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
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}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/fs/nfs/direct.c
4 *
5 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
6 *
7 * High-performance uncached I/O for the Linux NFS client
8 *
9 * There are important applications whose performance or correctness
10 * depends on uncached access to file data. Database clusters
11 * (multiple copies of the same instance running on separate hosts)
12 * implement their own cache coherency protocol that subsumes file
13 * system cache protocols. Applications that process datasets
14 * considerably larger than the client's memory do not always benefit
15 * from a local cache. A streaming video server, for instance, has no
16 * need to cache the contents of a file.
17 *
18 * When an application requests uncached I/O, all read and write requests
19 * are made directly to the server; data stored or fetched via these
20 * requests is not cached in the Linux page cache. The client does not
21 * correct unaligned requests from applications. All requested bytes are
22 * held on permanent storage before a direct write system call returns to
23 * an application.
24 *
25 * Solaris implements an uncached I/O facility called directio() that
26 * is used for backups and sequential I/O to very large files. Solaris
27 * also supports uncaching whole NFS partitions with "-o forcedirectio,"
28 * an undocumented mount option.
29 *
30 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
31 * help from Andrew Morton.
32 *
33 * 18 Dec 2001 Initial implementation for 2.4 --cel
34 * 08 Jul 2002 Version for 2.4.19, with bug fixes --trondmy
35 * 08 Jun 2003 Port to 2.5 APIs --cel
36 * 31 Mar 2004 Handle direct I/O without VFS support --cel
37 * 15 Sep 2004 Parallel async reads --cel
38 * 04 May 2005 support O_DIRECT with aio --cel
39 *
40 */
41
42#include <linux/errno.h>
43#include <linux/sched.h>
44#include <linux/kernel.h>
45#include <linux/file.h>
46#include <linux/pagemap.h>
47#include <linux/kref.h>
48#include <linux/slab.h>
49#include <linux/task_io_accounting_ops.h>
50#include <linux/module.h>
51
52#include <linux/nfs_fs.h>
53#include <linux/nfs_page.h>
54#include <linux/sunrpc/clnt.h>
55
56#include <linux/uaccess.h>
57#include <linux/atomic.h>
58
59#include "internal.h"
60#include "iostat.h"
61#include "pnfs.h"
62#include "fscache.h"
63#include "nfstrace.h"
64
65#define NFSDBG_FACILITY NFSDBG_VFS
66
67static struct kmem_cache *nfs_direct_cachep;
68
69static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
70static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
71static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
72static void nfs_direct_write_schedule_work(struct work_struct *work);
73
74static inline void get_dreq(struct nfs_direct_req *dreq)
75{
76 atomic_inc(&dreq->io_count);
77}
78
79static inline int put_dreq(struct nfs_direct_req *dreq)
80{
81 return atomic_dec_and_test(&dreq->io_count);
82}
83
84static void
85nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
86 const struct nfs_pgio_header *hdr,
87 ssize_t dreq_len)
88{
89 if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
90 test_bit(NFS_IOHDR_EOF, &hdr->flags)))
91 return;
92 if (dreq->max_count >= dreq_len) {
93 dreq->max_count = dreq_len;
94 if (dreq->count > dreq_len)
95 dreq->count = dreq_len;
96 }
97
98 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && !dreq->error)
99 dreq->error = hdr->error;
100}
101
102static void
103nfs_direct_count_bytes(struct nfs_direct_req *dreq,
104 const struct nfs_pgio_header *hdr)
105{
106 loff_t hdr_end = hdr->io_start + hdr->good_bytes;
107 ssize_t dreq_len = 0;
108
109 if (hdr_end > dreq->io_start)
110 dreq_len = hdr_end - dreq->io_start;
111
112 nfs_direct_handle_truncated(dreq, hdr, dreq_len);
113
114 if (dreq_len > dreq->max_count)
115 dreq_len = dreq->max_count;
116
117 if (dreq->count < dreq_len)
118 dreq->count = dreq_len;
119}
120
121static void nfs_direct_truncate_request(struct nfs_direct_req *dreq,
122 struct nfs_page *req)
123{
124 loff_t offs = req_offset(req);
125 size_t req_start = (size_t)(offs - dreq->io_start);
126
127 if (req_start < dreq->max_count)
128 dreq->max_count = req_start;
129 if (req_start < dreq->count)
130 dreq->count = req_start;
131}
132
133/**
134 * nfs_swap_rw - NFS address space operation for swap I/O
135 * @iocb: target I/O control block
136 * @iter: I/O buffer
137 *
138 * Perform IO to the swap-file. This is much like direct IO.
