, result); |
d9df8d6b TM |
373 | break; |
374 | } | |
375 | bytes -= pgbase; | |
584aa810 | 376 | npages = result; |
607f31e8 TM |
377 | } |
378 | ||
584aa810 FI |
379 | for (i = 0; i < npages; i++) { |
380 | struct nfs_page *req; | |
bf5fc402 | 381 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); |
584aa810 FI |
382 | /* XXX do we need to do the eof zeroing found in async_filler? */ |
383 | req = nfs_create_request(dreq->ctx, dreq->inode, | |
384 | pagevec[i], | |
385 | pgbase, req_len); | |
386 | if (IS_ERR(req)) { | |
584aa810 FI |
387 | result = PTR_ERR(req); |
388 | break; | |
389 | } | |
390 | req->wb_index = pos >> PAGE_SHIFT; | |
391 | req->wb_offset = pos & ~PAGE_MASK; | |
392 | if (!nfs_pageio_add_request(desc, req)) { | |
393 | result = desc->pg_error; | |
394 | nfs_release_request(req); | |
584aa810 FI |
395 | break; |
396 | } | |
397 | pgbase = 0; | |
398 | bytes -= req_len; | |
399 | started += req_len; | |
400 | user_addr += req_len; | |
401 | pos += req_len; | |
402 | count -= req_len; | |
35754bc0 | 403 | dreq->bytes_left -= req_len; |
584aa810 | 404 | } |
6d74743b TM |
405 | /* The nfs_page now hold references to these pages */ |
406 | nfs_direct_release_pages(pagevec, npages); | |
71e8cc00 | 407 | } while (count != 0 && result >= 0); |
607f31e8 | 408 | |
584aa810 FI |
409 | kfree(pagevec); |
410 | ||
607f31e8 | 411 | if (started) |
c216fd70 | 412 | return started; |
607f31e8 | 413 | return result < 0 ? (ssize_t) result : -EFAULT; |
1da177e4 LT |
414 | } |
415 | ||
19f73787 CL |
416 | static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq, |
417 | const struct iovec *iov, | |
418 | unsigned long nr_segs, | |
a564b8f0 | 419 | loff_t pos, bool uio) |
19f73787 | 420 | { |
584aa810 | 421 | struct nfs_pageio_descriptor desc; |
1f90ee27 | 422 | struct inode *inode = dreq->inode; |
19f73787 CL |
423 | ssize_t result = -EINVAL; |
424 | size_t requested_bytes = 0; | |
425 | unsigned long seg; | |
426 | ||
59948db3 | 427 | NFS_PROTO(dreq->inode)->read_pageio_init(&desc, dreq->inode, |
584aa810 | 428 | &nfs_direct_read_completion_ops); |
19f73787 | 429 | get_dreq(dreq); |
584aa810 | 430 | desc.pg_dreq = dreq; |
1f90ee27 | 431 | atomic_inc(&inode->i_dio_count); |
19f73787 CL |
432 | |
433 | for (seg = 0; seg < nr_segs; seg++) { | |
434 | const struct iovec *vec = &iov[seg]; | |
a564b8f0 | 435 | result = nfs_direct_read_schedule_segment(&desc, vec, pos, uio); |
19f73787 CL |
436 | if (result < 0) |
437 | break; | |
438 | requested_bytes += result; | |
439 | if ((size_t)result < vec->iov_len) | |
440 | break; | |
441 | pos += vec->iov_len; | |
442 | } | |
443 | ||
584aa810 FI |
444 | nfs_pageio_complete(&desc); |
445 | ||
839f7ad6 CL |
446 | /* |
447 | * If no bytes were started, return the error, and let the | |
448 | * generic layer handle the completion. | |
449 | */ | |
450 | if (requested_bytes == 0) { | |
1f90ee27 | 451 | inode_dio_done(inode); |
839f7ad6 CL |
452 | nfs_direct_req_release(dreq); |
453 | return result < 0 ? result : -EIO; | |
454 | } | |
455 | ||
19f73787 | 456 | if (put_dreq(dreq)) |
9811cd57 | 457 | nfs_direct_complete(dreq, false); |
839f7ad6 | 458 | return 0; |
19f73787 CL |
459 | } |
460 | ||
14a3ec79 CH |
461 | /** |
462 | * nfs_file_direct_read - file direct read operation for NFS files | |
