1/*
2 *	An async IO implementation for Linux
3 *	Written by Benjamin LaHaise <bcrl@kvack.org>
4 *
5 *	Implements an efficient asynchronous io interface.
6 *
7 *	Copyright 2000, 2001, 2002 Red Hat, Inc.  All Rights Reserved.
8 *
9 *	See ../COPYING for licensing terms.
10 */
11#define pr_fmt(fmt) "%s: " fmt, __func__
12
13#include <linux/kernel.h>
14#include <linux/init.h>
15#include <linux/errno.h>
16#include <linux/time.h>
17#include <linux/aio_abi.h>
18#include <linux/export.h>
19#include <linux/syscalls.h>
20#include <linux/backing-dev.h>
21#include <linux/uio.h>
22
23#include <linux/sched.h>
24#include <linux/fs.h>
25#include <linux/file.h>
26#include <linux/mm.h>
27#include <linux/mman.h>
28#include <linux/mmu_context.h>
29#include <linux/percpu.h>
30#include <linux/slab.h>
31#include <linux/timer.h>
32#include <linux/aio.h>
33#include <linux/highmem.h>
34#include <linux/workqueue.h>
35#include <linux/security.h>
36#include <linux/eventfd.h>
37#include <linux/blkdev.h>
38#include <linux/compat.h>
39#include <linux/migrate.h>
40#include <linux/ramfs.h>
41#include <linux/percpu-refcount.h>
42#include <linux/mount.h>
43
44#include <asm/kmap_types.h>
45#include <asm/uaccess.h>
46
47#include "internal.h"
48
49#define AIO_RING_MAGIC			0xa10a10a1
50#define AIO_RING_COMPAT_FEATURES	1
51#define AIO_RING_INCOMPAT_FEATURES	0
52struct aio_ring {
53	unsigned	id;	/* kernel internal index number */
54	unsigned	nr;	/* number of io_events */
55	unsigned	head;	/* Written to by userland or under ring_lock
56				 * mutex by aio_read_events_ring(). */
57	unsigned	tail;
58
59	unsigned	magic;
60	unsigned	compat_features;
61	unsigned	incompat_features;
62	unsigned	header_length;	/* size of aio_ring */
63
64
65	struct io_event		io_events[0];
66}; /* 128 bytes + ring size */
67
68#define AIO_RING_PAGES	8
69
70struct kioctx_table {
71	struct rcu_head	rcu;
72	unsigned	nr;
73	struct kioctx	*table[];
74};
75
76struct kioctx_cpu {
77	unsigned		reqs_available;
78};
79
80struct ctx_rq_wait {
81	struct completion comp;
82	atomic_t count;
83};
84
85struct kioctx {
86	struct percpu_ref	users;
87	atomic_t		dead;
88
89	struct percpu_ref	reqs;
90
91	unsigned long		user_id;
92
93	struct __percpu kioctx_cpu *cpu;
94
95	/*
96	 * For percpu reqs_available, number of slots we move to/from global
97	 * counter at a time:
98	 */
99	unsigned		req_batch;
100	/*
101	 * This is what userspace passed to io_setup(), it's not used for
102	 * anything but counting against the global max_reqs quota.
103	 *
104	 * The real limit is nr_events - 1, which will be larger (see
105	 * aio_setup_ring())
106	 */
107	unsigned		max_reqs;
108
109	/* Size of ringbuffer, in units of struct io_event */
110	unsigned		nr_events;
111
112	unsigned long		mmap_base;
113	unsigned long		mmap_size;
114
115	struct page		**ring_pages;
116	long			nr_pages;
117
118	struct work_struct	free_work;
119
120	/*
121	 * signals when all in-flight requests are done
122	 */
123	struct ctx_rq_wait	*rq_wait;
124
125	struct {
126		/*
127		 * This counts the number of available slots in the ringbuffer,
128		 * so we avoid overflowing it: it's decremented (if positive)
129		 * when allocating a kiocb and incremented when the resulting
130		 * io_event is pulled off the ringbuffer.
131		 *
132		 * We batch accesses to it with a percpu version.
133		 */
134		atomic_t	reqs_available;
135	} ____cacheline_aligned_in_smp;
136
137	struct {
138		spinlock_t	ctx_lock;
139		struct list_head active_reqs;	/* used for cancellation */
140	} ____cacheline_aligned_in_smp;
141
142	struct {
143		struct mutex	ring_lock;
144		wait_queue_head_t wait;
145	} ____cacheline_aligned_in_smp;
146
147	struct {
148		unsigned	tail;
149		unsigned	completed_events;
150		spinlock_t	completion_lock;
151	} ____cacheline_aligned_in_smp;
152
153	struct page		*internal_pages[AIO_RING_PAGES];
154	struct file		*aio_ring_file;
155
156	unsigned		id;
157};
158
159/*
160 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
161 * cancelled or completed (this makes a certain amount of sense because
162 * successful cancellation - io_cancel() - does deliver the completion to
163 * userspace).
164 *
165 * And since most things don't implement kiocb cancellation and we'd really like
166 * kiocb completion to be lockless when possible, we use ki_cancel to
167 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
168 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
169 */
170#define KIOCB_CANCELLED		((void *) (~0ULL))
171
172struct aio_kiocb {
173	struct kiocb		common;
174
175	struct kioctx		*ki_ctx;
176	kiocb_cancel_fn		*ki_cancel;
177
178	struct iocb __user	*ki_user_iocb;	/* user's aiocb */
179	__u64			ki_user_data;	/* user's data for completion */
180
181	struct list_head	ki_list;	/* the aio core uses this
182						 * for cancellation */
183
184	/*
185	 * If the aio_resfd field of the userspace iocb is not zero,
186	 * this is the underlying eventfd context to deliver events to.
187	 */
188	struct eventfd_ctx	*ki_eventfd;
189};
190
191/*------ sysctl variables----*/
192static DEFINE_SPINLOCK(aio_nr_lock);
193unsigned long aio_nr;		/* current system wide number of aio requests */
194unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
195/*----end sysctl variables---*/
196
197static struct kmem_cache	*kiocb_cachep;
198static struct kmem_cache	*kioctx_cachep;
199
200static struct vfsmount *aio_mnt;
201
202static const struct file_operations aio_ring_fops;
203static const struct address_space_operations aio_ctx_aops;
204
205static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
206{
207	struct qstr this = QSTR_INIT("[aio]", 5);
208	struct file *file;
209	struct path path;
210	struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
211	if (IS_ERR(inode))
212		return ERR_CAST(inode);
213
214	inode->i_mapping->a_ops = &aio_ctx_aops;
215	inode->i_mapping->private_data = ctx;
216	inode->i_size = PAGE_SIZE * nr_pages;
217
218	path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
219	if (!path.dentry) {
220		iput(inode);
221		return ERR_PTR(-ENOMEM);
222	}
223	path.mnt = mntget(aio_mnt);
224
225	d_instantiate(path.dentry, inode);
226	file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
227	if (IS_ERR(file)) {
228		path_put(&path);
229		return file;
230	}
231
232	file->f_flags = O_RDWR;
233	return file;
234}
235
236static struct dentry *aio_mount(struct file_system_type *fs_type,
237				int flags, const char *dev_name, void *data)
238{
239	static const struct dentry_operations ops = {
240		.d_dname	= simple_dname,
241	};
242	return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
243}
244
245/* aio_setup
246 *	Creates the slab caches used by the aio routines, panic on
247 *	failure as this is done early during the boot sequence.