139 */
140int nfs_swap_rw(struct kiocb *iocb, struct iov_iter *iter)
141{
142 ssize_t ret;
143
144 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
145
146 if (iov_iter_rw(iter) == READ)
147 ret = nfs_file_direct_read(iocb, iter, true);
148 else
149 ret = nfs_file_direct_write(iocb, iter, true);
150 if (ret < 0)
151 return ret;
152 return 0;
153}
154
155static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
156{
157 unsigned int i;
158 for (i = 0; i < npages; i++)
159 put_page(pages[i]);
160}
161
162void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
163 struct nfs_direct_req *dreq)
164{
165 cinfo->inode = dreq->inode;
166 cinfo->mds = &dreq->mds_cinfo;
167 cinfo->ds = &dreq->ds_cinfo;
168 cinfo->dreq = dreq;
169 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
170}
171
172static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
173{
174 struct nfs_direct_req *dreq;
175
176 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
177 if (!dreq)
178 return NULL;
179
180 kref_init(&dreq->kref);
181 kref_get(&dreq->kref);
182 init_completion(&dreq->completion);
183 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
184 pnfs_init_ds_commit_info(&dreq->ds_cinfo);
185 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
186 spin_lock_init(&dreq->lock);
187
188 return dreq;
189}
190
191static void nfs_direct_req_free(struct kref *kref)
192{
193 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
194
195 pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode);
196 if (dreq->l_ctx != NULL)
197 nfs_put_lock_context(dreq->l_ctx);
198 if (dreq->ctx != NULL)
199 put_nfs_open_context(dreq->ctx);
200 kmem_cache_free(nfs_direct_cachep, dreq);
201}
202
203static void nfs_direct_req_release(struct nfs_direct_req *dreq)
204{
205 kref_put(&dreq->kref, nfs_direct_req_free);
206}
207
208ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq, loff_t offset)
209{
210 loff_t start = offset - dreq->io_start;
211 return dreq->max_count - start;
212}
213EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
214
215/*
216 * Collects and returns the final error value/byte-count.
217 */
218static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
219{
220 ssize_t result = -EIOCBQUEUED;
221
222 /* Async requests don't wait here */
223 if (dreq->iocb)
224 goto out;
225
226 result = wait_for_completion_killable(&dreq->completion);
227
228 if (!result) {
229 result = dreq->count;
230 WARN_ON_ONCE(dreq->count < 0);
231 }
232 if (!result)
233 result = dreq->error;
234
235out:
236 return (ssize_t) result;
237}
238
239/*
240 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
241 * the iocb is still valid here if this is a synchronous request.
