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 
50 #include <linux/nfs_fs.h>
51 #include <linux/nfs_page.h>
52 #include <linux/sunrpc/clnt.h>
53 
54 #include <asm/uaccess.h>
55 #include <linux/atomic.h>
56 
57 #include "internal.h"
58 #include "iostat.h"
59 
60 #define NFSDBG_FACILITY		NFSDBG_VFS
61 
62 static struct kmem_cache *nfs_direct_cachep;
63 
64 /*
65  * This represents a set of asynchronous requests that we're waiting on
66  */
67 struct nfs_direct_req {
68 	struct kref		kref;		/* release manager */
69 
70 	/* I/O parameters */
71 	struct nfs_open_context	*ctx;		/* file open context info */
72 	struct nfs_lock_context *l_ctx;		/* Lock context info */
73 	struct kiocb *		iocb;		/* controlling i/o request */
74 	struct inode *		inode;		/* target file of i/o */
75 
76 	/* completion state */
77 	atomic_t		io_count;	/* i/os we're waiting for */
78 	spinlock_t		lock;		/* protect completion state */
79 	ssize_t			count,		/* bytes actually processed */
80 				error;		/* any reported error */
81 	struct completion	completion;	/* wait for i/o completion */
82 
83 	/* commit state */
84 	struct list_head	rewrite_list;	/* saved nfs_write_data structs */
85 	struct nfs_write_data *	commit_data;	/* special write_data for commits */
86 	int			flags;
87 #define NFS_ODIRECT_DO_COMMIT		(1)	/* an unstable reply was received */
88 #define NFS_ODIRECT_RESCHED_WRITES	(2)	/* write verification failed */
89 	struct nfs_writeverf	verf;		/* unstable write verifier */
90 };
91 
92 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
93 static const struct rpc_call_ops nfs_write_direct_ops;
94 
get_dreq(struct nfs_direct_req * dreq)95 static inline void get_dreq(struct nfs_direct_req *dreq)
96 {
97 	atomic_inc(&dreq->io_count);
98 }
99 
put_dreq(struct nfs_direct_req * dreq)100 static inline int put_dreq(struct nfs_direct_req *dreq)
101 {
102 	return atomic_dec_and_test(&dreq->io_count);
103 }
104 
105 /**
106  * nfs_direct_IO - NFS address space operation for direct I/O
107  * @rw: direction (read or write)
108  * @iocb: target I/O control block
109  * @iov: array of vectors that define I/O buffer
110  * @pos: offset in file to begin the operation
111  * @nr_segs: size of iovec array
112  *
113  * The presence of this routine in the address space ops vector means
114  * the NFS client supports direct I/O.  However, we shunt off direct
115  * read and write requests before the VFS gets them, so this method
116  * should never be called.
117  */
nfs_direct_IO(int rw,struct kiocb * iocb,const struct iovec * iov,loff_t pos,unsigned long nr_segs)118 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
119 {
120 	dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
121 			iocb->ki_filp->f_path.dentry->d_name.name,
122 			(long long) pos, nr_segs);
123 
124 	return -EINVAL;
125 }
126 
nfs_direct_dirty_pages(struct page ** pages,unsigned int pgbase,size_t count)127 static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
128 {
129 	unsigned int npages;
130 	unsigned int i;
131 
132 	if (count == 0)
133 		return;
134 	pages += (pgbase >> PAGE_SHIFT);
135 	npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
136 	for (i = 0; i < npages; i++) {
137 		struct page *page = pages[i];
138 		if (!PageCompound(page))
139 			set_page_dirty(page);
140 	}
141 }
142 
nfs_direct_release_pages(struct page ** pages,unsigned int npages)143 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
144 {
145 	unsigned int i;
146 	for (i = 0; i < npages; i++)
147 		page_cache_release(pages[i]);
148 }
149 
nfs_direct_req_alloc(void)150 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
151 {
152 	struct nfs_direct_req *dreq;
153 
154 	dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
155 	if (!dreq)
156 		return NULL;
157 
158 	kref_init(&dreq->kref);
159 	kref_get(&dreq->kref);
160 	init_completion(&dreq->completion);
161 	INIT_LIST_HEAD(&dreq->rewrite_list);
162 	dreq->iocb = NULL;
163 	dreq->ctx = NULL;
164 	dreq->l_ctx = NULL;
165 	spin_lock_init(&dreq->lock);
166 	atomic_set(&dreq->io_count, 0);
167 	dreq->count = 0;
168 	dreq->error = 0;
169 	dreq->flags = 0;
170 
171 	return dreq;
172 }
173 
nfs_direct_req_free(struct kref * kref)174 static void nfs_direct_req_free(struct kref *kref)
175 {
176 	struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
177 
178 	if (dreq->l_ctx != NULL)
179 		nfs_put_lock_context(dreq->l_ctx);
180 	if (dreq->ctx != NULL)
181 		put_nfs_open_context(dreq->ctx);
182 	kmem_cache_free(nfs_direct_cachep, dreq);
183 }
184 
nfs_direct_req_release(struct nfs_direct_req * dreq)185 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
186 {
187 	kref_put(&dreq->kref, nfs_direct_req_free);
188 }
189 
190 /*
191  * Collects and returns the final error value/byte-count.