248 */
249static int __init aio_setup(void)
250{
251	static struct file_system_type aio_fs = {
252		.name		= "aio",
253		.mount		= aio_mount,
254		.kill_sb	= kill_anon_super,
255	};
256	aio_mnt = kern_mount(&aio_fs);
257	if (IS_ERR(aio_mnt))
258		panic("Failed to create aio fs mount.");
259
260	kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
261	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
262
263	pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
264
265	return 0;
266}
267__initcall(aio_setup);
268
269static void put_aio_ring_file(struct kioctx *ctx)
270{
271	struct file *aio_ring_file = ctx->aio_ring_file;
272	if (aio_ring_file) {
273		truncate_setsize(aio_ring_file->f_inode, 0);
274
275		/* Prevent further access to the kioctx from migratepages */
276		spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
277		aio_ring_file->f_inode->i_mapping->private_data = NULL;
278		ctx->aio_ring_file = NULL;
279		spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
280
281		fput(aio_ring_file);
282	}
283}
284
285static void aio_free_ring(struct kioctx *ctx)
286{
287	int i;
288
289	/* Disconnect the kiotx from the ring file.  This prevents future
290	 * accesses to the kioctx from page migration.
291	 */
292	put_aio_ring_file(ctx);
293
294	for (i = 0; i < ctx->nr_pages; i++) {
295		struct page *page;
296		pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
297				page_count(ctx->ring_pages[i]));
298		page = ctx->ring_pages[i];
299		if (!page)
300			continue;
301		ctx->ring_pages[i] = NULL;
302		put_page(page);
303	}
304
305	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
306		kfree(ctx->ring_pages);
307		ctx->ring_pages = NULL;
308	}
309}
310
311static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
312{
313	vma->vm_flags |= VM_DONTEXPAND;
314	vma->vm_ops = &generic_file_vm_ops;
315	return 0;
316}
317
318static int aio_ring_remap(struct file *file, struct vm_area_struct *vma)
319{
320	struct mm_struct *mm = vma->vm_mm;
321	struct kioctx_table *table;
322	int i, res = -EINVAL;
323
324	spin_lock(&mm->ioctx_lock);
325	rcu_read_lock();
326	table = rcu_dereference(mm->ioctx_table);
327	for (i = 0; i < table->nr; i++) {
328		struct kioctx *ctx;
329
330		ctx = table->table[i];
331		if (ctx && ctx->aio_ring_file == file) {
332			if (!atomic_read(&ctx->dead)) {
333				ctx->user_id = ctx->mmap_base = vma->vm_start;
334				res = 0;
335			}
336			break;
337		}
338	}
339
340	rcu_read_unlock();
341	spin_unlock(&mm->ioctx_lock);
342	return res;
343}
344
345static const struct file_operations aio_ring_fops = {
346	.mmap = aio_ring_mmap,
347	.mremap = aio_ring_remap,
348};
349
350#if IS_ENABLED(CONFIG_MIGRATION)
351static int aio_migratepage(struct address_space *mapping, struct page *new,
352			struct page *old, enum migrate_mode mode)
353{
354	struct kioctx *ctx;
355	unsigned long flags;
356	pgoff_t idx;
357	int rc;
358
359	rc = 0;
360
361	/* mapping->private_lock here protects against the kioctx teardown.  */
362	spin_lock(&mapping->private_lock);
363	ctx = mapping->private_data;
364	if (!ctx) {
365		rc = -EINVAL;
366		goto out;
367	}
368
369	/* The ring_lock mutex.  The prevents aio_read_events() from writing
370	 * to the ring's head, and prevents page migration from mucking in
371	 * a partially initialized kiotx.
372	 */
373	if (!mutex_trylock(&ctx->ring_lock)) {
374		rc = -EAGAIN;
375		goto out;
376	}
377
378	idx = old->index;
379	if (idx < (pgoff_t)ctx->nr_pages) {
380		/* Make sure the old page hasn't already been changed */
381		if (ctx->ring_pages[idx] != old)
382			rc = -EAGAIN;
383	} else
384		rc = -EINVAL;
385
386	if (rc != 0)
387		goto out_unlock;
388
389	/* Writeback must be complete */
390	BUG_ON(PageWriteback(old));
391	get_page(new);
392
393	rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
394	if (rc != MIGRATEPAGE_SUCCESS) {
395		put_page(new);
396		goto out_unlock;
397	}
398
399	/* Take completion_lock to prevent other writes to the ring buffer
400	 * while the old page is copied to the new.  This prevents new
401	 * events from being lost.
