aio.c 50.6 KB
Newer Older
Linus Torvalds's avatar
Linus Torvalds committed
1
2
3
4
5
6
7
8
9
10
/*
 *	An async IO implementation for Linux
 *	Written by Benjamin LaHaise <bcrl@kvack.org>
 *
 *	Implements an efficient asynchronous io interface.
 *
 *	Copyright 2000, 2001, 2002 Red Hat, Inc.  All Rights Reserved.
 *
 *	See ../COPYING for licensing terms.
 */
Kent Overstreet's avatar
Kent Overstreet committed
11
12
#define pr_fmt(fmt) "%s: " fmt, __func__

Linus Torvalds's avatar
Linus Torvalds committed
13
14
15
16
17
#include <linux/kernel.h>
#include <linux/init.h>
#include <linux/errno.h>
#include <linux/time.h>
#include <linux/aio_abi.h>
18
#include <linux/export.h>
Linus Torvalds's avatar
Linus Torvalds committed
19
#include <linux/syscalls.h>
20
#include <linux/backing-dev.h>
21
#include <linux/uio.h>
Linus Torvalds's avatar
Linus Torvalds committed
22

23
#include <linux/sched/signal.h>
Linus Torvalds's avatar
Linus Torvalds committed
24
25
26
27
#include <linux/fs.h>
#include <linux/file.h>
#include <linux/mm.h>
#include <linux/mman.h>
28
#include <linux/mmu_context.h>
Kent Overstreet's avatar
Kent Overstreet committed
29
#include <linux/percpu.h>
Linus Torvalds's avatar
Linus Torvalds committed
30
31
32
33
34
35
#include <linux/slab.h>
#include <linux/timer.h>
#include <linux/aio.h>
#include <linux/highmem.h>
#include <linux/workqueue.h>
#include <linux/security.h>
36
#include <linux/eventfd.h>
Jeff Moyer's avatar
Jeff Moyer committed
37
#include <linux/blkdev.h>
38
#include <linux/compat.h>
39
40
#include <linux/migrate.h>
#include <linux/ramfs.h>
Kent Overstreet's avatar
Kent Overstreet committed
41
#include <linux/percpu-refcount.h>
42
#include <linux/mount.h>
Linus Torvalds's avatar
Linus Torvalds committed
43
44

#include <asm/kmap_types.h>
45
#include <linux/uaccess.h>
Linus Torvalds's avatar
Linus Torvalds committed
46

Al Viro's avatar
Al Viro committed
47
48
#include "internal.h"

49
50
51
52
53
54
#define AIO_RING_MAGIC			0xa10a10a1
#define AIO_RING_COMPAT_FEATURES	1
#define AIO_RING_INCOMPAT_FEATURES	0
struct aio_ring {
	unsigned	id;	/* kernel internal index number */
	unsigned	nr;	/* number of io_events */
55
56
	unsigned	head;	/* Written to by userland or under ring_lock
				 * mutex by aio_read_events_ring(). */
57
58
59
60
61
62
63
64
65
66
67
68
69
	unsigned	tail;

	unsigned	magic;
	unsigned	compat_features;
	unsigned	incompat_features;
	unsigned	header_length;	/* size of aio_ring */


	struct io_event		io_events[0];
}; /* 128 bytes + ring size */

#define AIO_RING_PAGES	8

70
struct kioctx_table {
71
72
73
	struct rcu_head		rcu;
	unsigned		nr;
	struct kioctx __rcu	*table[];
74
75
};

Kent Overstreet's avatar
Kent Overstreet committed
76
77
78
79
struct kioctx_cpu {
	unsigned		reqs_available;
};

80
81
82
83
84
struct ctx_rq_wait {
	struct completion comp;
	atomic_t count;
};

85
struct kioctx {
Kent Overstreet's avatar
Kent Overstreet committed
86
	struct percpu_ref	users;
Kent Overstreet's avatar
Kent Overstreet committed
87
	atomic_t		dead;
88

Kent Overstreet's avatar
Kent Overstreet committed
89
90
	struct percpu_ref	reqs;

91
92
	unsigned long		user_id;

Kent Overstreet's avatar
Kent Overstreet committed
93
94
95
96
97
98
99
	struct __percpu kioctx_cpu *cpu;

	/*
	 * For percpu reqs_available, number of slots we move to/from global
	 * counter at a time:
	 */
	unsigned		req_batch;
100
101
102
103
	/*
	 * This is what userspace passed to io_setup(), it's not used for
	 * anything but counting against the global max_reqs quota.
	 *
104
	 * The real limit is nr_events - 1, which will be larger (see
105
106
	 * aio_setup_ring())
	 */
107
108
	unsigned		max_reqs;

109
110
	/* Size of ringbuffer, in units of struct io_event */
	unsigned		nr_events;
111

112
113
114
115
116
117
	unsigned long		mmap_base;
	unsigned long		mmap_size;

	struct page		**ring_pages;
	long			nr_pages;

118
	struct rcu_work		free_rwork;	/* see free_ioctx() */
119

120
121
122
	/*
	 * signals when all in-flight requests are done
	 */
123
	struct ctx_rq_wait	*rq_wait;
124

125
	struct {
126
127
128
129
130
		/*
		 * This counts the number of available slots in the ringbuffer,
		 * so we avoid overflowing it: it's decremented (if positive)
		 * when allocating a kiocb and incremented when the resulting
		 * io_event is pulled off the ringbuffer.
Kent Overstreet's avatar
Kent Overstreet committed
131
132
		 *
		 * We batch accesses to it with a percpu version.
133
134
		 */
		atomic_t	reqs_available;
135
136
137
138
139
140
141
	} ____cacheline_aligned_in_smp;

	struct {
		spinlock_t	ctx_lock;
		struct list_head active_reqs;	/* used for cancellation */
	} ____cacheline_aligned_in_smp;

