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 | ||
5002c586 WAA |
111 | /* |
112 | * nfs_direct_select_verf - select the right verifier | |
113 | * @dreq - direct request possibly spanning multiple servers | |
114 | * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs | |
115 | * @ds_idx - index of data server in data server list, only valid if ds_clp set | |
116 | * | |
117 | * returns the correct verifier to use given the role of the server | |
118 | */ | |
119 | static struct nfs_writeverf * | |
120 | nfs_direct_select_verf(struct nfs_direct_req *dreq, | |
121 | struct nfs_client *ds_clp, | |
122 | int ds_idx) | |
123 | { | |
124 | struct nfs_writeverf *verfp = &dreq->verf; | |
125 | ||
126 | #ifdef CONFIG_NFS_V4_1 | |
127 | if (ds_clp) { | |
128 | /* pNFS is in use, use the DS verf */ | |
129 | if (ds_idx >= 0 && ds_idx < dreq->ds_cinfo.nbuckets) | |
130 | verfp = &dreq->ds_cinfo.buckets[ds_idx].direct_verf; | |
131 | else | |
132 | WARN_ON_ONCE(1); | |
133 | } | |
134 | #endif | |
135 | return verfp; | |
136 | } | |
137 | ||
138 | ||
139 | /* | |
140 | * nfs_direct_set_hdr_verf - set the write/commit verifier | |
141 | * @dreq - direct request possibly spanning multiple servers | |
142 | * @hdr - pageio header to validate against previously seen verfs | |
143 | * | |
144 | * Set the server's (MDS or DS) "seen" verifier | |
145 | */ | |
146 | static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq, | |
147 | struct nfs_pgio_header *hdr) | |
148 | { | |
149 | struct nfs_writeverf *verfp; | |
150 | ||
151 | verfp = nfs_direct_select_verf(dreq, hdr->data->ds_clp, | |
152 | hdr->data->ds_idx); | |
153 | WARN_ON_ONCE(verfp->committed >= 0); | |
154 | memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf)); | |
155 | WARN_ON_ONCE(verfp->committed < 0); | |
156 | } | |
157 | ||
158 | /* | |
159 | * nfs_direct_cmp_hdr_verf - compare verifier for pgio header | |
160 | * @dreq - direct request possibly spanning multiple servers | |
161 | * @hdr - pageio header to validate against previously seen verf | |
162 | * | |
163 | * set the server's "seen" verf if not initialized. | |
164 | * returns result of comparison between @hdr->verf and the "seen" | |
165 | * verf of the server used by @hdr (DS or MDS) | |
166 | */ | |
167 | static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq, | |
168 | struct nfs_pgio_header *hdr) | |
169 | { | |
170 | struct nfs_writeverf *verfp; | |
171 | ||
172 | verfp = nfs_direct_select_verf(dreq, hdr->data->ds_clp, | |
173 | hdr->data->ds_idx); | |
174 | if (verfp->committed < 0) { | |
175 | nfs_direct_set_hdr_verf(dreq, hdr); | |
176 | return 0; | |
177 | } | |
178 | return memcmp(verfp, &hdr->verf, sizeof(struct nfs_writeverf)); | |
179 | } | |
180 | ||
181 | #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) | |
182 | /* | |
183 | * nfs_direct_cmp_commit_data_verf - compare verifier for commit data | |
184 | * @dreq - direct request possibly spanning multiple servers | |
185 | * @data - commit data to validate against previously seen verf | |
186 | * | |
187 | * returns result of comparison between @data->verf and the verf of | |
188 | * the server used by @data (DS or MDS) | |
189 | */ | |
190 | static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq, | |
191 | struct nfs_commit_data *data) | |
192 | { | |
193 | struct nfs_writeverf *verfp; | |
194 | ||
195 | verfp = nfs_direct_select_verf(dreq, data->ds_clp, | |
196 | data->ds_commit_index); | |
197 | WARN_ON_ONCE(verfp->committed < 0); | |
198 | return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf)); | |
199 | } | |
200 | #endif | |
201 | ||
1da177e4 | 202 | /** |
b8a32e2b CL |
203 | * nfs_direct_IO - NFS address space operation for direct I/O |
204 | * @rw: direction (read or write) | |
205 | * @iocb: target I/O control block | |
206 | * @iov: array of vectors that define I/O buffer | |
207 | * @pos: offset in file to begin the operation | |
208 | * @nr_segs: size of iovec array | |
209 | * | |
210 | * The presence of this routine in the address space ops vector means | |
a564b8f0 MG |
211 | * the NFS client supports direct I/O. However, for most direct IO, we |
212 | * shunt off direct read and write requests before the VFS gets them, | |
213 | * so this method is only ever called for swap. | |
1da177e4 | 214 | */ |
b8a32e2b CL |
215 | ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs) |
216 | { | |
a564b8f0 | 217 | #ifndef CONFIG_NFS_SWAP |
6de1472f AV |
218 | dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n", |
219 | iocb->ki_filp, (long long) pos, nr_segs); | |
b8a32e2b CL |
220 | |
221 | return -EINVAL; | |
a564b8f0 | 222 | #else |
a564b8f0 MG |
223 | VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE); |
224 | ||
225 | if (rw == READ || rw == KERNEL_READ) | |
226 | return nfs_file_direct_read(iocb, iov, nr_segs, pos, | |
227 | rw == READ ? true : false); | |
228 | return nfs_file_direct_write(iocb, iov, nr_segs, pos, | |
229 | rw == WRITE ? true : false); | |
230 | #endif /* CONFIG_NFS_SWAP */ | |
b8a32e2b CL |
231 | } |
232 | ||
