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 *
   9 *	See ../COPYING for licensing terms.
  10 */
  11#define pr_fmt(fmt) "%s: " fmt, __func__
  12
  13#include <linux/kernel.h>
  14#include <linux/init.h>
  15#include <linux/errno.h>
  16#include <linux/time.h>
  17#include <linux/aio_abi.h>
  18#include <linux/export.h>
  19#include <linux/syscalls.h>
  20#include <linux/backing-dev.h>
  21#include <linux/uio.h>
  22
  23#include <linux/sched/signal.h>
  24#include <linux/fs.h>
  25#include <linux/file.h>
  26#include <linux/mm.h>
  27#include <linux/mman.h>
  28#include <linux/mmu_context.h>
  29#include <linux/percpu.h>
  30#include <linux/slab.h>
  31#include <linux/timer.h>
  32#include <linux/aio.h>
  33#include <linux/highmem.h>
  34#include <linux/workqueue.h>
  35#include <linux/security.h>
  36#include <linux/eventfd.h>
  37#include <linux/blkdev.h>
  38#include <linux/compat.h>
  39#include <linux/migrate.h>
  40#include <linux/ramfs.h>
  41#include <linux/percpu-refcount.h>
  42#include <linux/mount.h>
  43
  44#include <asm/kmap_types.h>
  45#include <linux/uaccess.h>
  46
  47#include "internal.h"
  48
  49#define AIO_RING_MAGIC			0xa10a10a1
  50#define AIO_RING_COMPAT_FEATURES	1
  51#define AIO_RING_INCOMPAT_FEATURES	0
  52struct aio_ring {
  53	unsigned	id;	/* kernel internal index number */
  54	unsigned	nr;	/* number of io_events */
  55	unsigned	head;	/* Written to by userland or under ring_lock
  56				 * mutex by aio_read_events_ring(). */
  57	unsigned	tail;
  58
  59	unsigned	magic;
  60	unsigned	compat_features;
  61	unsigned	incompat_features;
  62	unsigned	header_length;	/* size of aio_ring */
  63
  64
  65	struct io_event		io_events[0];
  66}; /* 128 bytes + ring size */
  67
  68#define AIO_RING_PAGES	8
  69
  70struct kioctx_table {
  71	struct rcu_head		rcu;
  72	unsigned		nr;
  73	struct kioctx __rcu	*table[];
  74};
  75
  76struct kioctx_cpu {
  77	unsigned		reqs_available;
  78};
  79
  80struct ctx_rq_wait {
  81	struct completion comp;
  82	atomic_t count;
  83};
  84
  85struct kioctx {
  86	struct percpu_ref	users;
  87	atomic_t		dead;
  88
  89	struct percpu_ref	reqs;
  90
  91	unsigned long		user_id;
  92
  93	struct __percpu kioctx_cpu *cpu;
  94
  95	/*
  96	 * For percpu reqs_available, number of slots we move to/from global
  97	 * counter at a time:
  98	 */
  99	unsigned		req_batch;
 100	/*
 101	 * This is what userspace passed to io_setup(), it's not used for
 102	 * anything but counting against the global max_reqs quota.
 103	 *
 104	 * The real limit is nr_events - 1, which will be larger (see
 105	 * aio_setup_ring())
 106	 */
 107	unsigned		max_reqs;
 108
 109	/* Size of ringbuffer, in units of struct io_event */
 110	unsigned		nr_events;
 111
 112	unsigned long		mmap_base;
 113	unsigned long		mmap_size;
 114
 115	struct page		**ring_pages;
 116	long			nr_pages;
 117
 118	struct rcu_work		free_rwork;	/* see free_ioctx() */
 119
 120	/*
 121	 * signals when all in-flight requests are done
 122	 */
 123	struct ctx_rq_wait	*rq_wait;
 124
 125	struct {
 126		/*
 127		 * This counts the number of available slots in the ringbuffer,
 128		 * so we avoid overflowing it: it's decremented (if positive)
 129		 * when allocating a kiocb and incremented when the resulting
 130		 * io_event is pulled off the ringbuffer.
 131		 *
 132		 * We batch accesses to it with a percpu version.
 133		 */
 134		atomic_t	reqs_available;
 135	} ____cacheline_aligned_in_smp;
 136
 137	struct {
 138		spinlock_t	ctx_lock;
 139		struct list_head active_reqs;	/* used for cancellation */
 140	} ____cacheline_aligned_in_smp;
 141
 142	struct {
 143		struct mutex	ring_lock;
 144		wait_queue_head_t wait;
 145	} ____cacheline_aligned_in_smp;
 146
 147	struct {
 148		unsigned	tail;
 149		unsigned	completed_events;
 150		spinlock_t	completion_lock;
 151	} ____cacheline_aligned_in_smp;
 152
 153	struct page		*internal_pages[AIO_RING_PAGES];
 154	struct file		*aio_ring_file;
 155
 156	unsigned		id;
 157};
 158
 159/*
 160 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
 161 * cancelled or completed (this makes a certain amount of sense because
 162 * successful cancellation - io_cancel() - does deliver the completion to
 163 * userspace).
 164 *
 165 * And since most things don't implement kiocb cancellation and we'd really like
 166 * kiocb completion to be lockless when possible, we use ki_cancel to
 167 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
 168 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
 169 */
 170#define KIOCB_CANCELLED		((void *) (~0ULL))
 171
 172struct aio_kiocb {
 173	struct kiocb		common;
 174
 175	struct kioctx		*ki_ctx;
 176	kiocb_cancel_fn		*ki_cancel;
 177
 178	struct iocb __user	*ki_user_iocb;	/* user's aiocb */
 179	__u64			ki_user_data;	/* user's data for completion */
 180
 181	struct list_head	ki_list;	/* the aio core uses this
 182						 * for cancellation */
 183
 184	/*
 185	 * If the aio_resfd field of the userspace iocb is not zero,
 186	 * this is the underlying eventfd context to deliver events to.
 187	 */
 188	struct eventfd_ctx	*ki_eventfd;
 189};
 190
 191/*------ sysctl variables----*/
 192static DEFINE_SPINLOCK(aio_nr_lock);
 193unsigned long aio_nr;		/* current system wide number of aio requests */
 194unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
 195/*----end sysctl variables---*/
 196
 197static struct kmem_cache	*kiocb_cachep;
 198static struct kmem_cache	*kioctx_cachep;
 199
 200static struct vfsmount *aio_mnt;
 201
 202static const struct file_operations aio_ring_fops;
 203static const struct address_space_operations aio_ctx_aops;
 204
 205static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
 206{
 207	struct qstr this = QSTR_INIT("[aio]", 5);
 208	struct file *file;
 209	struct path path;
 210	struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
 211	if (IS_ERR(inode))
 212		return ERR_CAST(inode);
 213
 214	inode->i_mapping->a_ops = &aio_ctx_aops;
 215	inode->i_mapping->private_data = ctx;
 216	inode->i_size = PAGE_SIZE * nr_pages;
 217
 218	path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
 219	if (!path.dentry) {
 220		iput(inode);
 221		return ERR_PTR(-ENOMEM);
 222	}
 223	path.mnt = mntget(aio_mnt);
 224
 225	d_instantiate(path.dentry, inode);
 226	file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
 227	if (IS_ERR(file)) {
 228		path_put(&path);
 229		return file;
 230	}
 231
 232	file->f_flags = O_RDWR;
 
