aio.c revision b542e383
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			done;
185	bool			cancelled;
186	struct wait_queue_entry	wait;
187	struct work_struct	work;
188};
189
190/*
191 * NOTE! Each of the iocb union members has the file pointer
192 * as the first entry in their struct definition. So you can
193 * access the file pointer through any of the sub-structs,
194 * or directly as just 'ki_filp' in this struct.
195 */
196struct aio_kiocb {
197	union {
198		struct file		*ki_filp;
199		struct kiocb		rw;
200		struct fsync_iocb	fsync;
201		struct poll_iocb	poll;
202	};
203
204	struct kioctx		*ki_ctx;
205	kiocb_cancel_fn		*ki_cancel;
206
207	struct io_event		ki_res;
208
209	struct list_head	ki_list;	/* the aio core uses this
210						 * for cancellation */
211	refcount_t		ki_refcnt;
212
213	/*
214	 * If the aio_resfd field of the userspace iocb is not zero,
215	 * this is the underlying eventfd context to deliver events to.
216	 */
217	struct eventfd_ctx	*ki_eventfd;
218};
219
220/*------ sysctl variables----*/
221static DEFINE_SPINLOCK(aio_nr_lock);
222unsigned long aio_nr;		/* current system wide number of aio requests */
223unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
224/*----end sysctl variables---*/
225
226static struct kmem_cache	*kiocb_cachep;
227static struct kmem_cache	*kioctx_cachep;
228
229static struct vfsmount *aio_mnt;
230
231static const struct file_operations aio_ring_fops;
232static const struct address_space_operations aio_ctx_aops;
233
234static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
235{
236	struct file *file;
237	struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
238	if (IS_ERR(inode))
239		return ERR_CAST(inode);
240
241	inode->i_mapping->a_ops = &aio_ctx_aops;
242	inode->i_mapping->private_data = ctx;
243	inode->i_size = PAGE_SIZE * nr_pages;
244
245	file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
246				O_RDWR, &aio_ring_fops);
247	if (IS_ERR(file))
248		iput(inode);
249	return file;
250}
251
252static int aio_init_fs_context(struct fs_context *fc)
253{
254	if (!init_pseudo(fc, AIO_RING_MAGIC))
255		return -ENOMEM;
256	fc->s_iflags |= SB_I_NOEXEC;
257	return 0;
258}
259
260/* aio_setup
261 *	Creates the slab caches used by the aio routines, panic on
262 *	failure as this is done early during the boot sequence.
263 */
264static int __init aio_setup(void)
265{
266	static struct file_system_type aio_fs = {
267		.name		= "aio",
268		.init_fs_context = aio_init_fs_context,
269		.kill_sb	= kill_anon_super,
270	};
271	aio_mnt = kern_mount(&aio_fs);
272	if (IS_ERR(aio_mnt))
273		panic("Failed to create aio fs mount.");
274
275	kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
276	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
277	return 0;
278}
279__initcall(aio_setup);
280
281static void put_aio_ring_file(struct kioctx *ctx)
282{
283	struct file *aio_ring_file = ctx->aio_ring_file;
284	struct address_space *i_mapping;
285
286	if (aio_ring_file) {
287		truncate_setsize(file_inode(aio_ring_file), 0);
288
289		/* Prevent further access to the kioctx from migratepages */
290		i_mapping = aio_ring_file->f_mapping;
291		spin_lock(&i_mapping->private_lock);
292		i_mapping->private_data = NULL;
293		ctx->aio_ring_file = NULL;
294		spin_unlock(&i_mapping->private_lock);
295
296		fput(aio_ring_file);
297	}
298}
299
300static void aio_free_ring(struct kioctx *ctx)
301{
302	int i;
303
304	/* Disconnect the kiotx from the ring file.  This prevents future
305	 * accesses to the kioctx from page migration.
306	 */
307	put_aio_ring_file(ctx);
308
309	for (i = 0; i < ctx->nr_pages; i++) {
310		struct page *page;
311		pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
312				page_count(ctx->ring_pages[i]));
313		page = ctx->ring_pages[i];
314		if (!page)
315			continue;
316		ctx->ring_pages[i] = NULL;
317		put_page(page);
318	}
319
320	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
321		kfree(ctx->ring_pages);
322		ctx->ring_pages = NULL;
323	}
324}
325
326static int aio_ring_mremap(struct vm_area_struct *vma)
327{
328	struct file *file = vma->vm_file;
329	struct mm_struct *mm = vma->vm_mm;
330	struct kioctx_table *table;
331	int i, res = -EINVAL;
332
333	spin_lock(&mm->ioctx_lock);
334	rcu_read_lock();
335	table = rcu_dereference(mm->ioctx_table);
336	for (i = 0; i < table->nr; i++) {
337		struct kioctx *ctx;
338
339		ctx = rcu_dereference(table->table[i]);
340		if (ctx && ctx->aio_ring_file == file) {
341			if (!atomic_read(&ctx->dead)) {
342				ctx->user_id = ctx->mmap_base = vma->vm_start;
343				res = 0;
344			}
345			break;
346		}
347	}
348
349	rcu_read_unlock();
350	spin_unlock(&mm->ioctx_lock);
351	return res;
352}
353
354static const struct vm_operations_struct aio_ring_vm_ops = {
355	.mremap		= aio_ring_mremap,
356#if IS_ENABLED(CONFIG_MMU)
357	.fault		= filemap_fault,
358	.map_pages	= filemap_map_pages,
359	.page_mkwrite	= filemap_page_mkwrite,
360#endif
361};
362
363static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
364{
365	vma->vm_flags |= VM_DONTEXPAND;
366	vma->vm_ops = &aio_ring_vm_ops;
367	return 0;
368}
369
370static const struct file_operations aio_ring_fops = {
371	.mmap = aio_ring_mmap,
372};
373
374#if IS_ENABLED(CONFIG_MIGRATION)
375static int aio_migratepage(struct address_space *mapping, struct page *new,
376			struct page *old, enum migrate_mode mode)
377{
378	struct kioctx *ctx;
379	unsigned long flags;
380	pgoff_t idx;
381	int rc;
382
383	/*
384	 * We cannot support the _NO_COPY case here, because copy needs to
385	 * happen under the ctx->completion_lock. That does not work with the
386	 * migration workflow of MIGRATE_SYNC_NO_COPY.
387	 */
388	if (mode == MIGRATE_SYNC_NO_COPY)
389		return -EINVAL;
390
391	rc = 0;
392
393	/* mapping->private_lock here protects against the kioctx teardown.  */
394	spin_lock(&mapping->private_lock);
395	ctx = mapping->private_data;
396	if (!ctx) {
397		rc = -EINVAL;
398		goto out;
399	}
400
401	/* The ring_lock mutex.  The prevents aio_read_events() from writing
402	 * to the ring's head, and prevents page migration from mucking in
403	 * a partially initialized kiotx.
404	 */
405	if (!mutex_trylock(&ctx->ring_lock)) {
406		rc = -EAGAIN;
407		goto out;
408	}
409
410	idx = old->index;
411	if (idx < (pgoff_t)ctx->nr_pages) {
412		/* Make sure the old page hasn't already been changed */
413		if (ctx->ring_pages[idx] != old)
414			rc = -EAGAIN;
415	} else
416		rc = -EINVAL;
417
418	if (rc != 0)
419		goto out_unlock;
420
421	/* Writeback must be complete */
422	BUG_ON(PageWriteback(old));
423	get_page(new);
424
425	rc = migrate_page_move_mapping(mapping, new, old, 1);
426	if (rc != MIGRATEPAGE_SUCCESS) {
427		put_page(new);
428		goto out_unlock;
429	}
430
431	/* Take completion_lock to prevent other writes to the ring buffer
432	 * while the old page is copied to the new.  This prevents new
433	 * events from being lost.
