xref: /linux/fs/nfs/direct.c (revision c537b994505099b7197e7d3125b942ecbcc51eb6)
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/smp_lock.h>
45 #include <linux/file.h>
46 #include <linux/pagemap.h>
47 #include <linux/kref.h>
48 
49 #include <linux/nfs_fs.h>
50 #include <linux/nfs_page.h>
51 #include <linux/sunrpc/clnt.h>
52 
53 #include <asm/system.h>
54 #include <asm/uaccess.h>
55 #include <asm/atomic.h>
56 
57 #include "iostat.h"
58 
59 #define NFSDBG_FACILITY		NFSDBG_VFS
60 
61 static struct kmem_cache *nfs_direct_cachep;
62 
63 /*
64  * This represents a set of asynchronous requests that we're waiting on
65  */
66 struct nfs_direct_req {
67 	struct kref		kref;		/* release manager */
68 
69 	/* I/O parameters */
70 	struct nfs_open_context	*ctx;		/* file open context info */
71 	struct kiocb *		iocb;		/* controlling i/o request */
72 	struct inode *		inode;		/* target file of i/o */
73 
74 	/* completion state */
75 	atomic_t		io_count;	/* i/os we're waiting for */
76 	spinlock_t		lock;		/* protect completion state */
77 	ssize_t			count,		/* bytes actually processed */
78 				error;		/* any reported error */
79 	struct completion	completion;	/* wait for i/o completion */
80 
81 	/* commit state */
82 	struct list_head	rewrite_list;	/* saved nfs_write_data structs */
83 	struct nfs_write_data *	commit_data;	/* special write_data for commits */
84 	int			flags;
85 #define NFS_ODIRECT_DO_COMMIT		(1)	/* an unstable reply was received */
86 #define NFS_ODIRECT_RESCHED_WRITES	(2)	/* write verification failed */
87 	struct nfs_writeverf	verf;		/* unstable write verifier */
88 };
89 
90 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
91 static const struct rpc_call_ops nfs_write_direct_ops;
92 
93 static inline void get_dreq(struct nfs_direct_req *dreq)
94 {
95 	atomic_inc(&dreq->io_count);
96 }
97 
98 static inline int put_dreq(struct nfs_direct_req *dreq)
99 {
100 	return atomic_dec_and_test(&dreq->io_count);
101 }
102 
103 /**
104  * nfs_direct_IO - NFS address space operation for direct I/O
105  * @rw: direction (read or write)
106  * @iocb: target I/O control block
107  * @iov: array of vectors that define I/O buffer
108  * @pos: offset in file to begin the operation
109  * @nr_segs: size of iovec array
110  *
111  * The presence of this routine in the address space ops vector means
112  * the NFS client supports direct I/O.  However, we shunt off direct
113  * read and write requests before the VFS gets them, so this method
114  * should never be called.
115  */
116 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
117 {
118 	dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
119 			iocb->ki_filp->f_path.dentry->d_name.name,
120 			(long long) pos, nr_segs);
121 
122 	return -EINVAL;
123 }
124 
125 static void nfs_direct_dirty_pages(struct page **pages, int npages)
126 {
127 	int i;
128 	for (i = 0; i < npages; i++) {
129 		struct page *page = pages[i];
130 		if (!PageCompound(page))
131 			set_page_dirty_lock(page);
132 	}
133 }
134 
135 static void nfs_direct_release_pages(struct page **pages, int npages)
136 {
137 	int i;
138 	for (i = 0; i < npages; i++)
139 		page_cache_release(pages[i]);
140 }
141 
142 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
143 {
144 	struct nfs_direct_req *dreq;
145 
146 	dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
147 	if (!dreq)
148 		return NULL;
149 
150 	kref_init(&dreq->kref);
151 	kref_get(&dreq->kref);
152 	init_completion(&dreq->completion);
153 	INIT_LIST_HEAD(&dreq->rewrite_list);
154 	dreq->iocb = NULL;
155 	dreq->ctx = NULL;
156 	spin_lock_init(&dreq->lock);
157 	atomic_set(&dreq->io_count, 0);
158 	dreq->count = 0;
159 	dreq->error = 0;
160 	dreq->flags = 0;
161 
162 	return dreq;
163 }
164 
165 static void nfs_direct_req_release(struct kref *kref)
166 {
167 	struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
168 
169 	if (dreq->ctx != NULL)
170 		put_nfs_open_context(dreq->ctx);
171 	kmem_cache_free(nfs_direct_cachep, dreq);
172 }
173 
174 /*
175  * Collects and returns the final error value/byte-count.
