aio.c 53.2 KB
Newer Older
Linus Torvalds's avatar
Linus Torvalds committed
1
2
3
4
5
6
7
/*
 *	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.
8
 *	Copyright 2018 Christoph Hellwig.
Linus Torvalds's avatar
Linus Torvalds committed
9
10
11
 *
 *	See ../COPYING for licensing terms.
 */
Kent Overstreet's avatar
Kent Overstreet committed
12
13
#define pr_fmt(fmt) "%s: " fmt, __func__

Linus Torvalds's avatar
Linus Torvalds committed
14
15
16
17
18
#include <linux/kernel.h>
#include <linux/init.h>
#include <linux/errno.h>
#include <linux/time.h>
#include <linux/aio_abi.h>
19
#include <linux/export.h>
Linus Torvalds's avatar
Linus Torvalds committed
20
#include <linux/syscalls.h>
21
#include <linux/backing-dev.h>
22
#include <linux/uio.h>
Linus Torvalds's avatar
Linus Torvalds committed
23

24
#include <linux/sched/signal.h>
Linus Torvalds's avatar
Linus Torvalds committed
25
26
27
28
#include <linux/fs.h>
#include <linux/file.h>
#include <linux/mm.h>
#include <linux/mman.h>
29
#include <linux/mmu_context.h>
Kent Overstreet's avatar
Kent Overstreet committed
30
#include <linux/percpu.h>
Linus Torvalds's avatar
Linus Torvalds committed
31
32
33
34
35
36
#include <linux/slab.h>
#include <linux/timer.h>
#include <linux/aio.h>
#include <linux/highmem.h>
#include <linux/workqueue.h>
#include <linux/security.h>
37
#include <linux/eventfd.h>
Jeff Moyer's avatar
Jeff Moyer committed
38
#include <linux/blkdev.h>
39
#include <linux/compat.h>
40
41
#include <linux/migrate.h>
#include <linux/ramfs.h>
Kent Overstreet's avatar
Kent Overstreet committed
42
#include <linux/percpu-refcount.h>
43
#include <linux/mount.h>
Linus Torvalds's avatar
Linus Torvalds committed
44
45

#include <asm/kmap_types.h>
46
#include <linux/uaccess.h>
Linus Torvalds's avatar
Linus Torvalds committed
47

Al Viro's avatar
Al Viro committed
48
49
#include "internal.h"

50
51
#define KIOCB_KEY		0

52
53
54
55
56
57
#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 */
58
59
	unsigned	head;	/* Written to by userland or under ring_lock
				 * mutex by aio_read_events_ring(). */
60
61
62
63
64
65
66
67
68
69
70
71
72
	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

73
struct kioctx_table {
74
75
76
	struct rcu_head		rcu;
	unsigned		nr;
	struct kioctx __rcu	*table[];
77
78
};

Kent Overstreet's avatar
Kent Overstreet committed
79
80
81
82
struct kioctx_cpu {
	unsigned		reqs_available;
};

83
84
85
86
87
struct ctx_rq_wait {
	struct completion comp;
	atomic_t count;
};

88
struct kioctx {
Kent Overstreet's avatar
Kent Overstreet committed
89
	struct percpu_ref	users;
Kent Overstreet's avatar
Kent Overstreet committed
90
	atomic_t		dead;
91

Kent Overstreet's avatar
Kent Overstreet committed
92
93
	struct percpu_ref	reqs;

94
95
	unsigned long		user_id;

Kent Overstreet's avatar
Kent Overstreet committed
96
97
98
99
100
101
102
	struct __percpu kioctx_cpu *cpu;

	/*
	 * For percpu reqs_available, number of slots we move to/from global
	 * counter at a time:
	 */
	unsigned		req_batch;
103
104
105
106
	/*
	 * This is what userspace passed to io_setup(), it's not used for
	 * anything but counting against the global max_reqs quota.
	 *
107
	 * The real limit is nr_events - 1, which will be larger (see
108
109
	 * aio_setup_ring())
	 */
110
111
	unsigned		max_reqs;

