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 <asm/uaccess.h>
56#include <linux/atomic.h>
57
58#include "internal.h"
59#include "iostat.h"
60#include "pnfs.h"
61
62#define NFSDBG_FACILITY NFSDBG_VFS
63
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_req {
70 struct kref kref; /* release manager */
71
72 /* I/O parameters */
73 struct nfs_open_context *ctx; /* file open context info */
74 struct nfs_lock_context *l_ctx; /* Lock context info */
75 struct kiocb * iocb; /* controlling i/o request */
76 struct inode * inode; /* target file of i/o */
77
78 /* completion state */
79 atomic_t io_count; /* i/os we're waiting for */
80 spinlock_t lock; /* protect completion state */
81 ssize_t count, /* bytes actually processed */
82 bytes_left, /* bytes left to be sent */
83 error; /* any reported error */
84 struct completion completion; /* wait for i/o completion */
85
86 /* commit state */
87 struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo */
88 struct pnfs_ds_commit_info ds_cinfo; /* Storage for cinfo */
89 struct work_struct work;
90 int flags;
91#define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
92#define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
93 struct nfs_writeverf verf; /* unstable write verifier */
94};
95
96static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
97static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
98static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
99static void nfs_direct_write_schedule_work(struct work_struct *work);
100
101static inline void get_dreq(struct nfs_direct_req *dreq)
102{
103 atomic_inc(&dreq->io_count);
104}
105
106static inline int put_dreq(struct nfs_direct_req *dreq)
107{
108 return atomic_dec_and_test(&dreq->io_count);
109}
110
111/**
112 * nfs_direct_IO - NFS address space operation for direct I/O
113 * @rw: direction (read or write)
114 * @iocb: target I/O control block
115 * @iov: array of vectors that define I/O buffer
116 * @pos: offset in file to begin the operation
117 * @nr_segs: size of iovec array
118 *
119 * The presence of this routine in the address space ops vector means
120 * the NFS client supports direct I/O. However, for most direct IO, we
121 * shunt off direct read and write requests before the VFS gets them,
122 * so this method is only ever called for swap.
123 */
124ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
125{
126#ifndef CONFIG_NFS_SWAP
127 dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n",
128 iocb->ki_filp, (long long) pos, nr_segs);
129
130 return -EINVAL;
131#else
132 VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
133
134 if (rw == READ || rw == KERNEL_READ)
135 return nfs_file_direct_read(iocb, iov, nr_segs, pos,
136 rw == READ ? true : false);
137 return nfs_file_direct_write(iocb, iov, nr_segs, pos,
138 rw == WRITE ? true : false);
139#endif /* CONFIG_NFS_SWAP */
140}
141
142static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
143{
144 unsigned int i;
145 for (i = 0; i < npages; i++)
146 page_cache_release(pages[i]);
147}
148
149void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
150 struct nfs_direct_req *dreq)
151{
152 cinfo->lock = &dreq->lock;
153 cinfo->mds = &dreq->mds_cinfo;
154 cinfo->ds = &dreq->ds_cinfo;
155 cinfo->dreq = dreq;
156 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
157}
158
159static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
160{
161 struct nfs_direct_req *dreq;
162
163 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
164 if (!dreq)
165 return NULL;
166
167 kref_init(&dreq->kref);
168 kref_get(&dreq->kref);
169 init_completion(&dreq->completion);
170 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
171 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
172 spin_lock_init(&dreq->lock);
173
174 return dreq;
175}
176
177static void nfs_direct_req_free(struct kref *kref)
178{
179 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
180
181 if (dreq->l_ctx != NULL)
182 nfs_put_lock_context(dreq->l_ctx);
183 if (dreq->ctx != NULL)
184 put_nfs_open_context(dreq->ctx);
185 kmem_cache_free(nfs_direct_cachep, dreq);
186}
187
188static void nfs_direct_req_release(struct nfs_direct_req *dreq)
189{
190 kref_put(&dreq->kref, nfs_direct_req_free);
191}
192
193ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
194{
195 return dreq->bytes_left;
196}
197EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
198
199/*
200 * Collects and returns the final error value/byte-count.
201 */
202static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
203{
204 ssize_t result = -EIOCBQUEUED;
205
206 /* Async requests don't wait here */
207 if (dreq->iocb)
208 goto out;
209
210 result = wait_for_completion_killable(&dreq->completion);
211
212 if (!result)
213 result = dreq->error;
214 if (!result)
215 result = dreq->count;
216
217out:
218 return (ssize_t) result;
219}
220
221/*
222 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
223 * the iocb is still valid here if this is a synchronous request.
224 */
225static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
226{
227 struct inode *inode = dreq->inode;
228
229 if (dreq->iocb && write) {
230 loff_t pos = dreq->iocb->ki_pos + dreq->count;
231
232 spin_lock(&inode->i_lock);
233 if (i_size_read(inode) < pos)
234 i_size_write(inode, pos);
235 spin_unlock(&inode->i_lock);
236 }
237
238 if (write)
239 nfs_zap_mapping(inode, inode->i_mapping);
240
241 inode_dio_done(inode);
242
243 if (dreq->iocb) {
244 long res = (long) dreq->error;
245 if (!res)
246 res = (long) dreq->count;
247 aio_complete(dreq->iocb, res, 0);
248 }
249
250 complete_all(&dreq->completion);
251
252 nfs_direct_req_release(dreq);
253}
254
255static void nfs_direct_readpage_release(struct nfs_page *req)
256{
257 dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
258 req->wb_context->dentry->d_inode->i_sb->s_id,
259 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
260 req->wb_bytes,
261 (long long)req_offset(req));
262 nfs_release_request(req);
263}
264
265static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
266{
267 unsigned long bytes = 0;
268 struct nfs_direct_req *dreq = hdr->dreq;
269
270 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
271 goto out_put;
272
273 spin_lock(&dreq->lock);
274 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
275 dreq->error = hdr->error;
276 else
277 dreq->count += hdr->good_bytes;
278 spin_unlock(&dreq->lock);
279
280 while (!list_empty(&hdr->pages)) {
281 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
282 struct page *page = req->wb_page;
283
284 if (!PageCompound(page) && bytes < hdr->good_bytes)
285 set_page_dirty(page);
286 bytes += req->wb_bytes;
287 nfs_list_remove_request(req);
288 nfs_direct_readpage_release(req);
289 }
290out_put:
291 if (put_dreq(dreq))
292 nfs_direct_complete(dreq, false);
293 hdr->release(hdr);
294}
295
296static void nfs_read_sync_pgio_error(struct list_head *head)
297{
298 struct nfs_page *req;
299
300 while (!list_empty(head)) {
301 req = nfs_list_entry(head->next);
302 nfs_list_remove_request(req);
303 nfs_release_request(req);
304 }
305}
306
307static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
308{
309 get_dreq(hdr->dreq);
310}
311
312static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
313 .error_cleanup = nfs_read_sync_pgio_error,
314 .init_hdr = nfs_direct_pgio_init,
315 .completion = nfs_direct_read_completion,
316};
317
318/*
319 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
320 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
321 * bail and stop sending more reads. Read length accounting is
322 * handled automatically by nfs_direct_read_result(). Otherwise, if
323 * no requests have been sent, just return an error.