192  */
nfs_direct_wait(struct nfs_direct_req * dreq)193 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
194 {
195 	ssize_t result = -EIOCBQUEUED;
196 
197 	/* Async requests don't wait here */
198 	if (dreq->iocb)
199 		goto out;
200 
201 	result = wait_for_completion_killable(&dreq->completion);
202 
203 	if (!result)
204 		result = dreq->error;
205 	if (!result)
206 		result = dreq->count;
207 
208 out:
209 	return (ssize_t) result;
210 }
211 
212 /*
213  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
214  * the iocb is still valid here if this is a synchronous request.
215  */
nfs_direct_complete(struct nfs_direct_req * dreq)216 static void nfs_direct_complete(struct nfs_direct_req *dreq)
217 {
218 	if (dreq->iocb) {
219 		long res = (long) dreq->error;
220 		if (!res)
221 			res = (long) dreq->count;
222 		aio_complete(dreq->iocb, res, 0);
223 	}
224 	complete_all(&dreq->completion);
225 
226 	nfs_direct_req_release(dreq);
227 }
228 
229 /*
230  * We must hold a reference to all the pages in this direct read request
231  * until the RPCs complete.  This could be long *after* we are woken up in
232  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
233  */
nfs_direct_read_result(struct rpc_task * task,void * calldata)234 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
235 {
236 	struct nfs_read_data *data = calldata;
237 
238 	nfs_readpage_result(task, data);
239 }
240 
nfs_direct_read_release(void * calldata)241 static void nfs_direct_read_release(void *calldata)
242 {
243 
244 	struct nfs_read_data *data = calldata;
245 	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
246 	int status = data->task.tk_status;
247 
248 	spin_lock(&dreq->lock);
249 	if (unlikely(status < 0)) {
250 		dreq->error = status;
251 		spin_unlock(&dreq->lock);
252 	} else {
253 		dreq->count += data->res.count;
254 		spin_unlock(&dreq->lock);
255 		nfs_direct_dirty_pages(data->pagevec,
256 				data->args.pgbase,
257 				data->res.count);
258 	}
259 	nfs_direct_release_pages(data->pagevec, data->npages);
260 
261 	if (put_dreq(dreq))
262 		nfs_direct_complete(dreq);
263 	nfs_readdata_free(data);
264 }
265 
266 static const struct rpc_call_ops nfs_read_direct_ops = {
267 	.rpc_call_prepare = nfs_read_prepare,
268 	.rpc_call_done = nfs_direct_read_result,
269 	.rpc_release = nfs_direct_read_release,
270 };
271 
272 /*
273  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
274  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
275  * bail and stop sending more reads.  Read length accounting is
276  * handled automatically by nfs_direct_read_result().  Otherwise, if
277  * no requests have been sent, just return an error.
