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 *	Copyright 2018 Christoph Hellwig.
9 *
10 *	See ../COPYING for licensing terms.
11 */
12#define pr_fmt(fmt) "%s: " fmt, __func__
13
14#include <linux/kernel.h>
15#include <linux/init.h>
16#include <linux/errno.h>
17#include <linux/time.h>
18#include <linux/aio_abi.h>
19#include <linux/export.h>
20#include <linux/syscalls.h>
21#include <linux/backing-dev.h>
22#include <linux/refcount.h>
23#include <linux/uio.h>
24
25#include <linux/sched/signal.h>
26#include <linux/fs.h>
27#include <linux/file.h>
28#include <linux/mm.h>
29#include <linux/mman.h>
30#include <linux/percpu.h>
31#include <linux/slab.h>
32#include <linux/timer.h>
33#include <linux/aio.h>
34#include <linux/highmem.h>
35#include <linux/workqueue.h>
36#include <linux/security.h>
37#include <linux/eventfd.h>
38#include <linux/blkdev.h>
39#include <linux/compat.h>
40#include <linux/migrate.h>
41#include <linux/ramfs.h>
42#include <linux/percpu-refcount.h>
43#include <linux/mount.h>
44#include <linux/pseudo_fs.h>
45
46#include <linux/uaccess.h>
47#include <linux/nospec.h>
48
49#include "internal.h"
50
51#define KIOCB_KEY		0
52
53#define AIO_RING_MAGIC			0xa10a10a1
54#define AIO_RING_COMPAT_FEATURES	1
55#define AIO_RING_INCOMPAT_FEATURES	0
56struct aio_ring {
57	unsigned	id;	/* kernel internal index number */
58	unsigned	nr;	/* number of io_events */
59	unsigned	head;	/* Written to by userland or under ring_lock
60				 * mutex by aio_read_events_ring(). */
61	unsigned	tail;
62
63	unsigned	magic;
64	unsigned	compat_features;
65	unsigned	incompat_features;
66	unsigned	header_length;	/* size of aio_ring */
67
68
69	struct io_event		io_events[];
70}; /* 128 bytes + ring size */
71
72/*
73 * Plugging is meant to work with larger batches of IOs. If we don't
74 * have more than the below, then don't bother setting up a plug.
75 */
76#define AIO_PLUG_THRESHOLD	2
77
78#define AIO_RING_PAGES	8
79
80struct kioctx_table {
81	struct rcu_head		rcu;
82	unsigned		nr;
83	struct kioctx __rcu	*table[];
84};
85
86struct kioctx_cpu {
87	unsigned		reqs_available;
88};
89
90struct ctx_rq_wait {
91	struct completion comp;
92	atomic_t count;
93};
94
95struct kioctx {
96	struct percpu_ref	users;
97	atomic_t		dead;
98
99	struct percpu_ref	reqs;
100
101	unsigned long		user_id;
102
103	struct __percpu kioctx_cpu *cpu;
104
105	/*
106	 * For percpu reqs_available, number of slots we move to/from global
107	 * counter at a time:
108	 */
109	unsigned		req_batch;
110	/*
111	 * This is what userspace passed to io_setup(), it's not used for
112	 * anything but counting against the global max_reqs quota.
113	 *
114	 * The real limit is nr_events - 1, which will be larger (see
115	 * aio_setup_ring())
116	 */
117	unsigned		max_reqs;
118
119	/* Size of ringbuffer, in units of struct io_event */
120	unsigned		nr_events;
121
122	unsigned long		mmap_base;
123	unsigned long		mmap_size;
124
125	struct page		**ring_pages;
126	long			nr_pages;
127
128	struct rcu_work		free_rwork;	/* see free_ioctx() */
129
130	/*
131	 * signals when all in-flight requests are done
132	 */
133	struct ctx_rq_wait	*rq_wait;
134
135	struct {
136		/*
137		 * This counts the number of available slots in the ringbuffer,
138		 * so we avoid overflowing it: it's decremented (if positive)
139		 * when allocating a kiocb and incremented when the resulting
140		 * io_event is pulled off the ringbuffer.
141		 *
142		 * We batch accesses to it with a percpu version.
143		 */
144		atomic_t	reqs_available;
145	} ____cacheline_aligned_in_smp;
146
147	struct {
148		spinlock_t	ctx_lock;
149		struct list_head active_reqs;	/* used for cancellation */
150	} ____cacheline_aligned_in_smp;
151
152	struct {
153		struct mutex	ring_lock;
154		wait_queue_head_t wait;
155	} ____cacheline_aligned_in_smp;
156
157	struct {
158		unsigned	tail;
159		unsigned	completed_events;
160		spinlock_t	completion_lock;
161	} ____cacheline_aligned_in_smp;
162
163	struct page		*internal_pages[AIO_RING_PAGES];
164	struct file		*aio_ring_file;
165
166	unsigned		id;
167};
168
169/*
170 * First field must be the file pointer in all the
171 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
172 */
173struct fsync_iocb {
174	struct file		*file;
175	struct work_struct	work;
176	bool			datasync;
177	struct cred		*creds;
178};
179
180struct poll_iocb {
181	struct file		*file;
182	struct wait_queue_head	*head;
183	__poll_t		events;
184	bool			cancelled;
185	bool			work_scheduled;
186	bool			work_need_resched;
187	struct wait_queue_entry	wait;
188	struct work_struct	work;
189};
190
191/*
192 * NOTE! Each of the iocb union members has the file pointer
193 * as the first entry in their struct definition. So you can
194 * access the file pointer through any of the sub-structs,
195 * or directly as just 'ki_filp' in this struct.
196 */
197struct aio_kiocb {
198	union {
199		struct file		*ki_filp;
200		struct kiocb		rw;
201		struct fsync_iocb	fsync;
202		struct poll_iocb	poll;
203	};
204
205	struct kioctx		*ki_ctx;
206	kiocb_cancel_fn		*ki_cancel;
207
208	struct io_event		ki_res;
209
210	struct list_head	ki_list;	/* the aio core uses this
211						 * for cancellation */
212	refcount_t		ki_refcnt;
213
214	/*
215	 * If the aio_resfd field of the userspace iocb is not zero,
216	 * this is the underlying eventfd context to deliver events to.
217	 */
218	struct eventfd_ctx	*ki_eventfd;
219};
220
221/*------ sysctl variables----*/
222static DEFINE_SPINLOCK(aio_nr_lock);
223static unsigned long aio_nr;		/* current system wide number of aio requests */
224static unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
225/*----end sysctl variables---*/
226#ifdef CONFIG_SYSCTL
227static struct ctl_table aio_sysctls[] = {
228	{
229		.procname	= "aio-nr",
230		.data		= &aio_nr,
231		.maxlen		= sizeof(aio_nr),
232		.mode		= 0444,
233		.proc_handler	= proc_doulongvec_minmax,
234	},
235	{
236		.procname	= "aio-max-nr",
237		.data		= &aio_max_nr,
238		.maxlen		= sizeof(aio_max_nr),
239		.mode		= 0644,
240		.proc_handler	= proc_doulongvec_minmax,
241	},
242	{}
243};
244
245static void __init aio_sysctl_init(void)
246{
247	register_sysctl_init("fs", aio_sysctls);
248}
249#else
250#define aio_sysctl_init() do { } while (0)
251#endif
252
253static struct kmem_cache	*kiocb_cachep;
254static struct kmem_cache	*kioctx_cachep;
255
256static struct vfsmount *aio_mnt;
257
258static const struct file_operations aio_ring_fops;
259static const struct address_space_operations aio_ctx_aops;
260
261static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
262{
263	struct file *file;
264	struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
265	if (IS_ERR(inode))
266		return ERR_CAST(inode);
267
268	inode->i_mapping->a_ops = &aio_ctx_aops;
269	inode->i_mapping->private_data = ctx;
270	inode->i_size = PAGE_SIZE * nr_pages;
271
272	file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
273				O_RDWR, &aio_ring_fops);
274	if (IS_ERR(file))
275		iput(inode);
276	return file;
277}
278
279static int aio_init_fs_context(struct fs_context *fc)
280{
281	if (!init_pseudo(fc, AIO_RING_MAGIC))
282		return -ENOMEM;
283	fc->s_iflags |= SB_I_NOEXEC;
284	return 0;
285}
286
287/* aio_setup
288 *	Creates the slab caches used by the aio routines, panic on
289 *	failure as this is done early during the boot sequence.
290 */
291static int __init aio_setup(void)
292{
293	static struct file_system_type aio_fs = {
294		.name		= "aio",
295		.init_fs_context = aio_init_fs_context,
296		.kill_sb	= kill_anon_super,
297	};
298	aio_mnt = kern_mount(&aio_fs);
299	if (IS_ERR(aio_mnt))
300		panic("Failed to create aio fs mount.");
301
302	kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
303	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
304	aio_sysctl_init();
305	return 0;
306}
307__initcall(aio_setup);
308
309static void put_aio_ring_file(struct kioctx *ctx)
310{
311	struct file *aio_ring_file = ctx->aio_ring_file;
312	struct address_space *i_mapping;
313
314	if (aio_ring_file) {
315		truncate_setsize(file_inode(aio_ring_file), 0);
316
317		/* Prevent further access to the kioctx from migratepages */
318		i_mapping = aio_ring_file->f_mapping;
319		spin_lock(&i_mapping->private_lock);
320		i_mapping->private_data = NULL;
321		ctx->aio_ring_file = NULL;
322		spin_unlock(&i_mapping->private_lock);
323
324		fput(aio_ring_file);
325	}
326}
327
328static void aio_free_ring(struct kioctx *ctx)
329{
330	int i;
331
332	/* Disconnect the kiotx from the ring file.  This prevents future
333	 * accesses to the kioctx from page migration.