324 */
325static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc,
326 const struct iovec *iov,
327 loff_t pos, bool uio)
328{
329 struct nfs_direct_req *dreq = desc->pg_dreq;
330 struct nfs_open_context *ctx = dreq->ctx;
331 struct inode *inode = ctx->dentry->d_inode;
332 unsigned long user_addr = (unsigned long)iov->iov_base;
333 size_t count = iov->iov_len;
334 size_t rsize = NFS_SERVER(inode)->rsize;
335 unsigned int pgbase;
336 int result;
337 ssize_t started = 0;
338 struct page **pagevec = NULL;
339 unsigned int npages;
340
341 do {
342 size_t bytes;
343 int i;
344
345 pgbase = user_addr & ~PAGE_MASK;
346 bytes = min(max_t(size_t, rsize, PAGE_SIZE), count);
347
348 result = -ENOMEM;
349 npages = nfs_page_array_len(pgbase, bytes);
350 if (!pagevec)
351 pagevec = kmalloc(npages * sizeof(struct page *),
352 GFP_KERNEL);
353 if (!pagevec)
354 break;
355 if (uio) {
356 down_read(¤t->mm->mmap_sem);
357 result = get_user_pages(current, current->mm, user_addr,
358 npages, 1, 0, pagevec, NULL);
359 up_read(¤t->mm->mmap_sem);
360 if (result < 0)
361 break;
362 } else {
363 WARN_ON(npages != 1);
364 result = get_kernel_page(user_addr, 1, pagevec);
365 if (WARN_ON(result != 1))
366 break;
367 }
368
369 if ((unsigned)result < npages) {
370 bytes = result * PAGE_SIZE;
371 if (bytes <= pgbase) {
372 nfs_direct_release_pages(pagevec, result);
373 break;
374 }
375 bytes -= pgbase;
376 npages = result;
377 }
378
379 for (i = 0; i < npages; i++) {
380 struct nfs_page *req;
381 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
382 /* XXX do we need to do the eof zeroing found in async_filler? */
383 req = nfs_create_request(dreq->ctx, dreq->inode,
384 pagevec[i],
385 pgbase, req_len);
386 if (IS_ERR(req)) {
387 result = PTR_ERR(req);
388 break;
389 }
390 req->wb_index = pos >> PAGE_SHIFT;
391 req->wb_offset = pos & ~PAGE_MASK;
392 if (!nfs_pageio_add_request(desc, req)) {
393 result = desc->pg_error;
394 nfs_release_request(req);
395 break;
396 }
397 pgbase = 0;
398 bytes -= req_len;
399 started += req_len;
400 user_addr += req_len;
401 pos += req_len;
402 count -= req_len;
403 dreq->bytes_left -= req_len;
404 }
405 /* The nfs_page now hold references to these pages */
406 nfs_direct_release_pages(pagevec, npages);
407 } while (count != 0 && result >= 0);
408
409 kfree(pagevec);
410
411 if (started)
412 return started;
413 return result < 0 ? (ssize_t) result : -EFAULT;
414}
415
416static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
417 const struct iovec *iov,
418 unsigned long nr_segs,
419 loff_t pos, bool uio)
420{
421 struct nfs_pageio_descriptor desc;
422 struct inode *inode = dreq->inode;
423 ssize_t result = -EINVAL;
424 size_t requested_bytes = 0;
425 unsigned long seg;
426
427 NFS_PROTO(dreq->inode)->read_pageio_init(&desc, dreq->inode,
428 &nfs_direct_read_completion_ops);
429 get_dreq(dreq);
430 desc.pg_dreq = dreq;
431 atomic_inc(&inode->i_dio_count);
432
433 for (seg = 0; seg < nr_segs; seg++) {
434 const struct iovec *vec = &iov[seg];
435 result = nfs_direct_read_schedule_segment(&desc, vec, pos, uio);
436 if (result < 0)
437 break;
438 requested_bytes += result;
439 if ((size_t)result < vec->iov_len)
440 break;
441 pos += vec->iov_len;
442 }
443
444 nfs_pageio_complete(&desc);
445
446 /*
447 * If no bytes were started, return the error, and let the
448 * generic layer handle the completion.
449 */
450 if (requested_bytes == 0) {
451 inode_dio_done(inode);
452 nfs_direct_req_release(dreq);
453 return result < 0 ? result : -EIO;
454 }
455
456 if (put_dreq(dreq))
457 nfs_direct_complete(dreq, false);
458 return 0;
459}
460
461/**
462 * nfs_file_direct_read - file direct read operation for NFS files
463 * @iocb: target I/O control block
464 * @iov: vector of user buffers into which to read data
465 * @nr_segs: size of iov vector
466 * @pos: byte offset in file where reading starts
467 *
468 * We use this function for direct reads instead of calling
469 * generic_file_aio_read() in order to avoid gfar's check to see if
470 * the request starts before the end of the file. For that check
471 * to work, we must generate a GETATTR before each direct read, and
472 * even then there is a window between the GETATTR and the subsequent
473 * READ where the file size could change. Our preference is simply
474 * to do all reads the application wants, and the server will take
475 * care of managing the end of file boundary.
476 *
477 * This function also eliminates unnecessarily updating the file's
478 * atime locally, as the NFS server sets the file's atime, and this
479 * client must read the updated atime from the server back into its
480 * cache.
