Loading...
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 <linux/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
64static struct kmem_cache *nfs_direct_cachep;
65
66/*
67 * This represents a set of asynchronous requests that we're waiting on
68 */
69struct nfs_direct_mirror {
70 ssize_t count;
71};
72
73struct nfs_direct_req {
74 struct kref kref; /* release manager */
75
76 /* I/O parameters */
77 struct nfs_open_context *ctx; /* file open context info */
78 struct nfs_lock_context *l_ctx; /* Lock context info */
79 struct kiocb * iocb; /* controlling i/o request */
80 struct inode * inode; /* target file of i/o */
81
82 /* completion state */
83 atomic_t io_count; /* i/os we're waiting for */
84 spinlock_t lock; /* protect completion state */
85
86 struct nfs_direct_mirror mirrors[NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX];
87 int mirror_count;
88
89 loff_t io_start; /* Start offset for I/O */
90 ssize_t count, /* bytes actually processed */
91 max_count, /* max expected count */
92 bytes_left, /* bytes left to be sent */
93 error; /* any reported error */
94 struct completion completion; /* wait for i/o completion */
95
96 /* commit state */
97 struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo */
98 struct pnfs_ds_commit_info ds_cinfo; /* Storage for cinfo */
99 struct work_struct work;
100 int flags;
101#define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
102#define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
103 struct nfs_writeverf verf; /* unstable write verifier */
104};
105
106static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
107static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
108static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
109static void nfs_direct_write_schedule_work(struct work_struct *work);
110
111static inline void get_dreq(struct nfs_direct_req *dreq)
112{
113 atomic_inc(&dreq->io_count);
114}
115
116static inline int put_dreq(struct nfs_direct_req *dreq)
117{
118 return atomic_dec_and_test(&dreq->io_count);
119}
120
121static void
122nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
123{
124 int i;
125 ssize_t count;
126
127 WARN_ON_ONCE(dreq->count >= dreq->max_count);
128
129 if (dreq->mirror_count == 1) {
130 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
131 dreq->count += hdr->good_bytes;
132 } else {
133 /* mirrored writes */
134 count = dreq->mirrors[hdr->pgio_mirror_idx].count;
135 if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
136 count = hdr->io_start + hdr->good_bytes - dreq->io_start;
137 dreq->mirrors[hdr->pgio_mirror_idx].count = count;
138 }
139 /* update the dreq->count by finding the minimum agreed count from all
140 * mirrors */
141 count = dreq->mirrors[0].count;
142
143 for (i = 1; i < dreq->mirror_count; i++)
144 count = min(count, dreq->mirrors[i].count);
145
146 dreq->count = count;
147 }
148}
149
150/*
151 * nfs_direct_select_verf - select the right verifier
152 * @dreq - direct request possibly spanning multiple servers
153 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
154 * @commit_idx - commit bucket index for the DS
155 *
156 * returns the correct verifier to use given the role of the server
157 */
158static struct nfs_writeverf *
159nfs_direct_select_verf(struct nfs_direct_req *dreq,
160 struct nfs_client *ds_clp,
161 int commit_idx)
162{
163 struct nfs_writeverf *verfp = &dreq->verf;
164
165#ifdef CONFIG_NFS_V4_1
166 /*
167 * pNFS is in use, use the DS verf except commit_through_mds is set
168 * for layout segment where nbuckets is zero.
169 */
170 if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
171 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
172 verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
173 else
174 WARN_ON_ONCE(1);
175 }
176#endif
177 return verfp;
178}
179
180
181/*
182 * nfs_direct_set_hdr_verf - set the write/commit verifier
183 * @dreq - direct request possibly spanning multiple servers
184 * @hdr - pageio header to validate against previously seen verfs
185 *
186 * Set the server's (MDS or DS) "seen" verifier
187 */
188static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
189 struct nfs_pgio_header *hdr)
190{
191 struct nfs_writeverf *verfp;
192
193 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
194 WARN_ON_ONCE(verfp->committed >= 0);
195 memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
196 WARN_ON_ONCE(verfp->committed < 0);
197}
198
199static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
200 const struct nfs_writeverf *v2)
201{
202 return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
203}
204
205/*
206 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
207 * @dreq - direct request possibly spanning multiple servers
208 * @hdr - pageio header to validate against previously seen verf
209 *
210 * set the server's "seen" verf if not initialized.
211 * returns result of comparison between @hdr->verf and the "seen"
212 * verf of the server used by @hdr (DS or MDS)
213 */
214static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
215 struct nfs_pgio_header *hdr)
216{
217 struct nfs_writeverf *verfp;
218
219 verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
220 if (verfp->committed < 0) {
221 nfs_direct_set_hdr_verf(dreq, hdr);
222 return 0;
223 }
224 return nfs_direct_cmp_verf(verfp, &hdr->verf);
225}
226
227/*
228 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
229 * @dreq - direct request possibly spanning multiple servers
230 * @data - commit data to validate against previously seen verf
231 *
232 * returns result of comparison between @data->verf and the verf of
233 * the server used by @data (DS or MDS)
234 */
235static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
236 struct nfs_commit_data *data)
237{
238 struct nfs_writeverf *verfp;
239
240 verfp = nfs_direct_select_verf(dreq, data->ds_clp,
241 data->ds_commit_index);
242
243 /* verifier not set so always fail */
244 if (verfp->committed < 0)
245 return 1;
246
247 return nfs_direct_cmp_verf(verfp, &data->verf);
248}
249
250/**
251 * nfs_direct_IO - NFS address space operation for direct I/O
252 * @iocb: target I/O control block
253 * @iter: I/O buffer
254 *
255 * The presence of this routine in the address space ops vector means
256 * the NFS client supports direct I/O. However, for most direct IO, we
257 * shunt off direct read and write requests before the VFS gets them,
258 * so this method is only ever called for swap.
