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
56	struct 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
80	struct kioctx_table {
81	struct rcu_head rcu;
82	unsigned nr;
83	struct kioctx __rcu *table[] __counted_by(nr);
84	};
85
86	struct kioctx_cpu {
87	unsigned reqs_available;
88	};
89
90	struct ctx_rq_wait {
91	struct completion comp;
92	atomic_t count;
93	};
94
95	struct 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	*/
173	struct fsync_iocb {
174	struct file *file;
175	struct work_struct work;
176	bool datasync;
177	struct cred *creds;
178	};
179
180	struct 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	*/
197	struct 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----*/
222	static DEFINE_SPINLOCK(aio_nr_lock);
223	static unsigned long aio_nr; /* current system wide number of aio requests */
224	static unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
225	/*----end sysctl variables---*/
226	#ifdef CONFIG_SYSCTL
227	static 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	static void __init aio_sysctl_init(void)
245	{
246	register_sysctl_init("fs", aio_sysctls);
247	}
248	#else
249	#define aio_sysctl_init() do { } while (0)
250	#endif
251
252	static struct kmem_cache *kiocb_cachep;
253	static struct kmem_cache *kioctx_cachep;
254
255	static struct vfsmount *aio_mnt;
256
257	static const struct file_operations aio_ring_fops;
258	static const struct address_space_operations aio_ctx_aops;
259
260	static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
261	{
262	struct file *file;
263	struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
264	if (IS_ERR(ptr: inode))
265	return ERR_CAST(ptr: inode);
266
267	inode->i_mapping->a_ops = &aio_ctx_aops;
268	inode->i_mapping->i_private_data = ctx;
269	inode->i_size = PAGE_SIZE * nr_pages;
270
271	file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
272	O_RDWR, &aio_ring_fops);
273	if (IS_ERR(ptr: file))
274	iput(inode);
275	return file;
276	}
277
278	static int aio_init_fs_context(struct fs_context *fc)
279	{
280	if (!init_pseudo(fc, AIO_RING_MAGIC))
281	return -ENOMEM;
282	fc->s_iflags |= SB_I_NOEXEC;
283	return 0;
284	}
285
286	/* aio_setup
287	* Creates the slab caches used by the aio routines, panic on
288	* failure as this is done early during the boot sequence.
289	*/
290	static int __init aio_setup(void)
291	{
292	static struct file_system_type aio_fs = {
293	.name = "aio",
294	.init_fs_context = aio_init_fs_context,
295	.kill_sb = kill_anon_super,
296	};
297	aio_mnt = kern_mount(&aio_fs);
298	if (IS_ERR(ptr: aio_mnt))
299	panic(fmt: "Failed to create aio fs mount.");
300
301	kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
302	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
303	aio_sysctl_init();
304	return 0;
305	}
306	__initcall(aio_setup);
307
308	static void put_aio_ring_file(struct kioctx *ctx)
309	{
310	struct file *aio_ring_file = ctx->aio_ring_file;
311	struct address_space *i_mapping;
312
313	if (aio_ring_file) {
314	truncate_setsize(inode: file_inode(f: aio_ring_file), newsize: 0);
315
316	/* Prevent further access to the kioctx from migratepages */
317	i_mapping = aio_ring_file->f_mapping;
318	spin_lock(lock: &i_mapping->i_private_lock);
319	i_mapping->i_private_data = NULL;
320	ctx->aio_ring_file = NULL;
321	spin_unlock(lock: &i_mapping->i_private_lock);
322
323	fput(aio_ring_file);
324	}
325	}
326
327	static void aio_free_ring(struct kioctx *ctx)
328	{
329	int i;
330
331	/* Disconnect the kiotx from the ring file. This prevents future
332	* accesses to the kioctx from page migration.
333	*/
334	put_aio_ring_file(ctx);
335
336	for (i = 0; i < ctx->nr_pages; i++) {
337	struct page *page;
338	pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
339	page_count(ctx->ring_pages[i]));
340	page = ctx->ring_pages[i];
341	if (!page)
342	continue;
343	ctx->ring_pages[i] = NULL;
344	put_page(page);
345	}
346
347	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
348	kfree(objp: ctx->ring_pages);
349	ctx->ring_pages = NULL;
350	}
351	}
352
353	static int aio_ring_mremap(struct vm_area_struct *vma)
354	{
355	struct file *file = vma->vm_file;
356	struct mm_struct *mm = vma->vm_mm;
357	struct kioctx_table *table;
358	int i, res = -EINVAL;
359
360	spin_lock(lock: &mm->ioctx_lock);
361	rcu_read_lock();
362	table = rcu_dereference(mm->ioctx_table);
363	if (!table)
364	goto out_unlock;
365
366	for (i = 0; i < table->nr; i++) {
367	struct kioctx *ctx;
368
369	ctx = rcu_dereference(table->table[i]);
370	if (ctx && ctx->aio_ring_file == file) {
371	if (!atomic_read(v: &ctx->dead)) {
372	ctx->user_id = ctx->mmap_base = vma->vm_start;
373	res = 0;
374	}
375	break;
376	}
377	}
378
379	out_unlock:
380	rcu_read_unlock();
381	spin_unlock(lock: &mm->ioctx_lock);
382	return res;
383	}
384
385	static const struct vm_operations_struct aio_ring_vm_ops = {
386	.mremap = aio_ring_mremap,
387	#if IS_ENABLED(CONFIG_MMU)
388	.fault = filemap_fault,
389	.map_pages = filemap_map_pages,
390	.page_mkwrite = filemap_page_mkwrite,
391	#endif
392	};
393
394	static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
395	{
396	vm_flags_set(vma, VM_DONTEXPAND);
397	vma->vm_ops = &aio_ring_vm_ops;
398	return 0;
399	}
400
401	static const struct file_operations aio_ring_fops = {
402	.mmap = aio_ring_mmap,
403	};
404
405	#if IS_ENABLED(CONFIG_MIGRATION)
406	static int aio_migrate_folio(struct address_space *mapping, struct folio *dst,
407	struct folio *src, enum migrate_mode mode)
408	{
409	struct kioctx *ctx;
410	unsigned long flags;
411	pgoff_t idx;
412	int rc;
413
414	/*
415	* We cannot support the _NO_COPY case here, because copy needs to
416	* happen under the ctx->completion_lock. That does not work with the
417	* migration workflow of MIGRATE_SYNC_NO_COPY.
418	*/
419	if (mode == MIGRATE_SYNC_NO_COPY)
420	return -EINVAL;
421
422	rc = 0;
423
424	/* mapping->i_private_lock here protects against the kioctx teardown. */
425	spin_lock(lock: &mapping->i_private_lock);
426	ctx = mapping->i_private_data;
427	if (!ctx) {
428	rc = -EINVAL;
429	goto out;
430	}
431
432	/* The ring_lock mutex. The prevents aio_read_events() from writing
433	* to the ring's head, and prevents page migration from mucking in
434	* a partially initialized kiotx.
435	*/
436	if (!mutex_trylock(lock: &ctx->ring_lock)) {
437	rc = -EAGAIN;
438	goto out;
439	}
440
441	idx = src->index;
442	if (idx < (pgoff_t)ctx->nr_pages) {
443	/* Make sure the old folio hasn't already been changed */
444	if (ctx->ring_pages[idx] != &src->page)
445	rc = -EAGAIN;
446	} else
447	rc = -EINVAL;
448
449	if (rc != 0)
450	goto out_unlock;
451
452	/* Writeback must be complete */
453	BUG_ON(folio_test_writeback(src));
454	folio_get(folio: dst);
455
456	rc = folio_migrate_mapping(mapping, newfolio: dst, folio: src, extra_count: 1);
457	if (rc != MIGRATEPAGE_SUCCESS) {
458	folio_put(folio: dst);
459	goto out_unlock;
460	}
461
462	/* Take completion_lock to prevent other writes to the ring buffer
463	* while the old folio is copied to the new. This prevents new
464	* events from being lost.