481 */
482ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
483 unsigned long nr_segs, loff_t pos, bool uio)
484{
485 struct file *file = iocb->ki_filp;
486 struct address_space *mapping = file->f_mapping;
487 struct inode *inode = mapping->host;
488 struct nfs_direct_req *dreq;
489 struct nfs_lock_context *l_ctx;
490 ssize_t result = -EINVAL;
491 size_t count;
492
493 count = iov_length(iov, nr_segs);
494 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
495
496 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
497 file, count, (long long) pos);
498
499 result = 0;
500 if (!count)
501 goto out;
502
503 mutex_lock(&inode->i_mutex);
504 result = nfs_sync_mapping(mapping);
505 if (result)
506 goto out_unlock;
507
508 task_io_account_read(count);
509
510 result = -ENOMEM;
511 dreq = nfs_direct_req_alloc();
512 if (dreq == NULL)
513 goto out_unlock;
514
515 dreq->inode = inode;
516 dreq->bytes_left = iov_length(iov, nr_segs);
517 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
518 l_ctx = nfs_get_lock_context(dreq->ctx);
519 if (IS_ERR(l_ctx)) {
520 result = PTR_ERR(l_ctx);
521 goto out_release;
522 }
523 dreq->l_ctx = l_ctx;
524 if (!is_sync_kiocb(iocb))
525 dreq->iocb = iocb;
526
527 NFS_I(inode)->read_io += iov_length(iov, nr_segs);
528 result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos, uio);
529
530 mutex_unlock(&inode->i_mutex);
531
532 if (!result) {
533 result = nfs_direct_wait(dreq);
534 if (result > 0)
535 iocb->ki_pos = pos + result;
536 }
537
538 nfs_direct_req_release(dreq);
539 return result;
540
541out_release:
542 nfs_direct_req_release(dreq);
543out_unlock:
544 mutex_unlock(&inode->i_mutex);
545out:
546 return result;
547}
548
549#if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
550static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
551{
552 struct nfs_pageio_descriptor desc;
553 struct nfs_page *req, *tmp;
554 LIST_HEAD(reqs);
555 struct nfs_commit_info cinfo;
556 LIST_HEAD(failed);
557
558 nfs_init_cinfo_from_dreq(&cinfo, dreq);
559 pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
560 spin_lock(cinfo.lock);
561 nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
562 spin_unlock(cinfo.lock);
563
564 dreq->count = 0;
565 get_dreq(dreq);
566
567 NFS_PROTO(dreq->inode)->write_pageio_init(&desc, dreq->inode, FLUSH_STABLE,
568 &nfs_direct_write_completion_ops);
569 desc.pg_dreq = dreq;
570
571 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
572 if (!nfs_pageio_add_request(&desc, req)) {
573 nfs_list_remove_request(req);
574 nfs_list_add_request(req, &failed);
575 spin_lock(cinfo.lock);
576 dreq->flags = 0;
577 dreq->error = -EIO;
578 spin_unlock(cinfo.lock);
579 }
580 nfs_release_request(req);
581 }
582 nfs_pageio_complete(&desc);
583
584 while (!list_empty(&failed)) {
585 req = nfs_list_entry(failed.next);
586 nfs_list_remove_request(req);
587 nfs_unlock_and_release_request(req);
588 }
589
590 if (put_dreq(dreq))
591 nfs_direct_write_complete(dreq, dreq->inode);
592}
593
594static void nfs_direct_commit_complete(struct nfs_commit_data *data)
595{
596 struct nfs_direct_req *dreq = data->dreq;
597 struct nfs_commit_info cinfo;
598 struct nfs_page *req;
599 int status = data->task.tk_status;
600
601 nfs_init_cinfo_from_dreq(&cinfo, dreq);
602 if (status < 0) {
603 dprintk("NFS: %5u commit failed with error %d.\n",
604 data->task.tk_pid, status);
605 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
606 } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
607 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
608 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
609 }
610
611 dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
612 while (!list_empty(&data->pages)) {
613 req = nfs_list_entry(data->pages.next);
614 nfs_list_remove_request(req);
615 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
616 /* Note the rewrite will go through mds */
617 nfs_mark_request_commit(req, NULL, &cinfo);
618 } else
619 nfs_release_request(req);
620 nfs_unlock_and_release_request(req);
621 }
622
623 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
624 nfs_direct_write_complete(dreq, data->inode);
625}
626
627static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
628{
629 /* There is no lock to clear */
630}
631
632static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
633 .completion = nfs_direct_commit_complete,
634 .error_cleanup = nfs_direct_error_cleanup,
635};
636
637static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
638{
639 int res;
640 struct nfs_commit_info cinfo;
641 LIST_HEAD(mds_list);
642
643 nfs_init_cinfo_from_dreq(&cinfo, dreq);
644 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
645 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
646 if (res < 0) /* res == -ENOMEM */
647 nfs_direct_write_reschedule(dreq);
648}
649
650static void nfs_direct_write_schedule_work(struct work_struct *work)
651{
652 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
653 int flags = dreq->flags;
654
655 dreq->flags = 0;
656 switch (flags) {
657 case NFS_ODIRECT_DO_COMMIT:
658 nfs_direct_commit_schedule(dreq);
659 break;
660 case NFS_ODIRECT_RESCHED_WRITES:
661 nfs_direct_write_reschedule(dreq);
662 break;
663 default:
664 nfs_direct_complete(dreq, true);
665 }
666}
667
668static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
669{
670 schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
671}
672
673#else
674static void nfs_direct_write_schedule_work(struct work_struct *work)
675{
676}
677
678static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
679{
680 nfs_direct_complete(dreq, true);
681}
682#endif
683
684/*
685 * NB: Return the value of the first error return code. Subsequent
686 * errors after the first one are ignored.
687 */
688/*
689 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
690 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
691 * bail and stop sending more writes. Write length accounting is
692 * handled automatically by nfs_direct_write_result(). Otherwise, if
693 * no requests have been sent, just return an error.
694 */
695static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc,
696 const struct iovec *iov,
697 loff_t pos, bool uio)
698{
699 struct nfs_direct_req *dreq = desc->pg_dreq;
700 struct nfs_open_context *ctx = dreq->ctx;
701 struct inode *inode = ctx->dentry->d_inode;
702 unsigned long user_addr = (unsigned long)iov->iov_base;
703 size_t count = iov->iov_len;
704 size_t wsize = NFS_SERVER(inode)->wsize;
705 unsigned int pgbase;
706 int result;
707 ssize_t started = 0;
708 struct page **pagevec = NULL;
709 unsigned int npages;
710
711 do {
712 size_t bytes;
713 int i;
714
715 pgbase = user_addr & ~PAGE_MASK;
716 bytes = min(max_t(size_t, wsize, PAGE_SIZE), count);
717
718 result = -ENOMEM;
719 npages = nfs_page_array_len(pgbase, bytes);
720 if (!pagevec)
721 pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
722 if (!pagevec)
723 break;
724
725 if (uio) {
726 down_read(¤t->mm->mmap_sem);
727 result = get_user_pages(current, current->mm, user_addr,
728 npages, 0, 0, pagevec, NULL);
729 up_read(¤t->mm->mmap_sem);
730 if (result < 0)
731 break;
732 } else {
733 WARN_ON(npages != 1);
734 result = get_kernel_page(user_addr, 0, pagevec);
735 if (WARN_ON(result != 1))
736 break;
737 }
738
739 if ((unsigned)result < npages) {
