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[] __counted_by(nr);
  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 kioctx_cpu __percpu *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 folio		**ring_folios;
 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 folio		*internal_folios[AIO_RING_PAGES];
 164	struct file		*aio_ring_file;
 165
 166	unsigned		id;
 167};
 168
 169/*
 170 * First field must be the file pointer in all the
 171 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
 172 */
 173struct fsync_iocb {
 174	struct file		*file;
 175	struct work_struct	work;
 176	bool			datasync;
 177	struct cred		*creds;
 178};
 179
 180struct poll_iocb {
 181	struct file		*file;
 182	struct wait_queue_head	*head;
 183	__poll_t		events;
 184	bool			cancelled;
 185	bool			work_scheduled;
 186	bool			work_need_resched;
 187	struct wait_queue_entry	wait;
 188	struct work_struct	work;
 189};
 190
 191/*
 192 * NOTE! Each of the iocb union members has the file pointer
 193 * as the first entry in their struct definition. So you can
 194 * access the file pointer through any of the sub-structs,
 195 * or directly as just 'ki_filp' in this struct.
 196 */
 197struct aio_kiocb {
 198	union {
 199		struct file		*ki_filp;
 200		struct kiocb		rw;
 201		struct fsync_iocb	fsync;
 202		struct poll_iocb	poll;
 203	};
 204
 205	struct kioctx		*ki_ctx;
 206	kiocb_cancel_fn		*ki_cancel;
 207
 208	struct io_event		ki_res;
 209
 210	struct list_head	ki_list;	/* the aio core uses this
 211						 * for cancellation */
 212	refcount_t		ki_refcnt;
 213
 214	/*
 215	 * If the aio_resfd field of the userspace iocb is not zero,
 216	 * this is the underlying eventfd context to deliver events to.
 217	 */
 218	struct eventfd_ctx	*ki_eventfd;
 219};
 220
 221/*------ sysctl variables----*/
 222static DEFINE_SPINLOCK(aio_nr_lock);
 223static unsigned long aio_nr;		/* current system wide number of aio requests */
 224static unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
 225/*----end sysctl variables---*/
 226#ifdef CONFIG_SYSCTL
 227static struct ctl_table aio_sysctls[] = {
 228	{
 229		.procname	= "aio-nr",
 230		.data		= &aio_nr,
 231		.maxlen		= sizeof(aio_nr),
 232		.mode		= 0444,
 233		.proc_handler	= proc_doulongvec_minmax,
 234	},
 235	{
 236		.procname	= "aio-max-nr",
 237		.data		= &aio_max_nr,
 238		.maxlen		= sizeof(aio_max_nr),
 239		.mode		= 0644,
 240		.proc_handler	= proc_doulongvec_minmax,
 241	},
 242};
 243
 244static 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
 252static struct kmem_cache	*kiocb_cachep;
 253static struct kmem_cache	*kioctx_cachep;
 254
 255static struct vfsmount *aio_mnt;
 256
 257static const struct file_operations aio_ring_fops;
 258static const struct address_space_operations aio_ctx_aops;
 259
 260static 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(inode))
 265		return ERR_CAST(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(file))
 274		iput(inode);
 275	return file;
 276}
 277
 278static 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 */
 290static 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(aio_mnt))
 299		panic("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
 308static 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(file_inode(aio_ring_file), 0);
 315
 316		/* Prevent further access to the kioctx from migratepages */
 317		i_mapping = aio_ring_file->f_mapping;
 318		spin_lock(&i_mapping->i_private_lock);
 319		i_mapping->i_private_data = NULL;
 320		ctx->aio_ring_file = NULL;
 321		spin_unlock(&i_mapping->i_private_lock);
 322
 323		fput(aio_ring_file);
 324	}
 325}
 326
 327static 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 folio *folio = ctx->ring_folios[i];
 338
 339		if (!folio)
 
 
 340			continue;
 341
 342		pr_debug("pid(%d) [%d] folio->count=%d\n", current->pid, i,
 343			 folio_ref_count(folio));
 344		ctx->ring_folios[i] = NULL;
 345		folio_put(folio);
 346	}
 347
 348	if (ctx->ring_folios && ctx->ring_folios != ctx->internal_folios) {
 349		kfree(ctx->ring_folios);
 350		ctx->ring_folios = NULL;
 351	}
 352}
 353
 354static int aio_ring_mremap(struct vm_area_struct *vma)
 355{
 356	struct file *file = vma->vm_file;
 357	struct mm_struct *mm = vma->vm_mm;
 358	struct kioctx_table *table;
 359	int i, res = -EINVAL;
 360
 361	spin_lock(&mm->ioctx_lock);
 362	rcu_read_lock();
 363	table = rcu_dereference(mm->ioctx_table);
 364	if (!table)
 365		goto out_unlock;
 366
 367	for (i = 0; i < table->nr; i++) {
 368		struct kioctx *ctx;
 369
 370		ctx = rcu_dereference(table->table[i]);
 371		if (ctx && ctx->aio_ring_file == file) {
 372			if (!atomic_read(&ctx->dead)) {
 373				ctx->user_id = ctx->mmap_base = vma->vm_start;
 374				res = 0;
 375			}
 376			break;
 377		}
 378	}
 379
 380out_unlock:
 381	rcu_read_unlock();
 382	spin_unlock(&mm->ioctx_lock);
 383	return res;
 384}
 385
 386static const struct vm_operations_struct aio_ring_vm_ops = {
 387	.mremap		= aio_ring_mremap,
 388#if IS_ENABLED(CONFIG_MMU)
 389	.fault		= filemap_fault,
 390	.map_pages	= filemap_map_pages,
 391	.page_mkwrite	= filemap_page_mkwrite,
 392#endif
 393};
 394
 395static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
 396{
 397	vm_flags_set(vma, VM_DONTEXPAND);
 398	vma->vm_ops = &aio_ring_vm_ops;
 399	return 0;
 400}
 401
 402static const struct file_operations aio_ring_fops = {
 403	.mmap = aio_ring_mmap,
 404};
 405
 406#if IS_ENABLED(CONFIG_MIGRATION)
 407static int aio_migrate_folio(struct address_space *mapping, struct folio *dst,
 408			struct folio *src, enum migrate_mode mode)
 409{
 410	struct kioctx *ctx;
 411	unsigned long flags;
 412	pgoff_t idx;
 413	int rc = 0;
 
 
 
 
 
 
 
 
 
 
 414
 415	/* mapping->i_private_lock here protects against the kioctx teardown.  */
 416	spin_lock(&mapping->i_private_lock);
 417	ctx = mapping->i_private_data;
 418	if (!ctx) {
 419		rc = -EINVAL;
 420		goto out;
 421	}
 422
 423	/* The ring_lock mutex.  The prevents aio_read_events() from writing
 424	 * to the ring's head, and prevents page migration from mucking in
 425	 * a partially initialized kiotx.
 426	 */
 427	if (!mutex_trylock(&ctx->ring_lock)) {
 428		rc = -EAGAIN;
 429		goto out;
 430	}
 431
 432	idx = src->index;
 433	if (idx < (pgoff_t)ctx->nr_pages) {
 434		/* Make sure the old folio hasn't already been changed */
 435		if (ctx->ring_folios[idx] != src)
 436			rc = -EAGAIN;
 437	} else
 438		rc = -EINVAL;
 439
 440	if (rc != 0)
 441		goto out_unlock;
 442
 443	/* Writeback must be complete */
 444	BUG_ON(folio_test_writeback(src));
 445	folio_get(dst);
 446
 447	rc = folio_migrate_mapping(mapping, dst, src, 1);
 448	if (rc != MIGRATEPAGE_SUCCESS) {
 449		folio_put(dst);
 450		goto out_unlock;
 451	}
 452
 453	/* Take completion_lock to prevent other writes to the ring buffer
 454	 * while the old folio is copied to the new.  This prevents new
 455	 * events from being lost.