434	 */
435	spin_lock_irqsave(&ctx->completion_lock, flags);
436	migrate_page_copy(new, old);
437	BUG_ON(ctx->ring_pages[idx] != old);
438	ctx->ring_pages[idx] = new;
439	spin_unlock_irqrestore(&ctx->completion_lock, flags);
440
441	/* The old page is no longer accessible. */
442	put_page(old);
443
444out_unlock:
445	mutex_unlock(&ctx->ring_lock);
446out:
447	spin_unlock(&mapping->private_lock);
448	return rc;
449}
450#endif
451
452static const struct address_space_operations aio_ctx_aops = {
453	.set_page_dirty = __set_page_dirty_no_writeback,
454#if IS_ENABLED(CONFIG_MIGRATION)
455	.migratepage	= aio_migratepage,
456#endif
457};
458
459static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
460{
461	struct aio_ring *ring;
462	struct mm_struct *mm = current->mm;
463	unsigned long size, unused;
464	int nr_pages;
465	int i;
466	struct file *file;
467
468	/* Compensate for the ring buffer's head/tail overlap entry */
469	nr_events += 2;	/* 1 is required, 2 for good luck */
470
471	size = sizeof(struct aio_ring);
472	size += sizeof(struct io_event) * nr_events;
473
474	nr_pages = PFN_UP(size);
475	if (nr_pages < 0)
476		return -EINVAL;
477
478	file = aio_private_file(ctx, nr_pages);
479	if (IS_ERR(file)) {
480		ctx->aio_ring_file = NULL;
481		return -ENOMEM;
482	}
483
484	ctx->aio_ring_file = file;
485	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
486			/ sizeof(struct io_event);
487
488	ctx->ring_pages = ctx->internal_pages;
489	if (nr_pages > AIO_RING_PAGES) {
490		ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
491					  GFP_KERNEL);
492		if (!ctx->ring_pages) {
493			put_aio_ring_file(ctx);
494			return -ENOMEM;
495		}
496	}
497
498	for (i = 0; i < nr_pages; i++) {
499		struct page *page;
500		page = find_or_create_page(file->f_mapping,
501					   i, GFP_HIGHUSER | __GFP_ZERO);
502		if (!page)
503			break;
504		pr_debug("pid(%d) page[%d]->count=%d\n",
505			 current->pid, i, page_count(page));
506		SetPageUptodate(page);
507		unlock_page(page);
508
509		ctx->ring_pages[i] = page;
510	}
511	ctx->nr_pages = i;
512
513	if (unlikely(i != nr_pages)) {
514		aio_free_ring(ctx);
515		return -ENOMEM;
516	}
517
518	ctx->mmap_size = nr_pages * PAGE_SIZE;
519	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
520
521	if (mmap_write_lock_killable(mm)) {
522		ctx->mmap_size = 0;
523		aio_free_ring(ctx);
524		return -EINTR;
525	}
526
527	ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
528				 PROT_READ | PROT_WRITE,
529				 MAP_SHARED, 0, &unused, NULL);
530	mmap_write_unlock(mm);
531	if (IS_ERR((void *)ctx->mmap_base)) {
532		ctx->mmap_size = 0;
533		aio_free_ring(ctx);
534		return -ENOMEM;
535	}
536
537	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
538
539	ctx->user_id = ctx->mmap_base;
540	ctx->nr_events = nr_events; /* trusted copy */
541
542	ring = kmap_atomic(ctx->ring_pages[0]);
543	ring->nr = nr_events;	/* user copy */
544	ring->id = ~0U;
545	ring->head = ring->tail = 0;
546	ring->magic = AIO_RING_MAGIC;
547	ring->compat_features = AIO_RING_COMPAT_FEATURES;
548	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
549	ring->header_length = sizeof(struct aio_ring);
550	kunmap_atomic(ring);
551	flush_dcache_page(ctx->ring_pages[0]);
552
553	return 0;
554}
555
556#define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
557#define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
558#define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
559
560void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
561{
562	struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
563	struct kioctx *ctx = req->ki_ctx;
564	unsigned long flags;
565
566	if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
567		return;
568
569	spin_lock_irqsave(&ctx->ctx_lock, flags);
570	list_add_tail(&req->ki_list, &ctx->active_reqs);
571	req->ki_cancel = cancel;
572	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
573}
574EXPORT_SYMBOL(kiocb_set_cancel_fn);
575
576/*
577 * free_ioctx() should be RCU delayed to synchronize against the RCU
578 * protected lookup_ioctx() and also needs process context to call
579 * aio_free_ring().  Use rcu_work.
580 */
581static void free_ioctx(struct work_struct *work)
582{
583	struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
584					  free_rwork);
585	pr_debug("freeing %p\n", ctx);
586
587	aio_free_ring(ctx);
588	free_percpu(ctx->cpu);
589	percpu_ref_exit(&ctx->reqs);
590	percpu_ref_exit(&ctx->users);
591	kmem_cache_free(kioctx_cachep, ctx);
592}
593
594static void free_ioctx_reqs(struct percpu_ref *ref)
595{
596	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
597
598	/* At this point we know that there are no any in-flight requests */
599	if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
600		complete(&ctx->rq_wait->comp);
601
602	/* Synchronize against RCU protected table->table[] dereferences */
603	INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
604	queue_rcu_work(system_wq, &ctx->free_rwork);
605}
606
607/*
608 * When this function runs, the kioctx has been removed from the "hash table"
609 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
610 * now it's safe to cancel any that need to be.
611 */
612static void free_ioctx_users(struct percpu_ref *ref)
613{
614	struct kioctx *ctx = container_of(ref, struct kioctx, users);
615	struct aio_kiocb *req;
616
617	spin_lock_irq(&ctx->ctx_lock);
618
619	while (!list_empty(&ctx->active_reqs)) {
620		req = list_first_entry(&ctx->active_reqs,
621				       struct aio_kiocb, ki_list);
622		req->ki_cancel(&req->rw);
623		list_del_init(&req->ki_list);
624	}
625
626	spin_unlock_irq(&ctx->ctx_lock);
627
628	percpu_ref_kill(&ctx->reqs);
629	percpu_ref_put(&ctx->reqs);
630}
631
632static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
633{
634	unsigned i, new_nr;
635	struct kioctx_table *table, *old;
636	struct aio_ring *ring;
637
638	spin_lock(&mm->ioctx_lock);
639	table = rcu_dereference_raw(mm->ioctx_table);
640
641	while (1) {
642		if (table)
643			for (i = 0; i < table->nr; i++)
644				if (!rcu_access_pointer(table->table[i])) {
645					ctx->id = i;
646					rcu_assign_pointer(table->table[i], ctx);
647					spin_unlock(&mm->ioctx_lock);
648
649					/* While kioctx setup is in progress,
650					 * we are protected from page migration
651					 * changes ring_pages by ->ring_lock.
652					 */
653					ring = kmap_atomic(ctx->ring_pages[0]);
654					ring->id = ctx->id;
655					kunmap_atomic(ring);
656					return 0;
657				}
658
659		new_nr = (table ? table->nr : 1) * 4;
660		spin_unlock(&mm->ioctx_lock);
661
662		table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
663				new_nr, GFP_KERNEL);
664		if (!table)
665			return -ENOMEM;
666
667		table->nr = new_nr;
668
669		spin_lock(&mm->ioctx_lock);
670		old = rcu_dereference_raw(mm->ioctx_table);
671
672		if (!old) {
673			rcu_assign_pointer(mm->ioctx_table, table);
674		} else if (table->nr > old->nr) {
675			memcpy(table->table, old->table,
676			       old->nr * sizeof(struct kioctx *));
677
678			rcu_assign_pointer(mm->ioctx_table, table);
679			kfree_rcu(old, rcu);
680		} else {
681			kfree(table);
682			table = old;
683		}
684	}
685}
686
687static void aio_nr_sub(unsigned nr)
688{
689	spin_lock(&aio_nr_lock);
690	if (WARN_ON(aio_nr - nr > aio_nr))
691		aio_nr = 0;
692	else
693		aio_nr -= nr;
694	spin_unlock(&aio_nr_lock);
695}
696
697/* ioctx_alloc
698 *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
699 */
700static struct kioctx *ioctx_alloc(unsigned nr_events)
701{
702	struct mm_struct *mm = current->mm;
703	struct kioctx *ctx;
704	int err = -ENOMEM;
705
706	/*
707	 * Store the original nr_events -- what userspace passed to io_setup(),
708	 * for counting against the global limit -- before it changes.
709	 */
710	unsigned int max_reqs = nr_events;
711
712	/*
713	 * We keep track of the number of available ringbuffer slots, to prevent
714	 * overflow (reqs_available), and we also use percpu counters for this.
