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 #include <linux/module.h>
50
51 #include <linux/nfs_fs.h>
52 #include <linux/nfs_page.h>
53 #include <linux/sunrpc/clnt.h>
54
55 #include <asm/uaccess.h>
56 #include <linux/atomic.h>
57
58 #include "internal.h"
59 #include "iostat.h"
60 #include "pnfs.h"
61
62 #define NFSDBG_FACILITY NFSDBG_VFS
63
64 static struct kmem_cache *nfs_direct_cachep;
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 */
71
72 /* I/O parameters */
73 struct nfs_open_context *ctx; /* file open context info */
74 struct nfs_lock_context *l_ctx; /* Lock context info */
75 struct kiocb * iocb; /* controlling i/o request */
76 struct inode * inode; /* target file of i/o */
77
78 /* completion state */
79 atomic_t io_count; /* i/os we're waiting for */
80 spinlock_t lock; /* protect completion state */
81 ssize_t count, /* bytes actually processed */
82 bytes_left, /* bytes left to be sent */
83 error; /* any reported error */
84 struct completion completion; /* wait for i/o completion */
85
86 /* commit state */
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;
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 */
94 };
95
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;
98 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
99 static void nfs_direct_write_schedule_work(struct work_struct *work);
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
111 /**
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
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.
123 */
124 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
125 {
126 struct inode *inode = iocb->ki_filp->f_mapping->host;
127
128 /* we only support swap file calling nfs_direct_IO */
129 if (!IS_SWAPFILE(inode))
130 return 0;
131
132 #ifndef CONFIG_NFS_SWAP
133 dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
134 iocb->ki_filp->f_path.dentry->d_name.name,
135 (long long) pos, nr_segs);
136
137 return -EINVAL;
138 #else
139 VM_BUG_ON(iocb->ki_left != PAGE_SIZE);
140 VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
141
142 if (rw == READ || rw == KERNEL_READ)
143 return nfs_file_direct_read(iocb, iov, nr_segs, pos,
144 rw == READ ? true : false);
145 return nfs_file_direct_write(iocb, iov, nr_segs, pos,
146 rw == WRITE ? true : false);
147 #endif /* CONFIG_NFS_SWAP */
148 }
149
150 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
151 {
152 unsigned int i;
153 for (i = 0; i < npages; i++)
154 page_cache_release(pages[i]);
155 }
156
157 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
158 struct nfs_direct_req *dreq)
159 {
160 cinfo->lock = &dreq->lock;
161 cinfo->mds = &dreq->mds_cinfo;
162 cinfo->ds = &dreq->ds_cinfo;
163 cinfo->dreq = dreq;
164 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
165 }
166
167 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
168 {
169 struct nfs_direct_req *dreq;
170
171 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
172 if (!dreq)
173 return NULL;
174
175 kref_init(&dreq->kref);
176 kref_get(&dreq->kref);
177 init_completion(&dreq->completion);
178 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
179 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
180 spin_lock_init(&dreq->lock);
181
182 return dreq;
183 }
184
185 static void nfs_direct_req_free(struct kref *kref)
186 {
187 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
188
189 nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
190 if (dreq->l_ctx != NULL)
191 nfs_put_lock_context(dreq->l_ctx);
192 if (dreq->ctx != NULL)
193 put_nfs_open_context(dreq->ctx);
194 kmem_cache_free(nfs_direct_cachep, dreq);
195 }
196
197 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
198 {
199 kref_put(&dreq->kref, nfs_direct_req_free);
200 }
201
202 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
203 {
204 return dreq->bytes_left;
205 }
206 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
207
208 /*
209 * Collects and returns the final error value/byte-count.
210 */
211 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
212 {
213 ssize_t result = -EIOCBQUEUED;
214
215 /* Async requests don't wait here */
216 if (dreq->iocb)
217 goto out;
218
219 result = wait_for_completion_killable(&dreq->completion);
220
221 if (!result)
222 result = dreq->error;
223 if (!result)
224 result = dreq->count;
225
226 out:
227 return (ssize_t) result;
228 }
229
230 /*
231 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
232 * the iocb is still valid here if this is a synchronous request.