259 */
260ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
261{
262 struct inode *inode = iocb->ki_filp->f_mapping->host;
263
264 /* we only support swap file calling nfs_direct_IO */
265 if (!IS_SWAPFILE(inode))
266 return 0;
267
268 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
269
270 if (iov_iter_rw(iter) == READ)
271 return nfs_file_direct_read(iocb, iter);
272 return nfs_file_direct_write(iocb, iter);
273}
274
275static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
276{
277 unsigned int i;
278 for (i = 0; i < npages; i++)
279 put_page(pages[i]);
280}
281
282void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
283 struct nfs_direct_req *dreq)
284{
285 cinfo->inode = dreq->inode;
286 cinfo->mds = &dreq->mds_cinfo;
287 cinfo->ds = &dreq->ds_cinfo;
288 cinfo->dreq = dreq;
289 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
290}
291
292static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
293 struct nfs_pageio_descriptor *pgio,
294 struct nfs_page *req)
295{
296 int mirror_count = 1;
297
298 if (pgio->pg_ops->pg_get_mirror_count)
299 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
300
301 dreq->mirror_count = mirror_count;
302}
303
304static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
305{
306 struct nfs_direct_req *dreq;
307
308 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
309 if (!dreq)
310 return NULL;
311
312 kref_init(&dreq->kref);
313 kref_get(&dreq->kref);
314 init_completion(&dreq->completion);
315 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
316 dreq->verf.committed = NFS_INVALID_STABLE_HOW; /* not set yet */
317 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
318 dreq->mirror_count = 1;
319 spin_lock_init(&dreq->lock);
320
321 return dreq;
322}
323
324static void nfs_direct_req_free(struct kref *kref)
325{
326 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
327
328 nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
329 if (dreq->l_ctx != NULL)
330 nfs_put_lock_context(dreq->l_ctx);
331 if (dreq->ctx != NULL)
332 put_nfs_open_context(dreq->ctx);
333 kmem_cache_free(nfs_direct_cachep, dreq);
334}
335
336static void nfs_direct_req_release(struct nfs_direct_req *dreq)
337{
338 kref_put(&dreq->kref, nfs_direct_req_free);
339}
340
341ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
342{
343 return dreq->bytes_left;
344}
345EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
346
347/*
348 * Collects and returns the final error value/byte-count.
349 */
350static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
351{
352 ssize_t result = -EIOCBQUEUED;
353
354 /* Async requests don't wait here */
355 if (dreq->iocb)
356 goto out;
357
358 result = wait_for_completion_killable(&dreq->completion);
359
360 if (!result) {
361 result = dreq->count;
362 WARN_ON_ONCE(dreq->count < 0);
363 }
364 if (!result)
365 result = dreq->error;
366
367out:
368 return (ssize_t) result;
369}
370
371/*
372 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
373 * the iocb is still valid here if this is a synchronous request.
374 */
375static void nfs_direct_complete(struct nfs_direct_req *dreq)
376{
377 struct inode *inode = dreq->inode;
378
379 inode_dio_end(inode);
380
381 if (dreq->iocb) {
382 long res = (long) dreq->error;
383 if (dreq->count != 0) {
384 res = (long) dreq->count;
385 WARN_ON_ONCE(dreq->count < 0);
386 }
387 dreq->iocb->ki_complete(dreq->iocb, res, 0);
388 }
389
390 complete(&dreq->completion);
391
392 nfs_direct_req_release(dreq);
393}
394
395static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
396{
397 unsigned long bytes = 0;
398 struct nfs_direct_req *dreq = hdr->dreq;
399
400 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
401 goto out_put;
402
403 spin_lock(&dreq->lock);
404 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
405 dreq->error = hdr->error;
406 else
407 nfs_direct_good_bytes(dreq, hdr);
408
409 spin_unlock(&dreq->lock);
410
411 while (!list_empty(&hdr->pages)) {
412 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
413 struct page *page = req->wb_page;
414
415 if (!PageCompound(page) && bytes < hdr->good_bytes)
416 set_page_dirty(page);
417 bytes += req->wb_bytes;
418 nfs_list_remove_request(req);
419 nfs_release_request(req);
420 }
421out_put:
422 if (put_dreq(dreq))
423 nfs_direct_complete(dreq);
424 hdr->release(hdr);
425}
426
427static void nfs_read_sync_pgio_error(struct list_head *head)
428{
429 struct nfs_page *req;
430
431 while (!list_empty(head)) {
432 req = nfs_list_entry(head->next);
433 nfs_list_remove_request(req);
434 nfs_release_request(req);
435 }
436}
437
438static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
439{
440 get_dreq(hdr->dreq);
441}
442
443static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
444 .error_cleanup = nfs_read_sync_pgio_error,
445 .init_hdr = nfs_direct_pgio_init,
446 .completion = nfs_direct_read_completion,
447};
448
449/*
450 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
451 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
452 * bail and stop sending more reads. Read length accounting is
453 * handled automatically by nfs_direct_read_result(). Otherwise, if
454 * no requests have been sent, just return an error.
455 */
456
457static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
458 struct iov_iter *iter,
459 loff_t pos)
460{
461 struct nfs_pageio_descriptor desc;
462 struct inode *inode = dreq->inode;
463 ssize_t result = -EINVAL;
464 size_t requested_bytes = 0;
465 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
466
467 nfs_pageio_init_read(&desc, dreq->inode, false,
468 &nfs_direct_read_completion_ops);
469 get_dreq(dreq);
470 desc.pg_dreq = dreq;
471 inode_dio_begin(inode);
472
473 while (iov_iter_count(iter)) {
474 struct page **pagevec;
475 size_t bytes;
476 size_t pgbase;
477 unsigned npages, i;
478
479 result = iov_iter_get_pages_alloc(iter, &pagevec,
480 rsize, &pgbase);
481 if (result < 0)
482 break;
483
484 bytes = result;
485 iov_iter_advance(iter, bytes);
486 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
487 for (i = 0; i < npages; i++) {
488 struct nfs_page *req;
489 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
490 /* XXX do we need to do the eof zeroing found in async_filler? */
491 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
492 pgbase, req_len);
493 if (IS_ERR(req)) {
494 result = PTR_ERR(req);
495 break;
496 }
497 req->wb_index = pos >> PAGE_SHIFT;
498 req->wb_offset = pos & ~PAGE_MASK;
499 if (!nfs_pageio_add_request(&desc, req)) {
500 result = desc.pg_error;
501 nfs_release_request(req);
502 break;
503 }
504 pgbase = 0;
505 bytes -= req_len;
506 requested_bytes += req_len;
507 pos += req_len;
508 dreq->bytes_left -= req_len;
509 }
510 nfs_direct_release_pages(pagevec, npages);
511 kvfree(pagevec);
512 if (result < 0)
513 break;
514 }
515
516 nfs_pageio_complete(&desc);
517
518 /*
519 * If no bytes were started, return the error, and let the
520 * generic layer handle the completion.
521 */
522 if (requested_bytes == 0) {
523 inode_dio_end(inode);
524 nfs_direct_req_release(dreq);
525 return result < 0 ? result : -EIO;
526 }
527
528 if (put_dreq(dreq))
529 nfs_direct_complete(dreq);
530 return requested_bytes;
531}
532
533/**
534 * nfs_file_direct_read - file direct read operation for NFS files
535 * @iocb: target I/O control block
536 * @iter: vector of user buffers into which to read data
537 *
538 * We use this function for direct reads instead of calling
539 * generic_file_aio_read() in order to avoid gfar's check to see if
540 * the request starts before the end of the file. For that check
541 * to work, we must generate a GETATTR before each direct read, and
542 * even then there is a window between the GETATTR and the subsequent
543 * READ where the file size could change. Our preference is simply
544 * to do all reads the application wants, and the server will take
545 * care of managing the end of file boundary.