278  */
nfs_direct_read_schedule_segment(struct nfs_direct_req * dreq,const struct iovec * iov,loff_t pos)279 static ssize_t nfs_direct_read_schedule_segment(struct nfs_direct_req *dreq,
280 						const struct iovec *iov,
281 						loff_t pos)
282 {
283 	struct nfs_open_context *ctx = dreq->ctx;
284 	struct inode *inode = ctx->dentry->d_inode;
285 	unsigned long user_addr = (unsigned long)iov->iov_base;
286 	size_t count = iov->iov_len;
287 	size_t rsize = NFS_SERVER(inode)->rsize;
288 	struct rpc_task *task;
289 	struct rpc_message msg = {
290 		.rpc_cred = ctx->cred,
291 	};
292 	struct rpc_task_setup task_setup_data = {
293 		.rpc_client = NFS_CLIENT(inode),
294 		.rpc_message = &msg,
295 		.callback_ops = &nfs_read_direct_ops,
296 		.workqueue = nfsiod_workqueue,
297 		.flags = RPC_TASK_ASYNC,
298 	};
299 	unsigned int pgbase;
300 	int result;
301 	ssize_t started = 0;
302 
303 	do {
304 		struct nfs_read_data *data;
305 		size_t bytes;
306 
307 		pgbase = user_addr & ~PAGE_MASK;
308 		bytes = min(rsize,count);
309 
310 		result = -ENOMEM;
311 		data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
312 		if (unlikely(!data))
313 			break;
314 
315 		down_read(&current->mm->mmap_sem);
316 		result = get_user_pages(current, current->mm, user_addr,
317 					data->npages, 1, 0, data->pagevec, NULL);
318 		up_read(&current->mm->mmap_sem);
319 		if (result < 0) {
320 			nfs_readdata_free(data);
321 			break;
322 		}
323 		if ((unsigned)result < data->npages) {
324 			bytes = result * PAGE_SIZE;
325 			if (bytes <= pgbase) {
326 				nfs_direct_release_pages(data->pagevec, result);
327 				nfs_readdata_free(data);
328 				break;
329 			}
330 			bytes -= pgbase;
331 			data->npages = result;
332 		}
333 
334 		get_dreq(dreq);
335 
336 		data->req = (struct nfs_page *) dreq;
337 		data->inode = inode;
338 		data->cred = msg.rpc_cred;
339 		data->args.fh = NFS_FH(inode);
340 		data->args.context = ctx;
341 		data->args.lock_context = dreq->l_ctx;
342 		data->args.offset = pos;
343 		data->args.pgbase = pgbase;
344 		data->args.pages = data->pagevec;
345 		data->args.count = bytes;
346 		data->res.fattr = &data->fattr;
347 		data->res.eof = 0;
348 		data->res.count = bytes;
349 		nfs_fattr_init(&data->fattr);
350 		msg.rpc_argp = &data->args;
351 		msg.rpc_resp = &data->res;
352 
353 		task_setup_data.task = &data->task;
354 		task_setup_data.callback_data = data;
355 		NFS_PROTO(inode)->read_setup(data, &msg);
356 
357 		task = rpc_run_task(&task_setup_data);
358 		if (IS_ERR(task))
359 			break;
360 		rpc_put_task(task);
361 
362 		dprintk("NFS: %5u initiated direct read call "
363 			"(req %s/%Ld, %zu bytes @ offset %Lu)\n",
364 				data->task.tk_pid,
365 				inode->i_sb->s_id,
366 				(long long)NFS_FILEID(inode),
367 				bytes,
368 				(unsigned long long)data->args.offset);
369 
370 		started += bytes;
371 		user_addr += bytes;
372 		pos += bytes;
373 		/* FIXME: Remove this unnecessary math from final patch */
374 		pgbase += bytes;
375 		pgbase &= ~PAGE_MASK;
376 		BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
377 
378 		count -= bytes;
379 	} while (count != 0);
380 
381 	if (started)
382 		return started;
383 	return result < 0 ? (ssize_t) result : -EFAULT;
384 }
385 
nfs_direct_read_schedule_iovec(struct nfs_direct_req * dreq,const struct iovec * iov,unsigned long nr_segs,loff_t pos)386 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
387 					      const struct iovec *iov,
388 					      unsigned long nr_segs,
389 					      loff_t pos)
390 {
391 	ssize_t result = -EINVAL;
392 	size_t requested_bytes = 0;
393 	unsigned long seg;
394 
395 	get_dreq(dreq);
396 
397 	for (seg = 0; seg < nr_segs; seg++) {
398 		const struct iovec *vec = &iov[seg];
399 		result = nfs_direct_read_schedule_segment(dreq, vec, pos);
400 		if (result < 0)
401 			break;
402 		requested_bytes += result;
403 		if ((size_t)result < vec->iov_len)
404 			break;
405 		pos += vec->iov_len;
406 	}
407 
408 	/*
409 	 * If no bytes were started, return the error, and let the
410 	 * generic layer handle the completion.