334	 */
335	put_aio_ring_file(ctx);
336
337	for (i = 0; i < ctx->nr_pages; i++) {
338		struct page *page;
339		pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
340				page_count(ctx->ring_pages[i]));
341		page = ctx->ring_pages[i];
342		if (!page)
343			continue;
344		ctx->ring_pages[i] = NULL;
345		put_page(page);
346	}
347
348	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
349		kfree(ctx->ring_pages);
350		ctx->ring_pages = NULL;
351	}
352}
353
354static int aio_ring_mremap(struct vm_area_struct *vma)
355{
356	struct file *file = vma->vm_file;
357	struct mm_struct *mm = vma->vm_mm;
358	struct kioctx_table *table;
359	int i, res = -EINVAL;
360
361	spin_lock(&mm->ioctx_lock);
362	rcu_read_lock();
363	table = rcu_dereference(mm->ioctx_table);
364	for (i = 0; i < table->nr; i++) {
365		struct kioctx *ctx;
366
367		ctx = rcu_dereference(table->table[i]);
368		if (ctx && ctx->aio_ring_file == file) {
369			if (!atomic_read(&ctx->dead)) {
370				ctx->user_id = ctx->mmap_base = vma->vm_start;
371				res = 0;
372			}
373			break;
374		}
375	}
376
377	rcu_read_unlock();
378	spin_unlock(&mm->ioctx_lock);
379	return res;
380}
381
382static const struct vm_operations_struct aio_ring_vm_ops = {
383	.mremap		= aio_ring_mremap,
384#if IS_ENABLED(CONFIG_MMU)
385	.fault		= filemap_fault,
386	.map_pages	= filemap_map_pages,
387	.page_mkwrite	= filemap_page_mkwrite,
388#endif
389};
390
391static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
392{
393	vma->vm_flags |= VM_DONTEXPAND;
394	vma->vm_ops = &aio_ring_vm_ops;
395	return 0;
396}
397
398static const struct file_operations aio_ring_fops = {
399	.mmap = aio_ring_mmap,
400};
401
402#if IS_ENABLED(CONFIG_MIGRATION)
403static int aio_migratepage(struct address_space *mapping, struct page *new,
404			struct page *old, enum migrate_mode mode)
405{
406	struct kioctx *ctx;
407	unsigned long flags;
408	pgoff_t idx;
409	int rc;
410
411	/*
412	 * We cannot support the _NO_COPY case here, because copy needs to
413	 * happen under the ctx->completion_lock. That does not work with the
414	 * migration workflow of MIGRATE_SYNC_NO_COPY.
415	 */
416	if (mode == MIGRATE_SYNC_NO_COPY)
417		return -EINVAL;
418
419	rc = 0;
420
421	/* mapping->private_lock here protects against the kioctx teardown.  */
422	spin_lock(&mapping->private_lock);
423	ctx = mapping->private_data;
424	if (!ctx) {
425		rc = -EINVAL;
426		goto out;
427	}
428
429	/* The ring_lock mutex.  The prevents aio_read_events() from writing
430	 * to the ring's head, and prevents page migration from mucking in
431	 * a partially initialized kiotx.
432	 */
433	if (!mutex_trylock(&ctx->ring_lock)) {
434		rc = -EAGAIN;
435		goto out;
436	}
437
438	idx = old->index;
439	if (idx < (pgoff_t)ctx->nr_pages) {
440		/* Make sure the old page hasn't already been changed */
441		if (ctx->ring_pages[idx] != old)
442			rc = -EAGAIN;
443	} else
444		rc = -EINVAL;
445
446	if (rc != 0)
447		goto out_unlock;
448
449	/* Writeback must be complete */
450	BUG_ON(PageWriteback(old));
451	get_page(new);
452
453	rc = migrate_page_move_mapping(mapping, new, old, 1);
454	if (rc != MIGRATEPAGE_SUCCESS) {
455		put_page(new);
456		goto out_unlock;
457	}
458
459	/* Take completion_lock to prevent other writes to the ring buffer
460	 * while the old page is copied to the new.  This prevents new
461	 * events from being lost.
462	 */
463	spin_lock_irqsave(&ctx->completion_lock, flags);
464	migrate_page_copy(new, old);
465	BUG_ON(ctx->ring_pages[idx] != old);
466	ctx->ring_pages[idx] = new;
467	spin_unlock_irqrestore(&ctx->completion_lock, flags);
468
469	/* The old page is no longer accessible. */
470	put_page(old);
471
472out_unlock:
473	mutex_unlock(&ctx->ring_lock);
474out:
475	spin_unlock(&mapping->private_lock);
476	return rc;
477}
478#endif
479
480static const struct address_space_operations aio_ctx_aops = {
481	.set_page_dirty = __set_page_dirty_no_writeback,
482#if IS_ENABLED(CONFIG_MIGRATION)
483	.migratepage	= aio_migratepage,
484#endif
485};
486
487static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
488{
489	struct aio_ring *ring;
490	struct mm_struct *mm = current->mm;
491	unsigned long size, unused;
492	int nr_pages;
493	int i;
494	struct file *file;
495
496	/* Compensate for the ring buffer's head/tail overlap entry */
497	nr_events += 2;	/* 1 is required, 2 for good luck */
498
499	size = sizeof(struct aio_ring);
500	size += sizeof(struct io_event) * nr_events;
501
502	nr_pages = PFN_UP(size);
503	if (nr_pages < 0)
504		return -EINVAL;
505
506	file = aio_private_file(ctx, nr_pages);
507	if (IS_ERR(file)) {
508		ctx->aio_ring_file = NULL;
509		return -ENOMEM;
510	}
511
512	ctx->aio_ring_file = file;
513	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
514			/ sizeof(struct io_event);
515
516	ctx->ring_pages = ctx->internal_pages;
517	if (nr_pages > AIO_RING_PAGES) {
518		ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
519					  GFP_KERNEL);
520		if (!ctx->ring_pages) {
521			put_aio_ring_file(ctx);
522			return -ENOMEM;
523		}
524	}
525
526	for (i = 0; i < nr_pages; i++) {
527		struct page *page;
528		page = find_or_create_page(file->f_mapping,
529					   i, GFP_HIGHUSER | __GFP_ZERO);
530		if (!page)
531			break;
532		pr_debug("pid(%d) page[%d]->count=%d\n",
533			 current->pid, i, page_count(page));
534		SetPageUptodate(page);
535		unlock_page(page);
536
537		ctx->ring_pages[i] = page;
538	}
539	ctx->nr_pages = i;
540
541	if (unlikely(i != nr_pages)) {
542		aio_free_ring(ctx);
543		return -ENOMEM;
544	}
545
546	ctx->mmap_size = nr_pages * PAGE_SIZE;
547	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
548
549	if (mmap_write_lock_killable(mm)) {
550		ctx->mmap_size = 0;
551		aio_free_ring(ctx);
552		return -EINTR;
553	}
554
555	ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
556				 PROT_READ | PROT_WRITE,
557				 MAP_SHARED, 0, &unused, NULL);
558	mmap_write_unlock(mm);
559	if (IS_ERR((void *)ctx->mmap_base)) {
560		ctx->mmap_size = 0;
561		aio_free_ring(ctx);
562		return -ENOMEM;
563	}
564
565	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
566
567	ctx->user_id = ctx->mmap_base;
568	ctx->nr_events = nr_events; /* trusted copy */
569
570	ring = kmap_atomic(ctx->ring_pages[0]);
571	ring->nr = nr_events;	/* user copy */
572	ring->id = ~0U;
573	ring->head = ring->tail = 0;
574	ring->magic = AIO_RING_MAGIC;
575	ring->compat_features = AIO_RING_COMPAT_FEATURES;
576	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
577	ring->header_length = sizeof(struct aio_ring);
578	kunmap_atomic(ring);
579	flush_dcache_page(ctx->ring_pages[0]);
580
581	return 0;
582}
583
584#define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
585#define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
586#define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
587
588void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
589{
590	struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
591	struct kioctx *ctx = req->ki_ctx;
592	unsigned long flags;
593
594	if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
595		return;
596
597	spin_lock_irqsave(&ctx->ctx_lock, flags);
598	list_add_tail(&req->ki_list, &ctx->active_reqs);
599	req->ki_cancel = cancel;
600	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
601}
602EXPORT_SYMBOL(kiocb_set_cancel_fn);
603
604/*
605 * free_ioctx() should be RCU delayed to synchronize against the RCU
606 * protected lookup_ioctx() and also needs process context to call
607 * aio_free_ring().  Use rcu_work.
608 */
609static void free_ioctx(struct work_struct *work)
610{
611	struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
612					  free_rwork);
613	pr_debug("freeing %p\n", ctx);
614
615	aio_free_ring(ctx);
616	free_percpu(ctx->cpu);
617	percpu_ref_exit(&ctx->reqs);
618	percpu_ref_exit(&ctx->users);
619	kmem_cache_free(kioctx_cachep, ctx);
620}
621
622static void free_ioctx_reqs(struct percpu_ref *ref)
623{
624	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
625
626	/* At this point we know that there are no any in-flight requests */
627	if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
628		complete(&ctx->rq_wait->comp);
629
630	/* Synchronize against RCU protected table->table[] dereferences */
631	INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
632	queue_rcu_work(system_wq, &ctx->free_rwork);
633}
634
635/*
636 * When this function runs, the kioctx has been removed from the "hash table"
637 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
638 * now it's safe to cancel any that need to be.