546 *
547 * This function also eliminates unnecessarily updating the file's
548 * atime locally, as the NFS server sets the file's atime, and this
549 * client must read the updated atime from the server back into its
550 * cache.
551 */
552ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
553{
554 struct file *file = iocb->ki_filp;
555 struct address_space *mapping = file->f_mapping;
556 struct inode *inode = mapping->host;
557 struct nfs_direct_req *dreq;
558 struct nfs_lock_context *l_ctx;
559 ssize_t result = -EINVAL, requested;
560 size_t count = iov_iter_count(iter);
561 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
562
563 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
564 file, count, (long long) iocb->ki_pos);
565
566 result = 0;
567 if (!count)
568 goto out;
569
570 task_io_account_read(count);
571
572 result = -ENOMEM;
573 dreq = nfs_direct_req_alloc();
574 if (dreq == NULL)
575 goto out;
576
577 dreq->inode = inode;
578 dreq->bytes_left = dreq->max_count = count;
579 dreq->io_start = iocb->ki_pos;
580 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
581 l_ctx = nfs_get_lock_context(dreq->ctx);
582 if (IS_ERR(l_ctx)) {
583 result = PTR_ERR(l_ctx);
584 goto out_release;
585 }
586 dreq->l_ctx = l_ctx;
587 if (!is_sync_kiocb(iocb))
588 dreq->iocb = iocb;
589
590 nfs_start_io_direct(inode);
591
592 NFS_I(inode)->read_io += count;
593 requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
594
595 nfs_end_io_direct(inode);
596
597 if (requested > 0) {
598 result = nfs_direct_wait(dreq);
599 if (result > 0) {
600 requested -= result;
601 iocb->ki_pos += result;
602 }
603 iov_iter_revert(iter, requested);
604 } else {
605 result = requested;
606 }
607
608out_release:
609 nfs_direct_req_release(dreq);
610out:
611 return result;
612}
613
614static void
615nfs_direct_write_scan_commit_list(struct inode *inode,
616 struct list_head *list,
617 struct nfs_commit_info *cinfo)
618{
619 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
620#ifdef CONFIG_NFS_V4_1
621 if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
622 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
623#endif
624 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
625 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
626}
627
628static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
629{
630 struct nfs_pageio_descriptor desc;
631 struct nfs_page *req, *tmp;
632 LIST_HEAD(reqs);
633 struct nfs_commit_info cinfo;
634 LIST_HEAD(failed);
635 int i;
636
637 nfs_init_cinfo_from_dreq(&cinfo, dreq);
638 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
639
640 dreq->count = 0;
641 dreq->verf.committed = NFS_INVALID_STABLE_HOW;
642 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
643 for (i = 0; i < dreq->mirror_count; i++)
644 dreq->mirrors[i].count = 0;
645 get_dreq(dreq);
646
647 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
648 &nfs_direct_write_completion_ops);
649 desc.pg_dreq = dreq;
650
651 req = nfs_list_entry(reqs.next);
652 nfs_direct_setup_mirroring(dreq, &desc, req);
653 if (desc.pg_error < 0) {
654 list_splice_init(&reqs, &failed);
655 goto out_failed;
656 }
657
658 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
659 if (!nfs_pageio_add_request(&desc, req)) {
660 nfs_list_remove_request(req);
661 nfs_list_add_request(req, &failed);
662 spin_lock(&cinfo.inode->i_lock);
663 dreq->flags = 0;
664 if (desc.pg_error < 0)
665 dreq->error = desc.pg_error;
666 else
667 dreq->error = -EIO;
668 spin_unlock(&cinfo.inode->i_lock);
669 }
670 nfs_release_request(req);
671 }
672 nfs_pageio_complete(&desc);
673
674out_failed:
675 while (!list_empty(&failed)) {
676 req = nfs_list_entry(failed.next);
677 nfs_list_remove_request(req);
678 nfs_unlock_and_release_request(req);
679 }
680
681 if (put_dreq(dreq))
682 nfs_direct_write_complete(dreq);
683}
684
685static void nfs_direct_commit_complete(struct nfs_commit_data *data)
686{
687 struct nfs_direct_req *dreq = data->dreq;
688 struct nfs_commit_info cinfo;
689 struct nfs_page *req;
690 int status = data->task.tk_status;
691
692 nfs_init_cinfo_from_dreq(&cinfo, dreq);
693 if (status < 0 || nfs_direct_cmp_commit_data_verf(dreq, data))
694 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
695
696 while (!list_empty(&data->pages)) {
697 req = nfs_list_entry(data->pages.next);
698 nfs_list_remove_request(req);
699 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
700 /* Note the rewrite will go through mds */
701 nfs_mark_request_commit(req, NULL, &cinfo, 0);
702 } else
703 nfs_release_request(req);
704 nfs_unlock_and_release_request(req);
705 }
706
707 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
708 nfs_direct_write_complete(dreq);
709}
710
711static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
712 struct nfs_page *req)
713{
714 struct nfs_direct_req *dreq = cinfo->dreq;
715
716 spin_lock(&dreq->lock);
717 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
718 spin_unlock(&dreq->lock);
719 nfs_mark_request_commit(req, NULL, cinfo, 0);
720}
721
722static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
723 .completion = nfs_direct_commit_complete,
724 .resched_write = nfs_direct_resched_write,
725};
726
727static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
728{
729 int res;
730 struct nfs_commit_info cinfo;
731 LIST_HEAD(mds_list);
732
733 nfs_init_cinfo_from_dreq(&cinfo, dreq);
734 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
735 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
736 if (res < 0) /* res == -ENOMEM */
737 nfs_direct_write_reschedule(dreq);
738}
739
740static void nfs_direct_write_schedule_work(struct work_struct *work)
741{
742 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
743 int flags = dreq->flags;
744
745 dreq->flags = 0;
746 switch (flags) {
747 case NFS_ODIRECT_DO_COMMIT:
748 nfs_direct_commit_schedule(dreq);
749 break;
750 case NFS_ODIRECT_RESCHED_WRITES:
751 nfs_direct_write_reschedule(dreq);
752 break;
753 default:
754 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
755 nfs_direct_complete(dreq);
756 }
757}
758
759static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
760{
761 schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
762}
763
764static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
765{
766 struct nfs_direct_req *dreq = hdr->dreq;
767 struct nfs_commit_info cinfo;
768 bool request_commit = false;
769 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
770
771 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
772 goto out_put;
773
774 nfs_init_cinfo_from_dreq(&cinfo, dreq);
775
776 spin_lock(&dreq->lock);
777
778 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
779 dreq->error = hdr->error;
780 if (dreq->error == 0) {
781 nfs_direct_good_bytes(dreq, hdr);
782 if (nfs_write_need_commit(hdr)) {
783 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
784 request_commit = true;
785 else if (dreq->flags == 0) {
786 nfs_direct_set_hdr_verf(dreq, hdr);
787 request_commit = true;
788 dreq->flags = NFS_ODIRECT_DO_COMMIT;
789 } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
790 request_commit = true;
791 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
792 dreq->flags =
793 NFS_ODIRECT_RESCHED_WRITES;
794 }
795 }
796 }
797 spin_unlock(&dreq->lock);
798
799 while (!list_empty(&hdr->pages)) {
800
801 req = nfs_list_entry(hdr->pages.next);
802 nfs_list_remove_request(req);
803 if (request_commit) {
804 kref_get(&req->wb_kref);
805 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
806 hdr->ds_commit_idx);
807 }
808 nfs_unlock_and_release_request(req);
809 }
810
811out_put:
812 if (put_dreq(dreq))
813 nfs_direct_write_complete(dreq);
814 hdr->release(hdr);
815}
816
817static void nfs_write_sync_pgio_error(struct list_head *head)
818{
819 struct nfs_page *req;
820
821 while (!list_empty(head)) {
822 req = nfs_list_entry(head->next);
823 nfs_list_remove_request(req);
824 nfs_unlock_and_release_request(req);
825 }
826}
827
828static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
829{
830 struct nfs_direct_req *dreq = hdr->dreq;
831
832 spin_lock(&dreq->lock);
833 if (dreq->error == 0) {
834 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
835 /* fake unstable write to let common nfs resend pages */
836 hdr->verf.committed = NFS_UNSTABLE;
837 hdr->good_bytes = hdr->args.count;
838 }
839 spin_unlock(&dreq->lock);
840}
841
842static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
843 .error_cleanup = nfs_write_sync_pgio_error,
844 .init_hdr = nfs_direct_pgio_init,
845 .completion = nfs_direct_write_completion,
846 .reschedule_io = nfs_direct_write_reschedule_io,
847};
848
849
850/*
851 * NB: Return the value of the first error return code. Subsequent
852 * errors after the first one are ignored.
