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 | ||
63ab46ab | 229 | if (dreq->iocb) { |
9811cd57 | 230 | loff_t pos = dreq->iocb->ki_pos + dreq->count; |
15ce4a0c | 231 | long res = (long) dreq->error; |
63ab46ab | 232 | if (!res) |
15ce4a0c | 233 | res = (long) dreq->count; |
9811cd57 CH |
234 | |
235 | if (write) { | |
236 | spin_lock(&inode->i_lock); | |
237 | if (i_size_read(inode) < pos) | |
238 | i_size_write(inode, pos); | |
239 | spin_unlock(&inode->i_lock); | |
240 | } | |
241 | ||
63ab46ab | 242 | aio_complete(dreq->iocb, res, 0); |
d72b7a6b TM |
243 | } |
244 | complete_all(&dreq->completion); | |
63ab46ab | 245 | |
b4946ffb | 246 | nfs_direct_req_release(dreq); |
63ab46ab CL |
247 | } |
248 | ||
1385b811 | 249 | static void nfs_direct_readpage_release(struct nfs_page *req) |
1da177e4 | 250 | { |
1e8968c5 | 251 | dprintk("NFS: direct read done (%s/%llu %d@%lld)\n", |
584aa810 | 252 | req->wb_context->dentry->d_inode->i_sb->s_id, |
1e8968c5 | 253 | (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode), |
584aa810 FI |
254 | req->wb_bytes, |
255 | (long long)req_offset(req)); | |
256 | nfs_release_request(req); | |
fdd1e74c TM |
257 | } |
258 | ||
584aa810 | 259 | static void nfs_direct_read_completion(struct nfs_pgio_header *hdr) |
fdd1e74c | 260 | { |
584aa810 FI |
261 | unsigned long bytes = 0; |
262 | struct nfs_direct_req *dreq = hdr->dreq; | |
fdd1e74c | 263 | |
584aa810 FI |
264 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) |
265 | goto out_put; | |
15ce4a0c CL |
266 | |
267 | spin_lock(&dreq->lock); | |
584aa810 FI |
268 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0)) |
269 | dreq->error = hdr->error; | |
270 | else | |
271 | dreq->count += hdr->good_bytes; | |
272 | spin_unlock(&dreq->lock); | |
273 | ||
4bd8b010 TM |
274 | while (!list_empty(&hdr->pages)) { |
275 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
276 | struct page *page = req->wb_page; | |
277 | ||
be7e9858 JL |
278 | if (!PageCompound(page) && bytes < hdr->good_bytes) |
279 | set_page_dirty(page); | |
4bd8b010 TM |
280 | bytes += req->wb_bytes; |
281 | nfs_list_remove_request(req); | |
282 | nfs_direct_readpage_release(req); | |
d4a8f367 | 283 | } |
584aa810 | 284 | out_put: |
607f31e8 | 285 | if (put_dreq(dreq)) |
9811cd57 | 286 | nfs_direct_complete(dreq, false); |
584aa810 | 287 | hdr->release(hdr); |
1da177e4 LT |
288 | } |
289 | ||
3e9e0ca3 | 290 | static void nfs_read_sync_pgio_error(struct list_head *head) |
cd841605 | 291 | { |
584aa810 | 292 | struct nfs_page *req; |
cd841605 | 293 | |
584aa810 FI |
294 | while (!list_empty(head)) { |
295 | req = nfs_list_entry(head->next); | |
296 | nfs_list_remove_request(req); | |
297 | nfs_release_request(req); | |
298 | } | |
cd841605 FI |
299 | } |
300 | ||
584aa810 FI |
301 | static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr) |
302 | { | |
303 | get_dreq(hdr->dreq); | |
304 | } | |
305 | ||
306 | static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = { | |
3e9e0ca3 | 307 | .error_cleanup = nfs_read_sync_pgio_error, |
584aa810 FI |
308 | .init_hdr = nfs_direct_pgio_init, |
309 | .completion = nfs_direct_read_completion, | |
310 | }; | |
311 | ||
d4cc948b | 312 | /* |
607f31e8 TM |
313 | * For each rsize'd chunk of the user's buffer, dispatch an NFS READ |
314 | * operation. If nfs_readdata_alloc() or get_user_pages() fails, | |
315 | * bail and stop sending more reads. Read length accounting is | |
316 | * handled automatically by nfs_direct_read_result(). Otherwise, if | |
317 | * no requests have been sent, just return an error. | |
1da177e4 | 318 | */ |
584aa810 | 319 | static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc, |
02fe4946 | 320 | const struct iovec *iov, |
a564b8f0 | 321 | loff_t pos, bool uio) |
1da177e4 | 322 | { |
584aa810 | 323 | struct nfs_direct_req *dreq = desc->pg_dreq; |
a8881f5a | 324 | struct nfs_open_context *ctx = dreq->ctx; |
3d4ff43d | 325 | struct inode *inode = ctx->dentry->d_inode; |