463 | * @iocb: target I/O control block | |
464 | * @iov: vector of user buffers into which to read data | |
465 | * @nr_segs: size of iov vector | |
466 | * @pos: byte offset in file where reading starts | |
467 | * | |
468 | * We use this function for direct reads instead of calling | |
469 | * generic_file_aio_read() in order to avoid gfar's check to see if | |
470 | * the request starts before the end of the file. For that check | |
471 | * to work, we must generate a GETATTR before each direct read, and | |
472 | * even then there is a window between the GETATTR and the subsequent | |
473 | * READ where the file size could change. Our preference is simply | |
474 | * to do all reads the application wants, and the server will take | |
475 | * care of managing the end of file boundary. | |
476 | * | |
477 | * This function also eliminates unnecessarily updating the file's | |
478 | * atime locally, as the NFS server sets the file's atime, and this | |
479 | * client must read the updated atime from the server back into its | |
480 | * cache. | |
481 | */ | |
482 | ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov, | |
483 | unsigned long nr_segs, loff_t pos, bool uio) | |
1da177e4 | 484 | { |
14a3ec79 CH |
485 | struct file *file = iocb->ki_filp; |
486 | struct address_space *mapping = file->f_mapping; | |
487 | struct inode *inode = mapping->host; | |
1da177e4 | 488 | struct nfs_direct_req *dreq; |
b3c54de6 | 489 | struct nfs_lock_context *l_ctx; |
14a3ec79 CH |
490 | ssize_t result = -EINVAL; |
491 | size_t count; | |
492 | ||
493 | count = iov_length(iov, nr_segs); | |
494 | nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); | |
495 | ||
496 | dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n", | |
497 | file, count, (long long) pos); | |
498 | ||
499 | result = 0; | |
500 | if (!count) | |
501 | goto out; | |
502 | ||
d0b9875d | 503 | mutex_lock(&inode->i_mutex); |
14a3ec79 CH |
504 | result = nfs_sync_mapping(mapping); |
505 | if (result) | |
d0b9875d | 506 | goto out_unlock; |
1da177e4 | 507 | |
14a3ec79 CH |
508 | task_io_account_read(count); |
509 | ||
510 | result = -ENOMEM; | |
607f31e8 | 511 | dreq = nfs_direct_req_alloc(); |
f11ac8db | 512 | if (dreq == NULL) |
d0b9875d | 513 | goto out_unlock; |
1da177e4 | 514 | |
91d5b470 | 515 | dreq->inode = inode; |
35754bc0 | 516 | dreq->bytes_left = iov_length(iov, nr_segs); |
cd3758e3 | 517 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
b3c54de6 TM |
518 | l_ctx = nfs_get_lock_context(dreq->ctx); |
519 | if (IS_ERR(l_ctx)) { | |
520 | result = PTR_ERR(l_ctx); | |
f11ac8db | 521 | goto out_release; |
b3c54de6 TM |
522 | } |
523 | dreq->l_ctx = l_ctx; | |
487b8372 CL |
524 | if (!is_sync_kiocb(iocb)) |
525 | dreq->iocb = iocb; | |
1da177e4 | 526 | |
7acdb026 | 527 | NFS_I(inode)->read_io += iov_length(iov, nr_segs); |
a564b8f0 | 528 | result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos, uio); |
d0b9875d CH |
529 | |
530 | mutex_unlock(&inode->i_mutex); | |
531 | ||
14a3ec79 | 532 | if (!result) { |
607f31e8 | 533 | result = nfs_direct_wait(dreq); |
14a3ec79 CH |
534 | if (result > 0) |
535 | iocb->ki_pos = pos + result; | |
536 | } | |
d0b9875d CH |
537 | |
538 | nfs_direct_req_release(dreq); | |
539 | return result; | |
540 | ||