233 */
234 static void nfs_direct_complete(struct nfs_direct_req *dreq)
235 {
236 if (dreq->iocb) {
237 long res = (long) dreq->error;
238 if (!res)
239 res = (long) dreq->count;
240 aio_complete(dreq->iocb, res, 0);
241 }
242 complete_all(&dreq->completion);
243
244 nfs_direct_req_release(dreq);
245 }
246
247 static void nfs_direct_readpage_release(struct nfs_page *req)
248 {
249 dprintk("NFS: direct read done (%s/%lld %d@%lld)\n",
250 req->wb_context->dentry->d_inode->i_sb->s_id,
251 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
252 req->wb_bytes,
253 (long long)req_offset(req));
254 nfs_release_request(req);
255 }
256
257 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
258 {
259 unsigned long bytes = 0;
260 struct nfs_direct_req *dreq = hdr->dreq;
261
262 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
263 goto out_put;
264
265 spin_lock(&dreq->lock);
266 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
267 dreq->error = hdr->error;
268 else
269 dreq->count += hdr->good_bytes;
270 spin_unlock(&dreq->lock);
271
272 while (!list_empty(&hdr->pages)) {
273 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
274 struct page *page = req->wb_page;
275
276 if (!PageCompound(page) && bytes < hdr->good_bytes)
277 set_page_dirty(page);
278 bytes += req->wb_bytes;
279 nfs_list_remove_request(req);
280 nfs_direct_readpage_release(req);
281 }
282 out_put:
283 if (put_dreq(dreq))
284 nfs_direct_complete(dreq);
285 hdr->release(hdr);
286 }
287
288 static void nfs_read_sync_pgio_error(struct list_head *head)
289 {
290 struct nfs_page *req;
291
292 while (!list_empty(head)) {
293 req = nfs_list_entry(head->next);
294 nfs_list_remove_request(req);
295 nfs_release_request(req);
296 }
297 }
298
299 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
300 {
301 get_dreq(hdr->dreq);
302 }
303
304 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
305 .error_cleanup = nfs_read_sync_pgio_error,
306 .init_hdr = nfs_direct_pgio_init,
307 .completion = nfs_direct_read_completion,
308 };
309
310 /*
311 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
312 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
313 * bail and stop sending more reads. Read length accounting is
314 * handled automatically by nfs_direct_read_result(). Otherwise, if
315 * no requests have been sent, just return an error.
316 */
317 static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc,
318 const struct iovec *iov,
319 loff_t pos, bool uio)
320 {
321 struct nfs_direct_req *dreq = desc->pg_dreq;
322 struct nfs_open_context *ctx = dreq->ctx;
323 struct inode *inode = ctx->dentry->d_inode;
324 unsigned long user_addr = (unsigned long)iov->iov_base;
325 size_t count = iov->iov_len;
326 size_t rsize = NFS_SERVER(inode)->rsize;
327 unsigned int pgbase;
328 int result;
329 ssize_t started = 0;
330 struct page **pagevec = NULL;
331 unsigned int npages;
332
333 do {
334 size_t bytes;
335 int i;
336
337 pgbase = user_addr & ~PAGE_MASK;
338 bytes = min(max_t(size_t, rsize, PAGE_SIZE), count);
339
340 result = -ENOMEM;
341 npages = nfs_page_array_len(pgbase, bytes);
342 if (!pagevec)
343 pagevec = kmalloc(npages * sizeof(struct page *),
344 GFP_KERNEL);
345 if (!pagevec)
346 break;
347 if (uio) {
348 down_read(¤t->mm->mmap_sem);
349 result = get_user_pages(current, current->mm, user_addr,
350 npages, 1, 0, pagevec, NULL);
351 up_read(¤t->mm->mmap_sem);
352 if (result < 0)
353 break;
354 } else {
355 WARN_ON(npages != 1);
356 result = get_kernel_page(user_addr, 1, pagevec);
357 if (WARN_ON(result != 1))
358 break;
359 }
360
361 if ((unsigned)result < npages) {
362 bytes = result * PAGE_SIZE;
363 if (bytes <= pgbase) {
364 nfs_direct_release_pages(pagevec, result);
365 break;
366 }
367 bytes -= pgbase;
368 npages = result;
369 }
370