 233	return file;
 234}
 235
 236static struct dentry *aio_mount(struct file_system_type *fs_type,
 237				int flags, const char *dev_name, void *data)
 238{
 239	static const struct dentry_operations ops = {
 240		.d_dname	= simple_dname,
 241	};
 242	struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops,
 243					   AIO_RING_MAGIC);
 244
 245	if (!IS_ERR(root))
 246		root->d_sb->s_iflags |= SB_I_NOEXEC;
 247	return root;
 248}
 249
 250/* aio_setup
 251 *	Creates the slab caches used by the aio routines, panic on
 252 *	failure as this is done early during the boot sequence.
 253 */
 254static int __init aio_setup(void)
 255{
 256	static struct file_system_type aio_fs = {
 257		.name		= "aio",
 258		.mount		= aio_mount,
 259		.kill_sb	= kill_anon_super,
 260	};
 261	aio_mnt = kern_mount(&aio_fs);
 262	if (IS_ERR(aio_mnt))
 263		panic("Failed to create aio fs mount.");
 264
 265	kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
 266	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
 267
 268	pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
 269
 270	return 0;
 271}
 272__initcall(aio_setup);
 273
 274static void put_aio_ring_file(struct kioctx *ctx)
 275{
 276	struct file *aio_ring_file = ctx->aio_ring_file;
 277	struct address_space *i_mapping;
 278
 279	if (aio_ring_file) {
 280		truncate_setsize(file_inode(aio_ring_file), 0);
 281
 282		/* Prevent further access to the kioctx from migratepages */
 283		i_mapping = aio_ring_file->f_mapping;
 284		spin_lock(&i_mapping->private_lock);
 285		i_mapping->private_data = NULL;
 286		ctx->aio_ring_file = NULL;
 287		spin_unlock(&i_mapping->private_lock);
 288
 289		fput(aio_ring_file);
 290	}
 291}
 292
 293static void aio_free_ring(struct kioctx *ctx)
 294{
 295	int i;
 296
 297	/* Disconnect the kiotx from the ring file.  This prevents future
 298	 * accesses to the kioctx from page migration.
 299	 */
 300	put_aio_ring_file(ctx);
 301
 302	for (i = 0; i < ctx->nr_pages; i++) {
 303		struct page *page;
 304		pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
 305				page_count(ctx->ring_pages[i]));
 306		page = ctx->ring_pages[i];
 307		if (!page)
 308			continue;
 309		ctx->ring_pages[i] = NULL;
 310		put_page(page);
 311	}
 312
 313	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
 314		kfree(ctx->ring_pages);
 315		ctx->ring_pages = NULL;
 316	}
 317}
 318
 319static int aio_ring_mremap(struct vm_area_struct *vma)
 320{
 321	struct file *file = vma->vm_file;
 322	struct mm_struct *mm = vma->vm_mm;
 323	struct kioctx_table *table;
 324	int i, res = -EINVAL;
 325
 326	spin_lock(&mm->ioctx_lock);
 327	rcu_read_lock();
 328	table = rcu_dereference(mm->ioctx_table);
 329	for (i = 0; i < table->nr; i++) {
 330		struct kioctx *ctx;
 331
 332		ctx = rcu_dereference(table->table[i]);
 333		if (ctx && ctx->aio_ring_file == file) {
 334			if (!atomic_read(&ctx->dead)) {
 335				ctx->user_id = ctx->mmap_base = vma->vm_start;
 336				res = 0;
 337			}
 338			break;
 339		}
 340	}
 341
 342	rcu_read_unlock();
 343	spin_unlock(&mm->ioctx_lock);
 344	return res;
 345}
 346
 347static const struct vm_operations_struct aio_ring_vm_ops = {
 348	.mremap		= aio_ring_mremap,
 349#if IS_ENABLED(CONFIG_MMU)
 350	.fault		= filemap_fault,
 351	.map_pages	= filemap_map_pages,
 352	.page_mkwrite	= filemap_page_mkwrite,
 353#endif
 354};
 355
 356static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
 357{
 358	vma->vm_flags |= VM_DONTEXPAND;
 359	vma->vm_ops = &aio_ring_vm_ops;
 360	return 0;
 361}
 362
 363static const struct file_operations aio_ring_fops = {
 364	.mmap = aio_ring_mmap,
 365};
 366
 
 
 
 
 
 367#if IS_ENABLED(CONFIG_MIGRATION)
 368static int aio_migratepage(struct address_space *mapping, struct page *new,
 369			struct page *old, enum migrate_mode mode)
 370{
 371	struct kioctx *ctx;
 372	unsigned long flags;
 373	pgoff_t idx;
 374	int rc;
 375
 376	/*
 377	 * We cannot support the _NO_COPY case here, because copy needs to
 378	 * happen under the ctx->completion_lock. That does not work with the
 379	 * migration workflow of MIGRATE_SYNC_NO_COPY.
 380	 */
 381	if (mode == MIGRATE_SYNC_NO_COPY)
 382		return -EINVAL;
 383
 384	rc = 0;
 385
 386	/* mapping->private_lock here protects against the kioctx teardown.  */
 387	spin_lock(&mapping->private_lock);
 388	ctx = mapping->private_data;
 389	if (!ctx) {
 390		rc = -EINVAL;
 391		goto out;
 392	}
 393
 394	/* The ring_lock mutex.  The prevents aio_read_events() from writing
 395	 * to the ring's head, and prevents page migration from mucking in
 396	 * a partially initialized kiotx.
 397	 */
 398	if (!mutex_trylock(&ctx->ring_lock)) {
 399		rc = -EAGAIN;
 400		goto out;
 401	}
 402
 403	idx = old->index;
 404	if (idx < (pgoff_t)ctx->nr_pages) {
 405		/* Make sure the old page hasn't already been changed */
 406		if (ctx->ring_pages[idx] != old)
 407			rc = -EAGAIN;
 408	} else
 409		rc = -EINVAL;
 410
 411	if (rc != 0)
 412		goto out_unlock;
 413
 414	/* Writeback must be complete */
 415	BUG_ON(PageWriteback(old));
 416	get_page(new);
 417
 418	rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
 419	if (rc != MIGRATEPAGE_SUCCESS) {
 420		put_page(new);
 421		goto out_unlock;
 422	}
 423
 424	/* Take completion_lock to prevent other writes to the ring buffer
 425	 * while the old page is copied to the new.  This prevents new
 426	 * events from being lost.
 427	 */
 428	spin_lock_irqsave(&ctx->completion_lock, flags);
 429	migrate_page_copy(new, old);
 430	BUG_ON(ctx->ring_pages[idx] != old);
 431	ctx->ring_pages[idx] = new;
 432	spin_unlock_irqrestore(&ctx->completion_lock, flags);
 433
 434	/* The old page is no longer accessible. */
 435	put_page(old);
 436
 437out_unlock:
 438	mutex_unlock(&ctx->ring_lock);
 439out:
 440	spin_unlock(&mapping->private_lock);
 441	return rc;
 442}
 443#endif
 444
 445static const struct address_space_operations aio_ctx_aops = {
 446	.set_page_dirty = __set_page_dirty_no_writeback,
 447#if IS_ENABLED(CONFIG_MIGRATION)
 448	.migratepage	= aio_migratepage,
 449#endif
 450};
 451
 452static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
 453{
 454	struct aio_ring *ring;
 