142
143
	struct {
		struct mutex	ring_lock;
144
145
		wait_queue_head_t wait;
	} ____cacheline_aligned_in_smp;
146
147
148

	struct {
		unsigned	tail;
149
		unsigned	completed_events;
150
		spinlock_t	completion_lock;
151
	} ____cacheline_aligned_in_smp;
152
153

	struct page		*internal_pages[AIO_RING_PAGES];
154
	struct file		*aio_ring_file;
155
156

	unsigned		id;
157
158
};

159
160
161
162
163
164
struct fsync_iocb {
	struct work_struct	work;
	struct file		*file;
	bool			datasync;
};

165
166
167
168
169
170
171
172
173
174
175
176
177
178
/*
 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
 * cancelled or completed (this makes a certain amount of sense because
 * successful cancellation - io_cancel() - does deliver the completion to
 * userspace).
 *
 * And since most things don't implement kiocb cancellation and we'd really like
 * kiocb completion to be lockless when possible, we use ki_cancel to
 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
 */
#define KIOCB_CANCELLED		((void *) (~0ULL))

struct aio_kiocb {
179
180
	union {
		struct kiocb		rw;
181
		struct fsync_iocb	fsync;
182
	};
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199

	struct kioctx		*ki_ctx;
	kiocb_cancel_fn		*ki_cancel;

	struct iocb __user	*ki_user_iocb;	/* user's aiocb */
	__u64			ki_user_data;	/* user's data for completion */

	struct list_head	ki_list;	/* the aio core uses this
						 * for cancellation */

	/*
	 * If the aio_resfd field of the userspace iocb is not zero,
	 * this is the underlying eventfd context to deliver events to.
	 */
	struct eventfd_ctx	*ki_eventfd;
};

Linus Torvalds's avatar
Linus Torvalds committed
200
/*------ sysctl variables----*/
201
202
203
static DEFINE_SPINLOCK(aio_nr_lock);
unsigned long aio_nr;		/* current system wide number of aio requests */
unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
Linus Torvalds's avatar
Linus Torvalds committed
204
205
/*----end sysctl variables---*/

206
207
static struct kmem_cache	*kiocb_cachep;
static struct kmem_cache	*kioctx_cachep;
Linus Torvalds's avatar
Linus Torvalds committed
208

209
210
211
212
213
214
215
216
217
218
219
static struct vfsmount *aio_mnt;

static const struct file_operations aio_ring_fops;
static const struct address_space_operations aio_ctx_aops;

static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
{
	struct qstr this = QSTR_INIT("[aio]", 5);
	struct file *file;
	struct path path;
	struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
220
221
	if (IS_ERR(inode))
		return ERR_CAST(inode);
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250

	inode->i_mapping->a_ops = &aio_ctx_aops;
	inode->i_mapping->private_data = ctx;
	inode->i_size = PAGE_SIZE * nr_pages;

	path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
	if (!path.dentry) {
		iput(inode);
		return ERR_PTR(-ENOMEM);
	}
	path.mnt = mntget(aio_mnt);

	d_instantiate(path.dentry, inode);
	file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
	if (IS_ERR(file)) {
		path_put(&path);
		return file;
	}

	file->f_flags = O_RDWR;
	return file;
}

static struct dentry *aio_mount(struct file_system_type *fs_type,
				int flags, const char *dev_name, void *data)
{
	static const struct dentry_operations ops = {
		.d_dname	= simple_dname,
	};
Jann Horn's avatar
Jann Horn committed
251
252
253
254
255
256
	struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops,
					   AIO_RING_MAGIC);

	if (!IS_ERR(root))
		root->d_sb->s_iflags |= SB_I_NOEXEC;
	return root;
257
258
}

Linus Torvalds's avatar
Linus Torvalds committed
259
260
261
262
263
264
/* aio_setup
 *	Creates the slab caches used by the aio routines, panic on
 *	failure as this is done early during the boot sequence.
 */
static int __init aio_setup(void)
{
265
266
267
268
269
270
271
272
273
	static struct file_system_type aio_fs = {
		.name		= "aio",
		.mount		= aio_mount,
		.kill_sb	= kill_anon_super,
	};
	aio_mnt = kern_mount(&aio_fs);
	if (IS_ERR(aio_mnt))
		panic("Failed to create aio fs mount.");

274
	kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
275
	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds's avatar
Linus Torvalds committed
276
277
	return 0;
}
278
__initcall(aio_setup);
Linus Torvalds's avatar
Linus Torvalds committed
279

280
281
282
static void put_aio_ring_file(struct kioctx *ctx)
{
	struct file *aio_ring_file = ctx->aio_ring_file;
283
284
	struct address_space *i_mapping;

285
	if (aio_ring_file) {
Al Viro's avatar
Al Viro committed
286
		truncate_setsize(file_inode(aio_ring_file), 0);
287
288

		/* Prevent further access to the kioctx from migratepages */
Al Viro's avatar
Al Viro committed
289
		i_mapping = aio_ring_file->f_mapping;
290
291
		spin_lock(&i_mapping->private_lock);
		i_mapping->private_data = NULL;
292
		ctx->aio_ring_file = NULL;
293
		spin_unlock(&i_mapping->private_lock);
294
295
296
297
298

		fput(aio_ring_file);
	}
}

Linus Torvalds's avatar
Linus Torvalds committed
299
300
static void aio_free_ring(struct kioctx *ctx)
{
301
	int i;
Linus Torvalds's avatar
Linus Torvalds committed
302