639 */
640static void free_ioctx_users(struct percpu_ref *ref)
641{
642	struct kioctx *ctx = container_of(ref, struct kioctx, users);
643	struct aio_kiocb *req;
644
645	spin_lock_irq(&ctx->ctx_lock);
646
647	while (!list_empty(&ctx->active_reqs)) {
648		req = list_first_entry(&ctx->active_reqs,
649				       struct aio_kiocb, ki_list);
650		req->ki_cancel(&req->rw);
651		list_del_init(&req->ki_list);
652	}
653
654	spin_unlock_irq(&ctx->ctx_lock);
655
656	percpu_ref_kill(&ctx->reqs);
657	percpu_ref_put(&ctx->reqs);
658}
659
660static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
661{
662	unsigned i, new_nr;
663	struct kioctx_table *table, *old;
664	struct aio_ring *ring;
665
666	spin_lock(&mm->ioctx_lock);
667	table = rcu_dereference_raw(mm->ioctx_table);
668
669	while (1) {
670		if (table)
671			for (i = 0; i < table->nr; i++)
672				if (!rcu_access_pointer(table->table[i])) {
673					ctx->id = i;
674					rcu_assign_pointer(table->table[i], ctx);
675					spin_unlock(&mm->ioctx_lock);
676
677					/* While kioctx setup is in progress,
678					 * we are protected from page migration
679					 * changes ring_pages by ->ring_lock.
680					 */
681					ring = kmap_atomic(ctx->ring_pages[0]);
682					ring->id = ctx->id;
683					kunmap_atomic(ring);
684					return 0;
685				}
686
687		new_nr = (table ? table->nr : 1) * 4;
688		spin_unlock(&mm->ioctx_lock);
689
690		table = kzalloc(struct_size(table, table, new_nr), GFP_KERNEL);
691		if (!table)
692			return -ENOMEM;
693
694		table->nr = new_nr;
695
696		spin_lock(&mm->ioctx_lock);
697		old = rcu_dereference_raw(mm->ioctx_table);
698
699		if (!old) {
700			rcu_assign_pointer(mm->ioctx_table, table);
701		} else if (table->nr > old->nr) {
702			memcpy(table->table, old->table,
703			       old->nr * sizeof(struct kioctx *));
704
705			rcu_assign_pointer(mm->ioctx_table, table);
706			kfree_rcu(old, rcu);
707		} else {
708			kfree(table);
709			table = old;
710		}
711	}
712}
713
714static void aio_nr_sub(unsigned nr)
715{
716	spin_lock(&aio_nr_lock);
717	if (WARN_ON(aio_nr - nr > aio_nr))
718		aio_nr = 0;
719	else
720		aio_nr -= nr;
721	spin_unlock(&aio_nr_lock);
722}
723
724/* ioctx_alloc
725 *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
726 */
727static struct kioctx *ioctx_alloc(unsigned nr_events)
728{
729	struct mm_struct *mm = current->mm;
730	struct kioctx *ctx;
731	int err = -ENOMEM;
732
733	/*
734	 * Store the original nr_events -- what userspace passed to io_setup(),
735	 * for counting against the global limit -- before it changes.
736	 */
737	unsigned int max_reqs = nr_events;
738
739	/*
740	 * We keep track of the number of available ringbuffer slots, to prevent
741	 * overflow (reqs_available), and we also use percpu counters for this.
742	 *
743	 * So since up to half the slots might be on other cpu's percpu counters
744	 * and unavailable, double nr_events so userspace sees what they
745	 * expected: additionally, we move req_batch slots to/from percpu
746	 * counters at a time, so make sure that isn't 0:
747	 */
748	nr_events = max(nr_events, num_possible_cpus() * 4);
749	nr_events *= 2;
750
751	/* Prevent overflows */
752	if (nr_events > (0x10000000U / sizeof(struct io_event))) {
753		pr_debug("ENOMEM: nr_events too high\n");
754		return ERR_PTR(-EINVAL);
755	}
756
757	if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
758		return ERR_PTR(-EAGAIN);
759
760	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
761	if (!ctx)
762		return ERR_PTR(-ENOMEM);
763
764	ctx->max_reqs = max_reqs;
765
766	spin_lock_init(&ctx->ctx_lock);
767	spin_lock_init(&ctx->completion_lock);
768	mutex_init(&ctx->ring_lock);
769	/* Protect against page migration throughout kiotx setup by keeping
770	 * the ring_lock mutex held until setup is complete. */
771	mutex_lock(&ctx->ring_lock);
772	init_waitqueue_head(&ctx->wait);
773
774	INIT_LIST_HEAD(&ctx->active_reqs);
775
776	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
777		goto err;
778
779	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
780		goto err;
781
782	ctx->cpu = alloc_percpu(struct kioctx_cpu);
783	if (!ctx->cpu)
784		goto err;
785
786	err = aio_setup_ring(ctx, nr_events);
787	if (err < 0)
788		goto err;
789
790	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
791	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
792	if (ctx->req_batch < 1)
793		ctx->req_batch = 1;
794
795	/* limit the number of system wide aios */
796	spin_lock(&aio_nr_lock);
797	if (aio_nr + ctx->max_reqs > aio_max_nr ||
798	    aio_nr + ctx->max_reqs < aio_nr) {
799		spin_unlock(&aio_nr_lock);
800		err = -EAGAIN;
801		goto err_ctx;
802	}
803	aio_nr += ctx->max_reqs;
804	spin_unlock(&aio_nr_lock);
805
806	percpu_ref_get(&ctx->users);	/* io_setup() will drop this ref */
807	percpu_ref_get(&ctx->reqs);	/* free_ioctx_users() will drop this */
808
809	err = ioctx_add_table(ctx, mm);
810	if (err)
811		goto err_cleanup;
812
813	/* Release the ring_lock mutex now that all setup is complete. */
814	mutex_unlock(&ctx->ring_lock);
815
816	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
817		 ctx, ctx->user_id, mm, ctx->nr_events);
818	return ctx;
819
820err_cleanup:
821	aio_nr_sub(ctx->max_reqs);
822err_ctx:
823	atomic_set(&ctx->dead, 1);
824	if (ctx->mmap_size)
825		vm_munmap(ctx->mmap_base, ctx->mmap_size);
826	aio_free_ring(ctx);
827err:
828	mutex_unlock(&ctx->ring_lock);
829	free_percpu(ctx->cpu);
830	percpu_ref_exit(&ctx->reqs);
831	percpu_ref_exit(&ctx->users);
832	kmem_cache_free(kioctx_cachep, ctx);
833	pr_debug("error allocating ioctx %d\n", err);
834	return ERR_PTR(err);
835}
836
837/* kill_ioctx
838 *	Cancels all outstanding aio requests on an aio context.  Used
839 *	when the processes owning a context have all exited to encourage
840 *	the rapid destruction of the kioctx.
841 */
842static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
843		      struct ctx_rq_wait *wait)
844{
845	struct kioctx_table *table;
846
847	spin_lock(&mm->ioctx_lock);
848	if (atomic_xchg(&ctx->dead, 1)) {
849		spin_unlock(&mm->ioctx_lock);
850		return -EINVAL;
851	}
852
853	table = rcu_dereference_raw(mm->ioctx_table);
854	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
855	RCU_INIT_POINTER(table->table[ctx->id], NULL);
856	spin_unlock(&mm->ioctx_lock);
857
858	/* free_ioctx_reqs() will do the necessary RCU synchronization */
859	wake_up_all(&ctx->wait);
860
861	/*
862	 * It'd be more correct to do this in free_ioctx(), after all
863	 * the outstanding kiocbs have finished - but by then io_destroy
864	 * has already returned, so io_setup() could potentially return
865	 * -EAGAIN with no ioctxs actually in use (as far as userspace
866	 *  could tell).
867	 */
868	aio_nr_sub(ctx->max_reqs);
869
870	if (ctx->mmap_size)
871		vm_munmap(ctx->mmap_base, ctx->mmap_size);
872
873	ctx->rq_wait = wait;
874	percpu_ref_kill(&ctx->users);
875	return 0;
876}
877
878/*
879 * exit_aio: called when the last user of mm goes away.  At this point, there is
880 * no way for any new requests to be submited or any of the io_* syscalls to be
881 * called on the context.
882 *
883 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
884 * them.
885 */
886void exit_aio(struct mm_struct *mm)
887{
888	struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
889	struct ctx_rq_wait wait;
890	int i, skipped;
891
892	if (!table)
893		return;
894
895	atomic_set(&wait.count, table->nr);
896	init_completion(&wait.comp);
897
898	skipped = 0;
899	for (i = 0; i < table->nr; ++i) {
900		struct kioctx *ctx =
901			rcu_dereference_protected(table->table[i], true);
902
903		if (!ctx) {
904			skipped++;
905			continue;
906		}
907
908		/*
909		 * We don't need to bother with munmap() here - exit_mmap(mm)
910		 * is coming and it'll unmap everything. And we simply can't,
911		 * this is not necessarily our ->mm.
912		 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
913		 * that it needs to unmap the area, just set it to 0.