176  */
177 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
178 {
179 	ssize_t result = -EIOCBQUEUED;
180 
181 	/* Async requests don't wait here */
182 	if (dreq->iocb)
183 		goto out;
184 
185 	result = wait_for_completion_interruptible(&dreq->completion);
186 
187 	if (!result)
188 		result = dreq->error;
189 	if (!result)
190 		result = dreq->count;
191 
192 out:
193 	kref_put(&dreq->kref, nfs_direct_req_release);
194 	return (ssize_t) result;
195 }
196 
197 /*
198  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
199  * the iocb is still valid here if this is a synchronous request.
200  */
201 static void nfs_direct_complete(struct nfs_direct_req *dreq)
202 {
203 	if (dreq->iocb) {
204 		long res = (long) dreq->error;
205 		if (!res)
206 			res = (long) dreq->count;
207 		aio_complete(dreq->iocb, res, 0);
208 	}
209 	complete_all(&dreq->completion);
210 
211 	kref_put(&dreq->kref, nfs_direct_req_release);
212 }
213 
214 /*
215  * We must hold a reference to all the pages in this direct read request
216  * until the RPCs complete.  This could be long *after* we are woken up in
217  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
218  */
219 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
220 {
221 	struct nfs_read_data *data = calldata;
222 	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
223 
224 	if (nfs_readpage_result(task, data) != 0)
225 		return;
226 
227 	nfs_direct_dirty_pages(data->pagevec, data->npages);
228 	nfs_direct_release_pages(data->pagevec, data->npages);
229 
230 	spin_lock(&dreq->lock);
231 
232 	if (likely(task->tk_status >= 0))
233 		dreq->count += data->res.count;
234 	else
235 		dreq->error = task->tk_status;
236 
237 	spin_unlock(&dreq->lock);
238 
239 	if (put_dreq(dreq))
240 		nfs_direct_complete(dreq);
241 }
242 
243 static const struct rpc_call_ops nfs_read_direct_ops = {
244 	.rpc_call_done = nfs_direct_read_result,
245 	.rpc_release = nfs_readdata_release,
246 };
247 
248 /*
249  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
250  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
251  * bail and stop sending more reads.  Read length accounting is
252  * handled automatically by nfs_direct_read_result().  Otherwise, if
253  * no requests have been sent, just return an error.
254  */
255 static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
256 {
257 	struct nfs_open_context *ctx = dreq->ctx;
258 	struct inode *inode = ctx->dentry->d_inode;
259 	size_t rsize = NFS_SERVER(inode)->rsize;
260 	unsigned int pgbase;
261 	int result;
262 	ssize_t started = 0;
263 
264 	get_dreq(dreq);
265 
266 	do {
267 		struct nfs_read_data *data;
268 		size_t bytes;
269 
270 		pgbase = user_addr & ~PAGE_MASK;
271 		bytes = min(rsize,count);
272 
273 		result = -ENOMEM;
274 		data = nfs_readdata_alloc(pgbase + bytes);
275 		if (unlikely(!data))
276 			break;
277 
278 		down_read(&current->mm->mmap_sem);
279 		result = get_user_pages(current, current->mm, user_addr,
280 					data->npages, 1, 0, data->pagevec, NULL);
281 		up_read(&current->mm->mmap_sem);
282 		if (unlikely(result < data->npages)) {
283 			if (result > 0)
284 				nfs_direct_release_pages(data->pagevec, result);
285 			nfs_readdata_release(data);
286 			break;
287 		}
288 
289 		get_dreq(dreq);
290 
291 		data->req = (struct nfs_page *) dreq;
292 		data->inode = inode;
293 		data->cred = ctx->cred;
294 		data->args.fh = NFS_FH(inode);