749e146e | 233 | static void nfs_direct_release_pages(struct page **pages, unsigned int npages) |
9c93ab7d | 234 | { |
749e146e | 235 | unsigned int i; |
607f31e8 TM |
236 | for (i = 0; i < npages; i++) |
237 | page_cache_release(pages[i]); | |
6b45d858 TM |
238 | } |
239 | ||
1763da12 FI |
240 | void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo, |
241 | struct nfs_direct_req *dreq) | |
242 | { | |
243 | cinfo->lock = &dreq->lock; | |
244 | cinfo->mds = &dreq->mds_cinfo; | |
245 | cinfo->ds = &dreq->ds_cinfo; | |
246 | cinfo->dreq = dreq; | |
247 | cinfo->completion_ops = &nfs_direct_commit_completion_ops; | |
248 | } | |
249 | ||
93619e59 | 250 | static inline struct nfs_direct_req *nfs_direct_req_alloc(void) |
1da177e4 | 251 | { |
93619e59 CL |
252 | struct nfs_direct_req *dreq; |
253 | ||
292f3eee | 254 | dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL); |
93619e59 CL |
255 | if (!dreq) |
256 | return NULL; | |
257 | ||
258 | kref_init(&dreq->kref); | |
607f31e8 | 259 | kref_get(&dreq->kref); |
d72b7a6b | 260 | init_completion(&dreq->completion); |
1763da12 | 261 | INIT_LIST_HEAD(&dreq->mds_cinfo.list); |
5002c586 | 262 | dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */ |
1763da12 | 263 | INIT_WORK(&dreq->work, nfs_direct_write_schedule_work); |
15ce4a0c | 264 | spin_lock_init(&dreq->lock); |
93619e59 CL |
265 | |
266 | return dreq; | |
1da177e4 LT |
267 | } |
268 | ||
b4946ffb | 269 | static void nfs_direct_req_free(struct kref *kref) |
1da177e4 LT |
270 | { |
271 | struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); | |
a8881f5a | 272 | |
f11ac8db TM |
273 | if (dreq->l_ctx != NULL) |
274 | nfs_put_lock_context(dreq->l_ctx); | |
a8881f5a TM |
275 | if (dreq->ctx != NULL) |
276 | put_nfs_open_context(dreq->ctx); | |
1da177e4 LT |
277 | kmem_cache_free(nfs_direct_cachep, dreq); |
278 | } | |
279 | ||
b4946ffb TM |
280 | static void nfs_direct_req_release(struct nfs_direct_req *dreq) |
281 | { | |
282 | kref_put(&dreq->kref, nfs_direct_req_free); | |
283 | } | |
284 | ||
6296556f PT |
285 | ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq) |
286 | { | |
287 | return dreq->bytes_left; | |
288 | } | |
289 | EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left); | |
290 | ||
bc0fb201 CL |
291 | /* |
292 | * Collects and returns the final error value/byte-count. | |
293 | */ | |
294 | static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) | |
295 | { | |
15ce4a0c | 296 | ssize_t result = -EIOCBQUEUED; |
bc0fb201 CL |
297 | |
298 | /* Async requests don't wait here */ | |
299 | if (dreq->iocb) | |
300 | goto out; | |
301 | ||
150030b7 | 302 | result = wait_for_completion_killable(&dreq->completion); |
bc0fb201 CL |
303 | |
304 | if (!result) | |
15ce4a0c | 305 | result = dreq->error; |
bc0fb201 | 306 | if (!result) |
15ce4a0c | 307 | result = dreq->count; |
bc0fb201 CL |
308 | |
309 | out: | |
bc0fb201 CL |
310 | return (ssize_t) result; |
311 | } | |
312 | ||
63ab46ab | 313 | /* |
607f31e8 TM |
314 | * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust |
315 | * the iocb is still valid here if this is a synchronous request. | |
63ab46ab | 316 | */ |
9811cd57 | 317 | static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write) |
63ab46ab | 318 | { |
9811cd57 CH |
319 | struct inode *inode = dreq->inode; |
320 | ||
2a009ec9 | 321 | if (dreq->iocb && write) { |
9811cd57 | 322 | loff_t pos = dreq->iocb->ki_pos + dreq->count; |
2a009ec9 CH |
323 | |
324 | spin_lock(&inode->i_lock); | |
325 | if (i_size_read(inode) < pos) | |
326 | i_size_write(inode, pos); | |
327 | spin_unlock(&inode->i_lock); | |
328 | } | |
329 | ||
1f90ee27 | 330 | if (write) |
2a009ec9 | 331 | nfs_zap_mapping(inode, inode->i_mapping); |
1f90ee27 CH |
332 | |
333 | inode_dio_done(inode); | |
2a009ec9 CH |
334 | |
335 | if (dreq->iocb) { | |
15ce4a0c | 336 | long res = (long) dreq->error; |
63ab46ab | 337 | if (!res) |
15ce4a0c | 338 | res = (long) dreq->count; |
63ab46ab | 339 | aio_complete(dreq->iocb, res, 0); |
d72b7a6b | 340 | } |
2a009ec9 | 341 | |
d72b7a6b | 342 | complete_all(&dreq->completion); |
63ab46ab | 343 | |
b4946ffb | 344 | nfs_direct_req_release(dreq); |
63ab46ab CL |
345 | } |
346 | ||
1385b811 | 347 | static void nfs_direct_readpage_release(struct nfs_page *req) |
1da177e4 | 348 | { |
1e8968c5 | 349 | dprintk("NFS: direct read done (%s/%llu %d@%lld)\n", |
584aa810 | 350 | req->wb_context->dentry->d_inode->i_sb->s_id, |
1e8968c5 | 351 | (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode), |
584aa810 FI |
352 | req->wb_bytes, |
353 | (long long)req_offset(req)); | |
354 | nfs_release_request(req); | |
fdd1e74c TM |
355 | } |
356 | ||
584aa810 | 357 | static void nfs_direct_read_completion(struct nfs_pgio_header *hdr) |
fdd1e74c | 358 | { |
584aa810 FI |
359 | unsigned long bytes = 0; |
360 | struct nfs_direct_req *dreq = hdr->dreq; | |
fdd1e74c | 361 | |
584aa810 FI |
362 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) |
363 | goto out_put; | |
15ce4a0c CL |
364 | |
365 | spin_lock(&dreq->lock); | |
584aa810 FI |
