aio.c 54.4 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>
Linus Torvalds's avatar
Linus Torvalds committed
48

Al Viro's avatar
Al Viro committed
49
50
#include "internal.h"

51
52
#define KIOCB_KEY		0

53
54
55
56
57
58
#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 */
59
60
	unsigned	head;	/* Written to by userland or under ring_lock
				 * mutex by aio_read_events_ring(). */
61
62
63
64
65
66
67
68
69
70
71
72
73
	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

74
struct kioctx_table {
75
76
77
	struct rcu_head		rcu;
	unsigned		nr;
	struct kioctx __rcu	*table[];
78
79
};

Kent Overstreet's avatar
Kent Overstreet committed
80
81
82
83
struct kioctx_cpu {
	unsigned		reqs_available;
};

84
85
86
87
88
struct ctx_rq_wait {
	struct completion comp;
	atomic_t count;
};

89
struct kioctx {
Kent Overstreet's avatar
Kent Overstreet committed
90
	struct percpu_ref	users;
Kent Overstreet's avatar
Kent Overstreet committed
91
	atomic_t		dead;
92

Kent Overstreet's avatar
Kent Overstreet committed
93
94
	struct percpu_ref	reqs;

95
96
	unsigned long		user_id;

Kent Overstreet's avatar
Kent Overstreet committed
97
98
99
100
101
102
103
	struct __percpu kioctx_cpu *cpu;

	/*
	 * For percpu reqs_available, number of slots we move to/from global
	 * counter at a time:
	 */
	unsigned		req_batch;
104
105
106
107
	/*
	 * This is what userspace passed to io_setup(), it's not used for
	 * anything but counting against the global max_reqs quota.
	 *
108
	 * The real limit is nr_events - 1, which will be larger (see
109
110
	 * aio_setup_ring())
	 */
111
112
	unsigned		max_reqs;

113
114
	/* Size of ringbuffer, in units of struct io_event */
	unsigned		nr_events;
115

116
117
118
119
120
121
	unsigned long		mmap_base;
	unsigned long		mmap_size;

	struct page		**ring_pages;
	long			nr_pages;

122
	struct rcu_work		free_rwork;	/* see free_ioctx() */
123

124
125
126
	/*
	 * signals when all in-flight requests are done
	 */
127
	struct ctx_rq_wait	*rq_wait;
128

129
	struct {
130
131
132
133
134
		/*
		 * 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
135
136
		 *
		 * We batch accesses to it with a percpu version.
137
138
		 */
		atomic_t	reqs_available;
139
140
141
142
143
144
145
	} ____cacheline_aligned_in_smp;

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

146
147
	struct {
		struct mutex	ring_lock;
148
149
		wait_queue_head_t wait;
	} ____cacheline_aligned_in_smp;
150
151
152

	struct {
		unsigned	tail;
153
		unsigned	completed_events;
154
		spinlock_t	completion_lock;
155
	} ____cacheline_aligned_in_smp;
156
157

	struct page		*internal_pages[AIO_RING_PAGES];
158
	struct file		*aio_ring_file;
159
160

	unsigned		id;
161
162
};

163
164
165
166
167
168
struct fsync_iocb {
	struct work_struct	work;
	struct file		*file;
	bool			datasync;
};

169
170
171
172
173
174
175
176
177
178
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;
};

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

	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
194
	refcount_t		ki_refcnt;
195
196
197
198
199
200
201
202

	/*
	 * 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
203
/*------ sysctl variables----*/
204
205
206
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
207
208
/*----end sysctl variables---*/

209
210
static struct kmem_cache	*kiocb_cachep;
static struct kmem_cache	*kioctx_cachep;
Linus Torvalds's avatar
Linus Torvalds committed
211

212
213
214
215
216
217
218
219
220
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);
221
222
	if (IS_ERR(inode))
		return ERR_CAST(inode);
223
224
225
226
227

	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
228
229
	file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
				O_RDWR, &aio_ring_fops);
230
	if (IS_ERR(file))
231
232
233
234
235
236
237
		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
238
	struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, NULL,
Jann Horn's avatar
Jann Horn committed
239
240
241
242
243
					   AIO_RING_MAGIC);

	if (!IS_ERR(root))
		root->d_sb->s_iflags |= SB_I_NOEXEC;
	return root;
244
245
}