402	 */
403	spin_lock_irqsave(&ctx->completion_lock, flags);
404	migrate_page_copy(new, old);
405	BUG_ON(ctx->ring_pages[idx] != old);
406	ctx->ring_pages[idx] = new;
407	spin_unlock_irqrestore(&ctx->completion_lock, flags);
408
409	/* The old page is no longer accessible. */
410	put_page(old);
411
412out_unlock:
413	mutex_unlock(&ctx->ring_lock);
414out:
415	spin_unlock(&mapping->private_lock);
416	return rc;
417}
418#endif
419
420static const struct address_space_operations aio_ctx_aops = {
421	.set_page_dirty = __set_page_dirty_no_writeback,
422#if IS_ENABLED(CONFIG_MIGRATION)
423	.migratepage	= aio_migratepage,
424#endif
425};
426
427static int aio_setup_ring(struct kioctx *ctx)
428{
429	struct aio_ring *ring;
430	unsigned nr_events = ctx->max_reqs;
431	struct mm_struct *mm = current->mm;
432	unsigned long size, unused;
433	int nr_pages;
434	int i;
435	struct file *file;
436
437	/* Compensate for the ring buffer's head/tail overlap entry */
438	nr_events += 2;	/* 1 is required, 2 for good luck */
439
440	size = sizeof(struct aio_ring);
441	size += sizeof(struct io_event) * nr_events;
442
443	nr_pages = PFN_UP(size);
444	if (nr_pages < 0)
445		return -EINVAL;
446
447	file = aio_private_file(ctx, nr_pages);
448	if (IS_ERR(file)) {
449		ctx->aio_ring_file = NULL;
450		return -ENOMEM;
451	}
452
453	ctx->aio_ring_file = file;
454	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
455			/ sizeof(struct io_event);
456
457	ctx->ring_pages = ctx->internal_pages;
458	if (nr_pages > AIO_RING_PAGES) {
459		ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
460					  GFP_KERNEL);
461		if (!ctx->ring_pages) {
462			put_aio_ring_file(ctx);
463			return -ENOMEM;
464		}
465	}
466
467	for (i = 0; i < nr_pages; i++) {
468		struct page *page;
469		page = find_or_create_page(file->f_inode->i_mapping,
470					   i, GFP_HIGHUSER | __GFP_ZERO);
471		if (!page)
472			break;
473		pr_debug("pid(%d) page[%d]->count=%d\n",
474			 current->pid, i, page_count(page));
475		SetPageUptodate(page);
476		unlock_page(page);
477
478		ctx->ring_pages[i] = page;
479	}
480	ctx->nr_pages = i;
481
482	if (unlikely(i != nr_pages)) {
483		aio_free_ring(ctx);
484		return -ENOMEM;
485	}
486
487	ctx->mmap_size = nr_pages * PAGE_SIZE;
488	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
489
490	down_write(&mm->mmap_sem);
491	ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
492				       PROT_READ | PROT_WRITE,
493				       MAP_SHARED, 0, &unused);
494	up_write(&mm->mmap_sem);
495	if (IS_ERR((void *)ctx->mmap_base)) {
496		ctx->mmap_size = 0;
497		aio_free_ring(ctx);
498		return -ENOMEM;
499	}
500
501	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
502
503	ctx->user_id = ctx->mmap_base;
504	ctx->nr_events = nr_events; /* trusted copy */
505
506	ring = kmap_atomic(ctx->ring_pages[0]);
507	ring->nr = nr_events;	/* user copy */
508	ring->id = ~0U;
509	ring->head = ring->tail = 0;
510	ring->magic = AIO_RING_MAGIC;
511	ring->compat_features = AIO_RING_COMPAT_FEATURES;
512	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
513	ring->header_length = sizeof(struct aio_ring);
514	kunmap_atomic(ring);
515	flush_dcache_page(ctx->ring_pages[0]);
516
517	return 0;
518}
519
520#define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
521#define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
522#define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
523
524void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
525{
526	struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
527	struct kioctx *ctx = req->ki_ctx;
528	unsigned long flags;
529
530	spin_lock_irqsave(&ctx->ctx_lock, flags);
531
532	if (!req->ki_list.next)
533		list_add(&req->ki_list, &ctx->active_reqs);
534
535	req->ki_cancel = cancel;
536
537	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
538}
539EXPORT_SYMBOL(kiocb_set_cancel_fn);
540
541static int kiocb_cancel(struct aio_kiocb *kiocb)
542{
543	kiocb_cancel_fn *old, *cancel;
544
545	/*
546	 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
547	 * actually has a cancel function, hence the cmpxchg()
548	 */
549
550	cancel = ACCESS_ONCE(kiocb->ki_cancel);
551	do {
552		if (!cancel || cancel == KIOCB_CANCELLED)
553			return -EINVAL;
554
555		old = cancel;
556		cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
557	} while (cancel != old);
558
559	return cancel(&kiocb->common);
560}
561
562static void free_ioctx(struct work_struct *work)
563{
564	struct kioctx *ctx = container_of(work, struct kioctx, free_work);
565
566	pr_debug("freeing %p\n", ctx);
567
568	aio_free_ring(ctx);
569	free_percpu(ctx->cpu);
570	percpu_ref_exit(&ctx->reqs);
571	percpu_ref_exit(&ctx->users);
572	kmem_cache_free(kioctx_cachep, ctx);
573}
574
575static void free_ioctx_reqs(struct percpu_ref *ref)
576{
577	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
578
579	/* At this point we know that there are no any in-flight requests */
580	if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
581		complete(&ctx->rq_wait->comp);
582
583	INIT_WORK(&ctx->free_work, free_ioctx);
584	schedule_work(&ctx->free_work);
585}
586
587/*
588 * When this function runs, the kioctx has been removed from the "hash table"
589 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
590 * now it's safe to cancel any that need to be.
591 */
592static void free_ioctx_users(struct percpu_ref *ref)
593{
594	struct kioctx *ctx = container_of(ref, struct kioctx, users);
595	struct aio_kiocb *req;
596
597	spin_lock_irq(&ctx->ctx_lock);
598
599	while (!list_empty(&ctx->active_reqs)) {
600		req = list_first_entry(&ctx->active_reqs,
601				       struct aio_kiocb, ki_list);
602
603		list_del_init(&req->ki_list);
604		kiocb_cancel(req);
605	}
606
607	spin_unlock_irq(&ctx->ctx_lock);
608
609	percpu_ref_kill(&ctx->reqs);
610	percpu_ref_put(&ctx->reqs);
611}
612
613static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
614{
615	unsigned i, new_nr;
616	struct kioctx_table *table, *old;
617	struct aio_ring *ring;
618
619	spin_lock(&mm->ioctx_lock);
620	table = rcu_dereference_raw(mm->ioctx_table);
621
622	while (1) {
623		if (table)
624			for (i = 0; i < table->nr; i++)
625				if (!table->table[i]) {
626					ctx->id = i;
627					table->table[i] = ctx;
628					spin_unlock(&mm->ioctx_lock);
629
630					/* While kioctx setup is in progress,
631					 * we are protected from page migration
632					 * changes ring_pages by ->ring_lock.
633					 */
634					ring = kmap_atomic(ctx->ring_pages[0]);
635					ring->id = ctx->id;
636					kunmap_atomic(ring);
637					return 0;
638				}
639
640		new_nr = (table ? table->nr : 1) * 4;
641		spin_unlock(&mm->ioctx_lock);
642
643		table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
644				new_nr, GFP_KERNEL);
645		if (!table)
646			return -ENOMEM;
647
648		table->nr = new_nr;
649
650		spin_lock(&mm->ioctx_lock);
651		old = rcu_dereference_raw(mm->ioctx_table);
652
653		if (!old) {
654			rcu_assign_pointer(mm->ioctx_table, table);
655		} else if (table->nr > old->nr) {
656			memcpy(table->table, old->table,
657			       old->nr * sizeof(struct kioctx *));
658
659			rcu_assign_pointer(mm->ioctx_table, table);
660			kfree_rcu(old, rcu);
661		} else {
662			kfree(table);
663			table = old;
664		}
665	}
666}
667
668static void aio_nr_sub(unsigned nr)
669{
670	spin_lock(&aio_nr_lock);
671	if (WARN_ON(aio_nr - nr > aio_nr))
672		aio_nr = 0;
673	else
674		aio_nr -= nr;
675	spin_unlock(&aio_nr_lock);
676}
677
678/* ioctx_alloc
679 *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
680 */
681static struct kioctx *ioctx_alloc(unsigned nr_events)
682{
683	struct mm_struct *mm = current->mm;
684	struct kioctx *ctx;
685	int err = -ENOMEM;
686
687	/*
688	 * We keep track of the number of available ringbuffer slots, to prevent
689	 * overflow (reqs_available), and we also use percpu counters for this.