366 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0)) |
367 | dreq->error = hdr->error; | |
368 | else | |
369 | dreq->count += hdr->good_bytes; | |
370 | spin_unlock(&dreq->lock); | |
371 | ||
4bd8b010 TM |
372 | while (!list_empty(&hdr->pages)) { |
373 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
374 | struct page *page = req->wb_page; | |
375 | ||
be7e9858 JL |
376 | if (!PageCompound(page) && bytes < hdr->good_bytes) |
377 | set_page_dirty(page); | |
4bd8b010 TM |
378 | bytes += req->wb_bytes; |
379 | nfs_list_remove_request(req); | |
380 | nfs_direct_readpage_release(req); | |
d4a8f367 | 381 | } |
584aa810 | 382 | out_put: |
607f31e8 | 383 | if (put_dreq(dreq)) |
9811cd57 | 384 | nfs_direct_complete(dreq, false); |
584aa810 | 385 | hdr->release(hdr); |
1da177e4 LT |
386 | } |
387 | ||
3e9e0ca3 | 388 | static void nfs_read_sync_pgio_error(struct list_head *head) |
cd841605 | 389 | { |
584aa810 | 390 | struct nfs_page *req; |
cd841605 | 391 | |
584aa810 FI |
392 | while (!list_empty(head)) { |
393 | req = nfs_list_entry(head->next); | |
394 | nfs_list_remove_request(req); | |
395 | nfs_release_request(req); | |
396 | } | |
cd841605 FI |
397 | } |
398 | ||
584aa810 FI |
399 | static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr) |
400 | { | |
401 | get_dreq(hdr->dreq); | |
402 | } | |
403 | ||
404 | static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = { | |
3e9e0ca3 | 405 | .error_cleanup = nfs_read_sync_pgio_error, |
584aa810 FI |
406 | .init_hdr = nfs_direct_pgio_init, |
407 | .completion = nfs_direct_read_completion, | |
408 | }; | |
409 | ||
d4cc948b | 410 | /* |
607f31e8 TM |
411 | * For each rsize'd chunk of the user's buffer, dispatch an NFS READ |
412 | * operation. If nfs_readdata_alloc() or get_user_pages() fails, | |
413 | * bail and stop sending more reads. Read length accounting is | |
414 | * handled automatically by nfs_direct_read_result(). Otherwise, if | |
415 | * no requests have been sent, just return an error. | |
1da177e4 | 416 | */ |
584aa810 | 417 | static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc, |
02fe4946 | 418 | const struct iovec *iov, |
a564b8f0 | 419 | loff_t pos, bool uio) |
1da177e4 | 420 | { |
584aa810 | 421 | struct nfs_direct_req *dreq = desc->pg_dreq; |
a8881f5a | 422 | struct nfs_open_context *ctx = dreq->ctx; |
3d4ff43d | 423 | struct inode *inode = ctx->dentry->d_inode; |
02fe4946 CL |
424 | unsigned long user_addr = (unsigned long)iov->iov_base; |
425 | size_t count = iov->iov_len; | |
5dd602f2 | 426 | size_t rsize = NFS_SERVER(inode)->rsize; |
607f31e8 TM |
427 | unsigned int pgbase; |
428 | int result; | |
429 | ssize_t started = 0; | |
584aa810 FI |
430 | struct page **pagevec = NULL; |
431 | unsigned int npages; | |
607f31e8 | 432 | |
1da177e4 | 433 | do { |
5dd602f2 | 434 | size_t bytes; |
584aa810 | 435 | int i; |
1da177e4 | 436 | |
e9f7bee1 | 437 | pgbase = user_addr & ~PAGE_MASK; |
bf5fc402 | 438 | bytes = min(max_t(size_t, rsize, PAGE_SIZE), count); |
e9f7bee1 | 439 | |
607f31e8 | 440 | result = -ENOMEM; |
584aa810 FI |
441 | npages = nfs_page_array_len(pgbase, bytes); |
442 | if (!pagevec) | |
443 | pagevec = kmalloc(npages * sizeof(struct page *), | |
444 | GFP_KERNEL); | |
445 | if (!pagevec) | |
4db6e0b7 | 446 | break; |
a564b8f0 MG |
447 | if (uio) { |
448 | down_read(¤t->mm->mmap_sem); | |
449 | result = get_user_pages(current, current->mm, user_addr, | |
584aa810 | 450 | npages, 1, 0, pagevec, NULL); |
a564b8f0 MG |
451 | up_read(¤t->mm->mmap_sem); |
452 | if (result < 0) | |
453 | break; | |
454 | } else { | |
455 | WARN_ON(npages != 1); | |
456 | result = get_kernel_page(user_addr, 1, pagevec); | |
457 | if (WARN_ON(result != 1)) | |
458 | break; | |
459 | } | |
460 | ||
584aa810 | 461 | if ((unsigned)result < npages) { |
d9df8d6b TM |
462 | bytes = result * PAGE_SIZE; |
463 | if (bytes <= pgbase) { | |
584aa810 | 464 | nfs_direct_release_pages(pagevec, result); |
d9df8d6b TM |
465 | break; |
466 | } | |
467 | bytes -= pgbase; | |
584aa810 | 468 | npages = result; |
607f31e8 TM |
469 | } |
470 | ||
584aa810 FI |
471 | for (i = 0; i < npages; i++) { |
472 | struct nfs_page *req; | |
bf5fc402 | 473 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); |
584aa810 | 474 | /* XXX do we need to do the eof zeroing found in async_filler? */ |
2bfc6e56 | 475 | req = nfs_create_request(dreq->ctx, pagevec[i], NULL, |
584aa810 FI |
476 | pgbase, req_len); |
477 | if (IS_ERR(req)) { | |
584aa810 FI |
478 | result = PTR_ERR(req); |
479 | break; | |
480 | } | |
481 | req->wb_index = pos >> PAGE_SHIFT; | |
482 | req->wb_offset = pos & ~PAGE_MASK; | |
483 | if (!nfs_pageio_add_request(desc, req)) { | |
484 | result = desc->pg_error; | |
485 | nfs_release_request(req); | |
584aa810 FI |
486 | break; |
487 | } | |
488 | pgbase = 0; | |
489 | bytes -= req_len; | |
490 | started += req_len; | |
491 | user_addr += req_len; | |
492 | pos += req_len; | |
493 | count -= req_len; | |
35754bc0 | 494 | dreq->bytes_left -= req_len; |
584aa810 | 495 | } |
6d74743b TM |
496 | /* The nfs_page now hold references to these pages */ |
