direct.c 24.0 KB
Newer Older
L
Linus Torvalds 已提交
1 2 3 4 5 6 7 8 9
/*
 * linux/fs/nfs/direct.c
 *
 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
 *
 * High-performance uncached I/O for the Linux NFS client
 *
 * There are important applications whose performance or correctness
 * depends on uncached access to file data.  Database clusters
10
 * (multiple copies of the same instance running on separate hosts)
L
Linus Torvalds 已提交
11
 * implement their own cache coherency protocol that subsumes file
12 13 14
 * system cache protocols.  Applications that process datasets
 * considerably larger than the client's memory do not always benefit
 * from a local cache.  A streaming video server, for instance, has no
L
Linus Torvalds 已提交
15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
 * need to cache the contents of a file.
 *
 * When an application requests uncached I/O, all read and write requests
 * are made directly to the server; data stored or fetched via these
 * requests is not cached in the Linux page cache.  The client does not
 * correct unaligned requests from applications.  All requested bytes are
 * held on permanent storage before a direct write system call returns to
 * an application.
 *
 * Solaris implements an uncached I/O facility called directio() that
 * is used for backups and sequential I/O to very large files.  Solaris
 * also supports uncaching whole NFS partitions with "-o forcedirectio,"
 * an undocumented mount option.
 *
 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
 * help from Andrew Morton.
 *
 * 18 Dec 2001	Initial implementation for 2.4  --cel
 * 08 Jul 2002	Version for 2.4.19, with bug fixes --trondmy
 * 08 Jun 2003	Port to 2.5 APIs  --cel
 * 31 Mar 2004	Handle direct I/O without VFS support  --cel
 * 15 Sep 2004	Parallel async reads  --cel
37
 * 04 May 2005	support O_DIRECT with aio  --cel
L
Linus Torvalds 已提交
38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55
 *
 */

#include <linux/errno.h>
#include <linux/sched.h>
#include <linux/kernel.h>
#include <linux/file.h>
#include <linux/pagemap.h>
#include <linux/kref.h>

#include <linux/nfs_fs.h>
#include <linux/nfs_page.h>
#include <linux/sunrpc/clnt.h>

#include <asm/system.h>
#include <asm/uaccess.h>
#include <asm/atomic.h>

56
#include "internal.h"
C
Chuck Lever 已提交
57 58
#include "iostat.h"

L
Linus Torvalds 已提交
59 60
#define NFSDBG_FACILITY		NFSDBG_VFS

61
static struct kmem_cache *nfs_direct_cachep;
L
Linus Torvalds 已提交
62 63 64 65 66 67

/*
 * This represents a set of asynchronous requests that we're waiting on
 */
struct nfs_direct_req {
	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 77
	spinlock_t		lock;		/* protect completion state */
	ssize_t			count,		/* bytes actually processed */
L
Linus Torvalds 已提交
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 84 85 86 87
	struct nfs_write_data *	commit_data;	/* special write_data for commits */
	int			flags;
#define NFS_ODIRECT_DO_COMMIT		(1)	/* an unstable reply was received */
#define NFS_ODIRECT_RESCHED_WRITES	(2)	/* write verification failed */
	struct nfs_writeverf	verf;		/* unstable write verifier */
L
Linus Torvalds 已提交
88 89
};

90
static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
91 92 93 94 95 96 97 98 99 100 101 102
static const struct rpc_call_ops nfs_write_direct_ops;

static inline void get_dreq(struct nfs_direct_req *dreq)
{
	atomic_inc(&dreq->io_count);
}

static inline int put_dreq(struct nfs_direct_req *dreq)
{
	return atomic_dec_and_test(&dreq->io_count);
}