715	 *
716	 * So since up to half the slots might be on other cpu's percpu counters
717	 * and unavailable, double nr_events so userspace sees what they
718	 * expected: additionally, we move req_batch slots to/from percpu
719	 * counters at a time, so make sure that isn't 0:
720	 */
721	nr_events = max(nr_events, num_possible_cpus() * 4);
722	nr_events *= 2;
723
724	/* Prevent overflows */
725	if (nr_events > (0x10000000U / sizeof(struct io_event))) {
726		pr_debug("ENOMEM: nr_events too high\n");
727		return ERR_PTR(-EINVAL);
728	}
729
730	if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
731		return ERR_PTR(-EAGAIN);
732
733	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
734	if (!ctx)
735		return ERR_PTR(-ENOMEM);
736
737	ctx->max_reqs = max_reqs;
738
739	spin_lock_init(&ctx->ctx_lock);
740	spin_lock_init(&ctx->completion_lock);
741	mutex_init(&ctx->ring_lock);
742	/* Protect against page migration throughout kiotx setup by keeping
743	 * the ring_lock mutex held until setup is complete. */
744	mutex_lock(&ctx->ring_lock);
745	init_waitqueue_head(&ctx->wait);
746
747	INIT_LIST_HEAD(&ctx->active_reqs);
748
749	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
750		goto err;
751
752	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
753		goto err;
754
755	ctx->cpu = alloc_percpu(struct kioctx_cpu);
756	if (!ctx->cpu)
757		goto err;
758
759	err = aio_setup_ring(ctx, nr_events);
760	if (err < 0)
761		goto err;
762
763	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
764	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
765	if (ctx->req_batch < 1)
766		ctx->req_batch = 1;
767
768	/* limit the number of system wide aios */
769	spin_lock(&aio_nr_lock);
770	if (aio_nr + ctx->max_reqs > aio_max_nr ||
771	    aio_nr + ctx->max_reqs < aio_nr) {
772		spin_unlock(&aio_nr_lock);
773		err = -EAGAIN;
774		goto err_ctx;
775	}
776	aio_nr += ctx->max_reqs;
777	spin_unlock(&aio_nr_lock);
778
779	percpu_ref_get(&ctx->users);	/* io_setup() will drop this ref */
780	percpu_ref_get(&ctx->reqs);	/* free_ioctx_users() will drop this */
781
782	err = ioctx_add_table(ctx, mm);
783	if (err)
784		goto err_cleanup;
785
786	/* Release the ring_lock mutex now that all setup is complete. */
787	mutex_unlock(&ctx->ring_lock);
788
789	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
790		 ctx, ctx->user_id, mm, ctx->nr_events);
791	return ctx;
792
793err_cleanup:
794	aio_nr_sub(ctx->max_reqs);
795err_ctx:
796	atomic_set(&ctx->dead, 1);
797	if (ctx->mmap_size)
798		vm_munmap(ctx->mmap_base, ctx->mmap_size);
799	aio_free_ring(ctx);
800err:
801	mutex_unlock(&ctx->ring_lock);
802	free_percpu(ctx->cpu);
803	percpu_ref_exit(&ctx->reqs);
804	percpu_ref_exit(&ctx->users);
805	kmem_cache_free(kioctx_cachep, ctx);
806	pr_debug("error allocating ioctx %d\n", err);
807	return ERR_PTR(err);
808}
809
810/* kill_ioctx
811 *	Cancels all outstanding aio requests on an aio context.  Used
812 *	when the processes owning a context have all exited to encourage
813 *	the rapid destruction of the kioctx.
814 */
815static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
816		      struct ctx_rq_wait *wait)
817{
818	struct kioctx_table *table;
819
820	spin_lock(&mm->ioctx_lock);
821	if (atomic_xchg(&ctx->dead, 1)) {
822		spin_unlock(&mm->ioctx_lock);
823		return -EINVAL;
824	}
825
826	table = rcu_dereference_raw(mm->ioctx_table);
827	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
828	RCU_INIT_POINTER(table->table[ctx->id], NULL);
829	spin_unlock(&mm->ioctx_lock);
830
831	/* free_ioctx_reqs() will do the necessary RCU synchronization */
832	wake_up_all(&ctx->wait);
833
834	/*
835	 * It'd be more correct to do this in free_ioctx(), after all
836	 * the outstanding kiocbs have finished - but by then io_destroy
837	 * has already returned, so io_setup() could potentially return
838	 * -EAGAIN with no ioctxs actually in use (as far as userspace
839	 *  could tell).
840	 */
841	aio_nr_sub(ctx->max_reqs);
842
843	if (ctx->mmap_size)
844		vm_munmap(ctx->mmap_base, ctx->mmap_size);
845
846	ctx->rq_wait = wait;
847	percpu_ref_kill(&ctx->users);
848	return 0;
849}
850
851/*
852 * exit_aio: called when the last user of mm goes away.  At this point, there is
853 * no way for any new requests to be submited or any of the io_* syscalls to be
854 * called on the context.
855 *
856 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
857 * them.
858 */
859void exit_aio(struct mm_struct *mm)
860{
861	struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
862	struct ctx_rq_wait wait;
863	int i, skipped;
864
865	if (!table)
866		return;
867
868	atomic_set(&wait.count, table->nr);
869	init_completion(&wait.comp);
870
871	skipped = 0;
872	for (i = 0; i < table->nr; ++i) {
873		struct kioctx *ctx =
874			rcu_dereference_protected(table->table[i], true);
875
876		if (!ctx) {
877			skipped++;
878			continue;
879		}
880
881		/*
882		 * We don't need to bother with munmap() here - exit_mmap(mm)
883		 * is coming and it'll unmap everything. And we simply can't,
884		 * this is not necessarily our ->mm.
885		 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
886		 * that it needs to unmap the area, just set it to 0.
887		 */
888		ctx->mmap_size = 0;
889		kill_ioctx(mm, ctx, &wait);
890	}
891
892	if (!atomic_sub_and_test(skipped, &wait.count)) {
893		/* Wait until all IO for the context are done. */
894		wait_for_completion(&wait.comp);
895	}
896
897	RCU_INIT_POINTER(mm->ioctx_table, NULL);
898	kfree(table);
899}
900
901static void put_reqs_available(struct kioctx *ctx, unsigned nr)
902{
903	struct kioctx_cpu *kcpu;
904	unsigned long flags;
905
906	local_irq_save(flags);
907	kcpu = this_cpu_ptr(ctx->cpu);
908	kcpu->reqs_available += nr;
909
910	while (kcpu->reqs_available >= ctx->req_batch * 2) {
911		kcpu->reqs_available -= ctx->req_batch;
912		atomic_add(ctx->req_batch, &ctx->reqs_available);
913	}
914
915	local_irq_restore(flags);
916}
917
918static bool __get_reqs_available(struct kioctx *ctx)
919{
920	struct kioctx_cpu *kcpu;
921	bool ret = false;
922	unsigned long flags;
923
924	local_irq_save(flags);
925	kcpu = this_cpu_ptr(ctx->cpu);
926	if (!kcpu->reqs_available) {
927		int old, avail = atomic_read(&ctx->reqs_available);
928
929		do {
930			if (avail < ctx->req_batch)
931				goto out;
932
933			old = avail;
934			avail = atomic_cmpxchg(&ctx->reqs_available,
935					       avail, avail - ctx->req_batch);
936		} while (avail != old);
937
938		kcpu->reqs_available += ctx->req_batch;
939	}
940
941	ret = true;
942	kcpu->reqs_available--;
943out:
944	local_irq_restore(flags);
945	return ret;
946}
947
948/* refill_reqs_available
949 *	Updates the reqs_available reference counts used for tracking the
950 *	number of free slots in the completion ring.  This can be called
951 *	from aio_complete() (to optimistically update reqs_available) or
952 *	from aio_get_req() (the we're out of events case).  It must be
953 *	called holding ctx->completion_lock.