497 | nfs_direct_release_pages(pagevec, npages); | |
71e8cc00 | 498 | } while (count != 0 && result >= 0); |
607f31e8 | 499 | |
584aa810 FI |
500 | kfree(pagevec); |
501 | ||
607f31e8 | 502 | if (started) |
c216fd70 | 503 | return started; |
607f31e8 | 504 | return result < 0 ? (ssize_t) result : -EFAULT; |
1da177e4 LT |
505 | } |
506 | ||
19f73787 CL |
507 | static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq, |
508 | const struct iovec *iov, | |
509 | unsigned long nr_segs, | |
a564b8f0 | 510 | loff_t pos, bool uio) |
19f73787 | 511 | { |
584aa810 | 512 | struct nfs_pageio_descriptor desc; |
1f90ee27 | 513 | struct inode *inode = dreq->inode; |
19f73787 CL |
514 | ssize_t result = -EINVAL; |
515 | size_t requested_bytes = 0; | |
516 | unsigned long seg; | |
517 | ||
fab5fc25 | 518 | nfs_pageio_init_read(&desc, dreq->inode, false, |
584aa810 | 519 | &nfs_direct_read_completion_ops); |
19f73787 | 520 | get_dreq(dreq); |
584aa810 | 521 | desc.pg_dreq = dreq; |
1f90ee27 | 522 | atomic_inc(&inode->i_dio_count); |
19f73787 CL |
523 | |
524 | for (seg = 0; seg < nr_segs; seg++) { | |
525 | const struct iovec *vec = &iov[seg]; | |
a564b8f0 | 526 | result = nfs_direct_read_schedule_segment(&desc, vec, pos, uio); |
19f73787 CL |
527 | if (result < 0) |
528 | break; | |
529 | requested_bytes += result; | |
530 | if ((size_t)result < vec->iov_len) | |
531 | break; | |
532 | pos += vec->iov_len; | |
533 | } | |
534 | ||
584aa810 FI |
535 | nfs_pageio_complete(&desc); |
536 | ||
839f7ad6 CL |
537 | /* |
538 | * If no bytes were started, return the error, and let the | |
539 | * generic layer handle the completion. | |
540 | */ | |
541 | if (requested_bytes == 0) { | |
1f90ee27 | 542 | inode_dio_done(inode); |
839f7ad6 CL |
543 | nfs_direct_req_release(dreq); |
544 | return result < 0 ? result : -EIO; | |
545 | } | |
546 | ||
19f73787 | 547 | if (put_dreq(dreq)) |
9811cd57 | 548 | nfs_direct_complete(dreq, false); |
839f7ad6 | 549 | return 0; |
19f73787 CL |
550 | } |
551 | ||
14a3ec79 CH |
552 | /** |
553 | * nfs_file_direct_read - file direct read operation for NFS files | |
554 | * @iocb: target I/O control block | |
555 | * @iov: vector of user buffers into which to read data | |
556 | * @nr_segs: size of iov vector | |
557 | * @pos: byte offset in file where reading starts | |
558 | * | |
559 | * We use this function for direct reads instead of calling | |
560 | * generic_file_aio_read() in order to avoid gfar's check to see if | |
561 | * the request starts before the end of the file. For that check | |
562 | * to work, we must generate a GETATTR before each direct read, and | |
563 | * even then there is a window between the GETATTR and the subsequent | |
564 | * READ where the file size could change. Our preference is simply | |
565 | * to do all reads the application wants, and the server will take | |
566 | * care of managing the end of file boundary. | |
567 | * | |
568 | * This function also eliminates unnecessarily updating the file's | |
569 | * atime locally, as the NFS server sets the file's atime, and this | |
570 | * client must read the updated atime from the server back into its | |
571 | * cache. | |
572 | */ | |
573 | ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov, | |
574 | unsigned long nr_segs, loff_t pos, bool uio) | |
1da177e4 | 575 | { |
14a3ec79 CH |
576 | struct file *file = iocb->ki_filp; |
577 | struct address_space *mapping = file->f_mapping; | |
578 | struct inode *inode = mapping->host; | |
1da177e4 | 579 | struct nfs_direct_req *dreq; |
b3c54de6 | 580 | struct nfs_lock_context *l_ctx; |
14a3ec79 CH |
581 | ssize_t result = -EINVAL; |
582 | size_t count; | |
583 | ||
584 | count = iov_length(iov, nr_segs); | |
585 | nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); | |
586 | ||
587 | dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n", | |
588 | file, count, (long long) pos); | |
589 | ||
590 | result = 0; | |
591 | if (!count) | |
592 | goto out; | |
593 | ||
d0b9875d | 594 | mutex_lock(&inode->i_mutex); |
14a3ec79 CH |
595 | result = nfs_sync_mapping(mapping); |
596 | if (result) | |
d0b9875d | 597 | goto out_unlock; |
1da177e4 | 598 | |
14a3ec79 CH |
599 | task_io_account_read(count); |
600 | ||
601 | result = -ENOMEM; | |
607f31e8 | 602 | dreq = nfs_direct_req_alloc(); |
f11ac8db | 603 | if (dreq == NULL) |
d0b9875d | 604 | goto out_unlock; |
1da177e4 | 605 | |
91d5b470 | 606 | dreq->inode = inode; |
35754bc0 | 607 | dreq->bytes_left = iov_length(iov, nr_segs); |
cd3758e3 | 608 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); |
b3c54de6 TM |
609 | l_ctx = nfs_get_lock_context(dreq->ctx); |
610 | if (IS_ERR(l_ctx)) { | |
611 | result = PTR_ERR(l_ctx); | |
f11ac8db | 612 | goto out_release; |
b3c54de6 TM |
613 | } |
614 | dreq->l_ctx = l_ctx; | |
487b8372 CL |
615 | if (!is_sync_kiocb(iocb)) |
616 | dreq->iocb = iocb; | |
1da177e4 | 617 | |
7acdb026 | 618 | NFS_I(inode)->read_io += iov_length(iov, nr_segs); |
a564b8f0 | 619 | result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos, uio); |
d0b9875d CH |