f11ac8db | 541 | out_release: |
b4946ffb | 542 | nfs_direct_req_release(dreq); |
d0b9875d CH |
543 | out_unlock: |
544 | mutex_unlock(&inode->i_mutex); | |
f11ac8db | 545 | out: |
1da177e4 LT |
546 | return result; |
547 | } | |
548 | ||
89d77c8f | 549 | #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) |
fad61490 TM |
550 | static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) |
551 | { | |
1763da12 FI |
552 | struct nfs_pageio_descriptor desc; |
553 | struct nfs_page *req, *tmp; | |
554 | LIST_HEAD(reqs); | |
555 | struct nfs_commit_info cinfo; | |
556 | LIST_HEAD(failed); | |
557 | ||
558 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
559 | pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo); | |
560 | spin_lock(cinfo.lock); | |
561 | nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0); | |
562 | spin_unlock(cinfo.lock); | |
1da177e4 | 563 | |
fad61490 | 564 | dreq->count = 0; |
607f31e8 TM |
565 | get_dreq(dreq); |
566 | ||
c95908e4 | 567 | NFS_PROTO(dreq->inode)->write_pageio_init(&desc, dreq->inode, FLUSH_STABLE, |
1763da12 FI |
568 | &nfs_direct_write_completion_ops); |
569 | desc.pg_dreq = dreq; | |
fedb595c | 570 | |
1763da12 FI |
571 | list_for_each_entry_safe(req, tmp, &reqs, wb_list) { |
572 | if (!nfs_pageio_add_request(&desc, req)) { | |
4035c248 | 573 | nfs_list_remove_request(req); |
1763da12 FI |
574 | nfs_list_add_request(req, &failed); |
575 | spin_lock(cinfo.lock); | |
576 | dreq->flags = 0; | |
577 | dreq->error = -EIO; | |
578 | spin_unlock(cinfo.lock); | |
579 | } | |
5a695da2 | 580 | nfs_release_request(req); |
1763da12 FI |
581 | } |
582 | nfs_pageio_complete(&desc); | |
fad61490 | 583 | |
4035c248 TM |
584 | while (!list_empty(&failed)) { |
585 | req = nfs_list_entry(failed.next); | |
586 | nfs_list_remove_request(req); | |
1d1afcbc | 587 | nfs_unlock_and_release_request(req); |
4035c248 | 588 | } |
fad61490 | 589 | |
1763da12 FI |
590 | if (put_dreq(dreq)) |
591 | nfs_direct_write_complete(dreq, dreq->inode); | |
c9d8f89d TM |
592 | } |
593 | ||
1763da12 | 594 | static void nfs_direct_commit_complete(struct nfs_commit_data *data) |
c9d8f89d | 595 | { |
0b7c0153 | 596 | struct nfs_direct_req *dreq = data->dreq; |
1763da12 FI |
597 | struct nfs_commit_info cinfo; |
598 | struct nfs_page *req; | |
c9d8f89d TM |
599 | int status = data->task.tk_status; |
600 | ||
1763da12 | 601 | nfs_init_cinfo_from_dreq(&cinfo, dreq); |
c9d8f89d | 602 | if (status < 0) { |
60fa3f76 | 603 | dprintk("NFS: %5u commit failed with error %d.\n", |
1763da12 | 604 | data->task.tk_pid, status); |
fad61490 | 605 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
60fa3f76 | 606 | } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) { |
c9d8f89d | 607 | dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid); |
fad61490 | 608 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
1da177e4 LT |
609 | } |
610 | ||
c9d8f89d | 611 | dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status); |
1763da12 FI |
612 | while (!list_empty(&data->pages)) { |
613 | req = nfs_list_entry(data->pages.next); | |
614 | nfs_list_remove_request(req); | |
615 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) { | |
616 | /* Note the rewrite will go through mds */ | |
617 | nfs_mark_request_commit(req, NULL, &cinfo); | |