954 */
955static void refill_reqs_available(struct kioctx *ctx, unsigned head,
956                                  unsigned tail)
957{
958	unsigned events_in_ring, completed;
959
960	/* Clamp head since userland can write to it. */
961	head %= ctx->nr_events;
962	if (head <= tail)
963		events_in_ring = tail - head;
964	else
965		events_in_ring = ctx->nr_events - (head - tail);
966
967	completed = ctx->completed_events;
968	if (events_in_ring < completed)
969		completed -= events_in_ring;
970	else
971		completed = 0;
972
973	if (!completed)
974		return;
975
976	ctx->completed_events -= completed;
977	put_reqs_available(ctx, completed);
978}
979
980/* user_refill_reqs_available
981 *	Called to refill reqs_available when aio_get_req() encounters an
982 *	out of space in the completion ring.
983 */
984static void user_refill_reqs_available(struct kioctx *ctx)
985{
986	spin_lock_irq(&ctx->completion_lock);
987	if (ctx->completed_events) {
988		struct aio_ring *ring;
989		unsigned head;
990
991		/* Access of ring->head may race with aio_read_events_ring()
992		 * here, but that's okay since whether we read the old version
993		 * or the new version, and either will be valid.  The important
994		 * part is that head cannot pass tail since we prevent
995		 * aio_complete() from updating tail by holding
996		 * ctx->completion_lock.  Even if head is invalid, the check
997		 * against ctx->completed_events below will make sure we do the
998		 * safe/right thing.
999		 */
1000		ring = kmap_atomic(ctx->ring_pages[0]);
1001		head = ring->head;
1002		kunmap_atomic(ring);
1003
1004		refill_reqs_available(ctx, head, ctx->tail);
1005	}
1006
1007	spin_unlock_irq(&ctx->completion_lock);
1008}
1009
1010static bool get_reqs_available(struct kioctx *ctx)
1011{
1012	if (__get_reqs_available(ctx))
1013		return true;
1014	user_refill_reqs_available(ctx);
1015	return __get_reqs_available(ctx);
1016}
1017
1018/* aio_get_req
1019 *	Allocate a slot for an aio request.
1020 * Returns NULL if no requests are free.
1021 *
1022 * The refcount is initialized to 2 - one for the async op completion,
1023 * one for the synchronous code that does this.
1024 */
1025static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1026{
1027	struct aio_kiocb *req;
1028
1029	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1030	if (unlikely(!req))
1031		return NULL;
1032
1033	if (unlikely(!get_reqs_available(ctx))) {
1034		kmem_cache_free(kiocb_cachep, req);
1035		return NULL;
1036	}
1037
1038	percpu_ref_get(&ctx->reqs);
1039	req->ki_ctx = ctx;
1040	INIT_LIST_HEAD(&req->ki_list);
1041	refcount_set(&req->ki_refcnt, 2);
1042	req->ki_eventfd = NULL;
1043	return req;
1044}
1045
1046static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1047{
1048	struct aio_ring __user *ring  = (void __user *)ctx_id;
1049	struct mm_struct *mm = current->mm;
1050	struct kioctx *ctx, *ret = NULL;
1051	struct kioctx_table *table;
1052	unsigned id;
1053
1054	if (get_user(id, &ring->id))
1055		return NULL;
1056
1057	rcu_read_lock();
1058	table = rcu_dereference(mm->ioctx_table);
1059
1060	if (!table || id >= table->nr)
1061		goto out;
1062
1063	id = array_index_nospec(id, table->nr);
1064	ctx = rcu_dereference(table->table[id]);
1065	if (ctx && ctx->user_id == ctx_id) {
1066		if (percpu_ref_tryget_live(&ctx->users))
1067			ret = ctx;
1068	}
1069out:
1070	rcu_read_unlock();
1071	return ret;
1072}
1073
1074static inline void iocb_destroy(struct aio_kiocb *iocb)
1075{
1076	if (iocb->ki_eventfd)
1077		eventfd_ctx_put(iocb->ki_eventfd);
1078	if (iocb->ki_filp)
1079		fput(iocb->ki_filp);
1080	percpu_ref_put(&iocb->ki_ctx->reqs);
1081	kmem_cache_free(kiocb_cachep, iocb);
1082}
1083
1084/* aio_complete
1085 *	Called when the io request on the given iocb is complete.
1086 */
1087static void aio_complete(struct aio_kiocb *iocb)
1088{
1089	struct kioctx	*ctx = iocb->ki_ctx;
1090	struct aio_ring	*ring;
1091	struct io_event	*ev_page, *event;
1092	unsigned tail, pos, head;
1093	unsigned long	flags;
1094
1095	/*
1096	 * Add a completion event to the ring buffer. Must be done holding
1097	 * ctx->completion_lock to prevent other code from messing with the tail
1098	 * pointer since we might be called from irq context.
1099	 */
1100	spin_lock_irqsave(&ctx->completion_lock, flags);
1101
1102	tail = ctx->tail;
1103	pos = tail + AIO_EVENTS_OFFSET;
1104
1105	if (++tail >= ctx->nr_events)
1106		tail = 0;
1107
1108	ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1109	event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1110
1111	*event = iocb->ki_res;
1112
1113	kunmap_atomic(ev_page);
1114	flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1115
1116	pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1117		 (void __user *)(unsigned long)iocb->ki_res.obj,
1118		 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1119
1120	/* after flagging the request as done, we
1121	 * must never even look at it again
1122	 */
1123	smp_wmb();	/* make event visible before updating tail */
1124
1125	ctx->tail = tail;
1126
1127	ring = kmap_atomic(ctx->ring_pages[0]);
1128	head = ring->head;
1129	ring->tail = tail;
1130	kunmap_atomic(ring);
1131	flush_dcache_page(ctx->ring_pages[0]);
1132
1133	ctx->completed_events++;
1134	if (ctx->completed_events > 1)
1135		refill_reqs_available(ctx, head, tail);
1136	spin_unlock_irqrestore(&ctx->completion_lock, flags);
1137
1138	pr_debug("added to ring %p at [%u]\n", iocb, tail);
1139
1140	/*
1141	 * Check if the user asked us to deliver the result through an
1142	 * eventfd. The eventfd_signal() function is safe to be called
1143	 * from IRQ context.
1144	 */
1145	if (iocb->ki_eventfd)
1146		eventfd_signal(iocb->ki_eventfd, 1);
1147
1148	/*
1149	 * We have to order our ring_info tail store above and test
1150	 * of the wait list below outside the wait lock.  This is
1151	 * like in wake_up_bit() where clearing a bit has to be
1152	 * ordered with the unlocked test.
1153	 */
1154	smp_mb();
1155
1156	if (waitqueue_active(&ctx->wait))
1157		wake_up(&ctx->wait);
1158}
1159
1160static inline void iocb_put(struct aio_kiocb *iocb)
1161{
1162	if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1163		aio_complete(iocb);
1164		iocb_destroy(iocb);
1165	}
1166}
1167
1168/* aio_read_events_ring
1169 *	Pull an event off of the ioctx's event ring.  Returns the number of
1170 *	events fetched
1171 */
1172static long aio_read_events_ring(struct kioctx *ctx,
1173				 struct io_event __user *event, long nr)
1174{
1175	struct aio_ring *ring;
1176	unsigned head, tail, pos;
1177	long ret = 0;
1178	int copy_ret;
1179
1180	/*
1181	 * The mutex can block and wake us up and that will cause
1182	 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1183	 * and repeat. This should be rare enough that it doesn't cause
1184	 * peformance issues. See the comment in read_events() for more detail.
1185	 */
1186	sched_annotate_sleep();
1187	mutex_lock(&ctx->ring_lock);
1188
1189	/* Access to ->ring_pages here is protected by ctx->ring_lock. */
1190	ring = kmap_atomic(ctx->ring_pages[0]);
1191	head = ring->head;
1192	tail = ring->tail;
1193	kunmap_atomic(ring);
1194
1195	/*
1196	 * Ensure that once we've read the current tail pointer, that
1197	 * we also see the events that were stored up to the tail.
