aio.c 54.5 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>
Christoph Hellwig's avatar
Christoph Hellwig committed
22
#include <linux/refcount.h>
23
#include <linux/uio.h>
Linus Torvalds's avatar
Linus Torvalds committed
24

25
#include <linux/sched/signal.h>
Linus Torvalds's avatar
Linus Torvalds committed
26
27
28
29
#include <linux/fs.h>
#include <linux/file.h>
#include <linux/mm.h>
#include <linux/mman.h>
30
#include <linux/mmu_context.h>
Kent Overstreet's avatar
Kent Overstreet committed
31
#include <linux/percpu.h>
Linus Torvalds's avatar
Linus Torvalds committed
32
33
34
35
36
37
#include <linux/slab.h>
#include <linux/timer.h>
#include <linux/aio.h>
#include <linux/highmem.h>
#include <linux/workqueue.h>
#include <linux/security.h>
38
#include <linux/eventfd.h>
Jeff Moyer's avatar
Jeff Moyer committed
39
#include <linux/blkdev.h>
40
#include <linux/compat.h>
41
42
#include <linux/migrate.h>
#include <linux/ramfs.h>
Kent Overstreet's avatar
Kent Overstreet committed
43
#include <linux/percpu-refcount.h>
44
#include <linux/mount.h>
Linus Torvalds's avatar
Linus Torvalds committed
45
46

#include <asm/kmap_types.h>
47
#include <linux/uaccess.h>
48
#include <linux/nospec.h>
Linus Torvalds's avatar
Linus Torvalds committed
49

Al Viro's avatar
Al Viro committed
50
51
#include "internal.h"

52
53
#define KIOCB_KEY		0

54
55
56
57
58
59
#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 */
60
61
	unsigned	head;	/* Written to by userland or under ring_lock
				 * mutex by aio_read_events_ring(). */
62
63
64
65
66
67
68
69
70
71
72
73
74
	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

75
struct kioctx_table {
76
77
78
	struct rcu_head		rcu;
	unsigned		nr;
	struct kioctx __rcu	*table[];
79
80
};

Kent Overstreet's avatar
Kent Overstreet committed
81
82
83
84
struct kioctx_cpu {
	unsigned		reqs_available;
};

85
86
87
88
89
struct ctx_rq_wait {
	struct completion comp;
	atomic_t count;
};

90
struct kioctx {
Kent Overstreet's avatar
Kent Overstreet committed
91
	struct percpu_ref	users;
Kent Overstreet's avatar
Kent Overstreet committed
92
	atomic_t		dead;
93

Kent Overstreet's avatar
Kent Overstreet committed
94
95
	struct percpu_ref	reqs;

96
97
	unsigned long		user_id;

Kent Overstreet's avatar
Kent Overstreet committed
98
99
100
101
102
103
104
	struct __percpu kioctx_cpu *cpu;

	/*
	 * For percpu reqs_available, number of slots we move to/from global
	 * counter at a time:
	 */
	unsigned		req_batch;
105
106
107
108
	/*
	 * This is what userspace passed to io_setup(), it's not used for
	 * anything but counting against the global max_reqs quota.
	 *
109
	 * The real limit is nr_events - 1, which will be larger (see
110
111
	 * aio_setup_ring())
	 */
112
113
	unsigned		max_reqs;

114
115
	/* Size of ringbuffer, in units of struct io_event */
	unsigned		nr_events;
116

117
118
119
120
121
122
	unsigned long		mmap_base;
	unsigned long		mmap_size;

	struct page		**ring_pages;
	long			nr_pages;

123
	struct rcu_work		free_rwork;	/* see free_ioctx() */
124

125
126
127
	/*
	 * signals when all in-flight requests are done
	 */
128
	struct ctx_rq_wait	*rq_wait;
129

130
	struct {
131
132
133
134
135
		/*
		 * 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
136
137
		 *
		 * We batch accesses to it with a percpu version.
138
139
		 */
		atomic_t	reqs_available;
140
141
142
143
144
145
146
	} ____cacheline_aligned_in_smp;