906369e4 FI |
618 | } else |
619 | nfs_release_request(req); | |
1d1afcbc | 620 | nfs_unlock_and_release_request(req); |
1763da12 FI |
621 | } |
622 | ||
623 | if (atomic_dec_and_test(&cinfo.mds->rpcs_out)) | |
624 | nfs_direct_write_complete(dreq, data->inode); | |
1da177e4 LT |
625 | } |
626 | ||
1763da12 FI |
627 | static void nfs_direct_error_cleanup(struct nfs_inode *nfsi) |
628 | { | |
629 | /* There is no lock to clear */ | |
630 | } | |
631 | ||
632 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = { | |
633 | .completion = nfs_direct_commit_complete, | |
634 | .error_cleanup = nfs_direct_error_cleanup, | |
fad61490 TM |
635 | }; |
636 | ||
637 | static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) | |
1da177e4 | 638 | { |
1763da12 FI |
639 | int res; |
640 | struct nfs_commit_info cinfo; | |
641 | LIST_HEAD(mds_list); | |
642 | ||
643 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
644 | nfs_scan_commit(dreq->inode, &mds_list, &cinfo); | |
645 | res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo); | |
646 | if (res < 0) /* res == -ENOMEM */ | |
647 | nfs_direct_write_reschedule(dreq); | |
fad61490 | 648 | } |
1da177e4 | 649 | |
1763da12 | 650 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
fad61490 | 651 | { |
1763da12 | 652 | struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work); |
fad61490 | 653 | int flags = dreq->flags; |
1da177e4 | 654 | |
fad61490 TM |
655 | dreq->flags = 0; |
656 | switch (flags) { | |
657 | case NFS_ODIRECT_DO_COMMIT: | |
658 | nfs_direct_commit_schedule(dreq); | |
1da177e4 | 659 | break; |
fad61490 TM |
660 | case NFS_ODIRECT_RESCHED_WRITES: |
661 | nfs_direct_write_reschedule(dreq); | |
662 | break; | |
663 | default: | |
9811cd57 | 664 | nfs_direct_complete(dreq, true); |
fad61490 TM |
665 | } |
666 | } | |
1da177e4 | 667 | |
1763da12 | 668 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
fad61490 | 669 | { |
1763da12 | 670 | schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */ |
fad61490 | 671 | } |
1763da12 | 672 | |
fad61490 | 673 | #else |
24fc9211 BS |
674 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
675 | { | |
676 | } | |
1da177e4 | 677 | |
fad61490 TM |
678 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
679 | { | |
9811cd57 | 680 | nfs_direct_complete(dreq, true); |
fad61490 TM |
681 | } |
682 | #endif | |
1da177e4 | 683 | |
c9d8f89d TM |
684 | /* |
685 | * NB: Return the value of the first error return code. Subsequent | |
686 | * errors after the first one are ignored. | |
687 | */ | |
462d5b32 | 688 | /* |
607f31e8 TM |
689 | * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE |
690 | * operation. If nfs_writedata_alloc() or get_user_pages() fails, | |
691 | * bail and stop sending more writes. Write length accounting is | |
692 | * handled automatically by nfs_direct_write_result(). Otherwise, if | |
693 | * no requests have been sent, just return an error. | |
462d5b32 | 694 | */ |
1763da12 | 695 | static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc, |
02fe4946 | 696 | const struct iovec *iov, |
a564b8f0 | 697 | loff_t pos, bool uio) |
462d5b32 | 698 | { |
1763da12 | 699 | struct nfs_direct_req *dreq = desc->pg_dreq; |
a8881f5a | 700 | struct nfs_open_context *ctx = dreq->ctx; |
