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