303
304
305
306
307
	/* Disconnect the kiotx from the ring file.  This prevents future
	 * accesses to the kioctx from page migration.
	 */
	put_aio_ring_file(ctx);

308
	for (i = 0; i < ctx->nr_pages; i++) {
309
		struct page *page;
310
311
		pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
				page_count(ctx->ring_pages[i]));
312
313
314
315
316
		page = ctx->ring_pages[i];
		if (!page)
			continue;
		ctx->ring_pages[i] = NULL;
		put_page(page);
317
	}
Linus Torvalds's avatar
Linus Torvalds committed
318

319
	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
320
		kfree(ctx->ring_pages);
321
322
		ctx->ring_pages = NULL;
	}
323
324
}

325
static int aio_ring_mremap(struct vm_area_struct *vma)
326
{
327
	struct file *file = vma->vm_file;
328
329
	struct mm_struct *mm = vma->vm_mm;
	struct kioctx_table *table;
Al Viro's avatar
Al Viro committed
330
	int i, res = -EINVAL;
331
332
333
334
335
336
337

	spin_lock(&mm->ioctx_lock);
	rcu_read_lock();
	table = rcu_dereference(mm->ioctx_table);
	for (i = 0; i < table->nr; i++) {
		struct kioctx *ctx;

338
		ctx = rcu_dereference(table->table[i]);
339
		if (ctx && ctx->aio_ring_file == file) {
Al Viro's avatar
Al Viro committed
340
341
342
343
			if (!atomic_read(&ctx->dead)) {
				ctx->user_id = ctx->mmap_base = vma->vm_start;
				res = 0;
			}
344
345
346
347
348
349
			break;
		}
	}

	rcu_read_unlock();
	spin_unlock(&mm->ioctx_lock);
Al Viro's avatar
Al Viro committed
350
	return res;
351
352
}

353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
static const struct vm_operations_struct aio_ring_vm_ops = {
	.mremap		= aio_ring_mremap,
#if IS_ENABLED(CONFIG_MMU)
	.fault		= filemap_fault,
	.map_pages	= filemap_map_pages,
	.page_mkwrite	= filemap_page_mkwrite,
#endif
};

static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
{
	vma->vm_flags |= VM_DONTEXPAND;
	vma->vm_ops = &aio_ring_vm_ops;
	return 0;
}

369
370
371
372
static const struct file_operations aio_ring_fops = {
	.mmap = aio_ring_mmap,
};

373
#if IS_ENABLED(CONFIG_MIGRATION)
374
375
376
static int aio_migratepage(struct address_space *mapping, struct page *new,
			struct page *old, enum migrate_mode mode)
{
377
	struct kioctx *ctx;
378
	unsigned long flags;
379
	pgoff_t idx;
380
381
	int rc;

382
383
384
385
386
387
388
389
	/*
	 * We cannot support the _NO_COPY case here, because copy needs to
	 * happen under the ctx->completion_lock. That does not work with the
	 * migration workflow of MIGRATE_SYNC_NO_COPY.
	 */
	if (mode == MIGRATE_SYNC_NO_COPY)
		return -EINVAL;

390
391
	rc = 0;

392
	/* mapping->private_lock here protects against the kioctx teardown.  */
393
394
	spin_lock(&mapping->private_lock);
	ctx = mapping->private_data;
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
	if (!ctx) {
		rc = -EINVAL;
		goto out;
	}

	/* The ring_lock mutex.  The prevents aio_read_events() from writing
	 * to the ring's head, and prevents page migration from mucking in
	 * a partially initialized kiotx.
	 */
	if (!mutex_trylock(&ctx->ring_lock)) {
		rc = -EAGAIN;
		goto out;
	}

	idx = old->index;
	if (idx < (pgoff_t)ctx->nr_pages) {
		/* Make sure the old page hasn't already been changed */
		if (ctx->ring_pages[idx] != old)
			rc = -EAGAIN;
414
415
416
417
	} else
		rc = -EINVAL;

	if (rc != 0)
418
		goto out_unlock;
419

420
421
	/* Writeback must be complete */
	BUG_ON(PageWriteback(old));
422
	get_page(new);
423

424
	rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
425
	if (rc != MIGRATEPAGE_SUCCESS) {
426
		put_page(new);
427
		goto out_unlock;
428
429
	}

430
431
432
	/* Take completion_lock to prevent other writes to the ring buffer
	 * while the old page is copied to the new.  This prevents new
	 * events from being lost.
433
	 */
434
435
436
437
438
	spin_lock_irqsave(&ctx->completion_lock, flags);
	migrate_page_copy(new, old);
	BUG_ON(ctx->ring_pages[idx] != old);
	ctx->ring_pages[idx] = new;
	spin_unlock_irqrestore(&ctx->completion_lock, flags);
439

440
441
	/* The old page is no longer accessible. */
	put_page(old);
442

443
444
445
446
out_unlock:
	mutex_unlock(&ctx->ring_lock);
out:
	spin_unlock(&mapping->private_lock);
447
	return rc;
Linus Torvalds's avatar
Linus Torvalds committed
448
}
449
#endif
Linus Torvalds's avatar
Linus Torvalds committed
450

451
static const struct address_space_operations aio_ctx_aops = {
452
	.set_page_dirty = __set_page_dirty_no_writeback,
453
#if IS_ENABLED(CONFIG_MIGRATION)
454
	.migratepage	= aio_migratepage,
455
#endif
456
457
};