L
Linus Torvalds 已提交
103
/**
104 105 106 107 108 109 110 111 112 113 114
 * nfs_direct_IO - NFS address space operation for direct I/O
 * @rw: direction (read or write)
 * @iocb: target I/O control block
 * @iov: array of vectors that define I/O buffer
 * @pos: offset in file to begin the operation
 * @nr_segs: size of iovec array
 *
 * The presence of this routine in the address space ops vector means
 * the NFS client supports direct I/O.  However, we shunt off direct
 * read and write requests before the VFS gets them, so this method
 * should never be called.
L
Linus Torvalds 已提交
115
 */
116 117 118
ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
{
	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);
121 122 123 124

	return -EINVAL;
}

125
static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
126
{
127
	unsigned int npages;
128
	unsigned int i;
129 130 131 132 133

	if (count == 0)
		return;
	pages += (pgbase >> PAGE_SHIFT);
	npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
134 135
	for (i = 0; i < npages; i++) {
		struct page *page = pages[i];
136
		if (!PageCompound(page))
137
			set_page_dirty(page);
138
	}
139 140
}

141
static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
142
{
143
	unsigned int i;
144 145
	for (i = 0; i < npages; i++)
		page_cache_release(pages[i]);
146 147
}

148
static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
L
Linus Torvalds 已提交
149
{
150 151
	struct nfs_direct_req *dreq;

152
	dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
153 154 155 156
	if (!dreq)
		return NULL;

	kref_init(&dreq->kref);
157
	kref_get(&dreq->kref);
158
	init_completion(&dreq->completion);
159
	INIT_LIST_HEAD(&dreq->rewrite_list);
160
	dreq->iocb = NULL;
161
	dreq->ctx = NULL;
162
	spin_lock_init(&dreq->lock);
163
	atomic_set(&dreq->io_count, 0);
164 165
	dreq->count = 0;
	dreq->error = 0;
166
	dreq->flags = 0;
167 168

	return dreq;
L
Linus Torvalds 已提交
169 170
}

171
static void nfs_direct_req_free(struct kref *kref)
L
Linus Torvalds 已提交
172 173
{
	struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
174 175 176

	if (dreq->ctx != NULL)
		put_nfs_open_context(dreq->ctx);
L
Linus Torvalds 已提交
177 178 179
	kmem_cache_free(nfs_direct_cachep, dreq);
}

180 181 182 183 184
static void nfs_direct_req_release(struct nfs_direct_req *dreq)
{
	kref_put(&dreq->kref, nfs_direct_req_free);
}

185 186 187 188 189
/*
 * Collects and returns the final error value/byte-count.
 */
static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
{
190
	ssize_t result = -EIOCBQUEUED;
191 192 193 194 195

	/* Async requests don't wait here */
	if (dreq->iocb)
		goto out;

196
	result = wait_for_completion_interruptible(&dreq->completion);
197 198

	if (!result)
199
		result = dreq->error;
200
	if (!result)
201
		result = dreq->count;
202 203 204 205 206

out:
	return (ssize_t) result;
}

207
/*
208 209
 * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 * the iocb is still valid here if this is a synchronous request.
210 211 212 213
 */
static void nfs_direct_complete(struct nfs_direct_req *dreq)
{
	if (dreq->iocb) {
214
		long res = (long) dreq->error;
215
		if (!res)
216
			res = (long) dreq->count;
217
		aio_complete(dreq->iocb, res, 0);
218 219
	}
	complete_all(&dreq->completion);
220

221
	nfs_direct_req_release(dreq);
222 223
}

224
/*
225 226 227
 * We must hold a reference to all the pages in this direct read request
 * until the RPCs complete.  This could be long *after* we are woken up in
 * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
228
 */
T
Trond Myklebust 已提交
229
static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
L
Linus Torvalds 已提交
230
{
T
Trond Myklebust 已提交
231
	struct nfs_read_data *data = calldata;
L
Linus Torvalds 已提交
232 233
	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;