465	*/
466	spin_lock_irqsave(&ctx->completion_lock, flags);
467	folio_migrate_copy(newfolio: dst, folio: src);
468	BUG_ON(ctx->ring_pages[idx] != &src->page);
469	ctx->ring_pages[idx] = &dst->page;
470	spin_unlock_irqrestore(lock: &ctx->completion_lock, flags);
471
472	/* The old folio is no longer accessible. */
473	folio_put(folio: src);
474
475	out_unlock:
476	mutex_unlock(lock: &ctx->ring_lock);
477	out:
478	spin_unlock(lock: &mapping->i_private_lock);
479	return rc;
480	}
481	#else
482	#define aio_migrate_folio NULL
483	#endif
484
485	static const struct address_space_operations aio_ctx_aops = {
486	.dirty_folio = noop_dirty_folio,
487	.migrate_folio = aio_migrate_folio,
488	};
489
490	static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
491	{
492	struct aio_ring *ring;
493	struct mm_struct *mm = current->mm;
494	unsigned long size, unused;
495	int nr_pages;
496	int i;
497	struct file *file;
498
499	/* Compensate for the ring buffer's head/tail overlap entry */
500	nr_events += 2; /* 1 is required, 2 for good luck */
501
502	size = sizeof(struct aio_ring);
503	size += sizeof(struct io_event) * nr_events;
504
505	nr_pages = PFN_UP(size);
506	if (nr_pages < 0)
507	return -EINVAL;
508
509	file = aio_private_file(ctx, nr_pages);
510	if (IS_ERR(ptr: file)) {
511	ctx->aio_ring_file = NULL;
512	return -ENOMEM;
513	}
514
515	ctx->aio_ring_file = file;
516	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
517	/ sizeof(struct io_event);
518
519	ctx->ring_pages = ctx->internal_pages;
520	if (nr_pages > AIO_RING_PAGES) {
521	ctx->ring_pages = kcalloc(n: nr_pages, size: sizeof(struct page *),
522	GFP_KERNEL);
523	if (!ctx->ring_pages) {
524	put_aio_ring_file(ctx);
525	return -ENOMEM;
526	}
527	}
528
529	for (i = 0; i < nr_pages; i++) {
530	struct page *page;
531	page = find_or_create_page(mapping: file->f_mapping,
532	index: i, GFP_USER | __GFP_ZERO);
533	if (!page)
534	break;
535	pr_debug("pid(%d) page[%d]->count=%d\n",
536	current->pid, i, page_count(page));
537	SetPageUptodate(page);
538	unlock_page(page);
539
540	ctx->ring_pages[i] = page;
541	}
542	ctx->nr_pages = i;
543
544	if (unlikely(i != nr_pages)) {
545	aio_free_ring(ctx);
546	return -ENOMEM;
547	}
548
549	ctx->mmap_size = nr_pages * PAGE_SIZE;
550	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
551
552	if (mmap_write_lock_killable(mm)) {
553	ctx->mmap_size = 0;
554	aio_free_ring(ctx);
555	return -EINTR;
556	}
557
558	ctx->mmap_base = do_mmap(file: ctx->aio_ring_file, addr: 0, len: ctx->mmap_size,
559	PROT_READ | PROT_WRITE,
560	MAP_SHARED, vm_flags: 0, pgoff: 0, populate: &unused, NULL);
561	mmap_write_unlock(mm);
562	if (IS_ERR(ptr: (void *)ctx->mmap_base)) {
563	ctx->mmap_size = 0;
564	aio_free_ring(ctx);
565	return -ENOMEM;
566	}
567
568	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
569
570	ctx->user_id = ctx->mmap_base;
571	ctx->nr_events = nr_events; /* trusted copy */
572
573	ring = page_address(ctx->ring_pages[0]);
574	ring->nr = nr_events; /* user copy */
575	ring->id = ~0U;
576	ring->head = ring->tail = 0;
577	ring->magic = AIO_RING_MAGIC;
578	ring->compat_features = AIO_RING_COMPAT_FEATURES;
579	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
580	ring->header_length = sizeof(struct aio_ring);
581	flush_dcache_page(page: ctx->ring_pages[0]);
582
583	return 0;
584	}
585
586	#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
587	#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
588	#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
589
590	void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
591	{
592	struct aio_kiocb *req;
593	struct kioctx *ctx;
594	unsigned long flags;
595
596	/*
597	* kiocb didn't come from aio or is neither a read nor a write, hence
598	* ignore it.
599	*/
600	if (!(iocb->ki_flags & IOCB_AIO_RW))
601	return;
602
603	req = container_of(iocb, struct aio_kiocb, rw);
604
605	if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
606	return;
607
608	ctx = req->ki_ctx;
609
610	spin_lock_irqsave(&ctx->ctx_lock, flags);
611	list_add_tail(new: &req->ki_list, head: &ctx->active_reqs);
612	req->ki_cancel = cancel;
613	spin_unlock_irqrestore(lock: &ctx->ctx_lock, flags);
614	}
615	EXPORT_SYMBOL(kiocb_set_cancel_fn);
616
617	/*
618	* free_ioctx() should be RCU delayed to synchronize against the RCU
619	* protected lookup_ioctx() and also needs process context to call
620	* aio_free_ring(). Use rcu_work.
621	*/
622	static void free_ioctx(struct work_struct *work)
623	{
624	struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
625	free_rwork);
626	pr_debug("freeing %p\n", ctx);
627
628	aio_free_ring(ctx);
629	free_percpu(pdata: ctx->cpu);
630	percpu_ref_exit(ref: &ctx->reqs);
631	percpu_ref_exit(ref: &ctx->users);
632	kmem_cache_free(s: kioctx_cachep, objp: ctx);
633	}
634
635	static void free_ioctx_reqs(struct percpu_ref *ref)
636	{
637	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
638
639	/* At this point we know that there are no any in-flight requests */
640	if (ctx->rq_wait && atomic_dec_and_test(v: &ctx->rq_wait->count))
641	complete(&ctx->rq_wait->comp);
642
643	/* Synchronize against RCU protected table->table[] dereferences */
644	INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
645	queue_rcu_work(wq: system_wq, rwork: &ctx->free_rwork);
646	}
647
648	/*
649	* When this function runs, the kioctx has been removed from the "hash table"
650	* and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
651	* now it's safe to cancel any that need to be.
652	*/
653	static void free_ioctx_users(struct percpu_ref *ref)
654	{
655	struct kioctx *ctx = container_of(ref, struct kioctx, users);
656	struct aio_kiocb *req;
657
658	spin_lock_irq(lock: &ctx->ctx_lock);
659
660	while (!list_empty(head: &ctx->active_reqs)) {
661	req = list_first_entry(&ctx->active_reqs,
662	struct aio_kiocb, ki_list);
663	req->ki_cancel(&req->rw);
664	list_del_init(entry: &req->ki_list);
665	}
666
667	spin_unlock_irq(lock: &ctx->ctx_lock);
668
669	percpu_ref_kill(ref: &ctx->reqs);
670	percpu_ref_put(ref: &ctx->reqs);
671	}
672
673	static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
674	{
675	unsigned i, new_nr;
676	struct kioctx_table *table, *old;
677	struct aio_ring *ring;
678
679	spin_lock(lock: &mm->ioctx_lock);
680	table = rcu_dereference_raw(mm->ioctx_table);
681
682	while (1) {
683	if (table)
684	for (i = 0; i < table->nr; i++)
685	if (!rcu_access_pointer(table->table[i])) {
686	ctx->id = i;
687	rcu_assign_pointer(table->table[i], ctx);
688	spin_unlock(lock: &mm->ioctx_lock);
689
690	/* While kioctx setup is in progress,
691	* we are protected from page migration
692	* changes ring_pages by ->ring_lock.
693	*/
694	ring = page_address(ctx->ring_pages[0]);
695	ring->id = ctx->id;
696	return 0;
697	}
698
699	new_nr = (table ? table->nr : 1) * 4;
700	spin_unlock(lock: &mm->ioctx_lock);
701
702	table = kzalloc(struct_size(table, table, new_nr), GFP_KERNEL);
703	if (!table)
704	return -ENOMEM;
705
706	table->nr = new_nr;
707
708	spin_lock(lock: &mm->ioctx_lock);
709	old = rcu_dereference_raw(mm->ioctx_table);
710
711	if (!old) {
712	rcu_assign_pointer(mm->ioctx_table, table);
713	} else if (table->nr > old->nr) {
714	memcpy(table->table, old->table,
715	old->nr * sizeof(struct kioctx *));
716
717	rcu_assign_pointer(mm->ioctx_table, table);
718	kfree_rcu(old, rcu);
719	} else {
720	kfree(objp: table);
721	table = old;
722	}
723	}
724	}
725
726	static void aio_nr_sub(unsigned nr)
727	{
728	spin_lock(lock: &aio_nr_lock);
729	if (WARN_ON(aio_nr - nr > aio_nr))
730	aio_nr = 0;
731	else
732	aio_nr -= nr;
733	spin_unlock(lock: &aio_nr_lock);
734	}
735
736	/* ioctx_alloc
737	* Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
738	*/
739	static struct kioctx *ioctx_alloc(unsigned nr_events)
740	{
741	struct mm_struct *mm = current->mm;
742	struct kioctx *ctx;
743	int err = -ENOMEM;
744
745	/*
746	* Store the original nr_events -- what userspace passed to io_setup(),
747	* for counting against the global limit -- before it changes.
748	*/
749	unsigned int max_reqs = nr_events;
750
751	/*
752	* We keep track of the number of available ringbuffer slots, to prevent
753	* overflow (reqs_available), and we also use percpu counters for this.