3d4ff43d | 701 | struct inode *inode = ctx->dentry->d_inode; |
02fe4946 CL |
702 | unsigned long user_addr = (unsigned long)iov->iov_base; |
703 | size_t count = iov->iov_len; | |
462d5b32 | 704 | size_t wsize = NFS_SERVER(inode)->wsize; |
607f31e8 TM |
705 | unsigned int pgbase; |
706 | int result; | |
707 | ssize_t started = 0; | |
1763da12 FI |
708 | struct page **pagevec = NULL; |
709 | unsigned int npages; | |
82b145c5 | 710 | |
1da177e4 | 711 | do { |
462d5b32 | 712 | size_t bytes; |
1763da12 | 713 | int i; |
462d5b32 | 714 | |
e9f7bee1 | 715 | pgbase = user_addr & ~PAGE_MASK; |
bf5fc402 | 716 | bytes = min(max_t(size_t, wsize, PAGE_SIZE), count); |
e9f7bee1 | 717 | |
607f31e8 | 718 | result = -ENOMEM; |
1763da12 FI |
719 | npages = nfs_page_array_len(pgbase, bytes); |
720 | if (!pagevec) | |
721 | pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL); | |
722 | if (!pagevec) | |
607f31e8 TM |
723 | break; |
724 | ||
a564b8f0 MG |
725 | if (uio) { |
726 | down_read(¤t->mm->mmap_sem); | |
727 | result = get_user_pages(current, current->mm, user_addr, | |
728 | npages, 0, 0, pagevec, NULL); | |
729 | up_read(¤t->mm->mmap_sem); | |
730 | if (result < 0) | |
731 | break; | |
732 | } else { | |
733 | WARN_ON(npages != 1); | |
734 | result = get_kernel_page(user_addr, 0, pagevec); | |
735 | if (WARN_ON(result != 1)) | |
736 | break; | |
737 | } | |
1763da12 FI |
738 | |
739 | if ((unsigned)result < npages) { | |
d9df8d6b TM |
740 | bytes = result * PAGE_SIZE; |
741 | if (bytes <= pgbase) { | |
1763da12 | 742 | nfs_direct_release_pages(pagevec, result); |
d9df8d6b TM |
743 | break; |
744 | } | |
745 | bytes -= pgbase; | |
1763da12 | 746 | npages = result; |
607f31e8 TM |
747 | } |
748 | ||
1763da12 FI |
749 | for (i = 0; i < npages; i++) { |
750 | struct nfs_page *req; | |
bf5fc402 | 751 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); |
1da177e4 | 752 | |
1763da12 FI |
753 | req = nfs_create_request(dreq->ctx, dreq->inode, |
754 | pagevec[i], | |
755 | pgbase, req_len); | |
756 | if (IS_ERR(req)) { | |
1763da12 FI |
757 | result = PTR_ERR(req); |
758 | break; | |
759 | } | |
760 | nfs_lock_request(req); | |
761 | req->wb_index = pos >> PAGE_SHIFT; | |
762 | req->wb_offset = pos & ~PAGE_MASK; | |
763 | if (!nfs_pageio_add_request(desc, req)) { | |
764 | result = desc->pg_error; | |
1d1afcbc | 765 | nfs_unlock_and_release_request(req); |
71e8cc00 | 766 | break; |
1763da12 FI |
767 | } |
768 | pgbase = 0; | |
769 | bytes -= req_len; | |
770 | started += req_len; | |
771 | user_addr += req_len; | |
772 | pos += req_len; | |
773 | count -= req_len; | |
35754bc0 | 774 | dreq->bytes_left -= req_len; |
1763da12 | 775 | } |
6d74743b TM |
776 | /* The nfs_page now hold references to these pages */ |
777 | nfs_direct_release_pages(pagevec, npages); | |
71e8cc00 | 778 | } while (count != 0 && result >= 0); |
607f31e8 | 779 | |
1763da12 FI |
780 | kfree(pagevec); |
781 | ||
607f31e8 | 782 | if (started) |
c216fd70 | 783 | return started; |
607f31e8 | 784 | return result < 0 ? (ssize_t) result : -EFAULT; |
462d5b32 | 785 | } |
1da177e4 | 786 | |
1763da12 FI |
787 | static void nfs_direct_write_completion(struct nfs_pgio_header *hdr) |
788 | { | |
789 | struct nfs_direct_req *dreq = hdr->dreq; | |
790 | struct nfs_commit_info cinfo; | |
791 | int bit = -1; | |
