xref: /openbmc/linux/fs/nfs/direct.c (revision 0e407915)
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 
67 static struct kmem_cache *nfs_direct_cachep;
68 
69 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
70 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
71 static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
72 static void nfs_direct_write_schedule_work(struct work_struct *work);
73 
74 static inline void get_dreq(struct nfs_direct_req *dreq)
75 {
76 	atomic_inc(&dreq->io_count);
77 }
78 
79 static inline int put_dreq(struct nfs_direct_req *dreq)
80 {
81 	return atomic_dec_and_test(&dreq->io_count);
82 }
83 
84 static void
85 nfs_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 		if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
98 			dreq->error = hdr->error;
99 		else /* Clear outstanding error if this is EOF */
100 			dreq->error = 0;
101 	}
102 }
103 
104 static void
105 nfs_direct_count_bytes(struct nfs_direct_req *dreq,
106 		       const struct nfs_pgio_header *hdr)
107 {
108 	loff_t hdr_end = hdr->io_start + hdr->good_bytes;
109 	ssize_t dreq_len = 0;
110 
111 	if (hdr_end > dreq->io_start)
112 		dreq_len = hdr_end - dreq->io_start;
113 
114 	nfs_direct_handle_truncated(dreq, hdr, dreq_len);
115 
116 	if (dreq_len > dreq->max_count)
117 		dreq_len = dreq->max_count;
118 
119 	if (dreq->count < dreq_len)
120 		dreq->count = dreq_len;
121 }
122 
123 /**
124  * nfs_swap_rw - NFS address space operation for swap I/O
125  * @iocb: target I/O control block
126  * @iter: I/O buffer
127  *
128  * Perform IO to the swap-file.  This is much like direct IO.
129  */
130 int nfs_swap_rw(struct kiocb *iocb, struct iov_iter *iter)
131 {
132 	ssize_t ret;
133 
134 	VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
135 
136 	if (iov_iter_rw(iter) == READ)
137 		ret = nfs_file_direct_read(iocb, iter, true);
138 	else
139 		ret = nfs_file_direct_write(iocb, iter, true);
140 	if (ret < 0)
141 		return ret;
142 	return 0;
143 }
144 
145 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
146 {
147 	unsigned int i;
148 	for (i = 0; i < npages; i++)
149 		put_page(pages[i]);
150 }
151 
152 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
153 			      struct nfs_direct_req *dreq)
154 {
155 	cinfo->inode = dreq->inode;
156 	cinfo->mds = &dreq->mds_cinfo;
157 	cinfo->ds = &dreq->ds_cinfo;
158 	cinfo->dreq = dreq;
159 	cinfo->completion_ops = &nfs_direct_commit_completion_ops;
160 }
161 
162 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
163 {
164 	struct nfs_direct_req *dreq;
165 
166 	dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
167 	if (!dreq)
168 		return NULL;
169 
170 	kref_init(&dreq->kref);
171 	kref_get(&dreq->kref);
172 	init_completion(&dreq->completion);
173 	INIT_LIST_HEAD(&dreq->mds_cinfo.list);
174 	pnfs_init_ds_commit_info(&dreq->ds_cinfo);
175 	INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
176 	spin_lock_init(&dreq->lock);
177 
178 	return dreq;
179 }
180 
181 static void nfs_direct_req_free(struct kref *kref)
182 {
183 	struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
184 
185 	pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode);
186 	if (dreq->l_ctx != NULL)
187 		nfs_put_lock_context(dreq->l_ctx);
188 	if (dreq->ctx != NULL)
189 		put_nfs_open_context(dreq->ctx);
190 	kmem_cache_free(nfs_direct_cachep, dreq);
191 }
192 
193 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
194 {
195 	kref_put(&dreq->kref, nfs_direct_req_free);
196 }
197 
198 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
199 {
200 	return dreq->bytes_left;
201 }
202 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
203 
204 /*
205  * Collects and returns the final error value/byte-count.