02fe4946 CL |
326 | unsigned long user_addr = (unsigned long)iov->iov_base; |
327 | size_t count = iov->iov_len; | |
5dd602f2 | 328 | size_t rsize = NFS_SERVER(inode)->rsize; |
607f31e8 TM |
329 | unsigned int pgbase; |
330 | int result; | |
331 | ssize_t started = 0; | |
584aa810 FI |
332 | struct page **pagevec = NULL; |
333 | unsigned int npages; | |
607f31e8 | 334 | |
1da177e4 | 335 | do { |
5dd602f2 | 336 | size_t bytes; |
584aa810 | 337 | int i; |
1da177e4 | 338 | |
e9f7bee1 | 339 | pgbase = user_addr & ~PAGE_MASK; |
bf5fc402 | 340 | bytes = min(max_t(size_t, rsize, PAGE_SIZE), count); |
e9f7bee1 | 341 | |
607f31e8 | 342 | result = -ENOMEM; |
584aa810 FI |
343 | npages = nfs_page_array_len(pgbase, bytes); |
344 | if (!pagevec) | |
345 | pagevec = kmalloc(npages * sizeof(struct page *), | |
346 | GFP_KERNEL); | |
347 | if (!pagevec) | |
4db6e0b7 | 348 | break; |
a564b8f0 MG |
349 | if (uio) { |
350 | down_read(¤t->mm->mmap_sem); | |
351 | result = get_user_pages(current, current->mm, user_addr, | |
584aa810 | 352 | npages, 1, 0, pagevec, NULL); |
a564b8f0 MG |
353 | up_read(¤t->mm->mmap_sem); |
354 | if (result < 0) | |
355 | break; | |
356 | } else { | |
357 | WARN_ON(npages != 1); | |
358 | result = get_kernel_page(user_addr, 1, pagevec); | |
359 | if (WARN_ON(result != 1)) | |
360 | break; | |
361 | } | |
362 | ||
584aa810 | 363 | if ((unsigned)result < npages) { |
d9df8d6b TM |
364 | bytes = result * PAGE_SIZE; |
365 | if (bytes <= pgbase) { | |
584aa810 | 366 | nfs_direct_release_pages(pagevec, result); |
d9df8d6b TM |
367 | break; |
368 | } | |
369 | bytes -= pgbase; | |
584aa810 | 370 | npages = result; |
607f31e8 TM |
371 | } |
372 | ||
584aa810 FI |
373 | for (i = 0; i < npages; i++) { |
374 | struct nfs_page *req; | |
bf5fc402 | 375 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); |
584aa810 FI |
376 | /* XXX do we need to do the eof zeroing found in async_filler? */ |
377 | req = nfs_create_request(dreq->ctx, dreq->inode, | |
378 | pagevec[i], | |
379 | pgbase, req_len); | |
380 | if (IS_ERR(req)) { | |
584aa810 FI |
381 | result = PTR_ERR(req); |
382 | break; | |
383 | } | |
384 | req->wb_index = pos >> PAGE_SHIFT; | |
385 | req->wb_offset = pos & ~PAGE_MASK; | |
386 | if (!nfs_pageio_add_request(desc, req)) { | |
387 | result = desc->pg_error; | |
388 | nfs_release_request(req); | |
584aa810 FI |
389 | break; |
390 | } | |
391 | pgbase = 0; | |
392 | bytes -= req_len; | |
393 | started += req_len; | |
394 | user_addr += req_len; | |
395 | pos += req_len; | |
396 | count -= req_len; | |
35754bc0 | 397 | dreq->bytes_left -= req_len; |
584aa810 | 398 | } |
6d74743b TM |
399 | /* The nfs_page now hold references to these pages */ |
400 | nfs_direct_release_pages(pagevec, npages); | |
71e8cc00 | 401 | } while (count != 0 && result >= 0); |
607f31e8 | 402 | |
584aa810 FI |
403 | kfree(pagevec); |
404 | ||
607f31e8 | 405 | if (started) |
c216fd70 | 406 | return started; |
607f31e8 | 407 | return result < 0 ? (ssize_t) result : -EFAULT; |
1da177e4 LT |
408 | } |
409 | ||
19f73787 CL |
410 | static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq, |
411 | const struct iovec *iov, | |
412 | unsigned long nr_segs, | |
a564b8f0 | 413 | loff_t pos, bool uio) |
19f73787 | 414 | { |
584aa810 | 415 | struct nfs_pageio_descriptor desc; |
19f73787 CL |
416 | ssize_t result = -EINVAL; |
417 | size_t requested_bytes = 0; | |
418 | unsigned long seg; | |
419 | ||
59948db3 | 420 | NFS_PROTO(dreq->inode)->read_pageio_init(&desc, dreq->inode, |
584aa810 | 421 | &nfs_direct_read_completion_ops); |
19f73787 | 422 | get_dreq(dreq); |
584aa810 | 423 | desc.pg_dreq = dreq; |
19f73787 CL |
424 | |
425 | for (seg = 0; seg < nr_segs; seg++) { | |
426 | const struct iovec *vec = &iov[seg]; | |
a564b8f0 | 427 | result = nfs_direct_read_schedule_segment(&desc, vec, pos, uio); |
19f73787 CL |
428 | if (result < 0) |
429 | break; | |
430 | requested_bytes += result; | |
431 | if ((size_t)result < vec->iov_len) | |