740 bytes = result * PAGE_SIZE;
741 if (bytes <= pgbase) {
742 nfs_direct_release_pages(pagevec, result);
743 break;
744 }
745 bytes -= pgbase;
746 npages = result;
747 }
748
749 for (i = 0; i < npages; i++) {
750 struct nfs_page *req;
751 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
752
753 req = nfs_create_request(dreq->ctx, dreq->inode,
754 pagevec[i],
755 pgbase, req_len);
756 if (IS_ERR(req)) {
757 result = PTR_ERR(req);
758 break;
759 }
760 nfs_lock_request(req);
761 req->wb_index = pos >> PAGE_SHIFT;
762 req->wb_offset = pos & ~PAGE_MASK;
763 if (!nfs_pageio_add_request(desc, req)) {
764 result = desc->pg_error;
765 nfs_unlock_and_release_request(req);
766 break;
767 }
768 pgbase = 0;
769 bytes -= req_len;
770 started += req_len;
771 user_addr += req_len;
772 pos += req_len;
773 count -= req_len;
774 dreq->bytes_left -= req_len;
775 }
776 /* The nfs_page now hold references to these pages */
777 nfs_direct_release_pages(pagevec, npages);
778 } while (count != 0 && result >= 0);
779
780 kfree(pagevec);
781
782 if (started)
783 return started;
784 return result < 0 ? (ssize_t) result : -EFAULT;
785}
786
787static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
788{
789 struct nfs_direct_req *dreq = hdr->dreq;
790 struct nfs_commit_info cinfo;
791 int bit = -1;
792 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
793
794 if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
795 goto out_put;
796
797 nfs_init_cinfo_from_dreq(&cinfo, dreq);
798
799 spin_lock(&dreq->lock);
800
801 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
802 dreq->flags = 0;
803 dreq->error = hdr->error;
804 }
805 if (dreq->error != 0)
806 bit = NFS_IOHDR_ERROR;
807 else {
808 dreq->count += hdr->good_bytes;
809 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
810 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
811 bit = NFS_IOHDR_NEED_RESCHED;
812 } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
813 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
814 bit = NFS_IOHDR_NEED_RESCHED;
815 else if (dreq->flags == 0) {
816 memcpy(&dreq->verf, hdr->verf,
817 sizeof(dreq->verf));
818 bit = NFS_IOHDR_NEED_COMMIT;
819 dreq->flags = NFS_ODIRECT_DO_COMMIT;
820 } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
821 if (memcmp(&dreq->verf, hdr->verf, sizeof(dreq->verf))) {
822 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
823 bit = NFS_IOHDR_NEED_RESCHED;
824 } else
825 bit = NFS_IOHDR_NEED_COMMIT;
826 }
827 }
828 }
829 spin_unlock(&dreq->lock);
830
831 while (!list_empty(&hdr->pages)) {
832 req = nfs_list_entry(hdr->pages.next);
833 nfs_list_remove_request(req);
834 switch (bit) {
835 case NFS_IOHDR_NEED_RESCHED:
836 case NFS_IOHDR_NEED_COMMIT:
837 kref_get(&req->wb_kref);
838 nfs_mark_request_commit(req, hdr->lseg, &cinfo);
839 }
840 nfs_unlock_and_release_request(req);
841 }
842
843out_put:
844 if (put_dreq(dreq))
845 nfs_direct_write_complete(dreq, hdr->inode);
846 hdr->release(hdr);
847}
848
849static void nfs_write_sync_pgio_error(struct list_head *head)
850{
851 struct nfs_page *req;
852
853 while (!list_empty(head)) {
854 req = nfs_list_entry(head->next);
855 nfs_list_remove_request(req);
856 nfs_unlock_and_release_request(req);
857 }
858}
859
860static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
861 .error_cleanup = nfs_write_sync_pgio_error,
862 .init_hdr = nfs_direct_pgio_init,
863 .completion = nfs_direct_write_completion,
864};
865
866static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
867 const struct iovec *iov,
868 unsigned long nr_segs,
869 loff_t pos, bool uio)
870{
871 struct nfs_pageio_descriptor desc;
872 struct inode *inode = dreq->inode;
873 ssize_t result = 0;
874 size_t requested_bytes = 0;
875 unsigned long seg;
876
877 NFS_PROTO(inode)->write_pageio_init(&desc, inode, FLUSH_COND_STABLE,
878 &nfs_direct_write_completion_ops);
879 desc.pg_dreq = dreq;
880 get_dreq(dreq);
881 atomic_inc(&inode->i_dio_count);
882
883 NFS_I(dreq->inode)->write_io += iov_length(iov, nr_segs);
884 for (seg = 0; seg < nr_segs; seg++) {
885 const struct iovec *vec = &iov[seg];
886 result = nfs_direct_write_schedule_segment(&desc, vec, pos, uio);
887 if (result < 0)
888 break;
889 requested_bytes += result;
890 if ((size_t)result < vec->iov_len)
891 break;
892 pos += vec->iov_len;
893 }
894 nfs_pageio_complete(&desc);
895
896 /*
897 * If no bytes were started, return the error, and let the
898 * generic layer handle the completion.
899 */
900 if (requested_bytes == 0) {
901 inode_dio_done(inode);
902 nfs_direct_req_release(dreq);
903 return result < 0 ? result : -EIO;
904 }
905
906 if (put_dreq(dreq))
907 nfs_direct_write_complete(dreq, dreq->inode);
908 return 0;
909}
910
911/**
912 * nfs_file_direct_write - file direct write operation for NFS files
913 * @iocb: target I/O control block
914 * @iov: vector of user buffers from which to write data
915 * @nr_segs: size of iov vector
916 * @pos: byte offset in file where writing starts
917 *
918 * We use this function for direct writes instead of calling
919 * generic_file_aio_write() in order to avoid taking the inode
920 * semaphore and updating the i_size. The NFS server will set
921 * the new i_size and this client must read the updated size
922 * back into its cache. We let the server do generic write
923 * parameter checking and report problems.
924 *
925 * We eliminate local atime updates, see direct read above.
926 *
927 * We avoid unnecessary page cache invalidations for normal cached
928 * readers of this file.
929 *
930 * Note that O_APPEND is not supported for NFS direct writes, as there
931 * is no atomic O_APPEND write facility in the NFS protocol.
932 */
933ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
934 unsigned long nr_segs, loff_t pos, bool uio)
935{
936 ssize_t result = -EINVAL;
937 struct file *file = iocb->ki_filp;
938 struct address_space *mapping = file->f_mapping;
939 struct inode *inode = mapping->host;
940 struct nfs_direct_req *dreq;
941 struct nfs_lock_context *l_ctx;
942 loff_t end;
943 size_t count;
944
945 count = iov_length(iov, nr_segs);
946 end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
947
948 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
949
950 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
951 file, count, (long long) pos);
952
953 result = generic_write_checks(file, &pos, &count, 0);
954 if (result)
955 goto out;
956
957 result = -EINVAL;
958 if ((ssize_t) count < 0)
959 goto out;
960 result = 0;
961 if (!count)
962 goto out;
963
964 mutex_lock(&inode->i_mutex);
965
966 result = nfs_sync_mapping(mapping);
967 if (result)
968 goto out_unlock;
969
970 if (mapping->nrpages) {
971 result = invalidate_inode_pages2_range(mapping,
972 pos >> PAGE_CACHE_SHIFT, end);
973 if (result)
974 goto out_unlock;
975 }
976
977 task_io_account_write(count);
978
979 result = -ENOMEM;
980 dreq = nfs_direct_req_alloc();
981 if (!dreq)
982 goto out_unlock;
983
984 dreq->inode = inode;
985 dreq->bytes_left = count;
986 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
987 l_ctx = nfs_get_lock_context(dreq->ctx);
988 if (IS_ERR(l_ctx)) {
989 result = PTR_ERR(l_ctx);
990 goto out_release;
991 }
992 dreq->l_ctx = l_ctx;
993 if (!is_sync_kiocb(iocb))
994 dreq->iocb = iocb;
995
996 result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, uio);