206  */
207 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
208 {
209 	ssize_t result = -EIOCBQUEUED;
210 
211 	/* Async requests don't wait here */
212 	if (dreq->iocb)
213 		goto out;
214 
215 	result = wait_for_completion_killable(&dreq->completion);
216 
217 	if (!result) {
218 		result = dreq->count;
219 		WARN_ON_ONCE(dreq->count < 0);
220 	}
221 	if (!result)
222 		result = dreq->error;
223 
224 out:
225 	return (ssize_t) result;
226 }
227 
228 /*
229  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
230  * the iocb is still valid here if this is a synchronous request.
231  */
232 static void nfs_direct_complete(struct nfs_direct_req *dreq)
233 {
234 	struct inode *inode = dreq->inode;
235 
236 	inode_dio_end(inode);
237 
238 	if (dreq->iocb) {
239 		long res = (long) dreq->error;
240 		if (dreq->count != 0) {
241 			res = (long) dreq->count;
242 			WARN_ON_ONCE(dreq->count < 0);
243 		}
244 		dreq->iocb->ki_complete(dreq->iocb, res);
245 	}
246 
247 	complete(&dreq->completion);
248 
249 	nfs_direct_req_release(dreq);
250 }
251 
252 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
253 {
254 	unsigned long bytes = 0;
255 	struct nfs_direct_req *dreq = hdr->dreq;
256 
257 	spin_lock(&dreq->lock);
258 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
259 		spin_unlock(&dreq->lock);
260 		goto out_put;
261 	}
262 
263 	nfs_direct_count_bytes(dreq, hdr);
264 	spin_unlock(&dreq->lock);
265 
266 	while (!list_empty(&hdr->pages)) {
267 		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
268 		struct page *page = req->wb_page;
269 
270 		if (!PageCompound(page) && bytes < hdr->good_bytes &&
271 		    (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
272 			set_page_dirty(page);
273 		bytes += req->wb_bytes;
274 		nfs_list_remove_request(req);
275 		nfs_release_request(req);
276 	}
277 out_put:
278 	if (put_dreq(dreq))
279 		nfs_direct_complete(dreq);
280 	hdr->release(hdr);
281 }
282 
283 static void nfs_read_sync_pgio_error(struct list_head *head, int error)
284 {
285 	struct nfs_page *req;
286 
287 	while (!list_empty(head)) {
288 		req = nfs_list_entry(head->next);
289 		nfs_list_remove_request(req);
290 		nfs_release_request(req);
291 	}
292 }
293 
294 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
295 {
296 	get_dreq(hdr->dreq);
297 }
298 
299 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
300 	.error_cleanup = nfs_read_sync_pgio_error,
301 	.init_hdr = nfs_direct_pgio_init,
302 	.completion = nfs_direct_read_completion,
303 };
304 
305 /*
306  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
307  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
308  * bail and stop sending more reads.  Read length accounting is
309  * handled automatically by nfs_direct_read_result().  Otherwise, if
310  * no requests have been sent, just return an error.
311  */
312 
313 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
314 					      struct iov_iter *iter,
315 					      loff_t pos)
316 {
317 	struct nfs_pageio_descriptor desc;
318 	struct inode *inode = dreq->inode;
319 	ssize_t result = -EINVAL;
320 	size_t requested_bytes = 0;
321 	size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
322 
323 	nfs_pageio_init_read(&desc, dreq->inode, false,
324 			     &nfs_direct_read_completion_ops);
325 	get_dreq(dreq);
326 	desc.pg_dreq = dreq;
327 	inode_dio_begin(inode);
328 
329 	while (iov_iter_count(iter)) {
330 		struct page **pagevec;
331 		size_t bytes;
332 		size_t pgbase;
333 		unsigned npages, i;
334 
335 		result = iov_iter_get_pages_alloc2(iter, &pagevec,
336 						  rsize, &pgbase);
337 		if (result < 0)
338 			break;
339 
340 		bytes = result;
341 		npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
342 		for (i = 0; i < npages; i++) {
343 			struct nfs_page *req;
344 			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
345 			/* XXX do we need to do the eof zeroing found in async_filler? */
346 			req = nfs_create_request(dreq->ctx, pagevec[i],
347 						 pgbase, req_len);
348 			if (IS_ERR(req)) {
349 				result = PTR_ERR(req);
350 				break;
351 			}
352 			req->wb_index = pos >> PAGE_SHIFT;
353 			req->wb_offset = pos & ~PAGE_MASK;
354 			if (!nfs_pageio_add_request(&desc, req)) {
355 				result = desc.pg_error;
356 				nfs_release_request(req);
357 				break;
358 			}
359 			pgbase = 0;
360 			bytes -= req_len;
361 			requested_bytes += req_len;
362 			pos += req_len;
363 			dreq->bytes_left -= req_len;
364 		}
365 		nfs_direct_release_pages(pagevec, npages);
366 		kvfree(pagevec);
367 		if (result < 0)
368 			break;
369 	}
370 
371 	nfs_pageio_complete(&desc);
372 
373 	/*
374 	 * If no bytes were started, return the error, and let the
375 	 * generic layer handle the completion.