295 		data->args.context = ctx;
296 		data->args.offset = pos;
297 		data->args.pgbase = pgbase;
298 		data->args.pages = data->pagevec;
299 		data->args.count = bytes;
300 		data->res.fattr = &data->fattr;
301 		data->res.eof = 0;
302 		data->res.count = bytes;
303 
304 		rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
305 				&nfs_read_direct_ops, data);
306 		NFS_PROTO(inode)->read_setup(data);
307 
308 		data->task.tk_cookie = (unsigned long) inode;
309 
310 		rpc_execute(&data->task);
311 
312 		dprintk("NFS: %5u initiated direct read call "
313 			"(req %s/%Ld, %zu bytes @ offset %Lu)\n",
314 				data->task.tk_pid,
315 				inode->i_sb->s_id,
316 				(long long)NFS_FILEID(inode),
317 				bytes,
318 				(unsigned long long)data->args.offset);
319 
320 		started += bytes;
321 		user_addr += bytes;
322 		pos += bytes;
323 		/* FIXME: Remove this unnecessary math from final patch */
324 		pgbase += bytes;
325 		pgbase &= ~PAGE_MASK;
326 		BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
327 
328 		count -= bytes;
329 	} while (count != 0);
330 
331 	if (put_dreq(dreq))
332 		nfs_direct_complete(dreq);
333 
334 	if (started)
335 		return 0;
336 	return result < 0 ? (ssize_t) result : -EFAULT;
337 }
338 
339 static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
340 {
341 	ssize_t result = 0;
342 	sigset_t oldset;
343 	struct inode *inode = iocb->ki_filp->f_mapping->host;
344 	struct rpc_clnt *clnt = NFS_CLIENT(inode);
345 	struct nfs_direct_req *dreq;
346 
347 	dreq = nfs_direct_req_alloc();
348 	if (!dreq)
349 		return -ENOMEM;
350 
351 	dreq->inode = inode;
352 	dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
353 	if (!is_sync_kiocb(iocb))
354 		dreq->iocb = iocb;
355 
356 	nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
357 	rpc_clnt_sigmask(clnt, &oldset);
358 	result = nfs_direct_read_schedule(dreq, user_addr, count, pos);
359 	if (!result)
360 		result = nfs_direct_wait(dreq);
361 	rpc_clnt_sigunmask(clnt, &oldset);
362 
363 	return result;
364 }
365 
366 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
367 {
368 	while (!list_empty(&dreq->rewrite_list)) {
369 		struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
370 		list_del(&data->pages);
371 		nfs_direct_release_pages(data->pagevec, data->npages);
372 		nfs_writedata_release(data);
373 	}
374 }
375 
376 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
377 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
378 {
379 	struct inode *inode = dreq->inode;
380 	struct list_head *p;
381 	struct nfs_write_data *data;
382 
383 	dreq->count = 0;
384 	get_dreq(dreq);
385 
386 	list_for_each(p, &dreq->rewrite_list) {
387 		data = list_entry(p, struct nfs_write_data, pages);
388 
389 		get_dreq(dreq);
390 
391 		/*
392 		 * Reset data->res.
393 		 */
394 		nfs_fattr_init(&data->fattr);
395 		data->res.count = data->args.count;
396 		memset(&data->verf, 0, sizeof(data->verf));
397 
398 		/*
399 		 * Reuse data->task; data->args should not have changed
400 		 * since the original request was sent.
401 		 */
402 		rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
403 				&nfs_write_direct_ops, data);
404 		NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
405 
406 		data->task.tk_priority = RPC_PRIORITY_NORMAL;
407 		data->task.tk_cookie = (unsigned long) inode;
408 
409 		/*
410 		 * We're called via an RPC callback, so BKL is already held.