 455	struct mm_struct *mm = current->mm;
 456	unsigned long size, unused;
 457	int nr_pages;
 458	int i;
 459	struct file *file;
 460
 461	/* Compensate for the ring buffer's head/tail overlap entry */
 462	nr_events += 2;	/* 1 is required, 2 for good luck */
 463
 464	size = sizeof(struct aio_ring);
 465	size += sizeof(struct io_event) * nr_events;
 466
 467	nr_pages = PFN_UP(size);
 468	if (nr_pages < 0)
 469		return -EINVAL;
 470
 471	file = aio_private_file(ctx, nr_pages);
 472	if (IS_ERR(file)) {
 473		ctx->aio_ring_file = NULL;
 474		return -ENOMEM;
 475	}
 476
 477	ctx->aio_ring_file = file;
 478	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
 479			/ sizeof(struct io_event);
 480
 481	ctx->ring_pages = ctx->internal_pages;
 482	if (nr_pages > AIO_RING_PAGES) {
 483		ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
 484					  GFP_KERNEL);
 485		if (!ctx->ring_pages) {
 486			put_aio_ring_file(ctx);
 487			return -ENOMEM;
 488		}
 489	}
 490
 491	for (i = 0; i < nr_pages; i++) {
 492		struct page *page;
 493		page = find_or_create_page(file->f_mapping,
 494					   i, GFP_HIGHUSER | __GFP_ZERO);
 495		if (!page)
 496			break;
 497		pr_debug("pid(%d) page[%d]->count=%d\n",
 498			 current->pid, i, page_count(page));
 499		SetPageUptodate(page);
 
 500		unlock_page(page);
 501
 502		ctx->ring_pages[i] = page;
 503	}
 504	ctx->nr_pages = i;
 505
 506	if (unlikely(i != nr_pages)) {
 507		aio_free_ring(ctx);
 508		return -ENOMEM;
 509	}
 510
 511	ctx->mmap_size = nr_pages * PAGE_SIZE;
 512	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
 513
 514	if (down_write_killable(&mm->mmap_sem)) {
 515		ctx->mmap_size = 0;
 516		aio_free_ring(ctx);
 517		return -EINTR;
 518	}
 519
 520	ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
 521				       PROT_READ | PROT_WRITE,
 522				       MAP_SHARED, 0, &unused, NULL);
 523	up_write(&mm->mmap_sem);
 524	if (IS_ERR((void *)ctx->mmap_base)) {
 525		ctx->mmap_size = 0;
 526		aio_free_ring(ctx);
 527		return -ENOMEM;
 528	}
 529
 530	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
 531
 532	ctx->user_id = ctx->mmap_base;
 533	ctx->nr_events = nr_events; /* trusted copy */
 534
 535	ring = kmap_atomic(ctx->ring_pages[0]);
 536	ring->nr = nr_events;	/* user copy */
 537	ring->id = ~0U;
 538	ring->head = ring->tail = 0;
 539	ring->magic = AIO_RING_MAGIC;
 540	ring->compat_features = AIO_RING_COMPAT_FEATURES;
 541	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
 542	ring->header_length = sizeof(struct aio_ring);
 543	kunmap_atomic(ring);
 544	flush_dcache_page(ctx->ring_pages[0]);
 545
 546	return 0;
 547}
 548
 549#define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
 550#define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
 551#define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
 552
 553void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
 554{
 555	struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
 556	struct kioctx *ctx = req->ki_ctx;
 557	unsigned long flags;
 558
 559	spin_lock_irqsave(&ctx->ctx_lock, flags);
 560
 561	if (!req->ki_list.next)
 562		list_add(&req->ki_list, &ctx->active_reqs);
 563
 564	req->ki_cancel = cancel;
 565
 566	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 567}
 568EXPORT_SYMBOL(kiocb_set_cancel_fn);
 569
 570static int kiocb_cancel(struct aio_kiocb *kiocb)
 571{
 572	kiocb_cancel_fn *old, *cancel;
 573
 574	/*
 575	 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
 576	 * actually has a cancel function, hence the cmpxchg()
 577	 */
 578
 579	cancel = READ_ONCE(kiocb->ki_cancel);
 580	do {
 581		if (!cancel || cancel == KIOCB_CANCELLED)
 582			return -EINVAL;
 583
 584		old = cancel;
 585		cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
 586	} while (cancel != old);
 587
 588	return cancel(&kiocb->common);
 589}
 590
 591/*
 592 * free_ioctx() should be RCU delayed to synchronize against the RCU
 593 * protected lookup_ioctx() and also needs process context to call
 594 * aio_free_ring().  Use rcu_work.
 595 */
 596static void free_ioctx(struct work_struct *work)
 597{
 598	struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
 599					  free_rwork);
 600	pr_debug("freeing %p\n", ctx);
 601
 602	aio_free_ring(ctx);
 603	free_percpu(ctx->cpu);
 604	percpu_ref_exit(&ctx->reqs);
 605	percpu_ref_exit(&ctx->users);
 606	kmem_cache_free(kioctx_cachep, ctx);
 607}
 608
 609static void free_ioctx_reqs(struct percpu_ref *ref)
 610{
 611	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
 612
 613	/* At this point we know that there are no any in-flight requests */
 614	if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
 615		complete(&ctx->rq_wait->comp);
 616
 617	/* Synchronize against RCU protected table->table[] dereferences */
 618	INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
 619	queue_rcu_work(system_wq, &ctx->free_rwork);
 620}
 621
 622/*
 623 * When this function runs, the kioctx has been removed from the "hash table"
 624 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
 625 * now it's safe to cancel any that need to be.
 626 */
 627static void free_ioctx_users(struct percpu_ref *ref)
 628{
 629	struct kioctx *ctx = container_of(ref, struct kioctx, users);
 630	struct aio_kiocb *req;
 631
 632	spin_lock_irq(&ctx->ctx_lock);
 633
 634	while (!list_empty(&ctx->active_reqs)) {
 635		req = list_first_entry(&ctx->active_reqs,
 636				       struct aio_kiocb, ki_list);
 637		kiocb_cancel(req);
 638		list_del_init(&req->ki_list);
 
 639	}
 640
 641	spin_unlock_irq(&ctx->ctx_lock);
 642
 643	percpu_ref_kill(&ctx->reqs);
 644	percpu_ref_put(&ctx->reqs);
 645}
 646
 647static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
 648{
 649	unsigned i, new_nr;
 650	struct kioctx_table *table, *old;
 651	struct aio_ring *ring;
 652
 653	spin_lock(&mm->ioctx_lock);
 654	table = rcu_dereference_raw(mm->ioctx_table);
 
 655
 656	while (1) {
 657		if (table)
 658			for (i = 0; i < table->nr; i++)
 659				if (!rcu_access_pointer(table->table[i])) {
 660					ctx->id = i;
 661					rcu_assign_pointer(table->table[i], ctx);
 
 662					spin_unlock(&mm->ioctx_lock);
 663
 664					/* While kioctx setup is in progress,
 665					 * we are protected from page migration
 666					 * changes ring_pages by ->ring_lock.
 667					 */
 668					ring = kmap_atomic(ctx->ring_pages[0]);
 669					ring->id = ctx->id;
 670					kunmap_atomic(ring);
 671					return 0;
 672				}
 673
 674		new_nr = (table ? table->nr : 1) * 4;
 
 
 675		spin_unlock(&mm->ioctx_lock);
 676
 677		table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
 678				new_nr, GFP_KERNEL);
 679		if (!table)
 680			return -ENOMEM;
 681
 682		table->nr = new_nr;
 683
 684		spin_lock(&mm->ioctx_lock);
 685		old = rcu_dereference_raw(mm->ioctx_table);
 