376 	 */
377 	if (requested_bytes == 0) {
378 		inode_dio_end(inode);
379 		nfs_direct_req_release(dreq);
380 		return result < 0 ? result : -EIO;
381 	}
382 
383 	if (put_dreq(dreq))
384 		nfs_direct_complete(dreq);
385 	return requested_bytes;
386 }
387 
388 /**
389  * nfs_file_direct_read - file direct read operation for NFS files
390  * @iocb: target I/O control block
391  * @iter: vector of user buffers into which to read data
392  * @swap: flag indicating this is swap IO, not O_DIRECT IO
393  *
394  * We use this function for direct reads instead of calling
395  * generic_file_aio_read() in order to avoid gfar's check to see if
396  * the request starts before the end of the file.  For that check
397  * to work, we must generate a GETATTR before each direct read, and
398  * even then there is a window between the GETATTR and the subsequent
399  * READ where the file size could change.  Our preference is simply
400  * to do all reads the application wants, and the server will take
401  * care of managing the end of file boundary.
402  *
403  * This function also eliminates unnecessarily updating the file's
404  * atime locally, as the NFS server sets the file's atime, and this
405  * client must read the updated atime from the server back into its
406  * cache.
407  */
408 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
409 			     bool swap)
410 {
411 	struct file *file = iocb->ki_filp;
412 	struct address_space *mapping = file->f_mapping;
413 	struct inode *inode = mapping->host;
414 	struct nfs_direct_req *dreq;
415 	struct nfs_lock_context *l_ctx;
416 	ssize_t result, requested;
417 	size_t count = iov_iter_count(iter);
418 	nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
419 
420 	dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
421 		file, count, (long long) iocb->ki_pos);
422 
423 	result = 0;
424 	if (!count)
425 		goto out;
426 
427 	task_io_account_read(count);
428 
429 	result = -ENOMEM;
430 	dreq = nfs_direct_req_alloc();
431 	if (dreq == NULL)
432 		goto out;
433 
434 	dreq->inode = inode;
435 	dreq->bytes_left = dreq->max_count = count;
436 	dreq->io_start = iocb->ki_pos;
437 	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
438 	l_ctx = nfs_get_lock_context(dreq->ctx);
439 	if (IS_ERR(l_ctx)) {
440 		result = PTR_ERR(l_ctx);
441 		nfs_direct_req_release(dreq);
442 		goto out_release;
443 	}
444 	dreq->l_ctx = l_ctx;
445 	if (!is_sync_kiocb(iocb))
446 		dreq->iocb = iocb;
447 
448 	if (user_backed_iter(iter))
449 		dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
450 
451 	if (!swap)
452 		nfs_start_io_direct(inode);
453 
454 	NFS_I(inode)->read_io += count;
455 	requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
456 
457 	if (!swap)
458 		nfs_end_io_direct(inode);
459 
460 	if (requested > 0) {
461 		result = nfs_direct_wait(dreq);
462 		if (result > 0) {
463 			requested -= result;
464 			iocb->ki_pos += result;
465 		}
466 		iov_iter_revert(iter, requested);
467 	} else {
468 		result = requested;
469 	}
470 
471 out_release:
472 	nfs_direct_req_release(dreq);
473 out:
474 	return result;
475 }
476 
477 static void
478 nfs_direct_join_group(struct list_head *list, struct inode *inode)
479 {
480 	struct nfs_page *req, *next;