1198	 */
1199	smp_rmb();
1200
1201	pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1202
1203	if (head == tail)
1204		goto out;
1205
1206	head %= ctx->nr_events;
1207	tail %= ctx->nr_events;
1208
1209	while (ret < nr) {
1210		long avail;
1211		struct io_event *ev;
1212		struct page *page;
1213
1214		avail = (head <= tail ?  tail : ctx->nr_events) - head;
1215		if (head == tail)
1216			break;
1217
1218		pos = head + AIO_EVENTS_OFFSET;
1219		page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1220		pos %= AIO_EVENTS_PER_PAGE;
1221
1222		avail = min(avail, nr - ret);
1223		avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1224
1225		ev = kmap(page);
1226		copy_ret = copy_to_user(event + ret, ev + pos,
1227					sizeof(*ev) * avail);
1228		kunmap(page);
1229
1230		if (unlikely(copy_ret)) {
1231			ret = -EFAULT;
1232			goto out;
1233		}
1234
1235		ret += avail;
1236		head += avail;
1237		head %= ctx->nr_events;
1238	}
1239
1240	ring = kmap_atomic(ctx->ring_pages[0]);
1241	ring->head = head;
1242	kunmap_atomic(ring);
1243	flush_dcache_page(ctx->ring_pages[0]);
1244
1245	pr_debug("%li  h%u t%u\n", ret, head, tail);
1246out:
1247	mutex_unlock(&ctx->ring_lock);
1248
1249	return ret;
1250}
1251
1252static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1253			    struct io_event __user *event, long *i)
1254{
1255	long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1256
1257	if (ret > 0)
1258		*i += ret;
1259
1260	if (unlikely(atomic_read(&ctx->dead)))
1261		ret = -EINVAL;
1262
1263	if (!*i)
1264		*i = ret;
1265
1266	return ret < 0 || *i >= min_nr;
1267}
1268
1269static long read_events(struct kioctx *ctx, long min_nr, long nr,
1270			struct io_event __user *event,
1271			ktime_t until)
1272{
1273	long ret = 0;
1274
1275	/*
1276	 * Note that aio_read_events() is being called as the conditional - i.e.
1277	 * we're calling it after prepare_to_wait() has set task state to
1278	 * TASK_INTERRUPTIBLE.
1279	 *
1280	 * But aio_read_events() can block, and if it blocks it's going to flip
1281	 * the task state back to TASK_RUNNING.
1282	 *
1283	 * This should be ok, provided it doesn't flip the state back to
1284	 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1285	 * will only happen if the mutex_lock() call blocks, and we then find
1286	 * the ringbuffer empty. So in practice we should be ok, but it's
1287	 * something to be aware of when touching this code.
1288	 */
1289	if (until == 0)
1290		aio_read_events(ctx, min_nr, nr, event, &ret);
1291	else
1292		wait_event_interruptible_hrtimeout(ctx->wait,
1293				aio_read_events(ctx, min_nr, nr, event, &ret),
1294				until);
1295	return ret;
1296}
1297
1298/* sys_io_setup:
1299 *	Create an aio_context capable of receiving at least nr_events.
1300 *	ctxp must not point to an aio_context that already exists, and
1301 *	must be initialized to 0 prior to the call.  On successful
1302 *	creation of the aio_context, *ctxp is filled in with the resulting
1303 *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1304 *	if the specified nr_events exceeds internal limits.  May fail
1305 *	with -EAGAIN if the specified nr_events exceeds the user's limit
1306 *	of available events.  May fail with -ENOMEM if insufficient kernel
1307 *	resources are available.  May fail with -EFAULT if an invalid
1308 *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1309 *	implemented.
1310 */
1311SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1312{
1313	struct kioctx *ioctx = NULL;
1314	unsigned long ctx;
1315	long ret;
1316
1317	ret = get_user(ctx, ctxp);
1318	if (unlikely(ret))
1319		goto out;
1320
1321	ret = -EINVAL;
1322	if (unlikely(ctx || nr_events == 0)) {
1323		pr_debug("EINVAL: ctx %lu nr_events %u\n",
1324		         ctx, nr_events);
1325		goto out;
1326	}
1327
1328	ioctx = ioctx_alloc(nr_events);
1329	ret = PTR_ERR(ioctx);
1330	if (!IS_ERR(ioctx)) {
1331		ret = put_user(ioctx->user_id, ctxp);
1332		if (ret)
1333			kill_ioctx(current->mm, ioctx, NULL);
1334		percpu_ref_put(&ioctx->users);
1335	}
1336
1337out:
1338	return ret;
1339}
1340
1341#ifdef CONFIG_COMPAT
1342COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1343{
1344	struct kioctx *ioctx = NULL;
1345	unsigned long ctx;
1346	long ret;
1347
1348	ret = get_user(ctx, ctx32p);
1349	if (unlikely(ret))
1350		goto out;
1351
1352	ret = -EINVAL;
1353	if (unlikely(ctx || nr_events == 0)) {
1354		pr_debug("EINVAL: ctx %lu nr_events %u\n",
1355		         ctx, nr_events);
1356		goto out;
1357	}
1358
1359	ioctx = ioctx_alloc(nr_events);
1360	ret = PTR_ERR(ioctx);
1361	if (!IS_ERR(ioctx)) {
1362		/* truncating is ok because it's a user address */
1363		ret = put_user((u32)ioctx->user_id, ctx32p);
1364		if (ret)
1365			kill_ioctx(current->mm, ioctx, NULL);
1366		percpu_ref_put(&ioctx->users);
1367	}
1368
1369out:
1370	return ret;
1371}
1372#endif
1373
1374/* sys_io_destroy:
1375 *	Destroy the aio_context specified.  May cancel any outstanding
1376 *	AIOs and block on completion.  Will fail with -ENOSYS if not
1377 *	implemented.  May fail with -EINVAL if the context pointed to
1378 *	is invalid.
1379 */
1380SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1381{
1382	struct kioctx *ioctx = lookup_ioctx(ctx);
1383	if (likely(NULL != ioctx)) {
1384		struct ctx_rq_wait wait;
1385		int ret;
1386
1387		init_completion(&wait.comp);
1388		atomic_set(&wait.count, 1);
1389
1390		/* Pass requests_done to kill_ioctx() where it can be set
1391		 * in a thread-safe way. If we try to set it here then we have
1392		 * a race condition if two io_destroy() called simultaneously.
1393		 */
1394		ret = kill_ioctx(current->mm, ioctx, &wait);
1395		percpu_ref_put(&ioctx->users);
1396
1397		/* Wait until all IO for the context are done. Otherwise kernel
1398		 * keep using user-space buffers even if user thinks the context
1399		 * is destroyed.
1400		 */
1401		if (!ret)
1402			wait_for_completion(&wait.comp);
1403
1404		return ret;
1405	}
1406	pr_debug("EINVAL: invalid context id\n");
1407	return -EINVAL;
1408}
1409
1410static void aio_remove_iocb(struct aio_kiocb *iocb)
1411{
1412	struct kioctx *ctx = iocb->ki_ctx;
1413	unsigned long flags;
1414
1415	spin_lock_irqsave(&ctx->ctx_lock, flags);
1416	list_del(&iocb->ki_list);
1417	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1418}
1419
1420static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1421{
1422	struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1423
1424	if (!list_empty_careful(&iocb->ki_list))
1425		aio_remove_iocb(iocb);
1426
1427	if (kiocb->ki_flags & IOCB_WRITE) {
1428		struct inode *inode = file_inode(kiocb->ki_filp);
1429
1430		/*
1431		 * Tell lockdep we inherited freeze protection from submission
1432		 * thread.
1433		 */
1434		if (S_ISREG(inode->i_mode))
1435			__sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1436		file_end_write(kiocb->ki_filp);
1437	}
1438
1439	iocb->ki_res.res = res;
1440	iocb->ki_res.res2 = res2;
1441	iocb_put(iocb);
1442}
1443
1444static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1445{
1446	int ret;
1447
1448	req->ki_complete = aio_complete_rw;
1449	req->private = NULL;
1450	req->ki_pos = iocb->aio_offset;
1451	req->ki_flags = iocb_flags(req->ki_filp);
1452	if (iocb->aio_flags & IOCB_FLAG_RESFD)
1453		req->ki_flags |= IOCB_EVENTFD;
1454	req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1455	if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1456		/*
1457		 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1458		 * aio_reqprio is interpreted as an I/O scheduling
1459		 * class and priority.