411 	 */
412 	if (requested_bytes == 0) {
413 		nfs_direct_req_release(dreq);
414 		return result < 0 ? result : -EIO;
415 	}
416 
417 	if (put_dreq(dreq))
418 		nfs_direct_complete(dreq);
419 	return 0;
420 }
421 
nfs_direct_read(struct kiocb * iocb,const struct iovec * iov,unsigned long nr_segs,loff_t pos)422 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
423 			       unsigned long nr_segs, loff_t pos)
424 {
425 	ssize_t result = -ENOMEM;
426 	struct inode *inode = iocb->ki_filp->f_mapping->host;
427 	struct nfs_direct_req *dreq;
428 
429 	dreq = nfs_direct_req_alloc();
430 	if (dreq == NULL)
431 		goto out;
432 
433 	dreq->inode = inode;
434 	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
435 	dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
436 	if (dreq->l_ctx == NULL)
437 		goto out_release;
438 	if (!is_sync_kiocb(iocb))
439 		dreq->iocb = iocb;
440 
441 	result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
442 	if (!result)
443 		result = nfs_direct_wait(dreq);
444 out_release:
445 	nfs_direct_req_release(dreq);
446 out:
447 	return result;
448 }
449 
nfs_direct_free_writedata(struct nfs_direct_req * dreq)450 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
451 {
452 	while (!list_empty(&dreq->rewrite_list)) {
453 		struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
454 		list_del(&data->pages);
455 		nfs_direct_release_pages(data->pagevec, data->npages);
456 		nfs_writedata_free(data);
457 	}
458 }
459 
460 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
nfs_direct_write_reschedule(struct nfs_direct_req * dreq)461 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
462 {
463 	struct inode *inode = dreq->inode;
464 	struct list_head *p;
465 	struct nfs_write_data *data;
466 	struct rpc_task *task;
467 	struct rpc_message msg = {
468 		.rpc_cred = dreq->ctx->cred,
469 	};
470 	struct rpc_task_setup task_setup_data = {
471 		.rpc_client = NFS_CLIENT(inode),
472 		.rpc_message = &msg,
473 		.callback_ops = &nfs_write_direct_ops,
474 		.workqueue = nfsiod_workqueue,
475 		.flags = RPC_TASK_ASYNC,
476 	};
477 
478 	dreq->count = 0;
479 	get_dreq(dreq);
480 
481 	list_for_each(p, &dreq->rewrite_list) {
482 		data = list_entry(p, struct nfs_write_data, pages);
483 
484 		get_dreq(dreq);
485 
486 		/* Use stable writes */
487 		data->args.stable = NFS_FILE_SYNC;
488 
489 		/*
490 		 * Reset data->res.
491 		 */
492 		nfs_fattr_init(&data->fattr);
493 		data->res.count = data->args.count;
494 		memset(&data->verf, 0, sizeof(data->verf));
495 
496 		/*
497 		 * Reuse data->task; data->args should not have changed
498 		 * since the original request was sent.
499 		 */
500 		task_setup_data.task = &data->task;
501 		task_setup_data.callback_data = data;
502 		msg.rpc_argp = &data->args;
503 		msg.rpc_resp = &data->res;
504 		NFS_PROTO(inode)->write_setup(data, &msg);
505 
506 		/*
507 		 * We're called via an RPC callback, so BKL is already held.