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

147
148
	struct {
		struct mutex	ring_lock;
149
150
		wait_queue_head_t wait;
	} ____cacheline_aligned_in_smp;
151
152
153

	struct {
		unsigned	tail;
154
		unsigned	completed_events;
155
		spinlock_t	completion_lock;
156
	} ____cacheline_aligned_in_smp;
157
158

	struct page		*internal_pages[AIO_RING_PAGES];
159
	struct file		*aio_ring_file;
160
161

	unsigned		id;
162
163
};

164
165
166
167
168
169
struct fsync_iocb {
	struct work_struct	work;
	struct file		*file;
	bool			datasync;
};

170
171
172
173
174
175
176
177
178
179
struct poll_iocb {
	struct file		*file;
	struct wait_queue_head	*head;
	__poll_t		events;
	bool			woken;
	bool			cancelled;
	struct wait_queue_entry	wait;
	struct work_struct	work;
};

180
struct aio_kiocb {
181
182
	union {
		struct kiocb		rw;
183
		struct fsync_iocb	fsync;
184
		struct poll_iocb	poll;
185
	};
186
187
188
189
190
191
192
193
194

	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 */
Christoph Hellwig's avatar
Christoph Hellwig committed
195
	refcount_t		ki_refcnt;
196
197
198
199
200
201
202
203

	/*
	 * 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
204
/*------ sysctl variables----*/
205
206
207
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
208
209
/*----end sysctl variables---*/

210
211
static struct kmem_cache	*kiocb_cachep;
static struct kmem_cache	*kioctx_cachep;
Linus Torvalds's avatar
Linus Torvalds committed
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 file *file;
	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

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

Al Viro's avatar
Al Viro committed
229
230
	file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
				O_RDWR, &aio_ring_fops);
231
	if (IS_ERR(file))
232
233
234
235
236
237
238
		iput(inode);
	return file;
}

static struct dentry *aio_mount(struct file_system_type *fs_type,
				int flags, const char *dev_name, void *data)
{
Al Viro's avatar
Al Viro committed
239
	struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, NULL,
Jann Horn's avatar
Jann Horn committed
240
241
242
243
244
					   AIO_RING_MAGIC);

	if (!IS_ERR(root))
		root->d_sb->s_iflags |= SB_I_NOEXEC;
	return root;
245
246
}

Linus Torvalds's avatar
Linus Torvalds committed
247
248
249
250
251
252
/* 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)
{
253
254
255
256
257
258
259
260
261
	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.");

262
	kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
263
	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds's avatar
Linus Torvalds committed
264
265
	return 0;
}
266
__initcall(aio_setup);
Linus Torvalds's avatar
Linus Torvalds committed
267

268
269
270
static void put_aio_ring_file(struct kioctx *ctx)
{
	struct file *aio_ring_file = ctx->aio_ring_file;
271
272
	struct address_space *i_mapping;

273
	if (aio_ring_file) {
Al Viro's avatar
Al Viro committed
274
		truncate_setsize(file_inode(aio_ring_file), 0);
275
276

		/* Prevent further access to the kioctx from migratepages */
Al Viro's avatar
Al Viro committed
277
		i_mapping = aio_ring_file->f_mapping;
278
279
		spin_lock(&i_mapping->private_lock);
		i_mapping->private_data = NULL;
280
		ctx->aio_ring_file = NULL;
281
		spin_unlock(&i_mapping->private_lock);
282
283
284
285
286

		fput(aio_ring_file);
	}
}

Linus Torvalds's avatar
Linus Torvalds committed
287
288
static void aio_free_ring(struct kioctx *ctx)
{
289
	int i;
Linus Torvalds's avatar
Linus Torvalds committed
290

291
292
293
294
295
	/* Disconnect the kiotx from the ring file.  This prevents future
	 * accesses to the kioctx from page migration.
	 */
	put_aio_ring_file(ctx);