620 | |
621 | mutex_unlock(&inode->i_mutex); | |
622 | ||
14a3ec79 | 623 | if (!result) { |
607f31e8 | 624 | result = nfs_direct_wait(dreq); |
14a3ec79 CH |
625 | if (result > 0) |
626 | iocb->ki_pos = pos + result; | |
627 | } | |
d0b9875d CH |
628 | |
629 | nfs_direct_req_release(dreq); | |
630 | return result; | |
631 | ||
f11ac8db | 632 | out_release: |
b4946ffb | 633 | nfs_direct_req_release(dreq); |
d0b9875d CH |
634 | out_unlock: |
635 | mutex_unlock(&inode->i_mutex); | |
f11ac8db | 636 | out: |
1da177e4 LT |
637 | return result; |
638 | } | |
639 | ||
89d77c8f | 640 | #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) |
fad61490 TM |
641 | static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) |
642 | { | |
1763da12 FI |
643 | struct nfs_pageio_descriptor desc; |
644 | struct nfs_page *req, *tmp; | |
645 | LIST_HEAD(reqs); | |
646 | struct nfs_commit_info cinfo; | |
647 | LIST_HEAD(failed); | |
648 | ||
649 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
650 | pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo); | |
651 | spin_lock(cinfo.lock); | |
652 | nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0); | |
653 | spin_unlock(cinfo.lock); | |
1da177e4 | 654 | |
fad61490 | 655 | dreq->count = 0; |
607f31e8 TM |
656 | get_dreq(dreq); |
657 | ||
a20c93e3 | 658 | nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false, |
1763da12 FI |
659 | &nfs_direct_write_completion_ops); |
660 | desc.pg_dreq = dreq; | |
fedb595c | 661 | |
1763da12 FI |
662 | list_for_each_entry_safe(req, tmp, &reqs, wb_list) { |
663 | if (!nfs_pageio_add_request(&desc, req)) { | |
4035c248 | 664 | nfs_list_remove_request(req); |
1763da12 FI |
665 | nfs_list_add_request(req, &failed); |
666 | spin_lock(cinfo.lock); | |
667 | dreq->flags = 0; | |
668 | dreq->error = -EIO; | |
669 | spin_unlock(cinfo.lock); | |
670 | } | |
5a695da2 | 671 | nfs_release_request(req); |
1763da12 FI |
672 | } |
673 | nfs_pageio_complete(&desc); | |
fad61490 | 674 | |
4035c248 TM |
675 | while (!list_empty(&failed)) { |
676 | req = nfs_list_entry(failed.next); | |
677 | nfs_list_remove_request(req); | |
1d1afcbc | 678 | nfs_unlock_and_release_request(req); |
4035c248 | 679 | } |
fad61490 | 680 | |
1763da12 FI |
681 | if (put_dreq(dreq)) |
682 | nfs_direct_write_complete(dreq, dreq->inode); | |
c9d8f89d TM |
683 | } |
684 | ||
1763da12 | 685 | static void nfs_direct_commit_complete(struct nfs_commit_data *data) |
c9d8f89d | 686 | { |
0b7c0153 | 687 | struct nfs_direct_req *dreq = data->dreq; |
1763da12 FI |
688 | struct nfs_commit_info cinfo; |
689 | struct nfs_page *req; | |
c9d8f89d TM |
690 | int status = data->task.tk_status; |
691 | ||
1763da12 | 692 | nfs_init_cinfo_from_dreq(&cinfo, dreq); |
c9d8f89d | 693 | if (status < 0) { |
60fa3f76 | 694 | dprintk("NFS: %5u commit failed with error %d.\n", |
1763da12 | 695 | data->task.tk_pid, status); |
fad61490 | 696 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
5002c586 | 697 | } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) { |
c9d8f89d | 698 | dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid); |
fad61490 | 699 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
1da177e4 LT |
700 | } |
701 | ||
c9d8f89d | 702 | dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status); |
1763da12 FI |
703 | while (!list_empty(&data->pages)) { |
704 | req = nfs_list_entry(data->pages.next); | |
705 | nfs_list_remove_request(req); | |
706 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) { | |
707 | /* Note the rewrite will go through mds */ | |
708 | nfs_mark_request_commit(req, NULL, &cinfo); | |
906369e4 FI |
709 | } else |
710 | nfs_release_request(req); | |
1d1afcbc | 711 | nfs_unlock_and_release_request(req); |
1763da12 FI |
712 | } |
713 | ||
714 | if (atomic_dec_and_test(&cinfo.mds->rpcs_out)) | |
715 | nfs_direct_write_complete(dreq, data->inode); | |
1da177e4 LT |
716 | } |
717 | ||
1763da12 FI |
718 | static void nfs_direct_error_cleanup(struct nfs_inode *nfsi) |
719 | { | |
720 | /* There is no lock to clear */ | |
721 | } | |
722 | ||
723 | static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = { | |
724 | .completion = nfs_direct_commit_complete, | |
725 | .error_cleanup = nfs_direct_error_cleanup, | |
fad61490 TM |
726 | }; |
727 | ||
728 | static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) | |
1da177e4 | 729 | { |
1763da12 FI |
730 | int res; |
731 | struct nfs_commit_info cinfo; | |
732 | LIST_HEAD(mds_list); | |
733 | ||
734 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
735 | nfs_scan_commit(dreq->inode, &mds_list, &cinfo); | |
736 | res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo); | |
737 | if (res < 0) /* res == -ENOMEM */ | |
738 | nfs_direct_write_reschedule(dreq); | |
fad61490 | 739 | } |
1da177e4 | 740 | |
1763da12 | 741 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
fad61490 | 742 | { |
1763da12 | 743 | struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work); |
fad61490 | 744 | int flags = dreq->flags; |
1da177e4 | 745 | |
fad61490 TM |
746 | dreq->flags = 0; |
