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_mirror {
70	ssize_t count;
71};
72
73struct nfs_direct_req {
74	struct kref		kref;		/* release manager */
75
76	/* I/O parameters */
77	struct nfs_open_context	*ctx;		/* file open context info */
78	struct nfs_lock_context *l_ctx;		/* Lock context info */
79	struct kiocb *		iocb;		/* controlling i/o request */
80	struct inode *		inode;		/* target file of i/o */
81
82	/* completion state */
83	atomic_t		io_count;	/* i/os we're waiting for */
84	spinlock_t		lock;		/* protect completion state */
85
86	struct nfs_direct_mirror mirrors[NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX];
87	int			mirror_count;
88
89	ssize_t			count,		/* bytes actually processed */
90				bytes_left,	/* bytes left to be sent */
91				io_start,	/* start of IO */
92				error;		/* any reported error */
93	struct completion	completion;	/* wait for i/o completion */
94
95	/* commit state */
96	struct nfs_mds_commit_info mds_cinfo;	/* Storage for cinfo */
97	struct pnfs_ds_commit_info ds_cinfo;	/* Storage for cinfo */
98	struct work_struct	work;
99	int			flags;
100#define NFS_ODIRECT_DO_COMMIT		(1)	/* an unstable reply was received */
101#define NFS_ODIRECT_RESCHED_WRITES	(2)	/* write verification failed */
102	struct nfs_writeverf	verf;		/* unstable write verifier */
103};
104
105static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
106static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
107static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
108static void nfs_direct_write_schedule_work(struct work_struct *work);
109
110static inline void get_dreq(struct nfs_direct_req *dreq)
111{
112	atomic_inc(&dreq->io_count);
113}
114
115static inline int put_dreq(struct nfs_direct_req *dreq)
116{
117	return atomic_dec_and_test(&dreq->io_count);
118}
119
120void nfs_direct_set_resched_writes(struct nfs_direct_req *dreq)
121{
122	dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
123}
124EXPORT_SYMBOL_GPL(nfs_direct_set_resched_writes);
125
126static void
127nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
128{
129	int i;
130	ssize_t count;
131
132	if (dreq->mirror_count == 1) {
133		dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
134		dreq->count += hdr->good_bytes;
135	} else {
136		/* mirrored writes */
137		count = dreq->mirrors[hdr->pgio_mirror_idx].count;
138		if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
139			count = hdr->io_start + hdr->good_bytes - dreq->io_start;
140			dreq->mirrors[hdr->pgio_mirror_idx].count = count;
141		}
142		/* update the dreq->count by finding the minimum agreed count from all
143		 * mirrors */
144		count = dreq->mirrors[0].count;
145
146		for (i = 1; i < dreq->mirror_count; i++)
147			count = min(count, dreq->mirrors[i].count);
148
149		dreq->count = count;
150	}
151}
152
153/*
154 * nfs_direct_select_verf - select the right verifier
155 * @dreq - direct request possibly spanning multiple servers
156 * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
157 * @commit_idx - commit bucket index for the DS
158 *
159 * returns the correct verifier to use given the role of the server
160 */
161static struct nfs_writeverf *
162nfs_direct_select_verf(struct nfs_direct_req *dreq,
163		       struct nfs_client *ds_clp,
164		       int commit_idx)
165{
166	struct nfs_writeverf *verfp = &dreq->verf;
167
168#ifdef CONFIG_NFS_V4_1
169	if (ds_clp) {
170		/* pNFS is in use, use the DS verf */
171		if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
172			verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
173		else
174			WARN_ON_ONCE(1);
175	}
176#endif
177	return verfp;
178}
179
180
181/*
182 * nfs_direct_set_hdr_verf - set the write/commit verifier
183 * @dreq - direct request possibly spanning multiple servers
184 * @hdr - pageio header to validate against previously seen verfs
185 *
186 * Set the server's (MDS or DS) "seen" verifier
187 */
188static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
189				    struct nfs_pgio_header *hdr)
190{
191	struct nfs_writeverf *verfp;
192
193	verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
194	WARN_ON_ONCE(verfp->committed >= 0);
195	memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
196	WARN_ON_ONCE(verfp->committed < 0);
197}
198
199/*
200 * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
201 * @dreq - direct request possibly spanning multiple servers
202 * @hdr - pageio header to validate against previously seen verf
203 *
204 * set the server's "seen" verf if not initialized.
