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 | 382 | /* XXX do we need to do the eof zeroing found in async_filler? */ |
8c8f1ac1 | 383 | req = nfs_create_request(dreq->ctx, pagevec[i], |
584aa810 FI |
384 | pgbase, req_len); |
385 | if (IS_ERR(req)) { | |
584aa810 FI |
386 | result = PTR_ERR(req); |
387 | break; | |
388 | } | |
389 | req->wb_index = pos >> PAGE_SHIFT; | |
390 | req->wb_offset = pos & ~PAGE_MASK; | |
391 | if (!nfs_pageio_add_request(desc, req)) { | |
392 | result = desc->pg_error; | |
393 | nfs_release_request(req); | |
584aa810 FI |
394 | break; |
395 | } | |
396 | pgbase = 0; | |
397 | bytes -= req_len; | |
398 | started += req_len; | |
399 | user_addr += req_len; | |
400 | pos += req_len; | |
401 | count -= req_len; | |
35754bc0 | 402 | dreq->bytes_left -= req_len; |
584aa810 | 403 | } |
6d74743b TM |
404 | /* The nfs_page now hold references to these pages */ |
405 | nfs_direct_release_pages(pagevec, npages); | |
71e8cc00 | 406 | } while (count != 0 && result >= 0); |
607f31e8 | 407 | |
584aa810 FI |
408 | kfree(pagevec); |
409 | ||
607f31e8 | 410 | if (started) |
c216fd70 | 411 | return started; |
607f31e8 | 412 | return result < 0 ? (ssize_t) result : -EFAULT; |
1da177e4 LT |
413 | } |
414 | ||
19f73787 CL |
415 | static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq, |
416 | const struct iovec *iov, | |
417 | unsigned long nr_segs, | |
a564b8f0 | 418 | loff_t pos, bool uio) |
19f73787 | 419 | { |
584aa810 | 420 | struct nfs_pageio_descriptor desc; |
1f90ee27 | 421 | struct inode *inode = dreq->inode; |
19f73787 CL |
422 | ssize_t result = -EINVAL; |
423 | size_t requested_bytes = 0; | |
424 | unsigned long seg; | |
425 | ||
fab5fc25 | 426 | nfs_pageio_init_read(&desc, dreq->inode, false, |
584aa810 | 427 | &nfs_direct_read_completion_ops); |
19f73787 | 428 | get_dreq(dreq); |
584aa810 | 429 | desc.pg_dreq = dreq; |
1f90ee27 | 430 | atomic_inc(&inode->i_dio_count); |
19f73787 CL |
431 | |
432 | for (seg = 0; seg < nr_segs; seg++) { | |
433 | const struct iovec *vec = &iov[seg]; | |
a564b8f0 | 434 | result = nfs_direct_read_schedule_segment(&desc, vec, pos, uio); |
19f73787 CL |
435 | if (result < 0) |
436 | break; | |
437 | requested_bytes += result; | |
438 | if ((size_t)result < vec->iov_len) | |
439 | break; | |
440 | pos += vec->iov_len; | |
441 | } | |
442 | ||
584aa810 FI |
443 | nfs_pageio_complete(&desc); |
444 | ||
839f7ad6 CL |
445 | /* |
446 | * If no bytes were started, return the error, and let the | |
447 | * generic layer handle the completion. | |
448 | */ | |
449 | if (requested_bytes == 0) { | |
1f90ee27 | 450 | inode_dio_done(inode); |
839f7ad6 CL |
451 | nfs_direct_req_release(dreq); |
452 | return result < 0 ? result : -EIO; | |
453 | } | |
454 | ||
19f73787 | 455 | if (put_dreq(dreq)) |
9811cd57 | 456 | nfs_direct_complete(dreq, false); |
839f7ad6 | 457 | return 0; |
19f73787 CL |
458 | } |
459 | ||
14a3ec79 CH |
460 | /** |
461 | * nfs_file_direct_read - file direct read operation for NFS files | |
462 | * @iocb: target I/O control block | |
463 | * @iov: vector of user buffers into which to read data | |
464 | * @nr_segs: size of iov vector | |
465 | * @pos: byte offset in file where reading starts | |
466 | * | |
467 | * We use this function for direct reads instead of calling | |
468 | * generic_file_aio_read() in order to avoid gfar's check to see if | |
469 | * the request starts before the end of the file. For that check | |
470 | * to work, we must generate a GETATTR before each direct read, and | |