914		 */
915		ctx->mmap_size = 0;
916		kill_ioctx(mm, ctx, &wait);
917	}
918
919	if (!atomic_sub_and_test(skipped, &wait.count)) {
920		/* Wait until all IO for the context are done. */
921		wait_for_completion(&wait.comp);
922	}
923
924	RCU_INIT_POINTER(mm->ioctx_table, NULL);
925	kfree(table);
926}
927
928static void put_reqs_available(struct kioctx *ctx, unsigned nr)
929{
930	struct kioctx_cpu *kcpu;
931	unsigned long flags;
932
933	local_irq_save(flags);
934	kcpu = this_cpu_ptr(ctx->cpu);
935	kcpu->reqs_available += nr;
936
937	while (kcpu->reqs_available >= ctx->req_batch * 2) {
938		kcpu->reqs_available -= ctx->req_batch;
939		atomic_add(ctx->req_batch, &ctx->reqs_available);
940	}
941
942	local_irq_restore(flags);
943}
944
945static bool __get_reqs_available(struct kioctx *ctx)
946{
947	struct kioctx_cpu *kcpu;
948	bool ret = false;
949	unsigned long flags;
950
951	local_irq_save(flags);
952	kcpu = this_cpu_ptr(ctx->cpu);
953	if (!kcpu->reqs_available) {
954		int old, avail = atomic_read(&ctx->reqs_available);
955
956		do {
957			if (avail < ctx->req_batch)
958				goto out;
959
960			old = avail;
961			avail = atomic_cmpxchg(&ctx->reqs_available,
962					       avail, avail - ctx->req_batch);
963		} while (avail != old);
964
965		kcpu->reqs_available += ctx->req_batch;
966	}
967
968	ret = true;
969	kcpu->reqs_available--;
970out:
971	local_irq_restore(flags);
972	return ret;
973}
974
975/* refill_reqs_available
976 *	Updates the reqs_available reference counts used for tracking the
977 *	number of free slots in the completion ring.  This can be called
978 *	from aio_complete() (to optimistically update reqs_available) or
979 *	from aio_get_req() (the we're out of events case).  It must be
980 *	called holding ctx->completion_lock.
981 */
982static void refill_reqs_available(struct kioctx *ctx, unsigned head,
983                                  unsigned tail)
984{
985	unsigned events_in_ring, completed;
986
987	/* Clamp head since userland can write to it. */
988	head %= ctx->nr_events;
989	if (head <= tail)
990		events_in_ring = tail - head;
991	else
992		events_in_ring = ctx->nr_events - (head - tail);
993
994	completed = ctx->completed_events;
995	if (events_in_ring < completed)
996		completed -= events_in_ring;
997	else
998		completed = 0;
999
1000	if (!completed)
1001		return;
1002
1003	ctx->completed_events -= completed;
1004	put_reqs_available(ctx, completed);
1005}
1006
1007/* user_refill_reqs_available
1008 *	Called to refill reqs_available when aio_get_req() encounters an
1009 *	out of space in the completion ring.
1010 */
1011static void user_refill_reqs_available(struct kioctx *ctx)
1012{
1013	spin_lock_irq(&ctx->completion_lock);
1014	if (ctx->completed_events) {
1015		struct aio_ring *ring;
1016		unsigned head;
1017
1018		/* Access of ring->head may race with aio_read_events_ring()
1019		 * here, but that's okay since whether we read the old version
1020		 * or the new version, and either will be valid.  The important
1021		 * part is that head cannot pass tail since we prevent
1022		 * aio_complete() from updating tail by holding
1023		 * ctx->completion_lock.  Even if head is invalid, the check
1024		 * against ctx->completed_events below will make sure we do the
1025		 * safe/right thing.
1026		 */
1027		ring = kmap_atomic(ctx->ring_pages[0]);
1028		head = ring->head;
1029		kunmap_atomic(ring);
1030
1031		refill_reqs_available(ctx, head, ctx->tail);
1032	}
1033
1034	spin_unlock_irq(&ctx->completion_lock);
1035}
1036
1037static bool get_reqs_available(struct kioctx *ctx)
1038{
1039	if (__get_reqs_available(ctx))
1040		return true;
1041	user_refill_reqs_available(ctx);
1042	return __get_reqs_available(ctx);
1043}
1044
1045/* aio_get_req
1046 *	Allocate a slot for an aio request.
1047 * Returns NULL if no requests are free.
1048 *
1049 * The refcount is initialized to 2 - one for the async op completion,
1050 * one for the synchronous code that does this.
1051 */
1052static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1053{
1054	struct aio_kiocb *req;
1055
1056	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1057	if (unlikely(!req))
1058		return NULL;
1059
1060	if (unlikely(!get_reqs_available(ctx))) {
1061		kmem_cache_free(kiocb_cachep, req);
1062		return NULL;
1063	}
1064
1065	percpu_ref_get(&ctx->reqs);
1066	req->ki_ctx = ctx;
1067	INIT_LIST_HEAD(&req->ki_list);
1068	refcount_set(&req->ki_refcnt, 2);
1069	req->ki_eventfd = NULL;
1070	return req;
1071}
1072
1073static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1074{
1075	struct aio_ring __user *ring  = (void __user *)ctx_id;
1076	struct mm_struct *mm = current->mm;
1077	struct kioctx *ctx, *ret = NULL;
1078	struct kioctx_table *table;
1079	unsigned id;
1080
1081	if (get_user(id, &ring->id))
1082		return NULL;
1083
1084	rcu_read_lock();
1085	table = rcu_dereference(mm->ioctx_table);
1086
1087	if (!table || id >= table->nr)
1088		goto out;
1089
1090	id = array_index_nospec(id, table->nr);
1091	ctx = rcu_dereference(table->table[id]);
1092	if (ctx && ctx->user_id == ctx_id) {
1093		if (percpu_ref_tryget_live(&ctx->users))
1094			ret = ctx;
1095	}
1096out:
1097	rcu_read_unlock();
1098	return ret;
1099}
1100
1101static inline void iocb_destroy(struct aio_kiocb *iocb)
1102{
1103	if (iocb->ki_eventfd)
1104		eventfd_ctx_put(iocb->ki_eventfd);
1105	if (iocb->ki_filp)
1106		fput(iocb->ki_filp);
1107	percpu_ref_put(&iocb->ki_ctx->reqs);
1108	kmem_cache_free(kiocb_cachep, iocb);
1109}
1110
1111/* aio_complete
1112 *	Called when the io request on the given iocb is complete.
1113 */
1114static void aio_complete(struct aio_kiocb *iocb)
1115{
1116	struct kioctx	*ctx = iocb->ki_ctx;
1117	struct aio_ring	*ring;
1118	struct io_event	*ev_page, *event;
1119	unsigned tail, pos, head;
1120	unsigned long	flags;
1121
1122	/*
1123	 * Add a completion event to the ring buffer. Must be done holding
1124	 * ctx->completion_lock to prevent other code from messing with the tail
1125	 * pointer since we might be called from irq context.
1126	 */
1127	spin_lock_irqsave(&ctx->completion_lock, flags);
1128
1129	tail = ctx->tail;
1130	pos = tail + AIO_EVENTS_OFFSET;
1131
1132	if (++tail >= ctx->nr_events)
1133		tail = 0;
1134
1135	ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1136	event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1137
1138	*event = iocb->ki_res;
1139
1140	kunmap_atomic(ev_page);
1141	flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1142
1143	pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1144		 (void __user *)(unsigned long)iocb->ki_res.obj,
1145		 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1146
1147	/* after flagging the request as done, we
1148	 * must never even look at it again
1149	 */
1150	smp_wmb();	/* make event visible before updating tail */
1151
1152	ctx->tail = tail;
1153
1154	ring = kmap_atomic(ctx->ring_pages[0]);
1155	head = ring->head;
1156	ring->tail = tail;
1157	kunmap_atomic(ring);
1158	flush_dcache_page(ctx->ring_pages[0]);
1159
1160	ctx->completed_events++;
1161	if (ctx->completed_events > 1)
1162		refill_reqs_available(ctx, head, tail);
1163	spin_unlock_irqrestore(&ctx->completion_lock, flags);
1164
1165	pr_debug("added to ring %p at [%u]\n", iocb, tail);
1166
1167	/*
1168	 * Check if the user asked us to deliver the result through an
1169	 * eventfd. The eventfd_signal() function is safe to be called
1170	 * from IRQ context.
1171	 */
1172	if (iocb->ki_eventfd)
1173		eventfd_signal(iocb->ki_eventfd, 1);
1174
1175	/*
1176	 * We have to order our ring_info tail store above and test
1177	 * of the wait list below outside the wait lock.  This is
1178	 * like in wake_up_bit() where clearing a bit has to be
1179	 * ordered with the unlocked test.
1180	 */
1181	smp_mb();
1182
1183	if (waitqueue_active(&ctx->wait))
1184		wake_up(&ctx->wait);
1185}
1186
1187static inline void iocb_put(struct aio_kiocb *iocb)
1188{
1189	if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1190		aio_complete(iocb);
1191		iocb_destroy(iocb);
1192	}
1193}
1194
1195/* aio_read_events_ring
1196 *	Pull an event off of the ioctx's event ring.  Returns the number of
1197 *	events fetched
1198 */
1199static long aio_read_events_ring(struct kioctx *ctx,
1200				 struct io_event __user *event, long nr)
1201{
1202	struct aio_ring *ring;
1203	unsigned head, tail, pos;
1204	long ret = 0;
1205	int copy_ret;
1206
1207	/*
1208	 * The mutex can block and wake us up and that will cause
1209	 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1210	 * and repeat. This should be rare enough that it doesn't cause
1211	 * peformance issues. See the comment in read_events() for more detail.
1212	 */
1213	sched_annotate_sleep();
1214	mutex_lock(&ctx->ring_lock);
1215
1216	/* Access to ->ring_pages here is protected by ctx->ring_lock. */
1217	ring = kmap_atomic(ctx->ring_pages[0]);
1218	head = ring->head;
1219	tail = ring->tail;
1220	kunmap_atomic(ring);
1221
1222	/*
1223	 * Ensure that once we've read the current tail pointer, that
1224	 * we also see the events that were stored up to the tail.