112
113
	/* Size of ringbuffer, in units of struct io_event */
	unsigned		nr_events;
114

115
116
117
118
119
120
	unsigned long		mmap_base;
	unsigned long		mmap_size;

	struct page		**ring_pages;
	long			nr_pages;

121
	struct rcu_work		free_rwork;	/* see free_ioctx() */
122

123
124
125
	/*
	 * signals when all in-flight requests are done
	 */
126
	struct ctx_rq_wait	*rq_wait;
127

128
	struct {
129
130
131
132
133
		/*
		 * 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
134
135
		 *
		 * We batch accesses to it with a percpu version.
136
137
		 */
		atomic_t	reqs_available;
138
139
140
141
142
143
144
	} ____cacheline_aligned_in_smp;

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

145
146
	struct {
		struct mutex	ring_lock;
147
148
		wait_queue_head_t wait;
	} ____cacheline_aligned_in_smp;
149
150
151

	struct {
		unsigned	tail;
152
		unsigned	completed_events;
153
		spinlock_t	completion_lock;
154
	} ____cacheline_aligned_in_smp;
155
156

	struct page		*internal_pages[AIO_RING_PAGES];
157
	struct file		*aio_ring_file;
158
159

	unsigned		id;
160
161
};

162
163
164
165
166
167
struct fsync_iocb {
	struct work_struct	work;
	struct file		*file;
	bool			datasync;
};

168
169
170
171
172
173
174
175
176
177
178
struct poll_iocb {
	struct file		*file;
	__poll_t		events;
	struct wait_queue_head	*head;

	union {
		struct wait_queue_entry	wait;
		struct work_struct	work;
	};
};

179
struct aio_kiocb {
180
181
	union {
		struct kiocb		rw;
182
		struct fsync_iocb	fsync;
183
		struct poll_iocb	poll;
184
	};
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201

	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
202
/*------ sysctl variables----*/
203
204
205
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
206
207
/*----end sysctl variables---*/

208
209
static struct kmem_cache	*kiocb_cachep;
static struct kmem_cache	*kioctx_cachep;
Linus Torvalds's avatar
Linus Torvalds committed
210

211
212
213
214
215
216
217
218
219
220
221
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);
222
223
	if (IS_ERR(inode))
		return ERR_CAST(inode);
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
251
252

	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
253
254
255
256
257
258
	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;
259
260
}

Linus Torvalds's avatar
Linus Torvalds committed
261
262
263
264
265
266
/* 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)
{
267
268
269
270
271
272
273
274
275
	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.");

276
	kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
277
	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds's avatar
Linus Torvalds committed
278
279
	return 0;
}
280
__initcall(aio_setup);
Linus Torvalds's avatar
Linus Torvalds committed
281

282
283
284
static void put_aio_ring_file(struct kioctx *ctx)
{
	struct file *aio_ring_file = ctx->aio_ring_file;
285
286
	struct address_space *i_mapping;

287
	if (aio_ring_file) {
Al Viro's avatar
Al Viro committed
288
		truncate_setsize(file_inode(aio_ring_file), 0);
289
290

		/* Prevent further access to the kioctx from migratepages */
Al Viro's avatar
Al Viro committed
291
		i_mapping = aio_ring_file->f_mapping;
292
293
		spin_lock(&i_mapping->private_lock);
		i_mapping->private_data = NULL;
294
		ctx->aio_ring_file = NULL;
295
		spin_unlock(&i_mapping->private_lock);
296
297
298
299
300

		fput(aio_ring_file);
	}
}

Linus Torvalds's avatar
Linus Torvalds committed
301
302
static void aio_free_ring(struct kioctx *ctx)
{
303
	int i;
Linus Torvalds's avatar
Linus Torvalds committed
304

305
306
307
308
309
	/* Disconnect the kiotx from the ring file.  This prevents future
	 * accesses to the kioctx from page migration.
	 */
	put_aio_ring_file(ctx);