508 		 */
509 		task = rpc_run_task(&task_setup_data);
510 		if (!IS_ERR(task))
511 			rpc_put_task(task);
512 
513 		dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
514 				data->task.tk_pid,
515 				inode->i_sb->s_id,
516 				(long long)NFS_FILEID(inode),
517 				data->args.count,
518 				(unsigned long long)data->args.offset);
519 	}
520 
521 	if (put_dreq(dreq))
522 		nfs_direct_write_complete(dreq, inode);
523 }
524 
nfs_direct_commit_result(struct rpc_task * task,void * calldata)525 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
526 {
527 	struct nfs_write_data *data = calldata;
528 
529 	/* Call the NFS version-specific code */
530 	NFS_PROTO(data->inode)->commit_done(task, data);
531 }
532 
nfs_direct_commit_release(void * calldata)533 static void nfs_direct_commit_release(void *calldata)
534 {
535 	struct nfs_write_data *data = calldata;
536 	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
537 	int status = data->task.tk_status;
538 
539 	if (status < 0) {
540 		dprintk("NFS: %5u commit failed with error %d.\n",
541 				data->task.tk_pid, status);
542 		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
543 	} else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
544 		dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
545 		dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
546 	}
547 
548 	dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
549 	nfs_direct_write_complete(dreq, data->inode);
550 	nfs_commit_free(data);
551 }
552 
553 static const struct rpc_call_ops nfs_commit_direct_ops = {
554 	.rpc_call_prepare = nfs_write_prepare,
555 	.rpc_call_done = nfs_direct_commit_result,
556 	.rpc_release = nfs_direct_commit_release,
557 };
558 
nfs_direct_commit_schedule(struct nfs_direct_req * dreq)559 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
560 {
561 	struct nfs_write_data *data = dreq->commit_data;
562 	struct rpc_task *task;
563 	struct rpc_message msg = {
564 		.rpc_argp = &data->args,
565 		.rpc_resp = &data->res,
566 		.rpc_cred = dreq->ctx->cred,
567 	};
568 	struct rpc_task_setup task_setup_data = {
569 		.task = &data->task,
570 		.rpc_client = NFS_CLIENT(dreq->inode),
571 		.rpc_message = &msg,
572 		.callback_ops = &nfs_commit_direct_ops,
573 		.callback_data = data,
574 		.workqueue = nfsiod_workqueue,
575 		.flags = RPC_TASK_ASYNC,
576 	};
577 
578 	data->inode = dreq->inode;
579 	data->cred = msg.rpc_cred;
580 
581 	data->args.fh = NFS_FH(data->inode);
582 	data->args.offset = 0;
583 	data->args.count = 0;
584 	data->args.context = dreq->ctx;
585 	data->args.lock_context = dreq->l_ctx;
586 	data->res.count = 0;
587 	data->res.fattr = &data->fattr;
588 	data->res.verf = &data->verf;
589 	nfs_fattr_init(&data->fattr);
590 
591 	NFS_PROTO(data->inode)->commit_setup(data, &msg);
592 
593 	/* Note: task.tk_ops->rpc_release will free dreq->commit_data */
594 	dreq->commit_data = NULL;
595 
596 	dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
597 
598 	task = rpc_run_task(&task_setup_data);
599 	if (!IS_ERR(task))
600 		rpc_put_task(task);
601 }
602 
nfs_direct_write_complete(struct nfs_direct_req * dreq,struct inode * inode)603 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
604 {
605 	int flags = dreq->flags;
606 
607 	dreq->flags = 0;
608 	switch (flags) {
609 		case NFS_ODIRECT_DO_COMMIT:
610 			nfs_direct_commit_schedule(dreq);
611 			break;
612 		case NFS_ODIRECT_RESCHED_WRITES:
613 			nfs_direct_write_reschedule(dreq);
614 			break;
615 		default:
616 			if (dreq->commit_data != NULL)
617 				nfs_commit_free(dreq->commit_data);
618 			nfs_direct_free_writedata(dreq);
619 			nfs_zap_mapping(inode, inode->i_mapping);
620 			nfs_direct_complete(dreq);
621 	}
622 }
623 
nfs_alloc_commit_data(struct nfs_direct_req * dreq)624 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
625 {
626 	dreq->commit_data = nfs_commitdata_alloc();
627 	if (dreq->commit_data != NULL)
628 		dreq->commit_data->req = (struct nfs_page *) dreq;
629 }
630 #else
nfs_alloc_commit_data(struct nfs_direct_req * dreq)631 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
632 {
633 	dreq->commit_data = NULL;
634 }
635 
nfs_direct_write_complete(struct nfs_direct_req * dreq,struct inode * inode)636 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
637 {
638 	nfs_direct_free_writedata(dreq);
639 	nfs_zap_mapping(inode, inode->i_mapping);
640 	nfs_direct_complete(dreq);
641 }
642 #endif
643 
nfs_direct_write_result(struct rpc_task * task,void * calldata)644 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
645 {
646 	struct nfs_write_data *data = calldata;
647 
648 	nfs_writeback_done(task, data);
649 }
650 
651 /*
652  * NB: Return the value of the first error return code.  Subsequent
653  *     errors after the first one are ignored.