 456	 */
 457	spin_lock_irqsave(&ctx->completion_lock, flags);
 458	folio_copy(dst, src);
 459	folio_migrate_flags(dst, src);
 460	BUG_ON(ctx->ring_folios[idx] != src);
 461	ctx->ring_folios[idx] = dst;
 462	spin_unlock_irqrestore(&ctx->completion_lock, flags);
 463
 464	/* The old folio is no longer accessible. */
 465	folio_put(src);
 466
 467out_unlock:
 468	mutex_unlock(&ctx->ring_lock);
 469out:
 470	spin_unlock(&mapping->i_private_lock);
 471	return rc;
 472}
 473#else
 474#define aio_migrate_folio NULL
 475#endif
 476
 477static const struct address_space_operations aio_ctx_aops = {
 478	.dirty_folio	= noop_dirty_folio,
 479	.migrate_folio	= aio_migrate_folio,
 480};
 481
 482static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
 483{
 484	struct aio_ring *ring;
 485	struct mm_struct *mm = current->mm;
 486	unsigned long size, unused;
 487	int nr_pages;
 488	int i;
 489	struct file *file;
 490
 491	/* Compensate for the ring buffer's head/tail overlap entry */
 492	nr_events += 2;	/* 1 is required, 2 for good luck */
 493
 494	size = sizeof(struct aio_ring);
 495	size += sizeof(struct io_event) * nr_events;
 496
 497	nr_pages = PFN_UP(size);
 498	if (nr_pages < 0)
 499		return -EINVAL;
 500
 501	file = aio_private_file(ctx, nr_pages);
 502	if (IS_ERR(file)) {
 503		ctx->aio_ring_file = NULL;
 504		return -ENOMEM;
 505	}
 506
 507	ctx->aio_ring_file = file;
 508	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
 509			/ sizeof(struct io_event);
 510
 511	ctx->ring_folios = ctx->internal_folios;
 512	if (nr_pages > AIO_RING_PAGES) {
 513		ctx->ring_folios = kcalloc(nr_pages, sizeof(struct folio *),
 514					   GFP_KERNEL);
 515		if (!ctx->ring_folios) {
 516			put_aio_ring_file(ctx);
 517			return -ENOMEM;
 518		}
 519	}
 520
 521	for (i = 0; i < nr_pages; i++) {
 522		struct folio *folio;
 523
 524		folio = __filemap_get_folio(file->f_mapping, i,
 525					    FGP_LOCK | FGP_ACCESSED | FGP_CREAT,
 526					    GFP_USER | __GFP_ZERO);
 527		if (IS_ERR(folio))
 528			break;
 
 
 
 
 529
 530		pr_debug("pid(%d) [%d] folio->count=%d\n", current->pid, i,
 531			 folio_ref_count(folio));
 532		folio_end_read(folio, true);
 533
 534		ctx->ring_folios[i] = folio;
 535	}
 536	ctx->nr_pages = i;
 537
 538	if (unlikely(i != nr_pages)) {
 539		aio_free_ring(ctx);
 540		return -ENOMEM;
 541	}
 542
 543	ctx->mmap_size = nr_pages * PAGE_SIZE;
 544	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
 545
 546	if (mmap_write_lock_killable(mm)) {
 547		ctx->mmap_size = 0;
 548		aio_free_ring(ctx);
 549		return -EINTR;
 550	}
 551
 552	ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
 553				 PROT_READ | PROT_WRITE,
 554				 MAP_SHARED, 0, 0, &unused, NULL);
 555	mmap_write_unlock(mm);
 556	if (IS_ERR((void *)ctx->mmap_base)) {
 557		ctx->mmap_size = 0;
 558		aio_free_ring(ctx);
 559		return -ENOMEM;
 560	}
 561
 562	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
 563
 564	ctx->user_id = ctx->mmap_base;
 565	ctx->nr_events = nr_events; /* trusted copy */
 566
 567	ring = folio_address(ctx->ring_folios[0]);
 568	ring->nr = nr_events;	/* user copy */
 569	ring->id = ~0U;
 570	ring->head = ring->tail = 0;
 571	ring->magic = AIO_RING_MAGIC;
 572	ring->compat_features = AIO_RING_COMPAT_FEATURES;
 573	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
 574	ring->header_length = sizeof(struct aio_ring);
 575	flush_dcache_folio(ctx->ring_folios[0]);
 576
 577	return 0;
 578}
 579
 580#define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
 581#define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
 582#define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
 583
 584void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
 585{
 586	struct aio_kiocb *req;
 587	struct kioctx *ctx;
 588	unsigned long flags;
 589
 590	/*
 591	 * kiocb didn't come from aio or is neither a read nor a write, hence
 592	 * ignore it.
 593	 */
 594	if (!(iocb->ki_flags & IOCB_AIO_RW))
 595		return;
 596
 597	req = container_of(iocb, struct aio_kiocb, rw);
 598
 599	if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
 600		return;
 601
 602	ctx = req->ki_ctx;
 603
 604	spin_lock_irqsave(&ctx->ctx_lock, flags);
 605	list_add_tail(&req->ki_list, &ctx->active_reqs);
 606	req->ki_cancel = cancel;
 607	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 608}
 609EXPORT_SYMBOL(kiocb_set_cancel_fn);
 610
 611/*
 612 * free_ioctx() should be RCU delayed to synchronize against the RCU
 613 * protected lookup_ioctx() and also needs process context to call
 614 * aio_free_ring().  Use rcu_work.
 615 */
 616static void free_ioctx(struct work_struct *work)
 617{
 618	struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
 619					  free_rwork);
 620	pr_debug("freeing %p\n", ctx);
 621
 622	aio_free_ring(ctx);
 623	free_percpu(ctx->cpu);
 624	percpu_ref_exit(&ctx->reqs);
 625	percpu_ref_exit(&ctx->users);
 626	kmem_cache_free(kioctx_cachep, ctx);
 627}
 628
 629static void free_ioctx_reqs(struct percpu_ref *ref)
 630{
 631	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
 632
 633	/* At this point we know that there are no any in-flight requests */
 634	if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
 635		complete(&ctx->rq_wait->comp);
 636
 637	/* Synchronize against RCU protected table->table[] dereferences */
 638	INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
 639	queue_rcu_work(system_wq, &ctx->free_rwork);
 640}
 641
 642/*
 643 * When this function runs, the kioctx has been removed from the "hash table"
 644 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
 645 * now it's safe to cancel any that need to be.
 646 */
 647static void free_ioctx_users(struct percpu_ref *ref)
 648{
 649	struct kioctx *ctx = container_of(ref, struct kioctx, users);
 650	struct aio_kiocb *req;
 651
 652	spin_lock_irq(&ctx->ctx_lock);
 653
 654	while (!list_empty(&ctx->active_reqs)) {
 655		req = list_first_entry(&ctx->active_reqs,
 656				       struct aio_kiocb, ki_list);
 657		req->ki_cancel(&req->rw);
 658		list_del_init(&req->ki_list);
 659	}
 660
 661	spin_unlock_irq(&ctx->ctx_lock);
 662
 663	percpu_ref_kill(&ctx->reqs);
 664	percpu_ref_put(&ctx->reqs);
 665}
 666
 667static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
 668{
 669	unsigned i, new_nr;
 670	struct kioctx_table *table, *old;
 671	struct aio_ring *ring;
 672
 673	spin_lock(&mm->ioctx_lock);
 674	table = rcu_dereference_raw(mm->ioctx_table);
 675
 676	while (1) {
 677		if (table)
 678			for (i = 0; i < table->nr; i++)
 679				if (!rcu_access_pointer(table->table[i])) {
 680					ctx->id = i;
 681					rcu_assign_pointer(table->table[i], ctx);
 682					spin_unlock(&mm->ioctx_lock);
 683
 684					/* While kioctx setup is in progress,
 685					 * we are protected from page migration
 686					 * changes ring_folios by ->ring_lock.
 687					 */
 688					ring = folio_address(ctx->ring_folios[0]);
 689					ring->id = ctx->id;
 690					return 0;
 691				}
 692
 693		new_nr = (table ? table->nr : 1) * 4;
 694		spin_unlock(&mm->ioctx_lock);
 695
 696		table = kzalloc(struct_size(table, table, new_nr), GFP_KERNEL);
 697		if (!table)
 698			return -ENOMEM;
 699
 700		table->nr = new_nr;
 701
 702		spin_lock(&mm->ioctx_lock);
 703		old = rcu_dereference_raw(mm->ioctx_table);
 704
 705		if (!old) {
 706			rcu_assign_pointer(mm->ioctx_table, table);
 707		} else if (table->nr > old->nr) {
 708			memcpy(table->table, old->table,
 709			       old->nr * sizeof(struct kioctx *));
 710
 711			rcu_assign_pointer(mm->ioctx_table, table);
 712			kfree_rcu(old, rcu);
 713		} else {
 714			kfree(table);
 715			table = old;
 716		}
 717	}
 718}
 719
 720static void aio_nr_sub(unsigned nr)
 721{
 722	spin_lock(&aio_nr_lock);
 723	if (WARN_ON(aio_nr - nr > aio_nr))
 724		aio_nr = 0;
 725	else
 726		aio_nr -= nr;
 727	spin_unlock(&aio_nr_lock);
 728}
 729
 730/* ioctx_alloc
 731 *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
 732 */
 733static struct kioctx *ioctx_alloc(unsigned nr_events)
 734{
 735	struct mm_struct *mm = current->mm;
 736	struct kioctx *ctx;
 737	int err = -ENOMEM;
 738
 739	/*
 740	 * Store the original nr_events -- what userspace passed to io_setup(),
 741	 * for counting against the global limit -- before it changes.
 742	 */
 743	unsigned int max_reqs = nr_events;
 744
 745	/*
 746	 * We keep track of the number of available ringbuffer slots, to prevent
 747	 * overflow (reqs_available), and we also use percpu counters for this.