471 | * even then there is a window between the GETATTR and the subsequent | |
472 | * READ where the file size could change. Our preference is simply | |
473 | * to do all reads the application wants, and the server will take | |
474 | * care of managing the end of file boundary. | |
475 | * | |
476 | * This function also eliminates unnecessarily updating the file's | |
477 | * atime locally, as the NFS server sets the file's atime, and this | |
478 | * client must read the updated atime from the server back into its | |
479 | * cache. | |
480 | */ | |
481 | ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov, | |
482 | unsigned long nr_segs, loff_t pos, bool uio) | |
1da177e4 | 483 | { |
14a3ec79 CH |
484 | struct file *file = iocb->ki_filp; |
485 | struct address_space *mapping = file->f_mapping; | |
486 | struct inode *inode = mapping->host; | |
1da177e4 | 487 | struct nfs_direct_req *dreq; |
b3c54de6 | 488 | struct nfs_lock_context *l_ctx; |
14a3ec79 CH |
489 | ssize_t result = -EINVAL; |
490 | size_t count; | |
491 | ||
492 | count = iov_length(iov, nr_segs); | |
493 | nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); | |
494 | ||
495 | dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n", | |
496 | file, count, (long long) pos); | |
497 | ||
498 | result = 0; | |
499 | if (!count) | |
500 | goto out; | |
501 | ||
d0b9875d | 502 | mutex_lock(&inode->i_mutex); |
14a3ec79 CH |
503 | result = nfs_sync_mapping(mapping); |
504 | if (result) | |
d0b9875d | 505 | goto out_unlock; |
1da177e4 | 506 | |
14a3ec79 CH |
507 | task_io_account_read(count); |
508 | ||
509 | result = -ENOMEM; | |
607f31e8 | 510 | dreq = nfs_direct_req_alloc(); |
f11ac8db | 511 | if (dreq == NULL) |
d0b9875d | 512 | goto out_unlock; |
1da177e4 | 513 | |
91d5b470 | 514 | dreq->inode = inode; |
35754bc0 | 515 | dreq->bytes_left = iov_length(iov, nr_segs); |
cd3758e3 | 516 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
b3c54de6 TM |
517 | l_ctx = nfs_get_lock_context(dreq->ctx); |
518 | if (IS_ERR(l_ctx)) { | |
519 | result = PTR_ERR(l_ctx); | |
f11ac8db | 520 | goto out_release; |
b3c54de6 TM |
521 | } |
522 | dreq->l_ctx = l_ctx; | |
487b8372 CL |
523 | if (!is_sync_kiocb(iocb)) |
524 | dreq->iocb = iocb; | |
1da177e4 | 525 | |
7acdb026 | 526 | NFS_I(inode)->read_io += iov_length(iov, nr_segs); |
a564b8f0 | 527 | result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos, uio); |
d0b9875d CH |
528 | |
529 | mutex_unlock(&inode->i_mutex); | |
530 | ||
14a3ec79 | 531 | if (!result) { |
607f31e8 | 532 | result = nfs_direct_wait(dreq); |
14a3ec79 CH |
533 | if (result > 0) |
534 | iocb->ki_pos = pos + result; | |
535 | } | |
d0b9875d CH |
536 | |
537 | nfs_direct_req_release(dreq); | |
538 | return result; | |
539 | ||
f11ac8db | 540 | out_release: |
b4946ffb | 541 | nfs_direct_req_release(dreq); |
d0b9875d CH |
542 | out_unlock: |
543 | mutex_unlock(&inode->i_mutex); | |
f11ac8db | 544 | out: |
1da177e4 LT |
545 | return result; |
546 | } | |
547 | ||
89d77c8f | 548 | #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) |
fad61490 TM |
549 | static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) |
550 | { | |
1763da12 FI |
551 | struct nfs_pageio_descriptor desc; |
552 | struct nfs_page *req, *tmp; | |
553 | LIST_HEAD(reqs); | |
554 | struct nfs_commit_info cinfo; | |
555 | LIST_HEAD(failed); | |
556 | ||
557 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
558 | pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo); | |
559 | spin_lock(cinfo.lock); | |
560 | nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0); | |