1225	 */
1226	smp_rmb();
1227
1228	pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1229
1230	if (head == tail)
1231		goto out;
1232
1233	head %= ctx->nr_events;
1234	tail %= ctx->nr_events;
1235
1236	while (ret < nr) {
1237		long avail;
1238		struct io_event *ev;
1239		struct page *page;
1240
1241		avail = (head <= tail ?  tail : ctx->nr_events) - head;
1242		if (head == tail)
1243			break;
1244
1245		pos = head + AIO_EVENTS_OFFSET;
1246		page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1247		pos %= AIO_EVENTS_PER_PAGE;
1248
1249		avail = min(avail, nr - ret);
1250		avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1251
1252		ev = kmap(page);
1253		copy_ret = copy_to_user(event + ret, ev + pos,
1254					sizeof(*ev) * avail);
1255		kunmap(page);
1256
1257		if (unlikely(copy_ret)) {
1258			ret = -EFAULT;
1259			goto out;
1260		}
1261
1262		ret += avail;
1263		head += avail;
1264		head %= ctx->nr_events;
1265	}
1266
1267	ring = kmap_atomic(ctx->ring_pages[0]);
1268	ring->head = head;
1269	kunmap_atomic(ring);
1270	flush_dcache_page(ctx->ring_pages[0]);
1271
1272	pr_debug("%li  h%u t%u\n", ret, head, tail);
1273out:
1274	mutex_unlock(&ctx->ring_lock);
1275
1276	return ret;
1277}
1278
1279static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1280			    struct io_event __user *event, long *i)
1281{
1282	long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1283
1284	if (ret > 0)
1285		*i += ret;
1286
1287	if (unlikely(atomic_read(&ctx->dead)))
1288		ret = -EINVAL;
1289
1290	if (!*i)
1291		*i = ret;
1292
1293	return ret < 0 || *i >= min_nr;
1294}
1295
1296static long read_events(struct kioctx *ctx, long min_nr, long nr,
1297			struct io_event __user *event,
1298			ktime_t until)
1299{
1300	long ret = 0;
1301
1302	/*
1303	 * Note that aio_read_events() is being called as the conditional - i.e.
1304	 * we're calling it after prepare_to_wait() has set task state to
1305	 * TASK_INTERRUPTIBLE.
1306	 *
1307	 * But aio_read_events() can block, and if it blocks it's going to flip
1308	 * the task state back to TASK_RUNNING.
1309	 *
1310	 * This should be ok, provided it doesn't flip the state back to
1311	 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1312	 * will only happen if the mutex_lock() call blocks, and we then find
1313	 * the ringbuffer empty. So in practice we should be ok, but it's
1314	 * something to be aware of when touching this code.
1315	 */
1316	if (until == 0)
1317		aio_read_events(ctx, min_nr, nr, event, &ret);
1318	else
1319		wait_event_interruptible_hrtimeout(ctx->wait,
1320				aio_read_events(ctx, min_nr, nr, event, &ret),
1321				until);
1322	return ret;
1323}
1324
1325/* sys_io_setup:
1326 *	Create an aio_context capable of receiving at least nr_events.
1327 *	ctxp must not point to an aio_context that already exists, and
1328 *	must be initialized to 0 prior to the call.  On successful
1329 *	creation of the aio_context, *ctxp is filled in with the resulting
1330 *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1331 *	if the specified nr_events exceeds internal limits.  May fail
1332 *	with -EAGAIN if the specified nr_events exceeds the user's limit
1333 *	of available events.  May fail with -ENOMEM if insufficient kernel
1334 *	resources are available.  May fail with -EFAULT if an invalid
1335 *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1336 *	implemented.
1337 */
1338SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1339{
1340	struct kioctx *ioctx = NULL;
1341	unsigned long ctx;
1342	long ret;
1343
1344	ret = get_user(ctx, ctxp);
1345	if (unlikely(ret))
1346		goto out;
1347
1348	ret = -EINVAL;
1349	if (unlikely(ctx || nr_events == 0)) {
1350		pr_debug("EINVAL: ctx %lu nr_events %u\n",
1351		         ctx, nr_events);
1352		goto out;
1353	}
1354
1355	ioctx = ioctx_alloc(nr_events);
1356	ret = PTR_ERR(ioctx);
1357	if (!IS_ERR(ioctx)) {
1358		ret = put_user(ioctx->user_id, ctxp);
1359		if (ret)
1360			kill_ioctx(current->mm, ioctx, NULL);
1361		percpu_ref_put(&ioctx->users);
1362	}
1363
1364out:
1365	return ret;
1366}
1367
1368#ifdef CONFIG_COMPAT
1369COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1370{
1371	struct kioctx *ioctx = NULL;
1372	unsigned long ctx;
1373	long ret;
1374
1375	ret = get_user(ctx, ctx32p);
1376	if (unlikely(ret))
1377		goto out;
1378
1379	ret = -EINVAL;
1380	if (unlikely(ctx || nr_events == 0)) {
1381		pr_debug("EINVAL: ctx %lu nr_events %u\n",
1382		         ctx, nr_events);
1383		goto out;
1384	}
1385
1386	ioctx = ioctx_alloc(nr_events);
1387	ret = PTR_ERR(ioctx);
1388	if (!IS_ERR(ioctx)) {
1389		/* truncating is ok because it's a user address */
1390		ret = put_user((u32)ioctx->user_id, ctx32p);
1391		if (ret)
1392			kill_ioctx(current->mm, ioctx, NULL);
1393		percpu_ref_put(&ioctx->users);
1394	}
1395
1396out:
1397	return ret;
1398}
1399#endif
1400
1401/* sys_io_destroy:
1402 *	Destroy the aio_context specified.  May cancel any outstanding
1403 *	AIOs and block on completion.  Will fail with -ENOSYS if not
1404 *	implemented.  May fail with -EINVAL if the context pointed to
1405 *	is invalid.
1406 */
1407SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1408{
1409	struct kioctx *ioctx = lookup_ioctx(ctx);
1410	if (likely(NULL != ioctx)) {
1411		struct ctx_rq_wait wait;
1412		int ret;
1413
1414		init_completion(&wait.comp);
1415		atomic_set(&wait.count, 1);
1416
1417		/* Pass requests_done to kill_ioctx() where it can be set
1418		 * in a thread-safe way. If we try to set it here then we have
1419		 * a race condition if two io_destroy() called simultaneously.
1420		 */
1421		ret = kill_ioctx(current->mm, ioctx, &wait);
1422		percpu_ref_put(&ioctx->users);
1423
1424		/* Wait until all IO for the context are done. Otherwise kernel
1425		 * keep using user-space buffers even if user thinks the context
1426		 * is destroyed.
1427		 */
1428		if (!ret)
1429			wait_for_completion(&wait.comp);
1430
1431		return ret;
1432	}
1433	pr_debug("EINVAL: invalid context id\n");
1434	return -EINVAL;
1435}
1436
1437static void aio_remove_iocb(struct aio_kiocb *iocb)
1438{
1439	struct kioctx *ctx = iocb->ki_ctx;
1440	unsigned long flags;
1441
1442	spin_lock_irqsave(&ctx->ctx_lock, flags);
1443	list_del(&iocb->ki_list);
1444	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1445}
1446
1447static void aio_complete_rw(struct kiocb *kiocb, long res)
1448{
1449	struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1450
1451	if (!list_empty_careful(&iocb->ki_list))
1452		aio_remove_iocb(iocb);
1453
1454	if (kiocb->ki_flags & IOCB_WRITE) {
1455		struct inode *inode = file_inode(kiocb->ki_filp);
1456
1457		/*
1458		 * Tell lockdep we inherited freeze protection from submission
1459		 * thread.
1460		 */
1461		if (S_ISREG(inode->i_mode))
1462			__sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1463		file_end_write(kiocb->ki_filp);
1464	}
1465
1466	iocb->ki_res.res = res;
1467	iocb->ki_res.res2 = 0;
1468	iocb_put(iocb);
1469}
1470
1471static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1472{
1473	int ret;
1474
1475	req->ki_complete = aio_complete_rw;
1476	req->private = NULL;
1477	req->ki_pos = iocb->aio_offset;
1478	req->ki_flags = iocb_flags(req->ki_filp);
1479	if (iocb->aio_flags & IOCB_FLAG_RESFD)
1480		req->ki_flags |= IOCB_EVENTFD;
1481	req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1482	if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1483		/*
1484		 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1485		 * aio_reqprio is interpreted as an I/O scheduling
1486		 * class and priority.
1487		 */
1488		ret = ioprio_check_cap(iocb->aio_reqprio);
1489		if (ret) {
1490			pr_debug("aio ioprio check cap error: %d\n", ret);
1491			return ret;
1492		}
1493
1494		req->ki_ioprio = iocb->aio_reqprio;
1495	} else
1496		req->ki_ioprio = get_current_ioprio();
1497
1498	ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1499	if (unlikely(ret))
1500		return ret;
1501
1502	req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1503	return 0;
1504}
1505
1506static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1507		struct iovec **iovec, bool vectored, bool compat,
1508		struct iov_iter *iter)
1509{
1510	void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1511	size_t len = iocb->aio_nbytes;
1512
1513	if (!vectored) {
1514		ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1515		*iovec = NULL;
1516		return ret;
1517	}
1518
1519	return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1520}
1521
1522static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1523{
1524	switch (ret) {
1525	case -EIOCBQUEUED:
1526		break;
1527	case -ERESTARTSYS:
1528	case -ERESTARTNOINTR:
1529	case -ERESTARTNOHAND:
1530	case -ERESTART_RESTARTBLOCK:
1531		/*
1532		 * There's no easy way to restart the syscall since other AIO's
1533		 * may be already running. Just fail this IO with EINTR.