Linus Torvalds's avatar
Linus Torvalds committed
246
247
248
249
250
251
/* 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)
{
252
253
254
255
256
257
258
259
260
	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.");

261
	kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
262
	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
Linus Torvalds's avatar
Linus Torvalds committed
263
264
	return 0;
}
265
__initcall(aio_setup);
Linus Torvalds's avatar
Linus Torvalds committed
266

267
268
269
static void put_aio_ring_file(struct kioctx *ctx)
{
	struct file *aio_ring_file = ctx->aio_ring_file;
270
271
	struct address_space *i_mapping;

272
	if (aio_ring_file) {
Al Viro's avatar
Al Viro committed
273
		truncate_setsize(file_inode(aio_ring_file), 0);
274
275

		/* Prevent further access to the kioctx from migratepages */
Al Viro's avatar
Al Viro committed
276
		i_mapping = aio_ring_file->f_mapping;
277
278
		spin_lock(&i_mapping->private_lock);
		i_mapping->private_data = NULL;
279
		ctx->aio_ring_file = NULL;
280
		spin_unlock(&i_mapping->private_lock);
281
282
283
284
285

		fput(aio_ring_file);
	}
}

Linus Torvalds's avatar
Linus Torvalds committed
286
287
static void aio_free_ring(struct kioctx *ctx)
{
288
	int i;
Linus Torvalds's avatar
Linus Torvalds committed
289

290
291
292
293
294
	/* Disconnect the kiotx from the ring file.  This prevents future
	 * accesses to the kioctx from page migration.
	 */
	put_aio_ring_file(ctx);

295
	for (i = 0; i < ctx->nr_pages; i++) {
296
		struct page *page;
297
298
		pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
				page_count(ctx->ring_pages[i]));
299
300
301
302
303
		page = ctx->ring_pages[i];
		if (!page)
			continue;
		ctx->ring_pages[i] = NULL;
		put_page(page);
304
	}
Linus Torvalds's avatar
Linus Torvalds committed
305

306
	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
307
		kfree(ctx->ring_pages);
308
309
		ctx->ring_pages = NULL;
	}
310
311
}

312
static int aio_ring_mremap(struct vm_area_struct *vma)
313
{
314
	struct file *file = vma->vm_file;
315
316
	struct mm_struct *mm = vma->vm_mm;
	struct kioctx_table *table;
Al Viro's avatar
Al Viro committed
317
	int i, res = -EINVAL;
318
319
320
321
322
323
324

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

325
		ctx = rcu_dereference(table->table[i]);
326
		if (ctx && ctx->aio_ring_file == file) {
Al Viro's avatar
Al Viro committed
327
328
329
330
			if (!atomic_read(&ctx->dead)) {
				ctx->user_id = ctx->mmap_base = vma->vm_start;
				res = 0;
			}
331
332
333
334
335
336
			break;
		}
	}

	rcu_read_unlock();
	spin_unlock(&mm->ioctx_lock);
Al Viro's avatar
Al Viro committed
337
	return res;
338
339
}

340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
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;
}

356
357
358
359
static const struct file_operations aio_ring_fops = {
	.mmap = aio_ring_mmap,
};

360
#if IS_ENABLED(CONFIG_MIGRATION)
361
362
363
static int aio_migratepage(struct address_space *mapping, struct page *new,
			struct page *old, enum migrate_mode mode)
{
364
	struct kioctx *ctx;
365
	unsigned long flags;
366
	pgoff_t idx;
367
368
	int rc;

369
370
371
372
373
374
375
376
	/*
	 * 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;

377
378
	rc = 0;

379
	/* mapping->private_lock here protects against the kioctx teardown.  */
380
381
	spin_lock(&mapping->private_lock);
	ctx = mapping->private_data;
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
	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;
401
402
403
404
	} else
		rc = -EINVAL;

	if (rc != 0)
405
		goto out_unlock;
406

407
408
	/* Writeback must be complete */
	BUG_ON(PageWriteback(old));
409
	get_page(new);
410

411
	rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
412
	if (rc != MIGRATEPAGE_SUCCESS) {
413
		put_page(new);
414
		goto out_unlock;
415
416
	}

417
418
419
	/* 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.
420
	 */
421
422
423
424
425
	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);
426