296
	for (i = 0; i < ctx->nr_pages; i++) {
297
		struct page *page;
298
299
		pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
				page_count(ctx->ring_pages[i]));
300
301
302
303
304
		page = ctx->ring_pages[i];
		if (!page)
			continue;
		ctx->ring_pages[i] = NULL;
		put_page(page);
305
	}
Linus Torvalds's avatar
Linus Torvalds committed
306

307
	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
308
		kfree(ctx->ring_pages);
309
310
		ctx->ring_pages = NULL;
	}
311
312
}

313
static int aio_ring_mremap(struct vm_area_struct *vma)
314
{
315
	struct file *file = vma->vm_file;
316
317
	struct mm_struct *mm = vma->vm_mm;
	struct kioctx_table *table;
Al Viro's avatar
Al Viro committed
318
	int i, res = -EINVAL;
319
320
321
322
323
324
325

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

326
		ctx = rcu_dereference(table->table[i]);
327
		if (ctx && ctx->aio_ring_file == file) {
Al Viro's avatar
Al Viro committed
328
329
330
331
			if (!atomic_read(&ctx->dead)) {
				ctx->user_id = ctx->mmap_base = vma->vm_start;
				res = 0;
			}
332
333
334
335
336
337
			break;
		}
	}

	rcu_read_unlock();
	spin_unlock(&mm->ioctx_lock);
Al Viro's avatar
Al Viro committed
338
	return res;
339
340
}

341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
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;
}

357
358
359
360
static const struct file_operations aio_ring_fops = {
	.mmap = aio_ring_mmap,
};

361
#if IS_ENABLED(CONFIG_MIGRATION)
362
363
364
static int aio_migratepage(struct address_space *mapping, struct page *new,
			struct page *old, enum migrate_mode mode)
{
365
	struct kioctx *ctx;
366
	unsigned long flags;
367
	pgoff_t idx;
368
369
	int rc;

370
371
372
373
374
375
376
377
	/*
	 * 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;

378
379
	rc = 0;

380
	/* mapping->private_lock here protects against the kioctx teardown.  */
381
382
	spin_lock(&mapping->private_lock);
	ctx = mapping->private_data;
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
	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;
402
403
404
405
	} else
		rc = -EINVAL;

	if (rc != 0)
406
		goto out_unlock;
407

408
409
	/* Writeback must be complete */
	BUG_ON(PageWriteback(old));
410
	get_page(new);
411

412
	rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
413
	if (rc != MIGRATEPAGE_SUCCESS) {
414
		put_page(new);
415
		goto out_unlock;
416
417
	}

418
419
420
	/* 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.
421
	 */
422
423
424
425
426
	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);
427

428
429
	/* The old page is no longer accessible. */
	put_page(old);
430

431
432
433
434
out_unlock:
	mutex_unlock(&ctx->ring_lock);
out:
	spin_unlock(&mapping->private_lock);
435
	return rc;
Linus Torvalds's avatar
Linus Torvalds committed
436
}
437
#endif
Linus Torvalds's avatar
Linus Torvalds committed
438

439
static const struct address_space_operations aio_ctx_aops = {
440
	.set_page_dirty = __set_page_dirty_no_writeback,
441
#if IS_ENABLED(CONFIG_MIGRATION)
442
	.migratepage	= aio_migratepage,
443
#endif
444
445
};

446
static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
Linus Torvalds's avatar
Linus Torvalds committed
447
448
{
	struct aio_ring *ring;
Zach Brown's avatar
Zach Brown committed
449
	struct mm_struct *mm = current->mm;
450
	unsigned long size, unused;
Linus Torvalds's avatar
Linus Torvalds committed
451
	int nr_pages;
452
453
	int i;
	struct file *file;
Linus Torvalds's avatar
Linus Torvalds committed
454
455
456
457
458
459
460

	/* 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;

461
	nr_pages = PFN_UP(size);
Linus Torvalds's avatar
Linus Torvalds committed
462
463
464
	if (nr_pages < 0)
		return -EINVAL;