 748	 *
 749	 * So since up to half the slots might be on other cpu's percpu counters
 750	 * and unavailable, double nr_events so userspace sees what they
 751	 * expected: additionally, we move req_batch slots to/from percpu
 752	 * counters at a time, so make sure that isn't 0:
 753	 */
 754	nr_events = max(nr_events, num_possible_cpus() * 4);
 755	nr_events *= 2;
 756
 757	/* Prevent overflows */
 758	if (nr_events > (0x10000000U / sizeof(struct io_event))) {
 759		pr_debug("ENOMEM: nr_events too high\n");
 760		return ERR_PTR(-EINVAL);
 761	}
 762
 763	if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
 764		return ERR_PTR(-EAGAIN);
 765
 766	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
 767	if (!ctx)
 768		return ERR_PTR(-ENOMEM);
 769
 770	ctx->max_reqs = max_reqs;
 771
 772	spin_lock_init(&ctx->ctx_lock);
 773	spin_lock_init(&ctx->completion_lock);
 774	mutex_init(&ctx->ring_lock);
 775	/* Protect against page migration throughout kiotx setup by keeping
 776	 * the ring_lock mutex held until setup is complete. */
 777	mutex_lock(&ctx->ring_lock);
 778	init_waitqueue_head(&ctx->wait);
 779
 780	INIT_LIST_HEAD(&ctx->active_reqs);
 781
 782	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
 783		goto err;
 784
 785	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
 786		goto err;
 787
 788	ctx->cpu = alloc_percpu(struct kioctx_cpu);
 789	if (!ctx->cpu)
 790		goto err;
 791
 792	err = aio_setup_ring(ctx, nr_events);
 793	if (err < 0)
 794		goto err;
 795
 796	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
 797	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
 798	if (ctx->req_batch < 1)
 799		ctx->req_batch = 1;
 800
 801	/* limit the number of system wide aios */
 802	spin_lock(&aio_nr_lock);
 803	if (aio_nr + ctx->max_reqs > aio_max_nr ||
 804	    aio_nr + ctx->max_reqs < aio_nr) {
 805		spin_unlock(&aio_nr_lock);
 806		err = -EAGAIN;
 807		goto err_ctx;
 808	}
 809	aio_nr += ctx->max_reqs;
 810	spin_unlock(&aio_nr_lock);
 811
 812	percpu_ref_get(&ctx->users);	/* io_setup() will drop this ref */
 813	percpu_ref_get(&ctx->reqs);	/* free_ioctx_users() will drop this */
 814
 815	err = ioctx_add_table(ctx, mm);
 816	if (err)
 817		goto err_cleanup;
 818
 819	/* Release the ring_lock mutex now that all setup is complete. */
 820	mutex_unlock(&ctx->ring_lock);
 821
 822	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
 823		 ctx, ctx->user_id, mm, ctx->nr_events);
 824	return ctx;
 825
 826err_cleanup:
 827	aio_nr_sub(ctx->max_reqs);
 828err_ctx:
 829	atomic_set(&ctx->dead, 1);
 830	if (ctx->mmap_size)
 831		vm_munmap(ctx->mmap_base, ctx->mmap_size);
 832	aio_free_ring(ctx);
 833err:
 834	mutex_unlock(&ctx->ring_lock);
 835	free_percpu(ctx->cpu);
 836	percpu_ref_exit(&ctx->reqs);
 837	percpu_ref_exit(&ctx->users);
 838	kmem_cache_free(kioctx_cachep, ctx);
 839	pr_debug("error allocating ioctx %d\n", err);
 840	return ERR_PTR(err);
 841}
 842
 843/* kill_ioctx
 844 *	Cancels all outstanding aio requests on an aio context.  Used
 845 *	when the processes owning a context have all exited to encourage
 846 *	the rapid destruction of the kioctx.
 847 */
 848static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
 849		      struct ctx_rq_wait *wait)
 850{
 851	struct kioctx_table *table;
 852
 853	spin_lock(&mm->ioctx_lock);
 854	if (atomic_xchg(&ctx->dead, 1)) {
 855		spin_unlock(&mm->ioctx_lock);
 856		return -EINVAL;
 857	}
 858
 859	table = rcu_dereference_raw(mm->ioctx_table);
 860	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
 861	RCU_INIT_POINTER(table->table[ctx->id], NULL);
 862	spin_unlock(&mm->ioctx_lock);
 863
 864	/* free_ioctx_reqs() will do the necessary RCU synchronization */
 865	wake_up_all(&ctx->wait);
 866
 867	/*
 868	 * It'd be more correct to do this in free_ioctx(), after all
 869	 * the outstanding kiocbs have finished - but by then io_destroy
 870	 * has already returned, so io_setup() could potentially return
 871	 * -EAGAIN with no ioctxs actually in use (as far as userspace
 872	 *  could tell).
 873	 */
 874	aio_nr_sub(ctx->max_reqs);
 875
 876	if (ctx->mmap_size)
 877		vm_munmap(ctx->mmap_base, ctx->mmap_size);
 878
 879	ctx->rq_wait = wait;
 880	percpu_ref_kill(&ctx->users);
 881	return 0;
 882}
 883
 884/*
 885 * exit_aio: called when the last user of mm goes away.  At this point, there is
 886 * no way for any new requests to be submited or any of the io_* syscalls to be
 887 * called on the context.
 888 *
 889 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
 890 * them.
 891 */
 892void exit_aio(struct mm_struct *mm)
 893{
 894	struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
 895	struct ctx_rq_wait wait;
 896	int i, skipped;
 897
 898	if (!table)
 899		return;
 900
 901	atomic_set(&wait.count, table->nr);
 902	init_completion(&wait.comp);
 903
 904	skipped = 0;
 905	for (i = 0; i < table->nr; ++i) {
 906		struct kioctx *ctx =
 907			rcu_dereference_protected(table->table[i], true);
 908
 909		if (!ctx) {
 910			skipped++;
 911			continue;
 912		}
 913
 914		/*
 915		 * We don't need to bother with munmap() here - exit_mmap(mm)
 916		 * is coming and it'll unmap everything. And we simply can't,
 917		 * this is not necessarily our ->mm.
 918		 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
 919		 * that it needs to unmap the area, just set it to 0.
 920		 */
 921		ctx->mmap_size = 0;
 922		kill_ioctx(mm, ctx, &wait);
 923	}
 924
 925	if (!atomic_sub_and_test(skipped, &wait.count)) {
 926		/* Wait until all IO for the context are done. */
 927		wait_for_completion(&wait.comp);
 928	}
 929
 930	RCU_INIT_POINTER(mm->ioctx_table, NULL);
 931	kfree(table);
 932}
 933
 934static void put_reqs_available(struct kioctx *ctx, unsigned nr)
 935{
 936	struct kioctx_cpu *kcpu;
 937	unsigned long flags;
 938
 939	local_irq_save(flags);
 940	kcpu = this_cpu_ptr(ctx->cpu);
 941	kcpu->reqs_available += nr;
 942
 943	while (kcpu->reqs_available >= ctx->req_batch * 2) {
 944		kcpu->reqs_available -= ctx->req_batch;
 945		atomic_add(ctx->req_batch, &ctx->reqs_available);
 946	}
 947
 948	local_irq_restore(flags);
 949}
 950
 951static bool __get_reqs_available(struct kioctx *ctx)
 952{
 953	struct kioctx_cpu *kcpu;
 954	bool ret = false;
 955	unsigned long flags;
 956
 957	local_irq_save(flags);
 958	kcpu = this_cpu_ptr(ctx->cpu);
 959	if (!kcpu->reqs_available) {
 960		int avail = atomic_read(&ctx->reqs_available);
 961
 962		do {
 963			if (avail < ctx->req_batch)
 964				goto out;
 965		} while (!atomic_try_cmpxchg(&ctx->reqs_available,
 966					     &avail, avail - ctx->req_batch));
 967
 968		kcpu->reqs_available += ctx->req_batch;
 969	}
 970
 971	ret = true;
 972	kcpu->reqs_available--;
 973out:
 974	local_irq_restore(flags);
 975	return ret;
 976}
 977
 978/* refill_reqs_available
 979 *	Updates the reqs_available reference counts used for tracking the
 980 *	number of free slots in the completion ring.  This can be called
 981 *	from aio_complete() (to optimistically update reqs_available) or
 982 *	from aio_get_req() (the we're out of events case).  It must be
 983 *	called holding ctx->completion_lock.
 984 */
 985static void refill_reqs_available(struct kioctx *ctx, unsigned head,
 986                                  unsigned tail)
 987{
 988	unsigned events_in_ring, completed;
 989
 990	/* Clamp head since userland can write to it. */
 991	head %= ctx->nr_events;
 992	if (head <= tail)
 993		events_in_ring = tail - head;
 994	else
 995		events_in_ring = ctx->nr_events - (head - tail);
 996
 997	completed = ctx->completed_events;
 998	if (events_in_ring < completed)
 999		completed -= events_in_ring;
1000	else
1001		completed = 0;
1002
1003	if (!completed)
1004		return;
1005
1006	ctx->completed_events -= completed;
1007	put_reqs_available(ctx, completed);
1008}
1009
1010/* user_refill_reqs_available
1011 *	Called to refill reqs_available when aio_get_req() encounters an
1012 *	out of space in the completion ring.
1013 */
1014static void user_refill_reqs_available(struct kioctx *ctx)
1015{
1016	spin_lock_irq(&ctx->completion_lock);
1017	if (ctx->completed_events) {
1018		struct aio_ring *ring;
1019		unsigned head;
1020
1021		/* Access of ring->head may race with aio_read_events_ring()
1022		 * here, but that's okay since whether we read the old version
1023		 * or the new version, and either will be valid.  The important
1024		 * part is that head cannot pass tail since we prevent
1025		 * aio_complete() from updating tail by holding
1026		 * ctx->completion_lock.  Even if head is invalid, the check
1027		 * against ctx->completed_events below will make sure we do the
1028		 * safe/right thing.