853 */
854/*
855 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
856 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
857 * bail and stop sending more writes. Write length accounting is
858 * handled automatically by nfs_direct_write_result(). Otherwise, if
859 * no requests have been sent, just return an error.
860 */
861static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
862 struct iov_iter *iter,
863 loff_t pos)
864{
865 struct nfs_pageio_descriptor desc;
866 struct inode *inode = dreq->inode;
867 ssize_t result = 0;
868 size_t requested_bytes = 0;
869 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
870
871 nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
872 &nfs_direct_write_completion_ops);
873 desc.pg_dreq = dreq;
874 get_dreq(dreq);
875 inode_dio_begin(inode);
876
877 NFS_I(inode)->write_io += iov_iter_count(iter);
878 while (iov_iter_count(iter)) {
879 struct page **pagevec;
880 size_t bytes;
881 size_t pgbase;
882 unsigned npages, i;
883
884 result = iov_iter_get_pages_alloc(iter, &pagevec,
885 wsize, &pgbase);
886 if (result < 0)
887 break;
888
889 bytes = result;
890 iov_iter_advance(iter, bytes);
891 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
892 for (i = 0; i < npages; i++) {
893 struct nfs_page *req;
894 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
895
896 req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
897 pgbase, req_len);
898 if (IS_ERR(req)) {
899 result = PTR_ERR(req);
900 break;
901 }
902
903 nfs_direct_setup_mirroring(dreq, &desc, req);
904 if (desc.pg_error < 0) {
905 nfs_free_request(req);
906 result = desc.pg_error;
907 break;
908 }
909
910 nfs_lock_request(req);
911 req->wb_index = pos >> PAGE_SHIFT;
912 req->wb_offset = pos & ~PAGE_MASK;
913 if (!nfs_pageio_add_request(&desc, req)) {
914 result = desc.pg_error;
915 nfs_unlock_and_release_request(req);
916 break;
917 }
918 pgbase = 0;
919 bytes -= req_len;
920 requested_bytes += req_len;
921 pos += req_len;
922 dreq->bytes_left -= req_len;
923 }
924 nfs_direct_release_pages(pagevec, npages);
925 kvfree(pagevec);
926 if (result < 0)
927 break;
928 }
929 nfs_pageio_complete(&desc);
930
931 /*
932 * If no bytes were started, return the error, and let the
933 * generic layer handle the completion.
934 */
935 if (requested_bytes == 0) {
936 inode_dio_end(inode);
937 nfs_direct_req_release(dreq);
938 return result < 0 ? result : -EIO;
939 }
940
941 if (put_dreq(dreq))
942 nfs_direct_write_complete(dreq);
943 return requested_bytes;
944}
945
946/**
947 * nfs_file_direct_write - file direct write operation for NFS files
948 * @iocb: target I/O control block
949 * @iter: vector of user buffers from which to write data
950 *
951 * We use this function for direct writes instead of calling
952 * generic_file_aio_write() in order to avoid taking the inode
953 * semaphore and updating the i_size. The NFS server will set
954 * the new i_size and this client must read the updated size
955 * back into its cache. We let the server do generic write
956 * parameter checking and report problems.
957 *
958 * We eliminate local atime updates, see direct read above.
959 *
960 * We avoid unnecessary page cache invalidations for normal cached
961 * readers of this file.
962 *
963 * Note that O_APPEND is not supported for NFS direct writes, as there
964 * is no atomic O_APPEND write facility in the NFS protocol.