690	 *
691	 * So since up to half the slots might be on other cpu's percpu counters
692	 * and unavailable, double nr_events so userspace sees what they
693	 * expected: additionally, we move req_batch slots to/from percpu
694	 * counters at a time, so make sure that isn't 0:
695	 */
696	nr_events = max(nr_events, num_possible_cpus() * 4);
697	nr_events *= 2;
698
699	/* Prevent overflows */
700	if (nr_events > (0x10000000U / sizeof(struct io_event))) {
701		pr_debug("ENOMEM: nr_events too high\n");
702		return ERR_PTR(-EINVAL);
703	}
704
705	if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
706		return ERR_PTR(-EAGAIN);
707
708	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
709	if (!ctx)
710		return ERR_PTR(-ENOMEM);
711
712	ctx->max_reqs = nr_events;
713
714	spin_lock_init(&ctx->ctx_lock);
715	spin_lock_init(&ctx->completion_lock);
716	mutex_init(&ctx->ring_lock);
717	/* Protect against page migration throughout kiotx setup by keeping
718	 * the ring_lock mutex held until setup is complete. */
719	mutex_lock(&ctx->ring_lock);
720	init_waitqueue_head(&ctx->wait);
721
722	INIT_LIST_HEAD(&ctx->active_reqs);
723
724	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
725		goto err;
726
727	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
728		goto err;
729
730	ctx->cpu = alloc_percpu(struct kioctx_cpu);
731	if (!ctx->cpu)
732		goto err;
733
734	err = aio_setup_ring(ctx);
735	if (err < 0)
736		goto err;
737
738	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
739	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
740	if (ctx->req_batch < 1)
741		ctx->req_batch = 1;
742
743	/* limit the number of system wide aios */
744	spin_lock(&aio_nr_lock);
745	if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
746	    aio_nr + nr_events < aio_nr) {
747		spin_unlock(&aio_nr_lock);
748		err = -EAGAIN;
749		goto err_ctx;
750	}
751	aio_nr += ctx->max_reqs;
752	spin_unlock(&aio_nr_lock);
753
754	percpu_ref_get(&ctx->users);	/* io_setup() will drop this ref */
755	percpu_ref_get(&ctx->reqs);	/* free_ioctx_users() will drop this */
756
757	err = ioctx_add_table(ctx, mm);
758	if (err)
759		goto err_cleanup;
760
761	/* Release the ring_lock mutex now that all setup is complete. */
762	mutex_unlock(&ctx->ring_lock);
763
764	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
765		 ctx, ctx->user_id, mm, ctx->nr_events);
766	return ctx;
767
768err_cleanup:
769	aio_nr_sub(ctx->max_reqs);
770err_ctx:
771	atomic_set(&ctx->dead, 1);
772	if (ctx->mmap_size)
773		vm_munmap(ctx->mmap_base, ctx->mmap_size);
774	aio_free_ring(ctx);
775err:
776	mutex_unlock(&ctx->ring_lock);
777	free_percpu(ctx->cpu);
778	percpu_ref_exit(&ctx->reqs);
779	percpu_ref_exit(&ctx->users);
780	kmem_cache_free(kioctx_cachep, ctx);
781	pr_debug("error allocating ioctx %d\n", err);
782	return ERR_PTR(err);
783}
784
785/* kill_ioctx
786 *	Cancels all outstanding aio requests on an aio context.  Used
787 *	when the processes owning a context have all exited to encourage
788 *	the rapid destruction of the kioctx.
789 */
790static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
791		      struct ctx_rq_wait *wait)
792{
793	struct kioctx_table *table;
794
795	spin_lock(&mm->ioctx_lock);
796	if (atomic_xchg(&ctx->dead, 1)) {
797		spin_unlock(&mm->ioctx_lock);
798		return -EINVAL;
799	}
800
801	table = rcu_dereference_raw(mm->ioctx_table);
802	WARN_ON(ctx != table->table[ctx->id]);
803	table->table[ctx->id] = NULL;
804	spin_unlock(&mm->ioctx_lock);
805
806	/* percpu_ref_kill() will do the necessary call_rcu() */
807	wake_up_all(&ctx->wait);
808
809	/*
810	 * It'd be more correct to do this in free_ioctx(), after all
811	 * the outstanding kiocbs have finished - but by then io_destroy
812	 * has already returned, so io_setup() could potentially return
813	 * -EAGAIN with no ioctxs actually in use (as far as userspace
814	 *  could tell).
815	 */
816	aio_nr_sub(ctx->max_reqs);
817
818	if (ctx->mmap_size)
819		vm_munmap(ctx->mmap_base, ctx->mmap_size);
820
821	ctx->rq_wait = wait;
822	percpu_ref_kill(&ctx->users);
823	return 0;
824}
825
826/*
827 * exit_aio: called when the last user of mm goes away.  At this point, there is
828 * no way for any new requests to be submited or any of the io_* syscalls to be
829 * called on the context.
830 *
831 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
832 * them.
833 */
834void exit_aio(struct mm_struct *mm)
835{
836	struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
837	struct ctx_rq_wait wait;
838	int i, skipped;
839
840	if (!table)
841		return;
842
843	atomic_set(&wait.count, table->nr);
844	init_completion(&wait.comp);
845
846	skipped = 0;
847	for (i = 0; i < table->nr; ++i) {
848		struct kioctx *ctx = table->table[i];
849
850		if (!ctx) {
851			skipped++;
852			continue;
853		}
854
855		/*
856		 * We don't need to bother with munmap() here - exit_mmap(mm)
857		 * is coming and it'll unmap everything. And we simply can't,
858		 * this is not necessarily our ->mm.
859		 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
860		 * that it needs to unmap the area, just set it to 0.