1460		 */
1461		ret = ioprio_check_cap(iocb->aio_reqprio);
1462		if (ret) {
1463			pr_debug("aio ioprio check cap error: %d\n", ret);
1464			return ret;
1465		}
1466
1467		req->ki_ioprio = iocb->aio_reqprio;
1468	} else
1469		req->ki_ioprio = get_current_ioprio();
1470
1471	ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1472	if (unlikely(ret))
1473		return ret;
1474
1475	req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1476	return 0;
1477}
1478
1479static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1480		struct iovec **iovec, bool vectored, bool compat,
1481		struct iov_iter *iter)
1482{
1483	void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1484	size_t len = iocb->aio_nbytes;
1485
1486	if (!vectored) {
1487		ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1488		*iovec = NULL;
1489		return ret;
1490	}
1491
1492	return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1493}
1494
1495static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1496{
1497	switch (ret) {
1498	case -EIOCBQUEUED:
1499		break;
1500	case -ERESTARTSYS:
1501	case -ERESTARTNOINTR:
1502	case -ERESTARTNOHAND:
1503	case -ERESTART_RESTARTBLOCK:
1504		/*
1505		 * There's no easy way to restart the syscall since other AIO's
1506		 * may be already running. Just fail this IO with EINTR.
1507		 */
1508		ret = -EINTR;
1509		fallthrough;
1510	default:
1511		req->ki_complete(req, ret, 0);
1512	}
1513}
1514
1515static int aio_read(struct kiocb *req, const struct iocb *iocb,
1516			bool vectored, bool compat)
1517{
1518	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1519	struct iov_iter iter;
1520	struct file *file;
1521	int ret;
1522
1523	ret = aio_prep_rw(req, iocb);
1524	if (ret)
1525		return ret;
1526	file = req->ki_filp;
1527	if (unlikely(!(file->f_mode & FMODE_READ)))
1528		return -EBADF;
1529	ret = -EINVAL;
1530	if (unlikely(!file->f_op->read_iter))
1531		return -EINVAL;
1532
1533	ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1534	if (ret < 0)
1535		return ret;
1536	ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1537	if (!ret)
1538		aio_rw_done(req, call_read_iter(file, req, &iter));
1539	kfree(iovec);
1540	return ret;
1541}
1542
1543static int aio_write(struct kiocb *req, const struct iocb *iocb,
1544			 bool vectored, bool compat)
1545{
1546	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1547	struct iov_iter iter;
1548	struct file *file;
1549	int ret;
1550
1551	ret = aio_prep_rw(req, iocb);
1552	if (ret)
1553		return ret;
1554	file = req->ki_filp;
1555
1556	if (unlikely(!(file->f_mode & FMODE_WRITE)))
1557		return -EBADF;
1558	if (unlikely(!file->f_op->write_iter))
1559		return -EINVAL;
1560
1561	ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1562	if (ret < 0)
1563		return ret;
1564	ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1565	if (!ret) {
1566		/*
1567		 * Open-code file_start_write here to grab freeze protection,
1568		 * which will be released by another thread in
1569		 * aio_complete_rw().  Fool lockdep by telling it the lock got
1570		 * released so that it doesn't complain about the held lock when
1571		 * we return to userspace.
1572		 */
1573		if (S_ISREG(file_inode(file)->i_mode)) {
1574			sb_start_write(file_inode(file)->i_sb);
1575			__sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1576		}
1577		req->ki_flags |= IOCB_WRITE;
1578		aio_rw_done(req, call_write_iter(file, req, &iter));
1579	}
1580	kfree(iovec);
1581	return ret;
1582}
1583
1584static void aio_fsync_work(struct work_struct *work)
1585{
1586	struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1587	const struct cred *old_cred = override_creds(iocb->fsync.creds);
1588
1589	iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1590	revert_creds(old_cred);
1591	put_cred(iocb->fsync.creds);
1592	iocb_put(iocb);
1593}
1594
1595static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1596		     bool datasync)
1597{
1598	if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1599			iocb->aio_rw_flags))
1600		return -EINVAL;
1601
1602	if (unlikely(!req->file->f_op->fsync))
1603		return -EINVAL;
1604
1605	req->creds = prepare_creds();
1606	if (!req->creds)
1607		return -ENOMEM;
1608
1609	req->datasync = datasync;
1610	INIT_WORK(&req->work, aio_fsync_work);
1611	schedule_work(&req->work);
1612	return 0;
1613}
1614
1615static void aio_poll_put_work(struct work_struct *work)
1616{
1617	struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1618	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1619
1620	iocb_put(iocb);
1621}
1622
1623static void aio_poll_complete_work(struct work_struct *work)
1624{
1625	struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1626	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1627	struct poll_table_struct pt = { ._key = req->events };
1628	struct kioctx *ctx = iocb->ki_ctx;
1629	__poll_t mask = 0;
1630
1631	if (!READ_ONCE(req->cancelled))
1632		mask = vfs_poll(req->file, &pt) & req->events;
1633
1634	/*
1635	 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1636	 * calling ->ki_cancel.  We need the ctx_lock roundtrip here to
1637	 * synchronize with them.  In the cancellation case the list_del_init
1638	 * itself is not actually needed, but harmless so we keep it in to
1639	 * avoid further branches in the fast path.
1640	 */
1641	spin_lock_irq(&ctx->ctx_lock);
1642	if (!mask && !READ_ONCE(req->cancelled)) {
1643		add_wait_queue(req->head, &req->wait);
1644		spin_unlock_irq(&ctx->ctx_lock);
1645		return;
1646	}
1647	list_del_init(&iocb->ki_list);
1648	iocb->ki_res.res = mangle_poll(mask);
1649	req->done = true;
1650	spin_unlock_irq(&ctx->ctx_lock);
1651
1652	iocb_put(iocb);
1653}
1654
1655/* assumes we are called with irqs disabled */
1656static int aio_poll_cancel(struct kiocb *iocb)
1657{
1658	struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1659	struct poll_iocb *req = &aiocb->poll;
1660
1661	spin_lock(&req->head->lock);
1662	WRITE_ONCE(req->cancelled, true);
1663	if (!list_empty(&req->wait.entry)) {
1664		list_del_init(&req->wait.entry);
1665		schedule_work(&aiocb->poll.work);
1666	}
1667	spin_unlock(&req->head->lock);
1668
1669	return 0;
1670}
1671
1672static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1673		void *key)
1674{
1675	struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1676	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1677	__poll_t mask = key_to_poll(key);
1678	unsigned long flags;
1679
1680	/* for instances that support it check for an event match first: */
1681	if (mask && !(mask & req->events))
1682		return 0;
1683
1684	list_del_init(&req->wait.entry);
1685
1686	if (mask && spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1687		struct kioctx *ctx = iocb->ki_ctx;
1688
1689		/*
1690		 * Try to complete the iocb inline if we can. Use
1691		 * irqsave/irqrestore because not all filesystems (e.g. fuse)
1692		 * call this function with IRQs disabled and because IRQs
1693		 * have to be disabled before ctx_lock is obtained.