T
Trond Myklebust 已提交
234 235
	if (nfs_readpage_result(task, data) != 0)
		return;
236 237

	spin_lock(&dreq->lock);
238
	if (unlikely(task->tk_status < 0)) {
239
		dreq->error = task->tk_status;
240 241 242 243 244 245 246 247 248
		spin_unlock(&dreq->lock);
	} else {
		dreq->count += data->res.count;
		spin_unlock(&dreq->lock);
		nfs_direct_dirty_pages(data->pagevec,
				data->args.pgbase,
				data->res.count);
	}
	nfs_direct_release_pages(data->pagevec, data->npages);
249 250 251

	if (put_dreq(dreq))
		nfs_direct_complete(dreq);
L
Linus Torvalds 已提交
252 253
}

T
Trond Myklebust 已提交
254 255 256 257 258
static const struct rpc_call_ops nfs_read_direct_ops = {
	.rpc_call_done = nfs_direct_read_result,
	.rpc_release = nfs_readdata_release,
};

259
/*
260 261 262 263 264
 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 * bail and stop sending more reads.  Read length accounting is
 * handled automatically by nfs_direct_read_result().  Otherwise, if
 * no requests have been sent, just return an error.
L
Linus Torvalds 已提交
265
 */
266
static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
L
Linus Torvalds 已提交
267
{
268 269
	struct nfs_open_context *ctx = dreq->ctx;
	struct inode *inode = ctx->dentry->d_inode;
270
	size_t rsize = NFS_SERVER(inode)->rsize;
271 272 273 274 275
	unsigned int pgbase;
	int result;
	ssize_t started = 0;

	get_dreq(dreq);
L
Linus Torvalds 已提交
276 277

	do {
278
		struct nfs_read_data *data;
279
		size_t bytes;
L
Linus Torvalds 已提交
280

281 282 283
		pgbase = user_addr & ~PAGE_MASK;
		bytes = min(rsize,count);

284
		result = -ENOMEM;
285
		data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
286 287 288 289 290 291 292
		if (unlikely(!data))
			break;

		down_read(&current->mm->mmap_sem);
		result = get_user_pages(current, current->mm, user_addr,
					data->npages, 1, 0, data->pagevec, NULL);
		up_read(&current->mm->mmap_sem);
293 294 295 296 297 298
		if (result < 0) {
			nfs_readdata_release(data);
			break;
		}
		if ((unsigned)result < data->npages) {
			nfs_direct_release_pages(data->pagevec, result);
299 300 301 302 303
			nfs_readdata_release(data);
			break;
		}

		get_dreq(dreq);
304

305
		data->req = (struct nfs_page *) dreq;
L
Linus Torvalds 已提交
306 307 308 309
		data->inode = inode;
		data->cred = ctx->cred;
		data->args.fh = NFS_FH(inode);
		data->args.context = ctx;
310
		data->args.offset = pos;
L
Linus Torvalds 已提交
311
		data->args.pgbase = pgbase;
312
		data->args.pages = data->pagevec;
L
Linus Torvalds 已提交
313 314 315 316 317
		data->args.count = bytes;
		data->res.fattr = &data->fattr;
		data->res.eof = 0;
		data->res.count = bytes;

T
Trond Myklebust 已提交
318 319
		rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
				&nfs_read_direct_ops, data);
L
Linus Torvalds 已提交
320 321 322 323 324 325
		NFS_PROTO(inode)->read_setup(data);

		data->task.tk_cookie = (unsigned long) inode;

		rpc_execute(&data->task);

C
Chuck Lever 已提交
326 327
		dprintk("NFS: %5u initiated direct read call "
			"(req %s/%Ld, %zu bytes @ offset %Lu)\n",
L
Linus Torvalds 已提交
328 329 330 331 332 333
				data->task.tk_pid,
				inode->i_sb->s_id,
				(long long)NFS_FILEID(inode),
				bytes,
				(unsigned long long)data->args.offset);

334 335
		started += bytes;
		user_addr += bytes;
336
		pos += bytes;
337
		/* FIXME: Remove this unnecessary math from final patch */
L
Linus Torvalds 已提交
338 339
		pgbase += bytes;
		pgbase &= ~PAGE_MASK;
340
		BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
L
Linus Torvalds 已提交
341 342 343

		count -= bytes;
	} while (count != 0);
344 345 346 347 348 349 350

	if (put_dreq(dreq))
		nfs_direct_complete(dreq);

	if (started)
		return 0;
	return result < 0 ? (ssize_t) result : -EFAULT;
L
Linus Torvalds 已提交
351 352
}