861		 */
862		ctx->mmap_size = 0;
863		kill_ioctx(mm, ctx, &wait);
864	}
865
866	if (!atomic_sub_and_test(skipped, &wait.count)) {
867		/* Wait until all IO for the context are done. */
868		wait_for_completion(&wait.comp);
869	}
870
871	RCU_INIT_POINTER(mm->ioctx_table, NULL);
872	kfree(table);
873}
874
875static void put_reqs_available(struct kioctx *ctx, unsigned nr)
876{
877	struct kioctx_cpu *kcpu;
878	unsigned long flags;
879
880	local_irq_save(flags);
881	kcpu = this_cpu_ptr(ctx->cpu);
882	kcpu->reqs_available += nr;
883
884	while (kcpu->reqs_available >= ctx->req_batch * 2) {
885		kcpu->reqs_available -= ctx->req_batch;
886		atomic_add(ctx->req_batch, &ctx->reqs_available);
887	}
888
889	local_irq_restore(flags);
890}
891
892static bool get_reqs_available(struct kioctx *ctx)
893{
894	struct kioctx_cpu *kcpu;
895	bool ret = false;
896	unsigned long flags;
897
898	local_irq_save(flags);
899	kcpu = this_cpu_ptr(ctx->cpu);
900	if (!kcpu->reqs_available) {
901		int old, avail = atomic_read(&ctx->reqs_available);
902
903		do {
904			if (avail < ctx->req_batch)
905				goto out;
906
907			old = avail;
908			avail = atomic_cmpxchg(&ctx->reqs_available,
909					       avail, avail - ctx->req_batch);
910		} while (avail != old);
911
912		kcpu->reqs_available += ctx->req_batch;
913	}
914
915	ret = true;
916	kcpu->reqs_available--;
917out:
918	local_irq_restore(flags);
919	return ret;
920}
921
922/* refill_reqs_available
923 *	Updates the reqs_available reference counts used for tracking the
924 *	number of free slots in the completion ring.  This can be called
925 *	from aio_complete() (to optimistically update reqs_available) or
926 *	from aio_get_req() (the we're out of events case).  It must be
927 *	called holding ctx->completion_lock.
928 */
929static void refill_reqs_available(struct kioctx *ctx, unsigned head,
930                                  unsigned tail)
931{
932	unsigned events_in_ring, completed;
933
934	/* Clamp head since userland can write to it. */
935	head %= ctx->nr_events;
936	if (head <= tail)
937		events_in_ring = tail - head;
938	else
939		events_in_ring = ctx->nr_events - (head - tail);
940
941	completed = ctx->completed_events;
942	if (events_in_ring < completed)
943		completed -= events_in_ring;
944	else
945		completed = 0;
946
947	if (!completed)
948		return;
949
950	ctx->completed_events -= completed;
951	put_reqs_available(ctx, completed);
952}
953
954/* user_refill_reqs_available
955 *	Called to refill reqs_available when aio_get_req() encounters an
956 *	out of space in the completion ring.
957 */
958static void user_refill_reqs_available(struct kioctx *ctx)
959{
960	spin_lock_irq(&ctx->completion_lock);
961	if (ctx->completed_events) {
962		struct aio_ring *ring;
963		unsigned head;
964
965		/* Access of ring->head may race with aio_read_events_ring()
966		 * here, but that's okay since whether we read the old version
967		 * or the new version, and either will be valid.  The important
968		 * part is that head cannot pass tail since we prevent
969		 * aio_complete() from updating tail by holding
970		 * ctx->completion_lock.  Even if head is invalid, the check
971		 * against ctx->completed_events below will make sure we do the
972		 * safe/right thing.
973		 */
974		ring = kmap_atomic(ctx->ring_pages[0]);
975		head = ring->head;
976		kunmap_atomic(ring);
977
978		refill_reqs_available(ctx, head, ctx->tail);
979	}
980
981	spin_unlock_irq(&ctx->completion_lock);
982}
983
984/* aio_get_req
985 *	Allocate a slot for an aio request.
986 * Returns NULL if no requests are free.
987 */
988static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
989{
990	struct aio_kiocb *req;
991
992	if (!get_reqs_available(ctx)) {
993		user_refill_reqs_available(ctx);
994		if (!get_reqs_available(ctx))
995			return NULL;
996	}
997
998	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
999	if (unlikely(!req))
1000		goto out_put;
1001
1002	percpu_ref_get(&ctx->reqs);
1003
1004	req->ki_ctx = ctx;
1005	return req;
1006out_put:
1007	put_reqs_available(ctx, 1);
1008	return NULL;
1009}
1010
1011static void kiocb_free(struct aio_kiocb *req)
1012{
1013	if (req->common.ki_filp)
1014		fput(req->common.ki_filp);
1015	if (req->ki_eventfd != NULL)
1016		eventfd_ctx_put(req->ki_eventfd);
1017	kmem_cache_free(kiocb_cachep, req);
1018}
1019
1020static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1021{
1022	struct aio_ring __user *ring  = (void __user *)ctx_id;
1023	struct mm_struct *mm = current->mm;
1024	struct kioctx *ctx, *ret = NULL;
1025	struct kioctx_table *table;
1026	unsigned id;
1027
1028	if (get_user(id, &ring->id))
1029		return NULL;
1030
1031	rcu_read_lock();
1032	table = rcu_dereference(mm->ioctx_table);
1033
1034	if (!table || id >= table->nr)
1035		goto out;
1036
1037	ctx = table->table[id];
1038	if (ctx && ctx->user_id == ctx_id) {
1039		percpu_ref_get(&ctx->users);
1040		ret = ctx;
1041	}
1042out:
1043	rcu_read_unlock();
1044	return ret;
1045}
1046
1047/* aio_complete
1048 *	Called when the io request on the given iocb is complete.
1049 */
1050static void aio_complete(struct kiocb *kiocb, long res, long res2)
1051{
1052	struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1053	struct kioctx	*ctx = iocb->ki_ctx;
1054	struct aio_ring	*ring;
1055	struct io_event	*ev_page, *event;
1056	unsigned tail, pos, head;
1057	unsigned long	flags;
1058
1059	/*
1060	 * Special case handling for sync iocbs:
1061	 *  - events go directly into the iocb for fast handling
1062	 *  - the sync task with the iocb in its stack holds the single iocb
1063	 *    ref, no other paths have a way to get another ref
1064	 *  - the sync task helpfully left a reference to itself in the iocb
1065	 */
1066	BUG_ON(is_sync_kiocb(kiocb));
1067
1068	if (iocb->ki_list.next) {
1069		unsigned long flags;
1070
1071		spin_lock_irqsave(&ctx->ctx_lock, flags);
1072		list_del(&iocb->ki_list);
1073		spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1074	}
1075
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	 * pointer since we might be called from irq context.
1080	 */
1081	spin_lock_irqsave(&ctx->completion_lock, flags);
1082
1083	tail = ctx->tail;
1084	pos = tail + AIO_EVENTS_OFFSET;
1085
1086	if (++tail >= ctx->nr_events)
1087		tail = 0;
1088
1089	ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1090	event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1091
1092	event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1093	event->data = iocb->ki_user_data;
1094	event->res = res;
1095	event->res2 = res2;
1096
1097	kunmap_atomic(ev_page);
1098	flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1099
1100	pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1101		 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1102		 res, res2);
1103
1104	/* after flagging the request as done, we
1105	 * must never even look at it again
1106	 */
1107	smp_wmb();	/* make event visible before updating tail */
1108
1109	ctx->tail = tail;
1110
1111	ring = kmap_atomic(ctx->ring_pages[0]);
1112	head = ring->head;
1113	ring->tail = tail;
1114	kunmap_atomic(ring);
1115	flush_dcache_page(ctx->ring_pages[0]);
1116
1117	ctx->completed_events++;
1118	if (ctx->completed_events > 1)
1119		refill_reqs_available(ctx, head, tail);
1120	spin_unlock_irqrestore(&ctx->completion_lock, flags);
1121
1122	pr_debug("added to ring %p at [%u]\n", iocb, tail);
1123
1124	/*
1125	 * Check if the user asked us to deliver the result through an
1126	 * eventfd. The eventfd_signal() function is safe to be called
1127	 * from IRQ context.