561 | spin_unlock(cinfo.lock); | |
1da177e4 | 562 | |
fad61490 | 563 | dreq->count = 0; |
607f31e8 TM |
564 | get_dreq(dreq); |
565 | ||
a20c93e3 | 566 | nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false, |
1763da12 FI |
567 | &nfs_direct_write_completion_ops); |
568 | desc.pg_dreq = dreq; | |
fedb595c | 569 | |
1763da12 FI |
570 | list_for_each_entry_safe(req, tmp, &reqs, wb_list) { |
571 | if (!nfs_pageio_add_request(&desc, req)) { | |
4035c248 | 572 | nfs_list_remove_request(req); |
1763da12 FI |
573 | nfs_list_add_request(req, &failed); |
574 | spin_lock(cinfo.lock); | |
575 | dreq->flags = 0; | |
576 | dreq->error = -EIO; | |
577 | spin_unlock(cinfo.lock); | |
578 | } | |
5a695da2 | 579 | nfs_release_request(req); |
1763da12 FI |
580 | } |
581 | nfs_pageio_complete(&desc); | |
fad61490 | 582 | |
4035c248 TM |
583 | while (!list_empty(&failed)) { |
584 | req = nfs_list_entry(failed.next); | |
585 | nfs_list_remove_request(req); | |
1d1afcbc | 586 | nfs_unlock_and_release_request(req); |
4035c248 | 587 | } |
fad61490 | 588 | |
1763da12 FI |
589 | if (put_dreq(dreq)) |
590 | nfs_direct_write_complete(dreq, dreq->inode); | |
c9d8f89d TM |
591 | } |
592 | ||
1763da12 | 593 | static void nfs_direct_commit_complete(struct nfs_commit_data *data) |
c9d8f89d | 594 | { |
0b7c0153 | 595 | struct nfs_direct_req *dreq = data->dreq; |
1763da12 FI |
596 | struct nfs_commit_info cinfo; |
597 | struct nfs_page *req; | |
c9d8f89d TM |
598 | int status = data->task.tk_status; |
599 | ||
1763da12 | 600 | nfs_init_cinfo_from_dreq(&cinfo, dreq); |
c9d8f89d | 601 | if (status < 0) { |
60fa3f76 | 602 | dprintk("NFS: %5u commit failed with error %d.\n", |
1763da12 | 603 | data->task.tk_pid, status); |
fad61490 | 604 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
60fa3f76 | 605 | } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) { |
c9d8f89d | 606 | dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid); |
fad61490 | 607 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
1da177e4 LT |
608 | } |
609 | ||
c9d8f89d | 610 | dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status); |
1763da12 FI |
611 | while (!list_empty(&data->pages)) { |
612 | req = nfs_list_entry(data->pages.next); | |
613 | nfs_list_remove_request(req); | |
614 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) { | |
615 | /* Note the rewrite will go through mds */ | |
616 | nfs_mark_request_commit(req, NULL, &cinfo); | |
906369e4 FI |
617 | } else |
618 | nfs_release_request(req); | |
1d1afcbc | 619 | nfs_unlock_and_release_request(req); |
1763da12 FI |
620 | } |
621 | ||
622 | if (atomic_dec_and_test(&cinfo.mds->rpcs_out)) | |
623 | nfs_direct_write_complete(dreq, data->inode); | |
1da177e4 LT |
624 | } |
625 | ||
1763da12 FI |
626 | static void nfs_direct_error_cleanup(struct nfs_inode *nfsi) |
627 | { | |
628 | /* There is no lock to clear */ | |
629 | } | |
630 | ||
631 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = { | |
632 | .completion = nfs_direct_commit_complete, | |
633 | .error_cleanup = nfs_direct_error_cleanup, | |
fad61490 TM |
634 | }; |
635 | ||
636 | static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) | |
1da177e4 | 637 | { |
1763da12 FI |
638 | int res; |
639 | struct nfs_commit_info cinfo; | |
640 | LIST_HEAD(mds_list); | |
641 | ||
642 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
643 | nfs_scan_commit(dreq->inode, &mds_list, &cinfo); | |
644 | res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo); | |
645 | if (res < 0) /* res == -ENOMEM */ | |