965 */
966ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
967{
968 ssize_t result = -EINVAL, requested;
969 size_t count;
970 struct file *file = iocb->ki_filp;
971 struct address_space *mapping = file->f_mapping;
972 struct inode *inode = mapping->host;
973 struct nfs_direct_req *dreq;
974 struct nfs_lock_context *l_ctx;
975 loff_t pos, end;
976
977 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
978 file, iov_iter_count(iter), (long long) iocb->ki_pos);
979
980 result = generic_write_checks(iocb, iter);
981 if (result <= 0)
982 return result;
983 count = result;
984 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
985
986 pos = iocb->ki_pos;
987 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
988
989 task_io_account_write(count);
990
991 result = -ENOMEM;
992 dreq = nfs_direct_req_alloc();
993 if (!dreq)
994 goto out;
995
996 dreq->inode = inode;
997 dreq->bytes_left = dreq->max_count = count;
998 dreq->io_start = pos;
999 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1000 l_ctx = nfs_get_lock_context(dreq->ctx);
1001 if (IS_ERR(l_ctx)) {
1002 result = PTR_ERR(l_ctx);
1003 goto out_release;
1004 }
1005 dreq->l_ctx = l_ctx;
1006 if (!is_sync_kiocb(iocb))
1007 dreq->iocb = iocb;
1008
1009 nfs_start_io_direct(inode);
1010
1011 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1012
1013 if (mapping->nrpages) {
1014 invalidate_inode_pages2_range(mapping,
1015 pos >> PAGE_SHIFT, end);
1016 }
1017
1018 nfs_end_io_direct(inode);
1019
1020 if (requested > 0) {
1021 result = nfs_direct_wait(dreq);
1022 if (result > 0) {
1023 requested -= result;
1024 iocb->ki_pos = pos + result;
1025 /* XXX: should check the generic_write_sync retval */
1026 generic_write_sync(iocb, result);
1027 }
1028 iov_iter_revert(iter, requested);
1029 } else {
1030 result = requested;
1031 }
1032out_release:
1033 nfs_direct_req_release(dreq);
1034out:
1035 return result;
1036}
1037
1038/**
1039 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1040 *
1041 */
1042int __init nfs_init_directcache(void)
1043{
1044 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1045 sizeof(struct nfs_direct_req),
1046 0, (SLAB_RECLAIM_ACCOUNT|
1047 SLAB_MEM_SPREAD),
1048 NULL);
1049 if (nfs_direct_cachep == NULL)
1050 return -ENOMEM;
1051
1052 return 0;
1053}
1054
1055/**
1056 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1057 *
1058 */
1059void nfs_destroy_directcache(void)
1060{
1061 kmem_cache_destroy(nfs_direct_cachep);
1062}
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
63static struct kmem_cache *nfs_direct_cachep;
64
65/*
66 * This represents a set of asynchronous requests that we're waiting on
67 */
68struct 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
94static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
95static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
96static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
97static void nfs_direct_write_schedule_work(struct work_struct *work);
98
99static inline void get_dreq(struct nfs_direct_req *dreq)
100{
101 atomic_inc(&dreq->io_count);
102}
103
104static 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, we shunt off direct
119 * read and write requests before the VFS gets them, so this method
120 * should never be called.
121 */
122ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
123{
124 dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
125 iocb->ki_filp->f_path.dentry->d_name.name,
126 (long long) pos, nr_segs);
127
128 return -EINVAL;
129}
130
131static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
132{
133 unsigned int i;
134 for (i = 0; i < npages; i++)
135 page_cache_release(pages[i]);
136}
137
138void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
139 struct nfs_direct_req *dreq)
140{
141 cinfo->lock = &dreq->lock;
142 cinfo->mds = &dreq->mds_cinfo;
143 cinfo->ds = &dreq->ds_cinfo;
144 cinfo->dreq = dreq;
145 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
146}
147
148static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
149{
150 struct nfs_direct_req *dreq;
151
152 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
153 if (!dreq)
154 return NULL;
155
156 kref_init(&dreq->kref);
157 kref_get(&dreq->kref);
158 init_completion(&dreq->completion);
159 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
160 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
161 spin_lock_init(&dreq->lock);
162
163 return dreq;
164}
165
166static void nfs_direct_req_free(struct kref *kref)
167{
168 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
169
170 if (dreq->l_ctx != NULL)
171 nfs_put_lock_context(dreq->l_ctx);
172 if (dreq->ctx != NULL)
173 put_nfs_open_context(dreq->ctx);
174 kmem_cache_free(nfs_direct_cachep, dreq);
175}
176
177static void nfs_direct_req_release(struct nfs_direct_req *dreq)
178{
179 kref_put(&dreq->kref, nfs_direct_req_free);
180}
181
182/*
183 * Collects and returns the final error value/byte-count.
184 */
185static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
186{
187 ssize_t result = -EIOCBQUEUED;
188
189 /* Async requests don't wait here */
190 if (dreq->iocb)
191 goto out;
192
193 result = wait_for_completion_killable(&dreq->completion);
194
195 if (!result)
196 result = dreq->error;
197 if (!result)
198 result = dreq->count;
199
200out:
201 return (ssize_t) result;
202}
203
204/*
205 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
206 * the iocb is still valid here if this is a synchronous request.
207 */
208static void nfs_direct_complete(struct nfs_direct_req *dreq)
209{
210 if (dreq->iocb) {
211 long res = (long) dreq->error;
212 if (!res)
213 res = (long) dreq->count;
214 aio_complete(dreq->iocb, res, 0);
215 }
216 complete_all(&dreq->completion);
217
218 nfs_direct_req_release(dreq);
219}
220
221static void nfs_direct_readpage_release(struct nfs_page *req)
222{
223 dprintk("NFS: direct read done (%s/%lld %d@%lld)\n",
224 req->wb_context->dentry->d_inode->i_sb->s_id,
225 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
226 req->wb_bytes,
227 (long long)req_offset(req));
228 nfs_release_request(req);
229}
230
231static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
232{
233 unsigned long bytes = 0;
234 struct nfs_direct_req *dreq = hdr->dreq;
235
236 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
237 goto out_put;
238
239 spin_lock(&dreq->lock);
240 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
241 dreq->error = hdr->error;
242 else
243 dreq->count += hdr->good_bytes;
244 spin_unlock(&dreq->lock);
245
246 while (!list_empty(&hdr->pages)) {
247 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
248 struct page *page = req->wb_page;
249
250 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
251 if (bytes > hdr->good_bytes)
252 zero_user(page, 0, PAGE_SIZE);
253 else if (hdr->good_bytes - bytes < PAGE_SIZE)
254 zero_user_segment(page,
255 hdr->good_bytes & ~PAGE_MASK,
256 PAGE_SIZE);
257 }
258 if (!PageCompound(page)) {
259 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
260 if (bytes < hdr->good_bytes)
261 set_page_dirty(page);
262 } else
263 set_page_dirty(page);
264 }
265 bytes += req->wb_bytes;
266 nfs_list_remove_request(req);
267 nfs_direct_readpage_release(req);
268 }
269out_put:
270 if (put_dreq(dreq))
271 nfs_direct_complete(dreq);
272 hdr->release(hdr);
273}
274
275static void nfs_read_sync_pgio_error(struct list_head *head)
276{
277 struct nfs_page *req;
278
279 while (!list_empty(head)) {
280 req = nfs_list_entry(head->next);
281 nfs_list_remove_request(req);
282 nfs_release_request(req);
283 }
284}
285
286static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
287{
288 get_dreq(hdr->dreq);
289}
290
291static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
292 .error_cleanup = nfs_read_sync_pgio_error,
293 .init_hdr = nfs_direct_pgio_init,
294 .completion = nfs_direct_read_completion,
295};
296
297/*
298 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
299 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
300 * bail and stop sending more reads. Read length accounting is
301 * handled automatically by nfs_direct_read_result(). Otherwise, if
302 * no requests have been sent, just return an error.