465
	file = aio_private_file(ctx, nr_pages);
466
467
	if (IS_ERR(file)) {
		ctx->aio_ring_file = NULL;
468
		return -ENOMEM;
469
470
	}

471
472
473
474
475
476
477
478
479
480
481
482
483
484
	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;
		}
	}

485
486
	for (i = 0; i < nr_pages; i++) {
		struct page *page;
Al Viro's avatar
Al Viro committed
487
		page = find_or_create_page(file->f_mapping,
488
489
490
491
492
493
494
					   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);
495
496

		ctx->ring_pages[i] = page;
497
	}
498
	ctx->nr_pages = i;
Linus Torvalds's avatar
Linus Torvalds committed
499

500
501
	if (unlikely(i != nr_pages)) {
		aio_free_ring(ctx);
502
		return -ENOMEM;
Linus Torvalds's avatar
Linus Torvalds committed
503
504
	}

505
506
	ctx->mmap_size = nr_pages * PAGE_SIZE;
	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
507

508
509
510
511
512
513
	if (down_write_killable(&mm->mmap_sem)) {
		ctx->mmap_size = 0;
		aio_free_ring(ctx);
		return -EINTR;
	}

514
515
	ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
				       PROT_READ | PROT_WRITE,
516
				       MAP_SHARED, 0, &unused, NULL);
517
	up_write(&mm->mmap_sem);
518
519
	if (IS_ERR((void *)ctx->mmap_base)) {
		ctx->mmap_size = 0;
Linus Torvalds's avatar
Linus Torvalds committed
520
		aio_free_ring(ctx);
521
		return -ENOMEM;
Linus Torvalds's avatar
Linus Torvalds committed
522
523
	}

524
	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
525

526
527
	ctx->user_id = ctx->mmap_base;
	ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds's avatar
Linus Torvalds committed
528

529
	ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds's avatar
Linus Torvalds committed
530
	ring->nr = nr_events;	/* user copy */
531
	ring->id = ~0U;
Linus Torvalds's avatar
Linus Torvalds committed
532
533
534
535
536
	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);
537
	kunmap_atomic(ring);
538
	flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds's avatar
Linus Torvalds committed
539
540
541
542
543
544
545
546

	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)

547
void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
548
{
549
	struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
550
551
552
	struct kioctx *ctx = req->ki_ctx;
	unsigned long flags;

553
554
	if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
		return;
555

556
557
	spin_lock_irqsave(&ctx->ctx_lock, flags);
	list_add_tail(&req->ki_list, &ctx->active_reqs);
558
559
560
561
562
	req->ki_cancel = cancel;
	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
}
EXPORT_SYMBOL(kiocb_set_cancel_fn);

563
564
565
/*
 * free_ioctx() should be RCU delayed to synchronize against the RCU
 * protected lookup_ioctx() and also needs process context to call
566
 * aio_free_ring().  Use rcu_work.
567
 */
Kent Overstreet's avatar
Kent Overstreet committed
568
static void free_ioctx(struct work_struct *work)
Kent Overstreet's avatar
Kent Overstreet committed
569
{
570
571
	struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
					  free_rwork);
Kent Overstreet's avatar
Kent Overstreet committed
572
	pr_debug("freeing %p\n", ctx);
Kent Overstreet's avatar
Kent Overstreet committed
573

Kent Overstreet's avatar
Kent Overstreet committed
574
	aio_free_ring(ctx);
Kent Overstreet's avatar
Kent Overstreet committed
575
	free_percpu(ctx->cpu);
576
577
	percpu_ref_exit(&ctx->reqs);
	percpu_ref_exit(&ctx->users);
Kent Overstreet's avatar
Kent Overstreet committed
578
579
580
	kmem_cache_free(kioctx_cachep, ctx);
}

Kent Overstreet's avatar
Kent Overstreet committed
581
582
583
584
static void free_ioctx_reqs(struct percpu_ref *ref)
{
	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);