481 
482 	list_for_each_entry(req, list, wb_list) {
483 		if (req->wb_head != req || req->wb_this_page == req)
484 			continue;
485 		for (next = req->wb_this_page;
486 				next != req->wb_head;
487 				next = next->wb_this_page) {
488 			nfs_list_remove_request(next);
489 			nfs_release_request(next);
490 		}
491 		nfs_join_page_group(req, inode);
492 	}
493 }
494 
495 static void
496 nfs_direct_write_scan_commit_list(struct inode *inode,
497 				  struct list_head *list,
498 				  struct nfs_commit_info *cinfo)
499 {
500 	mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
501 	pnfs_recover_commit_reqs(list, cinfo);
502 	nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
503 	mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
504 }
505 
506 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
507 {
508 	struct nfs_pageio_descriptor desc;
509 	struct nfs_page *req, *tmp;
510 	LIST_HEAD(reqs);
511 	struct nfs_commit_info cinfo;
512 	LIST_HEAD(failed);
513 
514 	nfs_init_cinfo_from_dreq(&cinfo, dreq);
515 	nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
516 
517 	nfs_direct_join_group(&reqs, dreq->inode);
518 
519 	dreq->count = 0;
520 	dreq->max_count = 0;
521 	list_for_each_entry(req, &reqs, wb_list)
522 		dreq->max_count += req->wb_bytes;
523 	nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
524 	get_dreq(dreq);
525 
526 	nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
527 			      &nfs_direct_write_completion_ops);
528 	desc.pg_dreq = dreq;
529 
530 	list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
531 		/* Bump the transmission count */
532 		req->wb_nio++;
533 		if (!nfs_pageio_add_request(&desc, req)) {
534 			nfs_list_move_request(req, &failed);
535 			spin_lock(&cinfo.inode->i_lock);
536 			dreq->flags = 0;
537 			if (desc.pg_error < 0)
538 				dreq->error = desc.pg_error;
539 			else
540 				dreq->error = -EIO;
541 			spin_unlock(&cinfo.inode->i_lock);
542 		}
543 		nfs_release_request(req);
544 	}
545 	nfs_pageio_complete(&desc);
546 
547 	while (!list_empty(&failed)) {
548 		req = nfs_list_entry(failed.next);
549 		nfs_list_remove_request(req);
550 		nfs_unlock_and_release_request(req);
551 	}
552 
553 	if (put_dreq(dreq))
554 		nfs_direct_write_complete(dreq);
555 }
556 
557 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
558 {
559 	const struct nfs_writeverf *verf = data->res.verf;
560 	struct nfs_direct_req *dreq = data->dreq;
561 	struct nfs_commit_info cinfo;
562 	struct nfs_page *req;
563 	int status = data->task.tk_status;
564 
565 	trace_nfs_direct_commit_complete(dreq);
566 
567 	if (status < 0) {
568 		/* Errors in commit are fatal */
569 		dreq->error = status;
570 		dreq->max_count = 0;
571 		dreq->count = 0;
572 		dreq->flags = NFS_ODIRECT_DONE;
573 	} else {
574 		status = dreq->error;
575 	}
576 
577 	nfs_init_cinfo_from_dreq(&cinfo, dreq);
578 
579 	while (!list_empty(&data->pages)) {
580 		req = nfs_list_entry(data->pages.next);
581 		nfs_list_remove_request(req);
582 		if (status >= 0 && !nfs_write_match_verf(verf, req)) {
583 			dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
584 			/*
585 			 * Despite the reboot, the write was successful,
586 			 * so reset wb_nio.