371 for (i = 0; i < npages; i++) {
372 struct nfs_page *req;
373 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
374 /* XXX do we need to do the eof zeroing found in async_filler? */
375 req = nfs_create_request(dreq->ctx, dreq->inode,
376 pagevec[i],
377 pgbase, req_len);
378 if (IS_ERR(req)) {
379 result = PTR_ERR(req);
380 break;
381 }
382 req->wb_index = pos >> PAGE_SHIFT;
383 req->wb_offset = pos & ~PAGE_MASK;
384 if (!nfs_pageio_add_request(desc, req)) {
385 result = desc->pg_error;
386 nfs_release_request(req);
387 break;
388 }
389 pgbase = 0;
390 bytes -= req_len;
391 started += req_len;
392 user_addr += req_len;
393 pos += req_len;
394 count -= req_len;
395 dreq->bytes_left -= req_len;
396 }
397 /* The nfs_page now hold references to these pages */
398 nfs_direct_release_pages(pagevec, npages);
399 } while (count != 0 && result >= 0);
400
401 kfree(pagevec);
402
403 if (started)
404 return started;
405 return result < 0 ? (ssize_t) result : -EFAULT;
406 }
407
408 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
409 const struct iovec *iov,
410 unsigned long nr_segs,
411 loff_t pos, bool uio)
412 {
413 struct nfs_pageio_descriptor desc;
414 ssize_t result = -EINVAL;
415 size_t requested_bytes = 0;
416 unsigned long seg;
417
418 NFS_PROTO(dreq->inode)->read_pageio_init(&desc, dreq->inode,
419 &nfs_direct_read_completion_ops);
420 get_dreq(dreq);
421 desc.pg_dreq = dreq;
422
423 for (seg = 0; seg < nr_segs; seg++) {
424 const struct iovec *vec = &iov[seg];
425 result = nfs_direct_read_schedule_segment(&desc, vec, pos, uio);
426 if (result < 0)
427 break;
428 requested_bytes += result;
429 if ((size_t)result < vec->iov_len)
430 break;
431 pos += vec->iov_len;
432 }
433
434 nfs_pageio_complete(&desc);
435
436 /*
437 * If no bytes were started, return the error, and let the
438 * generic layer handle the completion.
439 */
440 if (requested_bytes == 0) {
441 nfs_direct_req_release(dreq);
442 return result < 0 ? result : -EIO;
443 }
444
445 if (put_dreq(dreq))
446 nfs_direct_complete(dreq);
447 return 0;
448 }
449
450 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
451 unsigned long nr_segs, loff_t pos, bool uio)
452 {
453 ssize_t result = -ENOMEM;
454 struct inode *inode = iocb->ki_filp->f_mapping->host;
455 struct nfs_direct_req *dreq;
456 struct nfs_lock_context *l_ctx;
457
458 dreq = nfs_direct_req_alloc();
459 if (dreq == NULL)
460 goto out;
461
462 dreq->inode = inode;
463 dreq->bytes_left = iov_length(iov, nr_segs);
464 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
465 l_ctx = nfs_get_lock_context(dreq->ctx);
466 if (IS_ERR(l_ctx)) {
467 result = PTR_ERR(l_ctx);
468 goto out_release;
469 }
470 dreq->l_ctx = l_ctx;
471 if (!is_sync_kiocb(iocb))
472 dreq->iocb = iocb;
473
474 NFS_I(inode)->read_io += iov_length(iov, nr_segs);
475 result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos, uio);
476 if (!result)
477 result = nfs_direct_wait(dreq);
478 out_release:
479 nfs_direct_req_release(dreq);
480 out:
481 return result;
482 }
483
484 static void nfs_inode_dio_write_done(struct inode *inode)
485 {
486 nfs_zap_mapping(inode, inode->i_mapping);
487 inode_dio_done(inode);
488 }
489
490 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
491 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
492 {
493 struct nfs_pageio_descriptor desc;
494 struct nfs_page *req, *tmp;
495 LIST_HEAD(reqs);
496 struct nfs_commit_info cinfo;
497 LIST_HEAD(failed);
498
499 nfs_init_cinfo_from_dreq(&cinfo, dreq);
500 pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
501 spin_lock(cinfo.lock);
502 nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
503 spin_unlock(cinfo.lock);
504
505 dreq->count = 0;
506 get_dreq(dreq);
507
508 NFS_PROTO(dreq->inode)->write_pageio_init(&desc, dreq->inode, FLUSH_STABLE,