242 */
243static void nfs_direct_complete(struct nfs_direct_req *dreq)
244{
245 struct inode *inode = dreq->inode;
246
247 inode_dio_end(inode);
248
249 if (dreq->iocb) {
250 long res = (long) dreq->error;
251 if (dreq->count != 0) {
252 res = (long) dreq->count;
253 WARN_ON_ONCE(dreq->count < 0);
254 }
255 dreq->iocb->ki_complete(dreq->iocb, res);
256 }
257
258 complete(&dreq->completion);
259
260 nfs_direct_req_release(dreq);
261}
262
263static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
264{
265 unsigned long bytes = 0;
266 struct nfs_direct_req *dreq = hdr->dreq;
267
268 spin_lock(&dreq->lock);
269 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
270 spin_unlock(&dreq->lock);
271 goto out_put;
272 }
273
274 nfs_direct_count_bytes(dreq, hdr);
275 spin_unlock(&dreq->lock);
276
277 while (!list_empty(&hdr->pages)) {
278 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
279 struct page *page = req->wb_page;
280
281 if (!PageCompound(page) && bytes < hdr->good_bytes &&
282 (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
283 set_page_dirty(page);
284 bytes += req->wb_bytes;
285 nfs_list_remove_request(req);
286 nfs_release_request(req);
287 }
288out_put:
289 if (put_dreq(dreq))
290 nfs_direct_complete(dreq);
291 hdr->release(hdr);
292}
293
294static void nfs_read_sync_pgio_error(struct list_head *head, int error)
295{
296 struct nfs_page *req;
297
298 while (!list_empty(head)) {
299 req = nfs_list_entry(head->next);
300 nfs_list_remove_request(req);
301 nfs_release_request(req);
302 }
303}
304
305static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
306{
307 get_dreq(hdr->dreq);
308}
309
310static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
311 .error_cleanup = nfs_read_sync_pgio_error,
312 .init_hdr = nfs_direct_pgio_init,
313 .completion = nfs_direct_read_completion,
314};
315
316/*
317 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
318 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
319 * bail and stop sending more reads. Read length accounting is
320 * handled automatically by nfs_direct_read_result(). Otherwise, if
321 * no requests have been sent, just return an error.
322 */
323
324static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
325 struct iov_iter *iter,
326 loff_t pos)
327{
328 struct nfs_pageio_descriptor desc;
329 struct inode *inode = dreq->inode;
330 ssize_t result = -EINVAL;
331 size_t requested_bytes = 0;
332 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
333
334 nfs_pageio_init_read(&desc, dreq->inode, false,
335 &nfs_direct_read_completion_ops);
336 get_dreq(dreq);
337 desc.pg_dreq = dreq;
338 inode_dio_begin(inode);
339
340 while (iov_iter_count(iter)) {
341 struct page **pagevec;
342 size_t bytes;
343 size_t pgbase;
344 unsigned npages, i;
345
346 result = iov_iter_get_pages_alloc2(iter, &pagevec,
347 rsize, &pgbase);
348 if (result < 0)
349 break;
350
351 bytes = result;
352 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
353 for (i = 0; i < npages; i++) {
354 struct nfs_page *req;
355 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
356 /* XXX do we need to do the eof zeroing found in async_filler? */
357 req = nfs_page_create_from_page(dreq->ctx, pagevec[i],
358 pgbase, pos, req_len);
359 if (IS_ERR(req)) {
360 result = PTR_ERR(req);
361 break;
362 }
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 requested_bytes += req_len;
371 pos += req_len;
372 }
373 nfs_direct_release_pages(pagevec, npages);
374 kvfree(pagevec);
375 if (result < 0)
376 break;
377 }
378
379 nfs_pageio_complete(&desc);
380
381 /*
382 * If no bytes were started, return the error, and let the
383 * generic layer handle the completion.
384 */
385 if (requested_bytes == 0) {
386 inode_dio_end(inode);
387 nfs_direct_req_release(dreq);
388 return result < 0 ? result : -EIO;
389 }
390
391 if (put_dreq(dreq))
392 nfs_direct_complete(dreq);
393 return requested_bytes;
394}
395
396/**
397 * nfs_file_direct_read - file direct read operation for NFS files
398 * @iocb: target I/O control block
399 * @iter: vector of user buffers into which to read data
400 * @swap: flag indicating this is swap IO, not O_DIRECT IO
401 *
402 * We use this function for direct reads instead of calling
403 * generic_file_aio_read() in order to avoid gfar's check to see if
404 * the request starts before the end of the file. For that check
405 * to work, we must generate a GETATTR before each direct read, and
406 * even then there is a window between the GETATTR and the subsequent
407 * READ where the file size could change. Our preference is simply
408 * to do all reads the application wants, and the server will take
409 * care of managing the end of file boundary.