1534		 */
1535		ret = -EINTR;
1536		fallthrough;
1537	default:
1538		req->ki_complete(req, ret);
1539	}
1540}
1541
1542static int aio_read(struct kiocb *req, const struct iocb *iocb,
1543			bool vectored, bool compat)
1544{
1545	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1546	struct iov_iter iter;
1547	struct file *file;
1548	int ret;
1549
1550	ret = aio_prep_rw(req, iocb);
1551	if (ret)
1552		return ret;
1553	file = req->ki_filp;
1554	if (unlikely(!(file->f_mode & FMODE_READ)))
1555		return -EBADF;
1556	ret = -EINVAL;
1557	if (unlikely(!file->f_op->read_iter))
1558		return -EINVAL;
1559
1560	ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1561	if (ret < 0)
1562		return ret;
1563	ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1564	if (!ret)
1565		aio_rw_done(req, call_read_iter(file, req, &iter));
1566	kfree(iovec);
1567	return ret;
1568}
1569
1570static int aio_write(struct kiocb *req, const struct iocb *iocb,
1571			 bool vectored, bool compat)
1572{
1573	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1574	struct iov_iter iter;
1575	struct file *file;
1576	int ret;
1577
1578	ret = aio_prep_rw(req, iocb);
1579	if (ret)
1580		return ret;
1581	file = req->ki_filp;
1582
1583	if (unlikely(!(file->f_mode & FMODE_WRITE)))
1584		return -EBADF;
1585	if (unlikely(!file->f_op->write_iter))
1586		return -EINVAL;
1587
1588	ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1589	if (ret < 0)
1590		return ret;
1591	ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1592	if (!ret) {
1593		/*
1594		 * Open-code file_start_write here to grab freeze protection,
1595		 * which will be released by another thread in
1596		 * aio_complete_rw().  Fool lockdep by telling it the lock got
1597		 * released so that it doesn't complain about the held lock when
1598		 * we return to userspace.
1599		 */
1600		if (S_ISREG(file_inode(file)->i_mode)) {
1601			sb_start_write(file_inode(file)->i_sb);
1602			__sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1603		}
1604		req->ki_flags |= IOCB_WRITE;
1605		aio_rw_done(req, call_write_iter(file, req, &iter));
1606	}
1607	kfree(iovec);
1608	return ret;
1609}
1610
1611static void aio_fsync_work(struct work_struct *work)
1612{
1613	struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1614	const struct cred *old_cred = override_creds(iocb->fsync.creds);
1615
1616	iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1617	revert_creds(old_cred);
1618	put_cred(iocb->fsync.creds);
1619	iocb_put(iocb);
1620}
1621
1622static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1623		     bool datasync)
1624{
1625	if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1626			iocb->aio_rw_flags))
1627		return -EINVAL;
1628
1629	if (unlikely(!req->file->f_op->fsync))
1630		return -EINVAL;
1631
1632	req->creds = prepare_creds();
1633	if (!req->creds)
1634		return -ENOMEM;
1635
1636	req->datasync = datasync;
1637	INIT_WORK(&req->work, aio_fsync_work);
1638	schedule_work(&req->work);
1639	return 0;
1640}
1641
1642static void aio_poll_put_work(struct work_struct *work)
1643{
1644	struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1645	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1646
1647	iocb_put(iocb);
1648}
1649
1650/*
1651 * Safely lock the waitqueue which the request is on, synchronizing with the
1652 * case where the ->poll() provider decides to free its waitqueue early.
1653 *
1654 * Returns true on success, meaning that req->head->lock was locked, req->wait
1655 * is on req->head, and an RCU read lock was taken.  Returns false if the
1656 * request was already removed from its waitqueue (which might no longer exist).
1657 */
1658static bool poll_iocb_lock_wq(struct poll_iocb *req)
1659{
1660	wait_queue_head_t *head;
1661
1662	/*
1663	 * While we hold the waitqueue lock and the waitqueue is nonempty,
1664	 * wake_up_pollfree() will wait for us.  However, taking the waitqueue
1665	 * lock in the first place can race with the waitqueue being freed.
1666	 *
1667	 * We solve this as eventpoll does: by taking advantage of the fact that
1668	 * all users of wake_up_pollfree() will RCU-delay the actual free.  If
1669	 * we enter rcu_read_lock() and see that the pointer to the queue is
1670	 * non-NULL, we can then lock it without the memory being freed out from
1671	 * under us, then check whether the request is still on the queue.
1672	 *
1673	 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1674	 * case the caller deletes the entry from the queue, leaving it empty.
1675	 * In that case, only RCU prevents the queue memory from being freed.
1676	 */
1677	rcu_read_lock();
1678	head = smp_load_acquire(&req->head);
1679	if (head) {
1680		spin_lock(&head->lock);
1681		if (!list_empty(&req->wait.entry))
1682			return true;
1683		spin_unlock(&head->lock);
1684	}
1685	rcu_read_unlock();
1686	return false;
1687}
1688
1689static void poll_iocb_unlock_wq(struct poll_iocb *req)
1690{
1691	spin_unlock(&req->head->lock);
1692	rcu_read_unlock();
1693}
1694
1695static void aio_poll_complete_work(struct work_struct *work)
1696{
1697	struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1698	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1699	struct poll_table_struct pt = { ._key = req->events };
1700	struct kioctx *ctx = iocb->ki_ctx;
1701	__poll_t mask = 0;
1702
1703	if (!READ_ONCE(req->cancelled))
1704		mask = vfs_poll(req->file, &pt) & req->events;
1705
1706	/*
1707	 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1708	 * calling ->ki_cancel.  We need the ctx_lock roundtrip here to
1709	 * synchronize with them.  In the cancellation case the list_del_init
1710	 * itself is not actually needed, but harmless so we keep it in to
1711	 * avoid further branches in the fast path.
1712	 */
1713	spin_lock_irq(&ctx->ctx_lock);
1714	if (poll_iocb_lock_wq(req)) {
1715		if (!mask && !READ_ONCE(req->cancelled)) {
1716			/*
1717			 * The request isn't actually ready to be completed yet.
1718			 * Reschedule completion if another wakeup came in.
1719			 */
1720			if (req->work_need_resched) {
1721				schedule_work(&req->work);
1722				req->work_need_resched = false;
1723			} else {
1724				req->work_scheduled = false;
1725			}
1726			poll_iocb_unlock_wq(req);
1727			spin_unlock_irq(&ctx->ctx_lock);
1728			return;
1729		}
1730		list_del_init(&req->wait.entry);
1731		poll_iocb_unlock_wq(req);
1732	} /* else, POLLFREE has freed the waitqueue, so we must complete */
1733	list_del_init(&iocb->ki_list);
1734	iocb->ki_res.res = mangle_poll(mask);
1735	spin_unlock_irq(&ctx->ctx_lock);
1736
1737	iocb_put(iocb);
1738}
1739
1740/* assumes we are called with irqs disabled */
1741static int aio_poll_cancel(struct kiocb *iocb)
1742{
1743	struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1744	struct poll_iocb *req = &aiocb->poll;
1745
1746	if (poll_iocb_lock_wq(req)) {
1747		WRITE_ONCE(req->cancelled, true);
1748		if (!req->work_scheduled) {
1749			schedule_work(&aiocb->poll.work);
1750			req->work_scheduled = true;
1751		}
1752		poll_iocb_unlock_wq(req);
1753	} /* else, the request was force-cancelled by POLLFREE already */
1754
1755	return 0;
1756}
1757
1758static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1759		void *key)
1760{
1761	struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1762	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1763	__poll_t mask = key_to_poll(key);
1764	unsigned long flags;
1765
1766	/* for instances that support it check for an event match first: */
1767	if (mask && !(mask & req->events))
1768		return 0;
1769
1770	/*
1771	 * Complete the request inline if possible.  This requires that three
1772	 * conditions be met:
1773	 *   1. An event mask must have been passed.  If a plain wakeup was done
1774	 *	instead, then mask == 0 and we have to call vfs_poll() to get
1775	 *	the events, so inline completion isn't possible.
1776	 *   2. The completion work must not have already been scheduled.
1777	 *   3. ctx_lock must not be busy.  We have to use trylock because we
1778	 *	already hold the waitqueue lock, so this inverts the normal
1779	 *	locking order.  Use irqsave/irqrestore because not all
1780	 *	filesystems (e.g. fuse) call this function with IRQs disabled,
1781	 *	yet IRQs have to be disabled before ctx_lock is obtained.
1782	 */
1783	if (mask && !req->work_scheduled &&
1784	    spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1785		struct kioctx *ctx = iocb->ki_ctx;
1786
1787		list_del_init(&req->wait.entry);
1788		list_del(&iocb->ki_list);
1789		iocb->ki_res.res = mangle_poll(mask);
1790		if (iocb->ki_eventfd && !eventfd_signal_allowed()) {
1791			iocb = NULL;
1792			INIT_WORK(&req->work, aio_poll_put_work);
1793			schedule_work(&req->work);
1794		}
1795		spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1796		if (iocb)
1797			iocb_put(iocb);
1798	} else {
1799		/*
1800		 * Schedule the completion work if needed.  If it was already
1801		 * scheduled, record that another wakeup came in.
1802		 *
1803		 * Don't remove the request from the waitqueue here, as it might
1804		 * not actually be complete yet (we won't know until vfs_poll()
1805		 * is called), and we must not miss any wakeups.  POLLFREE is an
1806		 * exception to this; see below.