205 * returns result of comparison between @hdr->verf and the "seen"
206 * verf of the server used by @hdr (DS or MDS)
207 */
208static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
209					  struct nfs_pgio_header *hdr)
210{
211	struct nfs_writeverf *verfp;
212
213	verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
214	if (verfp->committed < 0) {
215		nfs_direct_set_hdr_verf(dreq, hdr);
216		return 0;
217	}
218	return memcmp(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
219}
220
221/*
222 * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
223 * @dreq - direct request possibly spanning multiple servers
224 * @data - commit data to validate against previously seen verf
225 *
226 * returns result of comparison between @data->verf and the verf of
227 * the server used by @data (DS or MDS)
228 */
229static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
230					   struct nfs_commit_data *data)
231{
232	struct nfs_writeverf *verfp;
233
234	verfp = nfs_direct_select_verf(dreq, data->ds_clp,
235					 data->ds_commit_index);
236
237	/* verifier not set so always fail */
238	if (verfp->committed < 0)
239		return 1;
240
241	return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf));
242}
243
244/**
245 * nfs_direct_IO - NFS address space operation for direct I/O
246 * @iocb: target I/O control block
247 * @iov: array of vectors that define I/O buffer
248 * @pos: offset in file to begin the operation
249 * @nr_segs: size of iovec array
250 *
251 * The presence of this routine in the address space ops vector means
252 * the NFS client supports direct I/O. However, for most direct IO, we
253 * shunt off direct read and write requests before the VFS gets them,
254 * so this method is only ever called for swap.
255 */
256ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter, loff_t pos)
257{
258	struct inode *inode = iocb->ki_filp->f_mapping->host;
259
260	/* we only support swap file calling nfs_direct_IO */
261	if (!IS_SWAPFILE(inode))
262		return 0;
263
264	VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
265
266	if (iov_iter_rw(iter) == READ)
267		return nfs_file_direct_read(iocb, iter, pos);
268	return nfs_file_direct_write(iocb, iter);
269}
270
271static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
272{
273	unsigned int i;
274	for (i = 0; i < npages; i++)
275		page_cache_release(pages[i]);
276}
277
278void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
279			      struct nfs_direct_req *dreq)
280{
281	cinfo->lock = &dreq->inode->i_lock;
282	cinfo->mds = &dreq->mds_cinfo;
283	cinfo->ds = &dreq->ds_cinfo;
284	cinfo->dreq = dreq;
285	cinfo->completion_ops = &nfs_direct_commit_completion_ops;
286}
287
288static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
289					     struct nfs_pageio_descriptor *pgio,
290					     struct nfs_page *req)
291{
292	int mirror_count = 1;
293
294	if (pgio->pg_ops->pg_get_mirror_count)
295		mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
296
297	dreq->mirror_count = mirror_count;
298}
299
300static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
301{
302	struct nfs_direct_req *dreq;
303
304	dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
305	if (!dreq)
306		return NULL;
307
308	kref_init(&dreq->kref);
309	kref_get(&dreq->kref);
310	init_completion(&dreq->completion);
311	INIT_LIST_HEAD(&dreq->mds_cinfo.list);
312	dreq->verf.committed = NFS_INVALID_STABLE_HOW;	/* not set yet */
313	INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
314	dreq->mirror_count = 1;
315	spin_lock_init(&dreq->lock);
316
317	return dreq;
318}
319
320static void nfs_direct_req_free(struct kref *kref)
321{
322	struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
323
324	nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
325	if (dreq->l_ctx != NULL)
326		nfs_put_lock_context(dreq->l_ctx);
327	if (dreq->ctx != NULL)
328		put_nfs_open_context(dreq->ctx);
329	kmem_cache_free(nfs_direct_cachep, dreq);
330}
331
332static void nfs_direct_req_release(struct nfs_direct_req *dreq)
333{
334	kref_put(&dreq->kref, nfs_direct_req_free);
335}
336
337ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
338{
339	return dreq->bytes_left;
340}
341EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
342
343/*
344 * Collects and returns the final error value/byte-count.