310
	for (i = 0; i < ctx->nr_pages; i++) {
311
		struct page *page;
312
313
		pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
				page_count(ctx->ring_pages[i]));
314
315
316
317
318
		page = ctx->ring_pages[i];
		if (!page)
			continue;
		ctx->ring_pages[i] = NULL;
		put_page(page);
319
	}
Linus Torvalds's avatar
Linus Torvalds committed
320

321
	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
322
		kfree(ctx->ring_pages);
323
324
		ctx->ring_pages = NULL;
	}
325
326
}

327
static int aio_ring_mremap(struct vm_area_struct *vma)
328
{
329
	struct file *file = vma->vm_file;
330
331
	struct mm_struct *mm = vma->vm_mm;
	struct kioctx_table *table;
Al Viro's avatar
Al Viro committed
332
	int i, res = -EINVAL;
333
334
335
336
337
338
339

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

340
		ctx = rcu_dereference(table->table[i]);
341
		if (ctx && ctx->aio_ring_file == file) {
Al Viro's avatar
Al Viro committed
342
343
344
345
			if (!atomic_read(&ctx->dead)) {
				ctx->user_id = ctx->mmap_base = vma->vm_start;
				res = 0;
			}
346
347
348
349
350
351
			break;
		}
	}

	rcu_read_unlock();
	spin_unlock(&mm->ioctx_lock);
Al Viro's avatar
Al Viro committed
352
	return res;
353
354
}

355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
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;
}

371
372
373
374
static const struct file_operations aio_ring_fops = {
	.mmap = aio_ring_mmap,
};

375
#if IS_ENABLED(CONFIG_MIGRATION)
376
377
378
static int aio_migratepage(struct address_space *mapping, struct page *new,
			struct page *old, enum migrate_mode mode)
{
379
	struct kioctx *ctx;
380
	unsigned long flags;
381
	pgoff_t idx;
382
383
	int rc;

384
385
386
387
388
389
390
391
	/*
	 * 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;

392
393
	rc = 0;

394
	/* mapping->private_lock here protects against the kioctx teardown.  */
395
396
	spin_lock(&mapping->private_lock);
	ctx = mapping->private_data;
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
	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;
416
417
418
419
	} else
		rc = -EINVAL;

	if (rc != 0)
420
		goto out_unlock;
421

422
423
	/* Writeback must be complete */
	BUG_ON(PageWriteback(old));
424
	get_page(new);
425

426
	rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
427
	if (rc != MIGRATEPAGE_SUCCESS) {
428
		put_page(new);
429
		goto out_unlock;
430
431
	}

432
433
434
	/* 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.
435
	 */
436
437
438
439
440
	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);
441

442
443
	/* The old page is no longer accessible. */
	put_page(old);
444

445
446
447
448
out_unlock:
	mutex_unlock(&ctx->ring_lock);
out:
	spin_unlock(&mapping->private_lock);
449
	return rc;
Linus Torvalds's avatar
Linus Torvalds committed
450
}
451
#endif
Linus Torvalds's avatar
Linus Torvalds committed
452

453
static const struct address_space_operations aio_ctx_aops = {
454
	.set_page_dirty = __set_page_dirty_no_writeback,
455
#if IS_ENABLED(CONFIG_MIGRATION)
456
	.migratepage	= aio_migratepage,
457
#endif
458
459
};

460
static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
Linus Torvalds's avatar
Linus Torvalds committed
461
462
{
	struct aio_ring *ring;
Zach Brown's avatar
Zach Brown committed
463
	struct mm_struct *mm = current->mm;
464
	unsigned long size, unused;
Linus Torvalds's avatar
Linus Torvalds committed
465
	int nr_pages;
466
467
	int i;
	struct file *file;
Linus Torvalds's avatar
Linus Torvalds committed
468
469
470
471
472
473
474

	/* 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;