303 */
304static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc,
305 const struct iovec *iov,
306 loff_t pos)
307{
308 struct nfs_direct_req *dreq = desc->pg_dreq;
309 struct nfs_open_context *ctx = dreq->ctx;
310 struct inode *inode = ctx->dentry->d_inode;
311 unsigned long user_addr = (unsigned long)iov->iov_base;
312 size_t count = iov->iov_len;
313 size_t rsize = NFS_SERVER(inode)->rsize;
314 unsigned int pgbase;
315 int result;
316 ssize_t started = 0;
317 struct page **pagevec = NULL;
318 unsigned int npages;
319
320 do {
321 size_t bytes;
322 int i;
323
324 pgbase = user_addr & ~PAGE_MASK;
325 bytes = min(max_t(size_t, rsize, PAGE_SIZE), count);
326
327 result = -ENOMEM;
328 npages = nfs_page_array_len(pgbase, bytes);
329 if (!pagevec)
330 pagevec = kmalloc(npages * sizeof(struct page *),
331 GFP_KERNEL);
332 if (!pagevec)
333 break;
334 down_read(¤t->mm->mmap_sem);
335 result = get_user_pages(current, current->mm, user_addr,
336 npages, 1, 0, pagevec, NULL);
337 up_read(¤t->mm->mmap_sem);
338 if (result < 0)
339 break;
340 if ((unsigned)result < npages) {
341 bytes = result * PAGE_SIZE;
342 if (bytes <= pgbase) {
343 nfs_direct_release_pages(pagevec, result);
344 break;
345 }
346 bytes -= pgbase;
347 npages = result;
348 }
349
350 for (i = 0; i < npages; i++) {
351 struct nfs_page *req;
352 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
353 /* XXX do we need to do the eof zeroing found in async_filler? */
354 req = nfs_create_request(dreq->ctx, dreq->inode,
355 pagevec[i],
356 pgbase, req_len);
357 if (IS_ERR(req)) {
358 result = PTR_ERR(req);
359 break;
360 }
361 req->wb_index = pos >> PAGE_SHIFT;
362 req->wb_offset = pos & ~PAGE_MASK;
363 if (!nfs_pageio_add_request(desc, req)) {
364 result = desc->pg_error;
365 nfs_release_request(req);
366 break;
367 }
368 pgbase = 0;
369 bytes -= req_len;
370 started += req_len;
371 user_addr += req_len;
372 pos += req_len;
373 count -= req_len;
374 }
375 /* The nfs_page now hold references to these pages */
376 nfs_direct_release_pages(pagevec, npages);
377 } while (count != 0 && result >= 0);
378
379 kfree(pagevec);
380
381 if (started)
382 return started;
383 return result < 0 ? (ssize_t) result : -EFAULT;
384}
385
386static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
387 const struct iovec *iov,
388 unsigned long nr_segs,
389 loff_t pos)
390{
391 struct nfs_pageio_descriptor desc;
392 ssize_t result = -EINVAL;
393 size_t requested_bytes = 0;
394 unsigned long seg;
395
396 nfs_pageio_init_read(&desc, dreq->inode,
397 &nfs_direct_read_completion_ops);
398 get_dreq(dreq);
399 desc.pg_dreq = dreq;
400
401 for (seg = 0; seg < nr_segs; seg++) {
402 const struct iovec *vec = &iov[seg];
403 result = nfs_direct_read_schedule_segment(&desc, vec, pos);
404 if (result < 0)
405 break;
406 requested_bytes += result;
407 if ((size_t)result < vec->iov_len)
408 break;
409 pos += vec->iov_len;
410 }
411
412 nfs_pageio_complete(&desc);
413
414 /*
415 * If no bytes were started, return the error, and let the
416 * generic layer handle the completion.
417 */
418 if (requested_bytes == 0) {
419 nfs_direct_req_release(dreq);
420 return result < 0 ? result : -EIO;
421 }
422
423 if (put_dreq(dreq))
424 nfs_direct_complete(dreq);
425 return 0;
426}
427
428static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
429 unsigned long nr_segs, loff_t pos)
430{
431 ssize_t result = -ENOMEM;
432 struct inode *inode = iocb->ki_filp->f_mapping->host;
433 struct nfs_direct_req *dreq;
434
435 dreq = nfs_direct_req_alloc();
436 if (dreq == NULL)
437 goto out;
438
439 dreq->inode = inode;
440 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
441 dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
442 if (dreq->l_ctx == NULL)
443 goto out_release;
444 if (!is_sync_kiocb(iocb))
445 dreq->iocb = iocb;
446
447 result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
448 if (!result)
449 result = nfs_direct_wait(dreq);
450 NFS_I(inode)->read_io += result;
451out_release:
452 nfs_direct_req_release(dreq);
453out:
454 return result;
455}
456
457static void nfs_inode_dio_write_done(struct inode *inode)
458{
459 nfs_zap_mapping(inode, inode->i_mapping);
460 inode_dio_done(inode);
461}
462
463#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
464static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
465{
466 struct nfs_pageio_descriptor desc;
467 struct nfs_page *req, *tmp;
468 LIST_HEAD(reqs);
469 struct nfs_commit_info cinfo;
470 LIST_HEAD(failed);
471
472 nfs_init_cinfo_from_dreq(&cinfo, dreq);
473 pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
474 spin_lock(cinfo.lock);
475 nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
476 spin_unlock(cinfo.lock);
477
478 dreq->count = 0;
479 get_dreq(dreq);
480
481 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE,
482 &nfs_direct_write_completion_ops);
483 desc.pg_dreq = dreq;
484
485 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
486 if (!nfs_pageio_add_request(&desc, req)) {
487 nfs_list_remove_request(req);
488 nfs_list_add_request(req, &failed);
489 spin_lock(cinfo.lock);
490 dreq->flags = 0;
491 dreq->error = -EIO;
492 spin_unlock(cinfo.lock);
493 }
494 nfs_release_request(req);
495 }
496 nfs_pageio_complete(&desc);
497
498 while (!list_empty(&failed)) {
499 req = nfs_list_entry(failed.next);
500 nfs_list_remove_request(req);
501 nfs_unlock_and_release_request(req);
502 }
503
504 if (put_dreq(dreq))
505 nfs_direct_write_complete(dreq, dreq->inode);
506}
507
508static void nfs_direct_commit_complete(struct nfs_commit_data *data)
509{
510 struct nfs_direct_req *dreq = data->dreq;
511 struct nfs_commit_info cinfo;
512 struct nfs_page *req;
513 int status = data->task.tk_status;
514
515 nfs_init_cinfo_from_dreq(&cinfo, dreq);
516 if (status < 0) {
517 dprintk("NFS: %5u commit failed with error %d.\n",
518 data->task.tk_pid, status);
519 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
520 } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
521 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
522 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
523 }
524
525 dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
526 while (!list_empty(&data->pages)) {
527 req = nfs_list_entry(data->pages.next);
528 nfs_list_remove_request(req);
529 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
530 /* Note the rewrite will go through mds */
531 nfs_mark_request_commit(req, NULL, &cinfo);
532 } else
533 nfs_release_request(req);
534 nfs_unlock_and_release_request(req);
535 }
536
537 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
538 nfs_direct_write_complete(dreq, data->inode);
539}
540
541static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
542{
543 /* There is no lock to clear */
544}
545
546static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
547 .completion = nfs_direct_commit_complete,
548 .error_cleanup = nfs_direct_error_cleanup,
549};
550
551static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
552{
553 int res;
554 struct nfs_commit_info cinfo;
555 LIST_HEAD(mds_list);
556
557 nfs_init_cinfo_from_dreq(&cinfo, dreq);
558 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
559 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
560 if (res < 0) /* res == -ENOMEM */
561 nfs_direct_write_reschedule(dreq);
562}
563
564static void nfs_direct_write_schedule_work(struct work_struct *work)
565{
566 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
567 int flags = dreq->flags;
568
569 dreq->flags = 0;
570 switch (flags) {
571 case NFS_ODIRECT_DO_COMMIT:
572 nfs_direct_commit_schedule(dreq);
573 break;
574 case NFS_ODIRECT_RESCHED_WRITES:
575 nfs_direct_write_reschedule(dreq);
576 break;
577 default:
578 nfs_inode_dio_write_done(dreq->inode);
579 nfs_direct_complete(dreq);
580 }
581}
582
583static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
584{
585 schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
586}
587
588#else
589static void nfs_direct_write_schedule_work(struct work_struct *work)
590{
591}
592
593static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
594{
595 nfs_inode_dio_write_done(inode);
596 nfs_direct_complete(dreq);
597}
598#endif
599
600/*
601 * NB: Return the value of the first error return code. Subsequent
602 * errors after the first one are ignored.