646 | nfs_direct_write_reschedule(dreq); | |
fad61490 | 647 | } |
1da177e4 | 648 | |
1763da12 | 649 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
fad61490 | 650 | { |
1763da12 | 651 | struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work); |
fad61490 | 652 | int flags = dreq->flags; |
1da177e4 | 653 | |
fad61490 TM |
654 | dreq->flags = 0; |
655 | switch (flags) { | |
656 | case NFS_ODIRECT_DO_COMMIT: | |
657 | nfs_direct_commit_schedule(dreq); | |
1da177e4 | 658 | break; |
fad61490 TM |
659 | case NFS_ODIRECT_RESCHED_WRITES: |
660 | nfs_direct_write_reschedule(dreq); | |
661 | break; | |
662 | default: | |
9811cd57 | 663 | nfs_direct_complete(dreq, true); |
fad61490 TM |
664 | } |
665 | } | |
1da177e4 | 666 | |
1763da12 | 667 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
fad61490 | 668 | { |
1763da12 | 669 | schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */ |
fad61490 | 670 | } |
1763da12 | 671 | |
fad61490 | 672 | #else |
24fc9211 BS |
673 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
674 | { | |
675 | } | |
1da177e4 | 676 | |
fad61490 TM |
677 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
678 | { | |
9811cd57 | 679 | nfs_direct_complete(dreq, true); |
fad61490 TM |
680 | } |
681 | #endif | |
1da177e4 | 682 | |
c9d8f89d TM |
683 | /* |
684 | * NB: Return the value of the first error return code. Subsequent | |
685 | * errors after the first one are ignored. | |
686 | */ | |
462d5b32 | 687 | /* |
607f31e8 TM |
688 | * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE |
689 | * operation. If nfs_writedata_alloc() or get_user_pages() fails, | |
690 | * bail and stop sending more writes. Write length accounting is | |
691 | * handled automatically by nfs_direct_write_result(). Otherwise, if | |
692 | * no requests have been sent, just return an error. | |
462d5b32 | 693 | */ |
1763da12 | 694 | static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc, |
02fe4946 | 695 | const struct iovec *iov, |
a564b8f0 | 696 | loff_t pos, bool uio) |
462d5b32 | 697 | { |
1763da12 | 698 | struct nfs_direct_req *dreq = desc->pg_dreq; |
a8881f5a | 699 | struct nfs_open_context *ctx = dreq->ctx; |
3d4ff43d | 700 | struct inode *inode = ctx->dentry->d_inode; |
02fe4946 CL |
701 | unsigned long user_addr = (unsigned long)iov->iov_base; |
702 | size_t count = iov->iov_len; | |
462d5b32 | 703 | size_t wsize = NFS_SERVER(inode)->wsize; |
607f31e8 TM |
704 | unsigned int pgbase; |
705 | int result; | |
706 | ssize_t started = 0; | |
1763da12 FI |
707 | struct page **pagevec = NULL; |
708 | unsigned int npages; | |
82b145c5 | 709 | |
1da177e4 | 710 | do { |
462d5b32 | 711 | size_t bytes; |
1763da12 | 712 | int i; |
462d5b32 | 713 | |
e9f7bee1 | 714 | pgbase = user_addr & ~PAGE_MASK; |
bf5fc402 | 715 | bytes = min(max_t(size_t, wsize, PAGE_SIZE), count); |
e9f7bee1 | 716 | |
607f31e8 | 717 | result = -ENOMEM; |
1763da12 FI |
718 | npages = nfs_page_array_len(pgbase, bytes); |
719 | if (!pagevec) | |
720 | pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL); | |
721 | if (!pagevec) | |
607f31e8 TM |
722 | break; |
723 | ||
a564b8f0 MG |
724 | if (uio) { |
725 | down_read(¤t->mm->mmap_sem); | |
726 | result = get_user_pages(current, current->mm, user_addr, | |
727 | npages, 0, 0, pagevec, NULL); | |
728 | up_read(¤t->mm->mmap_sem); | |
729 | if (result < 0) | |
730 | break; | |
731 | } else { | |
732 | WARN_ON(npages != 1); | |
733 | result = get_kernel_page(user_addr, 0, pagevec); | |
734 | if (WARN_ON(result != 1)) | |
735 | break; | |
736 | } | |
1763da12 FI |
737 | |