654  */
nfs_direct_write_release(void * calldata)655 static void nfs_direct_write_release(void *calldata)
656 {
657 	struct nfs_write_data *data = calldata;
658 	struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
659 	int status = data->task.tk_status;
660 
661 	spin_lock(&dreq->lock);
662 
663 	if (unlikely(status < 0)) {
664 		/* An error has occurred, so we should not commit */
665 		dreq->flags = 0;
666 		dreq->error = status;
667 	}
668 	if (unlikely(dreq->error != 0))
669 		goto out_unlock;
670 
671 	dreq->count += data->res.count;
672 
673 	if (data->res.verf->committed != NFS_FILE_SYNC) {
674 		switch (dreq->flags) {
675 			case 0:
676 				memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
677 				dreq->flags = NFS_ODIRECT_DO_COMMIT;
678 				break;
679 			case NFS_ODIRECT_DO_COMMIT:
680 				if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
681 					dprintk("NFS: %5u write verify failed\n", data->task.tk_pid);
682 					dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
683 				}
684 		}
685 	}
686 out_unlock:
687 	spin_unlock(&dreq->lock);
688 
689 	if (put_dreq(dreq))
690 		nfs_direct_write_complete(dreq, data->inode);
691 }
692 
693 static const struct rpc_call_ops nfs_write_direct_ops = {
694 	.rpc_call_prepare = nfs_write_prepare,
695 	.rpc_call_done = nfs_direct_write_result,
696 	.rpc_release = nfs_direct_write_release,
697 };
698 
699 /*
700  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
701  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
702  * bail and stop sending more writes.  Write length accounting is
703  * handled automatically by nfs_direct_write_result().  Otherwise, if
704  * no requests have been sent, just return an error.
705  */
nfs_direct_write_schedule_segment(struct nfs_direct_req * dreq,const struct iovec * iov,loff_t pos,int sync)706 static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
707 						 const struct iovec *iov,
708 						 loff_t pos, int sync)
709 {
710 	struct nfs_open_context *ctx = dreq->ctx;
711 	struct inode *inode = ctx->dentry->d_inode;
712 	unsigned long user_addr = (unsigned long)iov->iov_base;
713 	size_t count = iov->iov_len;
714 	struct rpc_task *task;
715 	struct rpc_message msg = {
716 		.rpc_cred = ctx->cred,
717 	};
718 	struct rpc_task_setup task_setup_data = {
719 		.rpc_client = NFS_CLIENT(inode),
720 		.rpc_message = &msg,
721 		.callback_ops = &nfs_write_direct_ops,
722 		.workqueue = nfsiod_workqueue,
723 		.flags = RPC_TASK_ASYNC,
724 	};
725 	size_t wsize = NFS_SERVER(inode)->wsize;
726 	unsigned int pgbase;
727 	int result;
728 	ssize_t started = 0;
729 
730 	do {
731 		struct nfs_write_data *data;
732 		size_t bytes;
733 
734 		pgbase = user_addr & ~PAGE_MASK;
735 		bytes = min(wsize,count);
736 
737 		result = -ENOMEM;
738 		data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
739 		if (unlikely(!data))
740 			break;
741 
742 		down_read(&current->mm->mmap_sem);
743 		result = get_user_pages(current, current->mm, user_addr,
744 					data->npages, 0, 0, data->pagevec, NULL);
745 		up_read(&current->mm->mmap_sem);
746 		if (result < 0) {
747 			nfs_writedata_free(data);
748 			break;
749 		}
750 		if ((unsigned)result < data->npages) {
751 			bytes = result * PAGE_SIZE;
752 			if (bytes <= pgbase) {
753 				nfs_direct_release_pages(data->pagevec, result);
754 				nfs_writedata_free(data);
755 				break;
756 			}
757 			bytes -= pgbase;
758 			data->npages = result;
759 		}
760 
761 		get_dreq(dreq);
762 
763 		list_move_tail(&data->pages, &dreq->rewrite_list);
764 
765 		data->req = (struct nfs_page *) dreq;
766 		data->inode = inode;
767 		data->cred = msg.rpc_cred;
768 		data->args.fh = NFS_FH(inode);
769 		data->args.context = ctx;
770 		data->args.lock_context = dreq->l_ctx;
771 		data->args.offset = pos;
772 		data->args.pgbase = pgbase;
773 		data->args.pages = data->pagevec;
774 		data->args.count = bytes;
775 		data->args.stable = sync;
776 		data->res.fattr = &data->fattr;
777 		data->res.count = bytes;