754	*
755	* So since up to half the slots might be on other cpu's percpu counters
756	* and unavailable, double nr_events so userspace sees what they
757	* expected: additionally, we move req_batch slots to/from percpu
758	* counters at a time, so make sure that isn't 0:
759	*/
760	nr_events = max(nr_events, num_possible_cpus() * 4);
761	nr_events *= 2;
762
763	/* Prevent overflows */
764	if (nr_events > (0x10000000U / sizeof(struct io_event))) {
765	pr_debug("ENOMEM: nr_events too high\n");
766	return ERR_PTR(error: -EINVAL);
767	}
768
769	if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
770	return ERR_PTR(error: -EAGAIN);
771
772	ctx = kmem_cache_zalloc(k: kioctx_cachep, GFP_KERNEL);
773	if (!ctx)
774	return ERR_PTR(error: -ENOMEM);
775
776	ctx->max_reqs = max_reqs;
777
778	spin_lock_init(&ctx->ctx_lock);
779	spin_lock_init(&ctx->completion_lock);
780	mutex_init(&ctx->ring_lock);
781	/* Protect against page migration throughout kiotx setup by keeping
782	* the ring_lock mutex held until setup is complete. */
783	mutex_lock(&ctx->ring_lock);
784	init_waitqueue_head(&ctx->wait);
785
786	INIT_LIST_HEAD(list: &ctx->active_reqs);
787
788	if (percpu_ref_init(ref: &ctx->users, release: free_ioctx_users, flags: 0, GFP_KERNEL))
789	goto err;
790
791	if (percpu_ref_init(ref: &ctx->reqs, release: free_ioctx_reqs, flags: 0, GFP_KERNEL))
792	goto err;
793
794	ctx->cpu = alloc_percpu(struct kioctx_cpu);
795	if (!ctx->cpu)
796	goto err;
797
798	err = aio_setup_ring(ctx, nr_events);
799	if (err < 0)
800	goto err;
801
802	atomic_set(v: &ctx->reqs_available, i: ctx->nr_events - 1);
803	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
804	if (ctx->req_batch < 1)
805	ctx->req_batch = 1;
806
807	/* limit the number of system wide aios */
808	spin_lock(lock: &aio_nr_lock);
809	if (aio_nr + ctx->max_reqs > aio_max_nr ||
810	aio_nr + ctx->max_reqs < aio_nr) {
811	spin_unlock(lock: &aio_nr_lock);
812	err = -EAGAIN;
813	goto err_ctx;
814	}
815	aio_nr += ctx->max_reqs;
816	spin_unlock(lock: &aio_nr_lock);
817
818	percpu_ref_get(ref: &ctx->users); /* io_setup() will drop this ref */
819	percpu_ref_get(ref: &ctx->reqs); /* free_ioctx_users() will drop this */
820
821	err = ioctx_add_table(ctx, mm);
822	if (err)
823	goto err_cleanup;
824
825	/* Release the ring_lock mutex now that all setup is complete. */
826	mutex_unlock(lock: &ctx->ring_lock);
827
828	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
829	ctx, ctx->user_id, mm, ctx->nr_events);
830	return ctx;
831
832	err_cleanup:
833	aio_nr_sub(nr: ctx->max_reqs);
834	err_ctx:
835	atomic_set(v: &ctx->dead, i: 1);
836	if (ctx->mmap_size)
837	vm_munmap(ctx->mmap_base, ctx->mmap_size);
838	aio_free_ring(ctx);
839	err:
840	mutex_unlock(lock: &ctx->ring_lock);
841	free_percpu(pdata: ctx->cpu);
842	percpu_ref_exit(ref: &ctx->reqs);
843	percpu_ref_exit(ref: &ctx->users);
844	kmem_cache_free(s: kioctx_cachep, objp: ctx);
845	pr_debug("error allocating ioctx %d\n", err);
846	return ERR_PTR(error: err);
847	}
848
849	/* kill_ioctx
850	* Cancels all outstanding aio requests on an aio context. Used
851	* when the processes owning a context have all exited to encourage
852	* the rapid destruction of the kioctx.
853	*/
854	static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
855	struct ctx_rq_wait *wait)
856	{
857	struct kioctx_table *table;
858
859	spin_lock(lock: &mm->ioctx_lock);
860	if (atomic_xchg(v: &ctx->dead, new: 1)) {
861	spin_unlock(lock: &mm->ioctx_lock);
862	return -EINVAL;
863	}
864
865	table = rcu_dereference_raw(mm->ioctx_table);
866	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
867	RCU_INIT_POINTER(table->table[ctx->id], NULL);
868	spin_unlock(lock: &mm->ioctx_lock);
869
870	/* free_ioctx_reqs() will do the necessary RCU synchronization */
871	wake_up_all(&ctx->wait);
872
873	/*
874	* It'd be more correct to do this in free_ioctx(), after all
875	* the outstanding kiocbs have finished - but by then io_destroy
876	* has already returned, so io_setup() could potentially return
877	* -EAGAIN with no ioctxs actually in use (as far as userspace
878	* could tell).
879	*/
880	aio_nr_sub(nr: ctx->max_reqs);
881
882	if (ctx->mmap_size)
883	vm_munmap(ctx->mmap_base, ctx->mmap_size);
884
885	ctx->rq_wait = wait;
886	percpu_ref_kill(ref: &ctx->users);
887	return 0;
888	}
889
890	/*
891	* exit_aio: called when the last user of mm goes away. At this point, there is
892	* no way for any new requests to be submited or any of the io_* syscalls to be
893	* called on the context.
894	*
895	* There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
896	* them.
897	*/
898	void exit_aio(struct mm_struct *mm)
899	{
900	struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
901	struct ctx_rq_wait wait;
902	int i, skipped;
903
904	if (!table)
905	return;
906
907	atomic_set(v: &wait.count, i: table->nr);
908	init_completion(x: &wait.comp);
909
910	skipped = 0;
911	for (i = 0; i < table->nr; ++i) {
912	struct kioctx *ctx =
913	rcu_dereference_protected(table->table[i], true);
914
915	if (!ctx) {
916	skipped++;
917	continue;
918	}
919
920	/*
921	* We don't need to bother with munmap() here - exit_mmap(mm)
922	* is coming and it'll unmap everything. And we simply can't,
923	* this is not necessarily our ->mm.
924	* Since kill_ioctx() uses non-zero ->mmap_size as indicator
925	* that it needs to unmap the area, just set it to 0.
926	*/
927	ctx->mmap_size = 0;
928	kill_ioctx(mm, ctx, wait: &wait);
929	}
930
931	if (!atomic_sub_and_test(i: skipped, v: &wait.count)) {
932	/* Wait until all IO for the context are done. */
933	wait_for_completion(&wait.comp);
934	}
935
936	RCU_INIT_POINTER(mm->ioctx_table, NULL);
937	kfree(objp: table);
938	}
939
940	static void put_reqs_available(struct kioctx *ctx, unsigned nr)
941	{
942	struct kioctx_cpu *kcpu;
943	unsigned long flags;
944
945	local_irq_save(flags);
946	kcpu = this_cpu_ptr(ctx->cpu);
947	kcpu->reqs_available += nr;
948
949	while (kcpu->reqs_available >= ctx->req_batch * 2) {
950	kcpu->reqs_available -= ctx->req_batch;
951	atomic_add(i: ctx->req_batch, v: &ctx->reqs_available);
952	}
953
954	local_irq_restore(flags);
955	}
956
957	static bool __get_reqs_available(struct kioctx *ctx)
958	{
959	struct kioctx_cpu *kcpu;
960	bool ret = false;
961	unsigned long flags;
962
963	local_irq_save(flags);
964	kcpu = this_cpu_ptr(ctx->cpu);
965	if (!kcpu->reqs_available) {
966	int avail = atomic_read(v: &ctx->reqs_available);
967
968	do {
969	if (avail < ctx->req_batch)
970	goto out;
971	} while (!atomic_try_cmpxchg(v: &ctx->reqs_available,
972	old: &avail, new: avail - ctx->req_batch));
973
974	kcpu->reqs_available += ctx->req_batch;
975	}
976
977	ret = true;
978	kcpu->reqs_available--;
979	out:
980	local_irq_restore(flags);
981	return ret;
982	}
983
984	/* refill_reqs_available
985	* Updates the reqs_available reference counts used for tracking the
986	* number of free slots in the completion ring. This can be called
987	* from aio_complete() (to optimistically update reqs_available) or
988	* from aio_get_req() (the we're out of events case). It must be
989	* called holding ctx->completion_lock.
990	*/
991	static void refill_reqs_available(struct kioctx *ctx, unsigned head,
992	unsigned tail)
993	{
994	unsigned events_in_ring, completed;
995
996	/* Clamp head since userland can write to it. */
997	head %= ctx->nr_events;
998	if (head <= tail)
999	events_in_ring = tail - head;
1000	else
1001	events_in_ring = ctx->nr_events - (head - tail);
1002
1003	completed = ctx->completed_events;
1004	if (events_in_ring < completed)
1005	completed -= events_in_ring;
1006	else
1007	completed = 0;
1008
1009	if (!completed)
1010	return;
1011
1012	ctx->completed_events -= completed;
1013	put_reqs_available(ctx, nr: completed);
1014	}
1015
1016	/* user_refill_reqs_available
1017	* Called to refill reqs_available when aio_get_req() encounters an
1018	* out of space in the completion ring.
1019	*/
1020	static void user_refill_reqs_available(struct kioctx *ctx)
1021	{
1022	spin_lock_irq(lock: &ctx->completion_lock);
1023	if (ctx->completed_events) {
1024	struct aio_ring *ring;
1025	unsigned head;
1026
1027	/* Access of ring->head may race with aio_read_events_ring()
1028	* here, but that's okay since whether we read the old version
1029	* or the new version, and either will be valid. The important
1030	* part is that head cannot pass tail since we prevent
1031	* aio_complete() from updating tail by holding
1032	* ctx->completion_lock. Even if head is invalid, the check
1033	* against ctx->completed_events below will make sure we do the
1034	* safe/right thing.
1035	*/
1036	ring = page_address(ctx->ring_pages[0]);
1037	head = ring->head;
1038
1039	refill_reqs_available(ctx, head, tail: ctx->tail);
1040	}
1041
1042	spin_unlock_irq(lock: &ctx->completion_lock);
1043	}
1044
1045	static bool get_reqs_available(struct kioctx *ctx)
1046	{
1047	if (__get_reqs_available(ctx))
1048	return true;
1049	user_refill_reqs_available(ctx);
1050	return __get_reqs_available(ctx);
1051	}
1052
1053	/* aio_get_req
1054	* Allocate a slot for an aio request.
1055	* Returns NULL if no requests are free.
1056	*
1057	* The refcount is initialized to 2 - one for the async op completion,
1058	* one for the synchronous code that does this.