458
static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
Linus Torvalds's avatar
Linus Torvalds committed
459
460
{
	struct aio_ring *ring;
Zach Brown's avatar
Zach Brown committed
461
	struct mm_struct *mm = current->mm;
462
	unsigned long size, unused;
Linus Torvalds's avatar
Linus Torvalds committed
463
	int nr_pages;
464
465
	int i;
	struct file *file;
Linus Torvalds's avatar
Linus Torvalds committed
466
467
468
469
470
471
472

	/* Compensate for the ring buffer's head/tail overlap entry */
	nr_events += 2;	/* 1 is required, 2 for good luck */

	size = sizeof(struct aio_ring);
	size += sizeof(struct io_event) * nr_events;

473
	nr_pages = PFN_UP(size);
Linus Torvalds's avatar
Linus Torvalds committed
474
475
476
	if (nr_pages < 0)
		return -EINVAL;

477
	file = aio_private_file(ctx, nr_pages);
478
479
	if (IS_ERR(file)) {
		ctx->aio_ring_file = NULL;
480
		return -ENOMEM;
481
482
	}

483
484
485
486
487
488
489
490
491
492
493
494
495
496
	ctx->aio_ring_file = file;
	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
			/ sizeof(struct io_event);

	ctx->ring_pages = ctx->internal_pages;
	if (nr_pages > AIO_RING_PAGES) {
		ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
					  GFP_KERNEL);
		if (!ctx->ring_pages) {
			put_aio_ring_file(ctx);
			return -ENOMEM;
		}
	}

497
498
	for (i = 0; i < nr_pages; i++) {
		struct page *page;
Al Viro's avatar
Al Viro committed
499
		page = find_or_create_page(file->f_mapping,
500
501
502
503
504
505
506
					   i, GFP_HIGHUSER | __GFP_ZERO);
		if (!page)
			break;
		pr_debug("pid(%d) page[%d]->count=%d\n",
			 current->pid, i, page_count(page));
		SetPageUptodate(page);
		unlock_page(page);
507
508

		ctx->ring_pages[i] = page;
509
	}
510
	ctx->nr_pages = i;
Linus Torvalds's avatar
Linus Torvalds committed
511

512
513
	if (unlikely(i != nr_pages)) {
		aio_free_ring(ctx);
514
		return -ENOMEM;
Linus Torvalds's avatar
Linus Torvalds committed
515
516
	}

517
518
	ctx->mmap_size = nr_pages * PAGE_SIZE;
	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
519

520
521
522
523
524
525
	if (down_write_killable(&mm->mmap_sem)) {
		ctx->mmap_size = 0;
		aio_free_ring(ctx);
		return -EINTR;
	}

526
527
	ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
				       PROT_READ | PROT_WRITE,
528
				       MAP_SHARED, 0, &unused, NULL);
529
	up_write(&mm->mmap_sem);
530
531
	if (IS_ERR((void *)ctx->mmap_base)) {
		ctx->mmap_size = 0;
Linus Torvalds's avatar
Linus Torvalds committed
532
		aio_free_ring(ctx);
533
		return -ENOMEM;
Linus Torvalds's avatar
Linus Torvalds committed
534
535
	}

536
	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
537

538
539
	ctx->user_id = ctx->mmap_base;
	ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds's avatar
Linus Torvalds committed
540

541
	ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds's avatar
Linus Torvalds committed
542
	ring->nr = nr_events;	/* user copy */
543
	ring->id = ~0U;
Linus Torvalds's avatar
Linus Torvalds committed
544
545
546
547
548
	ring->head = ring->tail = 0;
	ring->magic = AIO_RING_MAGIC;
	ring->compat_features = AIO_RING_COMPAT_FEATURES;
	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
	ring->header_length = sizeof(struct aio_ring);
549
	kunmap_atomic(ring);
550
	flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds's avatar
Linus Torvalds committed
551
552
553
554
555
556
557
558

	return 0;
}

#define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
#define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
#define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)

559
void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
560
{
561
	struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
562
563
564
	struct kioctx *ctx = req->ki_ctx;
	unsigned long flags;

565
566
	if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
		return;
567

568
569
	spin_lock_irqsave(&ctx->ctx_lock, flags);
	list_add_tail(&req->ki_list, &ctx->active_reqs);
570
571
572
573
574
	req->ki_cancel = cancel;
	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
}
EXPORT_SYMBOL(kiocb_set_cancel_fn);

575
static int kiocb_cancel(struct aio_kiocb *kiocb)
Kent Overstreet's avatar
Kent Overstreet committed
576
{
577
	kiocb_cancel_fn *old, *cancel;
Kent Overstreet's avatar
Kent Overstreet committed
578

579
580
581
582
583
	/*
	 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
	 * actually has a cancel function, hence the cmpxchg()
	 */

584
	cancel = READ_ONCE(kiocb->ki_cancel);
585
586
	do {
		if (!cancel || cancel == KIOCB_CANCELLED)
Kent Overstreet's avatar
Kent Overstreet committed
587
			return -EINVAL;
Kent Overstreet's avatar
Kent Overstreet committed
588

589
590
591
		old = cancel;
		cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
	} while (cancel != old);
Kent Overstreet's avatar
Kent Overstreet committed
592

593
	return cancel(&kiocb->rw);
Kent Overstreet's avatar
Kent Overstreet committed
594
595
}