427
428
	/* The old page is no longer accessible. */
	put_page(old);
429

430
431
432
433
out_unlock:
	mutex_unlock(&ctx->ring_lock);
out:
	spin_unlock(&mapping->private_lock);
434
	return rc;
Linus Torvalds's avatar
Linus Torvalds committed
435
}
436
#endif
Linus Torvalds's avatar
Linus Torvalds committed
437

438
static const struct address_space_operations aio_ctx_aops = {
439
	.set_page_dirty = __set_page_dirty_no_writeback,
440
#if IS_ENABLED(CONFIG_MIGRATION)
441
	.migratepage	= aio_migratepage,
442
#endif
443
444
};

445
static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
Linus Torvalds's avatar
Linus Torvalds committed
446
447
{
	struct aio_ring *ring;
Zach Brown's avatar
Zach Brown committed
448
	struct mm_struct *mm = current->mm;
449
	unsigned long size, unused;
Linus Torvalds's avatar
Linus Torvalds committed
450
	int nr_pages;
451
452
	int i;
	struct file *file;
Linus Torvalds's avatar
Linus Torvalds committed
453
454
455
456
457
458
459

	/* 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;

460
	nr_pages = PFN_UP(size);
Linus Torvalds's avatar
Linus Torvalds committed
461
462
463
	if (nr_pages < 0)
		return -EINVAL;

464
	file = aio_private_file(ctx, nr_pages);
465
466
	if (IS_ERR(file)) {
		ctx->aio_ring_file = NULL;
467
		return -ENOMEM;
468
469
	}

470
471
472
473
474
475
476
477
478
479
480
481
482
483
	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;
		}
	}

484
485
	for (i = 0; i < nr_pages; i++) {
		struct page *page;
Al Viro's avatar
Al Viro committed
486
		page = find_or_create_page(file->f_mapping,
487
488
489
490
491
492
493
					   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);
494
495

		ctx->ring_pages[i] = page;
496
	}
497
	ctx->nr_pages = i;
Linus Torvalds's avatar
Linus Torvalds committed
498

499
500
	if (unlikely(i != nr_pages)) {
		aio_free_ring(ctx);
501
		return -ENOMEM;
Linus Torvalds's avatar
Linus Torvalds committed
502
503
	}

504
505
	ctx->mmap_size = nr_pages * PAGE_SIZE;
	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
506

507
508
509
510
511
512
	if (down_write_killable(&mm->mmap_sem)) {
		ctx->mmap_size = 0;
		aio_free_ring(ctx);
		return -EINTR;
	}

513
514
	ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
				       PROT_READ | PROT_WRITE,
515
				       MAP_SHARED, 0, &unused, NULL);
516
	up_write(&mm->mmap_sem);
517
518
	if (IS_ERR((void *)ctx->mmap_base)) {
		ctx->mmap_size = 0;
Linus Torvalds's avatar
Linus Torvalds committed
519
		aio_free_ring(ctx);
520
		return -ENOMEM;
Linus Torvalds's avatar
Linus Torvalds committed
521
522
	}

523
	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
524

525
526
	ctx->user_id = ctx->mmap_base;
	ctx->nr_events = nr_events; /* trusted copy */
Linus Torvalds's avatar
Linus Torvalds committed
527

528
	ring = kmap_atomic(ctx->ring_pages[0]);
Linus Torvalds's avatar
Linus Torvalds committed
529
	ring->nr = nr_events;	/* user copy */
530
	ring->id = ~0U;
Linus Torvalds's avatar
Linus Torvalds committed
531
532
533
534
535
	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);
536
	kunmap_atomic(ring);
537
	flush_dcache_page(ctx->ring_pages[0]);
Linus Torvalds's avatar
Linus Torvalds committed
538
539
540
541
542
543
544
545

	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)

546
void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
547
{
548
	struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
549
550
551
	struct kioctx *ctx = req->ki_ctx;
	unsigned long flags;

552
553
	if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
		return;
554

555
556
	spin_lock_irqsave(&ctx->ctx_lock, flags);
	list_add_tail(&req->ki_list, &ctx->active_reqs);
557
558
559
560
561
	req->ki_cancel = cancel;
	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
}
EXPORT_SYMBOL(kiocb_set_cancel_fn);