587 			 */
588 			req->wb_nio = 0;
589 			nfs_mark_request_commit(req, NULL, &cinfo, 0);
590 		} else /* Error or match */
591 			nfs_release_request(req);
592 		nfs_unlock_and_release_request(req);
593 	}
594 
595 	if (nfs_commit_end(cinfo.mds))
596 		nfs_direct_write_complete(dreq);
597 }
598 
599 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
600 		struct nfs_page *req)
601 {
602 	struct nfs_direct_req *dreq = cinfo->dreq;
603 
604 	trace_nfs_direct_resched_write(dreq);
605 
606 	spin_lock(&dreq->lock);
607 	if (dreq->flags != NFS_ODIRECT_DONE)
608 		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
609 	spin_unlock(&dreq->lock);
610 	nfs_mark_request_commit(req, NULL, cinfo, 0);
611 }
612 
613 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
614 	.completion = nfs_direct_commit_complete,
615 	.resched_write = nfs_direct_resched_write,
616 };
617 
618 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
619 {
620 	int res;
621 	struct nfs_commit_info cinfo;
622 	LIST_HEAD(mds_list);
623 
624 	nfs_init_cinfo_from_dreq(&cinfo, dreq);
625 	nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
626 	res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
627 	if (res < 0) /* res == -ENOMEM */
628 		nfs_direct_write_reschedule(dreq);
629 }
630 
631 static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
632 {
633 	struct nfs_commit_info cinfo;
634 	struct nfs_page *req;
635 	LIST_HEAD(reqs);
636 
637 	nfs_init_cinfo_from_dreq(&cinfo, dreq);
638 	nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
639 
640 	while (!list_empty(&reqs)) {
641 		req = nfs_list_entry(reqs.next);
642 		nfs_list_remove_request(req);
643 		nfs_release_request(req);
644 		nfs_unlock_and_release_request(req);
645 	}
646 }
647 
648 static void nfs_direct_write_schedule_work(struct work_struct *work)
649 {
650 	struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
651 	int flags = dreq->flags;
652 
653 	dreq->flags = 0;
654 	switch (flags) {
655 		case NFS_ODIRECT_DO_COMMIT:
656 			nfs_direct_commit_schedule(dreq);
657 			break;
658 		case NFS_ODIRECT_RESCHED_WRITES:
659 			nfs_direct_write_reschedule(dreq);
660 			break;
661 		default:
662 			nfs_direct_write_clear_reqs(dreq);
663 			nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
664 			nfs_direct_complete(dreq);
665 	}
666 }
667 
668 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
669 {
670 	trace_nfs_direct_write_complete(dreq);
671 	queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
672 }
673 
674 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
675 {
676 	struct nfs_direct_req *dreq = hdr->dreq;
677 	struct nfs_commit_info cinfo;
678 	struct nfs_page *req = nfs_list_entry(hdr->pages.next);
679 	int flags = NFS_ODIRECT_DONE;
680 
681 	trace_nfs_direct_write_completion(dreq);
682 
683 	nfs_init_cinfo_from_dreq(&cinfo, dreq);
684 
685 	spin_lock(&dreq->lock);
686 	if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
687 		spin_unlock(&dreq->lock);
688 		goto out_put;
689 	}
690 
691 	nfs_direct_count_bytes(dreq, hdr);
692 	if (test_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags)) {
693 		if (!dreq->flags)
694 			dreq->flags = NFS_ODIRECT_DO_COMMIT;
695 		flags = dreq->flags;
696 	}
697 	spin_unlock(&dreq->lock);
698 
699 	while (!list_empty(&hdr->pages)) {
700 
701 		req = nfs_list_entry(hdr->pages.next);
702 		nfs_list_remove_request(req);
703 		if (flags == NFS_ODIRECT_DO_COMMIT) {
704 			kref_get(&req->wb_kref);
705 			memcpy(&req->wb_verf, &hdr->verf.verifier,
706 			       sizeof(req->wb_verf));
707 			nfs_mark_request_commit(req, hdr->lseg, &cinfo,
708 				hdr->ds_commit_idx);
709 		} else if (flags == NFS_ODIRECT_RESCHED_WRITES) {
710 			kref_get(&req->wb_kref);
711 			nfs_mark_request_commit(req, NULL, &cinfo, 0);
712 		}
713 		nfs_unlock_and_release_request(req);
714 	}
715 
716 out_put:
717 	if (put_dreq(dreq))
718 		nfs_direct_write_complete(dreq);
719 	hdr->release(hdr);
720 }
721 
722 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
723 {
724 	struct nfs_page *req;
725 
726 	while (!list_empty(head)) {
727 		req = nfs_list_entry(head->next);
728 		nfs_list_remove_request(req);
729 		nfs_unlock_and_release_request(req);
730 	}
731 }
732 
733 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
734 {
735 	struct nfs_direct_req *dreq = hdr->dreq;
736 
737 	trace_nfs_direct_write_reschedule_io(dreq);
738 
739 	spin_lock(&dreq->lock);
740 	if (dreq->error == 0) {
741 		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
742 		/* fake unstable write to let common nfs resend pages */
743 		hdr->verf.committed = NFS_UNSTABLE;
744 		hdr->good_bytes = hdr->args.offset + hdr->args.count -
745 			hdr->io_start;
746 	}
747 	spin_unlock(&dreq->lock);
748 }
749 
750 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
751 	.error_cleanup = nfs_write_sync_pgio_error,
752 	.init_hdr = nfs_direct_pgio_init,
753 	.completion = nfs_direct_write_completion,
754 	.reschedule_io = nfs_direct_write_reschedule_io,
755 };
756 
757 
758 /*
759  * NB: Return the value of the first error return code.  Subsequent
760  *     errors after the first one are ignored.