997
998 if (mapping->nrpages) {
999 invalidate_inode_pages2_range(mapping,
1000 pos >> PAGE_CACHE_SHIFT, end);
1001 }
1002
1003 mutex_unlock(&inode->i_mutex);
1004
1005 if (!result) {
1006 result = nfs_direct_wait(dreq);
1007 if (result > 0) {
1008 struct inode *inode = mapping->host;
1009
1010 iocb->ki_pos = pos + result;
1011 spin_lock(&inode->i_lock);
1012 if (i_size_read(inode) < iocb->ki_pos)
1013 i_size_write(inode, iocb->ki_pos);
1014 spin_unlock(&inode->i_lock);
1015 }
1016 }
1017 nfs_direct_req_release(dreq);
1018 return result;
1019
1020out_release:
1021 nfs_direct_req_release(dreq);
1022out_unlock:
1023 mutex_unlock(&inode->i_mutex);
1024out:
1025 return result;
1026}
1027
1028/**
1029 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1030 *
1031 */
1032int __init nfs_init_directcache(void)
1033{
1034 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1035 sizeof(struct nfs_direct_req),
1036 0, (SLAB_RECLAIM_ACCOUNT|
1037 SLAB_MEM_SPREAD),
1038 NULL);
1039 if (nfs_direct_cachep == NULL)
1040 return -ENOMEM;
1041
1042 return 0;
1043}
1044
1045/**
1046 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1047 *
1048 */
1049void nfs_destroy_directcache(void)
1050{
1051 kmem_cache_destroy(nfs_direct_cachep);
1052}
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
63#define NFSDBG_FACILITY NFSDBG_VFS
64
65static struct kmem_cache *nfs_direct_cachep;
66
67struct nfs_direct_req {
68 struct kref kref; /* release manager */
69
70 /* I/O parameters */
71 struct nfs_open_context *ctx; /* file open context info */
72 struct nfs_lock_context *l_ctx; /* Lock context info */
73 struct kiocb * iocb; /* controlling i/o request */
74 struct inode * inode; /* target file of i/o */
75
76 /* completion state */
77 atomic_t io_count; /* i/os we're waiting for */
78 spinlock_t lock; /* protect completion state */
79
80 loff_t io_start; /* Start offset for I/O */
81 ssize_t count, /* bytes actually processed */
82 max_count, /* max expected count */
83 bytes_left, /* bytes left to be sent */
84 error; /* any reported error */
85 struct completion completion; /* wait for i/o completion */
86
87 /* commit state */
88 struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo */
89 struct pnfs_ds_commit_info ds_cinfo; /* Storage for cinfo */
90 struct work_struct work;
91 int flags;
92 /* for write */
93#define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
94#define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
95 /* for read */
96#define NFS_ODIRECT_SHOULD_DIRTY (3) /* dirty user-space page after read */
97#define NFS_ODIRECT_DONE INT_MAX /* write verification failed */
98};
99
100static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
101static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
102static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
103static void nfs_direct_write_schedule_work(struct work_struct *work);
104
105static inline void get_dreq(struct nfs_direct_req *dreq)
106{
107 atomic_inc(&dreq->io_count);
108}
109
110static inline int put_dreq(struct nfs_direct_req *dreq)
111{
112 return atomic_dec_and_test(&dreq->io_count);
113}
114
115static void
116nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
117 const struct nfs_pgio_header *hdr,
118 ssize_t dreq_len)
119{
120 if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
121 test_bit(NFS_IOHDR_EOF, &hdr->flags)))
122 return;
123 if (dreq->max_count >= dreq_len) {
124 dreq->max_count = dreq_len;
125 if (dreq->count > dreq_len)
126 dreq->count = dreq_len;
127
128 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
129 dreq->error = hdr->error;
130 else /* Clear outstanding error if this is EOF */
131 dreq->error = 0;
132 }
133}
134
135static void
136nfs_direct_count_bytes(struct nfs_direct_req *dreq,
137 const struct nfs_pgio_header *hdr)
138{
139 loff_t hdr_end = hdr->io_start + hdr->good_bytes;
140 ssize_t dreq_len = 0;
141
142 if (hdr_end > dreq->io_start)
143 dreq_len = hdr_end - dreq->io_start;
144
145 nfs_direct_handle_truncated(dreq, hdr, dreq_len);
146
147 if (dreq_len > dreq->max_count)
148 dreq_len = dreq->max_count;
149
150 if (dreq->count < dreq_len)
151 dreq->count = dreq_len;
152}
153
154/**
155 * nfs_direct_IO - NFS address space operation for direct I/O
156 * @iocb: target I/O control block
157 * @iter: I/O buffer
158 *
159 * The presence of this routine in the address space ops vector means
160 * the NFS client supports direct I/O. However, for most direct IO, we
161 * shunt off direct read and write requests before the VFS gets them,
162 * so this method is only ever called for swap.
163 */
164ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
165{
166 struct inode *inode = iocb->ki_filp->f_mapping->host;
167
168 /* we only support swap file calling nfs_direct_IO */
169 if (!IS_SWAPFILE(inode))
170 return 0;
171
172 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
173
174 if (iov_iter_rw(iter) == READ)
175 return nfs_file_direct_read(iocb, iter);
176 return nfs_file_direct_write(iocb, iter);
177}
178
179static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
180{
181 unsigned int i;
182 for (i = 0; i < npages; i++)
183 put_page(pages[i]);
184}
185
186void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
187 struct nfs_direct_req *dreq)
188{
189 cinfo->inode = dreq->inode;
190 cinfo->mds = &dreq->mds_cinfo;
191 cinfo->ds = &dreq->ds_cinfo;
192 cinfo->dreq = dreq;
193 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
194}
195
196static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
197{
198 struct nfs_direct_req *dreq;
199
200 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
201 if (!dreq)
202 return NULL;
203
204 kref_init(&dreq->kref);
205 kref_get(&dreq->kref);
206 init_completion(&dreq->completion);
207 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
208 pnfs_init_ds_commit_info(&dreq->ds_cinfo);
209 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
210 spin_lock_init(&dreq->lock);
211
212 return dreq;
213}
214
215static void nfs_direct_req_free(struct kref *kref)
216{
217 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
218
219 pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode);
220 if (dreq->l_ctx != NULL)
221 nfs_put_lock_context(dreq->l_ctx);
222 if (dreq->ctx != NULL)
223 put_nfs_open_context(dreq->ctx);
224 kmem_cache_free(nfs_direct_cachep, dreq);
225}
226
227static void nfs_direct_req_release(struct nfs_direct_req *dreq)
228{
229 kref_put(&dreq->kref, nfs_direct_req_free);
230}
231
232ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
233{
234 return dreq->bytes_left;
235}
236EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
237
238/*
239 * Collects and returns the final error value/byte-count.
240 */
241static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
242{
243 ssize_t result = -EIOCBQUEUED;
244
245 /* Async requests don't wait here */
246 if (dreq->iocb)
247 goto out;
248
249 result = wait_for_completion_killable(&dreq->completion);
250
251 if (!result) {
252 result = dreq->count;
253 WARN_ON_ONCE(dreq->count < 0);
254 }
255 if (!result)
256 result = dreq->error;
257
258out:
259 return (ssize_t) result;
260}
261
262/*
263 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
264 * the iocb is still valid here if this is a synchronous request.