747 | switch (flags) { | |
748 | case NFS_ODIRECT_DO_COMMIT: | |
749 | nfs_direct_commit_schedule(dreq); | |
1da177e4 | 750 | break; |
fad61490 TM |
751 | case NFS_ODIRECT_RESCHED_WRITES: |
752 | nfs_direct_write_reschedule(dreq); | |
753 | break; | |
754 | default: | |
9811cd57 | 755 | nfs_direct_complete(dreq, true); |
fad61490 TM |
756 | } |
757 | } | |
1da177e4 | 758 | |
1763da12 | 759 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
fad61490 | 760 | { |
1763da12 | 761 | schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */ |
fad61490 | 762 | } |
1763da12 | 763 | |
fad61490 | 764 | #else |
24fc9211 BS |
765 | static void nfs_direct_write_schedule_work(struct work_struct *work) |
766 | { | |
767 | } | |
1da177e4 | 768 | |
fad61490 TM |
769 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
770 | { | |
9811cd57 | 771 | nfs_direct_complete(dreq, true); |
fad61490 TM |
772 | } |
773 | #endif | |
1da177e4 | 774 | |
c9d8f89d TM |
775 | /* |
776 | * NB: Return the value of the first error return code. Subsequent | |
777 | * errors after the first one are ignored. | |
778 | */ | |
462d5b32 | 779 | /* |
607f31e8 TM |
780 | * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE |
781 | * operation. If nfs_writedata_alloc() or get_user_pages() fails, | |
782 | * bail and stop sending more writes. Write length accounting is | |
783 | * handled automatically by nfs_direct_write_result(). Otherwise, if | |
784 | * no requests have been sent, just return an error. | |
462d5b32 | 785 | */ |
1763da12 | 786 | static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc, |
02fe4946 | 787 | const struct iovec *iov, |
a564b8f0 | 788 | loff_t pos, bool uio) |
462d5b32 | 789 | { |
1763da12 | 790 | struct nfs_direct_req *dreq = desc->pg_dreq; |
a8881f5a | 791 | struct nfs_open_context *ctx = dreq->ctx; |
3d4ff43d | 792 | struct inode *inode = ctx->dentry->d_inode; |
02fe4946 CL |
793 | unsigned long user_addr = (unsigned long)iov->iov_base; |
794 | size_t count = iov->iov_len; | |
462d5b32 | 795 | size_t wsize = NFS_SERVER(inode)->wsize; |
607f31e8 TM |
796 | unsigned int pgbase; |
797 | int result; | |
798 | ssize_t started = 0; | |
1763da12 FI |
799 | struct page **pagevec = NULL; |
800 | unsigned int npages; | |
82b145c5 | 801 | |
1da177e4 | 802 | do { |
462d5b32 | 803 | size_t bytes; |
1763da12 | 804 | int i; |
462d5b32 | 805 | |
e9f7bee1 | 806 | pgbase = user_addr & ~PAGE_MASK; |
bf5fc402 | 807 | bytes = min(max_t(size_t, wsize, PAGE_SIZE), count); |
e9f7bee1 | 808 | |
607f31e8 | 809 | result = -ENOMEM; |
1763da12 FI |
810 | npages = nfs_page_array_len(pgbase, bytes); |
811 | if (!pagevec) | |
812 | pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL); | |
813 | if (!pagevec) | |
607f31e8 TM |
814 | break; |
815 | ||
a564b8f0 MG |
816 | if (uio) { |
817 | down_read(¤t->mm->mmap_sem); | |
818 | result = get_user_pages(current, current->mm, user_addr, | |
819 | npages, 0, 0, pagevec, NULL); | |
820 | up_read(¤t->mm->mmap_sem); | |
821 | if (result < 0) | |
822 | break; | |
823 | } else { | |
824 | WARN_ON(npages != 1); | |
825 | result = get_kernel_page(user_addr, 0, pagevec); | |
826 | if (WARN_ON(result != 1)) | |
827 | break; | |
828 | } | |
1763da12 FI |
829 | |
830 | if ((unsigned)result < npages) { | |
d9df8d6b TM |
831 | bytes = result * PAGE_SIZE; |
832 | if (bytes <= pgbase) { | |
1763da12 | 833 | nfs_direct_release_pages(pagevec, result); |
d9df8d6b TM |
834 | break; |
835 | } | |
836 | bytes -= pgbase; | |
1763da12 | 837 | npages = result; |
607f31e8 TM |
838 | } |
839 | ||
1763da12 FI |
840 | for (i = 0; i < npages; i++) { |
841 | struct nfs_page *req; | |
bf5fc402 | 842 | unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); |
1da177e4 | 843 | |
2bfc6e56 | 844 | req = nfs_create_request(dreq->ctx, pagevec[i], NULL, |
1763da12 FI |
845 | pgbase, req_len); |
846 | if (IS_ERR(req)) { | |
1763da12 FI |
847 | result = PTR_ERR(req); |
848 | break; | |
849 | } | |
850 | nfs_lock_request(req); | |
851 | req->wb_index = pos >> PAGE_SHIFT; | |
852 | req->wb_offset = pos & ~PAGE_MASK; | |
853 | if (!nfs_pageio_add_request(desc, req)) { | |
854 | result = desc->pg_error; | |
1d1afcbc | 855 | nfs_unlock_and_release_request(req); |
71e8cc00 | 856 | break; |
1763da12 FI |
857 | } |
858 | pgbase = 0; | |
859 | bytes -= req_len; | |
860 | started += req_len; | |
861 | user_addr += req_len; | |
862 | pos += req_len; | |
863 | count -= req_len; | |
35754bc0 | 864 | dreq->bytes_left -= req_len; |
1763da12 | 865 | } |
6d74743b TM |
866 | /* The nfs_page now hold references to these pages */ |
867 | nfs_direct_release_pages(pagevec, npages); | |
71e8cc00 | 868 | } while (count != 0 && result >= 0); |
607f31e8 | 869 | |
1763da12 FI |
870 | kfree(pagevec); |
871 | ||
607f31e8 | 872 | if (started) |
c216fd70 | 873 | return started; |
607f31e8 | 874 | return result < 0 ? (ssize_t) result : -EFAULT; |
462d5b32 | 875 | } |
1da177e4 | 876 | |
1763da12 FI |
877 | static void nfs_direct_write_completion(struct nfs_pgio_header *hdr) |
878 | { | |
879 | struct nfs_direct_req *dreq = hdr->dreq; | |