410 *
411 * This function also eliminates unnecessarily updating the file's
412 * atime locally, as the NFS server sets the file's atime, and this
413 * client must read the updated atime from the server back into its
414 * cache.
415 */
416ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
417 bool swap)
418{
419 struct file *file = iocb->ki_filp;
420 struct address_space *mapping = file->f_mapping;
421 struct inode *inode = mapping->host;
422 struct nfs_direct_req *dreq;
423 struct nfs_lock_context *l_ctx;
424 ssize_t result, requested;
425 size_t count = iov_iter_count(iter);
426 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
427
428 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
429 file, count, (long long) iocb->ki_pos);
430
431 result = 0;
432 if (!count)
433 goto out;
434
435 task_io_account_read(count);
436
437 result = -ENOMEM;
438 dreq = nfs_direct_req_alloc();
439 if (dreq == NULL)
440 goto out;
441
442 dreq->inode = inode;
443 dreq->max_count = count;
444 dreq->io_start = iocb->ki_pos;
445 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
446 l_ctx = nfs_get_lock_context(dreq->ctx);
447 if (IS_ERR(l_ctx)) {
448 result = PTR_ERR(l_ctx);
449 nfs_direct_req_release(dreq);
450 goto out_release;
451 }
452 dreq->l_ctx = l_ctx;
453 if (!is_sync_kiocb(iocb))
454 dreq->iocb = iocb;
455
456 if (user_backed_iter(iter))
457 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
458
459 if (!swap)
460 nfs_start_io_direct(inode);
461
462 NFS_I(inode)->read_io += count;
463 requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
464
465 if (!swap)
466 nfs_end_io_direct(inode);
467
468 if (requested > 0) {
469 result = nfs_direct_wait(dreq);
470 if (result > 0) {
471 requested -= result;
472 iocb->ki_pos += result;
473 }
474 iov_iter_revert(iter, requested);
475 } else {
476 result = requested;
477 }
478
479out_release:
480 nfs_direct_req_release(dreq);
481out:
482 return result;
483}
484
485static void nfs_direct_add_page_head(struct list_head *list,
486 struct nfs_page *req)
487{
488 struct nfs_page *head = req->wb_head;
489
490 if (!list_empty(&head->wb_list) || !nfs_lock_request(head))
491 return;
492 if (!list_empty(&head->wb_list)) {
493 nfs_unlock_request(head);
494 return;
495 }
496 list_add(&head->wb_list, list);
497 kref_get(&head->wb_kref);
498 kref_get(&head->wb_kref);
499}
500
501static void nfs_direct_join_group(struct list_head *list,
502 struct nfs_commit_info *cinfo,
503 struct inode *inode)
504{
505 struct nfs_page *req, *subreq;
506
507 list_for_each_entry(req, list, wb_list) {
508 if (req->wb_head != req) {
509 nfs_direct_add_page_head(&req->wb_list, req);
510 continue;
511 }
512 subreq = req->wb_this_page;
513 if (subreq == req)
514 continue;
515 do {
516 /*
517 * Remove subrequests from this list before freeing
518 * them in the call to nfs_join_page_group().