345 */
346static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
347{
348	ssize_t result = -EIOCBQUEUED;
349
350	/* Async requests don't wait here */
351	if (dreq->iocb)
352		goto out;
353
354	result = wait_for_completion_killable(&dreq->completion);
355
356	if (!result)
357		result = dreq->error;
358	if (!result)
359		result = dreq->count;
360
361out:
362	return (ssize_t) result;
363}
364
365/*
366 * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
367 * the iocb is still valid here if this is a synchronous request.
368 */
369static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
370{
371	struct inode *inode = dreq->inode;
372
373	if (dreq->iocb && write) {
374		loff_t pos = dreq->iocb->ki_pos + dreq->count;
375
376		spin_lock(&inode->i_lock);
377		if (i_size_read(inode) < pos)
378			i_size_write(inode, pos);
379		spin_unlock(&inode->i_lock);
380	}
381
382	if (write)
383		nfs_zap_mapping(inode, inode->i_mapping);
384
385	inode_dio_end(inode);
386
387	if (dreq->iocb) {
388		long res = (long) dreq->error;
389		if (!res)
390			res = (long) dreq->count;
391		dreq->iocb->ki_complete(dreq->iocb, res, 0);
392	}
393
394	complete_all(&dreq->completion);
395
396	nfs_direct_req_release(dreq);
397}
398
399static void nfs_direct_readpage_release(struct nfs_page *req)
400{
401	dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
402		d_inode(req->wb_context->dentry)->i_sb->s_id,
403		(unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
404		req->wb_bytes,
405		(long long)req_offset(req));
406	nfs_release_request(req);
407}
408
409static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
410{
411	unsigned long bytes = 0;
412	struct nfs_direct_req *dreq = hdr->dreq;
413
414	if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
415		goto out_put;
416
417	spin_lock(&dreq->lock);
418	if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
419		dreq->error = hdr->error;
420	else
421		nfs_direct_good_bytes(dreq, hdr);
422
423	spin_unlock(&dreq->lock);
424
425	while (!list_empty(&hdr->pages)) {
426		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
427		struct page *page = req->wb_page;
428
429		if (!PageCompound(page) && bytes < hdr->good_bytes)
430			set_page_dirty(page);
431		bytes += req->wb_bytes;
432		nfs_list_remove_request(req);
433		nfs_direct_readpage_release(req);
434	}
435out_put:
436	if (put_dreq(dreq))
437		nfs_direct_complete(dreq, false);
438	hdr->release(hdr);
439}
440
441static void nfs_read_sync_pgio_error(struct list_head *head)
442{
443	struct nfs_page *req;
444
445	while (!list_empty(head)) {
446		req = nfs_list_entry(head->next);
447		nfs_list_remove_request(req);
448		nfs_release_request(req);
449	}
450}
451
452static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
453{
454	get_dreq(hdr->dreq);
455}
456
457static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
458	.error_cleanup = nfs_read_sync_pgio_error,
459	.init_hdr = nfs_direct_pgio_init,
460	.completion = nfs_direct_read_completion,
461};
462
463/*
464 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
465 * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
466 * bail and stop sending more reads.  Read length accounting is
467 * handled automatically by nfs_direct_read_result().  Otherwise, if
468 * no requests have been sent, just return an error.