475
	nr_pages = PFN_UP(size);
Linus Torvalds's avatar
Linus Torvalds committed
476
477
478
	if (nr_pages < 0)
		return -EINVAL;

479
	file = aio_private_file(ctx, nr_pages);
480
481
	if (IS_ERR(file)) {
		ctx->aio_ring_file = NULL;
482
		return -ENOMEM;
483
484
	}

485
486
487
488
489
490
491
492
493
494
495
496
497
498
	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;
		}
	}

499
500
	for (i = 0; i < nr_pages; i++) {
		struct page *page;
Al Viro's avatar
Al Viro committed
501
		page = find_or_create_page(file->f_mapping,
502
503
504
505
506
507
508
					   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);
509
510

		ctx->ring_pages[i] = page;
511
	}
512
	ctx->nr_pages = i;
Linus Torvalds's avatar
Linus Torvalds committed
513

514
515
	if (unlikely(i != nr_pages)) {
		aio_free_ring(ctx);
516
		return -ENOMEM;
Linus Torvalds's avatar
Linus Torvalds committed
517
518
	}

519
520
	ctx->mmap_size = nr_pages * PAGE_SIZE;
	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
521

522
523
524
525
526
527
	if (down_write_killable(&mm->mmap_sem)) {
		ctx->mmap_size = 0;
		aio_free_ring(ctx);
		return -EINTR;
	}

528
529
	ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
				       PROT_READ | PROT_WRITE,
530
				       MAP_SHARED, 0, &unused, NULL);
531
	up_write(&mm->mmap_sem);
532
533
	if (IS_ERR((void *)ctx->mmap_base)) {
		ctx->mmap_size = 0;
Linus Torvalds's avatar
Linus Torvalds committed
534
		aio_free_ring(ctx);
535
		return -ENOMEM;
Linus Torvalds's avatar
Linus Torvalds committed
536
537
	}

538
	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
539

540
541
	ctx->user_id = ctx->mmap_base;
	ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds's avatar
Linus Torvalds committed
542

543
	ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds's avatar
Linus Torvalds committed
544
	ring->nr = nr_events;	/* user copy */
545
	ring->id = ~0U;
Linus Torvalds's avatar
Linus Torvalds committed
546
547
548
549
550
	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);
551
	kunmap_atomic(ring);
552
	flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds's avatar
Linus Torvalds committed
553
554
555
556
557
558
559
560

	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)

561
void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
562
{
563
	struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
564
565
566
	struct kioctx *ctx = req->ki_ctx;
	unsigned long flags;

567
568
	if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
		return;
569

570
571
	spin_lock_irqsave(&ctx->ctx_lock, flags);
	list_add_tail(&req->ki_list, &ctx->active_reqs);
572
573
574
575
576
	req->ki_cancel = cancel;
	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
}
EXPORT_SYMBOL(kiocb_set_cancel_fn);

577
578
579
/*
 * free_ioctx() should be RCU delayed to synchronize against the RCU
 * protected lookup_ioctx() and also needs process context to call
580
 * aio_free_ring().  Use rcu_work.
581
 */
Kent Overstreet's avatar
Kent Overstreet committed
582
static void free_ioctx(struct work_struct *work)
Kent Overstreet's avatar
Kent Overstreet committed
583
{
584
585
	struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
					  free_rwork);
Kent Overstreet's avatar
Kent Overstreet committed
586
	pr_debug("freeing %p\n", ctx);
Kent Overstreet's avatar
Kent Overstreet committed
587

Kent Overstreet's avatar
Kent Overstreet committed
588
	aio_free_ring(ctx);
Kent Overstreet's avatar
Kent Overstreet committed
589
	free_percpu(ctx->cpu);
590
591
	percpu_ref_exit(&ctx->reqs);
	percpu_ref_exit(&ctx->users);
Kent Overstreet's avatar
Kent Overstreet committed
592
593
594
	kmem_cache_free(kioctx_cachep, ctx);
}