519 */
520 if (!list_empty(&subreq->wb_list)) {
521 nfs_list_remove_request(subreq);
522 nfs_release_request(subreq);
523 }
524 } while ((subreq = subreq->wb_this_page) != req);
525 nfs_join_page_group(req, cinfo, inode);
526 }
527}
528
529static void
530nfs_direct_write_scan_commit_list(struct inode *inode,
531 struct list_head *list,
532 struct nfs_commit_info *cinfo)
533{
534 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
535 pnfs_recover_commit_reqs(list, cinfo);
536 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
537 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
538}
539
540static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
541{
542 struct nfs_pageio_descriptor desc;
543 struct nfs_page *req;
544 LIST_HEAD(reqs);
545 struct nfs_commit_info cinfo;
546
547 nfs_init_cinfo_from_dreq(&cinfo, dreq);
548 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
549
550 nfs_direct_join_group(&reqs, &cinfo, dreq->inode);
551
552 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
553 get_dreq(dreq);
554
555 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
556 &nfs_direct_write_completion_ops);
557 desc.pg_dreq = dreq;
558
559 while (!list_empty(&reqs)) {
560 req = nfs_list_entry(reqs.next);
561 /* Bump the transmission count */
562 req->wb_nio++;
563 if (!nfs_pageio_add_request(&desc, req)) {
564 spin_lock(&dreq->lock);
565 if (dreq->error < 0) {
566 desc.pg_error = dreq->error;
567 } else if (desc.pg_error != -EAGAIN) {
568 dreq->flags = 0;
569 if (!desc.pg_error)
570 desc.pg_error = -EIO;
571 dreq->error = desc.pg_error;
572 } else
573 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
574 spin_unlock(&dreq->lock);
575 break;
576 }
577 nfs_release_request(req);
578 }
579 nfs_pageio_complete(&desc);
580
581 while (!list_empty(&reqs)) {
582 req = nfs_list_entry(reqs.next);
583 nfs_list_remove_request(req);
584 nfs_unlock_and_release_request(req);
585 if (desc.pg_error == -EAGAIN) {
586 nfs_mark_request_commit(req, NULL, &cinfo, 0);
587 } else {
588 spin_lock(&dreq->lock);
589 nfs_direct_truncate_request(dreq, req);
590 spin_unlock(&dreq->lock);
591 nfs_release_request(req);
592 }
593 }
594
595 if (put_dreq(dreq))
596 nfs_direct_write_complete(dreq);
597}
598
599static void nfs_direct_commit_complete(struct nfs_commit_data *data)
600{
601 const struct nfs_writeverf *verf = data->res.verf;
602 struct nfs_direct_req *dreq = data->dreq;
603 struct nfs_commit_info cinfo;
604 struct nfs_page *req;
605 int status = data->task.tk_status;
606
607 trace_nfs_direct_commit_complete(dreq);
608
609 spin_lock(&dreq->lock);
610 if (status < 0) {
611 /* Errors in commit are fatal */
612 dreq->error = status;
613 dreq->flags = NFS_ODIRECT_DONE;
614 } else {
615 status = dreq->error;
616 }
617 spin_unlock(&dreq->lock);
618
619 nfs_init_cinfo_from_dreq(&cinfo, dreq);
620
621 while (!list_empty(&data->pages)) {
622 req = nfs_list_entry(data->pages.next);
623 nfs_list_remove_request(req);
624 if (status < 0) {
625 spin_lock(&dreq->lock);
626 nfs_direct_truncate_request(dreq, req);
627 spin_unlock(&dreq->lock);
628 nfs_release_request(req);
629 } else if (!nfs_write_match_verf(verf, req)) {
630 spin_lock(&dreq->lock);
631 if (dreq->flags == 0)
632 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
633 spin_unlock(&dreq->lock);
634 /*
635 * Despite the reboot, the write was successful,
636 * so reset wb_nio.