738 | if ((unsigned)result < npages) { | |
d9df8d6b TM |
739 | bytes = result * PAGE_SIZE; |
740 | if (bytes <= pgbase) { | |
1763da12 | 741 | nfs_direct_release_pages(pagevec, result); |
d9df8d6b TM |
742 | break; |
743 | } | |
744 | bytes -= pgbase; | |
1763da12 | 745 | npages = result; |
607f31e8 TM |
746 | } |
747 | ||
1763da12 FI |
748 | for (i = 0; i < npages; i++) { |
749 | struct nfs_page *req; | |
bf5fc402 | 750 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); |
1da177e4 | 751 | |
8c8f1ac1 | 752 | req = nfs_create_request(dreq->ctx, pagevec[i], |
1763da12 FI |
753 | pgbase, req_len); |
754 | if (IS_ERR(req)) { | |
1763da12 FI |
755 | result = PTR_ERR(req); |
756 | break; | |
757 | } | |
758 | nfs_lock_request(req); | |
759 | req->wb_index = pos >> PAGE_SHIFT; | |
760 | req->wb_offset = pos & ~PAGE_MASK; | |
761 | if (!nfs_pageio_add_request(desc, req)) { | |
762 | result = desc->pg_error; | |
1d1afcbc | 763 | nfs_unlock_and_release_request(req); |
71e8cc00 | 764 | break; |
1763da12 FI |
765 | } |
766 | pgbase = 0; | |
767 | bytes -= req_len; | |
768 | started += req_len; | |
769 | user_addr += req_len; | |
770 | pos += req_len; | |
771 | count -= req_len; | |
35754bc0 | 772 | dreq->bytes_left -= req_len; |
1763da12 | 773 | } |
6d74743b TM |
774 | /* The nfs_page now hold references to these pages */ |
775 | nfs_direct_release_pages(pagevec, npages); | |
71e8cc00 | 776 | } while (count != 0 && result >= 0); |
607f31e8 | 777 | |
1763da12 FI |
778 | kfree(pagevec); |
779 | ||
607f31e8 | 780 | if (started) |
c216fd70 | 781 | return started; |
607f31e8 | 782 | return result < 0 ? (ssize_t) result : -EFAULT; |
462d5b32 | 783 | } |
1da177e4 | 784 | |
1763da12 FI |
785 | static void nfs_direct_write_completion(struct nfs_pgio_header *hdr) |
786 | { | |
787 | struct nfs_direct_req *dreq = hdr->dreq; | |
788 | struct nfs_commit_info cinfo; | |
789 | int bit = -1; | |
790 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
791 | ||
792 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) | |
793 | goto out_put; | |
794 | ||
795 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
796 | ||
797 | spin_lock(&dreq->lock); | |
798 | ||
799 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) { | |
800 | dreq->flags = 0; | |
801 | dreq->error = hdr->error; | |
802 | } | |
803 | if (dreq->error != 0) | |
804 | bit = NFS_IOHDR_ERROR; | |
805 | else { | |
806 | dreq->count += hdr->good_bytes; | |
807 | if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) { | |
808 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; | |
809 | bit = NFS_IOHDR_NEED_RESCHED; | |
810 | } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) { | |
811 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) | |
812 | bit = NFS_IOHDR_NEED_RESCHED; | |
813 | else if (dreq->flags == 0) { | |
f79d06f5 | 814 | memcpy(&dreq->verf, &hdr->verf, |
1763da12 FI |
815 | sizeof(dreq->verf)); |
816 | bit = NFS_IOHDR_NEED_COMMIT; | |
817 | dreq->flags = NFS_ODIRECT_DO_COMMIT; | |
818 | } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) { | |
f79d06f5 | 819 | if (memcmp(&dreq->verf, &hdr->verf, sizeof(dreq->verf))) { |
1763da12 FI |
820 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
821 | bit = NFS_IOHDR_NEED_RESCHED; | |
822 | } else | |
823 | bit = NFS_IOHDR_NEED_COMMIT; | |
824 | } | |
825 | } | |
826 | } | |
827 | spin_unlock(&dreq->lock); | |
828 | ||
829 | while (!list_empty(&hdr->pages)) { | |
830 | req = nfs_list_entry(hdr->pages.next); | |
831 | nfs_list_remove_request(req); | |
832 | switch (bit) { | |