603 */
604/*
605 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
606 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
607 * bail and stop sending more writes. Write length accounting is
608 * handled automatically by nfs_direct_write_result(). Otherwise, if
609 * no requests have been sent, just return an error.
610 */
611static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc,
612 const struct iovec *iov,
613 loff_t pos)
614{
615 struct nfs_direct_req *dreq = desc->pg_dreq;
616 struct nfs_open_context *ctx = dreq->ctx;
617 struct inode *inode = ctx->dentry->d_inode;
618 unsigned long user_addr = (unsigned long)iov->iov_base;
619 size_t count = iov->iov_len;
620 size_t wsize = NFS_SERVER(inode)->wsize;
621 unsigned int pgbase;
622 int result;
623 ssize_t started = 0;
624 struct page **pagevec = NULL;
625 unsigned int npages;
626
627 do {
628 size_t bytes;
629 int i;
630
631 pgbase = user_addr & ~PAGE_MASK;
632 bytes = min(max_t(size_t, wsize, PAGE_SIZE), count);
633
634 result = -ENOMEM;
635 npages = nfs_page_array_len(pgbase, bytes);
636 if (!pagevec)
637 pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
638 if (!pagevec)
639 break;
640
641 down_read(¤t->mm->mmap_sem);
642 result = get_user_pages(current, current->mm, user_addr,
643 npages, 0, 0, pagevec, NULL);
644 up_read(¤t->mm->mmap_sem);
645 if (result < 0)
646 break;
647
648 if ((unsigned)result < npages) {
649 bytes = result * PAGE_SIZE;
650 if (bytes <= pgbase) {
651 nfs_direct_release_pages(pagevec, result);
652 break;
653 }
654 bytes -= pgbase;
655 npages = result;
656 }
657
658 for (i = 0; i < npages; i++) {
659 struct nfs_page *req;
660 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
661
662 req = nfs_create_request(dreq->ctx, dreq->inode,
663 pagevec[i],
664 pgbase, req_len);
665 if (IS_ERR(req)) {
666 result = PTR_ERR(req);
667 break;
668 }
669 nfs_lock_request(req);
670 req->wb_index = pos >> PAGE_SHIFT;
671 req->wb_offset = pos & ~PAGE_MASK;
672 if (!nfs_pageio_add_request(desc, req)) {
673 result = desc->pg_error;
674 nfs_unlock_and_release_request(req);
675 break;
676 }
677 pgbase = 0;
678 bytes -= req_len;
679 started += req_len;
680 user_addr += req_len;
681 pos += req_len;
682 count -= req_len;
683 }
684 /* The nfs_page now hold references to these pages */
685 nfs_direct_release_pages(pagevec, npages);
686 } while (count != 0 && result >= 0);
687
688 kfree(pagevec);
689
690 if (started)
691 return started;
692 return result < 0 ? (ssize_t) result : -EFAULT;
693}
694
695static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
696{
697 struct nfs_direct_req *dreq = hdr->dreq;
698 struct nfs_commit_info cinfo;
699 int bit = -1;
700 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
701
702 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
703 goto out_put;
704
705 nfs_init_cinfo_from_dreq(&cinfo, dreq);
706
707 spin_lock(&dreq->lock);
708
709 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
710 dreq->flags = 0;
711 dreq->error = hdr->error;
712 }
713 if (dreq->error != 0)
714 bit = NFS_IOHDR_ERROR;
715 else {
716 dreq->count += hdr->good_bytes;
717 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
718 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
719 bit = NFS_IOHDR_NEED_RESCHED;
720 } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
721 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
722 bit = NFS_IOHDR_NEED_RESCHED;
723 else if (dreq->flags == 0) {
724 memcpy(&dreq->verf, hdr->verf,
725 sizeof(dreq->verf));
726 bit = NFS_IOHDR_NEED_COMMIT;
727 dreq->flags = NFS_ODIRECT_DO_COMMIT;
728 } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
729 if (memcmp(&dreq->verf, hdr->verf, sizeof(dreq->verf))) {
730 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
731 bit = NFS_IOHDR_NEED_RESCHED;
732 } else
733 bit = NFS_IOHDR_NEED_COMMIT;
734 }
735 }
736 }
737 spin_unlock(&dreq->lock);
738
739 while (!list_empty(&hdr->pages)) {
740 req = nfs_list_entry(hdr->pages.next);
741 nfs_list_remove_request(req);
742 switch (bit) {
743 case NFS_IOHDR_NEED_RESCHED:
744 case NFS_IOHDR_NEED_COMMIT:
745 kref_get(&req->wb_kref);
746 nfs_mark_request_commit(req, hdr->lseg, &cinfo);
747 }
748 nfs_unlock_and_release_request(req);
749 }
750
751out_put:
752 if (put_dreq(dreq))
753 nfs_direct_write_complete(dreq, hdr->inode);
754 hdr->release(hdr);
755}
756
757static void nfs_write_sync_pgio_error(struct list_head *head)
758{
759 struct nfs_page *req;
760
761 while (!list_empty(head)) {
762 req = nfs_list_entry(head->next);
763 nfs_list_remove_request(req);
764 nfs_unlock_and_release_request(req);
765 }
766}
767
768static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
769 .error_cleanup = nfs_write_sync_pgio_error,
770 .init_hdr = nfs_direct_pgio_init,
771 .completion = nfs_direct_write_completion,
772};
773
774static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
775 const struct iovec *iov,
776 unsigned long nr_segs,
777 loff_t pos)
778{
779 struct nfs_pageio_descriptor desc;
780 struct inode *inode = dreq->inode;
781 ssize_t result = 0;
782 size_t requested_bytes = 0;
783 unsigned long seg;
784
785 nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE,
786 &nfs_direct_write_completion_ops);
787 desc.pg_dreq = dreq;
788 get_dreq(dreq);
789 atomic_inc(&inode->i_dio_count);
790
791 for (seg = 0; seg < nr_segs; seg++) {
792 const struct iovec *vec = &iov[seg];
793 result = nfs_direct_write_schedule_segment(&desc, vec, pos);
794 if (result < 0)
795 break;
796 requested_bytes += result;
797 if ((size_t)result < vec->iov_len)
798 break;
799 pos += vec->iov_len;
800 }
801 nfs_pageio_complete(&desc);
802 NFS_I(dreq->inode)->write_io += desc.pg_bytes_written;
803
804 /*
805 * If no bytes were started, return the error, and let the
806 * generic layer handle the completion.