792 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
793 | ||
794 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) | |
795 | goto out_put; | |
796 | ||
797 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
798 | ||
799 | spin_lock(&dreq->lock); | |
800 | ||
801 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) { | |
802 | dreq->flags = 0; | |
803 | dreq->error = hdr->error; | |
804 | } | |
805 | if (dreq->error != 0) | |
806 | bit = NFS_IOHDR_ERROR; | |
807 | else { | |
808 | dreq->count += hdr->good_bytes; | |
809 | if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) { | |
810 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; | |
811 | bit = NFS_IOHDR_NEED_RESCHED; | |
812 | } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) { | |
813 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) | |
814 | bit = NFS_IOHDR_NEED_RESCHED; | |
815 | else if (dreq->flags == 0) { | |
9bce008b | 816 | memcpy(&dreq->verf, hdr->verf, |
1763da12 FI |
817 | sizeof(dreq->verf)); |
818 | bit = NFS_IOHDR_NEED_COMMIT; | |
819 | dreq->flags = NFS_ODIRECT_DO_COMMIT; | |
820 | } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) { | |
9bce008b | 821 | if (memcmp(&dreq->verf, hdr->verf, sizeof(dreq->verf))) { |
1763da12 FI |
822 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
823 | bit = NFS_IOHDR_NEED_RESCHED; | |
824 | } else | |
825 | bit = NFS_IOHDR_NEED_COMMIT; | |
826 | } | |
827 | } | |
828 | } | |
829 | spin_unlock(&dreq->lock); | |
830 | ||
831 | while (!list_empty(&hdr->pages)) { | |
832 | req = nfs_list_entry(hdr->pages.next); | |
833 | nfs_list_remove_request(req); | |
834 | switch (bit) { | |
835 | case NFS_IOHDR_NEED_RESCHED: | |
836 | case NFS_IOHDR_NEED_COMMIT: | |
04277086 | 837 | kref_get(&req->wb_kref); |
1763da12 | 838 | nfs_mark_request_commit(req, hdr->lseg, &cinfo); |
1763da12 | 839 | } |
1d1afcbc | 840 | nfs_unlock_and_release_request(req); |
1763da12 FI |
841 | } |
842 | ||
843 | out_put: | |
844 | if (put_dreq(dreq)) | |
845 | nfs_direct_write_complete(dreq, hdr->inode); | |
846 | hdr->release(hdr); | |
847 | } | |
848 | ||
3e9e0ca3 TM |
849 | static void nfs_write_sync_pgio_error(struct list_head *head) |
850 | { | |
851 | struct nfs_page *req; | |
852 | ||
853 | while (!list_empty(head)) { | |
854 | req = nfs_list_entry(head->next); | |
855 | nfs_list_remove_request(req); | |
1d1afcbc | 856 | nfs_unlock_and_release_request(req); |
3e9e0ca3 TM |
857 | } |
858 | } | |
859 | ||
1763da12 | 860 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = { |
3e9e0ca3 | 861 | .error_cleanup = nfs_write_sync_pgio_error, |
1763da12 FI |
862 | .init_hdr = nfs_direct_pgio_init, |
863 | .completion = nfs_direct_write_completion, | |
864 | }; | |
865 | ||
19f73787 CL |
866 | static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, |
867 | const struct iovec *iov, | |
868 | unsigned long nr_segs, | |
a564b8f0 | 869 | loff_t pos, bool uio) |
19f73787 | 870 | { |
1763da12 | 871 | struct nfs_pageio_descriptor desc; |
1d59d61f | 872 | struct inode *inode = dreq->inode; |
19f73787 CL |
873 | ssize_t result = 0; |
874 | size_t requested_bytes = 0; | |
875 | unsigned long seg; | |
876 | ||
c95908e4 | 877 | NFS_PROTO(inode)->write_pageio_init(&desc, inode, FLUSH_COND_STABLE, |
1763da12 FI |