562
563
564
/*
 * free_ioctx() should be RCU delayed to synchronize against the RCU
 * protected lookup_ioctx() and also needs process context to call
565
 * aio_free_ring().  Use rcu_work.
566
 */
Kent Overstreet's avatar
Kent Overstreet committed
567
static void free_ioctx(struct work_struct *work)
Kent Overstreet's avatar
Kent Overstreet committed
568
{
569
570
	struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
					  free_rwork);
Kent Overstreet's avatar
Kent Overstreet committed
571
	pr_debug("freeing %p\n", ctx);
Kent Overstreet's avatar
Kent Overstreet committed
572

Kent Overstreet's avatar
Kent Overstreet committed
573
	aio_free_ring(ctx);
Kent Overstreet's avatar
Kent Overstreet committed
574
	free_percpu(ctx->cpu);
575
576
	percpu_ref_exit(&ctx->reqs);
	percpu_ref_exit(&ctx->users);
Kent Overstreet's avatar
Kent Overstreet committed
577
578
579
	kmem_cache_free(kioctx_cachep, ctx);
}

Kent Overstreet's avatar
Kent Overstreet committed
580
581
582
583
static void free_ioctx_reqs(struct percpu_ref *ref)
{
	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);

584
	/* At this point we know that there are no any in-flight requests */
585
586
	if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
		complete(&ctx->rq_wait->comp);
587

588
	/* Synchronize against RCU protected table->table[] dereferences */
589
590
	INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
	queue_rcu_work(system_wq, &ctx->free_rwork);
Kent Overstreet's avatar
Kent Overstreet committed
591
592
}

Kent Overstreet's avatar
Kent Overstreet committed
593
594
595
596
597
/*
 * 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
598
static void free_ioctx_users(struct percpu_ref *ref)
Kent Overstreet's avatar
Kent Overstreet committed
599
{
Kent Overstreet's avatar
Kent Overstreet committed
600
	struct kioctx *ctx = container_of(ref, struct kioctx, users);
601
	struct aio_kiocb *req;
Kent Overstreet's avatar
Kent Overstreet committed
602
603
604
605
606

	spin_lock_irq(&ctx->ctx_lock);

	while (!list_empty(&ctx->active_reqs)) {
		req = list_first_entry(&ctx->active_reqs,
607
				       struct aio_kiocb, ki_list);
Christoph Hellwig's avatar
Christoph Hellwig committed
608
		req->ki_cancel(&req->rw);
Al Viro's avatar
Al Viro committed
609
		list_del_init(&req->ki_list);
Kent Overstreet's avatar
Kent Overstreet committed
610
611
612
613
	}

	spin_unlock_irq(&ctx->ctx_lock);

Kent Overstreet's avatar
Kent Overstreet committed
614
615
	percpu_ref_kill(&ctx->reqs);
	percpu_ref_put(&ctx->reqs);
Kent Overstreet's avatar
Kent Overstreet committed
616
617
}

618
619
620
621
622
623
624
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);
625
	table = rcu_dereference_raw(mm->ioctx_table);
626
627
628
629

	while (1) {
		if (table)
			for (i = 0; i < table->nr; i++)
630
				if (!rcu_access_pointer(table->table[i])) {
631
					ctx->id = i;
632
					rcu_assign_pointer(table->table[i], ctx);
633
634
					spin_unlock(&mm->ioctx_lock);

635
636
637
638
					/* While kioctx setup is in progress,
					 * we are protected from page migration
					 * changes ring_pages by ->ring_lock.
					 */
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
					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);
656
		old = rcu_dereference_raw(mm->ioctx_table);
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672

		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
673
674
675
676
677
678
679
680
681
682
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
683
684
685
686
687
/* 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
688
	struct mm_struct *mm = current->mm;
Linus Torvalds's avatar
Linus Torvalds committed
689
	struct kioctx *ctx;
690
	int err = -ENOMEM;
Linus Torvalds's avatar
Linus Torvalds committed
691

692
693
694
695
696
697
	/*
	 * 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
698
699
700
701
702
703
704
705
706
707
708
709
	/*
	 * 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
710
	/* Prevent overflows */
711
	if (nr_events > (0x10000000U / sizeof(struct io_event))) {
Linus Torvalds's avatar
Linus Torvalds committed
712
713
714
715
		pr_debug("ENOMEM: nr_events too high\n");
		return ERR_PTR(-EINVAL);
	}