432 | break; | |
433 | pos += vec->iov_len; | |
434 | } | |
435 | ||
584aa810 FI |
436 | nfs_pageio_complete(&desc); |
437 | ||
839f7ad6 CL |
438 | /* |
439 | * If no bytes were started, return the error, and let the | |
440 | * generic layer handle the completion. | |
441 | */ | |
442 | if (requested_bytes == 0) { | |
443 | nfs_direct_req_release(dreq); | |
444 | return result < 0 ? result : -EIO; | |
445 | } | |
446 | ||
19f73787 | 447 | if (put_dreq(dreq)) |
9811cd57 | 448 | nfs_direct_complete(dreq, false); |
839f7ad6 | 449 | return 0; |
19f73787 CL |
450 | } |
451 | ||
c216fd70 | 452 | static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov, |
a564b8f0 | 453 | unsigned long nr_segs, loff_t pos, bool uio) |
1da177e4 | 454 | { |
f11ac8db | 455 | ssize_t result = -ENOMEM; |
99514f8f | 456 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
1da177e4 | 457 | struct nfs_direct_req *dreq; |
b3c54de6 | 458 | struct nfs_lock_context *l_ctx; |
1da177e4 | 459 | |
607f31e8 | 460 | dreq = nfs_direct_req_alloc(); |
f11ac8db TM |
461 | if (dreq == NULL) |
462 | goto out; | |
1da177e4 | 463 | |
91d5b470 | 464 | dreq->inode = inode; |
35754bc0 | 465 | dreq->bytes_left = iov_length(iov, nr_segs); |
cd3758e3 | 466 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
b3c54de6 TM |
467 | l_ctx = nfs_get_lock_context(dreq->ctx); |
468 | if (IS_ERR(l_ctx)) { | |
469 | result = PTR_ERR(l_ctx); | |
f11ac8db | 470 | goto out_release; |
b3c54de6 TM |
471 | } |
472 | dreq->l_ctx = l_ctx; | |
487b8372 CL |
473 | if (!is_sync_kiocb(iocb)) |
474 | dreq->iocb = iocb; | |
1da177e4 | 475 | |
7acdb026 | 476 | NFS_I(inode)->read_io += iov_length(iov, nr_segs); |
a564b8f0 | 477 | result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos, uio); |
607f31e8 TM |
478 | if (!result) |
479 | result = nfs_direct_wait(dreq); | |
f11ac8db | 480 | out_release: |
b4946ffb | 481 | nfs_direct_req_release(dreq); |
f11ac8db | 482 | out: |
1da177e4 LT |
483 | return result; |
484 | } | |
485 | ||
1d59d61f TM |
486 | static void nfs_inode_dio_write_done(struct inode *inode) |
487 | { | |
488 | nfs_zap_mapping(inode, inode->i_mapping); | |
489 | inode_dio_done(inode); | |
490 | } | |
491 | ||
89d77c8f | 492 | #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) |
fad61490 TM |
493 | static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) |
494 | { | |
1763da12 FI |
495 | struct nfs_pageio_descriptor desc; |
496 | struct nfs_page *req, *tmp; | |
497 | LIST_HEAD(reqs); | |
498 | struct nfs_commit_info cinfo; | |
499 | LIST_HEAD(failed); | |
500 | ||
501 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
502 | pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo); | |
503 | spin_lock(cinfo.lock); | |
504 | nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0); | |
505 | spin_unlock(cinfo.lock); | |
1da177e4 | 506 | |
fad61490 | 507 | dreq->count = 0; |
607f31e8 TM |
508 | get_dreq(dreq); |
509 | ||
c95908e4 | 510 | NFS_PROTO(dreq->inode)->write_pageio_init(&desc, dreq->inode, FLUSH_STABLE, |
1763da12 FI |
511 | &nfs_direct_write_completion_ops); |
512 | desc.pg_dreq = dreq; | |
fedb595c | 513 | |
1763da12 FI |
514 | list_for_each_entry_safe(req, tmp, &reqs, wb_list) { |
515 | if (!nfs_pageio_add_request(&desc, req)) { | |
4035c248 | 516 | nfs_list_remove_request(req); |
1763da12 FI |
517 | nfs_list_add_request(req, &failed); |
518 | spin_lock(cinfo.lock); | |
519 | dreq->flags = 0; | |
520 | dreq->error = -EIO; | |
521 | spin_unlock(cinfo.lock); | |
522 | } | |
5a695da2 | 523 | nfs_release_request(req); |
1763da12 FI |
524 | } |
525 | nfs_pageio_complete(&desc); | |
fad61490 | 526 | |
4035c248 TM |
527 | while (!list_empty(&failed)) { |
528 | req = nfs_list_entry(failed.next); | |
529 | nfs_list_remove_request(req); | |
1d1afcbc | 530 | nfs_unlock_and_release_request(req); |
4035c248 | 531 | } |
fad61490 | 532 | |
1763da12 FI |
533 | if (put_dreq(dreq)) |
534 | nfs_direct_write_complete(dreq, dreq->inode); | |
c9d8f89d TM |