Kent Overstreet's avatar
Kent Overstreet committed
595
596
597
598
static void free_ioctx_reqs(struct percpu_ref *ref)
{
	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);

599
	/* At this point we know that there are no any in-flight requests */
600
601
	if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
		complete(&ctx->rq_wait->comp);
602

603
	/* Synchronize against RCU protected table->table[] dereferences */
604
605
	INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
	queue_rcu_work(system_wq, &ctx->free_rwork);
Kent Overstreet's avatar
Kent Overstreet committed
606
607
}

Kent Overstreet's avatar
Kent Overstreet committed
608
609
610
611
612
/*
 * 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
613
static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet's avatar
Kent Overstreet committed
614
{
Kent Overstreet's avatar
Kent Overstreet committed
615
	struct kioctx *ctx = container_of(ref, struct kioctx, users);
616
	struct aio_kiocb *req;
Kent Overstreet's avatar
Kent Overstreet committed
617
618
619
620
621

	spin_lock_irq(&ctx->ctx_lock);

	while (!list_empty(&ctx->active_reqs)) {
		req = list_first_entry(&ctx->active_reqs,
622
				       struct aio_kiocb, ki_list);
Christoph Hellwig's avatar
Christoph Hellwig committed
623
		req->ki_cancel(&req->rw);
Al Viro's avatar
Al Viro committed
624
		list_del_init(&req->ki_list);
Kent Overstreet's avatar
Kent Overstreet committed
625
626
627
628
	}

	spin_unlock_irq(&ctx->ctx_lock);

Kent Overstreet's avatar
Kent Overstreet committed
629
630
	percpu_ref_kill(&ctx->reqs);
	percpu_ref_put(&ctx->reqs);
Kent Overstreet's avatar
Kent Overstreet committed
631
632
}

633
634
635
636
637
638
639
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);
640
	table = rcu_dereference_raw(mm->ioctx_table);
641
642
643
644

	while (1) {
		if (table)
			for (i = 0; i < table->nr; i++)
645
				if (!rcu_access_pointer(table->table[i])) {
646
					ctx->id = i;
647
					rcu_assign_pointer(table->table[i], ctx);
648
649
					spin_unlock(&mm->ioctx_lock);

650
651
652
653
					/* While kioctx setup is in progress,
					 * we are protected from page migration
					 * changes ring_pages by ->ring_lock.
					 */
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
					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);
671
		old = rcu_dereference_raw(mm->ioctx_table);
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687

		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
688
689
690
691
692
693
694
695
696
697
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
698
699
700
701
702
/* 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
703
	struct mm_struct *mm = current->mm;
Linus Torvalds's avatar
Linus Torvalds committed
704
	struct kioctx *ctx;
705
	int err = -ENOMEM;
Linus Torvalds's avatar
Linus Torvalds committed
706

707
708
709
710
711
712
	/*
	 * 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
713
714
715
716
717
718
719
720
721
722
723
724
	/*
	 * 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
725
	/* Prevent overflows */
726
	if (nr_events > (0x10000000U / sizeof(struct io_event))) {
Linus Torvalds's avatar
Linus Torvalds committed
727
728
729
730
		pr_debug("ENOMEM: nr_events too high\n");
		return ERR_PTR(-EINVAL);
	}

731
	if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
Linus Torvalds's avatar
Linus Torvalds committed
732
733
		return ERR_PTR(-EAGAIN);

734
	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds's avatar
Linus Torvalds committed
735
736
737
	if (!ctx)
		return ERR_PTR(-ENOMEM);

738
	ctx->max_reqs = max_reqs;
Linus Torvalds's avatar
Linus Torvalds committed
739
740

	spin_lock_init(&ctx->ctx_lock);
741
	spin_lock_init(&ctx->completion_lock);
742
	mutex_init(&ctx->ring_lock);
743
744
745
	/* 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
746
747
748
749
	init_waitqueue_head(&ctx->wait);

	INIT_LIST_HEAD(&ctx->active_reqs);