716
	if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
Linus Torvalds's avatar
Linus Torvalds committed
717
718
		return ERR_PTR(-EAGAIN);

719
	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
Linus Torvalds's avatar
Linus Torvalds committed
720
721
722
	if (!ctx)
		return ERR_PTR(-ENOMEM);

723
	ctx->max_reqs = max_reqs;
Linus Torvalds's avatar
Linus Torvalds committed
724
725

	spin_lock_init(&ctx->ctx_lock);
726
	spin_lock_init(&ctx->completion_lock);
727
	mutex_init(&ctx->ring_lock);
728
729
730
	/* 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
731
732
733
734
	init_waitqueue_head(&ctx->wait);

	INIT_LIST_HEAD(&ctx->active_reqs);

735
	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
736
737
		goto err;

738
	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
739
740
		goto err;

Kent Overstreet's avatar
Kent Overstreet committed
741
742
	ctx->cpu = alloc_percpu(struct kioctx_cpu);
	if (!ctx->cpu)
Kent Overstreet's avatar
Kent Overstreet committed
743
		goto err;
Linus Torvalds's avatar
Linus Torvalds committed
744

745
	err = aio_setup_ring(ctx, nr_events);
746
	if (err < 0)
Kent Overstreet's avatar
Kent Overstreet committed
747
		goto err;
Kent Overstreet's avatar
Kent Overstreet committed
748

749
	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
Kent Overstreet's avatar
Kent Overstreet committed
750
	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
751
752
	if (ctx->req_batch < 1)
		ctx->req_batch = 1;
753

Linus Torvalds's avatar
Linus Torvalds committed
754
	/* limit the number of system wide aios */
755
	spin_lock(&aio_nr_lock);
756
757
	if (aio_nr + ctx->max_reqs > aio_max_nr ||
	    aio_nr + ctx->max_reqs < aio_nr) {
758
		spin_unlock(&aio_nr_lock);
Kent Overstreet's avatar
Kent Overstreet committed
759
		err = -EAGAIN;
760
		goto err_ctx;
761
762
	}
	aio_nr += ctx->max_reqs;
763
	spin_unlock(&aio_nr_lock);
Linus Torvalds's avatar
Linus Torvalds committed
764

765
766
	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
767

768
769
	err = ioctx_add_table(ctx, mm);
	if (err)
Kent Overstreet's avatar
Kent Overstreet committed
770
		goto err_cleanup;
771

772
773
774
	/* Release the ring_lock mutex now that all setup is complete. */
	mutex_unlock(&ctx->ring_lock);

Kent Overstreet's avatar
Kent Overstreet committed
775
	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
776
		 ctx, ctx->user_id, mm, ctx->nr_events);
Linus Torvalds's avatar
Linus Torvalds committed
777
778
	return ctx;

Kent Overstreet's avatar
Kent Overstreet committed
779
780
err_cleanup:
	aio_nr_sub(ctx->max_reqs);
781
err_ctx:
782
783
784
	atomic_set(&ctx->dead, 1);
	if (ctx->mmap_size)
		vm_munmap(ctx->mmap_base, ctx->mmap_size);
785
	aio_free_ring(ctx);
Kent Overstreet's avatar
Kent Overstreet committed
786
err:
787
	mutex_unlock(&ctx->ring_lock);
Kent Overstreet's avatar
Kent Overstreet committed
788
	free_percpu(ctx->cpu);
789
790
	percpu_ref_exit(&ctx->reqs);
	percpu_ref_exit(&ctx->users);
Linus Torvalds's avatar
Linus Torvalds committed
791
	kmem_cache_free(kioctx_cachep, ctx);
Kent Overstreet's avatar
Kent Overstreet committed
792
	pr_debug("error allocating ioctx %d\n", err);
793
	return ERR_PTR(err);
Linus Torvalds's avatar
Linus Torvalds committed
794
795
}

Kent Overstreet's avatar
Kent Overstreet committed
796
797
798
799
800
/* 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.
 */
801
static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
802
		      struct ctx_rq_wait *wait)
Kent Overstreet's avatar
Kent Overstreet committed
803
{
804
	struct kioctx_table *table;
805