535 | } |
536 | ||
1763da12 | 537 | static void nfs_direct_commit_complete(struct nfs_commit_data *data) |
c9d8f89d | 538 | { |
0b7c0153 | 539 | struct nfs_direct_req *dreq = data->dreq; |
1763da12 FI |
540 | struct nfs_commit_info cinfo; |
541 | struct nfs_page *req; | |
c9d8f89d TM |
542 | int status = data->task.tk_status; |
543 | ||
1763da12 | 544 | nfs_init_cinfo_from_dreq(&cinfo, dreq); |
c9d8f89d | 545 | if (status < 0) { |
60fa3f76 | 546 | dprintk("NFS: %5u commit failed with error %d.\n", |
1763da12 | 547 | data->task.tk_pid, status); |
fad61490 | 548 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
60fa3f76 | 549 | } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) { |
c9d8f89d | 550 | dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid); |
fad61490 | 551 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
1da177e4 LT |
552 | } |
553 | ||
c9d8f89d | 554 | dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status); |
1763da12 FI |
555 | while (!list_empty(&data->pages)) { |
556 | req = nfs_list_entry(data->pages.next); | |
557 | nfs_list_remove_request(req); | |
558 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) { | |
559 | /* Note the rewrite will go through mds */ | |
560 | nfs_mark_request_commit(req, NULL, &cinfo); | |
906369e4 FI |
561 | } else |
562 | nfs_release_request(req); | |
1d1afcbc | 563 | nfs_unlock_and_release_request(req); |
1763da12 FI |
564 | } |
565 | ||
566 | if (atomic_dec_and_test(&cinfo.mds->rpcs_out)) | |
567 | nfs_direct_write_complete(dreq, data->inode); | |
1da177e4 LT |
568 | } |
569 | ||
1763da12 FI |
570 | static void nfs_direct_error_cleanup(struct nfs_inode *nfsi) |
571 | { | |
572 | /* There is no lock to clear */ | |
573 | } | |
574 | ||
575 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = { | |
576 | .completion = nfs_direct_commit_complete, | |
577 | .error_cleanup = nfs_direct_error_cleanup, | |
fad61490 TM |
578 | }; |
579 | ||
580 | static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) | |
1da177e4 | 581 | { |
1763da12 FI |
582 | int res; |
583 | struct nfs_commit_info cinfo; | |
584 | LIST_HEAD(mds_list); | |
585 | ||
586 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
587 | nfs_scan_commit(dreq->inode, &mds_list, &cinfo); | |
588 | res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo); | |
589 | if (res < 0) /* res == -ENOMEM */ | |
590 | nfs_direct_write_reschedule(dreq); | |
fad61490 | 591 | } |
1da177e4 | 592 | |
1763da12 | 593 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
fad61490 | 594 | { |
1763da12 | 595 | struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work); |
fad61490 | 596 | int flags = dreq->flags; |
1da177e4 | 597 | |
fad61490 TM |
598 | dreq->flags = 0; |
599 | switch (flags) { | |
600 | case NFS_ODIRECT_DO_COMMIT: | |
601 | nfs_direct_commit_schedule(dreq); | |
1da177e4 | 602 | break; |
fad61490 TM |
603 | case NFS_ODIRECT_RESCHED_WRITES: |
604 | nfs_direct_write_reschedule(dreq); | |
605 | break; | |
606 | default: | |
1d59d61f | 607 | nfs_inode_dio_write_done(dreq->inode); |
9811cd57 | 608 | nfs_direct_complete(dreq, true); |
fad61490 TM |
609 | } |
610 | } | |
1da177e4 | 611 | |
1763da12 | 612 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
fad61490 | 613 | { |
1763da12 | 614 | schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */ |
fad61490 | 615 | } |
1763da12 | 616 | |
fad61490 | 617 | #else |
24fc9211 BS |
618 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
619 | { | |
620 | } | |
1da177e4 | 621 | |
fad61490 TM |
622 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
623 | { | |
1d59d61f | 624 | nfs_inode_dio_write_done(inode); |
9811cd57 | 625 | nfs_direct_complete(dreq, true); |
fad61490 TM |
626 | } |
627 | #endif | |
1da177e4 | 628 | |
c9d8f89d TM |
629 | /* |
630 | * NB: Return the value of the first error return code. Subsequent | |
631 | * errors after the first one are ignored. | |
632 | */ | |
462d5b32 | 633 | /* |
607f31e8 TM |