1059	*/
1060	static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1061	{
1062	struct aio_kiocb *req;
1063
1064	req = kmem_cache_alloc(cachep: kiocb_cachep, GFP_KERNEL);
1065	if (unlikely(!req))
1066	return NULL;
1067
1068	if (unlikely(!get_reqs_available(ctx))) {
1069	kmem_cache_free(s: kiocb_cachep, objp: req);
1070	return NULL;
1071	}
1072
1073	percpu_ref_get(ref: &ctx->reqs);
1074	req->ki_ctx = ctx;
1075	INIT_LIST_HEAD(list: &req->ki_list);
1076	refcount_set(r: &req->ki_refcnt, n: 2);
1077	req->ki_eventfd = NULL;
1078	return req;
1079	}
1080
1081	static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1082	{
1083	struct aio_ring __user *ring = (void __user *)ctx_id;
1084	struct mm_struct *mm = current->mm;
1085	struct kioctx *ctx, *ret = NULL;
1086	struct kioctx_table *table;
1087	unsigned id;
1088
1089	if (get_user(id, &ring->id))
1090	return NULL;
1091
1092	rcu_read_lock();
1093	table = rcu_dereference(mm->ioctx_table);
1094
1095	if (!table || id >= table->nr)
1096	goto out;
1097
1098	id = array_index_nospec(id, table->nr);
1099	ctx = rcu_dereference(table->table[id]);
1100	if (ctx && ctx->user_id == ctx_id) {
1101	if (percpu_ref_tryget_live(ref: &ctx->users))
1102	ret = ctx;
1103	}
1104	out:
1105	rcu_read_unlock();
1106	return ret;
1107	}
1108
1109	static inline void iocb_destroy(struct aio_kiocb *iocb)
1110	{
1111	if (iocb->ki_eventfd)
1112	eventfd_ctx_put(ctx: iocb->ki_eventfd);
1113	if (iocb->ki_filp)
1114	fput(iocb->ki_filp);
1115	percpu_ref_put(ref: &iocb->ki_ctx->reqs);
1116	kmem_cache_free(s: kiocb_cachep, objp: iocb);
1117	}
1118
1119	struct aio_waiter {
1120	struct wait_queue_entry w;
1121	size_t min_nr;
1122	};
1123
1124	/* aio_complete
1125	* Called when the io request on the given iocb is complete.
1126	*/
1127	static void aio_complete(struct aio_kiocb *iocb)
1128	{
1129	struct kioctx *ctx = iocb->ki_ctx;
1130	struct aio_ring *ring;
1131	struct io_event *ev_page, *event;
1132	unsigned tail, pos, head, avail;
1133	unsigned long flags;
1134
1135	/*
1136	* Add a completion event to the ring buffer. Must be done holding
1137	* ctx->completion_lock to prevent other code from messing with the tail
1138	* pointer since we might be called from irq context.
1139	*/
1140	spin_lock_irqsave(&ctx->completion_lock, flags);
1141
1142	tail = ctx->tail;
1143	pos = tail + AIO_EVENTS_OFFSET;
1144
1145	if (++tail >= ctx->nr_events)
1146	tail = 0;
1147
1148	ev_page = page_address(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1149	event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1150
1151	*event = iocb->ki_res;
1152
1153	flush_dcache_page(page: ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1154
1155	pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1156	(void __user *)(unsigned long)iocb->ki_res.obj,
1157	iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1158
1159	/* after flagging the request as done, we
1160	* must never even look at it again
1161	*/
1162	smp_wmb(); /* make event visible before updating tail */
1163
1164	ctx->tail = tail;
1165
1166	ring = page_address(ctx->ring_pages[0]);
1167	head = ring->head;
1168	ring->tail = tail;
1169	flush_dcache_page(page: ctx->ring_pages[0]);
1170
1171	ctx->completed_events++;
1172	if (ctx->completed_events > 1)
1173	refill_reqs_available(ctx, head, tail);
1174
1175	avail = tail > head
1176	? tail - head
1177	: tail + ctx->nr_events - head;
1178	spin_unlock_irqrestore(lock: &ctx->completion_lock, flags);
1179
1180	pr_debug("added to ring %p at [%u]\n", iocb, tail);
1181
1182	/*
1183	* Check if the user asked us to deliver the result through an
1184	* eventfd. The eventfd_signal() function is safe to be called
1185	* from IRQ context.
1186	*/
1187	if (iocb->ki_eventfd)
1188	eventfd_signal(ctx: iocb->ki_eventfd);
1189
1190	/*
1191	* We have to order our ring_info tail store above and test
1192	* of the wait list below outside the wait lock. This is
1193	* like in wake_up_bit() where clearing a bit has to be
1194	* ordered with the unlocked test.
1195	*/
1196	smp_mb();
1197
1198	if (waitqueue_active(wq_head: &ctx->wait)) {
1199	struct aio_waiter *curr, *next;
1200	unsigned long flags;
1201
1202	spin_lock_irqsave(&ctx->wait.lock, flags);
1203	list_for_each_entry_safe(curr, next, &ctx->wait.head, w.entry)
1204	if (avail >= curr->min_nr) {
1205	wake_up_process(tsk: curr->w.private);
1206	list_del_init_careful(entry: &curr->w.entry);
1207	}
1208	spin_unlock_irqrestore(lock: &ctx->wait.lock, flags);
1209	}
1210	}
1211
1212	static inline void iocb_put(struct aio_kiocb *iocb)
1213	{
1214	if (refcount_dec_and_test(r: &iocb->ki_refcnt)) {
1215	aio_complete(iocb);
1216	iocb_destroy(iocb);
1217	}
1218	}
1219
1220	/* aio_read_events_ring
1221	* Pull an event off of the ioctx's event ring. Returns the number of
1222	* events fetched
1223	*/
1224	static long aio_read_events_ring(struct kioctx *ctx,
1225	struct io_event __user *event, long nr)
1226	{
1227	struct aio_ring *ring;
1228	unsigned head, tail, pos;
1229	long ret = 0;
1230	int copy_ret;
1231
1232	/*
1233	* The mutex can block and wake us up and that will cause
1234	* wait_event_interruptible_hrtimeout() to schedule without sleeping
1235	* and repeat. This should be rare enough that it doesn't cause
1236	* peformance issues. See the comment in read_events() for more detail.
1237	*/
1238	sched_annotate_sleep();
1239	mutex_lock(&ctx->ring_lock);
1240
1241	/* Access to ->ring_pages here is protected by ctx->ring_lock. */
1242	ring = page_address(ctx->ring_pages[0]);
1243	head = ring->head;
1244	tail = ring->tail;
1245
1246	/*
1247	* Ensure that once we've read the current tail pointer, that
1248	* we also see the events that were stored up to the tail.
1249	*/
1250	smp_rmb();
1251
1252	pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1253
1254	if (head == tail)
1255	goto out;
1256
1257	head %= ctx->nr_events;
1258	tail %= ctx->nr_events;
1259
1260	while (ret < nr) {
1261	long avail;
1262	struct io_event *ev;
1263	struct page *page;
1264
1265	avail = (head <= tail ? tail : ctx->nr_events) - head;
1266	if (head == tail)
1267	break;
1268
1269	pos = head + AIO_EVENTS_OFFSET;
1270	page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1271	pos %= AIO_EVENTS_PER_PAGE;
1272
1273	avail = min(avail, nr - ret);
1274	avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1275
1276	ev = page_address(page);
1277	copy_ret = copy_to_user(to: event + ret, from: ev + pos,
1278	n: sizeof(*ev) * avail);
1279
1280	if (unlikely(copy_ret)) {
1281	ret = -EFAULT;
1282	goto out;
1283	}
1284
1285	ret += avail;
1286	head += avail;
1287	head %= ctx->nr_events;
1288	}
1289
1290	ring = page_address(ctx->ring_pages[0]);
1291	ring->head = head;
1292	flush_dcache_page(page: ctx->ring_pages[0]);
1293
1294	pr_debug("%li h%u t%u\n", ret, head, tail);
1295	out:
1296	mutex_unlock(lock: &ctx->ring_lock);
1297
1298	return ret;
1299	}
1300
1301	static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1302	struct io_event __user *event, long *i)
1303	{
1304	long ret = aio_read_events_ring(ctx, event: event + *i, nr: nr - *i);
1305
1306	if (ret > 0)
1307	*i += ret;
1308
1309	if (unlikely(atomic_read(&ctx->dead)))
1310	ret = -EINVAL;
1311
1312	if (!*i)
1313	*i = ret;
1314
1315	return ret < 0 || *i >= min_nr;
1316	}
1317
1318	static long read_events(struct kioctx *ctx, long min_nr, long nr,
1319	struct io_event __user *event,
1320	ktime_t until)
1321	{
1322	struct hrtimer_sleeper t;
1323	struct aio_waiter w;
1324	long ret = 0, ret2 = 0;
1325
1326	/*
1327	* Note that aio_read_events() is being called as the conditional - i.e.
1328	* we're calling it after prepare_to_wait() has set task state to
1329	* TASK_INTERRUPTIBLE.
1330	*
1331	* But aio_read_events() can block, and if it blocks it's going to flip
1332	* the task state back to TASK_RUNNING.
1333	*
1334	* This should be ok, provided it doesn't flip the state back to
1335	* TASK_RUNNING and return 0 too much - that causes us to spin. That
1336	* will only happen if the mutex_lock() call blocks, and we then find
1337	* the ringbuffer empty. So in practice we should be ok, but it's
1338	* something to be aware of when touching this code.
1339	*/
1340	aio_read_events(ctx, min_nr, nr, event, i: &ret);
1341	if (until == 0 || ret < 0 || ret >= min_nr)
1342	return ret;
1343
1344	hrtimer_init_sleeper_on_stack(sl: &t, CLOCK_MONOTONIC, mode: HRTIMER_MODE_REL);
1345	if (until != KTIME_MAX) {
1346	hrtimer_set_expires_range_ns(timer: &t.timer, time: until, current->timer_slack_ns);
1347	hrtimer_sleeper_start_expires(sl: &t, mode: HRTIMER_MODE_REL);
1348	}
1349
1350	init_wait(&w.w);
1351
1352	while (1) {
1353	unsigned long nr_got = ret;
1354
1355	w.min_nr = min_nr - ret;
1356
1357	ret2 = prepare_to_wait_event(wq_head: &ctx->wait, wq_entry: &w.w, TASK_INTERRUPTIBLE);
1358	if (!ret2 && !t.task)
1359	ret2 = -ETIME;
1360
1361	if (aio_read_events(ctx, min_nr, nr, event, i: &ret) || ret2)
1362	break;
1363
1364	if (nr_got == ret)
1365	schedule();
1366	}
1367
1368	finish_wait(wq_head: &ctx->wait, wq_entry: &w.w);
1369	hrtimer_cancel(timer: &t.timer);
1370	destroy_hrtimer_on_stack(timer: &t.timer);
1371
1372	return ret;
1373	}
1374
1375	/* sys_io_setup:
1376	* Create an aio_context capable of receiving at least nr_events.