469 */
470
471static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
472					      struct iov_iter *iter,
473					      loff_t pos)
474{
475	struct nfs_pageio_descriptor desc;
476	struct inode *inode = dreq->inode;
477	ssize_t result = -EINVAL;
478	size_t requested_bytes = 0;
479	size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
480
481	nfs_pageio_init_read(&desc, dreq->inode, false,
482			     &nfs_direct_read_completion_ops);
483	get_dreq(dreq);
484	desc.pg_dreq = dreq;
485	inode_dio_begin(inode);
486
487	while (iov_iter_count(iter)) {
488		struct page **pagevec;
489		size_t bytes;
490		size_t pgbase;
491		unsigned npages, i;
492
493		result = iov_iter_get_pages_alloc(iter, &pagevec,
494						  rsize, &pgbase);
495		if (result < 0)
496			break;
497
498		bytes = result;
499		iov_iter_advance(iter, bytes);
500		npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
501		for (i = 0; i < npages; i++) {
502			struct nfs_page *req;
503			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
504			/* XXX do we need to do the eof zeroing found in async_filler? */
505			req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
506						 pgbase, req_len);
507			if (IS_ERR(req)) {
508				result = PTR_ERR(req);
509				break;
510			}
511			req->wb_index = pos >> PAGE_SHIFT;
512			req->wb_offset = pos & ~PAGE_MASK;
513			if (!nfs_pageio_add_request(&desc, req)) {
514				result = desc.pg_error;
515				nfs_release_request(req);
516				break;
517			}
518			pgbase = 0;
519			bytes -= req_len;
520			requested_bytes += req_len;
521			pos += req_len;
522			dreq->bytes_left -= req_len;
523		}
524		nfs_direct_release_pages(pagevec, npages);
525		kvfree(pagevec);
526		if (result < 0)
527			break;
528	}
529
530	nfs_pageio_complete(&desc);
531
532	/*
533	 * If no bytes were started, return the error, and let the
534	 * generic layer handle the completion.
535	 */
536	if (requested_bytes == 0) {
537		inode_dio_end(inode);
538		nfs_direct_req_release(dreq);
539		return result < 0 ? result : -EIO;
540	}
541
542	if (put_dreq(dreq))
543		nfs_direct_complete(dreq, false);
544	return 0;
545}
546
547/**
548 * nfs_file_direct_read - file direct read operation for NFS files
549 * @iocb: target I/O control block
550 * @iter: vector of user buffers into which to read data
551 * @pos: byte offset in file where reading starts
552 *
553 * We use this function for direct reads instead of calling
554 * generic_file_aio_read() in order to avoid gfar's check to see if
555 * the request starts before the end of the file.  For that check
556 * to work, we must generate a GETATTR before each direct read, and
557 * even then there is a window between the GETATTR and the subsequent
558 * READ where the file size could change.  Our preference is simply
559 * to do all reads the application wants, and the server will take
560 * care of managing the end of file boundary.
561 *
562 * This function also eliminates unnecessarily updating the file's
563 * atime locally, as the NFS server sets the file's atime, and this
564 * client must read the updated atime from the server back into its
565 * cache.
566 */
567ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
568				loff_t pos)
569{
570	struct file *file = iocb->ki_filp;
571	struct address_space *mapping = file->f_mapping;
572	struct inode *inode = mapping->host;
573	struct nfs_direct_req *dreq;
574	struct nfs_lock_context *l_ctx;
575	ssize_t result = -EINVAL;
576	size_t count = iov_iter_count(iter);
577	nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
578
579	dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
580		file, count, (long long) pos);
581
582	result = 0;
583	if (!count)
584		goto out;
585
586	mutex_lock(&inode->i_mutex);
587	result = nfs_sync_mapping(mapping);
588	if (result)
589		goto out_unlock;
590
591	task_io_account_read(count);