1128	 */
1129	if (iocb->ki_eventfd != NULL)
1130		eventfd_signal(iocb->ki_eventfd, 1);
1131
1132	/* everything turned out well, dispose of the aiocb. */
1133	kiocb_free(iocb);
1134
1135	/*
1136	 * We have to order our ring_info tail store above and test
1137	 * of the wait list below outside the wait lock.  This is
1138	 * like in wake_up_bit() where clearing a bit has to be
1139	 * ordered with the unlocked test.
1140	 */
1141	smp_mb();
1142
1143	if (waitqueue_active(&ctx->wait))
1144		wake_up(&ctx->wait);
1145
1146	percpu_ref_put(&ctx->reqs);
1147}
1148
1149/* aio_read_events_ring
1150 *	Pull an event off of the ioctx's event ring.  Returns the number of
1151 *	events fetched
1152 */
1153static long aio_read_events_ring(struct kioctx *ctx,
1154				 struct io_event __user *event, long nr)
1155{
1156	struct aio_ring *ring;
1157	unsigned head, tail, pos;
1158	long ret = 0;
1159	int copy_ret;
1160
1161	/*
1162	 * The mutex can block and wake us up and that will cause
1163	 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1164	 * and repeat. This should be rare enough that it doesn't cause
1165	 * peformance issues. See the comment in read_events() for more detail.
1166	 */
1167	sched_annotate_sleep();
1168	mutex_lock(&ctx->ring_lock);
1169
1170	/* Access to ->ring_pages here is protected by ctx->ring_lock. */
1171	ring = kmap_atomic(ctx->ring_pages[0]);
1172	head = ring->head;
1173	tail = ring->tail;
1174	kunmap_atomic(ring);
1175
1176	/*
1177	 * Ensure that once we've read the current tail pointer, that
1178	 * we also see the events that were stored up to the tail.
1179	 */
1180	smp_rmb();
1181
1182	pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1183
1184	if (head == tail)
1185		goto out;
1186
1187	head %= ctx->nr_events;
1188	tail %= ctx->nr_events;
1189
1190	while (ret < nr) {
1191		long avail;
1192		struct io_event *ev;
1193		struct page *page;
1194
1195		avail = (head <= tail ?  tail : ctx->nr_events) - head;
1196		if (head == tail)
1197			break;
1198
1199		avail = min(avail, nr - ret);
1200		avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1201			    ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1202
1203		pos = head + AIO_EVENTS_OFFSET;
1204		page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1205		pos %= AIO_EVENTS_PER_PAGE;
1206
1207		ev = kmap(page);
1208		copy_ret = copy_to_user(event + ret, ev + pos,
1209					sizeof(*ev) * avail);
1210		kunmap(page);
1211
1212		if (unlikely(copy_ret)) {
1213			ret = -EFAULT;
1214			goto out;
1215		}
1216
1217		ret += avail;
1218		head += avail;
1219		head %= ctx->nr_events;
1220	}
1221
1222	ring = kmap_atomic(ctx->ring_pages[0]);
1223	ring->head = head;
1224	kunmap_atomic(ring);
1225	flush_dcache_page(ctx->ring_pages[0]);
1226
1227	pr_debug("%li  h%u t%u\n", ret, head, tail);
1228out:
1229	mutex_unlock(&ctx->ring_lock);
1230
1231	return ret;
1232}
1233
1234static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1235			    struct io_event __user *event, long *i)
1236{
1237	long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1238
1239	if (ret > 0)
1240		*i += ret;
1241
1242	if (unlikely(atomic_read(&ctx->dead)))
1243		ret = -EINVAL;
1244
1245	if (!*i)
1246		*i = ret;
1247
1248	return ret < 0 || *i >= min_nr;
1249}
1250
1251static long read_events(struct kioctx *ctx, long min_nr, long nr,
1252			struct io_event __user *event,
1253			struct timespec __user *timeout)
1254{
1255	ktime_t until = { .tv64 = KTIME_MAX };
1256	long ret = 0;
1257
1258	if (timeout) {
1259		struct timespec	ts;
1260
1261		if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1262			return -EFAULT;
1263
1264		until = timespec_to_ktime(ts);
1265	}
1266
1267	/*
1268	 * Note that aio_read_events() is being called as the conditional - i.e.
1269	 * we're calling it after prepare_to_wait() has set task state to
1270	 * TASK_INTERRUPTIBLE.
1271	 *
1272	 * But aio_read_events() can block, and if it blocks it's going to flip
1273	 * the task state back to TASK_RUNNING.
1274	 *
1275	 * This should be ok, provided it doesn't flip the state back to
1276	 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1277	 * will only happen if the mutex_lock() call blocks, and we then find
1278	 * the ringbuffer empty. So in practice we should be ok, but it's
1279	 * something to be aware of when touching this code.
1280	 */
1281	if (until.tv64 == 0)
1282		aio_read_events(ctx, min_nr, nr, event, &ret);
1283	else
1284		wait_event_interruptible_hrtimeout(ctx->wait,
1285				aio_read_events(ctx, min_nr, nr, event, &ret),
1286				until);
1287
1288	if (!ret && signal_pending(current))
1289		ret = -EINTR;
1290
1291	return ret;
1292}
1293
1294/* sys_io_setup:
1295 *	Create an aio_context capable of receiving at least nr_events.
1296 *	ctxp must not point to an aio_context that already exists, and
1297 *	must be initialized to 0 prior to the call.  On successful
1298 *	creation of the aio_context, *ctxp is filled in with the resulting
1299 *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1300 *	if the specified nr_events exceeds internal limits.  May fail
1301 *	with -EAGAIN if the specified nr_events exceeds the user's limit
1302 *	of available events.  May fail with -ENOMEM if insufficient kernel
1303 *	resources are available.  May fail with -EFAULT if an invalid
1304 *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1305 *	implemented.