 686
 687		if (!old) {
 688			rcu_assign_pointer(mm->ioctx_table, table);
 689		} else if (table->nr > old->nr) {
 690			memcpy(table->table, old->table,
 691			       old->nr * sizeof(struct kioctx *));
 692
 693			rcu_assign_pointer(mm->ioctx_table, table);
 694			kfree_rcu(old, rcu);
 695		} else {
 696			kfree(table);
 697			table = old;
 698		}
 699	}
 700}
 701
 702static void aio_nr_sub(unsigned nr)
 703{
 704	spin_lock(&aio_nr_lock);
 705	if (WARN_ON(aio_nr - nr > aio_nr))
 706		aio_nr = 0;
 707	else
 708		aio_nr -= nr;
 709	spin_unlock(&aio_nr_lock);
 710}
 711
 712/* ioctx_alloc
 713 *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
 714 */
 715static struct kioctx *ioctx_alloc(unsigned nr_events)
 716{
 717	struct mm_struct *mm = current->mm;
 718	struct kioctx *ctx;
 719	int err = -ENOMEM;
 720
 721	/*
 722	 * Store the original nr_events -- what userspace passed to io_setup(),
 723	 * for counting against the global limit -- before it changes.
 724	 */
 725	unsigned int max_reqs = nr_events;
 726
 727	/*
 728	 * We keep track of the number of available ringbuffer slots, to prevent
 729	 * overflow (reqs_available), and we also use percpu counters for this.
 730	 *
 731	 * So since up to half the slots might be on other cpu's percpu counters
 732	 * and unavailable, double nr_events so userspace sees what they
 733	 * expected: additionally, we move req_batch slots to/from percpu
 734	 * counters at a time, so make sure that isn't 0:
 735	 */
 736	nr_events = max(nr_events, num_possible_cpus() * 4);
 737	nr_events *= 2;
 738
 739	/* Prevent overflows */
 740	if (nr_events > (0x10000000U / sizeof(struct io_event))) {
 
 741		pr_debug("ENOMEM: nr_events too high\n");
 742		return ERR_PTR(-EINVAL);
 743	}
 744
 745	if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
 746		return ERR_PTR(-EAGAIN);
 747
 748	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
 749	if (!ctx)
 750		return ERR_PTR(-ENOMEM);
 751
 752	ctx->max_reqs = max_reqs;
 753
 754	spin_lock_init(&ctx->ctx_lock);
 755	spin_lock_init(&ctx->completion_lock);
 756	mutex_init(&ctx->ring_lock);
 757	/* Protect against page migration throughout kiotx setup by keeping
 758	 * the ring_lock mutex held until setup is complete. */
 759	mutex_lock(&ctx->ring_lock);
 760	init_waitqueue_head(&ctx->wait);
 761
 762	INIT_LIST_HEAD(&ctx->active_reqs);
 763
 764	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
 765		goto err;
 766
 767	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
 768		goto err;
 769
 770	ctx->cpu = alloc_percpu(struct kioctx_cpu);
 771	if (!ctx->cpu)
 772		goto err;
 773
 774	err = aio_setup_ring(ctx, nr_events);
 775	if (err < 0)
 776		goto err;
 777
 778	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
 779	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
 780	if (ctx->req_batch < 1)
 781		ctx->req_batch = 1;
 782
 783	/* limit the number of system wide aios */
 784	spin_lock(&aio_nr_lock);
 785	if (aio_nr + ctx->max_reqs > aio_max_nr ||
 786	    aio_nr + ctx->max_reqs < aio_nr) {
 787		spin_unlock(&aio_nr_lock);
 788		err = -EAGAIN;
 789		goto err_ctx;
 790	}
 791	aio_nr += ctx->max_reqs;
 792	spin_unlock(&aio_nr_lock);
 793
 794	percpu_ref_get(&ctx->users);	/* io_setup() will drop this ref */
 795	percpu_ref_get(&ctx->reqs);	/* free_ioctx_users() will drop this */
 796
 797	err = ioctx_add_table(ctx, mm);
 798	if (err)
 799		goto err_cleanup;
 800
 801	/* Release the ring_lock mutex now that all setup is complete. */
 802	mutex_unlock(&ctx->ring_lock);
 803
 804	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
 805		 ctx, ctx->user_id, mm, ctx->nr_events);
 806	return ctx;
 807
 808err_cleanup:
 809	aio_nr_sub(ctx->max_reqs);
 810err_ctx:
 811	atomic_set(&ctx->dead, 1);
 812	if (ctx->mmap_size)
 813		vm_munmap(ctx->mmap_base, ctx->mmap_size);
 814	aio_free_ring(ctx);
 815err:
 816	mutex_unlock(&ctx->ring_lock);
 817	free_percpu(ctx->cpu);
 818	percpu_ref_exit(&ctx->reqs);
 819	percpu_ref_exit(&ctx->users);
 820	kmem_cache_free(kioctx_cachep, ctx);
 821	pr_debug("error allocating ioctx %d\n", err);
 822	return ERR_PTR(err);
 823}
 824
 825/* kill_ioctx
 826 *	Cancels all outstanding aio requests on an aio context.  Used
 827 *	when the processes owning a context have all exited to encourage
 828 *	the rapid destruction of the kioctx.
 829 */
 830static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
 831		      struct ctx_rq_wait *wait)
 832{
 833	struct kioctx_table *table;
 
 834
 835	spin_lock(&mm->ioctx_lock);
 836	if (atomic_xchg(&ctx->dead, 1)) {
 
 
 
 
 
 837		spin_unlock(&mm->ioctx_lock);
 838		return -EINVAL;
 839	}
 840
 841	table = rcu_dereference_raw(mm->ioctx_table);
 842	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
 843	RCU_INIT_POINTER(table->table[ctx->id], NULL);
 844	spin_unlock(&mm->ioctx_lock);
 845
 846	/* free_ioctx_reqs() will do the necessary RCU synchronization */
 847	wake_up_all(&ctx->wait);
 
 
 
 
 
 
 848
 849	/*
 850	 * It'd be more correct to do this in free_ioctx(), after all
 851	 * the outstanding kiocbs have finished - but by then io_destroy
 852	 * has already returned, so io_setup() could potentially return
 853	 * -EAGAIN with no ioctxs actually in use (as far as userspace
 854	 *  could tell).
 855	 */
 856	aio_nr_sub(ctx->max_reqs);
 857
 858	if (ctx->mmap_size)
 859		vm_munmap(ctx->mmap_base, ctx->mmap_size);
 
 
 
 
 
 860
 861	ctx->rq_wait = wait;
 862	percpu_ref_kill(&ctx->users);
 863	return 0;
 
 
 
 
 
 
 
 
 
 
 864}
 
 865
 866/*
 867 * exit_aio: called when the last user of mm goes away.  At this point, there is
 868 * no way for any new requests to be submited or any of the io_* syscalls to be
 869 * called on the context.
 870 *
 871 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
 872 * them.
 873 */
 874void exit_aio(struct mm_struct *mm)
 875{
 876	struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
 877	struct ctx_rq_wait wait;
 878	int i, skipped;
 879
 880	if (!table)
 881		return;
 
 882
 883	atomic_set(&wait.count, table->nr);
 884	init_completion(&wait.comp);
 
 
 
 
 
 
 885
 886	skipped = 0;
 887	for (i = 0; i < table->nr; ++i) {
 888		struct kioctx *ctx =
 889			rcu_dereference_protected(table->table[i], true);
 890
 891		if (!ctx) {
 892			skipped++;
 893			continue;
 894		}
 895
 896		/*
 897		 * We don't need to bother with munmap() here - exit_mmap(mm)
 898		 * is coming and it'll unmap everything. And we simply can't,
 899		 * this is not necessarily our ->mm.
 900		 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
 901		 * that it needs to unmap the area, just set it to 0.
 