880 | struct nfs_commit_info cinfo; | |
881 | int bit = -1; | |
882 | struct nfs_page *req = nfs_list_entry(hdr->pages.next); | |
883 | ||
884 | if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) | |
885 | goto out_put; | |
886 | ||
887 | nfs_init_cinfo_from_dreq(&cinfo, dreq); | |
888 | ||
889 | spin_lock(&dreq->lock); | |
890 | ||
891 | if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) { | |
892 | dreq->flags = 0; | |
893 | dreq->error = hdr->error; | |
894 | } | |
895 | if (dreq->error != 0) | |
896 | bit = NFS_IOHDR_ERROR; | |
897 | else { | |
898 | dreq->count += hdr->good_bytes; | |
899 | if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) { | |
900 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; | |
901 | bit = NFS_IOHDR_NEED_RESCHED; | |
902 | } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) { | |
903 | if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) | |
904 | bit = NFS_IOHDR_NEED_RESCHED; | |
905 | else if (dreq->flags == 0) { | |
5002c586 | 906 | nfs_direct_set_hdr_verf(dreq, hdr); |
1763da12 FI |
907 | bit = NFS_IOHDR_NEED_COMMIT; |
908 | dreq->flags = NFS_ODIRECT_DO_COMMIT; | |
909 | } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) { | |
5002c586 WAA |
910 | if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr)) { |
911 | dreq->flags = | |
912 | NFS_ODIRECT_RESCHED_WRITES; | |
1763da12 FI |
913 | bit = NFS_IOHDR_NEED_RESCHED; |
914 | } else | |
915 | bit = NFS_IOHDR_NEED_COMMIT; | |
916 | } | |
917 | } | |
918 | } | |
919 | spin_unlock(&dreq->lock); | |
920 | ||
921 | while (!list_empty(&hdr->pages)) { | |
2bfc6e56 WAA |
922 | bool do_destroy = true; |
923 | ||
1763da12 FI |
924 | req = nfs_list_entry(hdr->pages.next); |
925 | nfs_list_remove_request(req); | |
926 | switch (bit) { | |
927 | case NFS_IOHDR_NEED_RESCHED: | |
928 | case NFS_IOHDR_NEED_COMMIT: | |
04277086 | 929 | kref_get(&req->wb_kref); |
1763da12 | 930 | nfs_mark_request_commit(req, hdr->lseg, &cinfo); |
2bfc6e56 | 931 | do_destroy = false; |
1763da12 | 932 | } |
1d1afcbc | 933 | nfs_unlock_and_release_request(req); |
1763da12 FI |
934 | } |
935 | ||
936 | out_put: | |
937 | if (put_dreq(dreq)) | |
938 | nfs_direct_write_complete(dreq, hdr->inode); | |
939 | hdr->release(hdr); | |
940 | } | |
941 | ||
3e9e0ca3 TM |
942 | static void nfs_write_sync_pgio_error(struct list_head *head) |
943 | { | |
944 | struct nfs_page *req; | |
945 | ||
946 | while (!list_empty(head)) { | |
947 | req = nfs_list_entry(head->next); | |
948 | nfs_list_remove_request(req); | |
1d1afcbc | 949 | nfs_unlock_and_release_request(req); |
3e9e0ca3 TM |
950 | } |
951 | } | |
952 | ||
1763da12 | 953 | static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = { |
3e9e0ca3 | 954 | .error_cleanup = nfs_write_sync_pgio_error, |
1763da12 FI |
955 | .init_hdr = nfs_direct_pgio_init, |
956 | .completion = nfs_direct_write_completion, | |
957 | }; | |
958 | ||
19f73787 CL |
959 | static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, |
960 | const struct iovec *iov, | |
961 | unsigned long nr_segs, | |
a564b8f0 | 962 | loff_t pos, bool uio) |
19f73787 | 963 | { |
1763da12 | 964 | struct nfs_pageio_descriptor desc; |
1d59d61f | 965 | struct inode *inode = dreq->inode; |
19f73787 CL |
966 | ssize_t result = 0; |
967 | size_t requested_bytes = 0; | |
968 | unsigned long seg; | |
969 | ||
a20c93e3 | 970 | nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false, |
1763da12 FI |
971 | &nfs_direct_write_completion_ops); |
972 | desc.pg_dreq = dreq; | |
19f73787 | 973 | get_dreq(dreq); |
1d59d61f | 974 | atomic_inc(&inode->i_dio_count); |
19f73787 | 975 | |
7acdb026 | 976 | NFS_I(dreq->inode)->write_io += iov_length(iov, nr_segs); |
19f73787 CL |
977 | for (seg = 0; seg < nr_segs; seg++) { |
978 | const struct iovec *vec = &iov[seg]; | |
a564b8f0 | 979 | result = nfs_direct_write_schedule_segment(&desc, vec, pos, uio); |
19f73787 CL |
980 | if (result < 0) |
981 | break; | |
982 | requested_bytes += result; | |
983 | if ((size_t)result < vec->iov_len) | |
984 | break; | |
985 | pos += vec->iov_len; | |
986 | } | |
1763da12 | 987 | nfs_pageio_complete(&desc); |
19f73787 | 988 | |
839f7ad6 CL |
989 | /* |
990 | * If no bytes were started, return the error, and let the | |
991 | * generic layer handle the completion. | |
992 | */ | |
993 | if (requested_bytes == 0) { | |
1d59d61f | 994 | inode_dio_done(inode); |
839f7ad6 CL |
995 | nfs_direct_req_release(dreq); |
996 | return result < 0 ? result : -EIO; | |
997 | } | |
998 | ||
19f73787 CL |
999 | if (put_dreq(dreq)) |
1000 | nfs_direct_write_complete(dreq, dreq->inode); | |
839f7ad6 | 1001 | return 0; |
19f73787 CL |
1002 | } |
1003 | ||
1da177e4 LT |
1004 | /** |
1005 | * nfs_file_direct_write - file direct write operation for NFS files | |
1006 | * @iocb: target I/O control block | |
027445c3 BP |
1007 | * @iov: vector of user buffers from which to write data |
1008 | * @nr_segs: size of iov vector | |
88467055 | 1009 | * @pos: byte offset in file where writing starts |
1da177e4 LT |
1010 | * |