1694		 */
1695		list_del(&iocb->ki_list);
1696		iocb->ki_res.res = mangle_poll(mask);
1697		req->done = true;
1698		if (iocb->ki_eventfd && eventfd_signal_allowed()) {
1699			iocb = NULL;
1700			INIT_WORK(&req->work, aio_poll_put_work);
1701			schedule_work(&req->work);
1702		}
1703		spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1704		if (iocb)
1705			iocb_put(iocb);
1706	} else {
1707		schedule_work(&req->work);
1708	}
1709	return 1;
1710}
1711
1712struct aio_poll_table {
1713	struct poll_table_struct	pt;
1714	struct aio_kiocb		*iocb;
1715	int				error;
1716};
1717
1718static void
1719aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1720		struct poll_table_struct *p)
1721{
1722	struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1723
1724	/* multiple wait queues per file are not supported */
1725	if (unlikely(pt->iocb->poll.head)) {
1726		pt->error = -EINVAL;
1727		return;
1728	}
1729
1730	pt->error = 0;
1731	pt->iocb->poll.head = head;
1732	add_wait_queue(head, &pt->iocb->poll.wait);
1733}
1734
1735static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1736{
1737	struct kioctx *ctx = aiocb->ki_ctx;
1738	struct poll_iocb *req = &aiocb->poll;
1739	struct aio_poll_table apt;
1740	bool cancel = false;
1741	__poll_t mask;
1742
1743	/* reject any unknown events outside the normal event mask. */
1744	if ((u16)iocb->aio_buf != iocb->aio_buf)
1745		return -EINVAL;
1746	/* reject fields that are not defined for poll */
1747	if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1748		return -EINVAL;
1749
1750	INIT_WORK(&req->work, aio_poll_complete_work);
1751	req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1752
1753	req->head = NULL;
1754	req->done = false;
1755	req->cancelled = false;
1756
1757	apt.pt._qproc = aio_poll_queue_proc;
1758	apt.pt._key = req->events;
1759	apt.iocb = aiocb;
1760	apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1761
1762	/* initialized the list so that we can do list_empty checks */
1763	INIT_LIST_HEAD(&req->wait.entry);
1764	init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1765
1766	mask = vfs_poll(req->file, &apt.pt) & req->events;
1767	spin_lock_irq(&ctx->ctx_lock);
1768	if (likely(req->head)) {
1769		spin_lock(&req->head->lock);
1770		if (unlikely(list_empty(&req->wait.entry))) {
1771			if (apt.error)
1772				cancel = true;
1773			apt.error = 0;
1774			mask = 0;
1775		}
1776		if (mask || apt.error) {
1777			list_del_init(&req->wait.entry);
1778		} else if (cancel) {
1779			WRITE_ONCE(req->cancelled, true);
1780		} else if (!req->done) { /* actually waiting for an event */
1781			list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1782			aiocb->ki_cancel = aio_poll_cancel;
1783		}
1784		spin_unlock(&req->head->lock);
1785	}
1786	if (mask) { /* no async, we'd stolen it */
1787		aiocb->ki_res.res = mangle_poll(mask);
1788		apt.error = 0;
1789	}
1790	spin_unlock_irq(&ctx->ctx_lock);
1791	if (mask)
1792		iocb_put(aiocb);
1793	return apt.error;
1794}
1795
1796static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1797			   struct iocb __user *user_iocb, struct aio_kiocb *req,
1798			   bool compat)
1799{
1800	req->ki_filp = fget(iocb->aio_fildes);
1801	if (unlikely(!req->ki_filp))
1802		return -EBADF;
1803
1804	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1805		struct eventfd_ctx *eventfd;
1806		/*
1807		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1808		 * instance of the file* now. The file descriptor must be
1809		 * an eventfd() fd, and will be signaled for each completed
1810		 * event using the eventfd_signal() function.
1811		 */
1812		eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1813		if (IS_ERR(eventfd))
1814			return PTR_ERR(eventfd);
1815
1816		req->ki_eventfd = eventfd;
1817	}
1818
1819	if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1820		pr_debug("EFAULT: aio_key\n");
1821		return -EFAULT;
1822	}
1823
1824	req->ki_res.obj = (u64)(unsigned long)user_iocb;
1825	req->ki_res.data = iocb->aio_data;
1826	req->ki_res.res = 0;
1827	req->ki_res.res2 = 0;
1828
1829	switch (iocb->aio_lio_opcode) {
1830	case IOCB_CMD_PREAD:
1831		return aio_read(&req->rw, iocb, false, compat);
1832	case IOCB_CMD_PWRITE:
1833		return aio_write(&req->rw, iocb, false, compat);
1834	case IOCB_CMD_PREADV:
1835		return aio_read(&req->rw, iocb, true, compat);
1836	case IOCB_CMD_PWRITEV:
1837		return aio_write(&req->rw, iocb, true, compat);
1838	case IOCB_CMD_FSYNC:
1839		return aio_fsync(&req->fsync, iocb, false);
1840	case IOCB_CMD_FDSYNC:
1841		return aio_fsync(&req->fsync, iocb, true);
1842	case IOCB_CMD_POLL:
1843		return aio_poll(req, iocb);
1844	default:
1845		pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1846		return -EINVAL;
1847	}
1848}
1849
1850static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1851			 bool compat)
1852{
1853	struct aio_kiocb *req;
1854	struct iocb iocb;
1855	int err;
1856
1857	if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1858		return -EFAULT;
1859
1860	/* enforce forwards compatibility on users */
1861	if (unlikely(iocb.aio_reserved2)) {
1862		pr_debug("EINVAL: reserve field set\n");
1863		return -EINVAL;
1864	}
1865
1866	/* prevent overflows */
1867	if (unlikely(
1868	    (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
1869	    (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
1870	    ((ssize_t)iocb.aio_nbytes < 0)
1871	   )) {
1872		pr_debug("EINVAL: overflow check\n");
1873		return -EINVAL;
1874	}
1875
1876	req = aio_get_req(ctx);
1877	if (unlikely(!req))
1878		return -EAGAIN;
1879
1880	err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
1881
1882	/* Done with the synchronous reference */
1883	iocb_put(req);
1884
1885	/*
1886	 * If err is 0, we'd either done aio_complete() ourselves or have
1887	 * arranged for that to be done asynchronously.  Anything non-zero
1888	 * means that we need to destroy req ourselves.
1889	 */
1890	if (unlikely(err)) {
1891		iocb_destroy(req);
1892		put_reqs_available(ctx, 1);
1893	}
1894	return err;
1895}
1896
1897/* sys_io_submit:
1898 *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1899 *	the number of iocbs queued.  May return -EINVAL if the aio_context
1900 *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
1901 *	*iocbpp[0] is not properly initialized, if the operation specified
1902 *	is invalid for the file descriptor in the iocb.  May fail with
1903 *	-EFAULT if any of the data structures point to invalid data.  May
1904 *	fail with -EBADF if the file descriptor specified in the first
1905 *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
1906 *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1907 *	fail with -ENOSYS if not implemented.
1908 */
1909SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1910		struct iocb __user * __user *, iocbpp)
1911{
1912	struct kioctx *ctx;
1913	long ret = 0;
1914	int i = 0;
1915	struct blk_plug plug;
1916
1917	if (unlikely(nr < 0))
1918		return -EINVAL;
1919
1920	ctx = lookup_ioctx(ctx_id);
1921	if (unlikely(!ctx)) {
1922		pr_debug("EINVAL: invalid context id\n");
1923		return -EINVAL;
1924	}
1925
1926	if (nr > ctx->nr_events)
1927		nr = ctx->nr_events;
1928
1929	if (nr > AIO_PLUG_THRESHOLD)
1930		blk_start_plug(&plug);
1931	for (i = 0; i < nr; i++) {
1932		struct iocb __user *user_iocb;
1933
1934		if (unlikely(get_user(user_iocb, iocbpp + i))) {
1935			ret = -EFAULT;
1936			break;
1937		}
1938
1939		ret = io_submit_one(ctx, user_iocb, false);
1940		if (ret)
1941			break;
1942	}
1943	if (nr > AIO_PLUG_THRESHOLD)
1944		blk_finish_plug(&plug);
1945
1946	percpu_ref_put(&ctx->users);
1947	return i ? i : ret;
1948}
1949
1950#ifdef CONFIG_COMPAT
1951COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1952		       int, nr, compat_uptr_t __user *, iocbpp)
1953{
1954	struct kioctx *ctx;
1955	long ret = 0;
1956	int i = 0;
1957	struct blk_plug plug;
1958
1959	if (unlikely(nr < 0))
1960		return -EINVAL;
1961
1962	ctx = lookup_ioctx(ctx_id);
1963	if (unlikely(!ctx)) {
1964		pr_debug("EINVAL: invalid context id\n");
1965		return -EINVAL;
1966	}
1967
1968	if (nr > ctx->nr_events)
1969		nr = ctx->nr_events;
1970
1971	if (nr > AIO_PLUG_THRESHOLD)
1972		blk_start_plug(&plug);
1973	for (i = 0; i < nr; i++) {
1974		compat_uptr_t user_iocb;
1975
1976		if (unlikely(get_user(user_iocb, iocbpp + i))) {
1977			ret = -EFAULT;
1978			break;
1979		}
1980
1981		ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
1982		if (ret)
1983			break;
1984	}
1985	if (nr > AIO_PLUG_THRESHOLD)
1986		blk_finish_plug(&plug);
1987
1988	percpu_ref_put(&ctx->users);
1989	return i ? i : ret;
1990}
1991#endif
1992
1993/* sys_io_cancel:
1994 *	Attempts to cancel an iocb previously passed to io_submit.  If
1995 *	the operation is successfully cancelled, the resulting event is
1996 *	copied into the memory pointed to by result without being placed
1997 *	into the completion queue and 0 is returned.  May fail with
1998 *	-EFAULT if any of the data structures pointed to are invalid.