 902		 */
 903		ctx->mmap_size = 0;
 904		kill_ioctx(mm, ctx, &wait);
 905	}
 906
 907	if (!atomic_sub_and_test(skipped, &wait.count)) {
 908		/* Wait until all IO for the context are done. */
 909		wait_for_completion(&wait.comp);
 910	}
 911
 912	RCU_INIT_POINTER(mm->ioctx_table, NULL);
 913	kfree(table);
 914}
 915
 916static void put_reqs_available(struct kioctx *ctx, unsigned nr)
 917{
 918	struct kioctx_cpu *kcpu;
 919	unsigned long flags;
 920
 921	local_irq_save(flags);
 922	kcpu = this_cpu_ptr(ctx->cpu);
 923	kcpu->reqs_available += nr;
 924
 
 925	while (kcpu->reqs_available >= ctx->req_batch * 2) {
 926		kcpu->reqs_available -= ctx->req_batch;
 927		atomic_add(ctx->req_batch, &ctx->reqs_available);
 928	}
 929
 930	local_irq_restore(flags);
 931}
 932
 933static bool get_reqs_available(struct kioctx *ctx)
 934{
 935	struct kioctx_cpu *kcpu;
 936	bool ret = false;
 937	unsigned long flags;
 938
 939	local_irq_save(flags);
 940	kcpu = this_cpu_ptr(ctx->cpu);
 
 941	if (!kcpu->reqs_available) {
 942		int old, avail = atomic_read(&ctx->reqs_available);
 943
 944		do {
 945			if (avail < ctx->req_batch)
 946				goto out;
 947
 948			old = avail;
 949			avail = atomic_cmpxchg(&ctx->reqs_available,
 950					       avail, avail - ctx->req_batch);
 951		} while (avail != old);
 952
 953		kcpu->reqs_available += ctx->req_batch;
 954	}
 955
 956	ret = true;
 957	kcpu->reqs_available--;
 958out:
 959	local_irq_restore(flags);
 960	return ret;
 961}
 962
 963/* refill_reqs_available
 964 *	Updates the reqs_available reference counts used for tracking the
 965 *	number of free slots in the completion ring.  This can be called
 966 *	from aio_complete() (to optimistically update reqs_available) or
 967 *	from aio_get_req() (the we're out of events case).  It must be
 968 *	called holding ctx->completion_lock.
 969 */
 970static void refill_reqs_available(struct kioctx *ctx, unsigned head,
 971                                  unsigned tail)
 972{
 973	unsigned events_in_ring, completed;
 974
 975	/* Clamp head since userland can write to it. */
 976	head %= ctx->nr_events;
 977	if (head <= tail)
 978		events_in_ring = tail - head;
 979	else
 980		events_in_ring = ctx->nr_events - (head - tail);
 981
 982	completed = ctx->completed_events;
 983	if (events_in_ring < completed)
 984		completed -= events_in_ring;
 985	else
 986		completed = 0;
 987
 988	if (!completed)
 989		return;
 990
 991	ctx->completed_events -= completed;
 992	put_reqs_available(ctx, completed);
 993}
 994
 995/* user_refill_reqs_available
 996 *	Called to refill reqs_available when aio_get_req() encounters an
 997 *	out of space in the completion ring.
 998 */
 999static void user_refill_reqs_available(struct kioctx *ctx)
1000{
1001	spin_lock_irq(&ctx->completion_lock);
1002	if (ctx->completed_events) {
1003		struct aio_ring *ring;
1004		unsigned head;
1005
1006		/* Access of ring->head may race with aio_read_events_ring()
1007		 * here, but that's okay since whether we read the old version
1008		 * or the new version, and either will be valid.  The important
1009		 * part is that head cannot pass tail since we prevent
1010		 * aio_complete() from updating tail by holding
1011		 * ctx->completion_lock.  Even if head is invalid, the check
1012		 * against ctx->completed_events below will make sure we do the
1013		 * safe/right thing.
1014		 */
1015		ring = kmap_atomic(ctx->ring_pages[0]);
1016		head = ring->head;
1017		kunmap_atomic(ring);
1018
1019		refill_reqs_available(ctx, head, ctx->tail);
1020	}
1021
1022	spin_unlock_irq(&ctx->completion_lock);
1023}
1024
1025/* aio_get_req
1026 *	Allocate a slot for an aio request.
1027 * Returns NULL if no requests are free.
1028 */
1029static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1030{
1031	struct aio_kiocb *req;
1032
1033	if (!get_reqs_available(ctx)) {
1034		user_refill_reqs_available(ctx);
1035		if (!get_reqs_available(ctx))
1036			return NULL;
1037	}
1038
1039	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1040	if (unlikely(!req))
1041		goto out_put;
1042
1043	percpu_ref_get(&ctx->reqs);
1044
1045	req->ki_ctx = ctx;
1046	return req;
1047out_put:
1048	put_reqs_available(ctx, 1);
1049	return NULL;
1050}
1051
1052static void kiocb_free(struct aio_kiocb *req)
1053{
1054	if (req->common.ki_filp)
1055		fput(req->common.ki_filp);
1056	if (req->ki_eventfd != NULL)
1057		eventfd_ctx_put(req->ki_eventfd);
1058	kmem_cache_free(kiocb_cachep, req);
1059}
1060
1061static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1062{
1063	struct aio_ring __user *ring  = (void __user *)ctx_id;
1064	struct mm_struct *mm = current->mm;
1065	struct kioctx *ctx, *ret = NULL;
1066	struct kioctx_table *table;
1067	unsigned id;
1068
1069	if (get_user(id, &ring->id))
1070		return NULL;
1071
1072	rcu_read_lock();
1073	table = rcu_dereference(mm->ioctx_table);
1074
1075	if (!table || id >= table->nr)
1076		goto out;
1077
1078	ctx = rcu_dereference(table->table[id]);
1079	if (ctx && ctx->user_id == ctx_id) {
1080		if (percpu_ref_tryget_live(&ctx->users))
1081			ret = ctx;
1082	}
1083out:
1084	rcu_read_unlock();
1085	return ret;
1086}
1087
1088/* aio_complete
1089 *	Called when the io request on the given iocb is complete.
1090 */
1091static void aio_complete(struct kiocb *kiocb, long res, long res2)
1092{
1093	struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1094	struct kioctx	*ctx = iocb->ki_ctx;
1095	struct aio_ring	*ring;
1096	struct io_event	*ev_page, *event;
1097	unsigned tail, pos, head;
1098	unsigned long	flags;
1099
1100	if (kiocb->ki_flags & IOCB_WRITE) {
1101		struct file *file = kiocb->ki_filp;
1102
1103		/*
1104		 * Tell lockdep we inherited freeze protection from submission
1105		 * thread.
1106		 */
1107		if (S_ISREG(file_inode(file)->i_mode))
1108			__sb_writers_acquired(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1109		file_end_write(file);
1110	}
1111
1112	/*
1113	 * Special case handling for sync iocbs:
1114	 *  - events go directly into the iocb for fast handling
1115	 *  - the sync task with the iocb in its stack holds the single iocb
1116	 *    ref, no other paths have a way to get another ref
1117	 *  - the sync task helpfully left a reference to itself in the iocb
1118	 */
1119	BUG_ON(is_sync_kiocb(kiocb));
 
 
 
 
 
 
1120
1121	if (iocb->ki_list.next) {
1122		unsigned long flags;
1123
1124		spin_lock_irqsave(&ctx->ctx_lock, flags);
1125		list_del(&iocb->ki_list);
1126		spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1127	}
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 = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1143	event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1144
1145	event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1146	event->data = iocb->ki_user_data;
1147	event->res = res;
1148	event->res2 = res2;
1149
1150	kunmap_atomic(ev_page);
1151	flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1152
1153	pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1154		 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1155		 res, res2);
1156
1157	/* after flagging the request as done, we
1158	 * must never even look at it again
1159	 */
1160	smp_wmb();	/* make event visible before updating tail */
1161
1162	ctx->tail = tail;
1163
1164	ring = kmap_atomic(ctx->ring_pages[0]);
1165	head = ring->head;
1166	ring->tail = tail;
1167	kunmap_atomic(ring);
1168	flush_dcache_page(ctx->ring_pages[0]);
1169
1170	ctx->completed_events++;
1171	if (ctx->completed_events > 1)
1172		refill_reqs_available(ctx, head, tail);
1173	spin_unlock_irqrestore(&ctx->completion_lock, flags);
1174
1175	pr_debug("added to ring %p at [%u]\n", iocb, tail);
1176
1177	/*
1178	 * Check if the user asked us to deliver the result through an
1179	 * eventfd. The eventfd_signal() function is safe to be called
1180	 * from IRQ context.
1181	 */
1182	if (iocb->ki_eventfd != NULL)
1183		eventfd_signal(iocb->ki_eventfd, 1);
1184
1185	/* everything turned out well, dispose of the aiocb. */
1186	kiocb_free(iocb);
1187
1188	/*
1189	 * We have to order our ring_info tail store above and test
1190	 * of the wait list below outside the wait lock.  This is
1191	 * like in wake_up_bit() where clearing a bit has to be
1192	 * ordered with the unlocked test.
1193	 */
1194	smp_mb();
1195
1196	if (waitqueue_active(&ctx->wait))
1197		wake_up(&ctx->wait);
1198
1199	percpu_ref_put(&ctx->reqs);
1200}
 