585
	/* At this point we know that there are no any in-flight requests */
586
587
	if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
		complete(&ctx->rq_wait->comp);
588

589
	/* Synchronize against RCU protected table->table[] dereferences */
590
591
	INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
	queue_rcu_work(system_wq, &ctx->free_rwork);
Kent Overstreet's avatar
Kent Overstreet committed
592
593
}

Kent Overstreet's avatar
Kent Overstreet committed
594
595
596
597
598
/*
 * 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
599
static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet's avatar
Kent Overstreet committed
600
{
Kent Overstreet's avatar
Kent Overstreet committed
601
	struct kioctx *ctx = container_of(ref, struct kioctx, users);
602
	struct aio_kiocb *req;
Kent Overstreet's avatar
Kent Overstreet committed
603
604
605
606
607

	spin_lock_irq(&ctx->ctx_lock);

	while (!list_empty(&ctx->active_reqs)) {
		req = list_first_entry(&ctx->active_reqs,
608
				       struct aio_kiocb, ki_list);
Christoph Hellwig's avatar
Christoph Hellwig committed
609
		req->ki_cancel(&req->rw);
Al Viro's avatar
Al Viro committed
610
		list_del_init(&req->ki_list);
Kent Overstreet's avatar
Kent Overstreet committed
611
612
613
614
	}

	spin_unlock_irq(&ctx->ctx_lock);

Kent Overstreet's avatar
Kent Overstreet committed
615
616
	percpu_ref_kill(&ctx->reqs);
	percpu_ref_put(&ctx->reqs);
Kent Overstreet's avatar
Kent Overstreet committed
617
618
}

619
620
621
622
623
624
625
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);
626
	table = rcu_dereference_raw(mm->ioctx_table);
627
628
629
630

	while (1) {
		if (table)
			for (i = 0; i < table->nr; i++)
631
				if (!rcu_access_pointer(table->table[i])) {
632
					ctx->id = i;
633
					rcu_assign_pointer(table->table[i], ctx);
634
635
					spin_unlock(&mm->ioctx_lock);

636
637
638
639
					/* While kioctx setup is in progress,
					 * we are protected from page migration
					 * changes ring_pages by ->ring_lock.
					 */
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
					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);
657
		old = rcu_dereference_raw(mm->ioctx_table);
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673

		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
674
675
676
677
678
679
680
681
682
683
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
684
685
686
687
688
/* 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
689
	struct mm_struct *mm = current->mm;
Linus Torvalds's avatar
Linus Torvalds committed
690
	struct kioctx *ctx;
691
	int err = -ENOMEM;
Linus Torvalds's avatar
Linus Torvalds committed
692

693
694
695
696
697
698
	/*
	 * 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
699
700
701
702
703
704
705
706
707
708
709
710
	/*
	 * 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
711
	/* Prevent overflows */
712
	if (nr_events > (0x10000000U / sizeof(struct io_event))) {
Linus Torvalds's avatar
Linus Torvalds committed
713
714
715
716
		pr_debug("ENOMEM: nr_events too high\n");
		return ERR_PTR(-EINVAL);
	}

717
	if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
Linus Torvalds's avatar
Linus Torvalds committed
718
719
		return ERR_PTR(-EAGAIN);

720
	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds's avatar
Linus Torvalds committed
721
722
723
	if (!ctx)
		return ERR_PTR(-ENOMEM);

724
	ctx->max_reqs = max_reqs;
Linus Torvalds's avatar
Linus Torvalds committed
725
726

	spin_lock_init(&ctx->ctx_lock);
727
	spin_lock_init(&ctx->completion_lock);
728
	mutex_init(&ctx->ring_lock);
729
730
731
	/* 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
732
733
734
735
	init_waitqueue_head(&ctx->wait);

	INIT_LIST_HEAD(&ctx->active_reqs);

736
	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
737
738
		goto err;

739
	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
740
741
		goto err;