833 | case NFS_IOHDR_NEED_RESCHED: | |
834 | case NFS_IOHDR_NEED_COMMIT: | |
04277086 | 835 | kref_get(&req->wb_kref); |
1763da12 | 836 | nfs_mark_request_commit(req, hdr->lseg, &cinfo); |
1763da12 | 837 | } |
1d1afcbc | 838 | nfs_unlock_and_release_request(req); |
1763da12 FI |
839 | } |
840 | ||
841 | out_put: | |
842 | if (put_dreq(dreq)) | |
843 | nfs_direct_write_complete(dreq, hdr->inode); | |
844 | hdr->release(hdr); | |
845 | } | |
846 | ||
3e9e0ca3 TM |
847 | static void nfs_write_sync_pgio_error(struct list_head *head) |
848 | { | |
849 | struct nfs_page *req; | |
850 | ||
851 | while (!list_empty(head)) { | |
852 | req = nfs_list_entry(head->next); | |
853 | nfs_list_remove_request(req); | |
1d1afcbc | 854 | nfs_unlock_and_release_request(req); |
3e9e0ca3 TM |
855 | } |
856 | } | |
857 | ||
1763da12 | 858 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = { |
3e9e0ca3 | 859 | .error_cleanup = nfs_write_sync_pgio_error, |
1763da12 FI |
860 | .init_hdr = nfs_direct_pgio_init, |
861 | .completion = nfs_direct_write_completion, | |
862 | }; | |
863 | ||
19f73787 CL |
864 | static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, |
865 | const struct iovec *iov, | |
866 | unsigned long nr_segs, | |
a564b8f0 | 867 | loff_t pos, bool uio) |
19f73787 | 868 | { |
1763da12 | 869 | struct nfs_pageio_descriptor desc; |
1d59d61f | 870 | struct inode *inode = dreq->inode; |
19f73787 CL |
871 | ssize_t result = 0; |
872 | size_t requested_bytes = 0; | |
873 | unsigned long seg; | |
874 | ||
a20c93e3 | 875 | nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false, |
1763da12 FI |
876 | &nfs_direct_write_completion_ops); |
877 | desc.pg_dreq = dreq; | |
19f73787 | 878 | get_dreq(dreq); |
1d59d61f | 879 | atomic_inc(&inode->i_dio_count); |
19f73787 | 880 | |
7acdb026 | 881 | NFS_I(dreq->inode)->write_io += iov_length(iov, nr_segs); |
19f73787 CL |
882 | for (seg = 0; seg < nr_segs; seg++) { |
883 | const struct iovec *vec = &iov[seg]; | |
a564b8f0 | 884 | result = nfs_direct_write_schedule_segment(&desc, vec, pos, uio); |
19f73787 CL |
885 | if (result < 0) |
886 | break; | |
887 | requested_bytes += result; | |
888 | if ((size_t)result < vec->iov_len) | |
889 | break; | |
890 | pos += vec->iov_len; | |
891 | } | |
1763da12 | 892 | nfs_pageio_complete(&desc); |
19f73787 | 893 | |
839f7ad6 CL |
894 | /* |
895 | * If no bytes were started, return the error, and let the | |
896 | * generic layer handle the completion. | |
897 | */ | |
898 | if (requested_bytes == 0) { | |
1d59d61f | 899 | inode_dio_done(inode); |
839f7ad6 CL |
900 | nfs_direct_req_release(dreq); |
901 | return result < 0 ? result : -EIO; | |
902 | } | |
903 | ||
19f73787 CL |
904 | if (put_dreq(dreq)) |
905 | nfs_direct_write_complete(dreq, dreq->inode); | |
839f7ad6 | 906 | return 0; |
19f73787 CL |
907 | } |
908 | ||
1da177e4 LT |
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 | 931 | ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov, |
a564b8f0 | 932 | unsigned long nr_segs, loff_t pos, bool uio) |
1da177e4 | 933 | { |
22cd1bf1 | 934 | ssize_t result = -EINVAL; |
1da177e4 | 935 | struct file *file = iocb->ki_filp; |
1da177e4 | 936 | struct address_space *mapping = file->f_mapping; |
22cd1bf1 CH |
937 | struct inode *inode = mapping->host; |
938 | struct nfs_direct_req *dreq; | |
939 | struct nfs_lock_context *l_ctx; | |
a9ab5e84 | 940 | loff_t end; |
c216fd70 | 941 | size_t count; |
1da177e4 | 942 | |
c216fd70 | 943 | count = iov_length(iov, nr_segs); |
a9ab5e84 CH |
944 | end = (pos + count - 1) >> PAGE_CACHE_SHIFT; |