1029		 */
1030		ring = folio_address(ctx->ring_folios[0]);
1031		head = ring->head;
1032
1033		refill_reqs_available(ctx, head, ctx->tail);
1034	}
1035
1036	spin_unlock_irq(&ctx->completion_lock);
1037}
1038
1039static bool get_reqs_available(struct kioctx *ctx)
1040{
1041	if (__get_reqs_available(ctx))
1042		return true;
1043	user_refill_reqs_available(ctx);
1044	return __get_reqs_available(ctx);
1045}
1046
1047/* aio_get_req
1048 *	Allocate a slot for an aio request.
1049 * Returns NULL if no requests are free.
1050 *
1051 * The refcount is initialized to 2 - one for the async op completion,
1052 * one for the synchronous code that does this.
1053 */
1054static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1055{
1056	struct aio_kiocb *req;
1057
1058	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1059	if (unlikely(!req))
1060		return NULL;
1061
1062	if (unlikely(!get_reqs_available(ctx))) {
1063		kmem_cache_free(kiocb_cachep, req);
1064		return NULL;
1065	}
1066
1067	percpu_ref_get(&ctx->reqs);
1068	req->ki_ctx = ctx;
1069	INIT_LIST_HEAD(&req->ki_list);
1070	refcount_set(&req->ki_refcnt, 2);
1071	req->ki_eventfd = NULL;
1072	return req;
1073}
1074
1075static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1076{
1077	struct aio_ring __user *ring  = (void __user *)ctx_id;
1078	struct mm_struct *mm = current->mm;
1079	struct kioctx *ctx, *ret = NULL;
1080	struct kioctx_table *table;
1081	unsigned id;
1082
1083	if (get_user(id, &ring->id))
1084		return NULL;
1085
1086	rcu_read_lock();
1087	table = rcu_dereference(mm->ioctx_table);
1088
1089	if (!table || id >= table->nr)
1090		goto out;
1091
1092	id = array_index_nospec(id, table->nr);
1093	ctx = rcu_dereference(table->table[id]);
1094	if (ctx && ctx->user_id == ctx_id) {
1095		if (percpu_ref_tryget_live(&ctx->users))
1096			ret = ctx;
1097	}
1098out:
1099	rcu_read_unlock();
1100	return ret;
1101}
1102
1103static inline void iocb_destroy(struct aio_kiocb *iocb)
1104{
1105	if (iocb->ki_eventfd)
1106		eventfd_ctx_put(iocb->ki_eventfd);
1107	if (iocb->ki_filp)
1108		fput(iocb->ki_filp);
1109	percpu_ref_put(&iocb->ki_ctx->reqs);
1110	kmem_cache_free(kiocb_cachep, iocb);
1111}
1112
1113struct aio_waiter {
1114	struct wait_queue_entry	w;
1115	size_t			min_nr;
1116};
1117
1118/* aio_complete
1119 *	Called when the io request on the given iocb is complete.
1120 */
1121static void aio_complete(struct aio_kiocb *iocb)
1122{
1123	struct kioctx	*ctx = iocb->ki_ctx;
1124	struct aio_ring	*ring;
1125	struct io_event	*ev_page, *event;
1126	unsigned tail, pos, head, avail;
1127	unsigned long	flags;
1128
1129	/*
1130	 * Add a completion event to the ring buffer. Must be done holding
1131	 * ctx->completion_lock to prevent other code from messing with the tail
1132	 * pointer since we might be called from irq context.
1133	 */
1134	spin_lock_irqsave(&ctx->completion_lock, flags);
1135
1136	tail = ctx->tail;
1137	pos = tail + AIO_EVENTS_OFFSET;
1138
1139	if (++tail >= ctx->nr_events)
1140		tail = 0;
1141
1142	ev_page = folio_address(ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE]);
1143	event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1144
1145	*event = iocb->ki_res;
1146
1147	flush_dcache_folio(ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE]);
1148
1149	pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1150		 (void __user *)(unsigned long)iocb->ki_res.obj,
1151		 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1152
1153	/* after flagging the request as done, we
1154	 * must never even look at it again
1155	 */
1156	smp_wmb();	/* make event visible before updating tail */
1157
1158	ctx->tail = tail;
1159
1160	ring = folio_address(ctx->ring_folios[0]);
1161	head = ring->head;
1162	ring->tail = tail;
1163	flush_dcache_folio(ctx->ring_folios[0]);
1164
1165	ctx->completed_events++;
1166	if (ctx->completed_events > 1)
1167		refill_reqs_available(ctx, head, tail);
1168
1169	avail = tail > head
1170		? tail - head
1171		: tail + ctx->nr_events - head;
1172	spin_unlock_irqrestore(&ctx->completion_lock, flags);
1173
1174	pr_debug("added to ring %p at [%u]\n", iocb, tail);
1175
1176	/*
1177	 * Check if the user asked us to deliver the result through an
1178	 * eventfd. The eventfd_signal() function is safe to be called
1179	 * from IRQ context.
1180	 */
1181	if (iocb->ki_eventfd)
1182		eventfd_signal(iocb->ki_eventfd);
1183
1184	/*
1185	 * We have to order our ring_info tail store above and test
1186	 * of the wait list below outside the wait lock.  This is
1187	 * like in wake_up_bit() where clearing a bit has to be
1188	 * ordered with the unlocked test.
1189	 */
1190	smp_mb();
1191
1192	if (waitqueue_active(&ctx->wait)) {
1193		struct aio_waiter *curr, *next;
1194		unsigned long flags;
1195
1196		spin_lock_irqsave(&ctx->wait.lock, flags);
1197		list_for_each_entry_safe(curr, next, &ctx->wait.head, w.entry)
1198			if (avail >= curr->min_nr) {
1199				wake_up_process(curr->w.private);
1200				list_del_init_careful(&curr->w.entry);
 
1201			}
1202		spin_unlock_irqrestore(&ctx->wait.lock, flags);
1203	}
1204}
1205
1206static inline void iocb_put(struct aio_kiocb *iocb)
1207{
1208	if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1209		aio_complete(iocb);
1210		iocb_destroy(iocb);
1211	}
1212}
1213
1214/* aio_read_events_ring
1215 *	Pull an event off of the ioctx's event ring.  Returns the number of
1216 *	events fetched
1217 */
1218static long aio_read_events_ring(struct kioctx *ctx,
1219				 struct io_event __user *event, long nr)
1220{
1221	struct aio_ring *ring;
1222	unsigned head, tail, pos;
1223	long ret = 0;
1224	int copy_ret;
1225
1226	/*
1227	 * The mutex can block and wake us up and that will cause
1228	 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1229	 * and repeat. This should be rare enough that it doesn't cause
1230	 * peformance issues. See the comment in read_events() for more detail.
1231	 */
1232	sched_annotate_sleep();
1233	mutex_lock(&ctx->ring_lock);
1234
1235	/* Access to ->ring_folios here is protected by ctx->ring_lock. */
1236	ring = folio_address(ctx->ring_folios[0]);
1237	head = ring->head;
1238	tail = ring->tail;
1239
1240	/*
1241	 * Ensure that once we've read the current tail pointer, that
1242	 * we also see the events that were stored up to the tail.
1243	 */
1244	smp_rmb();
1245
1246	pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1247
1248	if (head == tail)
1249		goto out;
1250
1251	head %= ctx->nr_events;
1252	tail %= ctx->nr_events;
1253
1254	while (ret < nr) {
1255		long avail;
1256		struct io_event *ev;
1257		struct folio *folio;
1258
1259		avail = (head <= tail ?  tail : ctx->nr_events) - head;
1260		if (head == tail)
1261			break;
1262
1263		pos = head + AIO_EVENTS_OFFSET;
1264		folio = ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE];
1265		pos %= AIO_EVENTS_PER_PAGE;
1266
1267		avail = min(avail, nr - ret);
1268		avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1269
1270		ev = folio_address(folio);
1271		copy_ret = copy_to_user(event + ret, ev + pos,
1272					sizeof(*ev) * avail);
1273
1274		if (unlikely(copy_ret)) {
1275			ret = -EFAULT;
1276			goto out;
1277		}
1278
1279		ret += avail;
1280		head += avail;
1281		head %= ctx->nr_events;
1282	}
1283
1284	ring = folio_address(ctx->ring_folios[0]);
1285	ring->head = head;
1286	flush_dcache_folio(ctx->ring_folios[0]);
1287
1288	pr_debug("%li  h%u t%u\n", ret, head, tail);
1289out:
1290	mutex_unlock(&ctx->ring_lock);
1291
1292	return ret;
1293}
1294
1295static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1296			    struct io_event __user *event, long *i)
1297{
1298	long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1299
1300	if (ret > 0)
1301		*i += ret;
1302
1303	if (unlikely(atomic_read(&ctx->dead)))
1304		ret = -EINVAL;
1305
1306	if (!*i)
1307		*i = ret;
1308
1309	return ret < 0 || *i >= min_nr;
1310}
1311
1312static long read_events(struct kioctx *ctx, long min_nr, long nr,
1313			struct io_event __user *event,
1314			ktime_t until)
1315{
1316	struct hrtimer_sleeper	t;
1317	struct aio_waiter	w;
1318	long ret = 0, ret2 = 0;
1319
1320	/*
1321	 * Note that aio_read_events() is being called as the conditional - i.e.