353
static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
L
Linus Torvalds 已提交
354
{
355
	ssize_t result = 0;
L
Linus Torvalds 已提交
356
	sigset_t oldset;
357
	struct inode *inode = iocb->ki_filp->f_mapping->host;
L
Linus Torvalds 已提交
358 359 360
	struct rpc_clnt *clnt = NFS_CLIENT(inode);
	struct nfs_direct_req *dreq;

361
	dreq = nfs_direct_req_alloc();
L
Linus Torvalds 已提交
362 363 364
	if (!dreq)
		return -ENOMEM;

C
Chuck Lever 已提交
365
	dreq->inode = inode;
366
	dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
367 368
	if (!is_sync_kiocb(iocb))
		dreq->iocb = iocb;
L
Linus Torvalds 已提交
369

C
Chuck Lever 已提交
370
	nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
L
Linus Torvalds 已提交
371
	rpc_clnt_sigmask(clnt, &oldset);
372 373 374
	result = nfs_direct_read_schedule(dreq, user_addr, count, pos);
	if (!result)
		result = nfs_direct_wait(dreq);
L
Linus Torvalds 已提交
375
	rpc_clnt_sigunmask(clnt, &oldset);
376
	nfs_direct_req_release(dreq);
L
Linus Torvalds 已提交
377 378 379 380

	return result;
}

381
static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
L
Linus Torvalds 已提交
382
{
383 384
	while (!list_empty(&dreq->rewrite_list)) {
		struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
385
		list_del(&data->pages);
386
		nfs_direct_release_pages(data->pagevec, data->npages);
387 388 389
		nfs_writedata_release(data);
	}
}
L
Linus Torvalds 已提交
390

391 392 393
#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
{
394 395 396
	struct inode *inode = dreq->inode;
	struct list_head *p;
	struct nfs_write_data *data;
L
Linus Torvalds 已提交
397

398
	dreq->count = 0;
399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435
	get_dreq(dreq);

	list_for_each(p, &dreq->rewrite_list) {
		data = list_entry(p, struct nfs_write_data, pages);

		get_dreq(dreq);

		/*
		 * Reset data->res.
		 */
		nfs_fattr_init(&data->fattr);
		data->res.count = data->args.count;
		memset(&data->verf, 0, sizeof(data->verf));

		/*
		 * Reuse data->task; data->args should not have changed
		 * since the original request was sent.
		 */
		rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
				&nfs_write_direct_ops, data);
		NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);

		data->task.tk_priority = RPC_PRIORITY_NORMAL;
		data->task.tk_cookie = (unsigned long) inode;

		/*
		 * We're called via an RPC callback, so BKL is already held.
		 */
		rpc_execute(&data->task);

		dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
				data->task.tk_pid,
				inode->i_sb->s_id,
				(long long)NFS_FILEID(inode),
				data->args.count,
				(unsigned long long)data->args.offset);
	}
436

437 438
	if (put_dreq(dreq))
		nfs_direct_write_complete(dreq, inode);
439 440 441 442 443 444 445 446 447 448 449
}

static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
{
	struct nfs_write_data *data = calldata;
	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;

	/* Call the NFS version-specific code */
	if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
		return;
	if (unlikely(task->tk_status < 0)) {
450 451
		dprintk("NFS: %5u commit failed with error %d.\n",
				task->tk_pid, task->tk_status);
452
		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
453
	} else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
454 455
		dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
L
Linus Torvalds 已提交
456 457
	}

458 459
	dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
	nfs_direct_write_complete(dreq, data->inode);
L
Linus Torvalds 已提交
460 461
}

462 463 464 465 466 467
static const struct rpc_call_ops nfs_commit_direct_ops = {
	.rpc_call_done = nfs_direct_commit_result,
	.rpc_release = nfs_commit_release,
};

static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
L
Linus Torvalds 已提交
468
{
469
	struct nfs_write_data *data = dreq->commit_data;
L
Linus Torvalds 已提交
470