1377	* ctxp must not point to an aio_context that already exists, and
1378	* must be initialized to 0 prior to the call. On successful
1379	* creation of the aio_context, *ctxp is filled in with the resulting
1380	* handle. May fail with -EINVAL if *ctxp is not initialized,
1381	* if the specified nr_events exceeds internal limits. May fail
1382	* with -EAGAIN if the specified nr_events exceeds the user's limit
1383	* of available events. May fail with -ENOMEM if insufficient kernel
1384	* resources are available. May fail with -EFAULT if an invalid
1385	* pointer is passed for ctxp. Will fail with -ENOSYS if not
1386	* implemented.
1387	*/
1388	SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1389	{
1390	struct kioctx *ioctx = NULL;
1391	unsigned long ctx;
1392	long ret;
1393
1394	ret = get_user(ctx, ctxp);
1395	if (unlikely(ret))
1396	goto out;
1397
1398	ret = -EINVAL;
1399	if (unlikely(ctx || nr_events == 0)) {
1400	pr_debug("EINVAL: ctx %lu nr_events %u\n",
1401	ctx, nr_events);
1402	goto out;
1403	}
1404
1405	ioctx = ioctx_alloc(nr_events);
1406	ret = PTR_ERR(ptr: ioctx);
1407	if (!IS_ERR(ptr: ioctx)) {
1408	ret = put_user(ioctx->user_id, ctxp);
1409	if (ret)
1410	kill_ioctx(current->mm, ctx: ioctx, NULL);
1411	percpu_ref_put(ref: &ioctx->users);
1412	}
1413
1414	out:
1415	return ret;
1416	}
1417
1418	#ifdef CONFIG_COMPAT
1419	COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1420	{
1421	struct kioctx *ioctx = NULL;
1422	unsigned long ctx;
1423	long ret;
1424
1425	ret = get_user(ctx, ctx32p);
1426	if (unlikely(ret))
1427	goto out;
1428
1429	ret = -EINVAL;
1430	if (unlikely(ctx || nr_events == 0)) {
1431	pr_debug("EINVAL: ctx %lu nr_events %u\n",
1432	ctx, nr_events);
1433	goto out;
1434	}
1435
1436	ioctx = ioctx_alloc(nr_events);
1437	ret = PTR_ERR(ptr: ioctx);
1438	if (!IS_ERR(ptr: ioctx)) {
1439	/* truncating is ok because it's a user address */
1440	ret = put_user((u32)ioctx->user_id, ctx32p);
1441	if (ret)
1442	kill_ioctx(current->mm, ctx: ioctx, NULL);
1443	percpu_ref_put(ref: &ioctx->users);
1444	}
1445
1446	out:
1447	return ret;
1448	}
1449	#endif
1450
1451	/* sys_io_destroy:
1452	* Destroy the aio_context specified. May cancel any outstanding
1453	* AIOs and block on completion. Will fail with -ENOSYS if not
1454	* implemented. May fail with -EINVAL if the context pointed to
1455	* is invalid.
1456	*/
1457	SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1458	{
1459	struct kioctx *ioctx = lookup_ioctx(ctx_id: ctx);
1460	if (likely(NULL != ioctx)) {
1461	struct ctx_rq_wait wait;
1462	int ret;
1463
1464	init_completion(x: &wait.comp);
1465	atomic_set(v: &wait.count, i: 1);
1466
1467	/* Pass requests_done to kill_ioctx() where it can be set
1468	* in a thread-safe way. If we try to set it here then we have
1469	* a race condition if two io_destroy() called simultaneously.
1470	*/
1471	ret = kill_ioctx(current->mm, ctx: ioctx, wait: &wait);
1472	percpu_ref_put(ref: &ioctx->users);
1473
1474	/* Wait until all IO for the context are done. Otherwise kernel
1475	* keep using user-space buffers even if user thinks the context
1476	* is destroyed.
1477	*/
1478	if (!ret)
1479	wait_for_completion(&wait.comp);
1480
1481	return ret;
1482	}
1483	pr_debug("EINVAL: invalid context id\n");
1484	return -EINVAL;
1485	}
1486
1487	static void aio_remove_iocb(struct aio_kiocb *iocb)
1488	{
1489	struct kioctx *ctx = iocb->ki_ctx;
1490	unsigned long flags;
1491
1492	spin_lock_irqsave(&ctx->ctx_lock, flags);
1493	list_del(entry: &iocb->ki_list);
1494	spin_unlock_irqrestore(lock: &ctx->ctx_lock, flags);
1495	}
1496
1497	static void aio_complete_rw(struct kiocb *kiocb, long res)
1498	{
1499	struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1500
1501	if (!list_empty_careful(head: &iocb->ki_list))
1502	aio_remove_iocb(iocb);
1503
1504	if (kiocb->ki_flags & IOCB_WRITE) {
1505	struct inode *inode = file_inode(f: kiocb->ki_filp);
1506
1507	if (S_ISREG(inode->i_mode))
1508	kiocb_end_write(iocb: kiocb);
1509	}
1510
1511	iocb->ki_res.res = res;
1512	iocb->ki_res.res2 = 0;
1513	iocb_put(iocb);
1514	}
1515
1516	static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1517	{
1518	int ret;
1519
1520	req->ki_complete = aio_complete_rw;
1521	req->private = NULL;
1522	req->ki_pos = iocb->aio_offset;
1523	req->ki_flags = req->ki_filp->f_iocb_flags | IOCB_AIO_RW;
1524	if (iocb->aio_flags & IOCB_FLAG_RESFD)
1525	req->ki_flags |= IOCB_EVENTFD;
1526	if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1527	/*
1528	* If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1529	* aio_reqprio is interpreted as an I/O scheduling
1530	* class and priority.
1531	*/
1532	ret = ioprio_check_cap(ioprio: iocb->aio_reqprio);
1533	if (ret) {
1534	pr_debug("aio ioprio check cap error: %d\n", ret);
1535	return ret;
1536	}
1537
1538	req->ki_ioprio = iocb->aio_reqprio;
1539	} else
1540	req->ki_ioprio = get_current_ioprio();
1541
1542	ret = kiocb_set_rw_flags(ki: req, flags: iocb->aio_rw_flags);
1543	if (unlikely(ret))
1544	return ret;
1545
1546	req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1547	return 0;
1548	}
1549
1550	static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1551	struct iovec **iovec, bool vectored, bool compat,
1552	struct iov_iter *iter)
1553	{
1554	void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1555	size_t len = iocb->aio_nbytes;
1556
1557	if (!vectored) {
1558	ssize_t ret = import_ubuf(type: rw, buf, len, i: iter);
1559	*iovec = NULL;
1560	return ret;
1561	}
1562
1563	return __import_iovec(type: rw, uvec: buf, nr_segs: len, UIO_FASTIOV, iovp: iovec, i: iter, compat);
1564	}
1565
1566	static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1567	{
1568	switch (ret) {
1569	case -EIOCBQUEUED:
1570	break;
1571	case -ERESTARTSYS:
1572	case -ERESTARTNOINTR:
1573	case -ERESTARTNOHAND:
1574	case -ERESTART_RESTARTBLOCK:
1575	/*
1576	* There's no easy way to restart the syscall since other AIO's
1577	* may be already running. Just fail this IO with EINTR.
1578	*/
1579	ret = -EINTR;
1580	fallthrough;
1581	default:
1582	req->ki_complete(req, ret);
1583	}
1584	}
1585
1586	static int aio_read(struct kiocb *req, const struct iocb *iocb,
1587	bool vectored, bool compat)
1588	{
1589	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1590	struct iov_iter iter;
1591	struct file *file;
1592	int ret;
1593
1594	ret = aio_prep_rw(req, iocb);
1595	if (ret)
1596	return ret;
1597	file = req->ki_filp;
1598	if (unlikely(!(file->f_mode & FMODE_READ)))
1599	return -EBADF;
1600	if (unlikely(!file->f_op->read_iter))
1601	return -EINVAL;
1602
1603	ret = aio_setup_rw(ITER_DEST, iocb, iovec: &iovec, vectored, compat, iter: &iter);
1604	if (ret < 0)
1605	return ret;
1606	ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(i: &iter));
1607	if (!ret)
1608	aio_rw_done(req, ret: call_read_iter(file, kio: req, iter: &iter));
1609	kfree(objp: iovec);
1610	return ret;
1611	}
1612
1613	static int aio_write(struct kiocb *req, const struct iocb *iocb,
1614	bool vectored, bool compat)
1615	{
1616	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1617	struct iov_iter iter;
1618	struct file *file;
1619	int ret;
1620
1621	ret = aio_prep_rw(req, iocb);
1622	if (ret)
1623	return ret;
1624	file = req->ki_filp;
1625
1626	if (unlikely(!(file->f_mode & FMODE_WRITE)))
1627	return -EBADF;
1628	if (unlikely(!file->f_op->write_iter))
1629	return -EINVAL;
1630
1631	ret = aio_setup_rw(ITER_SOURCE, iocb, iovec: &iovec, vectored, compat, iter: &iter);
1632	if (ret < 0)
1633	return ret;
1634	ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(i: &iter));
1635	if (!ret) {
1636	if (S_ISREG(file_inode(file)->i_mode))
1637	kiocb_start_write(iocb: req);
1638	req->ki_flags |= IOCB_WRITE;
1639	aio_rw_done(req, ret: call_write_iter(file, kio: req, iter: &iter));
1640	}
1641	kfree(objp: iovec);
1642	return ret;
1643	}
1644
1645	static void aio_fsync_work(struct work_struct *work)
1646	{
1647	struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1648	const struct cred *old_cred = override_creds(iocb->fsync.creds);
1649
1650	iocb->ki_res.res = vfs_fsync(file: iocb->fsync.file, datasync: iocb->fsync.datasync);
1651	revert_creds(old_cred);
1652	put_cred(cred: iocb->fsync.creds);
1653	iocb_put(iocb);
1654	}
1655
1656	static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1657	bool datasync)
1658	{
1659	if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1660	iocb->aio_rw_flags))
1661	return -EINVAL;
1662
1663	if (unlikely(!req->file->f_op->fsync))
1664	return -EINVAL;
1665
1666	req->creds = prepare_creds();
1667	if (!req->creds)
1668	return -ENOMEM;
1669
1670	req->datasync = datasync;
1671	INIT_WORK(&req->work, aio_fsync_work);
1672	schedule_work(work: &req->work);
1673	return 0;
1674	}
1675
1676	static void aio_poll_put_work(struct work_struct *work)
1677	{
1678	struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1679	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1680
1681	iocb_put(iocb);
1682	}
1683
1684	/*
1685	* Safely lock the waitqueue which the request is on, synchronizing with the
1686	* case where the ->poll() provider decides to free its waitqueue early.