1322	 * we're calling it after prepare_to_wait() has set task state to
1323	 * TASK_INTERRUPTIBLE.
1324	 *
1325	 * But aio_read_events() can block, and if it blocks it's going to flip
1326	 * the task state back to TASK_RUNNING.
1327	 *
1328	 * This should be ok, provided it doesn't flip the state back to
1329	 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1330	 * will only happen if the mutex_lock() call blocks, and we then find
1331	 * the ringbuffer empty. So in practice we should be ok, but it's
1332	 * something to be aware of when touching this code.
1333	 */
1334	aio_read_events(ctx, min_nr, nr, event, &ret);
1335	if (until == 0 || ret < 0 || ret >= min_nr)
1336		return ret;
1337
1338	hrtimer_setup_sleeper_on_stack(&t, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1339	if (until != KTIME_MAX) {
1340		hrtimer_set_expires_range_ns(&t.timer, until, current->timer_slack_ns);
1341		hrtimer_sleeper_start_expires(&t, HRTIMER_MODE_REL);
1342	}
1343
1344	init_wait(&w.w);
1345
1346	while (1) {
1347		unsigned long nr_got = ret;
1348
1349		w.min_nr = min_nr - ret;
1350
1351		ret2 = prepare_to_wait_event(&ctx->wait, &w.w, TASK_INTERRUPTIBLE);
1352		if (!ret2 && !t.task)
1353			ret2 = -ETIME;
1354
1355		if (aio_read_events(ctx, min_nr, nr, event, &ret) || ret2)
1356			break;
1357
1358		if (nr_got == ret)
1359			schedule();
1360	}
1361
1362	finish_wait(&ctx->wait, &w.w);
1363	hrtimer_cancel(&t.timer);
1364	destroy_hrtimer_on_stack(&t.timer);
1365
1366	return ret;
1367}
1368
1369/* sys_io_setup:
1370 *	Create an aio_context capable of receiving at least nr_events.
1371 *	ctxp must not point to an aio_context that already exists, and
1372 *	must be initialized to 0 prior to the call.  On successful
1373 *	creation of the aio_context, *ctxp is filled in with the resulting 
1374 *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1375 *	if the specified nr_events exceeds internal limits.  May fail 
1376 *	with -EAGAIN if the specified nr_events exceeds the user's limit 
1377 *	of available events.  May fail with -ENOMEM if insufficient kernel
1378 *	resources are available.  May fail with -EFAULT if an invalid
1379 *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1380 *	implemented.
1381 */
1382SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1383{
1384	struct kioctx *ioctx = NULL;
1385	unsigned long ctx;
1386	long ret;
1387
1388	ret = get_user(ctx, ctxp);
1389	if (unlikely(ret))
1390		goto out;
1391
1392	ret = -EINVAL;
1393	if (unlikely(ctx || nr_events == 0)) {
1394		pr_debug("EINVAL: ctx %lu nr_events %u\n",
1395		         ctx, nr_events);
1396		goto out;
1397	}
1398
1399	ioctx = ioctx_alloc(nr_events);
1400	ret = PTR_ERR(ioctx);
1401	if (!IS_ERR(ioctx)) {
1402		ret = put_user(ioctx->user_id, ctxp);
1403		if (ret)
1404			kill_ioctx(current->mm, ioctx, NULL);
1405		percpu_ref_put(&ioctx->users);
1406	}
1407
1408out:
1409	return ret;
1410}
1411
1412#ifdef CONFIG_COMPAT
1413COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1414{
1415	struct kioctx *ioctx = NULL;
1416	unsigned long ctx;
1417	long ret;
1418
1419	ret = get_user(ctx, ctx32p);
1420	if (unlikely(ret))
1421		goto out;
1422
1423	ret = -EINVAL;
1424	if (unlikely(ctx || nr_events == 0)) {
1425		pr_debug("EINVAL: ctx %lu nr_events %u\n",
1426		         ctx, nr_events);
1427		goto out;
1428	}
1429
1430	ioctx = ioctx_alloc(nr_events);
1431	ret = PTR_ERR(ioctx);
1432	if (!IS_ERR(ioctx)) {
1433		/* truncating is ok because it's a user address */
1434		ret = put_user((u32)ioctx->user_id, ctx32p);
1435		if (ret)
1436			kill_ioctx(current->mm, ioctx, NULL);
1437		percpu_ref_put(&ioctx->users);
1438	}
1439
1440out:
1441	return ret;
1442}
1443#endif
1444
1445/* sys_io_destroy:
1446 *	Destroy the aio_context specified.  May cancel any outstanding 
1447 *	AIOs and block on completion.  Will fail with -ENOSYS if not
1448 *	implemented.  May fail with -EINVAL if the context pointed to
1449 *	is invalid.
1450 */
1451SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1452{
1453	struct kioctx *ioctx = lookup_ioctx(ctx);
1454	if (likely(NULL != ioctx)) {
1455		struct ctx_rq_wait wait;
1456		int ret;
1457
1458		init_completion(&wait.comp);
1459		atomic_set(&wait.count, 1);
1460
1461		/* Pass requests_done to kill_ioctx() where it can be set
1462		 * in a thread-safe way. If we try to set it here then we have
1463		 * a race condition if two io_destroy() called simultaneously.
1464		 */
1465		ret = kill_ioctx(current->mm, ioctx, &wait);
1466		percpu_ref_put(&ioctx->users);
1467
1468		/* Wait until all IO for the context are done. Otherwise kernel
1469		 * keep using user-space buffers even if user thinks the context
1470		 * is destroyed.
1471		 */
1472		if (!ret)
1473			wait_for_completion(&wait.comp);
1474
1475		return ret;
1476	}
1477	pr_debug("EINVAL: invalid context id\n");
1478	return -EINVAL;
1479}
1480
1481static void aio_remove_iocb(struct aio_kiocb *iocb)
1482{
1483	struct kioctx *ctx = iocb->ki_ctx;
1484	unsigned long flags;
1485
1486	spin_lock_irqsave(&ctx->ctx_lock, flags);
1487	list_del(&iocb->ki_list);
1488	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1489}
1490
1491static void aio_complete_rw(struct kiocb *kiocb, long res)
1492{
1493	struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1494
1495	if (!list_empty_careful(&iocb->ki_list))
1496		aio_remove_iocb(iocb);
1497
1498	if (kiocb->ki_flags & IOCB_WRITE) {
1499		struct inode *inode = file_inode(kiocb->ki_filp);
1500
1501		if (S_ISREG(inode->i_mode))
1502			kiocb_end_write(kiocb);
1503	}
1504
1505	iocb->ki_res.res = res;
1506	iocb->ki_res.res2 = 0;
1507	iocb_put(iocb);
1508}
1509
1510static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb, int rw_type)
1511{
1512	int ret;
1513
1514	req->ki_complete = aio_complete_rw;
1515	req->private = NULL;
1516	req->ki_pos = iocb->aio_offset;
1517	req->ki_flags = req->ki_filp->f_iocb_flags | IOCB_AIO_RW;
1518	if (iocb->aio_flags & IOCB_FLAG_RESFD)
1519		req->ki_flags |= IOCB_EVENTFD;
1520	if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1521		/*
1522		 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1523		 * aio_reqprio is interpreted as an I/O scheduling
1524		 * class and priority.
1525		 */
1526		ret = ioprio_check_cap(iocb->aio_reqprio);
1527		if (ret) {
1528			pr_debug("aio ioprio check cap error: %d\n", ret);
1529			return ret;
1530		}
1531
1532		req->ki_ioprio = iocb->aio_reqprio;
1533	} else
1534		req->ki_ioprio = get_current_ioprio();
1535
1536	ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags, rw_type);
1537	if (unlikely(ret))
1538		return ret;
1539
1540	req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1541	return 0;
1542}
1543
1544static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1545		struct iovec **iovec, bool vectored, bool compat,
1546		struct iov_iter *iter)
1547{
1548	void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1549	size_t len = iocb->aio_nbytes;
1550
1551	if (!vectored) {
1552		ssize_t ret = import_ubuf(rw, buf, len, iter);
1553		*iovec = NULL;
1554		return ret;
1555	}
1556
1557	return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1558}
1559
1560static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1561{
1562	switch (ret) {
1563	case -EIOCBQUEUED:
1564		break;
1565	case -ERESTARTSYS:
1566	case -ERESTARTNOINTR:
1567	case -ERESTARTNOHAND:
1568	case -ERESTART_RESTARTBLOCK:
1569		/*
1570		 * There's no easy way to restart the syscall since other AIO's
1571		 * may be already running. Just fail this IO with EINTR.