471
	data->inode = dreq->inode;
472
	data->cred = dreq->ctx->cred;
L
Linus Torvalds 已提交
473

474
	data->args.fh = NFS_FH(data->inode);
475 476
	data->args.offset = 0;
	data->args.count = 0;
477 478 479
	data->res.count = 0;
	data->res.fattr = &data->fattr;
	data->res.verf = &data->verf;
L
Linus Torvalds 已提交
480

481 482 483
	rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
				&nfs_commit_direct_ops, data);
	NFS_PROTO(data->inode)->commit_setup(data, 0);
L
Linus Torvalds 已提交
484

485 486 487 488
	data->task.tk_priority = RPC_PRIORITY_NORMAL;
	data->task.tk_cookie = (unsigned long)data->inode;
	/* Note: task.tk_ops->rpc_release will free dreq->commit_data */
	dreq->commit_data = NULL;
L
Linus Torvalds 已提交
489

490
	dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
L
Linus Torvalds 已提交
491

492 493
	rpc_execute(&data->task);
}
L
Linus Torvalds 已提交
494

495 496 497
static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
{
	int flags = dreq->flags;
L
Linus Torvalds 已提交
498

499 500 501 502
	dreq->flags = 0;
	switch (flags) {
		case NFS_ODIRECT_DO_COMMIT:
			nfs_direct_commit_schedule(dreq);
L
Linus Torvalds 已提交
503
			break;
504 505 506 507 508 509 510 511
		case NFS_ODIRECT_RESCHED_WRITES:
			nfs_direct_write_reschedule(dreq);
			break;
		default:
			nfs_end_data_update(inode);
			if (dreq->commit_data != NULL)
				nfs_commit_free(dreq->commit_data);
			nfs_direct_free_writedata(dreq);
512
			nfs_zap_mapping(inode, inode->i_mapping);
513 514 515
			nfs_direct_complete(dreq);
	}
}
L
Linus Torvalds 已提交
516

517 518
static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
{
519
	dreq->commit_data = nfs_commit_alloc();
520 521 522 523 524 525 526 527
	if (dreq->commit_data != NULL)
		dreq->commit_data->req = (struct nfs_page *) dreq;
}
#else
static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
{
	dreq->commit_data = NULL;
}
L
Linus Torvalds 已提交
528

529 530 531 532
static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
{
	nfs_end_data_update(inode);
	nfs_direct_free_writedata(dreq);
533
	nfs_zap_mapping(inode, inode->i_mapping);
534 535 536
	nfs_direct_complete(dreq);
}
#endif
L
Linus Torvalds 已提交
537

538
static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
L
Linus Torvalds 已提交
539
{
540 541 542 543 544 545 546
	struct nfs_write_data *data = calldata;
	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
	int status = task->tk_status;

	if (nfs_writeback_done(task, data) != 0)
		return;

547
	spin_lock(&dreq->lock);
L
Linus Torvalds 已提交
548

549 550
	if (unlikely(dreq->error != 0))
		goto out_unlock;
551
	if (unlikely(status < 0)) {
552 553
		/* An error has occured, so we should not commit */
		dreq->flags = 0;
554 555 556 557
		dreq->error = status;
	}

	dreq->count += data->res.count;
L
Linus Torvalds 已提交
558

559 560 561 562 563
	if (data->res.verf->committed != NFS_FILE_SYNC) {
		switch (dreq->flags) {
			case 0:
				memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
				dreq->flags = NFS_ODIRECT_DO_COMMIT;
L
Linus Torvalds 已提交
564
				break;
565 566 567 568 569
			case NFS_ODIRECT_DO_COMMIT:
				if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
					dprintk("NFS: %5u write verify failed\n", task->tk_pid);
					dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
				}
L
Linus Torvalds 已提交
570 571
		}
	}
572
out_unlock:
573
	spin_unlock(&dreq->lock);
L
Linus Torvalds 已提交
574 575
}