637 */
638 req->wb_nio = 0;
639 nfs_mark_request_commit(req, NULL, &cinfo, 0);
640 } else
641 nfs_release_request(req);
642 nfs_unlock_and_release_request(req);
643 }
644
645 if (nfs_commit_end(cinfo.mds))
646 nfs_direct_write_complete(dreq);
647}
648
649static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
650 struct nfs_page *req)
651{
652 struct nfs_direct_req *dreq = cinfo->dreq;
653
654 trace_nfs_direct_resched_write(dreq);
655
656 spin_lock(&dreq->lock);
657 if (dreq->flags != NFS_ODIRECT_DONE)
658 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
659 spin_unlock(&dreq->lock);
660 nfs_mark_request_commit(req, NULL, cinfo, 0);
661}
662
663static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
664 .completion = nfs_direct_commit_complete,
665 .resched_write = nfs_direct_resched_write,
666};
667
668static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
669{
670 int res;
671 struct nfs_commit_info cinfo;
672 LIST_HEAD(mds_list);
673
674 nfs_init_cinfo_from_dreq(&cinfo, dreq);
675 nfs_commit_begin(cinfo.mds);
676 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
677 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
678 if (res < 0) { /* res == -ENOMEM */
679 spin_lock(&dreq->lock);
680 if (dreq->flags == 0)
681 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
682 spin_unlock(&dreq->lock);
683 }
684 if (nfs_commit_end(cinfo.mds))
685 nfs_direct_write_complete(dreq);
686}
687
688static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
689{
690 struct nfs_commit_info cinfo;
691 struct nfs_page *req;
692 LIST_HEAD(reqs);
693
694 nfs_init_cinfo_from_dreq(&cinfo, dreq);
695 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
696
697 while (!list_empty(&reqs)) {
698 req = nfs_list_entry(reqs.next);
699 nfs_list_remove_request(req);
700 nfs_direct_truncate_request(dreq, req);
701 nfs_release_request(req);
702 nfs_unlock_and_release_request(req);
703 }
704}
705
706static void nfs_direct_write_schedule_work(struct work_struct *work)
707{
708 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
709 int flags = dreq->flags;
710
711 dreq->flags = 0;
712 switch (flags) {
713 case NFS_ODIRECT_DO_COMMIT:
714 nfs_direct_commit_schedule(dreq);
715 break;
716 case NFS_ODIRECT_RESCHED_WRITES:
717 nfs_direct_write_reschedule(dreq);
718 break;
719 default:
720 nfs_direct_write_clear_reqs(dreq);
721 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
722 nfs_direct_complete(dreq);
723 }
724}
725
726static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
727{
728 trace_nfs_direct_write_complete(dreq);
729 queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
730}
731
732static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
733{
734 struct nfs_direct_req *dreq = hdr->dreq;
735 struct nfs_commit_info cinfo;
736 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
737 int flags = NFS_ODIRECT_DONE;
738
739 trace_nfs_direct_write_completion(dreq);
740
741 nfs_init_cinfo_from_dreq(&cinfo, dreq);
742
743 spin_lock(&dreq->lock);
744 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
745 spin_unlock(&dreq->lock);
746 goto out_put;
747 }
748
749 nfs_direct_count_bytes(dreq, hdr);
750 if (test_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags) &&
751 !test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
752 if (!dreq->flags)
753 dreq->flags = NFS_ODIRECT_DO_COMMIT;
754 flags = dreq->flags;
755 }
756 spin_unlock(&dreq->lock);
757
758 while (!list_empty(&hdr->pages)) {
759
760 req = nfs_list_entry(hdr->pages.next);
761 nfs_list_remove_request(req);
762 if (flags == NFS_ODIRECT_DO_COMMIT) {
763 kref_get(&req->wb_kref);
764 memcpy(&req->wb_verf, &hdr->verf.verifier,
765 sizeof(req->wb_verf));
766 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
767 hdr->ds_commit_idx);
768 } else if (flags == NFS_ODIRECT_RESCHED_WRITES) {
769 kref_get(&req->wb_kref);
770 nfs_mark_request_commit(req, NULL, &cinfo, 0);
771 }
772 nfs_unlock_and_release_request(req);
773 }
774
775out_put:
776 if (put_dreq(dreq))
777 nfs_direct_write_complete(dreq);
778 hdr->release(hdr);
779}
780
781static void nfs_write_sync_pgio_error(struct list_head *head, int error)
782{
783 struct nfs_page *req;
784
785 while (!list_empty(head)) {
786 req = nfs_list_entry(head->next);
787 nfs_list_remove_request(req);
788 nfs_unlock_and_release_request(req);
789 }
790}
791
792static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
793{
794 struct nfs_direct_req *dreq = hdr->dreq;
795 struct nfs_page *req;
796 struct nfs_commit_info cinfo;
797
798 trace_nfs_direct_write_reschedule_io(dreq);
799
800 nfs_init_cinfo_from_dreq(&cinfo, dreq);
801 spin_lock(&dreq->lock);
802 if (dreq->error == 0)
803 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
804 set_bit(NFS_IOHDR_REDO, &hdr->flags);
805 spin_unlock(&dreq->lock);
806 while (!list_empty(&hdr->pages)) {
807 req = nfs_list_entry(hdr->pages.next);
808 nfs_list_remove_request(req);
809 nfs_unlock_request(req);
810 nfs_mark_request_commit(req, NULL, &cinfo, 0);
811 }
812}
813
814static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
815 .error_cleanup = nfs_write_sync_pgio_error,
816 .init_hdr = nfs_direct_pgio_init,
817 .completion = nfs_direct_write_completion,
818 .reschedule_io = nfs_direct_write_reschedule_io,
819};
820
821
822/*
823 * NB: Return the value of the first error return code. Subsequent
824 * errors after the first one are ignored.