1572		 */
1573		ret = -EINTR;
1574		fallthrough;
1575	default:
1576		req->ki_complete(req, ret);
1577	}
1578}
1579
1580static int aio_read(struct kiocb *req, const struct iocb *iocb,
1581			bool vectored, bool compat)
1582{
1583	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1584	struct iov_iter iter;
1585	struct file *file;
1586	int ret;
1587
1588	ret = aio_prep_rw(req, iocb, READ);
1589	if (ret)
1590		return ret;
1591	file = req->ki_filp;
1592	if (unlikely(!(file->f_mode & FMODE_READ)))
1593		return -EBADF;
1594	if (unlikely(!file->f_op->read_iter))
1595		return -EINVAL;
1596
1597	ret = aio_setup_rw(ITER_DEST, iocb, &iovec, vectored, compat, &iter);
1598	if (ret < 0)
1599		return ret;
1600	ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1601	if (!ret)
1602		aio_rw_done(req, file->f_op->read_iter(req, &iter));
1603	kfree(iovec);
1604	return ret;
1605}
1606
1607static int aio_write(struct kiocb *req, const struct iocb *iocb,
1608			 bool vectored, bool compat)
1609{
1610	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1611	struct iov_iter iter;
1612	struct file *file;
1613	int ret;
1614
1615	ret = aio_prep_rw(req, iocb, WRITE);
1616	if (ret)
1617		return ret;
1618	file = req->ki_filp;
1619
1620	if (unlikely(!(file->f_mode & FMODE_WRITE)))
1621		return -EBADF;
1622	if (unlikely(!file->f_op->write_iter))
1623		return -EINVAL;
1624
1625	ret = aio_setup_rw(ITER_SOURCE, iocb, &iovec, vectored, compat, &iter);
1626	if (ret < 0)
1627		return ret;
1628	ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1629	if (!ret) {
1630		if (S_ISREG(file_inode(file)->i_mode))
1631			kiocb_start_write(req);
1632		req->ki_flags |= IOCB_WRITE;
1633		aio_rw_done(req, file->f_op->write_iter(req, &iter));
1634	}
1635	kfree(iovec);
1636	return ret;
1637}
1638
1639static void aio_fsync_work(struct work_struct *work)
1640{
1641	struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1642	const struct cred *old_cred = override_creds(iocb->fsync.creds);
1643
1644	iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1645	revert_creds(old_cred);
1646	put_cred(iocb->fsync.creds);
1647	iocb_put(iocb);
1648}
1649
1650static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1651		     bool datasync)
1652{
1653	if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1654			iocb->aio_rw_flags))
1655		return -EINVAL;
1656
1657	if (unlikely(!req->file->f_op->fsync))
1658		return -EINVAL;
1659
1660	req->creds = prepare_creds();
1661	if (!req->creds)
1662		return -ENOMEM;
1663
1664	req->datasync = datasync;
1665	INIT_WORK(&req->work, aio_fsync_work);
1666	schedule_work(&req->work);
1667	return 0;
1668}
1669
1670static void aio_poll_put_work(struct work_struct *work)
1671{
1672	struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1673	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1674
1675	iocb_put(iocb);
1676}
1677
1678/*
1679 * Safely lock the waitqueue which the request is on, synchronizing with the
1680 * case where the ->poll() provider decides to free its waitqueue early.
1681 *
1682 * Returns true on success, meaning that req->head->lock was locked, req->wait
1683 * is on req->head, and an RCU read lock was taken.  Returns false if the
1684 * request was already removed from its waitqueue (which might no longer exist).
1685 */
1686static bool poll_iocb_lock_wq(struct poll_iocb *req)
1687{
1688	wait_queue_head_t *head;
1689
1690	/*
1691	 * While we hold the waitqueue lock and the waitqueue is nonempty,
1692	 * wake_up_pollfree() will wait for us.  However, taking the waitqueue
1693	 * lock in the first place can race with the waitqueue being freed.
1694	 *
1695	 * We solve this as eventpoll does: by taking advantage of the fact that
1696	 * all users of wake_up_pollfree() will RCU-delay the actual free.  If
1697	 * we enter rcu_read_lock() and see that the pointer to the queue is
1698	 * non-NULL, we can then lock it without the memory being freed out from
1699	 * under us, then check whether the request is still on the queue.
1700	 *
1701	 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1702	 * case the caller deletes the entry from the queue, leaving it empty.
1703	 * In that case, only RCU prevents the queue memory from being freed.
1704	 */
1705	rcu_read_lock();
1706	head = smp_load_acquire(&req->head);
1707	if (head) {
1708		spin_lock(&head->lock);
1709		if (!list_empty(&req->wait.entry))
1710			return true;
1711		spin_unlock(&head->lock);
1712	}
1713	rcu_read_unlock();
1714	return false;
1715}
1716
1717static void poll_iocb_unlock_wq(struct poll_iocb *req)
1718{
1719	spin_unlock(&req->head->lock);
1720	rcu_read_unlock();
1721}
1722
1723static void aio_poll_complete_work(struct work_struct *work)
1724{
1725	struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1726	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1727	struct poll_table_struct pt = { ._key = req->events };
1728	struct kioctx *ctx = iocb->ki_ctx;
1729	__poll_t mask = 0;
1730
1731	if (!READ_ONCE(req->cancelled))
1732		mask = vfs_poll(req->file, &pt) & req->events;
1733
1734	/*
1735	 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1736	 * calling ->ki_cancel.  We need the ctx_lock roundtrip here to
1737	 * synchronize with them.  In the cancellation case the list_del_init
1738	 * itself is not actually needed, but harmless so we keep it in to
1739	 * avoid further branches in the fast path.
1740	 */
1741	spin_lock_irq(&ctx->ctx_lock);
1742	if (poll_iocb_lock_wq(req)) {
1743		if (!mask && !READ_ONCE(req->cancelled)) {
1744			/*
1745			 * The request isn't actually ready to be completed yet.
1746			 * Reschedule completion if another wakeup came in.
1747			 */
1748			if (req->work_need_resched) {
1749				schedule_work(&req->work);
1750				req->work_need_resched = false;
1751			} else {
1752				req->work_scheduled = false;
1753			}
1754			poll_iocb_unlock_wq(req);
1755			spin_unlock_irq(&ctx->ctx_lock);
1756			return;
1757		}
1758		list_del_init(&req->wait.entry);
1759		poll_iocb_unlock_wq(req);
1760	} /* else, POLLFREE has freed the waitqueue, so we must complete */
1761	list_del_init(&iocb->ki_list);
1762	iocb->ki_res.res = mangle_poll(mask);
1763	spin_unlock_irq(&ctx->ctx_lock);
1764
1765	iocb_put(iocb);
1766}
1767
1768/* assumes we are called with irqs disabled */
1769static int aio_poll_cancel(struct kiocb *iocb)
1770{
1771	struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1772	struct poll_iocb *req = &aiocb->poll;
1773
1774	if (poll_iocb_lock_wq(req)) {
1775		WRITE_ONCE(req->cancelled, true);
1776		if (!req->work_scheduled) {
1777			schedule_work(&aiocb->poll.work);
1778			req->work_scheduled = true;
1779		}
1780		poll_iocb_unlock_wq(req);
1781	} /* else, the request was force-cancelled by POLLFREE already */
1782
1783	return 0;
1784}
1785
1786static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1787		void *key)
1788{
1789	struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1790	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1791	__poll_t mask = key_to_poll(key);
1792	unsigned long flags;
1793
1794	/* for instances that support it check for an event match first: */
1795	if (mask && !(mask & req->events))
1796		return 0;
1797
1798	/*
1799	 * Complete the request inline if possible.  This requires that three
1800	 * conditions be met:
1801	 *   1. An event mask must have been passed.  If a plain wakeup was done
1802	 *	instead, then mask == 0 and we have to call vfs_poll() to get
1803	 *	the events, so inline completion isn't possible.
1804	 *   2. The completion work must not have already been scheduled.
1805	 *   3. ctx_lock must not be busy.  We have to use trylock because we
1806	 *	already hold the waitqueue lock, so this inverts the normal
1807	 *	locking order.  Use irqsave/irqrestore because not all
1808	 *	filesystems (e.g. fuse) call this function with IRQs disabled,
1809	 *	yet IRQs have to be disabled before ctx_lock is obtained.
1810	 */
1811	if (mask && !req->work_scheduled &&
1812	    spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1813		struct kioctx *ctx = iocb->ki_ctx;
1814
1815		list_del_init(&req->wait.entry);
1816		list_del(&iocb->ki_list);
1817		iocb->ki_res.res = mangle_poll(mask);
1818		if (iocb->ki_eventfd && !eventfd_signal_allowed()) {
1819			iocb = NULL;
1820			INIT_WORK(&req->work, aio_poll_put_work);
1821			schedule_work(&req->work);
1822		}
1823		spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1824		if (iocb)
1825			iocb_put(iocb);
1826	} else {
1827		/*
1828		 * Schedule the completion work if needed.  If it was already
1829		 * scheduled, record that another wakeup came in.
1830		 *
1831		 * Don't remove the request from the waitqueue here, as it might
1832		 * not actually be complete yet (we won't know until vfs_poll()
1833		 * is called), and we must not miss any wakeups.  POLLFREE is an
1834		 * exception to this; see below.
1835		 */
1836		if (req->work_scheduled) {
1837			req->work_need_resched = true;
1838		} else {
1839			schedule_work(&req->work);
1840			req->work_scheduled = true;
1841		}
1842
1843		/*
1844		 * If the waitqueue is being freed early but we can't complete
1845		 * the request inline, we have to tear down the request as best
1846		 * we can.  That means immediately removing the request from its
1847		 * waitqueue and preventing all further accesses to the
1848		 * waitqueue via the request.  We also need to schedule the
1849		 * completion work (done above).  Also mark the request as
1850		 * cancelled, to potentially skip an unneeded call to ->poll().