1201
1202/* aio_read_events_ring
1203 *	Pull an event off of the ioctx's event ring.  Returns the number of
1204 *	events fetched
1205 */
1206static long aio_read_events_ring(struct kioctx *ctx,
1207				 struct io_event __user *event, long nr)
1208{
1209	struct aio_ring *ring;
1210	unsigned head, tail, pos;
1211	long ret = 0;
1212	int copy_ret;
1213
1214	/*
1215	 * The mutex can block and wake us up and that will cause
1216	 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1217	 * and repeat. This should be rare enough that it doesn't cause
1218	 * peformance issues. See the comment in read_events() for more detail.
1219	 */
1220	sched_annotate_sleep();
1221	mutex_lock(&ctx->ring_lock);
1222
1223	/* Access to ->ring_pages here is protected by ctx->ring_lock. */
1224	ring = kmap_atomic(ctx->ring_pages[0]);
1225	head = ring->head;
1226	tail = ring->tail;
1227	kunmap_atomic(ring);
1228
1229	/*
1230	 * Ensure that once we've read the current tail pointer, that
1231	 * we also see the events that were stored up to the tail.
1232	 */
1233	smp_rmb();
1234
1235	pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1236
1237	if (head == tail)
1238		goto out;
1239
1240	head %= ctx->nr_events;
1241	tail %= ctx->nr_events;
1242
1243	while (ret < nr) {
1244		long avail;
1245		struct io_event *ev;
1246		struct page *page;
1247
1248		avail = (head <= tail ?  tail : ctx->nr_events) - head;
1249		if (head == tail)
1250			break;
1251
1252		avail = min(avail, nr - ret);
1253		avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1254			    ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1255
1256		pos = head + AIO_EVENTS_OFFSET;
1257		page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1258		pos %= AIO_EVENTS_PER_PAGE;
1259
1260		ev = kmap(page);
1261		copy_ret = copy_to_user(event + ret, ev + pos,
1262					sizeof(*ev) * avail);
1263		kunmap(page);
1264
1265		if (unlikely(copy_ret)) {
1266			ret = -EFAULT;
1267			goto out;
1268		}
1269
1270		ret += avail;
1271		head += avail;
1272		head %= ctx->nr_events;
1273	}
1274
1275	ring = kmap_atomic(ctx->ring_pages[0]);
1276	ring->head = head;
1277	kunmap_atomic(ring);
1278	flush_dcache_page(ctx->ring_pages[0]);
1279
1280	pr_debug("%li  h%u t%u\n", ret, head, tail);
 
 
1281out:
1282	mutex_unlock(&ctx->ring_lock);
1283
1284	return ret;
1285}
1286
1287static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1288			    struct io_event __user *event, long *i)
1289{
1290	long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1291
1292	if (ret > 0)
1293		*i += ret;
1294
1295	if (unlikely(atomic_read(&ctx->dead)))
1296		ret = -EINVAL;
1297
1298	if (!*i)
1299		*i = ret;
1300
1301	return ret < 0 || *i >= min_nr;
1302}
1303
1304static long read_events(struct kioctx *ctx, long min_nr, long nr,
1305			struct io_event __user *event,
1306			ktime_t until)
1307{
 
1308	long ret = 0;
1309
 
 
 
 
 
 
 
 
 
1310	/*
1311	 * Note that aio_read_events() is being called as the conditional - i.e.
1312	 * we're calling it after prepare_to_wait() has set task state to
1313	 * TASK_INTERRUPTIBLE.
1314	 *
1315	 * But aio_read_events() can block, and if it blocks it's going to flip
1316	 * the task state back to TASK_RUNNING.
1317	 *
1318	 * This should be ok, provided it doesn't flip the state back to
1319	 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1320	 * will only happen if the mutex_lock() call blocks, and we then find
1321	 * the ringbuffer empty. So in practice we should be ok, but it's
1322	 * something to be aware of when touching this code.
1323	 */
1324	if (until == 0)
1325		aio_read_events(ctx, min_nr, nr, event, &ret);
1326	else
1327		wait_event_interruptible_hrtimeout(ctx->wait,
1328				aio_read_events(ctx, min_nr, nr, event, &ret),
1329				until);
1330
1331	if (!ret && signal_pending(current))
1332		ret = -EINTR;
1333
1334	return ret;
1335}
1336
1337/* sys_io_setup:
1338 *	Create an aio_context capable of receiving at least nr_events.
1339 *	ctxp must not point to an aio_context that already exists, and
1340 *	must be initialized to 0 prior to the call.  On successful
1341 *	creation of the aio_context, *ctxp is filled in with the resulting 
1342 *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1343 *	if the specified nr_events exceeds internal limits.  May fail 
1344 *	with -EAGAIN if the specified nr_events exceeds the user's limit 
1345 *	of available events.  May fail with -ENOMEM if insufficient kernel
1346 *	resources are available.  May fail with -EFAULT if an invalid
1347 *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1348 *	implemented.
1349 */
1350SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1351{
1352	struct kioctx *ioctx = NULL;
1353	unsigned long ctx;
1354	long ret;
1355
1356	ret = get_user(ctx, ctxp);
1357	if (unlikely(ret))
1358		goto out;
1359
1360	ret = -EINVAL;
1361	if (unlikely(ctx || nr_events == 0)) {
1362		pr_debug("EINVAL: ctx %lu nr_events %u\n",
1363		         ctx, nr_events);
1364		goto out;
1365	}
1366
1367	ioctx = ioctx_alloc(nr_events);
1368	ret = PTR_ERR(ioctx);
1369	if (!IS_ERR(ioctx)) {
1370		ret = put_user(ioctx->user_id, ctxp);
1371		if (ret)
1372			kill_ioctx(current->mm, ioctx, NULL);
1373		percpu_ref_put(&ioctx->users);
1374	}
1375
1376out:
1377	return ret;
1378}
1379
1380#ifdef CONFIG_COMPAT
1381COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1382{
1383	struct kioctx *ioctx = NULL;
1384	unsigned long ctx;
1385	long ret;
1386
1387	ret = get_user(ctx, ctx32p);
1388	if (unlikely(ret))
1389		goto out;
1390
1391	ret = -EINVAL;
1392	if (unlikely(ctx || nr_events == 0)) {
1393		pr_debug("EINVAL: ctx %lu nr_events %u\n",
1394		         ctx, nr_events);
1395		goto out;
1396	}
1397
1398	ioctx = ioctx_alloc(nr_events);
1399	ret = PTR_ERR(ioctx);
1400	if (!IS_ERR(ioctx)) {
1401		/* truncating is ok because it's a user address */
1402		ret = put_user((u32)ioctx->user_id, ctx32p);
1403		if (ret)
1404			kill_ioctx(current->mm, ioctx, NULL);
1405		percpu_ref_put(&ioctx->users);
1406	}
1407
1408out:
1409	return ret;
1410}
1411#endif
1412
1413/* sys_io_destroy:
1414 *	Destroy the aio_context specified.  May cancel any outstanding 
1415 *	AIOs and block on completion.  Will fail with -ENOSYS if not
1416 *	implemented.  May fail with -EINVAL if the context pointed to
1417 *	is invalid.
1418 */
1419SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1420{
1421	struct kioctx *ioctx = lookup_ioctx(ctx);
1422	if (likely(NULL != ioctx)) {
1423		struct ctx_rq_wait wait;
1424		int ret;
1425
1426		init_completion(&wait.comp);
1427		atomic_set(&wait.count, 1);
1428
1429		/* Pass requests_done to kill_ioctx() where it can be set
1430		 * in a thread-safe way. If we try to set it here then we have
1431		 * a race condition if two io_destroy() called simultaneously.
1432		 */
1433		ret = kill_ioctx(current->mm, ioctx, &wait);
1434		percpu_ref_put(&ioctx->users);
1435
1436		/* Wait until all IO for the context are done. Otherwise kernel
1437		 * keep using user-space buffers even if user thinks the context
1438		 * is destroyed.
1439		 */
1440		if (!ret)
1441			wait_for_completion(&wait.comp);
1442
1443		return ret;
1444	}
1445	pr_debug("EINVAL: invalid context id\n");
1446	return -EINVAL;
1447}
1448
1449static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec,
1450		bool vectored, bool compat, struct iov_iter *iter)
 