576 577 578
/*
 * NB: Return the value of the first error return code.  Subsequent
 *     errors after the first one are ignored.
L
Linus Torvalds 已提交
579
 */
580
static void nfs_direct_write_release(void *calldata)
L
Linus Torvalds 已提交
581
{
582 583
	struct nfs_write_data *data = calldata;
	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
L
Linus Torvalds 已提交
584

585 586
	if (put_dreq(dreq))
		nfs_direct_write_complete(dreq, data->inode);
587 588 589 590
}

static const struct rpc_call_ops nfs_write_direct_ops = {
	.rpc_call_done = nfs_direct_write_result,
591
	.rpc_release = nfs_direct_write_release,
592 593 594
};

/*
595 596 597 598 599
 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 * bail and stop sending more writes.  Write length accounting is
 * handled automatically by nfs_direct_write_result().  Otherwise, if
 * no requests have been sent, just return an error.
600
 */
601
static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
602
{
603 604
	struct nfs_open_context *ctx = dreq->ctx;
	struct inode *inode = ctx->dentry->d_inode;
605
	size_t wsize = NFS_SERVER(inode)->wsize;
606 607 608
	unsigned int pgbase;
	int result;
	ssize_t started = 0;
609

610 611
	get_dreq(dreq);

L
Linus Torvalds 已提交
612
	do {
613
		struct nfs_write_data *data;
614 615
		size_t bytes;

616 617 618
		pgbase = user_addr & ~PAGE_MASK;
		bytes = min(wsize,count);

619
		result = -ENOMEM;
620
		data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
621 622 623 624 625 626 627
		if (unlikely(!data))
			break;

		down_read(&current->mm->mmap_sem);
		result = get_user_pages(current, current->mm, user_addr,
					data->npages, 0, 0, data->pagevec, NULL);
		up_read(&current->mm->mmap_sem);
628 629 630 631 632 633
		if (result < 0) {
			nfs_writedata_release(data);
			break;
		}
		if ((unsigned)result < data->npages) {
			nfs_direct_release_pages(data->pagevec, result);
634 635 636 637 638 639
			nfs_writedata_release(data);
			break;
		}

		get_dreq(dreq);

640
		list_move_tail(&data->pages, &dreq->rewrite_list);
641

642
		data->req = (struct nfs_page *) dreq;
643 644 645 646
		data->inode = inode;
		data->cred = ctx->cred;
		data->args.fh = NFS_FH(inode);
		data->args.context = ctx;
647
		data->args.offset = pos;
648
		data->args.pgbase = pgbase;
649
		data->args.pages = data->pagevec;
650 651 652
		data->args.count = bytes;
		data->res.fattr = &data->fattr;
		data->res.count = bytes;
653
		data->res.verf = &data->verf;
654 655 656

		rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
				&nfs_write_direct_ops, data);
657
		NFS_PROTO(inode)->write_setup(data, sync);
L
Linus Torvalds 已提交
658

659 660
		data->task.tk_priority = RPC_PRIORITY_NORMAL;
		data->task.tk_cookie = (unsigned long) inode;
L
Linus Torvalds 已提交
661

662
		rpc_execute(&data->task);
L
Linus Torvalds 已提交
663

C
Chuck Lever 已提交
664 665
		dprintk("NFS: %5u initiated direct write call "
			"(req %s/%Ld, %zu bytes @ offset %Lu)\n",
666 667 668 669 670
				data->task.tk_pid,
				inode->i_sb->s_id,
				(long long)NFS_FILEID(inode),
				bytes,
				(unsigned long long)data->args.offset);
L
Linus Torvalds 已提交
671

672 673
		started += bytes;
		user_addr += bytes;
674
		pos += bytes;
675 676

		/* FIXME: Remove this useless math from the final patch */
677 678
		pgbase += bytes;
		pgbase &= ~PAGE_MASK;
679
		BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
L
Linus Torvalds 已提交
680