750
	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
751
752
		goto err;

753
	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
754
755
		goto err;

Kent Overstreet's avatar
Kent Overstreet committed
756
757
	ctx->cpu = alloc_percpu(struct kioctx_cpu);
	if (!ctx->cpu)
Kent Overstreet's avatar
Kent Overstreet committed
758
		goto err;
Linus Torvalds's avatar
Linus Torvalds committed
759

760
	err = aio_setup_ring(ctx, nr_events);
761
	if (err < 0)
Kent Overstreet's avatar
Kent Overstreet committed
762
		goto err;
Kent Overstreet's avatar
Kent Overstreet committed
763

764
	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreet's avatar
Kent Overstreet committed
765
	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
766
767
	if (ctx->req_batch < 1)
		ctx->req_batch = 1;
768

Linus Torvalds's avatar
Linus Torvalds committed
769
	/* limit the number of system wide aios */
770
	spin_lock(&aio_nr_lock);
771
772
	if (aio_nr + ctx->max_reqs > aio_max_nr ||
	    aio_nr + ctx->max_reqs < aio_nr) {
773
		spin_unlock(&aio_nr_lock);
Kent Overstreet's avatar
Kent Overstreet committed
774
		err = -EAGAIN;
775
		goto err_ctx;
776
777
	}
	aio_nr += ctx->max_reqs;
778
	spin_unlock(&aio_nr_lock);
Linus Torvalds's avatar
Linus Torvalds committed
779

780
781
	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
782

783
784
	err = ioctx_add_table(ctx, mm);
	if (err)
Kent Overstreet's avatar
Kent Overstreet committed
785
		goto err_cleanup;
786

787
788
789
	/* Release the ring_lock mutex now that all setup is complete. */
	mutex_unlock(&ctx->ring_lock);

Kent Overstreet's avatar
Kent Overstreet committed
790
	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
791
		 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds's avatar
Linus Torvalds committed
792
793
	return ctx;

Kent Overstreet's avatar
Kent Overstreet committed
794
795
err_cleanup:
	aio_nr_sub(ctx->max_reqs);
796
err_ctx:
797
798
799
	atomic_set(&ctx->dead, 1);
	if (ctx->mmap_size)
		vm_munmap(ctx->mmap_base, ctx->mmap_size);
800
	aio_free_ring(ctx);
Kent Overstreet's avatar
Kent Overstreet committed
801
err:
802
	mutex_unlock(&ctx->ring_lock);
Kent Overstreet's avatar
Kent Overstreet committed
803
	free_percpu(ctx->cpu);
804
805
	percpu_ref_exit(&ctx->reqs);
	percpu_ref_exit(&ctx->users);
Linus Torvalds's avatar
Linus Torvalds committed
806
	kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreet's avatar
Kent Overstreet committed
807
	pr_debug("error allocating ioctx %d\n", err);
808
	return ERR_PTR(err);
Linus Torvalds's avatar
Linus Torvalds committed
809
810
}

Kent Overstreet's avatar
Kent Overstreet committed
811
812
813
814
815
/* 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.
 */
816
static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
817
		      struct ctx_rq_wait *wait)
Kent Overstreet's avatar
Kent Overstreet committed
818
{
819
	struct kioctx_table *table;
820

Al Viro's avatar
Al Viro committed
821
822
823
	spin_lock(&mm->ioctx_lock);
	if (atomic_xchg(&ctx->dead, 1)) {
		spin_unlock(&mm->ioctx_lock);
824
		return -EINVAL;
Al Viro's avatar
Al Viro committed
825
	}
826

827
	table = rcu_dereference_raw(mm->ioctx_table);
828
829
	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
	RCU_INIT_POINTER(table->table[ctx->id], NULL);
830
	spin_unlock(&mm->ioctx_lock);
831

832
	/* free_ioctx_reqs() will do the necessary RCU synchronization */
833
	wake_up_all(&ctx->wait);
834