Al Viro's avatar
Al Viro committed
806
807
808
	spin_lock(&mm->ioctx_lock);
	if (atomic_xchg(&ctx->dead, 1)) {
		spin_unlock(&mm->ioctx_lock);
809
		return -EINVAL;
Al Viro's avatar
Al Viro committed
810
	}
811

812
	table = rcu_dereference_raw(mm->ioctx_table);
813
814
	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
	RCU_INIT_POINTER(table->table[ctx->id], NULL);
815
	spin_unlock(&mm->ioctx_lock);
816

817
	/* free_ioctx_reqs() will do the necessary RCU synchronization */
818
	wake_up_all(&ctx->wait);
819

820
821
822
823
824
825
826
827
	/*
	 * 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);
828

829
830
	if (ctx->mmap_size)
		vm_munmap(ctx->mmap_base, ctx->mmap_size);
831

832
	ctx->rq_wait = wait;
833
834
	percpu_ref_kill(&ctx->users);
	return 0;
Linus Torvalds's avatar
Linus Torvalds committed
835
836
}

Kent Overstreet's avatar
Kent Overstreet committed
837
838
839
840
841
842
843
/*
 * 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
844
 */
845
void exit_aio(struct mm_struct *mm)
Linus Torvalds's avatar
Linus Torvalds committed
846
{
847
	struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
848
849
	struct ctx_rq_wait wait;
	int i, skipped;
850

851
852
	if (!table)
		return;
853

854
855
856
857
	atomic_set(&wait.count, table->nr);
	init_completion(&wait.comp);

	skipped = 0;
858
	for (i = 0; i < table->nr; ++i) {
859
860
		struct kioctx *ctx =
			rcu_dereference_protected(table->table[i], true);
861

862
863
		if (!ctx) {
			skipped++;
864
			continue;
865
866
		}

867
		/*
868
869
870
871
872
		 * 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.
873
		 */
874
		ctx->mmap_size = 0;
875
876
		kill_ioctx(mm, ctx, &wait);
	}
Kent Overstreet's avatar
Kent Overstreet committed
877

878
	if (!atomic_sub_and_test(skipped, &wait.count)) {
879
		/* Wait until all IO for the context are done. */
880
		wait_for_completion(&wait.comp);
Linus Torvalds's avatar
Linus Torvalds committed
881
	}
882
883
884

	RCU_INIT_POINTER(mm->ioctx_table, NULL);
	kfree(table);
Linus Torvalds's avatar
Linus Torvalds committed
885
886
}

Kent Overstreet's avatar
Kent Overstreet committed
887
888
889
static void put_reqs_available(struct kioctx *ctx, unsigned nr)
{
	struct kioctx_cpu *kcpu;
890
	unsigned long flags;
Kent Overstreet's avatar
Kent Overstreet committed
891

892
	local_irq_save(flags);
893
	kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreet's avatar
Kent Overstreet committed
894
	kcpu->reqs_available += nr;
895

Kent Overstreet's avatar
Kent Overstreet committed
896
897
898
899
900
	while (kcpu->reqs_available >= ctx->req_batch * 2) {
		kcpu->reqs_available -= ctx->req_batch;
		atomic_add(ctx->req_batch, &ctx->reqs_available);
	}

901
	local_irq_restore(flags);
Kent Overstreet's avatar
Kent Overstreet committed
902
903
904
905
906
907
}

static bool get_reqs_available(struct kioctx *ctx)
{
	struct kioctx_cpu *kcpu;
	bool ret = false;
908
	unsigned long flags;
Kent Overstreet's avatar
Kent Overstreet committed
909

910
	local_irq_save(flags);
911
	kcpu = this_cpu_ptr(ctx->cpu);
Kent Overstreet's avatar
Kent Overstreet committed
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
	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:
930
	local_irq_restore(flags);
Kent Overstreet's avatar
Kent Overstreet committed
931
932
933
	return ret;
}

934
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
/* 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
996
/* aio_get_req
Kent Overstreet's avatar
Kent Overstreet committed
997
998
 *	Allocate a slot for an aio request.
 * Returns NULL if no requests are free.
Linus Torvalds's avatar
Linus Torvalds committed
999
 */
1000
static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
Linus Torvalds's avatar
Linus Torvalds committed
1001
{
1002
	struct aio_kiocb *req;
Kent Overstreet's avatar
Kent Overstreet committed
1003