878 | &nfs_direct_write_completion_ops); |
879 | desc.pg_dreq = dreq; | |
19f73787 | 880 | get_dreq(dreq); |
1d59d61f | 881 | atomic_inc(&inode->i_dio_count); |
19f73787 | 882 | |
7acdb026 | 883 | NFS_I(dreq->inode)->write_io += iov_length(iov, nr_segs); |
19f73787 CL |
884 | for (seg = 0; seg < nr_segs; seg++) { |
885 | const struct iovec *vec = &iov[seg]; | |
a564b8f0 | 886 | result = nfs_direct_write_schedule_segment(&desc, vec, pos, uio); |
19f73787 CL |
887 | if (result < 0) |
888 | break; | |
889 | requested_bytes += result; | |
890 | if ((size_t)result < vec->iov_len) | |
891 | break; | |
892 | pos += vec->iov_len; | |
893 | } | |
1763da12 | 894 | nfs_pageio_complete(&desc); |
19f73787 | 895 | |
839f7ad6 CL |
896 | /* |
897 | * If no bytes were started, return the error, and let the | |
898 | * generic layer handle the completion. | |
899 | */ | |
900 | if (requested_bytes == 0) { | |
1d59d61f | 901 | inode_dio_done(inode); |
839f7ad6 CL |
902 | nfs_direct_req_release(dreq); |
903 | return result < 0 ? result : -EIO; | |
904 | } | |
905 | ||
19f73787 CL |
906 | if (put_dreq(dreq)) |
907 | nfs_direct_write_complete(dreq, dreq->inode); | |
839f7ad6 | 908 | return 0; |
19f73787 CL |
909 | } |
910 | ||
1da177e4 LT |
911 | /** |
912 | * nfs_file_direct_write - file direct write operation for NFS files | |
913 | * @iocb: target I/O control block | |
027445c3 BP |
914 | * @iov: vector of user buffers from which to write data |
915 | * @nr_segs: size of iov vector | |
88467055 | 916 | * @pos: byte offset in file where writing starts |
1da177e4 LT |
917 | * |
918 | * We use this function for direct writes instead of calling | |
919 | * generic_file_aio_write() in order to avoid taking the inode | |
920 | * semaphore and updating the i_size. The NFS server will set | |
921 | * the new i_size and this client must read the updated size | |
922 | * back into its cache. We let the server do generic write | |
923 | * parameter checking and report problems. | |
924 | * | |
1da177e4 LT |
925 | * We eliminate local atime updates, see direct read above. |
926 | * | |
927 | * We avoid unnecessary page cache invalidations for normal cached | |
928 | * readers of this file. | |
929 | * | |
930 | * Note that O_APPEND is not supported for NFS direct writes, as there | |
931 | * is no atomic O_APPEND write facility in the NFS protocol. | |
932 | */ | |
027445c3 | 933 | ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov, |
a564b8f0 | 934 | unsigned long nr_segs, loff_t pos, bool uio) |
1da177e4 | 935 | { |
22cd1bf1 | 936 | ssize_t result = -EINVAL; |
1da177e4 | 937 | struct file *file = iocb->ki_filp; |
1da177e4 | 938 | struct address_space *mapping = file->f_mapping; |
22cd1bf1 CH |
939 | struct inode *inode = mapping->host; |
940 | struct nfs_direct_req *dreq; | |
941 | struct nfs_lock_context *l_ctx; | |
a9ab5e84 | 942 | loff_t end; |
c216fd70 | 943 | size_t count; |
1da177e4 | 944 | |
c216fd70 | 945 | count = iov_length(iov, nr_segs); |
a9ab5e84 CH |
946 | end = (pos + count - 1) >> PAGE_CACHE_SHIFT; |
947 | ||
c216fd70 CL |
948 | nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count); |
949 | ||
6de1472f AV |
950 | dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n", |
951 | file, count, (long long) pos); | |
027445c3 | 952 | |
22cd1bf1 CH |