596
597
598
/*
 * free_ioctx() should be RCU delayed to synchronize against the RCU
 * protected lookup_ioctx() and also needs process context to call
599
 * aio_free_ring().  Use rcu_work.
600
 */
Kent Overstreet's avatar
Kent Overstreet committed
601
static void free_ioctx(struct work_struct *work)
Kent Overstreet's avatar
Kent Overstreet committed
602
{
603
604
	struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
					  free_rwork);
Kent Overstreet's avatar
Kent Overstreet committed
605
	pr_debug("freeing %p\n", ctx);
Kent Overstreet's avatar
Kent Overstreet committed
606

Kent Overstreet's avatar
Kent Overstreet committed
607
	aio_free_ring(ctx);
Kent Overstreet's avatar
Kent Overstreet committed
608
	free_percpu(ctx->cpu);
609
610
	percpu_ref_exit(&ctx->reqs);
	percpu_ref_exit(&ctx->users);
Kent Overstreet's avatar
Kent Overstreet committed
611
612
613
	kmem_cache_free(kioctx_cachep, ctx);
}

Kent Overstreet's avatar
Kent Overstreet committed
614
615
616
617
static void free_ioctx_reqs(struct percpu_ref *ref)
{
	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);

618
	/* At this point we know that there are no any in-flight requests */
619
620
	if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
		complete(&ctx->rq_wait->comp);
621

622
	/* Synchronize against RCU protected table->table[] dereferences */
623
624
	INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
	queue_rcu_work(system_wq, &ctx->free_rwork);
Kent Overstreet's avatar
Kent Overstreet committed
625
626
}

Kent Overstreet's avatar
Kent Overstreet committed
627
628
629
630
631
/*
 * When this function runs, the kioctx has been removed from the "hash table"
 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
 * now it's safe to cancel any that need to be.
 */
Kent Overstreet's avatar
Kent Overstreet committed
632
static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet's avatar
Kent Overstreet committed
633
{
Kent Overstreet's avatar
Kent Overstreet committed
634
	struct kioctx *ctx = container_of(ref, struct kioctx, users);
635
	struct aio_kiocb *req;
Kent Overstreet's avatar
Kent Overstreet committed
636
637
638
639
640

	spin_lock_irq(&ctx->ctx_lock);

	while (!list_empty(&ctx->active_reqs)) {
		req = list_first_entry(&ctx->active_reqs,
641
				       struct aio_kiocb, ki_list);
642
		kiocb_cancel(req);
Al Viro's avatar
Al Viro committed
643
		list_del_init(&req->ki_list);
Kent Overstreet's avatar
Kent Overstreet committed
644
645
646
647
	}

	spin_unlock_irq(&ctx->ctx_lock);

Kent Overstreet's avatar
Kent Overstreet committed
648
649
	percpu_ref_kill(&ctx->reqs);
	percpu_ref_put(&ctx->reqs);
Kent Overstreet's avatar
Kent Overstreet committed
650
651
}

652
653
654
655
656
657
658
static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
{
	unsigned i, new_nr;
	struct kioctx_table *table, *old;
	struct aio_ring *ring;

	spin_lock(&mm->ioctx_lock);
659
	table = rcu_dereference_raw(mm->ioctx_table);
660
661
662
663

	while (1) {
		if (table)
			for (i = 0; i < table->nr; i++)
664
				if (!rcu_access_pointer(table->table[i])) {
665
					ctx->id = i;
666
					rcu_assign_pointer(table->table[i], ctx);
667
668
					spin_unlock(&mm->ioctx_lock);

669
670
671
672
					/* While kioctx setup is in progress,
					 * we are protected from page migration
					 * changes ring_pages by ->ring_lock.
					 */
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
					ring = kmap_atomic(ctx->ring_pages[0]);
					ring->id = ctx->id;
					kunmap_atomic(ring);
					return 0;
				}

		new_nr = (table ? table->nr : 1) * 4;
		spin_unlock(&mm->ioctx_lock);

		table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
				new_nr, GFP_KERNEL);
		if (!table)
			return -ENOMEM;

		table->nr = new_nr;

		spin_lock(&mm->ioctx_lock);
690
		old = rcu_dereference_raw(mm->ioctx_table);
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706

		if (!old) {
			rcu_assign_pointer(mm->ioctx_table, table);
		} else if (table->nr > old->nr) {
			memcpy(table->table, old->table,
			       old->nr * sizeof(struct kioctx *));

			rcu_assign_pointer(mm->ioctx_table, table);
			kfree_rcu(old, rcu);
		} else {
			kfree(table);
			table = old;
		}
	}
}

Kent Overstreet's avatar
Kent Overstreet committed
707
708
709
710
711
712
713
714
715
716
static void aio_nr_sub(unsigned nr)
{
	spin_lock(&aio_nr_lock);
	if (WARN_ON(aio_nr - nr > aio_nr))
		aio_nr = 0;
	else
		aio_nr -= nr;
	spin_unlock(&aio_nr_lock);
}

Linus Torvalds's avatar
Linus Torvalds committed
717
718
719
720
721
/* ioctx_alloc
 *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
 */
static struct kioctx *ioctx_alloc(unsigned nr_events)
{
Zach Brown's avatar
Zach Brown committed
722
	struct mm_struct *mm = current->mm;
Linus Torvalds's avatar
Linus Torvalds committed
723
	struct kioctx *ctx;
724
	int err = -ENOMEM;
Linus Torvalds's avatar
Linus Torvalds committed
725

726
727
728
729
730
731
	/*
	 * Store the original nr_events -- what userspace passed to io_setup(),
	 * for counting against the global limit -- before it changes.
	 */
	unsigned int max_reqs = nr_events;