634 | * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE |
635 | * operation. If nfs_writedata_alloc() or get_user_pages() fails, | |
636 | * bail and stop sending more writes. Write length accounting is | |
637 | * handled automatically by nfs_direct_write_result(). Otherwise, if | |
638 | * no requests have been sent, just return an error. | |
462d5b32 | 639 | */ |
1763da12 | 640 | static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc, |
02fe4946 | 641 | const struct iovec *iov, |
a564b8f0 | 642 | loff_t pos, bool uio) |
462d5b32 | 643 | { |
1763da12 | 644 | struct nfs_direct_req *dreq = desc->pg_dreq; |
a8881f5a | 645 | struct nfs_open_context *ctx = dreq->ctx; |
3d4ff43d | 646 | struct inode *inode = ctx->dentry->d_inode; |
02fe4946 CL |
647 | unsigned long user_addr = (unsigned long)iov->iov_base; |
648 | size_t count = iov->iov_len; | |
462d5b32 | 649 | size_t wsize = NFS_SERVER(inode)->wsize; |
607f31e8 TM |
650 | unsigned int pgbase; |
651 | int result; | |
652 | ssize_t started = 0; | |
1763da12 FI |
653 | struct page **pagevec = NULL; |
654 | unsigned int npages; | |
82b145c5 | 655 | |
1da177e4 | 656 | do { |
462d5b32 | 657 | size_t bytes; |
1763da12 | 658 | int i; |
462d5b32 | 659 | |
e9f7bee1 | 660 | pgbase = user_addr & ~PAGE_MASK; |
bf5fc402 | 661 | bytes = min(max_t(size_t, wsize, PAGE_SIZE), count); |
e9f7bee1 | 662 | |
607f31e8 | 663 | result = -ENOMEM; |
1763da12 FI |
664 | npages = nfs_page_array_len(pgbase, bytes); |
665 | if (!pagevec) | |
666 | pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL); | |
667 | if (!pagevec) | |
607f31e8 TM |
668 | break; |
669 | ||
a564b8f0 MG |
670 | if (uio) { |
671 | down_read(¤t->mm->mmap_sem); | |
672 | result = get_user_pages(current, current->mm, user_addr, | |
673 | npages, 0, 0, pagevec, NULL); | |
674 | up_read(¤t->mm->mmap_sem); | |
675 | if (result < 0) | |
676 | break; | |
677 | } else { | |
678 | WARN_ON(npages != 1); | |
679 | result = get_kernel_page(user_addr, 0, pagevec); | |
680 | if (WARN_ON(result != 1)) | |
681 | break; | |
682 | } | |
1763da12 FI |
683 | |
684 | if ((unsigned)result < npages) { | |
d9df8d6b TM |
685 | bytes = result * PAGE_SIZE; |
686 | if (bytes <= pgbase) { | |
1763da12 | 687 | nfs_direct_release_pages(pagevec, result); |
d9df8d6b TM |
688 | break; |
689 | } | |
690 | bytes -= pgbase; | |
1763da12 | 691 | npages = result; |
607f31e8 TM |
692 | } |
693 | ||
1763da12 FI |
694 | for (i = 0; i < npages; i++) { |
695 | struct nfs_page *req; | |
bf5fc402 | 696 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); |
1da177e4 | 697 | |
1763da12 FI |
698 | req = nfs_create_request(dreq->ctx, dreq->inode, |
699 | pagevec[i], | |
700 | pgbase, req_len); | |
701 | if (IS_ERR(req)) { | |
1763da12 FI |
702 | result = PTR_ERR(req); |
703 | break; | |
704 | } | |
705 | nfs_lock_request(req); | |
706 | req->wb_index = pos >> PAGE_SHIFT; | |
707 | req->wb_offset = pos & ~PAGE_MASK; | |
708 | if (!nfs_pageio_add_request(desc, req)) { | |
709 | result = desc->pg_error; | |
1d1afcbc | 710 | nfs_unlock_and_release_request(req); |
71e8cc00 | 711 | break; |
1763da12 FI |
712 | } |
713 | pgbase = 0; | |
714 | bytes -= req_len; | |
715 | started += req_len; | |
716 | user_addr += req_len; | |
717 | pos += req_len; | |
718 | count -= req_len; | |
35754bc0 | 719 | dreq->bytes_left -= req_len; |
1763da12 | 720 | } |
6d74743b TM |
721 | /* The nfs_page now hold references to these pages */ |
722 | nfs_direct_release_pages(pagevec, npages); | |
71e8cc00 | 723 | } while (count != 0 && result >= 0); |
607f31e8 | 724 | |
1763da12 FI |
725 | kfree(pagevec); |
726 | ||
607f31e8 | 727 | if (started) |
c216fd70 | 728 | return started; |
607f31e8 | 729 | return result < 0 ? (ssize_t) result : -EFAULT; |
462d5b32 | 730 | } |
1da177e4 | 731 | |
1763da12 FI |
732 | static void nfs_direct_write_completion(struct nfs_pgio_header *hdr) |
733 | { | |
734 | struct nfs_direct_req *dreq = hdr->dreq; | |
735 | struct nfs_commit_info cinfo; | |
736 | int bit = -1; | |