778 		data->res.verf = &data->verf;
779 		nfs_fattr_init(&data->fattr);
780 
781 		task_setup_data.task = &data->task;
782 		task_setup_data.callback_data = data;
783 		msg.rpc_argp = &data->args;
784 		msg.rpc_resp = &data->res;
785 		NFS_PROTO(inode)->write_setup(data, &msg);
786 
787 		task = rpc_run_task(&task_setup_data);
788 		if (IS_ERR(task))
789 			break;
790 		rpc_put_task(task);
791 
792 		dprintk("NFS: %5u initiated direct write call "
793 			"(req %s/%Ld, %zu bytes @ offset %Lu)\n",
794 				data->task.tk_pid,
795 				inode->i_sb->s_id,
796 				(long long)NFS_FILEID(inode),
797 				bytes,
798 				(unsigned long long)data->args.offset);
799 
800 		started += bytes;
801 		user_addr += bytes;
802 		pos += bytes;
803 
804 		/* FIXME: Remove this useless math from the final patch */
805 		pgbase += bytes;
806 		pgbase &= ~PAGE_MASK;
807 		BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
808 
809 		count -= bytes;
810 	} while (count != 0);
811 
812 	if (started)
813 		return started;
814 	return result < 0 ? (ssize_t) result : -EFAULT;
815 }
816 
nfs_direct_write_schedule_iovec(struct nfs_direct_req * dreq,const struct iovec * iov,unsigned long nr_segs,loff_t pos,int sync)817 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
818 					       const struct iovec *iov,
819 					       unsigned long nr_segs,
820 					       loff_t pos, int sync)
821 {
822 	ssize_t result = 0;
823 	size_t requested_bytes = 0;
824 	unsigned long seg;
825 
826 	get_dreq(dreq);
827 
828 	for (seg = 0; seg < nr_segs; seg++) {
829 		const struct iovec *vec = &iov[seg];
830 		result = nfs_direct_write_schedule_segment(dreq, vec,
831 							   pos, sync);
832 		if (result < 0)
833 			break;
834 		requested_bytes += result;
835 		if ((size_t)result < vec->iov_len)
836 			break;
837 		pos += vec->iov_len;
838 	}
839 
840 	/*
841 	 * If no bytes were started, return the error, and let the
842 	 * generic layer handle the completion.
843 	 */
844 	if (requested_bytes == 0) {
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, dreq->inode);
851 	return 0;
852 }
853 
nfs_direct_write(struct kiocb * iocb,const struct iovec * iov,unsigned long nr_segs,loff_t pos,size_t count)854 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
855 				unsigned long nr_segs, loff_t pos,
856 				size_t count)
857 {
858 	ssize_t result = -ENOMEM;
859 	struct inode *inode = iocb->ki_filp->f_mapping->host;
860 	struct nfs_direct_req *dreq;
861 	size_t wsize = NFS_SERVER(inode)->wsize;
862 	int sync = NFS_UNSTABLE;
863 
864 	dreq = nfs_direct_req_alloc();
865 	if (!dreq)
866 		goto out;
867 	nfs_alloc_commit_data(dreq);
868 
869 	if (dreq->commit_data == NULL || count <= wsize)
870 		sync = NFS_FILE_SYNC;
871 
872 	dreq->inode = inode;
873 	dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
874 	dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
875 	if (dreq->l_ctx == NULL)
876 		goto out_release;
877 	if (!is_sync_kiocb(iocb))
878 		dreq->iocb = iocb;
879 
880 	result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
881 	if (!result)
882 		result = nfs_direct_wait(dreq);
883 out_release:
884 	nfs_direct_req_release(dreq);
885 out:
886 	return result;
887 }
888 
889 /**
890  * nfs_file_direct_read - file direct read operation for NFS files
891  * @iocb: target I/O control block
892  * @iov: vector of user buffers into which to read data
893  * @nr_segs: size of iov vector
894  * @pos: byte offset in file where reading starts
895  *
896  * We use this function for direct reads instead of calling
897  * generic_file_aio_read() in order to avoid gfar's check to see if
898  * the request starts before the end of the file.  For that check
899  * to work, we must generate a GETATTR before each direct read, and
900  * even then there is a window between the GETATTR and the subsequent
901  * READ where the file size could change.  Our preference is simply
902  * to do all reads the application wants, and the server will take
903  * care of managing the end of file boundary.