Kent Overstreet's avatar
Kent Overstreet committed
732
733
734
735
736
737
738
739
740
741
742
743
	/*
	 * We keep track of the number of available ringbuffer slots, to prevent
	 * overflow (reqs_available), and we also use percpu counters for this.
	 *
	 * So since up to half the slots might be on other cpu's percpu counters
	 * and unavailable, double nr_events so userspace sees what they
	 * expected: additionally, we move req_batch slots to/from percpu
	 * counters at a time, so make sure that isn't 0:
	 */
	nr_events = max(nr_events, num_possible_cpus() * 4);
	nr_events *= 2;

Linus Torvalds's avatar
Linus Torvalds committed
744
	/* Prevent overflows */
745
	if (nr_events > (0x10000000U / sizeof(struct io_event))) {
Linus Torvalds's avatar
Linus Torvalds committed
746
747
748
749
		pr_debug("ENOMEM: nr_events too high\n");
		return ERR_PTR(-EINVAL);
	}

750
	if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
Linus Torvalds's avatar
Linus Torvalds committed
751
752
		return ERR_PTR(-EAGAIN);

753
	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds's avatar
Linus Torvalds committed
754
755
756
	if (!ctx)
		return ERR_PTR(-ENOMEM);

757
	ctx->max_reqs = max_reqs;
Linus Torvalds's avatar
Linus Torvalds committed
758
759

	spin_lock_init(&ctx->ctx_lock);
760
	spin_lock_init(&ctx->completion_lock);
761
	mutex_init(&ctx->ring_lock);
762
763
764
	/* Protect against page migration throughout kiotx setup by keeping
	 * the ring_lock mutex held until setup is complete. */
	mutex_lock(&ctx->ring_lock);
Linus Torvalds's avatar
Linus Torvalds committed
765
766
767
768
	init_waitqueue_head(&ctx->wait);

	INIT_LIST_HEAD(&ctx->active_reqs);

769
	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
770
771
		goto err;

772
	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
773
774
		goto err;

Kent Overstreet's avatar
Kent Overstreet committed
775
776
	ctx->cpu = alloc_percpu(struct kioctx_cpu);
	if (!ctx->cpu)
Kent Overstreet's avatar
Kent Overstreet committed
777
		goto err;
Linus Torvalds's avatar
Linus Torvalds committed
778

779
	err = aio_setup_ring(ctx, nr_events);
780
	if (err < 0)
Kent Overstreet's avatar
Kent Overstreet committed
781
		goto err;
Kent Overstreet's avatar
Kent Overstreet committed
782

783
	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreet's avatar
Kent Overstreet committed
784
	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
785
786
	if (ctx->req_batch < 1)
		ctx->req_batch = 1;
787

Linus Torvalds's avatar
Linus Torvalds committed
788
	/* limit the number of system wide aios */
789
	spin_lock(&aio_nr_lock);
790
791
	if (aio_nr + ctx->max_reqs > aio_max_nr ||
	    aio_nr + ctx->max_reqs < aio_nr) {
792
		spin_unlock(&aio_nr_lock);
Kent Overstreet's avatar
Kent Overstreet committed
793
		err = -EAGAIN;
794
		goto err_ctx;
795
796
	}
	aio_nr += ctx->max_reqs;
797
	spin_unlock(&aio_nr_lock);
Linus Torvalds's avatar
Linus Torvalds committed
798

799
800
	percpu_ref_get(&ctx->users);	/* io_setup() will drop this ref */
	percpu_ref_get(&ctx->reqs);	/* free_ioctx_users() will drop this */
Kent Overstreet's avatar
Kent Overstreet committed
801

802
803
	err = ioctx_add_table(ctx, mm);
	if (err)
Kent Overstreet's avatar
Kent Overstreet committed
804
		goto err_cleanup;
805

806
807
808
	/* Release the ring_lock mutex now that all setup is complete. */
	mutex_unlock(&ctx->ring_lock);

Kent Overstreet's avatar
Kent Overstreet committed
809
	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
810
		 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds's avatar
Linus Torvalds committed
811
812
	return ctx;

Kent Overstreet's avatar
Kent Overstreet committed
813
814
err_cleanup:
	aio_nr_sub(ctx->max_reqs);
815
err_ctx:
816
817
818
	atomic_set(&ctx->dead, 1);
	if (ctx->mmap_size)
		vm_munmap(ctx->mmap_base, ctx->mmap_size);
819
	aio_free_ring(ctx);
Kent Overstreet's avatar
Kent Overstreet committed
820
err:
821
	mutex_unlock(&ctx->ring_lock);
Kent Overstreet's avatar
Kent Overstreet committed
822
	free_percpu(ctx->cpu);
823
824
	percpu_ref_exit(&ctx->reqs);
	percpu_ref_exit(&ctx->users);
Linus Torvalds's avatar
Linus Torvalds committed
825
	kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreet's avatar
Kent Overstreet committed
826
	pr_debug("error allocating ioctx %d\n", err);
827
	return ERR_PTR(err);
Linus Torvalds's avatar
Linus Torvalds committed
828
829
}

Kent Overstreet's avatar
Kent Overstreet committed
830
831
832
833
834
/* kill_ioctx
 *	Cancels all outstanding aio requests on an aio context.  Used
 *	when the processes owning a context have all exited to encourage
 *	the rapid destruction of the kioctx.
 */
835
static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
836
		      struct ctx_rq_wait *wait)
Kent Overstreet's avatar
Kent Overstreet committed
837
{
838
	struct kioctx_table *table;
839