Kent Overstreet's avatar
Kent Overstreet committed
742
743
	ctx->cpu = alloc_percpu(struct kioctx_cpu);
	if (!ctx->cpu)
Kent Overstreet's avatar
Kent Overstreet committed
744
		goto err;
Linus Torvalds's avatar
Linus Torvalds committed
745

746
	err = aio_setup_ring(ctx, nr_events);
747
	if (err < 0)
Kent Overstreet's avatar
Kent Overstreet committed
748
		goto err;
Kent Overstreet's avatar
Kent Overstreet committed
749

750
	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreet's avatar
Kent Overstreet committed
751
	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
752
753
	if (ctx->req_batch < 1)
		ctx->req_batch = 1;
754

Linus Torvalds's avatar
Linus Torvalds committed
755
	/* limit the number of system wide aios */
756
	spin_lock(&aio_nr_lock);
757
758
	if (aio_nr + ctx->max_reqs > aio_max_nr ||
	    aio_nr + ctx->max_reqs < aio_nr) {
759
		spin_unlock(&aio_nr_lock);
Kent Overstreet's avatar
Kent Overstreet committed
760
		err = -EAGAIN;
761
		goto err_ctx;
762
763
	}
	aio_nr += ctx->max_reqs;
764
	spin_unlock(&aio_nr_lock);
Linus Torvalds's avatar
Linus Torvalds committed
765

766
767
	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
768

769
770
	err = ioctx_add_table(ctx, mm);
	if (err)
Kent Overstreet's avatar
Kent Overstreet committed
771
		goto err_cleanup;
772

773
774
775
	/* Release the ring_lock mutex now that all setup is complete. */
	mutex_unlock(&ctx->ring_lock);

Kent Overstreet's avatar
Kent Overstreet committed
776
	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
777
		 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds's avatar
Linus Torvalds committed
778
779
	return ctx;

Kent Overstreet's avatar
Kent Overstreet committed
780
781
err_cleanup:
	aio_nr_sub(ctx->max_reqs);
782
err_ctx:
783
784
785
	atomic_set(&ctx->dead, 1);
	if (ctx->mmap_size)
		vm_munmap(ctx->mmap_base, ctx->mmap_size);
786
	aio_free_ring(ctx);
Kent Overstreet's avatar
Kent Overstreet committed
787
err:
788
	mutex_unlock(&ctx->ring_lock);
Kent Overstreet's avatar
Kent Overstreet committed
789
	free_percpu(ctx->cpu);
790
791
	percpu_ref_exit(&ctx->reqs);
	percpu_ref_exit(&ctx->users);
Linus Torvalds's avatar
Linus Torvalds committed
792
	kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreet's avatar
Kent Overstreet committed
793
	pr_debug("error allocating ioctx %d\n", err);
794
	return ERR_PTR(err);
Linus Torvalds's avatar
Linus Torvalds committed
795
796
}

Kent Overstreet's avatar
Kent Overstreet committed
797
798
799
800
801
/* 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.
 */
802
static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
803
		      struct ctx_rq_wait *wait)
Kent Overstreet's avatar
Kent Overstreet committed
804
{
805
	struct kioctx_table *table;
806

Al Viro's avatar
Al Viro committed
807
808
809
	spin_lock(&mm->ioctx_lock);
	if (atomic_xchg(&ctx->dead, 1)) {
		spin_unlock(&mm->ioctx_lock);
810
		return -EINVAL;
Al Viro's avatar
Al Viro committed
811
	}
812

813
	table = rcu_dereference_raw(mm->ioctx_table);
814
815
	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
	RCU_INIT_POINTER(table->table[ctx->id], NULL);
816
	spin_unlock(&mm->ioctx_lock);
817

818
	/* free_ioctx_reqs() will do the necessary RCU synchronization */
819
	wake_up_all(&ctx->wait);
820

821
822
823
824
825
826
827
828
	/*
	 * 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);
829