1851		 */
1852		if (mask & POLLFREE) {
1853			WRITE_ONCE(req->cancelled, true);
1854			list_del_init(&req->wait.entry);
1855
1856			/*
1857			 * Careful: this *must* be the last step, since as soon
1858			 * as req->head is NULL'ed out, the request can be
1859			 * completed and freed, since aio_poll_complete_work()
1860			 * will no longer need to take the waitqueue lock.
1861			 */
1862			smp_store_release(&req->head, NULL);
1863		}
1864	}
1865	return 1;
1866}
1867
1868struct aio_poll_table {
1869	struct poll_table_struct	pt;
1870	struct aio_kiocb		*iocb;
1871	bool				queued;
1872	int				error;
1873};
1874
1875static void
1876aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1877		struct poll_table_struct *p)
1878{
1879	struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1880
1881	/* multiple wait queues per file are not supported */
1882	if (unlikely(pt->queued)) {
1883		pt->error = -EINVAL;
1884		return;
1885	}
1886
1887	pt->queued = true;
1888	pt->error = 0;
1889	pt->iocb->poll.head = head;
1890	add_wait_queue(head, &pt->iocb->poll.wait);
1891}
1892
1893static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1894{
1895	struct kioctx *ctx = aiocb->ki_ctx;
1896	struct poll_iocb *req = &aiocb->poll;
1897	struct aio_poll_table apt;
1898	bool cancel = false;
1899	__poll_t mask;
1900
1901	/* reject any unknown events outside the normal event mask. */
1902	if ((u16)iocb->aio_buf != iocb->aio_buf)
1903		return -EINVAL;
1904	/* reject fields that are not defined for poll */
1905	if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1906		return -EINVAL;
1907
1908	INIT_WORK(&req->work, aio_poll_complete_work);
1909	req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1910
1911	req->head = NULL;
1912	req->cancelled = false;
1913	req->work_scheduled = false;
1914	req->work_need_resched = false;
1915
1916	apt.pt._qproc = aio_poll_queue_proc;
1917	apt.pt._key = req->events;
1918	apt.iocb = aiocb;
1919	apt.queued = false;
1920	apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1921
1922	/* initialized the list so that we can do list_empty checks */
1923	INIT_LIST_HEAD(&req->wait.entry);
1924	init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1925
1926	mask = vfs_poll(req->file, &apt.pt) & req->events;
1927	spin_lock_irq(&ctx->ctx_lock);
1928	if (likely(apt.queued)) {
1929		bool on_queue = poll_iocb_lock_wq(req);
1930
1931		if (!on_queue || req->work_scheduled) {
1932			/*
1933			 * aio_poll_wake() already either scheduled the async
1934			 * completion work, or completed the request inline.
1935			 */
1936			if (apt.error) /* unsupported case: multiple queues */
1937				cancel = true;
1938			apt.error = 0;
1939			mask = 0;
1940		}
1941		if (mask || apt.error) {
1942			/* Steal to complete synchronously. */
1943			list_del_init(&req->wait.entry);
1944		} else if (cancel) {
1945			/* Cancel if possible (may be too late though). */
1946			WRITE_ONCE(req->cancelled, true);
1947		} else if (on_queue) {
1948			/*
1949			 * Actually waiting for an event, so add the request to
1950			 * active_reqs so that it can be cancelled if needed.
1951			 */
1952			list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1953			aiocb->ki_cancel = aio_poll_cancel;
1954		}
1955		if (on_queue)
1956			poll_iocb_unlock_wq(req);
1957	}
1958	if (mask) { /* no async, we'd stolen it */
1959		aiocb->ki_res.res = mangle_poll(mask);
1960		apt.error = 0;
1961	}
1962	spin_unlock_irq(&ctx->ctx_lock);
1963	if (mask)
1964		iocb_put(aiocb);
1965	return apt.error;
1966}
1967
1968static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1969			   struct iocb __user *user_iocb, struct aio_kiocb *req,
1970			   bool compat)
1971{
1972	req->ki_filp = fget(iocb->aio_fildes);
1973	if (unlikely(!req->ki_filp))
1974		return -EBADF;
1975
1976	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1977		struct eventfd_ctx *eventfd;
1978		/*
1979		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1980		 * instance of the file* now. The file descriptor must be
1981		 * an eventfd() fd, and will be signaled for each completed
1982		 * event using the eventfd_signal() function.
1983		 */
1984		eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1985		if (IS_ERR(eventfd))
1986			return PTR_ERR(eventfd);
1987
1988		req->ki_eventfd = eventfd;
1989	}
1990
1991	if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1992		pr_debug("EFAULT: aio_key\n");
1993		return -EFAULT;
1994	}
1995
1996	req->ki_res.obj = (u64)(unsigned long)user_iocb;
1997	req->ki_res.data = iocb->aio_data;
1998	req->ki_res.res = 0;
1999	req->ki_res.res2 = 0;
2000
2001	switch (iocb->aio_lio_opcode) {
2002	case IOCB_CMD_PREAD:
2003		return aio_read(&req->rw, iocb, false, compat);
2004	case IOCB_CMD_PWRITE:
2005		return aio_write(&req->rw, iocb, false, compat);
2006	case IOCB_CMD_PREADV:
2007		return aio_read(&req->rw, iocb, true, compat);
2008	case IOCB_CMD_PWRITEV:
2009		return aio_write(&req->rw, iocb, true, compat);
2010	case IOCB_CMD_FSYNC:
2011		return aio_fsync(&req->fsync, iocb, false);
2012	case IOCB_CMD_FDSYNC:
2013		return aio_fsync(&req->fsync, iocb, true);
2014	case IOCB_CMD_POLL:
2015		return aio_poll(req, iocb);
2016	default:
2017		pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
2018		return -EINVAL;
2019	}
2020}
2021
2022static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
2023			 bool compat)
2024{
2025	struct aio_kiocb *req;
2026	struct iocb iocb;
2027	int err;
2028
2029	if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
2030		return -EFAULT;
2031
2032	/* enforce forwards compatibility on users */
2033	if (unlikely(iocb.aio_reserved2)) {
2034		pr_debug("EINVAL: reserve field set\n");
2035		return -EINVAL;
2036	}
2037
2038	/* prevent overflows */
2039	if (unlikely(
2040	    (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
2041	    (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
2042	    ((ssize_t)iocb.aio_nbytes < 0)
2043	   )) {
2044		pr_debug("EINVAL: overflow check\n");
2045		return -EINVAL;
2046	}
2047
2048	req = aio_get_req(ctx);
2049	if (unlikely(!req))
2050		return -EAGAIN;
2051
2052	err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2053
2054	/* Done with the synchronous reference */
2055	iocb_put(req);
2056
2057	/*
2058	 * If err is 0, we'd either done aio_complete() ourselves or have
2059	 * arranged for that to be done asynchronously.  Anything non-zero
2060	 * means that we need to destroy req ourselves.
2061	 */
2062	if (unlikely(err)) {
2063		iocb_destroy(req);
2064		put_reqs_available(ctx, 1);
2065	}
2066	return err;
2067}
2068
2069/* sys_io_submit:
2070 *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
2071 *	the number of iocbs queued.  May return -EINVAL if the aio_context
2072 *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
2073 *	*iocbpp[0] is not properly initialized, if the operation specified
2074 *	is invalid for the file descriptor in the iocb.  May fail with
2075 *	-EFAULT if any of the data structures point to invalid data.  May
2076 *	fail with -EBADF if the file descriptor specified in the first
2077 *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
2078 *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
2079 *	fail with -ENOSYS if not implemented.
2080 */
2081SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2082		struct iocb __user * __user *, iocbpp)
2083{
2084	struct kioctx *ctx;
2085	long ret = 0;
2086	int i = 0;
2087	struct blk_plug plug;
2088
2089	if (unlikely(nr < 0))
2090		return -EINVAL;
2091
2092	ctx = lookup_ioctx(ctx_id);
2093	if (unlikely(!ctx)) {
2094		pr_debug("EINVAL: invalid context id\n");
2095		return -EINVAL;
2096	}
2097
2098	if (nr > ctx->nr_events)
2099		nr = ctx->nr_events;
2100
2101	if (nr > AIO_PLUG_THRESHOLD)
2102		blk_start_plug(&plug);
2103	for (i = 0; i < nr; i++) {
2104		struct iocb __user *user_iocb;
2105
2106		if (unlikely(get_user(user_iocb, iocbpp + i))) {
2107			ret = -EFAULT;
2108			break;
2109		}
2110
2111		ret = io_submit_one(ctx, user_iocb, false);
2112		if (ret)
2113			break;
2114	}
2115	if (nr > AIO_PLUG_THRESHOLD)
2116		blk_finish_plug(&plug);
2117
2118	percpu_ref_put(&ctx->users);
2119	return i ? i : ret;
2120}
2121
2122#ifdef CONFIG_COMPAT
2123COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2124		       int, nr, compat_uptr_t __user *, iocbpp)
2125{
2126	struct kioctx *ctx;
2127	long ret = 0;
2128	int i = 0;
2129	struct blk_plug plug;
2130
2131	if (unlikely(nr < 0))
2132		return -EINVAL;
2133
2134	ctx = lookup_ioctx(ctx_id);
2135	if (unlikely(!ctx)) {
2136		pr_debug("EINVAL: invalid context id\n");
2137		return -EINVAL;
2138	}
2139
2140	if (nr > ctx->nr_events)
2141		nr = ctx->nr_events;
2142
2143	if (nr > AIO_PLUG_THRESHOLD)
2144		blk_start_plug(&plug);
2145	for (i = 0; i < nr; i++) {
2146		compat_uptr_t user_iocb;
2147
2148		if (unlikely(get_user(user_iocb, iocbpp + i))) {
2149			ret = -EFAULT;
2150			break;
2151		}
2152
2153		ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2154		if (ret)
2155			break;
2156	}
2157	if (nr > AIO_PLUG_THRESHOLD)
2158		blk_finish_plug(&plug);
2159
2160	percpu_ref_put(&ctx->users);
2161	return i ? i : ret;
2162}
2163#endif
2164
2165/* sys_io_cancel:
2166 *	Attempts to cancel an iocb previously passed to io_submit.  If
2167 *	the operation is successfully cancelled, the resulting event is
2168 *	copied into the memory pointed to by result without being placed
2169 *	into the completion queue and 0 is returned.  May fail with
2170 *	-EFAULT if any of the data structures pointed to are invalid.