737 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
738 | ||
739 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) | |
740 | goto out_put; | |
741 | ||
742 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
743 | ||
744 | spin_lock(&dreq->lock); | |
745 | ||
746 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) { | |
747 | dreq->flags = 0; | |
748 | dreq->error = hdr->error; | |
749 | } | |
750 | if (dreq->error != 0) | |
751 | bit = NFS_IOHDR_ERROR; | |
752 | else { | |
753 | dreq->count += hdr->good_bytes; | |
754 | if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) { | |
755 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; | |
756 | bit = NFS_IOHDR_NEED_RESCHED; | |
757 | } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) { | |
758 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) | |
759 | bit = NFS_IOHDR_NEED_RESCHED; | |
760 | else if (dreq->flags == 0) { | |
9bce008b | 761 | memcpy(&dreq->verf, hdr->verf, |
1763da12 FI |
762 | sizeof(dreq->verf)); |
763 | bit = NFS_IOHDR_NEED_COMMIT; | |
764 | dreq->flags = NFS_ODIRECT_DO_COMMIT; | |
765 | } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) { | |
9bce008b | 766 | if (memcmp(&dreq->verf, hdr->verf, sizeof(dreq->verf))) { |
1763da12 FI |
767 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
768 | bit = NFS_IOHDR_NEED_RESCHED; | |
769 | } else | |
770 | bit = NFS_IOHDR_NEED_COMMIT; | |
771 | } | |
772 | } | |
773 | } | |
774 | spin_unlock(&dreq->lock); | |
775 | ||
776 | while (!list_empty(&hdr->pages)) { | |
777 | req = nfs_list_entry(hdr->pages.next); | |
778 | nfs_list_remove_request(req); | |
779 | switch (bit) { | |
780 | case NFS_IOHDR_NEED_RESCHED: | |
781 | case NFS_IOHDR_NEED_COMMIT: | |
04277086 | 782 | kref_get(&req->wb_kref); |
1763da12 | 783 | nfs_mark_request_commit(req, hdr->lseg, &cinfo); |
1763da12 | 784 | } |
1d1afcbc | 785 | nfs_unlock_and_release_request(req); |
1763da12 FI |
786 | } |
787 | ||
788 | out_put: | |
789 | if (put_dreq(dreq)) | |
790 | nfs_direct_write_complete(dreq, hdr->inode); | |
791 | hdr->release(hdr); | |
792 | } | |
793 | ||
3e9e0ca3 TM |
794 | static void nfs_write_sync_pgio_error(struct list_head *head) |
795 | { | |
796 | struct nfs_page *req; | |
797 | ||
798 | while (!list_empty(head)) { | |
799 | req = nfs_list_entry(head->next); | |
800 | nfs_list_remove_request(req); | |
1d1afcbc | 801 | nfs_unlock_and_release_request(req); |
3e9e0ca3 TM |
802 | } |
803 | } | |
804 | ||
1763da12 | 805 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = { |
3e9e0ca3 | 806 | .error_cleanup = nfs_write_sync_pgio_error, |
1763da12 FI |
807 | .init_hdr = nfs_direct_pgio_init, |
808 | .completion = nfs_direct_write_completion, | |
809 | }; | |
810 | ||
19f73787 CL |
811 | static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, |
812 | const struct iovec *iov, | |
813 | unsigned long nr_segs, | |
a564b8f0 | 814 | loff_t pos, bool uio) |
19f73787 | 815 | { |
1763da12 | 816 | struct nfs_pageio_descriptor desc; |
1d59d61f | 817 | struct inode *inode = dreq->inode; |
19f73787 CL |
818 | ssize_t result = 0; |
819 | size_t requested_bytes = 0; | |
820 | unsigned long seg; | |
821 | ||
c95908e4 | 822 | NFS_PROTO(inode)->write_pageio_init(&desc, inode, FLUSH_COND_STABLE, |
1763da12 FI |
823 | &nfs_direct_write_completion_ops); |
824 | desc.pg_dreq = dreq; | |
19f73787 | 825 | get_dreq(dreq); |
1d59d61f | 826 | atomic_inc(&inode->i_dio_count); |
19f73787 | 827 | |
7acdb026 | 828 | NFS_I(dreq->inode)->write_io += iov_length(iov, nr_segs); |
19f73787 CL |
829 | for (seg = 0; seg < nr_segs; seg++) { |
830 | const struct iovec *vec = &iov[seg]; | |
a564b8f0 | 831 | result = nfs_direct_write_schedule_segment(&desc, vec, pos, uio); |
19f73787 CL |
832 | if (result < 0) |
833 | break; | |
834 | requested_bytes += result; | |
835 | if ((size_t)result < vec->iov_len) | |
836 | break; | |
837 | pos += vec->iov_len; | |
838 | } | |
1763da12 | 839 | nfs_pageio_complete(&desc); |