 
 
 
 
 
1451{
1452	void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1453	size_t len = iocb->aio_nbytes;
 
1454
1455	if (!vectored) {
1456		ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1457		*iovec = NULL;
1458		return ret;
1459	}
1460#ifdef CONFIG_COMPAT
1461	if (compat)
1462		return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1463				iter);
 
 
1464#endif
1465	return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
 
 
 
 
 
 
 
 
1466}
1467
1468static inline ssize_t aio_ret(struct kiocb *req, ssize_t ret)
 
 
 
1469{
1470	switch (ret) {
1471	case -EIOCBQUEUED:
1472		return ret;
1473	case -ERESTARTSYS:
1474	case -ERESTARTNOINTR:
1475	case -ERESTARTNOHAND:
1476	case -ERESTART_RESTARTBLOCK:
1477		/*
1478		 * There's no easy way to restart the syscall since other AIO's
1479		 * may be already running. Just fail this IO with EINTR.
1480		 */
1481		ret = -EINTR;
1482		/*FALLTHRU*/
1483	default:
1484		aio_complete(req, ret, 0);
1485		return 0;
1486	}
1487}
1488
1489static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored,
1490		bool compat)
 
 
 
 
 
1491{
1492	struct file *file = req->ki_filp;
1493	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1494	struct iov_iter iter;
1495	ssize_t ret;
 
 
 
 
 
1496
1497	if (unlikely(!(file->f_mode & FMODE_READ)))
1498		return -EBADF;
1499	if (unlikely(!file->f_op->read_iter))
1500		return -EINVAL;
 
 
 
1501
1502	ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1503	if (ret)
1504		return ret;
1505	ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1506	if (!ret)
1507		ret = aio_ret(req, call_read_iter(file, req, &iter));
1508	kfree(iovec);
1509	return ret;
1510}
1511
1512static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored,
1513		bool compat)
1514{
1515	struct file *file = req->ki_filp;
1516	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1517	struct iov_iter iter;
1518	ssize_t ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1519
1520	if (unlikely(!(file->f_mode & FMODE_WRITE)))
1521		return -EBADF;
1522	if (unlikely(!file->f_op->write_iter))
 
 
 
 
 
 
 
 
 
 
 
 
 
1523		return -EINVAL;
 
1524
1525	ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1526	if (ret)
1527		return ret;
1528	ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1529	if (!ret) {
1530		req->ki_flags |= IOCB_WRITE;
1531		file_start_write(file);
1532		ret = aio_ret(req, call_write_iter(file, req, &iter));
1533		/*
1534		 * We release freeze protection in aio_complete().  Fool lockdep
1535		 * by telling it the lock got released so that it doesn't
1536		 * complain about held lock when we return to userspace.
1537		 */
1538		if (S_ISREG(file_inode(file)->i_mode))
1539			__sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
 
 
 
1540	}
1541	kfree(iovec);
1542	return ret;
1543}
1544
1545static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1546			 struct iocb *iocb, bool compat)
1547{
1548	struct aio_kiocb *req;
1549	struct file *file;
1550	ssize_t ret;
1551
1552	/* enforce forwards compatibility on users */
1553	if (unlikely(iocb->aio_reserved2)) {
1554		pr_debug("EINVAL: reserve field set\n");
1555		return -EINVAL;
1556	}
1557
1558	/* prevent overflows */
1559	if (unlikely(
1560	    (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1561	    (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1562	    ((ssize_t)iocb->aio_nbytes < 0)
1563	   )) {
1564		pr_debug("EINVAL: overflow check\n");
1565		return -EINVAL;
1566	}
1567
1568	req = aio_get_req(ctx);
1569	if (unlikely(!req))
1570		return -EAGAIN;
1571
1572	req->common.ki_filp = file = fget(iocb->aio_fildes);
1573	if (unlikely(!req->common.ki_filp)) {
1574		ret = -EBADF;
1575		goto out_put_req;
1576	}
1577	req->common.ki_pos = iocb->aio_offset;
1578	req->common.ki_complete = aio_complete;
1579	req->common.ki_flags = iocb_flags(req->common.ki_filp);
1580	req->common.ki_hint = file_write_hint(file);
1581
1582	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1583		/*
1584		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1585		 * instance of the file* now. The file descriptor must be
1586		 * an eventfd() fd, and will be signaled for each completed
1587		 * event using the eventfd_signal() function.
1588		 */
1589		req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1590		if (IS_ERR(req->ki_eventfd)) {
1591			ret = PTR_ERR(req->ki_eventfd);
1592			req->ki_eventfd = NULL;
1593			goto out_put_req;
1594		}
1595
1596		req->common.ki_flags |= IOCB_EVENTFD;
1597	}
1598
1599	ret = kiocb_set_rw_flags(&req->common, iocb->aio_rw_flags);
1600	if (unlikely(ret)) {
1601		pr_debug("EINVAL: aio_rw_flags\n");
1602		goto out_put_req;
1603	}
1604
1605	ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1606	if (unlikely(ret)) {
1607		pr_debug("EFAULT: aio_key\n");
1608		goto out_put_req;
1609	}
1610
1611	req->ki_user_iocb = user_iocb;
1612	req->ki_user_data = iocb->aio_data;
 
 
1613
1614	get_file(file);
1615	switch (iocb->aio_lio_opcode) {
1616	case IOCB_CMD_PREAD:
1617		ret = aio_read(&req->common, iocb, false, compat);
1618		break;
1619	case IOCB_CMD_PWRITE:
1620		ret = aio_write(&req->common, iocb, false, compat);
1621		break;
1622	case IOCB_CMD_PREADV:
1623		ret = aio_read(&req->common, iocb, true, compat);
1624		break;
1625	case IOCB_CMD_PWRITEV:
1626		ret = aio_write(&req->common, iocb, true, compat);
1627		break;
1628	default:
1629		pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1630		ret = -EINVAL;
1631		break;
1632	}
1633	fput(file);
1634
1635	if (ret && ret != -EIOCBQUEUED)
1636		goto out_put_req;
 