265 */
266static void nfs_direct_complete(struct nfs_direct_req *dreq)
267{
268 struct inode *inode = dreq->inode;
269
270 inode_dio_end(inode);
271
272 if (dreq->iocb) {
273 long res = (long) dreq->error;
274 if (dreq->count != 0) {
275 res = (long) dreq->count;
276 WARN_ON_ONCE(dreq->count < 0);
277 }
278 dreq->iocb->ki_complete(dreq->iocb, res, 0);
279 }
280
281 complete(&dreq->completion);
282
283 nfs_direct_req_release(dreq);
284}
285
286static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
287{
288 unsigned long bytes = 0;
289 struct nfs_direct_req *dreq = hdr->dreq;
290
291 spin_lock(&dreq->lock);
292 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
293 spin_unlock(&dreq->lock);
294 goto out_put;
295 }
296
297 nfs_direct_count_bytes(dreq, hdr);
298 spin_unlock(&dreq->lock);
299
300 while (!list_empty(&hdr->pages)) {
301 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
302 struct page *page = req->wb_page;
303
304 if (!PageCompound(page) && bytes < hdr->good_bytes &&
305 (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
306 set_page_dirty(page);
307 bytes += req->wb_bytes;
308 nfs_list_remove_request(req);
309 nfs_release_request(req);
310 }
311out_put:
312 if (put_dreq(dreq))
313 nfs_direct_complete(dreq);
314 hdr->release(hdr);
315}
316
317static void nfs_read_sync_pgio_error(struct list_head *head, int error)
318{
319 struct nfs_page *req;
320
321 while (!list_empty(head)) {
322 req = nfs_list_entry(head->next);
323 nfs_list_remove_request(req);
324 nfs_release_request(req);
325 }
326}
327
328static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
329{
330 get_dreq(hdr->dreq);
331}
332
333static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
334 .error_cleanup = nfs_read_sync_pgio_error,
335 .init_hdr = nfs_direct_pgio_init,
336 .completion = nfs_direct_read_completion,
337};
338
339/*
340 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
341 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
342 * bail and stop sending more reads. Read length accounting is
343 * handled automatically by nfs_direct_read_result(). Otherwise, if
344 * no requests have been sent, just return an error.
345 */
346
347static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
348 struct iov_iter *iter,
349 loff_t pos)
350{
351 struct nfs_pageio_descriptor desc;
352 struct inode *inode = dreq->inode;
353 ssize_t result = -EINVAL;
354 size_t requested_bytes = 0;
355 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
356
357 nfs_pageio_init_read(&desc, dreq->inode, false,
358 &nfs_direct_read_completion_ops);
359 get_dreq(dreq);
360 desc.pg_dreq = dreq;
361 inode_dio_begin(inode);
362
363 while (iov_iter_count(iter)) {
364 struct page **pagevec;
365 size_t bytes;
366 size_t pgbase;
367 unsigned npages, i;
368
369 result = iov_iter_get_pages_alloc(iter, &pagevec,
370 rsize, &pgbase);
371 if (result < 0)
372 break;
373
374 bytes = result;
375 iov_iter_advance(iter, bytes);
376 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
377 for (i = 0; i < npages; i++) {
378 struct nfs_page *req;
379 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
380 /* XXX do we need to do the eof zeroing found in async_filler? */
381 req = nfs_create_request(dreq->ctx, pagevec[i],
382 pgbase, req_len);
383 if (IS_ERR(req)) {
384 result = PTR_ERR(req);
385 break;
386 }
387 req->wb_index = pos >> PAGE_SHIFT;
388 req->wb_offset = pos & ~PAGE_MASK;
389 if (!nfs_pageio_add_request(&desc, req)) {
390 result = desc.pg_error;
391 nfs_release_request(req);
392 break;
393 }
394 pgbase = 0;
395 bytes -= req_len;
396 requested_bytes += req_len;
397 pos += req_len;
398 dreq->bytes_left -= req_len;
399 }
400 nfs_direct_release_pages(pagevec, npages);
401 kvfree(pagevec);
402 if (result < 0)
403 break;
404 }
405
406 nfs_pageio_complete(&desc);
407
408 /*
409 * If no bytes were started, return the error, and let the
410 * generic layer handle the completion.
411 */
412 if (requested_bytes == 0) {
413 inode_dio_end(inode);
414 nfs_direct_req_release(dreq);
415 return result < 0 ? result : -EIO;
416 }
417
418 if (put_dreq(dreq))
419 nfs_direct_complete(dreq);
420 return requested_bytes;
421}
422
423/**
424 * nfs_file_direct_read - file direct read operation for NFS files
425 * @iocb: target I/O control block
426 * @iter: vector of user buffers into which to read data
427 *
428 * We use this function for direct reads instead of calling
429 * generic_file_aio_read() in order to avoid gfar's check to see if
430 * the request starts before the end of the file. For that check
431 * to work, we must generate a GETATTR before each direct read, and
432 * even then there is a window between the GETATTR and the subsequent
433 * READ where the file size could change. Our preference is simply
434 * to do all reads the application wants, and the server will take
435 * care of managing the end of file boundary.
436 *
437 * This function also eliminates unnecessarily updating the file's
438 * atime locally, as the NFS server sets the file's atime, and this
439 * client must read the updated atime from the server back into its
440 * cache.