945 | ||
c216fd70 CL |
946 | nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count); |
947 | ||
6de1472f AV |
948 | dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n", |
949 | file, count, (long long) pos); | |
027445c3 | 950 | |
22cd1bf1 CH |
951 | result = generic_write_checks(file, &pos, &count, 0); |
952 | if (result) | |
1da177e4 | 953 | goto out; |
ce1a8e67 | 954 | |
22cd1bf1 | 955 | result = -EINVAL; |
ce1a8e67 | 956 | if ((ssize_t) count < 0) |
1da177e4 | 957 | goto out; |
22cd1bf1 | 958 | result = 0; |
1da177e4 LT |
959 | if (!count) |
960 | goto out; | |
ce1a8e67 | 961 | |
a9ab5e84 CH |
962 | mutex_lock(&inode->i_mutex); |
963 | ||
22cd1bf1 CH |
964 | result = nfs_sync_mapping(mapping); |
965 | if (result) | |
a9ab5e84 CH |
966 | goto out_unlock; |
967 | ||
968 | if (mapping->nrpages) { | |
969 | result = invalidate_inode_pages2_range(mapping, | |
970 | pos >> PAGE_CACHE_SHIFT, end); | |
971 | if (result) | |
972 | goto out_unlock; | |
973 | } | |
1da177e4 | 974 | |
7ec10f26 KK |
975 | task_io_account_write(count); |
976 | ||
22cd1bf1 CH |
977 | result = -ENOMEM; |
978 | dreq = nfs_direct_req_alloc(); | |
979 | if (!dreq) | |
a9ab5e84 | 980 | goto out_unlock; |
9eafa8cc | 981 | |
22cd1bf1 CH |
982 | dreq->inode = inode; |
983 | dreq->bytes_left = count; | |
984 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); | |
985 | l_ctx = nfs_get_lock_context(dreq->ctx); | |
986 | if (IS_ERR(l_ctx)) { | |
987 | result = PTR_ERR(l_ctx); | |
988 | goto out_release; | |
989 | } | |
990 | dreq->l_ctx = l_ctx; | |
991 | if (!is_sync_kiocb(iocb)) | |
992 | dreq->iocb = iocb; | |
993 | ||
994 | result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, uio); | |
a9ab5e84 CH |
995 | |
996 | if (mapping->nrpages) { | |
997 | invalidate_inode_pages2_range(mapping, | |
998 | pos >> PAGE_CACHE_SHIFT, end); | |
999 | } | |
1000 | ||
1001 | mutex_unlock(&inode->i_mutex); | |
1002 | ||
22cd1bf1 CH |
1003 | if (!result) { |
1004 | result = nfs_direct_wait(dreq); | |
1005 | if (result > 0) { | |
1006 | struct inode *inode = mapping->host; | |
1007 | ||
1008 | iocb->ki_pos = pos + result; | |
1009 | spin_lock(&inode->i_lock); | |
1010 | if (i_size_read(inode) < iocb->ki_pos) | |
1011 | i_size_write(inode, iocb->ki_pos); | |
1012 | spin_unlock(&inode->i_lock); | |
1013 | } | |
1763da12 | 1014 | } |
a9ab5e84 CH |
1015 | nfs_direct_req_release(dreq); |
1016 | return result; | |
1017 | ||
22cd1bf1 CH |
1018 | out_release: |
1019 | nfs_direct_req_release(dreq); | |
a9ab5e84 CH |
1020 | out_unlock: |
1021 | mutex_unlock(&inode->i_mutex); | |
1da177e4 | 1022 | out: |
22cd1bf1 | 1023 | return result; |
1da177e4 LT |
1024 | } |
1025 | ||
88467055 CL |
1026 | /** |
1027 | * nfs_init_directcache - create a slab cache for nfs_direct_req structures | |
1028 | * | |
1029 | */ | |
f7b422b1 | 1030 | int __init nfs_init_directcache(void) |
1da177e4 LT |
1031 | { |
1032 | nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", | |
1033 | sizeof(struct nfs_direct_req), | |
fffb60f9 PJ |
1034 | 0, (SLAB_RECLAIM_ACCOUNT| |
1035 | SLAB_MEM_SPREAD), | |
20c2df83 | 1036 | NULL); |
1da177e4 LT |
1037 | if (nfs_direct_cachep == NULL) |
1038 | return -ENOMEM; | |
1039 | ||
1040 | return 0; | |
1041 | } | |
1042 | ||
88467055 | 1043 | /** |
f7b422b1 | 1044 | * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures |
88467055 CL |
1045 | * |
1046 | */ | |
266bee88 | 1047 | void nfs_destroy_directcache(void) |
1da177e4 | 1048 | { |
1a1d92c1 | 1049 | kmem_cache_destroy(nfs_direct_cachep); |
1da177e4 | 1050 | } |