509 &nfs_direct_write_completion_ops);
510 desc.pg_dreq = dreq;
511
512 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
513 if (!nfs_pageio_add_request(&desc, req)) {
514 nfs_list_remove_request(req);
515 nfs_list_add_request(req, &failed);
516 spin_lock(cinfo.lock);
517 dreq->flags = 0;
518 dreq->error = -EIO;
519 spin_unlock(cinfo.lock);
520 }
521 nfs_release_request(req);
522 }
523 nfs_pageio_complete(&desc);
524
525 while (!list_empty(&failed)) {
526 req = nfs_list_entry(failed.next);
527 nfs_list_remove_request(req);
528 nfs_unlock_and_release_request(req);
529 }
530
531 if (put_dreq(dreq))
532 nfs_direct_write_complete(dreq, dreq->inode);
533 }
534
535 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
536 {
537 struct nfs_direct_req *dreq = data->dreq;
538 struct nfs_commit_info cinfo;
539 struct nfs_page *req;
540 int status = data->task.tk_status;
541
542 nfs_init_cinfo_from_dreq(&cinfo, dreq);
543 if (status < 0) {
544 dprintk("NFS: %5u commit failed with error %d.\n",
545 data->task.tk_pid, status);
546 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
547 } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
548 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
549 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
550 }
551
552 dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
553 while (!list_empty(&data->pages)) {
554 req = nfs_list_entry(data->pages.next);
555 nfs_list_remove_request(req);
556 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
557 /* Note the rewrite will go through mds */
558 nfs_mark_request_commit(req, NULL, &cinfo);
559 } else
560 nfs_release_request(req);
561 nfs_unlock_and_release_request(req);
562 }
563
564 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
565 nfs_direct_write_complete(dreq, data->inode);
566 }
567
568 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
569 {
570 /* There is no lock to clear */
571 }
572
573 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
574 .completion = nfs_direct_commit_complete,
575 .error_cleanup = nfs_direct_error_cleanup,
576 };
577
578 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
579 {
580 int res;
581 struct nfs_commit_info cinfo;
582 LIST_HEAD(mds_list);
583
584 nfs_init_cinfo_from_dreq(&cinfo, dreq);
585 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
586 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
587 if (res < 0) /* res == -ENOMEM */
588 nfs_direct_write_reschedule(dreq);
589 }
590
591 static void nfs_direct_write_schedule_work(struct work_struct *work)
592 {
593 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
594 int flags = dreq->flags;
595
596 dreq->flags = 0;
597 switch (flags) {
598 case NFS_ODIRECT_DO_COMMIT:
599 nfs_direct_commit_schedule(dreq);
600 break;
601 case NFS_ODIRECT_RESCHED_WRITES:
602 nfs_direct_write_reschedule(dreq);
603 break;
604 default:
605 nfs_inode_dio_write_done(dreq->inode);
606 nfs_direct_complete(dreq);
607 }
608 }
609
610 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
611 {
612 schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
613 }
614
615 #else
616 static void nfs_direct_write_schedule_work(struct work_struct *work)
617 {
618 }
619
620 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
621 {
622 nfs_inode_dio_write_done(inode);
623 nfs_direct_complete(dreq);
624 }
625 #endif
626
627 /*
628 * NB: Return the value of the first error return code. Subsequent
629 * errors after the first one are ignored.
630 */
631 /*
632 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
633 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
634 * bail and stop sending more writes. Write length accounting is
635 * handled automatically by nfs_direct_write_result(). Otherwise, if
636 * no requests have been sent, just return an error.