953 | result = generic_write_checks(file, &pos, &count, 0); |
954 | if (result) | |
1da177e4 | 955 | goto out; |
ce1a8e67 | 956 | |
22cd1bf1 | 957 | result = -EINVAL; |
ce1a8e67 | 958 | if ((ssize_t) count < 0) |
1da177e4 | 959 | goto out; |
22cd1bf1 | 960 | result = 0; |
1da177e4 LT |
961 | if (!count) |
962 | goto out; | |
ce1a8e67 | 963 | |
a9ab5e84 CH |
964 | mutex_lock(&inode->i_mutex); |
965 | ||
22cd1bf1 CH |
966 | result = nfs_sync_mapping(mapping); |
967 | if (result) | |
a9ab5e84 CH |
968 | goto out_unlock; |
969 | ||
970 | if (mapping->nrpages) { | |
971 | result = invalidate_inode_pages2_range(mapping, | |
972 | pos >> PAGE_CACHE_SHIFT, end); | |
973 | if (result) | |
974 | goto out_unlock; | |
975 | } | |
1da177e4 | 976 | |
7ec10f26 KK |
977 | task_io_account_write(count); |
978 | ||
22cd1bf1 CH |
979 | result = -ENOMEM; |
980 | dreq = nfs_direct_req_alloc(); | |
981 | if (!dreq) | |
a9ab5e84 | 982 | goto out_unlock; |
9eafa8cc | 983 | |
22cd1bf1 CH |
984 | dreq->inode = inode; |
985 | dreq->bytes_left = count; | |
986 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); | |
987 | l_ctx = nfs_get_lock_context(dreq->ctx); | |
988 | if (IS_ERR(l_ctx)) { | |
989 | result = PTR_ERR(l_ctx); | |
990 | goto out_release; | |
991 | } | |
992 | dreq->l_ctx = l_ctx; | |
993 | if (!is_sync_kiocb(iocb)) | |
994 | dreq->iocb = iocb; | |
995 | ||
996 | result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, uio); | |
a9ab5e84 CH |
997 | |
998 | if (mapping->nrpages) { | |
999 | invalidate_inode_pages2_range(mapping, | |
1000 | pos >> PAGE_CACHE_SHIFT, end); | |
1001 | } | |
1002 | ||
1003 | mutex_unlock(&inode->i_mutex); | |
1004 | ||
22cd1bf1 CH |
1005 | if (!result) { |
1006 | result = nfs_direct_wait(dreq); | |
1007 | if (result > 0) { | |
1008 | struct inode *inode = mapping->host; | |
1009 | ||
1010 | iocb->ki_pos = pos + result; | |
1011 | spin_lock(&inode->i_lock); | |
1012 | if (i_size_read(inode) < iocb->ki_pos) | |
1013 | i_size_write(inode, iocb->ki_pos); | |
1014 | spin_unlock(&inode->i_lock); | |
1015 | } | |
1763da12 | 1016 | } |
a9ab5e84 CH |
1017 | nfs_direct_req_release(dreq); |
1018 | return result; | |
1019 | ||
22cd1bf1 CH |
1020 | out_release: |
1021 | nfs_direct_req_release(dreq); | |
a9ab5e84 CH |
1022 | out_unlock: |
1023 | mutex_unlock(&inode->i_mutex); | |
1da177e4 | 1024 | out: |
22cd1bf1 | 1025 | return result; |
1da177e4 LT |
1026 | } |
1027 | ||
88467055 CL |
1028 | /** |
1029 | * nfs_init_directcache - create a slab cache for nfs_direct_req structures | |
1030 | * | |
1031 | */ | |
f7b422b1 | 1032 | int __init nfs_init_directcache(void) |
1da177e4 LT |
1033 | { |
1034 | nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", | |
1035 | sizeof(struct nfs_direct_req), | |
fffb60f9 PJ |
1036 | 0, (SLAB_RECLAIM_ACCOUNT| |
1037 | SLAB_MEM_SPREAD), | |
20c2df83 | 1038 | NULL); |
1da177e4 LT |
1039 | if (nfs_direct_cachep == NULL) |
1040 | return -ENOMEM; | |
1041 | ||
1042 | return 0; | |
1043 | } | |
1044 | ||
88467055 | 1045 | /** |
f7b422b1 | 1046 | * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures |
88467055 CL |
1047 | * |
1048 | */ | |
266bee88 | 1049 | void nfs_destroy_directcache(void) |
1da177e4 | 1050 | { |
1a1d92c1 | 1051 | kmem_cache_destroy(nfs_direct_cachep); |
1da177e4 | 1052 | } |