19f73787 | 840 | |
839f7ad6 CL |
841 | /* |
842 | * If no bytes were started, return the error, and let the | |
843 | * generic layer handle the completion. | |
844 | */ | |
845 | if (requested_bytes == 0) { | |
1d59d61f | 846 | inode_dio_done(inode); |
839f7ad6 CL |
847 | nfs_direct_req_release(dreq); |
848 | return result < 0 ? result : -EIO; | |
849 | } | |
850 | ||
19f73787 CL |
851 | if (put_dreq(dreq)) |
852 | nfs_direct_write_complete(dreq, dreq->inode); | |
839f7ad6 | 853 | return 0; |
19f73787 CL |
854 | } |
855 | ||
c216fd70 CL |
856 | static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov, |
857 | unsigned long nr_segs, loff_t pos, | |
a564b8f0 | 858 | size_t count, bool uio) |
462d5b32 | 859 | { |
f11ac8db | 860 | ssize_t result = -ENOMEM; |
c89f2ee5 | 861 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
462d5b32 | 862 | struct nfs_direct_req *dreq; |
b3c54de6 | 863 | struct nfs_lock_context *l_ctx; |
1da177e4 | 864 | |
607f31e8 | 865 | dreq = nfs_direct_req_alloc(); |
462d5b32 | 866 | if (!dreq) |
f11ac8db | 867 | goto out; |
1da177e4 | 868 | |
c89f2ee5 | 869 | dreq->inode = inode; |
35754bc0 | 870 | dreq->bytes_left = count; |
cd3758e3 | 871 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
b3c54de6 TM |
872 | l_ctx = nfs_get_lock_context(dreq->ctx); |
873 | if (IS_ERR(l_ctx)) { | |
874 | result = PTR_ERR(l_ctx); | |
f11ac8db | 875 | goto out_release; |
b3c54de6 TM |
876 | } |
877 | dreq->l_ctx = l_ctx; | |
c89f2ee5 CL |
878 | if (!is_sync_kiocb(iocb)) |
879 | dreq->iocb = iocb; | |
1da177e4 | 880 | |
a564b8f0 | 881 | result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, uio); |
607f31e8 TM |
882 | if (!result) |
883 | result = nfs_direct_wait(dreq); | |
f11ac8db | 884 | out_release: |
b4946ffb | 885 | nfs_direct_req_release(dreq); |
f11ac8db | 886 | out: |
1da177e4 LT |
887 | return result; |
888 | } | |
889 | ||
890 | /** | |
891 | * nfs_file_direct_read - file direct read operation for NFS files | |
892 | * @iocb: target I/O control block | |
027445c3 BP |
893 | * @iov: vector of user buffers into which to read data |
894 | * @nr_segs: size of iov vector | |
88467055 | 895 | * @pos: byte offset in file where reading starts |
1da177e4 LT |
896 | * |
897 | * We use this function for direct reads instead of calling | |
898 | * generic_file_aio_read() in order to avoid gfar's check to see if | |
899 | * the request starts before the end of the file. For that check | |
900 | * to work, we must generate a GETATTR before each direct read, and | |
901 | * even then there is a window between the GETATTR and the subsequent | |
88467055 | 902 | * READ where the file size could change. Our preference is simply |
1da177e4 LT |
903 | * to do all reads the application wants, and the server will take |
904 | * care of managing the end of file boundary. | |
88467055 | 905 | * |
1da177e4 LT |
906 | * This function also eliminates unnecessarily updating the file's |
907 | * atime locally, as the NFS server sets the file's atime, and this | |
908 | * client must read the updated atime from the server back into its | |
909 | * cache. | |
910 | */ | |
027445c3 | 911 | ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov, |
a564b8f0 | 912 | unsigned long nr_segs, loff_t pos, bool uio) |
1da177e4 LT |
913 | { |
914 | ssize_t retval = -EINVAL; | |
1da177e4 | 915 | struct file *file = iocb->ki_filp; |
1da177e4 | 916 | struct address_space *mapping = file->f_mapping; |
c216fd70 CL |
917 | size_t count; |
918 | ||
919 | count = iov_length(iov, nr_segs); | |
920 | nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); | |
1da177e4 | 921 | |
6de1472f AV |
922 | dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n", |
923 | file, count, (long long) pos); | |
1da177e4 | 924 | |
1da177e4 LT |
925 | retval = 0; |
926 | if (!count) | |
927 | goto out; | |
928 | ||
29884df0 TM |
929 | retval = nfs_sync_mapping(mapping); |
930 | if (retval) | |
931 | goto out; | |
1da177e4 | 932 | |
7ec10f26 KK |