1807		 */
1808		if (req->work_scheduled) {
1809			req->work_need_resched = true;
1810		} else {
1811			schedule_work(&req->work);
1812			req->work_scheduled = true;
1813		}
1814
1815		/*
1816		 * If the waitqueue is being freed early but we can't complete
1817		 * the request inline, we have to tear down the request as best
1818		 * we can.  That means immediately removing the request from its
1819		 * waitqueue and preventing all further accesses to the
1820		 * waitqueue via the request.  We also need to schedule the
1821		 * completion work (done above).  Also mark the request as
1822		 * cancelled, to potentially skip an unneeded call to ->poll().
1823		 */
1824		if (mask & POLLFREE) {
1825			WRITE_ONCE(req->cancelled, true);
1826			list_del_init(&req->wait.entry);
1827
1828			/*
1829			 * Careful: this *must* be the last step, since as soon
1830			 * as req->head is NULL'ed out, the request can be
1831			 * completed and freed, since aio_poll_complete_work()
1832			 * will no longer need to take the waitqueue lock.
1833			 */
1834			smp_store_release(&req->head, NULL);
1835		}
1836	}
1837	return 1;
1838}
1839
1840struct aio_poll_table {
1841	struct poll_table_struct	pt;
1842	struct aio_kiocb		*iocb;
1843	bool				queued;
1844	int				error;
1845};
1846
1847static void
1848aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1849		struct poll_table_struct *p)
1850{
1851	struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1852
1853	/* multiple wait queues per file are not supported */
1854	if (unlikely(pt->queued)) {
1855		pt->error = -EINVAL;
1856		return;
1857	}
1858
1859	pt->queued = true;
1860	pt->error = 0;
1861	pt->iocb->poll.head = head;
1862	add_wait_queue(head, &pt->iocb->poll.wait);
1863}
1864
1865static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1866{
1867	struct kioctx *ctx = aiocb->ki_ctx;
1868	struct poll_iocb *req = &aiocb->poll;
1869	struct aio_poll_table apt;
1870	bool cancel = false;
1871	__poll_t mask;
1872
1873	/* reject any unknown events outside the normal event mask. */
1874	if ((u16)iocb->aio_buf != iocb->aio_buf)
1875		return -EINVAL;
1876	/* reject fields that are not defined for poll */
1877	if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1878		return -EINVAL;
1879
1880	INIT_WORK(&req->work, aio_poll_complete_work);
1881	req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1882
1883	req->head = NULL;
1884	req->cancelled = false;
1885	req->work_scheduled = false;
1886	req->work_need_resched = false;
1887
1888	apt.pt._qproc = aio_poll_queue_proc;
1889	apt.pt._key = req->events;
1890	apt.iocb = aiocb;
1891	apt.queued = false;
1892	apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1893
1894	/* initialized the list so that we can do list_empty checks */
1895	INIT_LIST_HEAD(&req->wait.entry);
1896	init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1897
1898	mask = vfs_poll(req->file, &apt.pt) & req->events;
1899	spin_lock_irq(&ctx->ctx_lock);
1900	if (likely(apt.queued)) {
1901		bool on_queue = poll_iocb_lock_wq(req);
1902
1903		if (!on_queue || req->work_scheduled) {
1904			/*
1905			 * aio_poll_wake() already either scheduled the async
1906			 * completion work, or completed the request inline.
1907			 */
1908			if (apt.error) /* unsupported case: multiple queues */
1909				cancel = true;
1910			apt.error = 0;
1911			mask = 0;
1912		}
1913		if (mask || apt.error) {
1914			/* Steal to complete synchronously. */
1915			list_del_init(&req->wait.entry);
1916		} else if (cancel) {
1917			/* Cancel if possible (may be too late though). */
1918			WRITE_ONCE(req->cancelled, true);
1919		} else if (on_queue) {
1920			/*
1921			 * Actually waiting for an event, so add the request to
1922			 * active_reqs so that it can be cancelled if needed.
1923			 */
1924			list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1925			aiocb->ki_cancel = aio_poll_cancel;
1926		}
1927		if (on_queue)
1928			poll_iocb_unlock_wq(req);
1929	}
1930	if (mask) { /* no async, we'd stolen it */
1931		aiocb->ki_res.res = mangle_poll(mask);
1932		apt.error = 0;
1933	}
1934	spin_unlock_irq(&ctx->ctx_lock);
1935	if (mask)
1936		iocb_put(aiocb);
1937	return apt.error;
1938}
1939
1940static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1941			   struct iocb __user *user_iocb, struct aio_kiocb *req,
1942			   bool compat)
1943{
1944	req->ki_filp = fget(iocb->aio_fildes);
1945	if (unlikely(!req->ki_filp))
1946		return -EBADF;
1947
1948	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1949		struct eventfd_ctx *eventfd;
1950		/*
1951		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1952		 * instance of the file* now. The file descriptor must be
1953		 * an eventfd() fd, and will be signaled for each completed
1954		 * event using the eventfd_signal() function.
1955		 */
1956		eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1957		if (IS_ERR(eventfd))
1958			return PTR_ERR(eventfd);
1959
1960		req->ki_eventfd = eventfd;
1961	}
1962
1963	if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1964		pr_debug("EFAULT: aio_key\n");
1965		return -EFAULT;
1966	}
1967
1968	req->ki_res.obj = (u64)(unsigned long)user_iocb;
1969	req->ki_res.data = iocb->aio_data;
1970	req->ki_res.res = 0;
1971	req->ki_res.res2 = 0;
1972
1973	switch (iocb->aio_lio_opcode) {
1974	case IOCB_CMD_PREAD:
1975		return aio_read(&req->rw, iocb, false, compat);
1976	case IOCB_CMD_PWRITE:
1977		return aio_write(&req->rw, iocb, false, compat);
1978	case IOCB_CMD_PREADV:
1979		return aio_read(&req->rw, iocb, true, compat);
1980	case IOCB_CMD_PWRITEV:
1981		return aio_write(&req->rw, iocb, true, compat);
1982	case IOCB_CMD_FSYNC:
1983		return aio_fsync(&req->fsync, iocb, false);
1984	case IOCB_CMD_FDSYNC:
1985		return aio_fsync(&req->fsync, iocb, true);
1986	case IOCB_CMD_POLL:
1987		return aio_poll(req, iocb);
1988	default:
1989		pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1990		return -EINVAL;
1991	}
1992}
1993
1994static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1995			 bool compat)
1996{
1997	struct aio_kiocb *req;
1998	struct iocb iocb;
1999	int err;
2000
2001	if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
2002		return -EFAULT;
2003
2004	/* enforce forwards compatibility on users */
2005	if (unlikely(iocb.aio_reserved2)) {
2006		pr_debug("EINVAL: reserve field set\n");
2007		return -EINVAL;
2008	}
2009
2010	/* prevent overflows */
2011	if (unlikely(
2012	    (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
2013	    (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
2014	    ((ssize_t)iocb.aio_nbytes < 0)
2015	   )) {
2016		pr_debug("EINVAL: overflow check\n");
2017		return -EINVAL;
2018	}
2019
2020	req = aio_get_req(ctx);
2021	if (unlikely(!req))
2022		return -EAGAIN;
2023
2024	err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2025
2026	/* Done with the synchronous reference */
2027	iocb_put(req);
2028
2029	/*
2030	 * If err is 0, we'd either done aio_complete() ourselves or have
2031	 * arranged for that to be done asynchronously.  Anything non-zero
2032	 * means that we need to destroy req ourselves.
2033	 */
2034	if (unlikely(err)) {
2035		iocb_destroy(req);
2036		put_reqs_available(ctx, 1);
2037	}
2038	return err;
2039}
2040
2041/* sys_io_submit:
2042 *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
2043 *	the number of iocbs queued.  May return -EINVAL if the aio_context
2044 *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
2045 *	*iocbpp[0] is not properly initialized, if the operation specified
2046 *	is invalid for the file descriptor in the iocb.  May fail with
2047 *	-EFAULT if any of the data structures point to invalid data.  May
2048 *	fail with -EBADF if the file descriptor specified in the first
2049 *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
2050 *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
2051 *	fail with -ENOSYS if not implemented.