835
836
837
838
839
840
841
842
	/*
	 * 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);
843

844
845
	if (ctx->mmap_size)
		vm_munmap(ctx->mmap_base, ctx->mmap_size);
846

847
	ctx->rq_wait = wait;
848
849
	percpu_ref_kill(&ctx->users);
	return 0;
Linus Torvalds's avatar
Linus Torvalds committed
850
851
}

Kent Overstreet's avatar
Kent Overstreet committed
852
853
854
855
856
857
858
/*
 * 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
859
 */
860
void exit_aio(struct mm_struct *mm)
Linus Torvalds's avatar
Linus Torvalds committed
861
{
862
	struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
863
864
	struct ctx_rq_wait wait;
	int i, skipped;
865

866
867
	if (!table)
		return;
868

869
870
871
872
	atomic_set(&wait.count, table->nr);
	init_completion(&wait.comp);

	skipped = 0;
873
	for (i = 0; i < table->nr; ++i) {
874
875
		struct kioctx *ctx =
			rcu_dereference_protected(table->table[i], true);
876

877
878
		if (!ctx) {
			skipped++;
879
			continue;
880
881
		}

882
		/*
883
884
885
886
887
		 * 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.
888
		 */
889
		ctx->mmap_size = 0;
890
891
		kill_ioctx(mm, ctx, &wait);
	}
Kent Overstreet's avatar
Kent Overstreet committed
892

893
	if (!atomic_sub_and_test(skipped, &wait.count)) {
894
		/* Wait until all IO for the context are done. */
895
		wait_for_completion(&wait.comp);
Linus Torvalds's avatar
Linus Torvalds committed
896
	}
897
898
899

	RCU_INIT_POINTER(mm->ioctx_table, NULL);
	kfree(table);
Linus Torvalds's avatar
Linus Torvalds committed
900
901
}

Kent Overstreet's avatar
Kent Overstreet committed
902
903
904
static void put_reqs_available(struct kioctx *ctx, unsigned nr)
{
	struct kioctx_cpu *kcpu;
905
	unsigned long flags;
Kent Overstreet's avatar
Kent Overstreet committed
906

907
	local_irq_save(flags);
908
	kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreet's avatar
Kent Overstreet committed
909
	kcpu->reqs_available += nr;
910

Kent Overstreet's avatar
Kent Overstreet committed
911
912
913
914
915
	while (kcpu->reqs_available >= ctx->req_batch * 2) {
		kcpu->reqs_available -= ctx->req_batch;
		atomic_add(ctx->req_batch, &ctx->reqs_available);
	}

916
	local_irq_restore(flags);
Kent Overstreet's avatar
Kent Overstreet committed
917
918
919
920
921
922
}

static bool get_reqs_available(struct kioctx *ctx)
{
	struct kioctx_cpu *kcpu;
	bool ret = false;
923
	unsigned long flags;
Kent Overstreet's avatar
Kent Overstreet committed
924

925
	local_irq_save(flags);
926
	kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreet's avatar
Kent Overstreet committed
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
	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:
945
	local_irq_restore(flags);
Kent Overstreet's avatar
Kent Overstreet committed
946
947
948
	return ret;
}

949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
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
/* 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
1011
/* aio_get_req
Kent Overstreet's avatar
Kent Overstreet committed
1012
1013
 *	Allocate a slot for an aio request.
 * Returns NULL if no requests are free.
Linus Torvalds's avatar
Linus Torvalds committed
1014
 */
1015
static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds's avatar
Linus Torvalds committed
1016
{
1017
	struct aio_kiocb *req;
Kent Overstreet's avatar
Kent Overstreet committed
1018

1019
1020
1021
1022
1023
	if (!get_reqs_available(ctx)) {
		user_refill_reqs_available(ctx);
		if (!get_reqs_available(ctx))
			return NULL;
	}
Kent Overstreet's avatar
Kent Overstreet committed
1024