825 */
826/*
827 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
828 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
829 * bail and stop sending more writes. Write length accounting is
830 * handled automatically by nfs_direct_write_result(). Otherwise, if
831 * no requests have been sent, just return an error.
832 */
833static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
834 struct iov_iter *iter,
835 loff_t pos, int ioflags)
836{
837 struct nfs_pageio_descriptor desc;
838 struct inode *inode = dreq->inode;
839 struct nfs_commit_info cinfo;
840 ssize_t result = 0;
841 size_t requested_bytes = 0;
842 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
843 bool defer = false;
844
845 trace_nfs_direct_write_schedule_iovec(dreq);
846
847 nfs_pageio_init_write(&desc, inode, ioflags, false,
848 &nfs_direct_write_completion_ops);
849 desc.pg_dreq = dreq;
850 get_dreq(dreq);
851 inode_dio_begin(inode);
852
853 NFS_I(inode)->write_io += iov_iter_count(iter);
854 while (iov_iter_count(iter)) {
855 struct page **pagevec;
856 size_t bytes;
857 size_t pgbase;
858 unsigned npages, i;
859
860 result = iov_iter_get_pages_alloc2(iter, &pagevec,
861 wsize, &pgbase);
862 if (result < 0)
863 break;
864
865 bytes = result;
866 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
867 for (i = 0; i < npages; i++) {
868 struct nfs_page *req;
869 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
870
871 req = nfs_page_create_from_page(dreq->ctx, pagevec[i],
872 pgbase, pos, req_len);
873 if (IS_ERR(req)) {
874 result = PTR_ERR(req);
875 break;
876 }
877
878 if (desc.pg_error < 0) {
879 nfs_free_request(req);
880 result = desc.pg_error;
881 break;
882 }
883
884 pgbase = 0;
885 bytes -= req_len;
886 requested_bytes += req_len;
887 pos += req_len;
888
889 if (defer) {
890 nfs_mark_request_commit(req, NULL, &cinfo, 0);
891 continue;
892 }
893
894 nfs_lock_request(req);
895 if (nfs_pageio_add_request(&desc, req))
896 continue;
897
898 /* Exit on hard errors */
899 if (desc.pg_error < 0 && desc.pg_error != -EAGAIN) {
900 result = desc.pg_error;
901 nfs_unlock_and_release_request(req);
902 break;
903 }
904
905 /* If the error is soft, defer remaining requests */
906 nfs_init_cinfo_from_dreq(&cinfo, dreq);
907 spin_lock(&dreq->lock);
908 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
909 spin_unlock(&dreq->lock);
910 nfs_unlock_request(req);
911 nfs_mark_request_commit(req, NULL, &cinfo, 0);
912 desc.pg_error = 0;
913 defer = true;
914 }
915 nfs_direct_release_pages(pagevec, npages);
916 kvfree(pagevec);
917 if (result < 0)
918 break;
919 }
920 nfs_pageio_complete(&desc);
921
922 /*
923 * If no bytes were started, return the error, and let the
924 * generic layer handle the completion.
