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