1025
	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
Linus Torvalds's avatar
Linus Torvalds committed
1026
	if (unlikely(!req))
Kent Overstreet's avatar
Kent Overstreet committed
1027
		goto out_put;
Linus Torvalds's avatar
Linus Torvalds committed
1028

Kent Overstreet's avatar
Kent Overstreet committed
1029
	percpu_ref_get(&ctx->reqs);
1030
	INIT_LIST_HEAD(&req->ki_list);
Linus Torvalds's avatar
Linus Torvalds committed
1031
	req->ki_ctx = ctx;
Jeff Moyer's avatar
Jeff Moyer committed
1032
	return req;
Kent Overstreet's avatar
Kent Overstreet committed
1033
out_put:
Kent Overstreet's avatar
Kent Overstreet committed
1034
	put_reqs_available(ctx, 1);
Kent Overstreet's avatar
Kent Overstreet committed
1035
	return NULL;
Linus Torvalds's avatar
Linus Torvalds committed
1036
1037
}

1038
static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds's avatar
Linus Torvalds committed
1039
{
1040
	struct aio_ring __user *ring  = (void __user *)ctx_id;
1041
	struct mm_struct *mm = current->mm;
1042
	struct kioctx *ctx, *ret = NULL;
1043
1044
1045
1046
1047
	struct kioctx_table *table;
	unsigned id;

	if (get_user(id, &ring->id))
		return NULL;
Linus Torvalds's avatar
Linus Torvalds committed
1048

1049
	rcu_read_lock();
1050
	table = rcu_dereference(mm->ioctx_table);
1051

1052
1053
	if (!table || id >= table->nr)
		goto out;
Linus Torvalds's avatar
Linus Torvalds committed
1054

1055
	ctx = rcu_dereference(table->table[id]);
1056
	if (ctx && ctx->user_id == ctx_id) {
1057
1058
		if (percpu_ref_tryget_live(&ctx->users))
			ret = ctx;
1059
1060
	}
out:
1061
	rcu_read_unlock();
1062
	return ret;
Linus Torvalds's avatar
Linus Torvalds committed
1063
1064
1065
1066
1067
}

/* aio_complete
 *	Called when the io request on the given iocb is complete.
 */
1068
static void aio_complete(struct aio_kiocb *iocb, long res, long res2)
Linus Torvalds's avatar
Linus Torvalds committed
1069
1070
1071
{
	struct kioctx	*ctx = iocb->ki_ctx;
	struct aio_ring	*ring;
Kent Overstreet's avatar
Kent Overstreet committed
1072
	struct io_event	*ev_page, *event;
1073
	unsigned tail, pos, head;
Linus Torvalds's avatar
Linus Torvalds committed
1074
1075
	unsigned long	flags;

1076
1077
	/*
	 * Add a completion event to the ring buffer. Must be done holding
1078
	 * ctx->completion_lock to prevent other code from messing with the tail
1079
1080
1081
1082
	 * pointer since we might be called from irq context.
	 */
	spin_lock_irqsave(&ctx->completion_lock, flags);

1083
	tail = ctx->tail;
Kent Overstreet's avatar
Kent Overstreet committed
1084
1085
	pos = tail + AIO_EVENTS_OFFSET;

1086
	if (++tail >= ctx->nr_events)
1087
		tail = 0;
Linus Torvalds's avatar
Linus Torvalds committed
1088

1089
	ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet's avatar
Kent Overstreet committed
1090
1091
	event = ev_page + pos % AIO_EVENTS_PER_PAGE;

1092
	event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
Linus Torvalds's avatar
Linus Torvalds committed
1093
1094
1095
1096
	event->data = iocb->ki_user_data;
	event->res = res;
	event->res2 = res2;

Kent Overstreet's avatar
Kent Overstreet committed
1097
	kunmap_atomic(ev_page);
1098
	flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet's avatar
Kent Overstreet committed
1099
1100

	pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1101
		 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
Kent Overstreet's avatar
Kent Overstreet committed