925 */
926 if (requested_bytes == 0) {
927 inode_dio_end(inode);
928 nfs_direct_req_release(dreq);
929 return result < 0 ? result : -EIO;
930 }
931
932 if (put_dreq(dreq))
933 nfs_direct_write_complete(dreq);
934 return requested_bytes;
935}
936
937/**
938 * nfs_file_direct_write - file direct write operation for NFS files
939 * @iocb: target I/O control block
940 * @iter: vector of user buffers from which to write data
941 * @swap: flag indicating this is swap IO, not O_DIRECT IO
942 *
943 * We use this function for direct writes instead of calling
944 * generic_file_aio_write() in order to avoid taking the inode
945 * semaphore and updating the i_size. The NFS server will set
946 * the new i_size and this client must read the updated size
947 * back into its cache. We let the server do generic write
948 * parameter checking and report problems.
949 *
950 * We eliminate local atime updates, see direct read above.
951 *
952 * We avoid unnecessary page cache invalidations for normal cached
953 * readers of this file.
954 *
955 * Note that O_APPEND is not supported for NFS direct writes, as there
956 * is no atomic O_APPEND write facility in the NFS protocol.
957 */
958ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
959 bool swap)
960{
961 ssize_t result, requested;
962 size_t count;
963 struct file *file = iocb->ki_filp;
964 struct address_space *mapping = file->f_mapping;
965 struct inode *inode = mapping->host;
966 struct nfs_direct_req *dreq;
967 struct nfs_lock_context *l_ctx;
968 loff_t pos, end;
969
970 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
971 file, iov_iter_count(iter), (long long) iocb->ki_pos);
972
973 if (swap)
974 /* bypass generic checks */
975 result = iov_iter_count(iter);
976 else
977 result = generic_write_checks(iocb, iter);
978 if (result <= 0)
979 return result;
980 count = result;
981 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
982
983 pos = iocb->ki_pos;
984 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
985
986 task_io_account_write(count);
987
988 result = -ENOMEM;
989 dreq = nfs_direct_req_alloc();
990 if (!dreq)
991 goto out;
992
993 dreq->inode = inode;
994 dreq->max_count = count;
995 dreq->io_start = pos;
996 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
997 l_ctx = nfs_get_lock_context(dreq->ctx);
998 if (IS_ERR(l_ctx)) {
999 result = PTR_ERR(l_ctx);
1000 nfs_direct_req_release(dreq);
1001 goto out_release;
1002 }
1003 dreq->l_ctx = l_ctx;
1004 if (!is_sync_kiocb(iocb))
1005 dreq->iocb = iocb;
1006 pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode);
1007
1008 if (swap) {
1009 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1010 FLUSH_STABLE);
1011 } else {
1012 nfs_start_io_direct(inode);
1013
1014 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1015 FLUSH_COND_STABLE);
1016
1017 if (mapping->nrpages) {
1018 invalidate_inode_pages2_range(mapping,
1019 pos >> PAGE_SHIFT, end);
1020 }
1021
1022 nfs_end_io_direct(inode);
1023 }
1024
1025 if (requested > 0) {
1026 result = nfs_direct_wait(dreq);
1027 if (result > 0) {
1028 requested -= result;
1029 iocb->ki_pos = pos + result;
1030 /* XXX: should check the generic_write_sync retval */
1031 generic_write_sync(iocb, result);
1032 }
1033 iov_iter_revert(iter, requested);
1034 } else {
1035 result = requested;
1036 }
1037 nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE);
1038out_release:
1039 nfs_direct_req_release(dreq);
1040out:
1041 return result;
1042}
1043
1044/**
1045 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1046 *
1047 */
1048int __init nfs_init_directcache(void)
1049{
1050 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1051 sizeof(struct nfs_direct_req),
1052 0, SLAB_RECLAIM_ACCOUNT,
1053 NULL);
1054 if (nfs_direct_cachep == NULL)
1055 return -ENOMEM;
1056
1057 return 0;
1058}
1059
1060/**
1061 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1062 *
1063 */
1064void nfs_destroy_directcache(void)
1065{
1066 kmem_cache_destroy(nfs_direct_cachep);
1067}