Al Viro's avatar
Al Viro committed
840
841
842
	spin_lock(&mm->ioctx_lock);
	if (atomic_xchg(&ctx->dead, 1)) {
		spin_unlock(&mm->ioctx_lock);
843
		return -EINVAL;
Al Viro's avatar
Al Viro committed
844
	}
845

846
	table = rcu_dereference_raw(mm->ioctx_table);
847
848
	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
	RCU_INIT_POINTER(table->table[ctx->id], NULL);
849
	spin_unlock(&mm->ioctx_lock);
850

851
	/* free_ioctx_reqs() will do the necessary RCU synchronization */
852
	wake_up_all(&ctx->wait);
853

854
855
856
857
858
859
860
861
	/*
	 * It'd be more correct to do this in free_ioctx(), after all
	 * the outstanding kiocbs have finished - but by then io_destroy
	 * has already returned, so io_setup() could potentially return
	 * -EAGAIN with no ioctxs actually in use (as far as userspace
	 *  could tell).
	 */
	aio_nr_sub(ctx->max_reqs);
862

863
864
	if (ctx->mmap_size)
		vm_munmap(ctx->mmap_base, ctx->mmap_size);
865

866
	ctx->rq_wait = wait;
867
868
	percpu_ref_kill(&ctx->users);
	return 0;
Linus Torvalds's avatar
Linus Torvalds committed
869
870
}

Kent Overstreet's avatar
Kent Overstreet committed
871
872
873
874
875
876
877
/*
 * exit_aio: called when the last user of mm goes away.  At this point, there is
 * no way for any new requests to be submited or any of the io_* syscalls to be
 * called on the context.
 *
 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
 * them.
Linus Torvalds's avatar
Linus Torvalds committed
878
 */
879
void exit_aio(struct mm_struct *mm)
Linus Torvalds's avatar
Linus Torvalds committed
880
{
881
	struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
882
883
	struct ctx_rq_wait wait;
	int i, skipped;
884

885
886
	if (!table)
		return;
887

888
889
890
891
	atomic_set(&wait.count, table->nr);
	init_completion(&wait.comp);

	skipped = 0;
892
	for (i = 0; i < table->nr; ++i) {
893
894
		struct kioctx *ctx =
			rcu_dereference_protected(table->table[i], true);
895

896
897
		if (!ctx) {
			skipped++;
898
			continue;
899
900
		}

901
		/*
902
903
904
905
906
		 * We don't need to bother with munmap() here - exit_mmap(mm)
		 * is coming and it'll unmap everything. And we simply can't,
		 * this is not necessarily our ->mm.
		 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
		 * that it needs to unmap the area, just set it to 0.
907
		 */
908
		ctx->mmap_size = 0;
909
910
		kill_ioctx(mm, ctx, &wait);
	}
Kent Overstreet's avatar
Kent Overstreet committed
911

912
	if (!atomic_sub_and_test(skipped, &wait.count)) {
913
		/* Wait until all IO for the context are done. */
914
		wait_for_completion(&wait.comp);
Linus Torvalds's avatar
Linus Torvalds committed
915
	}
916
917
918

	RCU_INIT_POINTER(mm->ioctx_table, NULL);
	kfree(table);
Linus Torvalds's avatar
Linus Torvalds committed
919
920
}

Kent Overstreet's avatar
Kent Overstreet committed
921
922
923
static void put_reqs_available(struct kioctx *ctx, unsigned nr)
{
	struct kioctx_cpu *kcpu;
924
	unsigned long flags;
Kent Overstreet's avatar
Kent Overstreet committed
925

926
	local_irq_save(flags);
927
	kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreet's avatar
Kent Overstreet committed
928
	kcpu->reqs_available += nr;
929

Kent Overstreet's avatar
Kent Overstreet committed
930
931
932
933
934
	while (kcpu->reqs_available >= ctx->req_batch * 2) {
		kcpu->reqs_available -= ctx->req_batch;
		atomic_add(ctx->req_batch, &ctx->reqs_available);
	}

935
	local_irq_restore(flags);
Kent Overstreet's avatar
Kent Overstreet committed
936
937
938
939
940
941
}

static bool get_reqs_available(struct kioctx *ctx)
{
	struct kioctx_cpu *kcpu;
	bool ret = false;
942
	unsigned long flags;
Kent Overstreet's avatar
Kent Overstreet committed
943

944
	local_irq_save(flags);
945
	kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreet's avatar
Kent Overstreet committed
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
	if (!kcpu->reqs_available) {
		int old, avail = atomic_read(&ctx->reqs_available);

		do {
			if (avail < ctx->req_batch)
				goto out;

			old = avail;
			avail = atomic_cmpxchg(&ctx->reqs_available,
					       avail, avail - ctx->req_batch);
		} while (avail != old);

		kcpu->reqs_available += ctx->req_batch;
	}

	ret = true;
	kcpu->reqs_available--;
out:
964
	local_irq_restore(flags);
Kent Overstreet's avatar
Kent Overstreet committed
965
966
967
	return ret;
}

968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
/* refill_reqs_available
 *	Updates the reqs_available reference counts used for tracking the
 *	number of free slots in the completion ring.  This can be called
 *	from aio_complete() (to optimistically update reqs_available) or
 *	from aio_get_req() (the we're out of events case).  It must be
 *	called holding ctx->completion_lock.
 */
static void refill_reqs_available(struct kioctx *ctx, unsigned head,
                                  unsigned tail)
{
	unsigned events_in_ring, completed;