637 */
638 static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc,
639 const struct iovec *iov,
640 loff_t pos, bool uio)
641 {
642 struct nfs_direct_req *dreq = desc->pg_dreq;
643 struct nfs_open_context *ctx = dreq->ctx;
644 struct inode *inode = ctx->dentry->d_inode;
645 unsigned long user_addr = (unsigned long)iov->iov_base;
646 size_t count = iov->iov_len;
647 size_t wsize = NFS_SERVER(inode)->wsize;
648 unsigned int pgbase;
649 int result;
650 ssize_t started = 0;
651 struct page **pagevec = NULL;
652 unsigned int npages;
653
654 do {
655 size_t bytes;
656 int i;
657
658 pgbase = user_addr & ~PAGE_MASK;
659 bytes = min(max_t(size_t, wsize, PAGE_SIZE), count);
660
661 result = -ENOMEM;
662 npages = nfs_page_array_len(pgbase, bytes);
663 if (!pagevec)
664 pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
665 if (!pagevec)
666 break;
667
668 if (uio) {
669 down_read(¤t->mm->mmap_sem);
670 result = get_user_pages(current, current->mm, user_addr,
671 npages, 0, 0, pagevec, NULL);
672 up_read(¤t->mm->mmap_sem);
673 if (result < 0)
674 break;
675 } else {
676 WARN_ON(npages != 1);
677 result = get_kernel_page(user_addr, 0, pagevec);
678 if (WARN_ON(result != 1))
679 break;
680 }
681
682 if ((unsigned)result < npages) {
683 bytes = result * PAGE_SIZE;
684 if (bytes <= pgbase) {
685 nfs_direct_release_pages(pagevec, result);
686 break;
687 }
688 bytes -= pgbase;
689 npages = result;
690 }
691
692 for (i = 0; i < npages; i++) {
693 struct nfs_page *req;
694 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
695
696 req = nfs_create_request(dreq->ctx, dreq->inode,
697 pagevec[i],
698 pgbase, req_len);
699 if (IS_ERR(req)) {
700 result = PTR_ERR(req);
701 break;
702 }
703 nfs_lock_request(req);
704 req->wb_index = pos >> PAGE_SHIFT;
705 req->wb_offset = pos & ~PAGE_MASK;
706 if (!nfs_pageio_add_request(desc, req)) {
707 result = desc->pg_error;
708 nfs_unlock_and_release_request(req);
709 break;
710 }
711 pgbase = 0;
712 bytes -= req_len;
713 started += req_len;
714 user_addr += req_len;
715 pos += req_len;
716 count -= req_len;
717 dreq->bytes_left -= req_len;
718 }
719 /* The nfs_page now hold references to these pages */
720 nfs_direct_release_pages(pagevec, npages);
721 } while (count != 0 && result >= 0);
722
723 kfree(pagevec);
724
725 if (started)
726 return started;
727 return result < 0 ? (ssize_t) result : -EFAULT;
728 }
729
730 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
731 {
732 struct nfs_direct_req *dreq = hdr->dreq;
733 struct nfs_commit_info cinfo;
734 int bit = -1;
735 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
736
737 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
738 goto out_put;
739
740 nfs_init_cinfo_from_dreq(&cinfo, dreq);
741
742 spin_lock(&dreq->lock);
743
744 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
745 dreq->flags = 0;
746 dreq->error = hdr->error;
747 }
748 if (dreq->error != 0)
749 bit = NFS_IOHDR_ERROR;
750 else {
751 dreq->count += hdr->good_bytes;
752 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
753 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
754 bit = NFS_IOHDR_NEED_RESCHED;
755 } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
756 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
757 bit = NFS_IOHDR_NEED_RESCHED;
758 else if (dreq->flags == 0) {
759 memcpy(&dreq->verf, hdr->verf,
760 sizeof(dreq->verf));
761 bit = NFS_IOHDR_NEED_COMMIT;
762 dreq->flags = NFS_ODIRECT_DO_COMMIT;
763 } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
764 if (memcmp(&dreq->verf, hdr->verf, sizeof(dreq->verf))) {
765 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
766 bit = NFS_IOHDR_NEED_RESCHED;
767 } else
768 bit = NFS_IOHDR_NEED_COMMIT;
769 }
770 }
771 }
772 spin_unlock(&dreq->lock);
773
774 while (!list_empty(&hdr->pages)) {