681 682
		count -= bytes;
	} while (count != 0);
683 684 685 686 687 688 689

	if (put_dreq(dreq))
		nfs_direct_write_complete(dreq, inode);

	if (started)
		return 0;
	return result < 0 ? (ssize_t) result : -EFAULT;
690
}
L
Linus Torvalds 已提交
691

692
static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
693
{
694
	ssize_t result = 0;
695
	sigset_t oldset;
696
	struct inode *inode = iocb->ki_filp->f_mapping->host;
697 698
	struct rpc_clnt *clnt = NFS_CLIENT(inode);
	struct nfs_direct_req *dreq;
699 700
	size_t wsize = NFS_SERVER(inode)->wsize;
	int sync = 0;
L
Linus Torvalds 已提交
701

702
	dreq = nfs_direct_req_alloc();
703 704
	if (!dreq)
		return -ENOMEM;
705 706
	nfs_alloc_commit_data(dreq);

707 708
	if (dreq->commit_data == NULL || count < wsize)
		sync = FLUSH_STABLE;
L
Linus Torvalds 已提交
709

710
	dreq->inode = inode;
711
	dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
712 713
	if (!is_sync_kiocb(iocb))
		dreq->iocb = iocb;
L
Linus Torvalds 已提交
714

715 716
	nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);

717
	nfs_begin_data_update(inode);
L
Linus Torvalds 已提交
718

719
	rpc_clnt_sigmask(clnt, &oldset);
720 721 722
	result = nfs_direct_write_schedule(dreq, user_addr, count, pos, sync);
	if (!result)
		result = nfs_direct_wait(dreq);
723
	rpc_clnt_sigunmask(clnt, &oldset);
724
	nfs_direct_req_release(dreq);
L
Linus Torvalds 已提交
725 726 727 728 729 730 731

	return result;
}

/**
 * nfs_file_direct_read - file direct read operation for NFS files
 * @iocb: target I/O control block
732 733
 * @iov: vector of user buffers into which to read data
 * @nr_segs: size of iov vector
734
 * @pos: byte offset in file where reading starts
L
Linus Torvalds 已提交
735 736 737 738 739 740
 *
 * We use this function for direct reads instead of calling
 * generic_file_aio_read() in order to avoid gfar's check to see if
 * the request starts before the end of the file.  For that check
 * to work, we must generate a GETATTR before each direct read, and
 * even then there is a window between the GETATTR and the subsequent
741
 * READ where the file size could change.  Our preference is simply
L
Linus Torvalds 已提交
742 743
 * to do all reads the application wants, and the server will take
 * care of managing the end of file boundary.
744
 *
L
Linus Torvalds 已提交
745 746 747 748 749
 * This function also eliminates unnecessarily updating the file's
 * atime locally, as the NFS server sets the file's atime, and this
 * client must read the updated atime from the server back into its
 * cache.
 */
750 751
ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
				unsigned long nr_segs, loff_t pos)
L
Linus Torvalds 已提交
752 753 754 755
{
	ssize_t retval = -EINVAL;
	struct file *file = iocb->ki_filp;
	struct address_space *mapping = file->f_mapping;
756 757 758
	/* XXX: temporary */
	const char __user *buf = iov[0].iov_base;
	size_t count = iov[0].iov_len;
L
Linus Torvalds 已提交
759

760
	dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
761 762
		file->f_path.dentry->d_parent->d_name.name,
		file->f_path.dentry->d_name.name,
763
		(unsigned long) count, (long long) pos);
L
Linus Torvalds 已提交
764

765
	if (nr_segs != 1)
L
Linus Torvalds 已提交
766
		goto out;
C
Chuck Lever 已提交
767

L
Linus Torvalds 已提交
768
	retval = -EFAULT;
769
	if (!access_ok(VERIFY_WRITE, buf, count))
L
Linus Torvalds 已提交
770 771 772 773 774
		goto out;
	retval = 0;
	if (!count)
		goto out;

T
Trond Myklebust 已提交
775 776 777
	retval = nfs_sync_mapping(mapping);
	if (retval)
		goto out;
L
Linus Torvalds 已提交
778