	/* Clamp head since userland can write to it. */
	head %= ctx->nr_events;
	if (head <= tail)
		events_in_ring = tail - head;
	else
		events_in_ring = ctx->nr_events - (head - tail);

	completed = ctx->completed_events;
	if (events_in_ring < completed)
		completed -= events_in_ring;
	else
		completed = 0;

	if (!completed)
		return;

	ctx->completed_events -= completed;
	put_reqs_available(ctx, completed);
}

/* user_refill_reqs_available
 *	Called to refill reqs_available when aio_get_req() encounters an
 *	out of space in the completion ring.
 */
static void user_refill_reqs_available(struct kioctx *ctx)
{
	spin_lock_irq(&ctx->completion_lock);
	if (ctx->completed_events) {
		struct aio_ring *ring;
		unsigned head;

		/* Access of ring->head may race with aio_read_events_ring()
		 * here, but that's okay since whether we read the old version
		 * or the new version, and either will be valid.  The important
		 * part is that head cannot pass tail since we prevent
		 * aio_complete() from updating tail by holding
		 * ctx->completion_lock.  Even if head is invalid, the check
		 * against ctx->completed_events below will make sure we do the
		 * safe/right thing.
		 */
		ring = kmap_atomic(ctx->ring_pages[0]);
		head = ring->head;
		kunmap_atomic(ring);

		refill_reqs_available(ctx, head, ctx->tail);
	}

	spin_unlock_irq(&ctx->completion_lock);
}

Linus Torvalds's avatar
Linus Torvalds committed
1030
/* aio_get_req
Kent Overstreet's avatar
Kent Overstreet committed
1031
1032
 *	Allocate a slot for an aio request.
 * Returns NULL if no requests are free.
Linus Torvalds's avatar
Linus Torvalds committed
1033
 */
1034
static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds's avatar
Linus Torvalds committed
1035
{
1036
	struct aio_kiocb *req;
Kent Overstreet's avatar
Kent Overstreet committed
1037

1038
1039
1040
1041
1042
	if (!get_reqs_available(ctx)) {
		user_refill_reqs_available(ctx);
		if (!get_reqs_available(ctx))
			return NULL;
	}
Kent Overstreet's avatar
Kent Overstreet committed
1043

1044
	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
Linus Torvalds's avatar
Linus Torvalds committed
1045
	if (unlikely(!req))
Kent Overstreet's avatar
Kent Overstreet committed
1046
		goto out_put;
Linus Torvalds's avatar
Linus Torvalds committed
1047

Kent Overstreet's avatar
Kent Overstreet committed
1048
	percpu_ref_get(&ctx->reqs);
1049
	INIT_LIST_HEAD(&req->ki_list);
Linus Torvalds's avatar
Linus Torvalds committed
1050
	req->ki_ctx = ctx;
Jeff Moyer's avatar
Jeff Moyer committed
1051
	return req;
Kent Overstreet's avatar
Kent Overstreet committed
1052
out_put:
Kent Overstreet's avatar
Kent Overstreet committed
1053
	put_reqs_available(ctx, 1);
Kent Overstreet's avatar
Kent Overstreet committed
1054
	return NULL;
Linus Torvalds's avatar
Linus Torvalds committed
1055
1056
}

1057
static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds's avatar
Linus Torvalds committed
1058
{
1059
	struct aio_ring __user *ring  = (void __user *)ctx_id;
1060
	struct mm_struct *mm = current->mm;
1061
	struct kioctx *ctx, *ret = NULL;
1062
1063
1064
1065
1066
	struct kioctx_table *table;
	unsigned id;

	if (get_user(id, &ring->id))
		return NULL;
Linus Torvalds's avatar
Linus Torvalds committed
1067

1068
	rcu_read_lock();
1069
	table = rcu_dereference(mm->ioctx_table);
1070

1071
1072
	if (!table || id >= table->nr)
		goto out;
Linus Torvalds's avatar
Linus Torvalds committed
1073

1074
	ctx = rcu_dereference(table->table[id]);
1075
	if (ctx && ctx->user_id == ctx_id) {
1076
1077
		if (percpu_ref_tryget_live(&ctx->users))
			ret = ctx;
1078
1079
	}
out:
1080
	rcu_read_unlock();
1081
	return ret;
Linus Torvalds's avatar
Linus Torvalds committed
1082
1083
1084
1085
1086
}

/* aio_complete
 *	Called when the io request on the given iocb is complete.
 */
1087
static void aio_complete(struct aio_kiocb *iocb, long res, long res2)
Linus Torvalds's avatar
Linus Torvalds committed
1088
1089
1090
{
	struct kioctx	*ctx = iocb->ki_ctx;
	struct aio_ring	*ring;
Kent Overstreet's avatar
Kent Overstreet committed
1091
	struct io_event	*ev_page, *event;
1092
	unsigned tail, pos, head;
Linus Torvalds's avatar
Linus Torvalds committed
1093
1094
	unsigned long	flags;

1095
	if (!list_empty_careful(&iocb->ki_list)) {
1096
1097
1098
1099
1100
1101
		unsigned long flags;

		spin_lock_irqsave(&ctx->ctx_lock, flags);
		list_del(&iocb->ki_list);
		spin_unlock_irqrestore(&ctx->ctx_lock, flags);
	}
1102

1103
1104
	/*
	 * Add a completion event to the ring buffer. Must be done holding
1105
	 * ctx->completion_lock to prevent other code from messing with the tail
1106
1107
1108
1109
	 * pointer since we might be called from irq context.
	 */
	spin_lock_irqsave(&ctx->completion_lock, flags);

1110