592
593	result = -ENOMEM;
594	dreq = nfs_direct_req_alloc();
595	if (dreq == NULL)
596		goto out_unlock;
597
598	dreq->inode = inode;
599	dreq->bytes_left = count;
600	dreq->io_start = pos;
601	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
602	l_ctx = nfs_get_lock_context(dreq->ctx);
603	if (IS_ERR(l_ctx)) {
604		result = PTR_ERR(l_ctx);
605		goto out_release;
606	}
607	dreq->l_ctx = l_ctx;
608	if (!is_sync_kiocb(iocb))
609		dreq->iocb = iocb;
610
611	NFS_I(inode)->read_io += count;
612	result = nfs_direct_read_schedule_iovec(dreq, iter, pos);
613
614	mutex_unlock(&inode->i_mutex);
615
616	if (!result) {
617		result = nfs_direct_wait(dreq);
618		if (result > 0)
619			iocb->ki_pos = pos + result;
620	}
621
622	nfs_direct_req_release(dreq);
623	return result;
624
625out_release:
626	nfs_direct_req_release(dreq);
627out_unlock:
628	mutex_unlock(&inode->i_mutex);
629out:
630	return result;
631}
632
633static void
634nfs_direct_write_scan_commit_list(struct inode *inode,
635				  struct list_head *list,
636				  struct nfs_commit_info *cinfo)
637{
638	spin_lock(cinfo->lock);
639#ifdef CONFIG_NFS_V4_1
640	if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
641		NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
642#endif
643	nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
644	spin_unlock(cinfo->lock);
645}
646
647static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
648{
649	struct nfs_pageio_descriptor desc;
650	struct nfs_page *req, *tmp;
651	LIST_HEAD(reqs);
652	struct nfs_commit_info cinfo;
653	LIST_HEAD(failed);
654	int i;
655
656	nfs_init_cinfo_from_dreq(&cinfo, dreq);
657	nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
658
659	dreq->count = 0;
660	for (i = 0; i < dreq->mirror_count; i++)
661		dreq->mirrors[i].count = 0;
662	get_dreq(dreq);
663
664	nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
665			      &nfs_direct_write_completion_ops);
666	desc.pg_dreq = dreq;
667
668	req = nfs_list_entry(reqs.next);
669	nfs_direct_setup_mirroring(dreq, &desc, req);
670
671	list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
672		if (!nfs_pageio_add_request(&desc, req)) {
673			nfs_list_remove_request(req);
674			nfs_list_add_request(req, &failed);
675			spin_lock(cinfo.lock);
676			dreq->flags = 0;
677			dreq->error = -EIO;
678			spin_unlock(cinfo.lock);
679		}
680		nfs_release_request(req);
681	}
682	nfs_pageio_complete(&desc);
683
684	while (!list_empty(&failed)) {
685		req = nfs_list_entry(failed.next);
686		nfs_list_remove_request(req);
687		nfs_unlock_and_release_request(req);
688	}
689
690	if (put_dreq(dreq))
691		nfs_direct_write_complete(dreq, dreq->inode);
692}
693
694static void nfs_direct_commit_complete(struct nfs_commit_data *data)
695{
696	struct nfs_direct_req *dreq = data->dreq;
697	struct nfs_commit_info cinfo;
698	struct nfs_page *req;
699	int status = data->task.tk_status;
700
701	nfs_init_cinfo_from_dreq(&cinfo, dreq);
702	if (status < 0) {
703		dprintk("NFS: %5u commit failed with error %d.\n",
704			data->task.tk_pid, status);
705		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
706	} else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
707		dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
708		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
709	}
710
711	dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
712	while (!list_empty(&data->pages)) {
713		req = nfs_list_entry(data->pages.next);
714		nfs_list_remove_request(req);
715		if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
716			/* Note the rewrite will go through mds */
717			nfs_mark_request_commit(req, NULL, &cinfo, 0);
718		} else
719			nfs_release_request(req);
720		nfs_unlock_and_release_request(req);
721	}
722
723	if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