807 */
808 if (requested_bytes == 0) {
809 inode_dio_done(inode);
810 nfs_direct_req_release(dreq);
811 return result < 0 ? result : -EIO;
812 }
813
814 if (put_dreq(dreq))
815 nfs_direct_write_complete(dreq, dreq->inode);
816 return 0;
817}
818
819static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
820 unsigned long nr_segs, loff_t pos,
821 size_t count)
822{
823 ssize_t result = -ENOMEM;
824 struct inode *inode = iocb->ki_filp->f_mapping->host;
825 struct nfs_direct_req *dreq;
826
827 dreq = nfs_direct_req_alloc();
828 if (!dreq)
829 goto out;
830
831 dreq->inode = inode;
832 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
833 dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
834 if (dreq->l_ctx == NULL)
835 goto out_release;
836 if (!is_sync_kiocb(iocb))
837 dreq->iocb = iocb;
838
839 result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos);
840 if (!result)
841 result = nfs_direct_wait(dreq);
842out_release:
843 nfs_direct_req_release(dreq);
844out:
845 return result;
846}
847
848/**
849 * nfs_file_direct_read - file direct read operation for NFS files
850 * @iocb: target I/O control block
851 * @iov: vector of user buffers into which to read data
852 * @nr_segs: size of iov vector
853 * @pos: byte offset in file where reading starts
854 *
855 * We use this function for direct reads instead of calling
856 * generic_file_aio_read() in order to avoid gfar's check to see if
857 * the request starts before the end of the file. For that check
858 * to work, we must generate a GETATTR before each direct read, and
859 * even then there is a window between the GETATTR and the subsequent
860 * READ where the file size could change. Our preference is simply
861 * to do all reads the application wants, and the server will take
862 * care of managing the end of file boundary.
863 *
864 * This function also eliminates unnecessarily updating the file's
865 * atime locally, as the NFS server sets the file's atime, and this
866 * client must read the updated atime from the server back into its
867 * cache.
868 */
869ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
870 unsigned long nr_segs, loff_t pos)
871{
872 ssize_t retval = -EINVAL;
873 struct file *file = iocb->ki_filp;
874 struct address_space *mapping = file->f_mapping;
875 size_t count;
876
877 count = iov_length(iov, nr_segs);
878 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
879
880 dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
881 file->f_path.dentry->d_parent->d_name.name,
882 file->f_path.dentry->d_name.name,
883 count, (long long) pos);
884
885 retval = 0;
886 if (!count)
887 goto out;
888
889 retval = nfs_sync_mapping(mapping);
890 if (retval)
891 goto out;
892
893 task_io_account_read(count);
894
895 retval = nfs_direct_read(iocb, iov, nr_segs, pos);
896 if (retval > 0)
897 iocb->ki_pos = pos + retval;
898
899out:
900 return retval;
901}
902
903/**
904 * nfs_file_direct_write - file direct write operation for NFS files
905 * @iocb: target I/O control block
906 * @iov: vector of user buffers from which to write data
907 * @nr_segs: size of iov vector
908 * @pos: byte offset in file where writing starts
909 *
910 * We use this function for direct writes instead of calling
911 * generic_file_aio_write() in order to avoid taking the inode
912 * semaphore and updating the i_size. The NFS server will set
913 * the new i_size and this client must read the updated size
914 * back into its cache. We let the server do generic write
915 * parameter checking and report problems.
916 *
917 * We eliminate local atime updates, see direct read above.
918 *
919 * We avoid unnecessary page cache invalidations for normal cached
920 * readers of this file.
921 *
922 * Note that O_APPEND is not supported for NFS direct writes, as there
923 * is no atomic O_APPEND write facility in the NFS protocol.
924 */
925ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
926 unsigned long nr_segs, loff_t pos)
927{
928 ssize_t retval = -EINVAL;
929 struct file *file = iocb->ki_filp;
930 struct address_space *mapping = file->f_mapping;
931 size_t count;
932
933 count = iov_length(iov, nr_segs);
934 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
935
936 dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
937 file->f_path.dentry->d_parent->d_name.name,
938 file->f_path.dentry->d_name.name,
939 count, (long long) pos);
940
941 retval = generic_write_checks(file, &pos, &count, 0);
942 if (retval)
943 goto out;
944
945 retval = -EINVAL;
946 if ((ssize_t) count < 0)
947 goto out;
948 retval = 0;
949 if (!count)
950 goto out;
951
952 retval = nfs_sync_mapping(mapping);
953 if (retval)
954 goto out;
955
956 task_io_account_write(count);
957
958 retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
959 if (retval > 0) {
960 struct inode *inode = mapping->host;
961
962 iocb->ki_pos = pos + retval;
963 spin_lock(&inode->i_lock);
964 if (i_size_read(inode) < iocb->ki_pos)
965 i_size_write(inode, iocb->ki_pos);
966 spin_unlock(&inode->i_lock);
967 }
968out:
969 return retval;
970}
971
972/**
973 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
974 *
975 */
976int __init nfs_init_directcache(void)
977{
978 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
979 sizeof(struct nfs_direct_req),
980 0, (SLAB_RECLAIM_ACCOUNT|
981 SLAB_MEM_SPREAD),
982 NULL);
983 if (nfs_direct_cachep == NULL)
984 return -ENOMEM;
985
986 return 0;
987}
988
989/**
990 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
991 *
992 */
993void nfs_destroy_directcache(void)
994{
995 kmem_cache_destroy(nfs_direct_cachep);
996}