830
831
	if (ctx->mmap_size)
		vm_munmap(ctx->mmap_base, ctx->mmap_size);
832

833
	ctx->rq_wait = wait;
834
835
	percpu_ref_kill(&ctx->users);
	return 0;
Linus Torvalds's avatar
Linus Torvalds committed
836
837
}

Kent Overstreet's avatar
Kent Overstreet committed
838
839
840
841
842
843
844
/*
 * 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
845
 */
846
void exit_aio(struct mm_struct *mm)
Linus Torvalds's avatar
Linus Torvalds committed
847
{
848
	struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
849
850
	struct ctx_rq_wait wait;
	int i, skipped;
851

852
853
	if (!table)
		return;
854

855
856
857
858
	atomic_set(&wait.count, table->nr);
	init_completion(&wait.comp);

	skipped = 0;
859
	for (i = 0; i < table->nr; ++i) {
860
861
		struct kioctx *ctx =
			rcu_dereference_protected(table->table[i], true);
862

863
864
		if (!ctx) {
			skipped++;
865
			continue;
866
867
		}

868
		/*
869
870
871
872
873
		 * 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.
874
		 */
875
		ctx->mmap_size = 0;
876
877
		kill_ioctx(mm, ctx, &wait);
	}
Kent Overstreet's avatar
Kent Overstreet committed
878

879
	if (!atomic_sub_and_test(skipped, &wait.count)) {
880
		/* Wait until all IO for the context are done. */
881
		wait_for_completion(&wait.comp);
Linus Torvalds's avatar
Linus Torvalds committed
882
	}
883
884
885

	RCU_INIT_POINTER(mm->ioctx_table, NULL);
	kfree(table);
Linus Torvalds's avatar
Linus Torvalds committed
886
887
}

Kent Overstreet's avatar
Kent Overstreet committed
888
889
890
static void put_reqs_available(struct kioctx *ctx, unsigned nr)
{
	struct kioctx_cpu *kcpu;
891
	unsigned long flags;
Kent Overstreet's avatar
Kent Overstreet committed
892

893
	local_irq_save(flags);
894
	kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreet's avatar
Kent Overstreet committed
895
	kcpu->reqs_available += nr;
896

Kent Overstreet's avatar
Kent Overstreet committed
897
898
899
900
901
	while (kcpu->reqs_available >= ctx->req_batch * 2) {
		kcpu->reqs_available -= ctx->req_batch;
		atomic_add(ctx->req_batch, &ctx->reqs_available);
	}

902
	local_irq_restore(flags);
Kent Overstreet's avatar
Kent Overstreet committed
903
904
905
906
907
908
}

static bool get_reqs_available(struct kioctx *ctx)
{
	struct kioctx_cpu *kcpu;
	bool ret = false;
909
	unsigned long flags;
Kent Overstreet's avatar
Kent Overstreet committed
910

911
	local_irq_save(flags);
912
	kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreet's avatar
Kent Overstreet committed
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
	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:
931
	local_irq_restore(flags);
Kent Overstreet's avatar
Kent Overstreet committed
932
933
934
	return ret;
}

935
936
937
938
939
940
941
942
943
944
945
946
947
948
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
/* 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
997
/* aio_get_req
Kent Overstreet's avatar
Kent Overstreet committed
998
999
 *	Allocate a slot for an aio request.
 * Returns NULL if no requests are free.
Linus Torvalds's avatar
Linus Torvalds committed
1000
 */
1001
static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds's avatar
Linus Torvalds committed
1002
{
1003
	struct aio_kiocb *req;
Kent Overstreet's avatar
Kent Overstreet committed
1004

1005
1006
1007
1008
1009
	if (!get_reqs_available(ctx)) {
		user_refill_reqs_available(ctx);
		if (!get_reqs_available(ctx))
			return NULL;
	}
Kent Overstreet's avatar
Kent Overstreet committed
1010

1011
	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
Linus Torvalds's avatar
Linus Torvalds committed
1012
	if (unlikely(!req))
Kent Overstreet's avatar
Kent Overstreet committed
1013
		goto out_put;
Linus Torvalds's avatar
Linus Torvalds committed
1014