2052 */
2053SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2054		struct iocb __user * __user *, iocbpp)
2055{
2056	struct kioctx *ctx;
2057	long ret = 0;
2058	int i = 0;
2059	struct blk_plug plug;
2060
2061	if (unlikely(nr < 0))
2062		return -EINVAL;
2063
2064	ctx = lookup_ioctx(ctx_id);
2065	if (unlikely(!ctx)) {
2066		pr_debug("EINVAL: invalid context id\n");
2067		return -EINVAL;
2068	}
2069
2070	if (nr > ctx->nr_events)
2071		nr = ctx->nr_events;
2072
2073	if (nr > AIO_PLUG_THRESHOLD)
2074		blk_start_plug(&plug);
2075	for (i = 0; i < nr; i++) {
2076		struct iocb __user *user_iocb;
2077
2078		if (unlikely(get_user(user_iocb, iocbpp + i))) {
2079			ret = -EFAULT;
2080			break;
2081		}
2082
2083		ret = io_submit_one(ctx, user_iocb, false);
2084		if (ret)
2085			break;
2086	}
2087	if (nr > AIO_PLUG_THRESHOLD)
2088		blk_finish_plug(&plug);
2089
2090	percpu_ref_put(&ctx->users);
2091	return i ? i : ret;
2092}
2093
2094#ifdef CONFIG_COMPAT
2095COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2096		       int, nr, compat_uptr_t __user *, iocbpp)
2097{
2098	struct kioctx *ctx;
2099	long ret = 0;
2100	int i = 0;
2101	struct blk_plug plug;
2102
2103	if (unlikely(nr < 0))
2104		return -EINVAL;
2105
2106	ctx = lookup_ioctx(ctx_id);
2107	if (unlikely(!ctx)) {
2108		pr_debug("EINVAL: invalid context id\n");
2109		return -EINVAL;
2110	}
2111
2112	if (nr > ctx->nr_events)
2113		nr = ctx->nr_events;
2114
2115	if (nr > AIO_PLUG_THRESHOLD)
2116		blk_start_plug(&plug);
2117	for (i = 0; i < nr; i++) {
2118		compat_uptr_t user_iocb;
2119
2120		if (unlikely(get_user(user_iocb, iocbpp + i))) {
2121			ret = -EFAULT;
2122			break;
2123		}
2124
2125		ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2126		if (ret)
2127			break;
2128	}
2129	if (nr > AIO_PLUG_THRESHOLD)
2130		blk_finish_plug(&plug);
2131
2132	percpu_ref_put(&ctx->users);
2133	return i ? i : ret;
2134}
2135#endif
2136
2137/* sys_io_cancel:
2138 *	Attempts to cancel an iocb previously passed to io_submit.  If
2139 *	the operation is successfully cancelled, the resulting event is
2140 *	copied into the memory pointed to by result without being placed
2141 *	into the completion queue and 0 is returned.  May fail with
2142 *	-EFAULT if any of the data structures pointed to are invalid.
2143 *	May fail with -EINVAL if aio_context specified by ctx_id is
2144 *	invalid.  May fail with -EAGAIN if the iocb specified was not
2145 *	cancelled.  Will fail with -ENOSYS if not implemented.
2146 */
2147SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2148		struct io_event __user *, result)
2149{
2150	struct kioctx *ctx;
2151	struct aio_kiocb *kiocb;
2152	int ret = -EINVAL;
2153	u32 key;
2154	u64 obj = (u64)(unsigned long)iocb;
2155
2156	if (unlikely(get_user(key, &iocb->aio_key)))
2157		return -EFAULT;
2158	if (unlikely(key != KIOCB_KEY))
2159		return -EINVAL;
2160
2161	ctx = lookup_ioctx(ctx_id);
2162	if (unlikely(!ctx))
2163		return -EINVAL;
2164
2165	spin_lock_irq(&ctx->ctx_lock);
2166	/* TODO: use a hash or array, this sucks. */
2167	list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2168		if (kiocb->ki_res.obj == obj) {
2169			ret = kiocb->ki_cancel(&kiocb->rw);
2170			list_del_init(&kiocb->ki_list);
2171			break;
2172		}
2173	}
2174	spin_unlock_irq(&ctx->ctx_lock);
2175
2176	if (!ret) {
2177		/*
2178		 * The result argument is no longer used - the io_event is
2179		 * always delivered via the ring buffer. -EINPROGRESS indicates
2180		 * cancellation is progress:
2181		 */
2182		ret = -EINPROGRESS;
2183	}
2184
2185	percpu_ref_put(&ctx->users);
2186
2187	return ret;
2188}
2189
2190static long do_io_getevents(aio_context_t ctx_id,
2191		long min_nr,
2192		long nr,
2193		struct io_event __user *events,
2194		struct timespec64 *ts)
2195{
2196	ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2197	struct kioctx *ioctx = lookup_ioctx(ctx_id);
2198	long ret = -EINVAL;
2199
2200	if (likely(ioctx)) {
2201		if (likely(min_nr <= nr && min_nr >= 0))
2202			ret = read_events(ioctx, min_nr, nr, events, until);
2203		percpu_ref_put(&ioctx->users);
2204	}
2205
2206	return ret;
2207}
2208
2209/* io_getevents:
2210 *	Attempts to read at least min_nr events and up to nr events from
2211 *	the completion queue for the aio_context specified by ctx_id. If
2212 *	it succeeds, the number of read events is returned. May fail with
2213 *	-EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2214 *	out of range, if timeout is out of range.  May fail with -EFAULT
2215 *	if any of the memory specified is invalid.  May return 0 or
2216 *	< min_nr if the timeout specified by timeout has elapsed
2217 *	before sufficient events are available, where timeout == NULL
2218 *	specifies an infinite timeout. Note that the timeout pointed to by
2219 *	timeout is relative.  Will fail with -ENOSYS if not implemented.
2220 */
2221#ifdef CONFIG_64BIT
2222
2223SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2224		long, min_nr,
2225		long, nr,
2226		struct io_event __user *, events,
2227		struct __kernel_timespec __user *, timeout)
2228{
2229	struct timespec64	ts;
2230	int			ret;
2231
2232	if (timeout && unlikely(get_timespec64(&ts, timeout)))
2233		return -EFAULT;
2234
2235	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2236	if (!ret && signal_pending(current))
2237		ret = -EINTR;
2238	return ret;
2239}
2240
2241#endif
2242
2243struct __aio_sigset {
2244	const sigset_t __user	*sigmask;
2245	size_t		sigsetsize;
2246};
2247
2248SYSCALL_DEFINE6(io_pgetevents,
2249		aio_context_t, ctx_id,
2250		long, min_nr,
2251		long, nr,
2252		struct io_event __user *, events,
2253		struct __kernel_timespec __user *, timeout,
2254		const struct __aio_sigset __user *, usig)
2255{
2256	struct __aio_sigset	ksig = { NULL, };
2257	struct timespec64	ts;
2258	bool interrupted;
2259	int ret;
2260
2261	if (timeout && unlikely(get_timespec64(&ts, timeout)))
2262		return -EFAULT;
2263
2264	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2265		return -EFAULT;
2266
2267	ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2268	if (ret)
2269		return ret;
2270
2271	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2272
2273	interrupted = signal_pending(current);
2274	restore_saved_sigmask_unless(interrupted);
2275	if (interrupted && !ret)
2276		ret = -ERESTARTNOHAND;
2277
2278	return ret;
2279}
2280
2281#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2282
2283SYSCALL_DEFINE6(io_pgetevents_time32,
2284		aio_context_t, ctx_id,
2285		long, min_nr,
2286		long, nr,
2287		struct io_event __user *, events,
2288		struct old_timespec32 __user *, timeout,
2289		const struct __aio_sigset __user *, usig)
2290{
2291	struct __aio_sigset	ksig = { NULL, };
2292	struct timespec64	ts;
2293	bool interrupted;
2294	int ret;
2295
2296	if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2297		return -EFAULT;
2298
2299	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2300		return -EFAULT;
2301
2302
2303	ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2304	if (ret)
2305		return ret;
2306
2307	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2308
2309	interrupted = signal_pending(current);
2310	restore_saved_sigmask_unless(interrupted);
2311	if (interrupted && !ret)
2312		ret = -ERESTARTNOHAND;
2313
2314	return ret;
2315}
2316
2317#endif
2318
2319#if defined(CONFIG_COMPAT_32BIT_TIME)
2320
2321SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2322		__s32, min_nr,
2323		__s32, nr,
2324		struct io_event __user *, events,
2325		struct old_timespec32 __user *, timeout)
2326{
2327	struct timespec64 t;
2328	int ret;
2329
2330	if (timeout && get_old_timespec32(&t, timeout))
2331		return -EFAULT;
2332
2333	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2334	if (!ret && signal_pending(current))
2335		ret = -EINTR;
2336	return ret;
2337}
2338
2339#endif
2340
2341#ifdef CONFIG_COMPAT
2342
2343struct __compat_aio_sigset {
2344	compat_uptr_t		sigmask;
2345	compat_size_t		sigsetsize;
2346};
2347
2348#if defined(CONFIG_COMPAT_32BIT_TIME)
2349
2350COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2351		compat_aio_context_t, ctx_id,
2352		compat_long_t, min_nr,
2353		compat_long_t, nr,
2354		struct io_event __user *, events,
2355		struct old_timespec32 __user *, timeout,
2356		const struct __compat_aio_sigset __user *, usig)
2357{
2358	struct __compat_aio_sigset ksig = { 0, };
2359	struct timespec64 t;
2360	bool interrupted;
2361	int ret;
2362
2363	if (timeout && get_old_timespec32(&t, timeout))
2364		return -EFAULT;
2365
2366	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2367		return -EFAULT;
2368
2369	ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2370	if (ret)
2371		return ret;
2372
2373	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2374
2375	interrupted = signal_pending(current);
2376	restore_saved_sigmask_unless(interrupted);
2377	if (interrupted && !ret)
2378		ret = -ERESTARTNOHAND;
2379
2380	return ret;
2381}
2382
2383#endif
2384
2385COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2386		compat_aio_context_t, ctx_id,
2387		compat_long_t, min_nr,
2388		compat_long_t, nr,
2389		struct io_event __user *, events,
2390		struct __kernel_timespec __user *, timeout,
2391		const struct __compat_aio_sigset __user *, usig)
2392{
2393	struct __compat_aio_sigset ksig = { 0, };
2394	struct timespec64 t;
2395	bool interrupted;
2396	int ret;
2397
2398	if (timeout && get_timespec64(&t, timeout))
2399		return -EFAULT;
2400
2401	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2402		return -EFAULT;
2403
2404	ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2405	if (ret)
2406		return ret;
2407
2408	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2409
2410	interrupted = signal_pending(current);
2411	restore_saved_sigmask_unless(interrupted);
2412	if (interrupted && !ret)
2413		ret = -ERESTARTNOHAND;
2414
2415	return ret;
2416}
2417#endif
2418