1999 *	May fail with -EINVAL if aio_context specified by ctx_id is
2000 *	invalid.  May fail with -EAGAIN if the iocb specified was not
2001 *	cancelled.  Will fail with -ENOSYS if not implemented.
2002 */
2003SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2004		struct io_event __user *, result)
2005{
2006	struct kioctx *ctx;
2007	struct aio_kiocb *kiocb;
2008	int ret = -EINVAL;
2009	u32 key;
2010	u64 obj = (u64)(unsigned long)iocb;
2011
2012	if (unlikely(get_user(key, &iocb->aio_key)))
2013		return -EFAULT;
2014	if (unlikely(key != KIOCB_KEY))
2015		return -EINVAL;
2016
2017	ctx = lookup_ioctx(ctx_id);
2018	if (unlikely(!ctx))
2019		return -EINVAL;
2020
2021	spin_lock_irq(&ctx->ctx_lock);
2022	/* TODO: use a hash or array, this sucks. */
2023	list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2024		if (kiocb->ki_res.obj == obj) {
2025			ret = kiocb->ki_cancel(&kiocb->rw);
2026			list_del_init(&kiocb->ki_list);
2027			break;
2028		}
2029	}
2030	spin_unlock_irq(&ctx->ctx_lock);
2031
2032	if (!ret) {
2033		/*
2034		 * The result argument is no longer used - the io_event is
2035		 * always delivered via the ring buffer. -EINPROGRESS indicates
2036		 * cancellation is progress:
2037		 */
2038		ret = -EINPROGRESS;
2039	}
2040
2041	percpu_ref_put(&ctx->users);
2042
2043	return ret;
2044}
2045
2046static long do_io_getevents(aio_context_t ctx_id,
2047		long min_nr,
2048		long nr,
2049		struct io_event __user *events,
2050		struct timespec64 *ts)
2051{
2052	ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2053	struct kioctx *ioctx = lookup_ioctx(ctx_id);
2054	long ret = -EINVAL;
2055
2056	if (likely(ioctx)) {
2057		if (likely(min_nr <= nr && min_nr >= 0))
2058			ret = read_events(ioctx, min_nr, nr, events, until);
2059		percpu_ref_put(&ioctx->users);
2060	}
2061
2062	return ret;
2063}
2064
2065/* io_getevents:
2066 *	Attempts to read at least min_nr events and up to nr events from
2067 *	the completion queue for the aio_context specified by ctx_id. If
2068 *	it succeeds, the number of read events is returned. May fail with
2069 *	-EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2070 *	out of range, if timeout is out of range.  May fail with -EFAULT
2071 *	if any of the memory specified is invalid.  May return 0 or
2072 *	< min_nr if the timeout specified by timeout has elapsed
2073 *	before sufficient events are available, where timeout == NULL
2074 *	specifies an infinite timeout. Note that the timeout pointed to by
2075 *	timeout is relative.  Will fail with -ENOSYS if not implemented.
2076 */
2077#ifdef CONFIG_64BIT
2078
2079SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2080		long, min_nr,
2081		long, nr,
2082		struct io_event __user *, events,
2083		struct __kernel_timespec __user *, timeout)
2084{
2085	struct timespec64	ts;
2086	int			ret;
2087
2088	if (timeout && unlikely(get_timespec64(&ts, timeout)))
2089		return -EFAULT;
2090
2091	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2092	if (!ret && signal_pending(current))
2093		ret = -EINTR;
2094	return ret;
2095}
2096
2097#endif
2098
2099struct __aio_sigset {
2100	const sigset_t __user	*sigmask;
2101	size_t		sigsetsize;
2102};
2103
2104SYSCALL_DEFINE6(io_pgetevents,
2105		aio_context_t, ctx_id,
2106		long, min_nr,
2107		long, nr,
2108		struct io_event __user *, events,
2109		struct __kernel_timespec __user *, timeout,
2110		const struct __aio_sigset __user *, usig)
2111{
2112	struct __aio_sigset	ksig = { NULL, };
2113	struct timespec64	ts;
2114	bool interrupted;
2115	int ret;
2116
2117	if (timeout && unlikely(get_timespec64(&ts, timeout)))
2118		return -EFAULT;
2119
2120	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2121		return -EFAULT;
2122
2123	ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2124	if (ret)
2125		return ret;
2126
2127	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2128
2129	interrupted = signal_pending(current);
2130	restore_saved_sigmask_unless(interrupted);
2131	if (interrupted && !ret)
2132		ret = -ERESTARTNOHAND;
2133
2134	return ret;
2135}
2136
2137#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2138
2139SYSCALL_DEFINE6(io_pgetevents_time32,
2140		aio_context_t, ctx_id,
2141		long, min_nr,
2142		long, nr,
2143		struct io_event __user *, events,
2144		struct old_timespec32 __user *, timeout,
2145		const struct __aio_sigset __user *, usig)
2146{
2147	struct __aio_sigset	ksig = { NULL, };
2148	struct timespec64	ts;
2149	bool interrupted;
2150	int ret;
2151
2152	if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2153		return -EFAULT;
2154
2155	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2156		return -EFAULT;
2157
2158
2159	ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2160	if (ret)
2161		return ret;
2162
2163	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2164
2165	interrupted = signal_pending(current);
2166	restore_saved_sigmask_unless(interrupted);
2167	if (interrupted && !ret)
2168		ret = -ERESTARTNOHAND;
2169
2170	return ret;
2171}
2172
2173#endif
2174
2175#if defined(CONFIG_COMPAT_32BIT_TIME)
2176
2177SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2178		__s32, min_nr,
2179		__s32, nr,
2180		struct io_event __user *, events,
2181		struct old_timespec32 __user *, timeout)
2182{
2183	struct timespec64 t;
2184	int ret;
2185
2186	if (timeout && get_old_timespec32(&t, timeout))
2187		return -EFAULT;
2188
2189	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2190	if (!ret && signal_pending(current))
2191		ret = -EINTR;
2192	return ret;
2193}
2194
2195#endif
2196
2197#ifdef CONFIG_COMPAT
2198
2199struct __compat_aio_sigset {
2200	compat_uptr_t		sigmask;
2201	compat_size_t		sigsetsize;
2202};
2203
2204#if defined(CONFIG_COMPAT_32BIT_TIME)
2205
2206COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2207		compat_aio_context_t, ctx_id,
2208		compat_long_t, min_nr,
2209		compat_long_t, nr,
2210		struct io_event __user *, events,
2211		struct old_timespec32 __user *, timeout,
2212		const struct __compat_aio_sigset __user *, usig)
2213{
2214	struct __compat_aio_sigset ksig = { 0, };
2215	struct timespec64 t;
2216	bool interrupted;
2217	int ret;
2218
2219	if (timeout && get_old_timespec32(&t, timeout))
2220		return -EFAULT;
2221
2222	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2223		return -EFAULT;
2224
2225	ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2226	if (ret)
2227		return ret;
2228
2229	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2230
2231	interrupted = signal_pending(current);
2232	restore_saved_sigmask_unless(interrupted);
2233	if (interrupted && !ret)
2234		ret = -ERESTARTNOHAND;
2235
2236	return ret;
2237}
2238
2239#endif
2240
2241COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2242		compat_aio_context_t, ctx_id,
2243		compat_long_t, min_nr,
2244		compat_long_t, nr,
2245		struct io_event __user *, events,
2246		struct __kernel_timespec __user *, timeout,
2247		const struct __compat_aio_sigset __user *, usig)
2248{
2249	struct __compat_aio_sigset ksig = { 0, };
2250	struct timespec64 t;
2251	bool interrupted;
2252	int ret;
2253
2254	if (timeout && get_timespec64(&t, timeout))
2255		return -EFAULT;
2256
2257	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2258		return -EFAULT;
2259
2260	ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2261	if (ret)
2262		return ret;
2263
2264	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2265
2266	interrupted = signal_pending(current);
2267	restore_saved_sigmask_unless(interrupted);
2268	if (interrupted && !ret)
2269		ret = -ERESTARTNOHAND;
2270
2271	return ret;
2272}
2273#endif
2274