775 req = nfs_list_entry(hdr->pages.next);
776 nfs_list_remove_request(req);
777 switch (bit) {
778 case NFS_IOHDR_NEED_RESCHED:
779 case NFS_IOHDR_NEED_COMMIT:
780 kref_get(&req->wb_kref);
781 nfs_mark_request_commit(req, hdr->lseg, &cinfo);
782 }
783 nfs_unlock_and_release_request(req);
784 }
785
786 out_put:
787 if (put_dreq(dreq))
788 nfs_direct_write_complete(dreq, hdr->inode);
789 hdr->release(hdr);
790 }
791
792 static void nfs_write_sync_pgio_error(struct list_head *head)
793 {
794 struct nfs_page *req;
795
796 while (!list_empty(head)) {
797 req = nfs_list_entry(head->next);
798 nfs_list_remove_request(req);
799 nfs_unlock_and_release_request(req);
800 }
801 }
802
803 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
804 .error_cleanup = nfs_write_sync_pgio_error,
805 .init_hdr = nfs_direct_pgio_init,
806 .completion = nfs_direct_write_completion,
807 };
808
809 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
810 const struct iovec *iov,
811 unsigned long nr_segs,
812 loff_t pos, bool uio)
813 {
814 struct nfs_pageio_descriptor desc;
815 struct inode *inode = dreq->inode;
816 ssize_t result = 0;
817 size_t requested_bytes = 0;
818 unsigned long seg;
819
820 NFS_PROTO(inode)->write_pageio_init(&desc, inode, FLUSH_COND_STABLE,
821 &nfs_direct_write_completion_ops);
822 desc.pg_dreq = dreq;
823 get_dreq(dreq);
824 atomic_inc(&inode->i_dio_count);
825
826 NFS_I(dreq->inode)->write_io += iov_length(iov, nr_segs);
827 for (seg = 0; seg < nr_segs; seg++) {
828 const struct iovec *vec = &iov[seg];
829 result = nfs_direct_write_schedule_segment(&desc, vec, pos, uio);
830 if (result < 0)
831 break;
832 requested_bytes += result;
833 if ((size_t)result < vec->iov_len)
834 break;
835 pos += vec->iov_len;
836 }
837 nfs_pageio_complete(&desc);
838
839 /*
840 * If no bytes were started, return the error, and let the
841 * generic layer handle the completion.
842 */
843 if (requested_bytes == 0) {
844 inode_dio_done(inode);
845 nfs_direct_req_release(dreq);
846 return result < 0 ? result : -EIO;
847 }
848
849 if (put_dreq(dreq))
850 nfs_direct_write_complete(dreq, dreq->inode);
851 return 0;
852 }
853
854 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
855 unsigned long nr_segs, loff_t pos,
856 size_t count, bool uio)
857 {
858 ssize_t result = -ENOMEM;
859 struct inode *inode = iocb->ki_filp->f_mapping->host;
860 struct nfs_direct_req *dreq;
861 struct nfs_lock_context *l_ctx;
862
863 dreq = nfs_direct_req_alloc();
864 if (!dreq)
865 goto out;
866
867 dreq->inode = inode;
868 dreq->bytes_left = count;
869 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
870 l_ctx = nfs_get_lock_context(dreq->ctx);
871 if (IS_ERR(l_ctx)) {
872 result = PTR_ERR(l_ctx);
873 goto out_release;
874 }
875 dreq->l_ctx = l_ctx;
876 if (!is_sync_kiocb(iocb))
877 dreq->iocb = iocb;
878
879 result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, uio);
880 if (!result)
881 result = nfs_direct_wait(dreq);
882 out_release:
883 nfs_direct_req_release(dreq);
884 out:
885 return result;
886 }
887
888 /**
889 * nfs_file_direct_read - file direct read operation for NFS files
890 * @iocb: target I/O control block
891 * @iov: vector of user buffers into which to read data
892 * @nr_segs: size of iov vector
893 * @pos: byte offset in file where reading starts
894 *
895 * We use this function for direct reads instead of calling
896 * generic_file_aio_read() in order to avoid gfar's check to see if
897 * the request starts before the end of the file. For that check
898 * to work, we must generate a GETATTR before each direct read, and
899 * even then there is a window between the GETATTR and the subsequent
900 * READ where the file size could change. Our preference is simply
901 * to do all reads the application wants, and the server will take
902 * care of managing the end of file boundary.