724		nfs_direct_write_complete(dreq, data->inode);
725}
726
727static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
728{
729	/* There is no lock to clear */
730}
731
732static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
733	.completion = nfs_direct_commit_complete,
734	.error_cleanup = nfs_direct_error_cleanup,
735};
736
737static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
738{
739	int res;
740	struct nfs_commit_info cinfo;
741	LIST_HEAD(mds_list);
742
743	nfs_init_cinfo_from_dreq(&cinfo, dreq);
744	nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
745	res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
746	if (res < 0) /* res == -ENOMEM */
747		nfs_direct_write_reschedule(dreq);
748}
749
750static void nfs_direct_write_schedule_work(struct work_struct *work)
751{
752	struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
753	int flags = dreq->flags;
754
755	dreq->flags = 0;
756	switch (flags) {
757		case NFS_ODIRECT_DO_COMMIT:
758			nfs_direct_commit_schedule(dreq);
759			break;
760		case NFS_ODIRECT_RESCHED_WRITES:
761			nfs_direct_write_reschedule(dreq);
762			break;
763		default:
764			nfs_direct_complete(dreq, true);
765	}
766}
767
768static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
769{
770	schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
771}
772
773static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
774{
775	struct nfs_direct_req *dreq = hdr->dreq;
776	struct nfs_commit_info cinfo;
777	bool request_commit = false;
778	struct nfs_page *req = nfs_list_entry(hdr->pages.next);
779
780	if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
781		goto out_put;
782
783	nfs_init_cinfo_from_dreq(&cinfo, dreq);
784
785	spin_lock(&dreq->lock);
786
787	if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
788		dreq->flags = 0;
789		dreq->error = hdr->error;
790	}
791	if (dreq->error == 0) {
792		nfs_direct_good_bytes(dreq, hdr);
793		if (nfs_write_need_commit(hdr)) {
794			if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
795				request_commit = true;
796			else if (dreq->flags == 0) {
797				nfs_direct_set_hdr_verf(dreq, hdr);
798				request_commit = true;
799				dreq->flags = NFS_ODIRECT_DO_COMMIT;
800			} else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
801				request_commit = true;
802				if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
803					dreq->flags =
804						NFS_ODIRECT_RESCHED_WRITES;
805			}
806		}
807	}
808	spin_unlock(&dreq->lock);
809
810	while (!list_empty(&hdr->pages)) {
811
812		req = nfs_list_entry(hdr->pages.next);
813		nfs_list_remove_request(req);
814		if (request_commit) {
815			kref_get(&req->wb_kref);
816			nfs_mark_request_commit(req, hdr->lseg, &cinfo,
817				hdr->ds_commit_idx);
818		}
819		nfs_unlock_and_release_request(req);
820	}
821
822out_put:
823	if (put_dreq(dreq))
824		nfs_direct_write_complete(dreq, hdr->inode);
825	hdr->release(hdr);
826}
827
828static void nfs_write_sync_pgio_error(struct list_head *head)
829{
830	struct nfs_page *req;
831
832	while (!list_empty(head)) {
833		req = nfs_list_entry(head->next);
834		nfs_list_remove_request(req);
835		nfs_unlock_and_release_request(req);
836	}
837}
838
839static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
840	.error_cleanup = nfs_write_sync_pgio_error,
841	.init_hdr = nfs_direct_pgio_init,
842	.completion = nfs_direct_write_completion,
843};
844
845
846/*
847 * NB: Return the value of the first error return code.  Subsequent
848 *     errors after the first one are ignored.
849 */
850/*
851 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
852 * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
853 * bail and stop sending more writes.  Write length accounting is
854 * handled automatically by nfs_direct_write_result().  Otherwise, if
855 * no requests have been sent, just return an error.