761  */
762 /*
763  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
764  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
765  * bail and stop sending more writes.  Write length accounting is
766  * handled automatically by nfs_direct_write_result().  Otherwise, if
767  * no requests have been sent, just return an error.
768  */
769 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
770 					       struct iov_iter *iter,
771 					       loff_t pos, int ioflags)
772 {
773 	struct nfs_pageio_descriptor desc;
774 	struct inode *inode = dreq->inode;
775 	ssize_t result = 0;
776 	size_t requested_bytes = 0;
777 	size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
778 
779 	trace_nfs_direct_write_schedule_iovec(dreq);
780 
781 	nfs_pageio_init_write(&desc, inode, ioflags, false,
782 			      &nfs_direct_write_completion_ops);
783 	desc.pg_dreq = dreq;
784 	get_dreq(dreq);
785 	inode_dio_begin(inode);
786 
787 	NFS_I(inode)->write_io += iov_iter_count(iter);
788 	while (iov_iter_count(iter)) {
789 		struct page **pagevec;
790 		size_t bytes;
791 		size_t pgbase;
792 		unsigned npages, i;
793 
794 		result = iov_iter_get_pages_alloc2(iter, &pagevec,
795 						  wsize, &pgbase);
796 		if (result < 0)
797 			break;
798 
799 		bytes = result;
800 		npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
801 		for (i = 0; i < npages; i++) {
802 			struct nfs_page *req;
803 			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
804 
805 			req = nfs_create_request(dreq->ctx, pagevec[i],
806 						 pgbase, req_len);
807 			if (IS_ERR(req)) {
808 				result = PTR_ERR(req);
809 				break;
810 			}
811 
812 			if (desc.pg_error < 0) {
813 				nfs_free_request(req);
814 				result = desc.pg_error;
815 				break;
816 			}
817 
818 			nfs_lock_request(req);
819 			req->wb_index = pos >> PAGE_SHIFT;
820 			req->wb_offset = pos & ~PAGE_MASK;
821 			if (!nfs_pageio_add_request(&desc, req)) {
822 				result = desc.pg_error;
823 				nfs_unlock_and_release_request(req);
824 				break;
825 			}
826 			pgbase = 0;
827 			bytes -= req_len;
828 			requested_bytes += req_len;
829 			pos += req_len;
830 			dreq->bytes_left -= req_len;
831 		}
832 		nfs_direct_release_pages(pagevec, npages);
833 		kvfree(pagevec);
834 		if (result < 0)
835 			break;
836 	}
837 	nfs_pageio_complete(&desc);
838 
839 	/*
840 	 * If no bytes were started, return the error, and let the
841 	 * generic layer handle the completion.