933 | task_io_account_read(count); |
934 | ||
a564b8f0 | 935 | retval = nfs_direct_read(iocb, iov, nr_segs, pos, uio); |
1da177e4 | 936 | if (retval > 0) |
0cdd80d0 | 937 | iocb->ki_pos = pos + retval; |
1da177e4 LT |
938 | |
939 | out: | |
940 | return retval; | |
941 | } | |
942 | ||
943 | /** | |
944 | * nfs_file_direct_write - file direct write operation for NFS files | |
945 | * @iocb: target I/O control block | |
027445c3 BP |
946 | * @iov: vector of user buffers from which to write data |
947 | * @nr_segs: size of iov vector | |
88467055 | 948 | * @pos: byte offset in file where writing starts |
1da177e4 LT |
949 | * |
950 | * We use this function for direct writes instead of calling | |
951 | * generic_file_aio_write() in order to avoid taking the inode | |
952 | * semaphore and updating the i_size. The NFS server will set | |
953 | * the new i_size and this client must read the updated size | |
954 | * back into its cache. We let the server do generic write | |
955 | * parameter checking and report problems. | |
956 | * | |
1da177e4 LT |
957 | * We eliminate local atime updates, see direct read above. |
958 | * | |
959 | * We avoid unnecessary page cache invalidations for normal cached | |
960 | * readers of this file. | |
961 | * | |
962 | * Note that O_APPEND is not supported for NFS direct writes, as there | |
963 | * is no atomic O_APPEND write facility in the NFS protocol. | |
964 | */ | |
027445c3 | 965 | ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov, |
a564b8f0 | 966 | unsigned long nr_segs, loff_t pos, bool uio) |
1da177e4 | 967 | { |
070ea602 | 968 | ssize_t retval = -EINVAL; |
1da177e4 | 969 | struct file *file = iocb->ki_filp; |
1da177e4 | 970 | struct address_space *mapping = file->f_mapping; |
c216fd70 | 971 | size_t count; |
1da177e4 | 972 | |
c216fd70 CL |
973 | count = iov_length(iov, nr_segs); |
974 | nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count); | |
975 | ||
6de1472f AV |
976 | dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n", |
977 | file, count, (long long) pos); | |
027445c3 | 978 | |
ce1a8e67 CL |
979 | retval = generic_write_checks(file, &pos, &count, 0); |
980 | if (retval) | |
1da177e4 | 981 | goto out; |
ce1a8e67 CL |
982 | |
983 | retval = -EINVAL; | |
984 | if ((ssize_t) count < 0) | |
1da177e4 | 985 | goto out; |
1da177e4 LT |
986 | retval = 0; |
987 | if (!count) | |
988 | goto out; | |
ce1a8e67 | 989 | |
29884df0 TM |
990 | retval = nfs_sync_mapping(mapping); |
991 | if (retval) | |
992 | goto out; | |
1da177e4 | 993 | |
7ec10f26 KK |
994 | task_io_account_write(count); |
995 | ||
a564b8f0 | 996 | retval = nfs_direct_write(iocb, iov, nr_segs, pos, count, uio); |
1763da12 FI |
997 | if (retval > 0) { |
998 | struct inode *inode = mapping->host; | |
9eafa8cc | 999 | |
ce1a8e67 | 1000 | iocb->ki_pos = pos + retval; |
1763da12 FI |
1001 | spin_lock(&inode->i_lock); |
1002 | if (i_size_read(inode) < iocb->ki_pos) | |
1003 | i_size_write(inode, iocb->ki_pos); | |
1004 | spin_unlock(&inode->i_lock); | |
1005 | } | |
1da177e4 LT |
1006 | out: |
1007 | return retval; | |
1008 | } | |
1009 | ||
88467055 CL |
1010 | /** |
1011 | * nfs_init_directcache - create a slab cache for nfs_direct_req structures | |
1012 | * | |
1013 | */ | |
f7b422b1 | 1014 | int __init nfs_init_directcache(void) |
1da177e4 LT |
1015 | { |
1016 | nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", | |
1017 | sizeof(struct nfs_direct_req), | |
fffb60f9 PJ |
1018 | 0, (SLAB_RECLAIM_ACCOUNT| |
1019 | SLAB_MEM_SPREAD), | |
20c2df83 | 1020 | NULL); |
1da177e4 LT |
1021 | if (nfs_direct_cachep == NULL) |
1022 | return -ENOMEM; | |
1023 | ||
1024 | return 0; | |
1025 | } | |
1026 | ||
88467055 | 1027 | /** |
f7b422b1 | 1028 | * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures |
88467055 CL |
1029 | * |
1030 | */ | |
266bee88 | 1031 | void nfs_destroy_directcache(void) |
1da177e4 | 1032 | { |
1a1d92c1 | 1033 | kmem_cache_destroy(nfs_direct_cachep); |
1da177e4 | 1034 | } |