441 */
442ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
443{
444 struct file *file = iocb->ki_filp;
445 struct address_space *mapping = file->f_mapping;
446 struct inode *inode = mapping->host;
447 struct nfs_direct_req *dreq;
448 struct nfs_lock_context *l_ctx;
449 ssize_t result, requested;
450 size_t count = iov_iter_count(iter);
451 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
452
453 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
454 file, count, (long long) iocb->ki_pos);
455
456 result = 0;
457 if (!count)
458 goto out;
459
460 task_io_account_read(count);
461
462 result = -ENOMEM;
463 dreq = nfs_direct_req_alloc();
464 if (dreq == NULL)
465 goto out;
466
467 dreq->inode = inode;
468 dreq->bytes_left = dreq->max_count = count;
469 dreq->io_start = iocb->ki_pos;
470 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
471 l_ctx = nfs_get_lock_context(dreq->ctx);
472 if (IS_ERR(l_ctx)) {
473 result = PTR_ERR(l_ctx);
474 nfs_direct_req_release(dreq);
475 goto out_release;
476 }
477 dreq->l_ctx = l_ctx;
478 if (!is_sync_kiocb(iocb))
479 dreq->iocb = iocb;
480
481 if (iter_is_iovec(iter))
482 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
483
484 nfs_start_io_direct(inode);
485
486 NFS_I(inode)->read_io += count;
487 requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
488
489 nfs_end_io_direct(inode);
490
491 if (requested > 0) {
492 result = nfs_direct_wait(dreq);
493 if (result > 0) {
494 requested -= result;
495 iocb->ki_pos += result;
496 }
497 iov_iter_revert(iter, requested);
498 } else {
499 result = requested;
500 }
501
502out_release:
503 nfs_direct_req_release(dreq);
504out:
505 return result;
506}
507
508static void
509nfs_direct_join_group(struct list_head *list, struct inode *inode)
510{
511 struct nfs_page *req, *next;
512
513 list_for_each_entry(req, list, wb_list) {
514 if (req->wb_head != req || req->wb_this_page == req)
515 continue;
516 for (next = req->wb_this_page;
517 next != req->wb_head;
518 next = next->wb_this_page) {
519 nfs_list_remove_request(next);
520 nfs_release_request(next);
521 }
522 nfs_join_page_group(req, inode);
523 }
524}
525
526static void
527nfs_direct_write_scan_commit_list(struct inode *inode,
528 struct list_head *list,
529 struct nfs_commit_info *cinfo)
530{
531 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
532 pnfs_recover_commit_reqs(list, cinfo);
533 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
534 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
535}
536
537static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
538{
539 struct nfs_pageio_descriptor desc;
540 struct nfs_page *req, *tmp;
541 LIST_HEAD(reqs);
542 struct nfs_commit_info cinfo;
543 LIST_HEAD(failed);
544
545 nfs_init_cinfo_from_dreq(&cinfo, dreq);
546 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
547
548 nfs_direct_join_group(&reqs, dreq->inode);
549
550 dreq->count = 0;
551 dreq->max_count = 0;
552 list_for_each_entry(req, &reqs, wb_list)
553 dreq->max_count += req->wb_bytes;
554 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
555 get_dreq(dreq);
556
557 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
558 &nfs_direct_write_completion_ops);
559 desc.pg_dreq = dreq;
560
561 list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
562 /* Bump the transmission count */
563 req->wb_nio++;
564 if (!nfs_pageio_add_request(&desc, req)) {
565 nfs_list_move_request(req, &failed);
566 spin_lock(&cinfo.inode->i_lock);
567 dreq->flags = 0;
568 if (desc.pg_error < 0)
569 dreq->error = desc.pg_error;
570 else
571 dreq->error = -EIO;
572 spin_unlock(&cinfo.inode->i_lock);
573 }
574 nfs_release_request(req);
575 }
576 nfs_pageio_complete(&desc);
577
578 while (!list_empty(&failed)) {
579 req = nfs_list_entry(failed.next);
580 nfs_list_remove_request(req);
581 nfs_unlock_and_release_request(req);
582 }
583
584 if (put_dreq(dreq))
585 nfs_direct_write_complete(dreq);
586}
587
588static void nfs_direct_commit_complete(struct nfs_commit_data *data)
589{
590 const struct nfs_writeverf *verf = data->res.verf;
591 struct nfs_direct_req *dreq = data->dreq;
592 struct nfs_commit_info cinfo;
593 struct nfs_page *req;
594 int status = data->task.tk_status;
595
596 if (status < 0) {
597 /* Errors in commit are fatal */
598 dreq->error = status;
599 dreq->max_count = 0;
600 dreq->count = 0;
601 dreq->flags = NFS_ODIRECT_DONE;
602 } else if (dreq->flags == NFS_ODIRECT_DONE)
603 status = dreq->error;
604
605 nfs_init_cinfo_from_dreq(&cinfo, dreq);
606
607 while (!list_empty(&data->pages)) {
608 req = nfs_list_entry(data->pages.next);
609 nfs_list_remove_request(req);
610 if (status >= 0 && !nfs_write_match_verf(verf, req)) {
611 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
612 /*
613 * Despite the reboot, the write was successful,
614 * so reset wb_nio.
615 */
616 req->wb_nio = 0;
617 nfs_mark_request_commit(req, NULL, &cinfo, 0);
618 } else /* Error or match */
619 nfs_release_request(req);
620 nfs_unlock_and_release_request(req);
621 }
622
623 if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
624 nfs_direct_write_complete(dreq);
625}
626
627static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
628 struct nfs_page *req)
629{
630 struct nfs_direct_req *dreq = cinfo->dreq;
631
632 spin_lock(&dreq->lock);
633 if (dreq->flags != NFS_ODIRECT_DONE)
634 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
635 spin_unlock(&dreq->lock);
636 nfs_mark_request_commit(req, NULL, cinfo, 0);
637}
638
639static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
640 .completion = nfs_direct_commit_complete,
641 .resched_write = nfs_direct_resched_write,
642};
643
644static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
645{
646 int res;
647 struct nfs_commit_info cinfo;
648 LIST_HEAD(mds_list);
649
650 nfs_init_cinfo_from_dreq(&cinfo, dreq);
651 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
652 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
653 if (res < 0) /* res == -ENOMEM */
654 nfs_direct_write_reschedule(dreq);
655}
656
657static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
658{
659 struct nfs_commit_info cinfo;
660 struct nfs_page *req;
661 LIST_HEAD(reqs);
662
663 nfs_init_cinfo_from_dreq(&cinfo, dreq);
664 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
665
666 while (!list_empty(&reqs)) {
667 req = nfs_list_entry(reqs.next);
668 nfs_list_remove_request(req);
669 nfs_release_request(req);
670 nfs_unlock_and_release_request(req);
671 }
672}
673
674static void nfs_direct_write_schedule_work(struct work_struct *work)
675{
676 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
677 int flags = dreq->flags;
678
679 dreq->flags = 0;
680 switch (flags) {
681 case NFS_ODIRECT_DO_COMMIT:
682 nfs_direct_commit_schedule(dreq);
683 break;
684 case NFS_ODIRECT_RESCHED_WRITES:
685 nfs_direct_write_reschedule(dreq);
686 break;
687 default:
688 nfs_direct_write_clear_reqs(dreq);
689 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
690 nfs_direct_complete(dreq);
691 }
692}
693
694static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
695{
696 queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
697}
698
699static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
700{
701 struct nfs_direct_req *dreq = hdr->dreq;
702 struct nfs_commit_info cinfo;
703 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
704 int flags = NFS_ODIRECT_DONE;
705
706 nfs_init_cinfo_from_dreq(&cinfo, dreq);
707
708 spin_lock(&dreq->lock);
709 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
710 spin_unlock(&dreq->lock);
711 goto out_put;
712 }
713
714 nfs_direct_count_bytes(dreq, hdr);
715 if (hdr->good_bytes != 0 && nfs_write_need_commit(hdr)) {
716 if (!dreq->flags)
717 dreq->flags = NFS_ODIRECT_DO_COMMIT;
718 flags = dreq->flags;
719 }
720 spin_unlock(&dreq->lock);
721
722 while (!list_empty(&hdr->pages)) {
723
724 req = nfs_list_entry(hdr->pages.next);
725 nfs_list_remove_request(req);
726 if (flags == NFS_ODIRECT_DO_COMMIT) {
727 kref_get(&req->wb_kref);
728 memcpy(&req->wb_verf, &hdr->verf.verifier,
729 sizeof(req->wb_verf));
730 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
731 hdr->ds_commit_idx);
732 } else if (flags == NFS_ODIRECT_RESCHED_WRITES) {
733 kref_get(&req->wb_kref);
734 nfs_mark_request_commit(req, NULL, &cinfo, 0);
735 }
736 nfs_unlock_and_release_request(req);
737 }
738
739out_put:
740 if (put_dreq(dreq))
741 nfs_direct_write_complete(dreq);
742 hdr->release(hdr);
743}
744
745static void nfs_write_sync_pgio_error(struct list_head *head, int error)
746{
747 struct nfs_page *req;
748
749 while (!list_empty(head)) {
750 req = nfs_list_entry(head->next);
751 nfs_list_remove_request(req);
752 nfs_unlock_and_release_request(req);
753 }
754}
755
756static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
757{
758 struct nfs_direct_req *dreq = hdr->dreq;
759
760 spin_lock(&dreq->lock);
761 if (dreq->error == 0) {
762 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
763 /* fake unstable write to let common nfs resend pages */
764 hdr->verf.committed = NFS_UNSTABLE;
765 hdr->good_bytes = hdr->args.offset + hdr->args.count -
766 hdr->io_start;
767 }
768 spin_unlock(&dreq->lock);
769}
770
771static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
772 .error_cleanup = nfs_write_sync_pgio_error,
773 .init_hdr = nfs_direct_pgio_init,
774 .completion = nfs_direct_write_completion,
775 .reschedule_io = nfs_direct_write_reschedule_io,
776};
777
778
779/*
780 * NB: Return the value of the first error return code. Subsequent
781 * errors after the first one are ignored.