411 		 */
412 		rpc_execute(&data->task);
413 
414 		dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
415 				data->task.tk_pid,
416 				inode->i_sb->s_id,
417 				(long long)NFS_FILEID(inode),
418 				data->args.count,
419 				(unsigned long long)data->args.offset);
420 	}
421 
422 	if (put_dreq(dreq))
423 		nfs_direct_write_complete(dreq, inode);
424 }
425 
426 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
427 {
428 	struct nfs_write_data *data = calldata;
429 	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
430 
431 	/* Call the NFS version-specific code */
432 	if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
433 		return;
434 	if (unlikely(task->tk_status < 0)) {
435 		dreq->error = task->tk_status;
436 		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
437 	}
438 	if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
439 		dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
440 		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
441 	}
442 
443 	dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
444 	nfs_direct_write_complete(dreq, data->inode);
445 }
446 
447 static const struct rpc_call_ops nfs_commit_direct_ops = {
448 	.rpc_call_done = nfs_direct_commit_result,
449 	.rpc_release = nfs_commit_release,
450 };
451 
452 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
453 {
454 	struct nfs_write_data *data = dreq->commit_data;
455 
456 	data->inode = dreq->inode;
457 	data->cred = dreq->ctx->cred;
458 
459 	data->args.fh = NFS_FH(data->inode);
460 	data->args.offset = 0;
461 	data->args.count = 0;
462 	data->res.count = 0;
463 	data->res.fattr = &data->fattr;
464 	data->res.verf = &data->verf;
465 
466 	rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
467 				&nfs_commit_direct_ops, data);
468 	NFS_PROTO(data->inode)->commit_setup(data, 0);
469 
470 	data->task.tk_priority = RPC_PRIORITY_NORMAL;
471 	data->task.tk_cookie = (unsigned long)data->inode;
472 	/* Note: task.tk_ops->rpc_release will free dreq->commit_data */
473 	dreq->commit_data = NULL;
474 
475 	dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
476 
477 	rpc_execute(&data->task);
478 }
479 
480 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
481 {
482 	int flags = dreq->flags;
483 
484 	dreq->flags = 0;
485 	switch (flags) {
486 		case NFS_ODIRECT_DO_COMMIT:
487 			nfs_direct_commit_schedule(dreq);
488 			break;
489 		case NFS_ODIRECT_RESCHED_WRITES:
490 			nfs_direct_write_reschedule(dreq);
491 			break;
492 		default:
493 			nfs_end_data_update(inode);
494 			if (dreq->commit_data != NULL)
495 				nfs_commit_free(dreq->commit_data);
496 			nfs_direct_free_writedata(dreq);
497 			nfs_zap_mapping(inode, inode->i_mapping);
498 			nfs_direct_complete(dreq);
499 	}
500 }
501 
502 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
503 {
504 	dreq->commit_data = nfs_commit_alloc();
505 	if (dreq->commit_data != NULL)
506 		dreq->commit_data->req = (struct nfs_page *) dreq;
507 }
508 #else
509 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
510 {
511 	dreq->commit_data = NULL;
512 }
513 
514 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
515 {
516 	nfs_end_data_update(inode);
517 	nfs_direct_free_writedata(dreq);
518 	nfs_zap_mapping(inode, inode->i_mapping);
519 	nfs_direct_complete(dreq);
520 }
521 #endif
522 
523 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
524 {
525 	struct nfs_write_data *data = calldata;
526 	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
527 	int status = task->tk_status;
528 
529 	if (nfs_writeback_done(task, data) != 0)
530 		return;
531 
532 	spin_lock(&dreq->lock);
533 
534 	if (unlikely(status < 0)) {
535 		dreq->error = status;
536 		goto out_unlock;
537 	}
538 
539 	dreq->count += data->res.count;
540 
541 	if (data->res.verf->committed != NFS_FILE_SYNC) {
542 		switch (dreq->flags) {
543 			case 0:
544 				memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
545 				dreq->flags = NFS_ODIRECT_DO_COMMIT;
546 				break;
547 			case NFS_ODIRECT_DO_COMMIT:
548 				if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
549 					dprintk("NFS: %5u write verify failed\n", task->tk_pid);
550 					dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
551 				}
552 		}
553 	}
554 out_unlock:
555 	spin_unlock(&dreq->lock);
556 }
557 
558 /*
559  * NB: Return the value of the first error return code.  Subsequent
560  *     errors after the first one are ignored.
561  */
562 static void nfs_direct_write_release(void *calldata)
563 {
564 	struct nfs_write_data *data = calldata;
565 	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
566 
567 	if (put_dreq(dreq))
568 		nfs_direct_write_complete(dreq, data->inode);
569 }
570 
571 static const struct rpc_call_ops nfs_write_direct_ops = {
572 	.rpc_call_done = nfs_direct_write_result,
573 	.rpc_release = nfs_direct_write_release,
574 };
575 
576 /*
577  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
578  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
579  * bail and stop sending more writes.  Write length accounting is
580  * handled automatically by nfs_direct_write_result().  Otherwise, if
581  * no requests have been sent, just return an error.