779
	retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos);
L
Linus Torvalds 已提交
780
	if (retval > 0)
781
		iocb->ki_pos = pos + retval;
L
Linus Torvalds 已提交
782 783 784 785 786 787 788 789

out:
	return retval;
}

/**
 * nfs_file_direct_write - file direct write operation for NFS files
 * @iocb: target I/O control block
790 791
 * @iov: vector of user buffers from which to write data
 * @nr_segs: size of iov vector
792
 * @pos: byte offset in file where writing starts
L
Linus Torvalds 已提交
793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811
 *
 * We use this function for direct writes instead of calling
 * generic_file_aio_write() in order to avoid taking the inode
 * semaphore and updating the i_size.  The NFS server will set
 * the new i_size and this client must read the updated size
 * back into its cache.  We let the server do generic write
 * parameter checking and report problems.
 *
 * We also avoid an unnecessary invocation of generic_osync_inode(),
 * as it is fairly meaningless to sync the metadata of an NFS file.
 *
 * We eliminate local atime updates, see direct read above.
 *
 * We avoid unnecessary page cache invalidations for normal cached
 * readers of this file.
 *
 * Note that O_APPEND is not supported for NFS direct writes, as there
 * is no atomic O_APPEND write facility in the NFS protocol.
 */
812 813
ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
				unsigned long nr_segs, loff_t pos)
L
Linus Torvalds 已提交
814
{
C
Chuck Lever 已提交
815
	ssize_t retval = -EINVAL;
L
Linus Torvalds 已提交
816 817
	struct file *file = iocb->ki_filp;
	struct address_space *mapping = file->f_mapping;
818 819 820
	/* XXX: temporary */
	const char __user *buf = iov[0].iov_base;
	size_t count = iov[0].iov_len;
L
Linus Torvalds 已提交
821

C
Chuck Lever 已提交
822
	dprintk("nfs: direct write(%s/%s, %lu@%Ld)\n",
823 824
		file->f_path.dentry->d_parent->d_name.name,
		file->f_path.dentry->d_name.name,
825
		(unsigned long) count, (long long) pos);
L
Linus Torvalds 已提交
826

827
	if (nr_segs != 1)
C
Chuck Lever 已提交
828
		goto out;
829

830 831
	retval = generic_write_checks(file, &pos, &count, 0);
	if (retval)
L
Linus Torvalds 已提交
832
		goto out;
833 834 835

	retval = -EINVAL;
	if ((ssize_t) count < 0)
L
Linus Torvalds 已提交
836 837 838 839
		goto out;
	retval = 0;
	if (!count)
		goto out;
840 841

	retval = -EFAULT;
842
	if (!access_ok(VERIFY_READ, buf, count))
843
		goto out;
L
Linus Torvalds 已提交
844

T
Trond Myklebust 已提交
845 846 847
	retval = nfs_sync_mapping(mapping);
	if (retval)
		goto out;
L
Linus Torvalds 已提交
848

849
	retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos);
850

L
Linus Torvalds 已提交
851
	if (retval > 0)
852
		iocb->ki_pos = pos + retval;
L
Linus Torvalds 已提交
853 854 855 856 857

out:
	return retval;
}

858 859 860 861
/**
 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 *
 */
D
David Howells 已提交
862
int __init nfs_init_directcache(void)
L
Linus Torvalds 已提交
863 864 865
{
	nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
						sizeof(struct nfs_direct_req),
866 867
						0, (SLAB_RECLAIM_ACCOUNT|
							SLAB_MEM_SPREAD),
L
Linus Torvalds 已提交
868 869 870 871 872 873 874
						NULL, NULL);
	if (nfs_direct_cachep == NULL)
		return -ENOMEM;

	return 0;
}

875
/**
D
David Howells 已提交
876
 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
877 878
 *
 */
879
void nfs_destroy_directcache(void)
L
Linus Torvalds 已提交
880
{
881
	kmem_cache_destroy(nfs_direct_cachep);
L
Linus Torvalds 已提交
882
}