782 */
783/*
784 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
785 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
786 * bail and stop sending more writes. Write length accounting is
787 * handled automatically by nfs_direct_write_result(). Otherwise, if
788 * no requests have been sent, just return an error.
789 */
790static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
791 struct iov_iter *iter,
792 loff_t pos)
793{
794 struct nfs_pageio_descriptor desc;
795 struct inode *inode = dreq->inode;
796 ssize_t result = 0;
797 size_t requested_bytes = 0;
798 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
799
800 nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
801 &nfs_direct_write_completion_ops);
802 desc.pg_dreq = dreq;
803 get_dreq(dreq);
804 inode_dio_begin(inode);
805
806 NFS_I(inode)->write_io += iov_iter_count(iter);
807 while (iov_iter_count(iter)) {
808 struct page **pagevec;
809 size_t bytes;
810 size_t pgbase;
811 unsigned npages, i;
812
813 result = iov_iter_get_pages_alloc(iter, &pagevec,
814 wsize, &pgbase);
815 if (result < 0)
816 break;
817
818 bytes = result;
819 iov_iter_advance(iter, bytes);
820 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
821 for (i = 0; i < npages; i++) {
822 struct nfs_page *req;
823 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
824
825 req = nfs_create_request(dreq->ctx, pagevec[i],
826 pgbase, req_len);
827 if (IS_ERR(req)) {
828 result = PTR_ERR(req);
829 break;
830 }
831
832 if (desc.pg_error < 0) {
833 nfs_free_request(req);
834 result = desc.pg_error;
835 break;
836 }
837
838 nfs_lock_request(req);
839 req->wb_index = pos >> PAGE_SHIFT;
840 req->wb_offset = pos & ~PAGE_MASK;
841 if (!nfs_pageio_add_request(&desc, req)) {
842 result = desc.pg_error;
843 nfs_unlock_and_release_request(req);
844 break;
845 }
846 pgbase = 0;
847 bytes -= req_len;
848 requested_bytes += req_len;
849 pos += req_len;
850 dreq->bytes_left -= req_len;
851 }
852 nfs_direct_release_pages(pagevec, npages);
853 kvfree(pagevec);
854 if (result < 0)
855 break;
856 }
857 nfs_pageio_complete(&desc);
858
859 /*
860 * If no bytes were started, return the error, and let the
861 * generic layer handle the completion.
862 */
863 if (requested_bytes == 0) {
864 inode_dio_end(inode);
865 nfs_direct_req_release(dreq);
866 return result < 0 ? result : -EIO;
867 }
868
869 if (put_dreq(dreq))
870 nfs_direct_write_complete(dreq);
871 return requested_bytes;
872}
873
874/**
875 * nfs_file_direct_write - file direct write operation for NFS files
876 * @iocb: target I/O control block
877 * @iter: vector of user buffers from which to write data
878 *
879 * We use this function for direct writes instead of calling
880 * generic_file_aio_write() in order to avoid taking the inode
881 * semaphore and updating the i_size. The NFS server will set
882 * the new i_size and this client must read the updated size
883 * back into its cache. We let the server do generic write
884 * parameter checking and report problems.
885 *
886 * We eliminate local atime updates, see direct read above.
887 *
888 * We avoid unnecessary page cache invalidations for normal cached
889 * readers of this file.
890 *
891 * Note that O_APPEND is not supported for NFS direct writes, as there
892 * is no atomic O_APPEND write facility in the NFS protocol.
893 */
894ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
895{
896 ssize_t result, requested;
897 size_t count;
898 struct file *file = iocb->ki_filp;
899 struct address_space *mapping = file->f_mapping;
900 struct inode *inode = mapping->host;
901 struct nfs_direct_req *dreq;
902 struct nfs_lock_context *l_ctx;
903 loff_t pos, end;
904
905 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
906 file, iov_iter_count(iter), (long long) iocb->ki_pos);
907
908 result = generic_write_checks(iocb, iter);
909 if (result <= 0)
910 return result;
911 count = result;
912 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
913
914 pos = iocb->ki_pos;
915 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
916
917 task_io_account_write(count);
918
919 result = -ENOMEM;
920 dreq = nfs_direct_req_alloc();
921 if (!dreq)
922 goto out;
923
924 dreq->inode = inode;
925 dreq->bytes_left = dreq->max_count = count;
926 dreq->io_start = pos;
927 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
928 l_ctx = nfs_get_lock_context(dreq->ctx);
929 if (IS_ERR(l_ctx)) {
930 result = PTR_ERR(l_ctx);
931 nfs_direct_req_release(dreq);
932 goto out_release;
933 }
934 dreq->l_ctx = l_ctx;
935 if (!is_sync_kiocb(iocb))
936 dreq->iocb = iocb;
937 pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode);
938
939 nfs_start_io_direct(inode);
940
941 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos);
942
943 if (mapping->nrpages) {
944 invalidate_inode_pages2_range(mapping,
945 pos >> PAGE_SHIFT, end);
946 }
947
948 nfs_end_io_direct(inode);
949
950 if (requested > 0) {
951 result = nfs_direct_wait(dreq);
952 if (result > 0) {
953 requested -= result;
954 iocb->ki_pos = pos + result;
955 /* XXX: should check the generic_write_sync retval */
956 generic_write_sync(iocb, result);
957 }
958 iov_iter_revert(iter, requested);
959 } else {
960 result = requested;
961 }
962out_release:
963 nfs_direct_req_release(dreq);
964out:
965 return result;
966}
967
968/**
969 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
970 *
971 */
972int __init nfs_init_directcache(void)
973{
974 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
975 sizeof(struct nfs_direct_req),
976 0, (SLAB_RECLAIM_ACCOUNT|
977 SLAB_MEM_SPREAD),
978 NULL);
979 if (nfs_direct_cachep == NULL)
980 return -ENOMEM;
981
982 return 0;
983}
984
985/**
986 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
987 *
988 */
989void nfs_destroy_directcache(void)
990{
991 kmem_cache_destroy(nfs_direct_cachep);
992}