582  */
583 static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
584 {
585 	struct nfs_open_context *ctx = dreq->ctx;
586 	struct inode *inode = ctx->dentry->d_inode;
587 	size_t wsize = NFS_SERVER(inode)->wsize;
588 	unsigned int pgbase;
589 	int result;
590 	ssize_t started = 0;
591 
592 	get_dreq(dreq);
593 
594 	do {
595 		struct nfs_write_data *data;
596 		size_t bytes;
597 
598 		pgbase = user_addr & ~PAGE_MASK;
599 		bytes = min(wsize,count);
600 
601 		result = -ENOMEM;
602 		data = nfs_writedata_alloc(pgbase + bytes);
603 		if (unlikely(!data))
604 			break;
605 
606 		down_read(&current->mm->mmap_sem);
607 		result = get_user_pages(current, current->mm, user_addr,
608 					data->npages, 0, 0, data->pagevec, NULL);
609 		up_read(&current->mm->mmap_sem);
610 		if (unlikely(result < data->npages)) {
611 			if (result > 0)
612 				nfs_direct_release_pages(data->pagevec, result);
613 			nfs_writedata_release(data);
614 			break;
615 		}
616 
617 		get_dreq(dreq);
618 
619 		list_move_tail(&data->pages, &dreq->rewrite_list);
620 
621 		data->req = (struct nfs_page *) dreq;
622 		data->inode = inode;
623 		data->cred = ctx->cred;
624 		data->args.fh = NFS_FH(inode);
625 		data->args.context = ctx;
626 		data->args.offset = pos;
627 		data->args.pgbase = pgbase;
628 		data->args.pages = data->pagevec;
629 		data->args.count = bytes;
630 		data->res.fattr = &data->fattr;
631 		data->res.count = bytes;
632 		data->res.verf = &data->verf;
633 
634 		rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
635 				&nfs_write_direct_ops, data);
636 		NFS_PROTO(inode)->write_setup(data, sync);
637 
638 		data->task.tk_priority = RPC_PRIORITY_NORMAL;
639 		data->task.tk_cookie = (unsigned long) inode;
640 
641 		rpc_execute(&data->task);
642 
643 		dprintk("NFS: %5u initiated direct write call "
644 			"(req %s/%Ld, %zu bytes @ offset %Lu)\n",
645 				data->task.tk_pid,
646 				inode->i_sb->s_id,
647 				(long long)NFS_FILEID(inode),
648 				bytes,
649 				(unsigned long long)data->args.offset);
650 
651 		started += bytes;
652 		user_addr += bytes;
653 		pos += bytes;
654 
655 		/* FIXME: Remove this useless math from the final patch */
656 		pgbase += bytes;
657 		pgbase &= ~PAGE_MASK;
658 		BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
659 
660 		count -= bytes;
661 	} while (count != 0);
662 
663 	if (put_dreq(dreq))
664 		nfs_direct_write_complete(dreq, inode);
665 
666 	if (started)
667 		return 0;
668 	return result < 0 ? (ssize_t) result : -EFAULT;
669 }
670 
671 static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
672 {
673 	ssize_t result = 0;
674 	sigset_t oldset;
675 	struct inode *inode = iocb->ki_filp->f_mapping->host;
676 	struct rpc_clnt *clnt = NFS_CLIENT(inode);
677 	struct nfs_direct_req *dreq;
678 	size_t wsize = NFS_SERVER(inode)->wsize;
679 	int sync = 0;
680 
681 	dreq = nfs_direct_req_alloc();
682 	if (!dreq)
683 		return -ENOMEM;
684 	nfs_alloc_commit_data(dreq);
685 
686 	if (dreq->commit_data == NULL || count < wsize)
687 		sync = FLUSH_STABLE;
688 
689 	dreq->inode = inode;
690 	dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
691 	if (!is_sync_kiocb(iocb))
692 		dreq->iocb = iocb;
693 
694 	nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
695 
696 	nfs_begin_data_update(inode);
697 
698 	rpc_clnt_sigmask(clnt, &oldset);
699 	result = nfs_direct_write_schedule(dreq, user_addr, count, pos, sync);
700 	if (!result)
701 		result = nfs_direct_wait(dreq);
702 	rpc_clnt_sigunmask(clnt, &oldset);
703 
704 	return result;
705 }
706 
707 /**
708  * nfs_file_direct_read - file direct read operation for NFS files
709  * @iocb: target I/O control block
710  * @iov: vector of user buffers into which to read data
711  * @nr_segs: size of iov vector
712  * @pos: byte offset in file where reading starts
713  *
714  * We use this function for direct reads instead of calling
715  * generic_file_aio_read() in order to avoid gfar's check to see if
716  * the request starts before the end of the file.  For that check
717  * to work, we must generate a GETATTR before each direct read, and
718  * even then there is a window between the GETATTR and the subsequent
719  * READ where the file size could change.  Our preference is simply
720  * to do all reads the application wants, and the server will take
721  * care of managing the end of file boundary.