1011 | * We use this function for direct writes instead of calling | |
1012 | * generic_file_aio_write() in order to avoid taking the inode | |
1013 | * semaphore and updating the i_size. The NFS server will set | |
1014 | * the new i_size and this client must read the updated size | |
1015 | * back into its cache. We let the server do generic write | |
1016 | * parameter checking and report problems. | |
1017 | * | |
1da177e4 LT |
1018 | * We eliminate local atime updates, see direct read above. |
1019 | * | |
1020 | * We avoid unnecessary page cache invalidations for normal cached | |
1021 | * readers of this file. | |
1022 | * | |
1023 | * Note that O_APPEND is not supported for NFS direct writes, as there | |
1024 | * is no atomic O_APPEND write facility in the NFS protocol. | |
1025 | */ | |
027445c3 | 1026 | ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov, |
a564b8f0 | 1027 | unsigned long nr_segs, loff_t pos, bool uio) |
1da177e4 | 1028 | { |
22cd1bf1 | 1029 | ssize_t result = -EINVAL; |
1da177e4 | 1030 | struct file *file = iocb->ki_filp; |
1da177e4 | 1031 | struct address_space *mapping = file->f_mapping; |
22cd1bf1 CH |
1032 | struct inode *inode = mapping->host; |
1033 | struct nfs_direct_req *dreq; | |
1034 | struct nfs_lock_context *l_ctx; | |
a9ab5e84 | 1035 | loff_t end; |
c216fd70 | 1036 | size_t count; |
1da177e4 | 1037 | |
c216fd70 | 1038 | count = iov_length(iov, nr_segs); |
a9ab5e84 CH |
1039 | end = (pos + count - 1) >> PAGE_CACHE_SHIFT; |
1040 | ||
c216fd70 CL |
1041 | nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count); |
1042 | ||
6de1472f AV |
1043 | dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n", |
1044 | file, count, (long long) pos); | |
027445c3 | 1045 | |
22cd1bf1 CH |
1046 | result = generic_write_checks(file, &pos, &count, 0); |
1047 | if (result) | |
1da177e4 | 1048 | goto out; |
ce1a8e67 | 1049 | |
22cd1bf1 | 1050 | result = -EINVAL; |
ce1a8e67 | 1051 | if ((ssize_t) count < 0) |
1da177e4 | 1052 | goto out; |
22cd1bf1 | 1053 | result = 0; |
1da177e4 LT |
1054 | if (!count) |
1055 | goto out; | |
ce1a8e67 | 1056 | |
a9ab5e84 CH |
1057 | mutex_lock(&inode->i_mutex); |
1058 | ||
22cd1bf1 CH |
1059 | result = nfs_sync_mapping(mapping); |
1060 | if (result) | |
a9ab5e84 CH |
1061 | goto out_unlock; |
1062 | ||
1063 | if (mapping->nrpages) { | |
1064 | result = invalidate_inode_pages2_range(mapping, | |
1065 | pos >> PAGE_CACHE_SHIFT, end); | |
1066 | if (result) | |
1067 | goto out_unlock; | |
1068 | } | |
1da177e4 | 1069 | |
7ec10f26 KK |
1070 | task_io_account_write(count); |
1071 | ||
22cd1bf1 CH |
1072 | result = -ENOMEM; |
1073 | dreq = nfs_direct_req_alloc(); | |
1074 | if (!dreq) | |
a9ab5e84 | 1075 | goto out_unlock; |
9eafa8cc | 1076 | |
22cd1bf1 CH |
1077 | dreq->inode = inode; |
1078 | dreq->bytes_left = count; | |
1079 | dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); | |
1080 | l_ctx = nfs_get_lock_context(dreq->ctx); | |
1081 | if (IS_ERR(l_ctx)) { | |
1082 | result = PTR_ERR(l_ctx); | |
1083 | goto out_release; | |
1084 | } | |
1085 | dreq->l_ctx = l_ctx; | |
1086 | if (!is_sync_kiocb(iocb)) | |
1087 | dreq->iocb = iocb; | |
1088 | ||
1089 | result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, uio); | |
a9ab5e84 CH |
1090 | |
1091 | if (mapping->nrpages) { | |
1092 | invalidate_inode_pages2_range(mapping, | |
1093 | pos >> PAGE_CACHE_SHIFT, end); | |
1094 | } | |
1095 | ||
1096 | mutex_unlock(&inode->i_mutex); | |
1097 | ||
22cd1bf1 CH |
1098 | if (!result) { |
1099 | result = nfs_direct_wait(dreq); | |
1100 | if (result > 0) { | |
1101 | struct inode *inode = mapping->host; | |
1102 | ||
1103 | iocb->ki_pos = pos + result; | |
1104 | spin_lock(&inode->i_lock); | |
1105 | if (i_size_read(inode) < iocb->ki_pos) | |
1106 | i_size_write(inode, iocb->ki_pos); | |
1107 | spin_unlock(&inode->i_lock); | |
1108 | } | |
1763da12 | 1109 | } |
a9ab5e84 CH |
1110 | nfs_direct_req_release(dreq); |
1111 | return result; | |
1112 | ||
22cd1bf1 CH |
1113 | out_release: |
1114 | nfs_direct_req_release(dreq); | |
a9ab5e84 CH |
1115 | out_unlock: |
1116 | mutex_unlock(&inode->i_mutex); | |
1da177e4 | 1117 | out: |
22cd1bf1 | 1118 | return result; |
1da177e4 LT |
1119 | } |
1120 | ||
88467055 CL |
1121 | /** |
1122 | * nfs_init_directcache - create a slab cache for nfs_direct_req structures | |
1123 | * | |
1124 | */ | |
f7b422b1 | 1125 | int __init nfs_init_directcache(void) |
1da177e4 LT |
1126 | { |
1127 | nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", | |
1128 | sizeof(struct nfs_direct_req), | |
fffb60f9 PJ |
1129 | 0, (SLAB_RECLAIM_ACCOUNT| |
1130 | SLAB_MEM_SPREAD), | |
20c2df83 | 1131 | NULL); |
1da177e4 LT |
1132 | if (nfs_direct_cachep == NULL) |
1133 | return -ENOMEM; | |
1134 | ||
1135 | return 0; | |
1136 | } | |
1137 | ||
88467055 | 1138 | /** |
f7b422b1 | 1139 | * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures |
88467055 CL |
1140 | * |
1141 | */ | |
266bee88 | 1142 | void nfs_destroy_directcache(void) |
1da177e4 | 1143 | { |
1a1d92c1 | 1144 | kmem_cache_destroy(nfs_direct_cachep); |
1da177e4 | 1145 | } |