1306 */
1307SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1308{
1309	struct kioctx *ioctx = NULL;
1310	unsigned long ctx;
1311	long ret;
1312
1313	ret = get_user(ctx, ctxp);
1314	if (unlikely(ret))
1315		goto out;
1316
1317	ret = -EINVAL;
1318	if (unlikely(ctx || nr_events == 0)) {
1319		pr_debug("EINVAL: ctx %lu nr_events %u\n",
1320		         ctx, nr_events);
1321		goto out;
1322	}
1323
1324	ioctx = ioctx_alloc(nr_events);
1325	ret = PTR_ERR(ioctx);
1326	if (!IS_ERR(ioctx)) {
1327		ret = put_user(ioctx->user_id, ctxp);
1328		if (ret)
1329			kill_ioctx(current->mm, ioctx, NULL);
1330		percpu_ref_put(&ioctx->users);
1331	}
1332
1333out:
1334	return ret;
1335}
1336
1337/* sys_io_destroy:
1338 *	Destroy the aio_context specified.  May cancel any outstanding
1339 *	AIOs and block on completion.  Will fail with -ENOSYS if not
1340 *	implemented.  May fail with -EINVAL if the context pointed to
1341 *	is invalid.
1342 */
1343SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1344{
1345	struct kioctx *ioctx = lookup_ioctx(ctx);
1346	if (likely(NULL != ioctx)) {
1347		struct ctx_rq_wait wait;
1348		int ret;
1349
1350		init_completion(&wait.comp);
1351		atomic_set(&wait.count, 1);
1352
1353		/* Pass requests_done to kill_ioctx() where it can be set
1354		 * in a thread-safe way. If we try to set it here then we have
1355		 * a race condition if two io_destroy() called simultaneously.
1356		 */
1357		ret = kill_ioctx(current->mm, ioctx, &wait);
1358		percpu_ref_put(&ioctx->users);
1359
1360		/* Wait until all IO for the context are done. Otherwise kernel
1361		 * keep using user-space buffers even if user thinks the context
1362		 * is destroyed.
1363		 */
1364		if (!ret)
1365			wait_for_completion(&wait.comp);
1366
1367		return ret;
1368	}
1369	pr_debug("EINVAL: invalid context id\n");
1370	return -EINVAL;
1371}
1372
1373typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1374
1375static int aio_setup_vectored_rw(int rw, char __user *buf, size_t len,
1376				 struct iovec **iovec,
1377				 bool compat,
1378				 struct iov_iter *iter)
1379{
1380#ifdef CONFIG_COMPAT
1381	if (compat)
1382		return compat_import_iovec(rw,
1383				(struct compat_iovec __user *)buf,
1384				len, UIO_FASTIOV, iovec, iter);
1385#endif
1386	return import_iovec(rw, (struct iovec __user *)buf,
1387				len, UIO_FASTIOV, iovec, iter);
1388}
1389
1390/*
1391 * aio_run_iocb:
1392 *	Performs the initial checks and io submission.
1393 */
1394static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1395			    char __user *buf, size_t len, bool compat)
1396{
1397	struct file *file = req->ki_filp;
1398	ssize_t ret;
1399	int rw;
1400	fmode_t mode;
1401	rw_iter_op *iter_op;
1402	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1403	struct iov_iter iter;
1404
1405	switch (opcode) {
1406	case IOCB_CMD_PREAD:
1407	case IOCB_CMD_PREADV:
1408		mode	= FMODE_READ;
1409		rw	= READ;
1410		iter_op	= file->f_op->read_iter;
1411		goto rw_common;
1412
1413	case IOCB_CMD_PWRITE:
1414	case IOCB_CMD_PWRITEV:
1415		mode	= FMODE_WRITE;
1416		rw	= WRITE;
1417		iter_op	= file->f_op->write_iter;
1418		goto rw_common;
1419rw_common:
1420		if (unlikely(!(file->f_mode & mode)))
1421			return -EBADF;
1422
1423		if (!iter_op)
1424			return -EINVAL;
1425
1426		if (opcode == IOCB_CMD_PREADV || opcode == IOCB_CMD_PWRITEV)
1427			ret = aio_setup_vectored_rw(rw, buf, len,
1428						&iovec, compat, &iter);
1429		else {
1430			ret = import_single_range(rw, buf, len, iovec, &iter);
1431			iovec = NULL;
1432		}
1433		if (!ret)
1434			ret = rw_verify_area(rw, file, &req->ki_pos,
1435					     iov_iter_count(&iter));
1436		if (ret < 0) {
1437			kfree(iovec);
1438			return ret;
1439		}
1440
1441		len = ret;
1442
1443		if (rw == WRITE)
1444			file_start_write(file);
1445
1446		ret = iter_op(req, &iter);
1447
1448		if (rw == WRITE)
1449			file_end_write(file);
1450		kfree(iovec);
1451		break;
1452
1453	case IOCB_CMD_FDSYNC:
1454		if (!file->f_op->aio_fsync)
1455			return -EINVAL;
1456
1457		ret = file->f_op->aio_fsync(req, 1);
1458		break;
1459
1460	case IOCB_CMD_FSYNC:
1461		if (!file->f_op->aio_fsync)
1462			return -EINVAL;
1463
1464		ret = file->f_op->aio_fsync(req, 0);
1465		break;
1466
1467	default:
1468		pr_debug("EINVAL: no operation provided\n");
1469		return -EINVAL;
1470	}
1471
1472	if (ret != -EIOCBQUEUED) {
1473		/*
1474		 * There's no easy way to restart the syscall since other AIO's
1475		 * may be already running. Just fail this IO with EINTR.
1476		 */
1477		if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1478			     ret == -ERESTARTNOHAND ||
1479			     ret == -ERESTART_RESTARTBLOCK))
1480			ret = -EINTR;
1481		aio_complete(req, ret, 0);
1482	}
1483
1484	return 0;
1485}
1486
1487static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1488			 struct iocb *iocb, bool compat)
1489{
1490	struct aio_kiocb *req;
1491	ssize_t ret;
1492
1493	/* enforce forwards compatibility on users */
1494	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1495		pr_debug("EINVAL: reserve field set\n");
1496		return -EINVAL;
1497	}
1498
1499	/* prevent overflows */
1500	if (unlikely(
1501	    (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1502	    (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1503	    ((ssize_t)iocb->aio_nbytes < 0)
1504	   )) {
1505		pr_debug("EINVAL: overflow check\n");
1506		return -EINVAL;
1507	}
1508
1509	req = aio_get_req(ctx);
1510	if (unlikely(!req))
1511		return -EAGAIN;
1512
1513	req->common.ki_filp = fget(iocb->aio_fildes);
1514	if (unlikely(!req->common.ki_filp)) {
1515		ret = -EBADF;
1516		goto out_put_req;
1517	}
1518	req->common.ki_pos = iocb->aio_offset;
1519	req->common.ki_complete = aio_complete;
1520	req->common.ki_flags = iocb_flags(req->common.ki_filp);
1521
1522	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1523		/*
1524		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1525		 * instance of the file* now. The file descriptor must be
1526		 * an eventfd() fd, and will be signaled for each completed
1527		 * event using the eventfd_signal() function.