1687	*
1688	* Returns true on success, meaning that req->head->lock was locked, req->wait
1689	* is on req->head, and an RCU read lock was taken. Returns false if the
1690	* request was already removed from its waitqueue (which might no longer exist).
1691	*/
1692	static bool poll_iocb_lock_wq(struct poll_iocb *req)
1693	{
1694	wait_queue_head_t *head;
1695
1696	/*
1697	* While we hold the waitqueue lock and the waitqueue is nonempty,
1698	* wake_up_pollfree() will wait for us. However, taking the waitqueue
1699	* lock in the first place can race with the waitqueue being freed.
1700	*
1701	* We solve this as eventpoll does: by taking advantage of the fact that
1702	* all users of wake_up_pollfree() will RCU-delay the actual free. If
1703	* we enter rcu_read_lock() and see that the pointer to the queue is
1704	* non-NULL, we can then lock it without the memory being freed out from
1705	* under us, then check whether the request is still on the queue.
1706	*
1707	* Keep holding rcu_read_lock() as long as we hold the queue lock, in
1708	* case the caller deletes the entry from the queue, leaving it empty.
1709	* In that case, only RCU prevents the queue memory from being freed.
1710	*/
1711	rcu_read_lock();
1712	head = smp_load_acquire(&req->head);
1713	if (head) {
1714	spin_lock(lock: &head->lock);
1715	if (!list_empty(head: &req->wait.entry))
1716	return true;
1717	spin_unlock(lock: &head->lock);
1718	}
1719	rcu_read_unlock();
1720	return false;
1721	}
1722
1723	static void poll_iocb_unlock_wq(struct poll_iocb *req)
1724	{
1725	spin_unlock(lock: &req->head->lock);
1726	rcu_read_unlock();
1727	}
1728
1729	static void aio_poll_complete_work(struct work_struct *work)
1730	{
1731	struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1732	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1733	struct poll_table_struct pt = { ._key = req->events };
1734	struct kioctx *ctx = iocb->ki_ctx;
1735	__poll_t mask = 0;
1736
1737	if (!READ_ONCE(req->cancelled))
1738	mask = vfs_poll(file: req->file, pt: &pt) & req->events;
1739
1740	/*
1741	* Note that ->ki_cancel callers also delete iocb from active_reqs after
1742	* calling ->ki_cancel. We need the ctx_lock roundtrip here to
1743	* synchronize with them. In the cancellation case the list_del_init
1744	* itself is not actually needed, but harmless so we keep it in to
1745	* avoid further branches in the fast path.
1746	*/
1747	spin_lock_irq(lock: &ctx->ctx_lock);
1748	if (poll_iocb_lock_wq(req)) {
1749	if (!mask && !READ_ONCE(req->cancelled)) {
1750	/*
1751	* The request isn't actually ready to be completed yet.
1752	* Reschedule completion if another wakeup came in.
1753	*/
1754	if (req->work_need_resched) {
1755	schedule_work(work: &req->work);
1756	req->work_need_resched = false;
1757	} else {
1758	req->work_scheduled = false;
1759	}
1760	poll_iocb_unlock_wq(req);
1761	spin_unlock_irq(lock: &ctx->ctx_lock);
1762	return;
1763	}
1764	list_del_init(entry: &req->wait.entry);
1765	poll_iocb_unlock_wq(req);
1766	} /* else, POLLFREE has freed the waitqueue, so we must complete */
1767	list_del_init(entry: &iocb->ki_list);
1768	iocb->ki_res.res = mangle_poll(val: mask);
1769	spin_unlock_irq(lock: &ctx->ctx_lock);
1770
1771	iocb_put(iocb);
1772	}
1773
1774	/* assumes we are called with irqs disabled */
1775	static int aio_poll_cancel(struct kiocb *iocb)
1776	{
1777	struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1778	struct poll_iocb *req = &aiocb->poll;
1779
1780	if (poll_iocb_lock_wq(req)) {
1781	WRITE_ONCE(req->cancelled, true);
1782	if (!req->work_scheduled) {
1783	schedule_work(work: &aiocb->poll.work);
1784	req->work_scheduled = true;
1785	}
1786	poll_iocb_unlock_wq(req);
1787	} /* else, the request was force-cancelled by POLLFREE already */
1788
1789	return 0;
1790	}
1791
1792	static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1793	void *key)
1794	{
1795	struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1796	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1797	__poll_t mask = key_to_poll(key);
1798	unsigned long flags;
1799
1800	/* for instances that support it check for an event match first: */
1801	if (mask && !(mask & req->events))
1802	return 0;
1803
1804	/*
1805	* Complete the request inline if possible. This requires that three
1806	* conditions be met:
1807	* 1. An event mask must have been passed. If a plain wakeup was done
1808	* instead, then mask == 0 and we have to call vfs_poll() to get
1809	* the events, so inline completion isn't possible.
1810	* 2. The completion work must not have already been scheduled.
1811	* 3. ctx_lock must not be busy. We have to use trylock because we
1812	* already hold the waitqueue lock, so this inverts the normal
1813	* locking order. Use irqsave/irqrestore because not all
1814	* filesystems (e.g. fuse) call this function with IRQs disabled,
1815	* yet IRQs have to be disabled before ctx_lock is obtained.
1816	*/
1817	if (mask && !req->work_scheduled &&
1818	spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1819	struct kioctx *ctx = iocb->ki_ctx;
1820
1821	list_del_init(entry: &req->wait.entry);
1822	list_del(entry: &iocb->ki_list);
1823	iocb->ki_res.res = mangle_poll(val: mask);
1824	if (iocb->ki_eventfd && !eventfd_signal_allowed()) {
1825	iocb = NULL;
1826	INIT_WORK(&req->work, aio_poll_put_work);
1827	schedule_work(work: &req->work);
1828	}
1829	spin_unlock_irqrestore(lock: &ctx->ctx_lock, flags);
1830	if (iocb)
1831	iocb_put(iocb);
1832	} else {
1833	/*
1834	* Schedule the completion work if needed. If it was already
1835	* scheduled, record that another wakeup came in.
1836	*
1837	* Don't remove the request from the waitqueue here, as it might
1838	* not actually be complete yet (we won't know until vfs_poll()
1839	* is called), and we must not miss any wakeups. POLLFREE is an
1840	* exception to this; see below.
1841	*/
1842	if (req->work_scheduled) {
1843	req->work_need_resched = true;
1844	} else {
1845	schedule_work(work: &req->work);
1846	req->work_scheduled = true;
1847	}
1848
1849	/*
1850	* If the waitqueue is being freed early but we can't complete
1851	* the request inline, we have to tear down the request as best
1852	* we can. That means immediately removing the request from its
1853	* waitqueue and preventing all further accesses to the
1854	* waitqueue via the request. We also need to schedule the
1855	* completion work (done above). Also mark the request as
1856	* cancelled, to potentially skip an unneeded call to ->poll().
1857	*/
1858	if (mask & POLLFREE) {
1859	WRITE_ONCE(req->cancelled, true);
1860	list_del_init(entry: &req->wait.entry);
1861
1862	/*
1863	* Careful: this *must* be the last step, since as soon
1864	* as req->head is NULL'ed out, the request can be
1865	* completed and freed, since aio_poll_complete_work()
1866	* will no longer need to take the waitqueue lock.