903 *
904 * This function also eliminates unnecessarily updating the file's
905 * atime locally, as the NFS server sets the file's atime, and this
906 * client must read the updated atime from the server back into its
907 * cache.
908 */
909 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
910 unsigned long nr_segs, loff_t pos, bool uio)
911 {
912 ssize_t retval = -EINVAL;
913 struct file *file = iocb->ki_filp;
914 struct address_space *mapping = file->f_mapping;
915 size_t count;
916
917 count = iov_length(iov, nr_segs);
918 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
919
920 dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
921 file->f_path.dentry->d_parent->d_name.name,
922 file->f_path.dentry->d_name.name,
923 count, (long long) pos);
924
925 retval = 0;
926 if (!count)
927 goto out;
928
929 retval = nfs_sync_mapping(mapping);
930 if (retval)
931 goto out;
932
933 task_io_account_read(count);
934
935 retval = nfs_direct_read(iocb, iov, nr_segs, pos, uio);
936 if (retval > 0)
937 iocb->ki_pos = pos + retval;
938
939 out:
940 return retval;
941 }
942
943 /**
944 * nfs_file_direct_write - file direct write operation for NFS files
945 * @iocb: target I/O control block
946 * @iov: vector of user buffers from which to write data
947 * @nr_segs: size of iov vector
948 * @pos: byte offset in file where writing starts
949 *
950 * We use this function for direct writes instead of calling
951 * generic_file_aio_write() in order to avoid taking the inode
952 * semaphore and updating the i_size. The NFS server will set
953 * the new i_size and this client must read the updated size
954 * back into its cache. We let the server do generic write
955 * parameter checking and report problems.
956 *
957 * We eliminate local atime updates, see direct read above.
958 *
959 * We avoid unnecessary page cache invalidations for normal cached
960 * readers of this file.
961 *
962 * Note that O_APPEND is not supported for NFS direct writes, as there
963 * is no atomic O_APPEND write facility in the NFS protocol.
964 */
965 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
966 unsigned long nr_segs, loff_t pos, bool uio)
967 {
968 ssize_t retval = -EINVAL;
969 struct file *file = iocb->ki_filp;
970 struct address_space *mapping = file->f_mapping;
971 size_t count;
972
973 count = iov_length(iov, nr_segs);
974 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
975
976 dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
977 file->f_path.dentry->d_parent->d_name.name,
978 file->f_path.dentry->d_name.name,
979 count, (long long) pos);
980
981 retval = generic_write_checks(file, &pos, &count, 0);
982 if (retval)
983 goto out;
984
985 retval = -EINVAL;
986 if ((ssize_t) count < 0)
987 goto out;
988 retval = 0;
989 if (!count)
990 goto out;
991
992 retval = nfs_sync_mapping(mapping);
993 if (retval)
994 goto out;
995
996 task_io_account_write(count);
997
998 retval = nfs_direct_write(iocb, iov, nr_segs, pos, count, uio);
999 if (retval > 0) {
1000 struct inode *inode = mapping->host;
1001
1002 iocb->ki_pos = pos + retval;
1003 spin_lock(&inode->i_lock);
1004 if (i_size_read(inode) < iocb->ki_pos)
1005 i_size_write(inode, iocb->ki_pos);
1006 spin_unlock(&inode->i_lock);
1007 }
1008 out:
1009 return retval;
1010 }
1011
1012 /**
1013 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1014 *
1015 */
1016 int __init nfs_init_directcache(void)
1017 {
1018 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1019 sizeof(struct nfs_direct_req),
1020 0, (SLAB_RECLAIM_ACCOUNT|
1021 SLAB_MEM_SPREAD),
1022 NULL);
1023 if (nfs_direct_cachep == NULL)
1024 return -ENOMEM;
1025
1026 return 0;
1027 }
1028
1029 /**
1030 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1031 *
1032 */
1033 void nfs_destroy_directcache(void)
1034 {
1035 kmem_cache_destroy(nfs_direct_cachep);
1036 }
1037
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.