904  *
905  * This function also eliminates unnecessarily updating the file's
906  * atime locally, as the NFS server sets the file's atime, and this
907  * client must read the updated atime from the server back into its
908  * cache.
909  */
nfs_file_direct_read(struct kiocb * iocb,const struct iovec * iov,unsigned long nr_segs,loff_t pos)910 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
911 				unsigned long nr_segs, loff_t pos)
912 {
913 	ssize_t retval = -EINVAL;
914 	struct file *file = iocb->ki_filp;
915 	struct address_space *mapping = file->f_mapping;
916 	size_t count;
917 
918 	count = iov_length(iov, nr_segs);
919 	nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
920 
921 	dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
922 		file->f_path.dentry->d_parent->d_name.name,
923 		file->f_path.dentry->d_name.name,
924 		count, (long long) pos);
925 
926 	retval = 0;
927 	if (!count)
928 		goto out;
929 
930 	retval = nfs_sync_mapping(mapping);
931 	if (retval)
932 		goto out;
933 
934 	task_io_account_read(count);
935 
936 	retval = nfs_direct_read(iocb, iov, nr_segs, pos);
937 	if (retval > 0)
938 		iocb->ki_pos = pos + retval;
939 
940 out:
941 	return retval;
942 }
943 
944 /**
945  * nfs_file_direct_write - file direct write operation for NFS files
946  * @iocb: target I/O control block
947  * @iov: vector of user buffers from which to write data
948  * @nr_segs: size of iov vector
949  * @pos: byte offset in file where writing starts
950  *
951  * We use this function for direct writes instead of calling
952  * generic_file_aio_write() in order to avoid taking the inode
953  * semaphore and updating the i_size.  The NFS server will set
954  * the new i_size and this client must read the updated size
955  * back into its cache.  We let the server do generic write
956  * parameter checking and report problems.
957  *
958  * We eliminate local atime updates, see direct read above.
959  *
960  * We avoid unnecessary page cache invalidations for normal cached
961  * readers of this file.
962  *
963  * Note that O_APPEND is not supported for NFS direct writes, as there
964  * is no atomic O_APPEND write facility in the NFS protocol.
965  */
nfs_file_direct_write(struct kiocb * iocb,const struct iovec * iov,unsigned long nr_segs,loff_t pos)966 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
967 				unsigned long nr_segs, loff_t pos)
968 {
969 	ssize_t retval = -EINVAL;
970 	struct file *file = iocb->ki_filp;
971 	struct address_space *mapping = file->f_mapping;
972 	size_t count;
973 
974 	count = iov_length(iov, nr_segs);
975 	nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
976 
977 	dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
978 		file->f_path.dentry->d_parent->d_name.name,
979 		file->f_path.dentry->d_name.name,
980 		count, (long long) pos);
981 
982 	retval = generic_write_checks(file, &pos, &count, 0);
983 	if (retval)
984 		goto out;
985 
986 	retval = -EINVAL;
987 	if ((ssize_t) count < 0)
988 		goto out;
989 	retval = 0;
990 	if (!count)
991 		goto out;
992 
993 	retval = nfs_sync_mapping(mapping);
994 	if (retval)
995 		goto out;
996 
997 	task_io_account_write(count);
998 
999 	retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
1000 
1001 	if (retval > 0)
1002 		iocb->ki_pos = pos + retval;
1003 
1004 out:
1005 	return retval;
1006 }
1007 
1008 /**
1009  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1010  *
1011  */
nfs_init_directcache(void)1012 int __init nfs_init_directcache(void)
1013 {
1014 	nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1015 						sizeof(struct nfs_direct_req),
1016 						0, (SLAB_RECLAIM_ACCOUNT|
1017 							SLAB_MEM_SPREAD),
1018 						NULL);
1019 	if (nfs_direct_cachep == NULL)
1020 		return -ENOMEM;
1021 
1022 	return 0;
1023 }
1024 
1025 /**
1026  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1027  *
1028  */
nfs_destroy_directcache(void)1029 void nfs_destroy_directcache(void)
1030 {
1031 	kmem_cache_destroy(nfs_direct_cachep);
1032 }
1033