1004
1005
1006
1007
1008
	if (!get_reqs_available(ctx)) {
		user_refill_reqs_available(ctx);
		if (!get_reqs_available(ctx))
			return NULL;
	}
Kent Overstreet's avatar
Kent Overstreet committed
1009

1010
	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
Linus Torvalds's avatar
Linus Torvalds committed
1011
	if (unlikely(!req))
Kent Overstreet's avatar
Kent Overstreet committed
1012
		goto out_put;
Linus Torvalds's avatar
Linus Torvalds committed
1013

Kent Overstreet's avatar
Kent Overstreet committed
1014
	percpu_ref_get(&ctx->reqs);
1015
	INIT_LIST_HEAD(&req->ki_list);
Christoph Hellwig's avatar
Christoph Hellwig committed
1016
	refcount_set(&req->ki_refcnt, 0);
Linus Torvalds's avatar
Linus Torvalds committed
1017
	req->ki_ctx = ctx;
Jeff Moyer's avatar
Jeff Moyer committed
1018
	return req;
Kent Overstreet's avatar
Kent Overstreet committed
1019
out_put:
Kent Overstreet's avatar
Kent Overstreet committed
1020
	put_reqs_available(ctx, 1);
Kent Overstreet's avatar
Kent Overstreet committed
1021
	return NULL;
Linus Torvalds's avatar
Linus Torvalds committed
1022
1023
}

1024
static struct kioctx *lookup_ioctx(unsigned long ctx_id)
Linus Torvalds's avatar
Linus Torvalds committed
1025
{
1026
	struct aio_ring __user *ring  = (void __user *)ctx_id;
1027
	struct mm_struct *mm = current->mm;
1028
	struct kioctx *ctx, *ret = NULL;
1029
1030
1031
1032
1033
	struct kioctx_table *table;
	unsigned id;

	if (get_user(id, &ring->id))
		return NULL;
Linus Torvalds's avatar
Linus Torvalds committed
1034

1035
	rcu_read_lock();
1036
	table = rcu_dereference(mm->ioctx_table);
1037

1038
1039
	if (!table || id >= table->nr)
		goto out;
Linus Torvalds's avatar
Linus Torvalds committed
1040

1041
	ctx = rcu_dereference(table->table[id]);
1042
	if (ctx && ctx->user_id == ctx_id) {
1043
1044
		if (percpu_ref_tryget_live(&ctx->users))
			ret = ctx;
1045
1046
	}
out:
1047
	rcu_read_unlock();
1048
	return ret;
Linus Torvalds's avatar
Linus Torvalds committed
1049
1050
}

Christoph Hellwig's avatar
Christoph Hellwig committed
1051
1052
1053
1054
1055
1056
1057
1058
1059
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
1060
1061
1062
/* aio_complete
 *	Called when the io request on the given iocb is complete.
 */
1063
static void aio_complete(struct aio_kiocb *iocb, long res, long res2)
Linus Torvalds's avatar
Linus Torvalds committed
1064
1065
1066
{
	struct kioctx	*ctx = iocb->ki_ctx;
	struct aio_ring	*ring;
Kent Overstreet's avatar
Kent Overstreet committed
1067
	struct io_event	*ev_page, *event;
1068
	unsigned tail, pos, head;
Linus Torvalds's avatar
Linus Torvalds committed
1069
1070
	unsigned long	flags;

1071
1072
	/*
	 * Add a completion event to the ring buffer. Must be done holding
1073
	 * ctx->completion_lock to prevent other code from messing with the tail
1074
1075
1076
1077
	 * pointer since we might be called from irq context.
	 */
	spin_lock_irqsave(&ctx->completion_lock, flags);

1078
	tail = ctx->tail;
Kent Overstreet's avatar
Kent Overstreet committed
1079
1080
	pos = tail + AIO_EVENTS_OFFSET;

1081
	if (++tail >= ctx->nr_events)
1082
		tail = 0;
Linus Torvalds's avatar
Linus Torvalds committed
1083

1084
	ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
Kent Overstreet's avatar
Kent Overstreet committed
1085
1086
	event = ev_page + pos % AIO_EVENTS_PER_PAGE;

1087