722  *
723  * This function also eliminates unnecessarily updating the file's
724  * atime locally, as the NFS server sets the file's atime, and this
725  * client must read the updated atime from the server back into its
726  * cache.
727  */
728 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
729 				unsigned long nr_segs, loff_t pos)
730 {
731 	ssize_t retval = -EINVAL;
732 	struct file *file = iocb->ki_filp;
733 	struct address_space *mapping = file->f_mapping;
734 	/* XXX: temporary */
735 	const char __user *buf = iov[0].iov_base;
736 	size_t count = iov[0].iov_len;
737 
738 	dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
739 		file->f_path.dentry->d_parent->d_name.name,
740 		file->f_path.dentry->d_name.name,
741 		(unsigned long) count, (long long) pos);
742 
743 	if (nr_segs != 1)
744 		return -EINVAL;
745 
746 	if (count < 0)
747 		goto out;
748 	retval = -EFAULT;
749 	if (!access_ok(VERIFY_WRITE, buf, count))
750 		goto out;
751 	retval = 0;
752 	if (!count)
753 		goto out;
754 
755 	retval = nfs_sync_mapping(mapping);
756 	if (retval)
757 		goto out;
758 
759 	retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos);
760 	if (retval > 0)
761 		iocb->ki_pos = pos + retval;
762 
763 out:
764 	return retval;
765 }
766 
767 /**
768  * nfs_file_direct_write - file direct write operation for NFS files
769  * @iocb: target I/O control block
770  * @iov: vector of user buffers from which to write data
771  * @nr_segs: size of iov vector
772  * @pos: byte offset in file where writing starts
773  *
774  * We use this function for direct writes instead of calling
775  * generic_file_aio_write() in order to avoid taking the inode
776  * semaphore and updating the i_size.  The NFS server will set
777  * the new i_size and this client must read the updated size
778  * back into its cache.  We let the server do generic write
779  * parameter checking and report problems.
780  *
781  * We also avoid an unnecessary invocation of generic_osync_inode(),
782  * as it is fairly meaningless to sync the metadata of an NFS file.
783  *
784  * We eliminate local atime updates, see direct read above.
785  *
786  * We avoid unnecessary page cache invalidations for normal cached
787  * readers of this file.
788  *
789  * Note that O_APPEND is not supported for NFS direct writes, as there
790  * is no atomic O_APPEND write facility in the NFS protocol.
791  */
792 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
793 				unsigned long nr_segs, loff_t pos)
794 {
795 	ssize_t retval;
796 	struct file *file = iocb->ki_filp;
797 	struct address_space *mapping = file->f_mapping;
798 	/* XXX: temporary */
799 	const char __user *buf = iov[0].iov_base;
800 	size_t count = iov[0].iov_len;
801 
802 	dprintk("nfs: direct write(%s/%s, %lu@%Ld)\n",
803 		file->f_path.dentry->d_parent->d_name.name,
804 		file->f_path.dentry->d_name.name,
805 		(unsigned long) count, (long long) pos);
806 
807 	if (nr_segs != 1)
808 		return -EINVAL;
809 
810 	retval = generic_write_checks(file, &pos, &count, 0);
811 	if (retval)
812 		goto out;
813 
814 	retval = -EINVAL;
815 	if ((ssize_t) count < 0)
816 		goto out;
817 	retval = 0;
818 	if (!count)
819 		goto out;
820 
821 	retval = -EFAULT;
822 	if (!access_ok(VERIFY_READ, buf, count))
823 		goto out;
824 
825 	retval = nfs_sync_mapping(mapping);
826 	if (retval)
827 		goto out;
828 
829 	retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos);
830 
831 	if (retval > 0)
832 		iocb->ki_pos = pos + retval;
833 
834 out:
835 	return retval;
836 }
837 
838 /**
839  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
840  *
841  */
842 int __init nfs_init_directcache(void)
843 {
844 	nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
845 						sizeof(struct nfs_direct_req),
846 						0, (SLAB_RECLAIM_ACCOUNT|
847 							SLAB_MEM_SPREAD),
848 						NULL, NULL);
849 	if (nfs_direct_cachep == NULL)
850 		return -ENOMEM;
851 
852 	return 0;
853 }
854 
855 /**
856  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
857  *
858  */
859 void nfs_destroy_directcache(void)
860 {
861 	kmem_cache_destroy(nfs_direct_cachep);
862 }
863