856 */
857static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
858					       struct iov_iter *iter,
859					       loff_t pos)
860{
861	struct nfs_pageio_descriptor desc;
862	struct inode *inode = dreq->inode;
863	ssize_t result = 0;
864	size_t requested_bytes = 0;
865	size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
866
867	nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
868			      &nfs_direct_write_completion_ops);
869	desc.pg_dreq = dreq;
870	get_dreq(dreq);
871	inode_dio_begin(inode);
872
873	NFS_I(inode)->write_io += iov_iter_count(iter);
874	while (iov_iter_count(iter)) {
875		struct page **pagevec;
876		size_t bytes;
877		size_t pgbase;
878		unsigned npages, i;
879
880		result = iov_iter_get_pages_alloc(iter, &pagevec,
881						  wsize, &pgbase);
882		if (result < 0)
883			break;
884
885		bytes = result;
886		iov_iter_advance(iter, bytes);
887		npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
888		for (i = 0; i < npages; i++) {
889			struct nfs_page *req;
890			unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
891
892			req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
893						 pgbase, req_len);
894			if (IS_ERR(req)) {
895				result = PTR_ERR(req);
896				break;
897			}
898
899			nfs_direct_setup_mirroring(dreq, &desc, req);
900
901			nfs_lock_request(req);
902			req->wb_index = pos >> PAGE_SHIFT;
903			req->wb_offset = pos & ~PAGE_MASK;
904			if (!nfs_pageio_add_request(&desc, req)) {
905				result = desc.pg_error;
906				nfs_unlock_and_release_request(req);
907				break;
908			}
909			pgbase = 0;
910			bytes -= req_len;
911			requested_bytes += req_len;
912			pos += req_len;
913			dreq->bytes_left -= req_len;
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, dreq->inode);
934	return 0;
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 * @pos: byte offset in file where writing starts
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{
960	ssize_t result = -EINVAL;
961	struct file *file = iocb->ki_filp;
962	struct address_space *mapping = file->f_mapping;
963	struct inode *inode = mapping->host;
964	struct nfs_direct_req *dreq;
965	struct nfs_lock_context *l_ctx;
966	loff_t pos, end;
967
968	dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
969		file, iov_iter_count(iter), (long long) iocb->ki_pos);
970
971	nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES,
972		      iov_iter_count(iter));
973
974	pos = iocb->ki_pos;
975	end = (pos + iov_iter_count(iter) - 1) >> PAGE_CACHE_SHIFT;
976
977	mutex_lock(&inode->i_mutex);
978
979	result = nfs_sync_mapping(mapping);
980	if (result)
981		goto out_unlock;
982
983	if (mapping->nrpages) {
984		result = invalidate_inode_pages2_range(mapping,
985					pos >> PAGE_CACHE_SHIFT, end);
986		if (result)
987			goto out_unlock;
988	}
989
990	task_io_account_write(iov_iter_count(iter));
991
992	result = -ENOMEM;
993	dreq = nfs_direct_req_alloc();
994	if (!dreq)
995		goto out_unlock;
996
997	dreq->inode = inode;
998	dreq->bytes_left = iov_iter_count(iter);
999	dreq->io_start = pos;
1000	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1001	l_ctx = nfs_get_lock_context(dreq->ctx);
1002	if (IS_ERR(l_ctx)) {
1003		result = PTR_ERR(l_ctx);
1004		goto out_release;
1005	}
1006	dreq->l_ctx = l_ctx;
1007	if (!is_sync_kiocb(iocb))
1008		dreq->iocb = iocb;
1009
1010	result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1011
1012	if (mapping->nrpages) {
1013		invalidate_inode_pages2_range(mapping,
1014					      pos >> PAGE_CACHE_SHIFT, end);
1015	}
1016
1017	mutex_unlock(&inode->i_mutex);
1018
1019	if (!result) {
1020		result = nfs_direct_wait(dreq);
1021		if (result > 0) {
1022			struct inode *inode = mapping->host;
1023
1024			iocb->ki_pos = pos + result;
1025			spin_lock(&inode->i_lock);
1026			if (i_size_read(inode) < iocb->ki_pos)
1027				i_size_write(inode, iocb->ki_pos);
1028			spin_unlock(&inode->i_lock);
1029			generic_write_sync(file, pos, result);
1030		}
1031	}
1032	nfs_direct_req_release(dreq);
1033	return result;
1034
1035out_release:
1036	nfs_direct_req_release(dreq);
1037out_unlock:
1038	mutex_unlock(&inode->i_mutex);
1039	return result;
1040}
1041
1042/**
1043 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1044 *
1045 */
1046int __init nfs_init_directcache(void)
1047{
1048	nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1049						sizeof(struct nfs_direct_req),
1050						0, (SLAB_RECLAIM_ACCOUNT|
1051							SLAB_MEM_SPREAD),
1052						NULL);
1053	if (nfs_direct_cachep == NULL)
1054		return -ENOMEM;
1055
1056	return 0;
1057}
1058
1059/**
1060 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1061 *
1062 */
1063void nfs_destroy_directcache(void)
1064{
1065	kmem_cache_destroy(nfs_direct_cachep);
1066}
1067