842 	 */
843 	if (requested_bytes == 0) {
844 		inode_dio_end(inode);
845 		nfs_direct_req_release(dreq);
846 		return result < 0 ? result : -EIO;
847 	}
848 
849 	if (put_dreq(dreq))
850 		nfs_direct_write_complete(dreq);
851 	return requested_bytes;
852 }
853 
854 /**
855  * nfs_file_direct_write - file direct write operation for NFS files
856  * @iocb: target I/O control block
857  * @iter: vector of user buffers from which to write data
858  * @swap: flag indicating this is swap IO, not O_DIRECT IO
859  *
860  * We use this function for direct writes instead of calling
861  * generic_file_aio_write() in order to avoid taking the inode
862  * semaphore and updating the i_size.  The NFS server will set
863  * the new i_size and this client must read the updated size
864  * back into its cache.  We let the server do generic write
865  * parameter checking and report problems.
866  *
867  * We eliminate local atime updates, see direct read above.
868  *
869  * We avoid unnecessary page cache invalidations for normal cached
870  * readers of this file.
871  *
872  * Note that O_APPEND is not supported for NFS direct writes, as there
873  * is no atomic O_APPEND write facility in the NFS protocol.
874  */
875 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
876 			      bool swap)
877 {
878 	ssize_t result, requested;
879 	size_t count;
880 	struct file *file = iocb->ki_filp;
881 	struct address_space *mapping = file->f_mapping;
882 	struct inode *inode = mapping->host;
883 	struct nfs_direct_req *dreq;
884 	struct nfs_lock_context *l_ctx;
885 	loff_t pos, end;
886 
887 	dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
888 		file, iov_iter_count(iter), (long long) iocb->ki_pos);
889 
890 	if (swap)
891 		/* bypass generic checks */
892 		result =  iov_iter_count(iter);
893 	else
894 		result = generic_write_checks(iocb, iter);
895 	if (result <= 0)
896 		return result;
897 	count = result;
898 	nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
899 
900 	pos = iocb->ki_pos;
901 	end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
902 
903 	task_io_account_write(count);
904 
905 	result = -ENOMEM;
906 	dreq = nfs_direct_req_alloc();
907 	if (!dreq)
908 		goto out;
909 
910 	dreq->inode = inode;
911 	dreq->bytes_left = dreq->max_count = count;
912 	dreq->io_start = pos;
913 	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
914 	l_ctx = nfs_get_lock_context(dreq->ctx);
915 	if (IS_ERR(l_ctx)) {
916 		result = PTR_ERR(l_ctx);
917 		nfs_direct_req_release(dreq);
918 		goto out_release;
919 	}
920 	dreq->l_ctx = l_ctx;
921 	if (!is_sync_kiocb(iocb))
922 		dreq->iocb = iocb;
923 	pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode);
924 
925 	if (swap) {
926 		requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
927 							    FLUSH_STABLE);
928 	} else {
929 		nfs_start_io_direct(inode);
930 
931 		requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
932 							    FLUSH_COND_STABLE);
933 
934 		if (mapping->nrpages) {
935 			invalidate_inode_pages2_range(mapping,
936 						      pos >> PAGE_SHIFT, end);
937 		}
938 
939 		nfs_end_io_direct(inode);
940 	}
941 
942 	if (requested > 0) {
943 		result = nfs_direct_wait(dreq);
944 		if (result > 0) {
945 			requested -= result;
946 			iocb->ki_pos = pos + result;
947 			/* XXX: should check the generic_write_sync retval */
948 			generic_write_sync(iocb, result);
949 		}
950 		iov_iter_revert(iter, requested);
951 	} else {
952 		result = requested;
953 	}
954 	nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE);
955 out_release:
956 	nfs_direct_req_release(dreq);
957 out:
958 	return result;
959 }
960 
961 /**
962  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
963  *
964  */
965 int __init nfs_init_directcache(void)
966 {
967 	nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
968 						sizeof(struct nfs_direct_req),
969 						0, (SLAB_RECLAIM_ACCOUNT|
970 							SLAB_MEM_SPREAD),
971 						NULL);
972 	if (nfs_direct_cachep == NULL)
973 		return -ENOMEM;
974 
975 	return 0;
976 }
977 
978 /**
979  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
980  *
981  */
982 void nfs_destroy_directcache(void)
983 {
984 	kmem_cache_destroy(nfs_direct_cachep);
985 }
986