1867	*/
1868	smp_store_release(&req->head, NULL);
1869	}
1870	}
1871	return 1;
1872	}
1873
1874	struct aio_poll_table {
1875	struct poll_table_struct pt;
1876	struct aio_kiocb *iocb;
1877	bool queued;
1878	int error;
1879	};
1880
1881	static void
1882	aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1883	struct poll_table_struct *p)
1884	{
1885	struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1886
1887	/* multiple wait queues per file are not supported */
1888	if (unlikely(pt->queued)) {
1889	pt->error = -EINVAL;
1890	return;
1891	}
1892
1893	pt->queued = true;
1894	pt->error = 0;
1895	pt->iocb->poll.head = head;
1896	add_wait_queue(wq_head: head, wq_entry: &pt->iocb->poll.wait);
1897	}
1898
1899	static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1900	{
1901	struct kioctx *ctx = aiocb->ki_ctx;
1902	struct poll_iocb *req = &aiocb->poll;
1903	struct aio_poll_table apt;
1904	bool cancel = false;
1905	__poll_t mask;
1906
1907	/* reject any unknown events outside the normal event mask. */
1908	if ((u16)iocb->aio_buf != iocb->aio_buf)
1909	return -EINVAL;
1910	/* reject fields that are not defined for poll */
1911	if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1912	return -EINVAL;
1913
1914	INIT_WORK(&req->work, aio_poll_complete_work);
1915	req->events = demangle_poll(val: iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1916
1917	req->head = NULL;
1918	req->cancelled = false;
1919	req->work_scheduled = false;
1920	req->work_need_resched = false;
1921
1922	apt.pt._qproc = aio_poll_queue_proc;
1923	apt.pt._key = req->events;
1924	apt.iocb = aiocb;
1925	apt.queued = false;
1926	apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1927
1928	/* initialized the list so that we can do list_empty checks */
1929	INIT_LIST_HEAD(list: &req->wait.entry);
1930	init_waitqueue_func_entry(wq_entry: &req->wait, func: aio_poll_wake);
1931
1932	mask = vfs_poll(file: req->file, pt: &apt.pt) & req->events;
1933	spin_lock_irq(lock: &ctx->ctx_lock);
1934	if (likely(apt.queued)) {
1935	bool on_queue = poll_iocb_lock_wq(req);
1936
1937	if (!on_queue || req->work_scheduled) {
1938	/*
1939	* aio_poll_wake() already either scheduled the async
1940	* completion work, or completed the request inline.
1941	*/
1942	if (apt.error) /* unsupported case: multiple queues */
1943	cancel = true;
1944	apt.error = 0;
1945	mask = 0;
1946	}
1947	if (mask || apt.error) {
1948	/* Steal to complete synchronously. */
1949	list_del_init(entry: &req->wait.entry);
1950	} else if (cancel) {
1951	/* Cancel if possible (may be too late though). */
1952	WRITE_ONCE(req->cancelled, true);
1953	} else if (on_queue) {
1954	/*
1955	* Actually waiting for an event, so add the request to
1956	* active_reqs so that it can be cancelled if needed.
1957	*/
1958	list_add_tail(new: &aiocb->ki_list, head: &ctx->active_reqs);
1959	aiocb->ki_cancel = aio_poll_cancel;
1960	}
1961	if (on_queue)
1962	poll_iocb_unlock_wq(req);
1963	}
1964	if (mask) { /* no async, we'd stolen it */
1965	aiocb->ki_res.res = mangle_poll(val: mask);
1966	apt.error = 0;
1967	}
1968	spin_unlock_irq(lock: &ctx->ctx_lock);
1969	if (mask)
1970	iocb_put(iocb: aiocb);
1971	return apt.error;
1972	}
1973
1974	static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1975	struct iocb __user *user_iocb, struct aio_kiocb *req,
1976	bool compat)
1977	{
1978	req->ki_filp = fget(fd: iocb->aio_fildes);
1979	if (unlikely(!req->ki_filp))
1980	return -EBADF;
1981
1982	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1983	struct eventfd_ctx *eventfd;
1984	/*
1985	* If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1986	* instance of the file* now. The file descriptor must be
1987	* an eventfd() fd, and will be signaled for each completed
1988	* event using the eventfd_signal() function.
1989	*/
1990	eventfd = eventfd_ctx_fdget(fd: iocb->aio_resfd);
1991	if (IS_ERR(ptr: eventfd))
1992	return PTR_ERR(ptr: eventfd);
1993
1994	req->ki_eventfd = eventfd;
1995	}
1996
1997	if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1998	pr_debug("EFAULT: aio_key\n");
1999	return -EFAULT;
2000	}
2001
2002	req->ki_res.obj = (u64)(unsigned long)user_iocb;
2003	req->ki_res.data = iocb->aio_data;
2004	req->ki_res.res = 0;
2005	req->ki_res.res2 = 0;
2006
2007	switch (iocb->aio_lio_opcode) {
2008	case IOCB_CMD_PREAD:
2009	return aio_read(req: &req->rw, iocb, vectored: false, compat);
2010	case IOCB_CMD_PWRITE:
2011	return aio_write(req: &req->rw, iocb, vectored: false, compat);
2012	case IOCB_CMD_PREADV:
2013	return aio_read(req: &req->rw, iocb, vectored: true, compat);
2014	case IOCB_CMD_PWRITEV:
2015	return aio_write(req: &req->rw, iocb, vectored: true, compat);
2016	case IOCB_CMD_FSYNC:
2017	return aio_fsync(req: &req->fsync, iocb, datasync: false);
2018	case IOCB_CMD_FDSYNC:
2019	return aio_fsync(req: &req->fsync, iocb, datasync: true);
2020	case IOCB_CMD_POLL:
2021	return aio_poll(aiocb: req, iocb);
2022	default:
2023	pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
2024	return -EINVAL;
2025	}
2026	}
2027
2028	static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
2029	bool compat)
2030	{
2031	struct aio_kiocb *req;
2032	struct iocb iocb;
2033	int err;
2034
2035	if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
2036	return -EFAULT;
2037
2038	/* enforce forwards compatibility on users */
2039	if (unlikely(iocb.aio_reserved2)) {
2040	pr_debug("EINVAL: reserve field set\n");
2041	return -EINVAL;
2042	}
2043
2044	/* prevent overflows */
2045	if (unlikely(
2046	(iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
2047	(iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
2048	((ssize_t)iocb.aio_nbytes < 0)
2049	)) {
2050	pr_debug("EINVAL: overflow check\n");
2051	return -EINVAL;
2052	}
2053
2054	req = aio_get_req(ctx);
2055	if (unlikely(!req))
2056	return -EAGAIN;
2057
2058	err = __io_submit_one(ctx, iocb: &iocb, user_iocb, req, compat);
2059
2060	/* Done with the synchronous reference */
2061	iocb_put(iocb: req);
2062
2063	/*
2064	* If err is 0, we'd either done aio_complete() ourselves or have
2065	* arranged for that to be done asynchronously. Anything non-zero
2066	* means that we need to destroy req ourselves.
2067	*/
2068	if (unlikely(err)) {
2069	iocb_destroy(iocb: req);
2070	put_reqs_available(ctx, nr: 1);
2071	}
2072	return err;
2073	}
2074
2075	/* sys_io_submit:
2076	* Queue the nr iocbs pointed to by iocbpp for processing. Returns
2077	* the number of iocbs queued. May return -EINVAL if the aio_context
2078	* specified by ctx_id is invalid, if nr is < 0, if the iocb at
2079	* *iocbpp[0] is not properly initialized, if the operation specified
2080	* is invalid for the file descriptor in the iocb. May fail with
2081	* -EFAULT if any of the data structures point to invalid data. May
2082	* fail with -EBADF if the file descriptor specified in the first
2083	* iocb is invalid. May fail with -EAGAIN if insufficient resources
2084	* are available to queue any iocbs. Will return 0 if nr is 0. Will
2085	* fail with -ENOSYS if not implemented.
2086	*/
2087	SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2088	struct iocb __user * __user *, iocbpp)
2089	{
2090	struct kioctx *ctx;
2091	long ret = 0;
2092	int i = 0;
2093	struct blk_plug plug;
2094
2095	if (unlikely(nr < 0))
2096	return -EINVAL;
2097
2098	ctx = lookup_ioctx(ctx_id);
2099	if (unlikely(!ctx)) {
2100	pr_debug("EINVAL: invalid context id\n");
2101	return -EINVAL;
2102	}
2103
2104	if (nr > ctx->nr_events)
2105	nr = ctx->nr_events;
2106
2107	if (nr > AIO_PLUG_THRESHOLD)
2108	blk_start_plug(&plug);
2109	for (i = 0; i < nr; i++) {
2110	struct iocb __user *user_iocb;
2111
2112	if (unlikely(get_user(user_iocb, iocbpp + i))) {
2113	ret = -EFAULT;
2114	break;
2115	}
2116
2117	ret = io_submit_one(ctx, user_iocb, compat: false);
2118	if (ret)
2119	break;
2120	}
2121	if (nr > AIO_PLUG_THRESHOLD)
2122	blk_finish_plug(&plug);
2123
2124	percpu_ref_put(ref: &ctx->users);
2125	return i ? i : ret;
2126	}
2127
2128	#ifdef CONFIG_COMPAT
2129	COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2130	int, nr, compat_uptr_t __user *, iocbpp)
2131	{
2132	struct kioctx *ctx;
2133	long ret = 0;
2134	int i = 0;
2135	struct blk_plug plug;
2136
2137	if (unlikely(nr < 0))
2138	return -EINVAL;
2139
2140	ctx = lookup_ioctx(ctx_id);
2141	if (unlikely(!ctx)) {
2142	pr_debug("EINVAL: invalid context id\n");
2143	return -EINVAL;
2144	}
2145
2146	if (nr > ctx->nr_events)
2147	nr = ctx->nr_events;
2148
2149	if (nr > AIO_PLUG_THRESHOLD)
2150	blk_start_plug(&plug);
2151	for (i = 0; i < nr; i++) {
2152	compat_uptr_t user_iocb;
2153
2154	if (unlikely(get_user(user_iocb, iocbpp + i))) {
2155	ret = -EFAULT;
2156	break;
2157	}
2158
2159	ret = io_submit_one(ctx, user_iocb: compat_ptr(uptr: user_iocb), compat: true);
2160	if (ret)
2161	break;
2162	}
2163	if (nr > AIO_PLUG_THRESHOLD)
2164	blk_finish_plug(&plug);
2165
2166	percpu_ref_put(ref: &ctx->users);
2167	return i ? i : ret;
2168	}
2169	#endif
2170
2171	/* sys_io_cancel:
2172	* Attempts to cancel an iocb previously passed to io_submit. If
2173	* the operation is successfully cancelled, the resulting event is
2174	* copied into the memory pointed to by result without being placed
2175	* into the completion queue and 0 is returned. May fail with
2176	* -EFAULT if any of the data structures pointed to are invalid.