2171 *	May fail with -EINVAL if aio_context specified by ctx_id is
2172 *	invalid.  May fail with -EAGAIN if the iocb specified was not
2173 *	cancelled.  Will fail with -ENOSYS if not implemented.
2174 */
2175SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2176		struct io_event __user *, result)
2177{
2178	struct kioctx *ctx;
2179	struct aio_kiocb *kiocb;
2180	int ret = -EINVAL;
2181	u32 key;
2182	u64 obj = (u64)(unsigned long)iocb;
2183
2184	if (unlikely(get_user(key, &iocb->aio_key)))
2185		return -EFAULT;
2186	if (unlikely(key != KIOCB_KEY))
2187		return -EINVAL;
2188
2189	ctx = lookup_ioctx(ctx_id);
2190	if (unlikely(!ctx))
2191		return -EINVAL;
2192
2193	spin_lock_irq(&ctx->ctx_lock);
 
2194	list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2195		if (kiocb->ki_res.obj == obj) {
2196			ret = kiocb->ki_cancel(&kiocb->rw);
2197			list_del_init(&kiocb->ki_list);
2198			break;
2199		}
2200	}
2201	spin_unlock_irq(&ctx->ctx_lock);
2202
2203	if (!ret) {
2204		/*
2205		 * The result argument is no longer used - the io_event is
2206		 * always delivered via the ring buffer. -EINPROGRESS indicates
2207		 * cancellation is progress:
2208		 */
2209		ret = -EINPROGRESS;
2210	}
2211
2212	percpu_ref_put(&ctx->users);
2213
2214	return ret;
2215}
2216
2217static long do_io_getevents(aio_context_t ctx_id,
2218		long min_nr,
2219		long nr,
2220		struct io_event __user *events,
2221		struct timespec64 *ts)
2222{
2223	ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2224	struct kioctx *ioctx = lookup_ioctx(ctx_id);
2225	long ret = -EINVAL;
2226
2227	if (likely(ioctx)) {
2228		if (likely(min_nr <= nr && min_nr >= 0))
2229			ret = read_events(ioctx, min_nr, nr, events, until);
2230		percpu_ref_put(&ioctx->users);
2231	}
2232
2233	return ret;
2234}
2235
2236/* io_getevents:
2237 *	Attempts to read at least min_nr events and up to nr events from
2238 *	the completion queue for the aio_context specified by ctx_id. If
2239 *	it succeeds, the number of read events is returned. May fail with
2240 *	-EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2241 *	out of range, if timeout is out of range.  May fail with -EFAULT
2242 *	if any of the memory specified is invalid.  May return 0 or
2243 *	< min_nr if the timeout specified by timeout has elapsed
2244 *	before sufficient events are available, where timeout == NULL
2245 *	specifies an infinite timeout. Note that the timeout pointed to by
2246 *	timeout is relative.  Will fail with -ENOSYS if not implemented.
2247 */
2248#ifdef CONFIG_64BIT
2249
2250SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2251		long, min_nr,
2252		long, nr,
2253		struct io_event __user *, events,
2254		struct __kernel_timespec __user *, timeout)
2255{
2256	struct timespec64	ts;
2257	int			ret;
2258
2259	if (timeout && unlikely(get_timespec64(&ts, timeout)))
2260		return -EFAULT;
2261
2262	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2263	if (!ret && signal_pending(current))
2264		ret = -EINTR;
2265	return ret;
2266}
2267
2268#endif
2269
2270struct __aio_sigset {
2271	const sigset_t __user	*sigmask;
2272	size_t		sigsetsize;
2273};
2274
2275SYSCALL_DEFINE6(io_pgetevents,
2276		aio_context_t, ctx_id,
2277		long, min_nr,
2278		long, nr,
2279		struct io_event __user *, events,
2280		struct __kernel_timespec __user *, timeout,
2281		const struct __aio_sigset __user *, usig)
2282{
2283	struct __aio_sigset	ksig = { NULL, };
2284	struct timespec64	ts;
2285	bool interrupted;
2286	int ret;
2287
2288	if (timeout && unlikely(get_timespec64(&ts, timeout)))
2289		return -EFAULT;
2290
2291	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2292		return -EFAULT;
2293
2294	ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2295	if (ret)
2296		return ret;
2297
2298	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2299
2300	interrupted = signal_pending(current);
2301	restore_saved_sigmask_unless(interrupted);
2302	if (interrupted && !ret)
2303		ret = -ERESTARTNOHAND;
2304
2305	return ret;
2306}
2307
2308#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2309
2310SYSCALL_DEFINE6(io_pgetevents_time32,
2311		aio_context_t, ctx_id,
2312		long, min_nr,
2313		long, nr,
2314		struct io_event __user *, events,
2315		struct old_timespec32 __user *, timeout,
2316		const struct __aio_sigset __user *, usig)
2317{
2318	struct __aio_sigset	ksig = { NULL, };
2319	struct timespec64	ts;
2320	bool interrupted;
2321	int ret;
2322
2323	if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2324		return -EFAULT;
2325
2326	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2327		return -EFAULT;
2328
2329
2330	ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2331	if (ret)
2332		return ret;
2333
2334	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2335
2336	interrupted = signal_pending(current);
2337	restore_saved_sigmask_unless(interrupted);
2338	if (interrupted && !ret)
2339		ret = -ERESTARTNOHAND;
2340
2341	return ret;
2342}
2343
2344#endif
2345
2346#if defined(CONFIG_COMPAT_32BIT_TIME)
2347
2348SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2349		__s32, min_nr,
2350		__s32, nr,
2351		struct io_event __user *, events,
2352		struct old_timespec32 __user *, timeout)
2353{
2354	struct timespec64 t;
2355	int ret;
2356
2357	if (timeout && get_old_timespec32(&t, timeout))
2358		return -EFAULT;
2359
2360	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2361	if (!ret && signal_pending(current))
2362		ret = -EINTR;
2363	return ret;
2364}
2365
2366#endif
2367
2368#ifdef CONFIG_COMPAT
2369
2370struct __compat_aio_sigset {
2371	compat_uptr_t		sigmask;
2372	compat_size_t		sigsetsize;
2373};
2374
2375#if defined(CONFIG_COMPAT_32BIT_TIME)
2376
2377COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2378		compat_aio_context_t, ctx_id,
2379		compat_long_t, min_nr,
2380		compat_long_t, nr,
2381		struct io_event __user *, events,
2382		struct old_timespec32 __user *, timeout,
2383		const struct __compat_aio_sigset __user *, usig)
2384{
2385	struct __compat_aio_sigset ksig = { 0, };
2386	struct timespec64 t;
2387	bool interrupted;
2388	int ret;
2389
2390	if (timeout && get_old_timespec32(&t, timeout))
2391		return -EFAULT;
2392
2393	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2394		return -EFAULT;
2395
2396	ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2397	if (ret)
2398		return ret;
2399
2400	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2401
2402	interrupted = signal_pending(current);
2403	restore_saved_sigmask_unless(interrupted);
2404	if (interrupted && !ret)
2405		ret = -ERESTARTNOHAND;
2406
2407	return ret;
2408}
2409
2410#endif
2411
2412COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2413		compat_aio_context_t, ctx_id,
2414		compat_long_t, min_nr,
2415		compat_long_t, nr,
2416		struct io_event __user *, events,
2417		struct __kernel_timespec __user *, timeout,
2418		const struct __compat_aio_sigset __user *, usig)
2419{
2420	struct __compat_aio_sigset ksig = { 0, };
2421	struct timespec64 t;
2422	bool interrupted;
2423	int ret;
2424
2425	if (timeout && get_timespec64(&t, timeout))
2426		return -EFAULT;
2427
2428	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2429		return -EFAULT;
2430
2431	ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2432	if (ret)
2433		return ret;
2434
2435	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2436
2437	interrupted = signal_pending(current);
2438	restore_saved_sigmask_unless(interrupted);
2439	if (interrupted && !ret)
2440		ret = -ERESTARTNOHAND;
2441
2442	return ret;
2443}
2444#endif