Kent Overstreet's avatar
Kent Overstreet committed
1015
	percpu_ref_get(&ctx->reqs);
1016
	INIT_LIST_HEAD(&req->ki_list);
Christoph Hellwig's avatar
Christoph Hellwig committed
1017
	refcount_set(&req->ki_refcnt, 0);
Linus Torvalds's avatar
Linus Torvalds committed
1018
	req->ki_ctx = ctx;
Jeff Moyer's avatar
Jeff Moyer committed
1019
	return req;
Kent Overstreet's avatar
Kent Overstreet committed
1020
out_put:
Kent Overstreet's avatar
Kent Overstreet committed
1021
	put_reqs_available(ctx, 1);
Kent Overstreet's avatar
Kent Overstreet committed
1022
	return NULL;
Linus Torvalds's avatar
Linus Torvalds committed
1023
1024
}

1025
static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds's avatar
Linus Torvalds committed
1026
{
1027
	struct aio_ring __user *ring  = (void __user *)ctx_id;
1028
	struct mm_struct *mm = current->mm;
1029
	struct kioctx *ctx, *ret = NULL;
1030
1031
1032
1033
1034
	struct kioctx_table *table;
	unsigned id;

	if (get_user(id, &ring->id))
		return NULL;
Linus Torvalds's avatar
Linus Torvalds committed
1035

1036
	rcu_read_lock();
1037
	table = rcu_dereference(mm->ioctx_table);
1038

1039
1040
	if (!table || id >= table->nr)
		goto out;
Linus Torvalds's avatar
Linus Torvalds committed
1041

1042
	id = array_index_nospec(id, table->nr);
1043
	ctx = rcu_dereference(table->table[id]);
1044
	if (ctx && ctx->user_id == ctx_id) {
1045
1046
		if (percpu_ref_tryget_live(&ctx->users))
			ret = ctx;
1047
1048
	}
out:
1049
	rcu_read_unlock();
1050
	return ret;
Linus Torvalds's avatar
Linus Torvalds committed
1051
1052
}

Christoph Hellwig's avatar
Christoph Hellwig committed
1053
1054
1055
1056
1057
1058
1059
1060
1061
static inline void iocb_put(struct aio_kiocb *iocb)
{
	if (refcount_read(&iocb->ki_refcnt) == 0 ||
	    refcount_dec_and_test(&iocb->ki_refcnt)) {
		percpu_ref_put(&iocb->ki_ctx->reqs);
		kmem_cache_free(kiocb_cachep, iocb);
	}
}

Linus Torvalds's avatar
Linus Torvalds committed
1062
1063
1064
/* aio_complete
 *	Called when the io request on the given iocb is complete.
 */
1065
static void aio_complete(struct aio_kiocb *iocb, long res, long res2)
Linus Torvalds's avatar
Linus Torvalds committed
1066
1067
1068
{
	struct kioctx	*ctx = iocb->ki_ctx;
	struct aio_ring	*ring;
Kent Overstreet's avatar
Kent Overstreet committed
1069
	struct io_event	*ev_page, *event;
1070
	unsigned tail, pos, head;
Linus Torvalds's avatar
Linus Torvalds committed
1071
1072
	unsigned long	flags;

1073
1074
	/*
	 * Add a completion event to the ring buffer. Must be done holding
1075
	 * ctx->completion_lock to prevent other code from messing with the tail
1076
1077
1078
1079
	 * pointer since we might be called from irq context.
	 */
	spin_lock_irqsave(&ctx->completion_lock, flags);

1080
	tail = ctx->tail;
Kent Overstreet's avatar
Kent Overstreet committed
1081
1082
	pos = tail + AIO_EVENTS_OFFSET;

1083
	if (++tail >= ctx->nr_events)
1084
		tail = 0;
Linus Torvalds's avatar
Linus Torvalds committed
1085

1086
	ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet's avatar
Kent Overstreet committed
1087
1088