2177	* May fail with -EINVAL if aio_context specified by ctx_id is
2178	* invalid. May fail with -EAGAIN if the iocb specified was not
2179	* cancelled. Will fail with -ENOSYS if not implemented.
2180	*/
2181	SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2182	struct io_event __user *, result)
2183	{
2184	struct kioctx *ctx;
2185	struct aio_kiocb *kiocb;
2186	int ret = -EINVAL;
2187	u32 key;
2188	u64 obj = (u64)(unsigned long)iocb;
2189
2190	if (unlikely(get_user(key, &iocb->aio_key)))
2191	return -EFAULT;
2192	if (unlikely(key != KIOCB_KEY))
2193	return -EINVAL;
2194
2195	ctx = lookup_ioctx(ctx_id);
2196	if (unlikely(!ctx))
2197	return -EINVAL;
2198
2199	spin_lock_irq(lock: &ctx->ctx_lock);
2200	/* TODO: use a hash or array, this sucks. */
2201	list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2202	if (kiocb->ki_res.obj == obj) {
2203	ret = kiocb->ki_cancel(&kiocb->rw);
2204	list_del_init(entry: &kiocb->ki_list);
2205	break;
2206	}
2207	}
2208	spin_unlock_irq(lock: &ctx->ctx_lock);
2209
2210	if (!ret) {
2211	/*
2212	* The result argument is no longer used - the io_event is
2213	* always delivered via the ring buffer. -EINPROGRESS indicates
2214	* cancellation is progress:
2215	*/
2216	ret = -EINPROGRESS;
2217	}
2218
2219	percpu_ref_put(ref: &ctx->users);
2220
2221	return ret;
2222	}
2223
2224	static long do_io_getevents(aio_context_t ctx_id,
2225	long min_nr,
2226	long nr,
2227	struct io_event __user *events,
2228	struct timespec64 *ts)
2229	{
2230	ktime_t until = ts ? timespec64_to_ktime(ts: *ts) : KTIME_MAX;
2231	struct kioctx *ioctx = lookup_ioctx(ctx_id);
2232	long ret = -EINVAL;
2233
2234	if (likely(ioctx)) {
2235	if (likely(min_nr <= nr && min_nr >= 0))
2236	ret = read_events(ctx: ioctx, min_nr, nr, event: events, until);
2237	percpu_ref_put(ref: &ioctx->users);
2238	}
2239
2240	return ret;
2241	}
2242
2243	/* io_getevents:
2244	* Attempts to read at least min_nr events and up to nr events from
2245	* the completion queue for the aio_context specified by ctx_id. If
2246	* it succeeds, the number of read events is returned. May fail with
2247	* -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2248	* out of range, if timeout is out of range. May fail with -EFAULT
2249	* if any of the memory specified is invalid. May return 0 or
2250	* < min_nr if the timeout specified by timeout has elapsed
2251	* before sufficient events are available, where timeout == NULL
2252	* specifies an infinite timeout. Note that the timeout pointed to by
2253	* timeout is relative. Will fail with -ENOSYS if not implemented.
2254	*/
2255	#ifdef CONFIG_64BIT
2256
2257	SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2258	long, min_nr,
2259	long, nr,
2260	struct io_event __user *, events,
2261	struct __kernel_timespec __user *, timeout)
2262	{
2263	struct timespec64 ts;
2264	int ret;
2265
2266	if (timeout && unlikely(get_timespec64(&ts, timeout)))
2267	return -EFAULT;
2268
2269	ret = do_io_getevents(ctx_id, min_nr, nr, events, ts: timeout ? &ts : NULL);
2270	if (!ret && signal_pending(current))
2271	ret = -EINTR;
2272	return ret;
2273	}
2274
2275	#endif
2276
2277	struct __aio_sigset {
2278	const sigset_t __user *sigmask;
2279	size_t sigsetsize;
2280	};
2281
2282	SYSCALL_DEFINE6(io_pgetevents,
2283	aio_context_t, ctx_id,
2284	long, min_nr,
2285	long, nr,
2286	struct io_event __user *, events,
2287	struct __kernel_timespec __user *, timeout,
2288	const struct __aio_sigset __user *, usig)
2289	{
2290	struct __aio_sigset ksig = { NULL, };
2291	struct timespec64 ts;
2292	bool interrupted;
2293	int ret;
2294
2295	if (timeout && unlikely(get_timespec64(&ts, timeout)))
2296	return -EFAULT;
2297
2298	if (usig && copy_from_user(to: &ksig, from: usig, n: sizeof(ksig)))
2299	return -EFAULT;
2300
2301	ret = set_user_sigmask(umask: ksig.sigmask, sigsetsize: ksig.sigsetsize);
2302	if (ret)
2303	return ret;
2304
2305	ret = do_io_getevents(ctx_id, min_nr, nr, events, ts: timeout ? &ts : NULL);
2306
2307	interrupted = signal_pending(current);
2308	restore_saved_sigmask_unless(interrupted);
2309	if (interrupted && !ret)
2310	ret = -ERESTARTNOHAND;
2311
2312	return ret;
2313	}
2314
2315	#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2316
2317	SYSCALL_DEFINE6(io_pgetevents_time32,
2318	aio_context_t, ctx_id,
2319	long, min_nr,
2320	long, nr,
2321	struct io_event __user *, events,
2322	struct old_timespec32 __user *, timeout,
2323	const struct __aio_sigset __user *, usig)
2324	{
2325	struct __aio_sigset ksig = { NULL, };
2326	struct timespec64 ts;
2327	bool interrupted;
2328	int ret;
2329
2330	if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2331	return -EFAULT;
2332
2333	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2334	return -EFAULT;
2335
2336
2337	ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2338	if (ret)
2339	return ret;
2340
2341	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2342
2343	interrupted = signal_pending(current);
2344	restore_saved_sigmask_unless(interrupted);
2345	if (interrupted && !ret)
2346	ret = -ERESTARTNOHAND;
2347
2348	return ret;
2349	}
2350
2351	#endif
2352
2353	#if defined(CONFIG_COMPAT_32BIT_TIME)
2354
2355	SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2356	__s32, min_nr,
2357	__s32, nr,
2358	struct io_event __user *, events,
2359	struct old_timespec32 __user *, timeout)
2360	{
2361	struct timespec64 t;
2362	int ret;
2363
2364	if (timeout && get_old_timespec32(&t, timeout))
2365	return -EFAULT;
2366
2367	ret = do_io_getevents(ctx_id, min_nr, nr, events, ts: timeout ? &t : NULL);
2368	if (!ret && signal_pending(current))
2369	ret = -EINTR;
2370	return ret;
2371	}
2372
2373	#endif
2374
2375	#ifdef CONFIG_COMPAT
2376
2377	struct __compat_aio_sigset {
2378	compat_uptr_t sigmask;
2379	compat_size_t sigsetsize;
2380	};
2381
2382	#if defined(CONFIG_COMPAT_32BIT_TIME)
2383
2384	COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2385	compat_aio_context_t, ctx_id,
2386	compat_long_t, min_nr,
2387	compat_long_t, nr,
2388	struct io_event __user *, events,
2389	struct old_timespec32 __user *, timeout,
2390	const struct __compat_aio_sigset __user *, usig)
2391	{
2392	struct __compat_aio_sigset ksig = { 0, };
2393	struct timespec64 t;
2394	bool interrupted;
2395	int ret;
2396
2397	if (timeout && get_old_timespec32(&t, timeout))
2398	return -EFAULT;
2399
2400	if (usig && copy_from_user(to: &ksig, from: usig, n: sizeof(ksig)))
2401	return -EFAULT;
2402
2403	ret = set_compat_user_sigmask(umask: compat_ptr(uptr: ksig.sigmask), sigsetsize: ksig.sigsetsize);
2404	if (ret)
2405	return ret;
2406
2407	ret = do_io_getevents(ctx_id, min_nr, nr, events, ts: timeout ? &t : NULL);
2408
2409	interrupted = signal_pending(current);
2410	restore_saved_sigmask_unless(interrupted);
2411	if (interrupted && !ret)
2412	ret = -ERESTARTNOHAND;
2413
2414	return ret;
2415	}
2416
2417	#endif
2418
2419	COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2420	compat_aio_context_t, ctx_id,
2421	compat_long_t, min_nr,
2422	compat_long_t, nr,
2423	struct io_event __user *, events,
2424	struct __kernel_timespec __user *, timeout,
2425	const struct __compat_aio_sigset __user *, usig)
2426	{
2427	struct __compat_aio_sigset ksig = { 0, };
2428	struct timespec64 t;
2429	bool interrupted;
2430	int ret;
2431
2432	if (timeout && get_timespec64(ts: &t, uts: timeout))
2433	return -EFAULT;
2434
2435	if (usig && copy_from_user(to: &ksig, from: usig, n: sizeof(ksig)))
2436	return -EFAULT;
2437
2438	ret = set_compat_user_sigmask(umask: compat_ptr(uptr: ksig.sigmask), sigsetsize: ksig.sigsetsize);
2439	if (ret)
2440	return ret;
2441
2442	ret = do_io_getevents(ctx_id, min_nr, nr, events, ts: timeout ? &t : NULL);
2443
2444	interrupted = signal_pending(current);
2445	restore_saved_sigmask_unless(interrupted);
2446	if (interrupted && !ret)
2447	ret = -ERESTARTNOHAND;
2448
2449	return ret;
2450	}
2451	#endif
2452