1528		 */
1529		req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1530		if (IS_ERR(req->ki_eventfd)) {
1531			ret = PTR_ERR(req->ki_eventfd);
1532			req->ki_eventfd = NULL;
1533			goto out_put_req;
1534		}
1535
1536		req->common.ki_flags |= IOCB_EVENTFD;
1537	}
1538
1539	ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1540	if (unlikely(ret)) {
1541		pr_debug("EFAULT: aio_key\n");
1542		goto out_put_req;
1543	}
1544
1545	req->ki_user_iocb = user_iocb;
1546	req->ki_user_data = iocb->aio_data;
1547
1548	ret = aio_run_iocb(&req->common, iocb->aio_lio_opcode,
1549			   (char __user *)(unsigned long)iocb->aio_buf,
1550			   iocb->aio_nbytes,
1551			   compat);
1552	if (ret)
1553		goto out_put_req;
1554
1555	return 0;
1556out_put_req:
1557	put_reqs_available(ctx, 1);
1558	percpu_ref_put(&ctx->reqs);
1559	kiocb_free(req);
1560	return ret;
1561}
1562
1563long do_io_submit(aio_context_t ctx_id, long nr,
1564		  struct iocb __user *__user *iocbpp, bool compat)
1565{
1566	struct kioctx *ctx;
1567	long ret = 0;
1568	int i = 0;
1569	struct blk_plug plug;
1570
1571	if (unlikely(nr < 0))
1572		return -EINVAL;
1573
1574	if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1575		nr = LONG_MAX/sizeof(*iocbpp);
1576
1577	if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1578		return -EFAULT;
1579
1580	ctx = lookup_ioctx(ctx_id);
1581	if (unlikely(!ctx)) {
1582		pr_debug("EINVAL: invalid context id\n");
1583		return -EINVAL;
1584	}
1585
1586	blk_start_plug(&plug);
1587
1588	/*
1589	 * AKPM: should this return a partial result if some of the IOs were
1590	 * successfully submitted?
1591	 */
1592	for (i=0; i<nr; i++) {
1593		struct iocb __user *user_iocb;
1594		struct iocb tmp;
1595
1596		if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1597			ret = -EFAULT;
1598			break;
1599		}
1600
1601		if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1602			ret = -EFAULT;
1603			break;
1604		}
1605
1606		ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1607		if (ret)
1608			break;
1609	}
1610	blk_finish_plug(&plug);
1611
1612	percpu_ref_put(&ctx->users);
1613	return i ? i : ret;
1614}
1615
1616/* sys_io_submit:
1617 *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1618 *	the number of iocbs queued.  May return -EINVAL if the aio_context
1619 *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
1620 *	*iocbpp[0] is not properly initialized, if the operation specified
1621 *	is invalid for the file descriptor in the iocb.  May fail with
1622 *	-EFAULT if any of the data structures point to invalid data.  May
1623 *	fail with -EBADF if the file descriptor specified in the first
1624 *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
1625 *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1626 *	fail with -ENOSYS if not implemented.
1627 */
1628SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1629		struct iocb __user * __user *, iocbpp)
1630{
1631	return do_io_submit(ctx_id, nr, iocbpp, 0);
1632}
1633
1634/* lookup_kiocb
1635 *	Finds a given iocb for cancellation.
1636 */
1637static struct aio_kiocb *
1638lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1639{
1640	struct aio_kiocb *kiocb;
1641
1642	assert_spin_locked(&ctx->ctx_lock);
1643
1644	if (key != KIOCB_KEY)
1645		return NULL;
1646
1647	/* TODO: use a hash or array, this sucks. */
1648	list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1649		if (kiocb->ki_user_iocb == iocb)
1650			return kiocb;
1651	}
1652	return NULL;
1653}
1654
1655/* sys_io_cancel:
1656 *	Attempts to cancel an iocb previously passed to io_submit.  If
1657 *	the operation is successfully cancelled, the resulting event is
1658 *	copied into the memory pointed to by result without being placed
1659 *	into the completion queue and 0 is returned.  May fail with
1660 *	-EFAULT if any of the data structures pointed to are invalid.
1661 *	May fail with -EINVAL if aio_context specified by ctx_id is
1662 *	invalid.  May fail with -EAGAIN if the iocb specified was not
1663 *	cancelled.  Will fail with -ENOSYS if not implemented.
1664 */
1665SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1666		struct io_event __user *, result)
1667{
1668	struct kioctx *ctx;
1669	struct aio_kiocb *kiocb;
1670	u32 key;
1671	int ret;
1672
1673	ret = get_user(key, &iocb->aio_key);
1674	if (unlikely(ret))
1675		return -EFAULT;
1676
1677	ctx = lookup_ioctx(ctx_id);
1678	if (unlikely(!ctx))
1679		return -EINVAL;
1680
1681	spin_lock_irq(&ctx->ctx_lock);
1682
1683	kiocb = lookup_kiocb(ctx, iocb, key);
1684	if (kiocb)
1685		ret = kiocb_cancel(kiocb);
1686	else
1687		ret = -EINVAL;
1688
1689	spin_unlock_irq(&ctx->ctx_lock);
1690
1691	if (!ret) {
1692		/*
1693		 * The result argument is no longer used - the io_event is
1694		 * always delivered via the ring buffer. -EINPROGRESS indicates
1695		 * cancellation is progress:
1696		 */
1697		ret = -EINPROGRESS;
1698	}
1699
1700	percpu_ref_put(&ctx->users);
1701
1702	return ret;
1703}
1704
1705/* io_getevents:
1706 *	Attempts to read at least min_nr events and up to nr events from
1707 *	the completion queue for the aio_context specified by ctx_id. If
1708 *	it succeeds, the number of read events is returned. May fail with
1709 *	-EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1710 *	out of range, if timeout is out of range.  May fail with -EFAULT
1711 *	if any of the memory specified is invalid.  May return 0 or
1712 *	< min_nr if the timeout specified by timeout has elapsed
1713 *	before sufficient events are available, where timeout == NULL
1714 *	specifies an infinite timeout. Note that the timeout pointed to by
1715 *	timeout is relative.  Will fail with -ENOSYS if not implemented.
1716 */
1717SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1718		long, min_nr,
1719		long, nr,
1720		struct io_event __user *, events,
1721		struct timespec __user *, timeout)
1722{
1723	struct kioctx *ioctx = lookup_ioctx(ctx_id);
1724	long ret = -EINVAL;
1725
1726	if (likely(ioctx)) {
1727		if (likely(min_nr <= nr && min_nr >= 0))
1728			ret = read_events(ioctx, min_nr, nr, events, timeout);
1729		percpu_ref_put(&ioctx->users);
1730	}
1731	return ret;
1732}
1733