1637	return 0;
1638out_put_req:
1639	put_reqs_available(ctx, 1);
1640	percpu_ref_put(&ctx->reqs);
1641	kiocb_free(req);
1642	return ret;
1643}
1644
1645static long do_io_submit(aio_context_t ctx_id, long nr,
1646			  struct iocb __user *__user *iocbpp, bool compat)
1647{
1648	struct kioctx *ctx;
1649	long ret = 0;
1650	int i = 0;
1651	struct blk_plug plug;
1652
1653	if (unlikely(nr < 0))
1654		return -EINVAL;
1655
1656	if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1657		nr = LONG_MAX/sizeof(*iocbpp);
1658
1659	if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1660		return -EFAULT;
1661
1662	ctx = lookup_ioctx(ctx_id);
1663	if (unlikely(!ctx)) {
1664		pr_debug("EINVAL: invalid context id\n");
1665		return -EINVAL;
1666	}
1667
1668	blk_start_plug(&plug);
1669
1670	/*
1671	 * AKPM: should this return a partial result if some of the IOs were
1672	 * successfully submitted?
1673	 */
1674	for (i=0; i<nr; i++) {
1675		struct iocb __user *user_iocb;
1676		struct iocb tmp;
1677
1678		if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1679			ret = -EFAULT;
1680			break;
1681		}
1682
1683		if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1684			ret = -EFAULT;
1685			break;
1686		}
1687
1688		ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1689		if (ret)
1690			break;
1691	}
1692	blk_finish_plug(&plug);
1693
1694	percpu_ref_put(&ctx->users);
1695	return i ? i : ret;
1696}
1697
1698/* sys_io_submit:
1699 *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1700 *	the number of iocbs queued.  May return -EINVAL if the aio_context
1701 *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
1702 *	*iocbpp[0] is not properly initialized, if the operation specified
1703 *	is invalid for the file descriptor in the iocb.  May fail with
1704 *	-EFAULT if any of the data structures point to invalid data.  May
1705 *	fail with -EBADF if the file descriptor specified in the first
1706 *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
1707 *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1708 *	fail with -ENOSYS if not implemented.
1709 */
1710SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1711		struct iocb __user * __user *, iocbpp)
1712{
1713	return do_io_submit(ctx_id, nr, iocbpp, 0);
1714}
1715
1716#ifdef CONFIG_COMPAT
1717static inline long
1718copy_iocb(long nr, u32 __user *ptr32, struct iocb __user * __user *ptr64)
1719{
1720	compat_uptr_t uptr;
1721	int i;
1722
1723	for (i = 0; i < nr; ++i) {
1724		if (get_user(uptr, ptr32 + i))
1725			return -EFAULT;
1726		if (put_user(compat_ptr(uptr), ptr64 + i))
1727			return -EFAULT;
1728	}
1729	return 0;
1730}
1731
1732#define MAX_AIO_SUBMITS 	(PAGE_SIZE/sizeof(struct iocb *))
1733
1734COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1735		       int, nr, u32 __user *, iocb)
1736{
1737	struct iocb __user * __user *iocb64;
1738	long ret;
1739
1740	if (unlikely(nr < 0))
1741		return -EINVAL;
1742
1743	if (nr > MAX_AIO_SUBMITS)
1744		nr = MAX_AIO_SUBMITS;
1745
1746	iocb64 = compat_alloc_user_space(nr * sizeof(*iocb64));
1747	ret = copy_iocb(nr, iocb, iocb64);
1748	if (!ret)
1749		ret = do_io_submit(ctx_id, nr, iocb64, 1);
1750	return ret;
1751}
1752#endif
1753
1754/* lookup_kiocb
1755 *	Finds a given iocb for cancellation.
1756 */
1757static struct aio_kiocb *
1758lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1759{
1760	struct aio_kiocb *kiocb;
1761
1762	assert_spin_locked(&ctx->ctx_lock);
1763
1764	if (key != KIOCB_KEY)
1765		return NULL;
1766
1767	/* TODO: use a hash or array, this sucks. */
1768	list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1769		if (kiocb->ki_user_iocb == iocb)
 
1770			return kiocb;
1771	}
1772	return NULL;
1773}
1774
1775/* sys_io_cancel:
1776 *	Attempts to cancel an iocb previously passed to io_submit.  If
1777 *	the operation is successfully cancelled, the resulting event is
1778 *	copied into the memory pointed to by result without being placed
1779 *	into the completion queue and 0 is returned.  May fail with
1780 *	-EFAULT if any of the data structures pointed to are invalid.
1781 *	May fail with -EINVAL if aio_context specified by ctx_id is
1782 *	invalid.  May fail with -EAGAIN if the iocb specified was not
1783 *	cancelled.  Will fail with -ENOSYS if not implemented.
1784 */
1785SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1786		struct io_event __user *, result)
1787{
1788	struct kioctx *ctx;
1789	struct aio_kiocb *kiocb;
1790	u32 key;
1791	int ret;
1792
1793	ret = get_user(key, &iocb->aio_key);
1794	if (unlikely(ret))
1795		return -EFAULT;
1796
1797	ctx = lookup_ioctx(ctx_id);
1798	if (unlikely(!ctx))
1799		return -EINVAL;
1800
1801	spin_lock_irq(&ctx->ctx_lock);
1802
1803	kiocb = lookup_kiocb(ctx, iocb, key);
1804	if (kiocb)
1805		ret = kiocb_cancel(kiocb);
1806	else
1807		ret = -EINVAL;
1808
1809	spin_unlock_irq(&ctx->ctx_lock);
1810
1811	if (!ret) {
1812		/*
1813		 * The result argument is no longer used - the io_event is
1814		 * always delivered via the ring buffer. -EINPROGRESS indicates
1815		 * cancellation is progress:
1816		 */
1817		ret = -EINPROGRESS;
1818	}
1819
1820	percpu_ref_put(&ctx->users);
1821
1822	return ret;
1823}
1824
1825static long do_io_getevents(aio_context_t ctx_id,
1826		long min_nr,
1827		long nr,
1828		struct io_event __user *events,
1829		struct timespec64 *ts)
1830{
1831	ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
1832	struct kioctx *ioctx = lookup_ioctx(ctx_id);
1833	long ret = -EINVAL;
1834
1835	if (likely(ioctx)) {
1836		if (likely(min_nr <= nr && min_nr >= 0))
1837			ret = read_events(ioctx, min_nr, nr, events, until);
1838		percpu_ref_put(&ioctx->users);
1839	}
1840
1841	return ret;
1842}
1843
1844/* io_getevents:
1845 *	Attempts to read at least min_nr events and up to nr events from
1846 *	the completion queue for the aio_context specified by ctx_id. If
1847 *	it succeeds, the number of read events is returned. May fail with
1848 *	-EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1849 *	out of range, if timeout is out of range.  May fail with -EFAULT
1850 *	if any of the memory specified is invalid.  May return 0 or
1851 *	< min_nr if the timeout specified by timeout has elapsed
1852 *	before sufficient events are available, where timeout == NULL
1853 *	specifies an infinite timeout. Note that the timeout pointed to by
1854 *	timeout is relative.  Will fail with -ENOSYS if not implemented.
1855 */
1856SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1857		long, min_nr,
1858		long, nr,
1859		struct io_event __user *, events,
1860		struct timespec __user *, timeout)
1861{
1862	struct timespec64	ts;
1863
1864	if (timeout) {
1865		if (unlikely(get_timespec64(&ts, timeout)))
1866			return -EFAULT;
1867	}
1868
1869	return do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
1870}
1871
1872#ifdef CONFIG_COMPAT
1873COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id,
1874		       compat_long_t, min_nr,
1875		       compat_long_t, nr,
1876		       struct io_event __user *, events,
1877		       struct compat_timespec __user *, timeout)
1878{
1879	struct timespec64 t;
1880
1881	if (timeout) {
1882		if (compat_get_timespec64(&t, timeout))
1883			return -EFAULT;
1884
 
 
 
 
1885	}
1886
1887	return do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
1888}
1889#endif