Linux Audio

Check our new training course

Loading...
v3.1
   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#include <linux/kernel.h>
  12#include <linux/init.h>
  13#include <linux/errno.h>
  14#include <linux/time.h>
  15#include <linux/aio_abi.h>
  16#include <linux/module.h>
  17#include <linux/syscalls.h>
  18#include <linux/backing-dev.h>
  19#include <linux/uio.h>
  20
  21#define DEBUG 0
  22
  23#include <linux/sched.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/slab.h>
  30#include <linux/timer.h>
  31#include <linux/aio.h>
  32#include <linux/highmem.h>
  33#include <linux/workqueue.h>
  34#include <linux/security.h>
  35#include <linux/eventfd.h>
  36#include <linux/blkdev.h>
  37#include <linux/compat.h>
 
 
 
 
  38
  39#include <asm/kmap_types.h>
  40#include <asm/uaccess.h>
  41
  42#if DEBUG > 1
  43#define dprintk		printk
  44#else
  45#define dprintk(x...)	do { ; } while (0)
  46#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  47
  48/*------ sysctl variables----*/
  49static DEFINE_SPINLOCK(aio_nr_lock);
  50unsigned long aio_nr;		/* current system wide number of aio requests */
  51unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
  52/*----end sysctl variables---*/
  53
  54static struct kmem_cache	*kiocb_cachep;
  55static struct kmem_cache	*kioctx_cachep;
  56
  57static struct workqueue_struct *aio_wq;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  58
  59/* Used for rare fput completion. */
  60static void aio_fput_routine(struct work_struct *);
  61static DECLARE_WORK(fput_work, aio_fput_routine);
 
 
 
  62
  63static DEFINE_SPINLOCK(fput_lock);
  64static LIST_HEAD(fput_head);
 
 
  65
  66static void aio_kick_handler(struct work_struct *);
  67static void aio_queue_work(struct kioctx *);
 
 
 
 
 
 
  68
  69/* aio_setup
  70 *	Creates the slab caches used by the aio routines, panic on
  71 *	failure as this is done early during the boot sequence.
  72 */
  73static int __init aio_setup(void)
  74{
 
 
 
 
 
 
 
 
 
  75	kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
  76	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
  77
  78	aio_wq = alloc_workqueue("aio", 0, 1);	/* used to limit concurrency */
  79	BUG_ON(!aio_wq);
  80
  81	pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
  82
  83	return 0;
  84}
  85__initcall(aio_setup);
  86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  87static void aio_free_ring(struct kioctx *ctx)
  88{
  89	struct aio_ring_info *info = &ctx->ring_info;
  90	long i;
  91
  92	for (i=0; i<info->nr_pages; i++)
  93		put_page(info->ring_pages[i]);
 
 
 
 
 
 
 
 
 
 
 
 
 
  94
  95	if (info->mmap_size) {
  96		down_write(&ctx->mm->mmap_sem);
  97		do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
  98		up_write(&ctx->mm->mmap_sem);
  99	}
 100
 101	if (info->ring_pages && info->ring_pages != info->internal_pages)
 102		kfree(info->ring_pages);
 103	info->ring_pages = NULL;
 104	info->nr = 0;
 105}
 106
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 107static int aio_setup_ring(struct kioctx *ctx)
 108{
 109	struct aio_ring *ring;
 110	struct aio_ring_info *info = &ctx->ring_info;
 111	unsigned nr_events = ctx->max_reqs;
 112	unsigned long size;
 
 113	int nr_pages;
 
 
 114
 115	/* Compensate for the ring buffer's head/tail overlap entry */
 116	nr_events += 2;	/* 1 is required, 2 for good luck */
 117
 118	size = sizeof(struct aio_ring);
 119	size += sizeof(struct io_event) * nr_events;
 120	nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
 121
 
 122	if (nr_pages < 0)
 123		return -EINVAL;
 124
 125	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
 
 
 
 
 
 
 
 
 126
 127	info->nr = 0;
 128	info->ring_pages = info->internal_pages;
 129	if (nr_pages > AIO_RING_PAGES) {
 130		info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
 131		if (!info->ring_pages)
 
 
 132			return -ENOMEM;
 
 133	}
 134
 135	info->mmap_size = nr_pages * PAGE_SIZE;
 136	dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
 137	down_write(&ctx->mm->mmap_sem);
 138	info->mmap_base = do_mmap(NULL, 0, info->mmap_size, 
 139				  PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
 140				  0);
 141	if (IS_ERR((void *)info->mmap_base)) {
 142		up_write(&ctx->mm->mmap_sem);
 143		info->mmap_size = 0;
 
 
 
 
 
 
 
 
 144		aio_free_ring(ctx);
 145		return -EAGAIN;
 146	}
 147
 148	dprintk("mmap address: 0x%08lx\n", info->mmap_base);
 149	info->nr_pages = get_user_pages(current, ctx->mm,
 150					info->mmap_base, nr_pages, 
 151					1, 0, info->ring_pages, NULL);
 152	up_write(&ctx->mm->mmap_sem);
 153
 154	if (unlikely(info->nr_pages != nr_pages)) {
 
 
 
 
 
 
 155		aio_free_ring(ctx);
 156		return -EAGAIN;
 157	}
 158
 159	ctx->user_id = info->mmap_base;
 160
 161	info->nr = nr_events;		/* trusted copy */
 
 162
 163	ring = kmap_atomic(info->ring_pages[0], KM_USER0);
 164	ring->nr = nr_events;	/* user copy */
 165	ring->id = ctx->user_id;
 166	ring->head = ring->tail = 0;
 167	ring->magic = AIO_RING_MAGIC;
 168	ring->compat_features = AIO_RING_COMPAT_FEATURES;
 169	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
 170	ring->header_length = sizeof(struct aio_ring);
 171	kunmap_atomic(ring, KM_USER0);
 
 172
 173	return 0;
 174}
 175
 176
 177/* aio_ring_event: returns a pointer to the event at the given index from
 178 * kmap_atomic(, km).  Release the pointer with put_aio_ring_event();
 179 */
 180#define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
 181#define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
 182#define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
 183
 184#define aio_ring_event(info, nr, km) ({					\
 185	unsigned pos = (nr) + AIO_EVENTS_OFFSET;			\
 186	struct io_event *__event;					\
 187	__event = kmap_atomic(						\
 188			(info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
 189	__event += pos % AIO_EVENTS_PER_PAGE;				\
 190	__event;							\
 191})
 192
 193#define put_aio_ring_event(event, km) do {	\
 194	struct io_event *__event = (event);	\
 195	(void)__event;				\
 196	kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
 197} while(0)
 198
 199static void ctx_rcu_free(struct rcu_head *head)
 
 
 
 
 
 
 200{
 201	struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
 202	unsigned nr_events = ctx->max_reqs;
 203
 204	kmem_cache_free(kioctx_cachep, ctx);
 
 
 
 205
 206	if (nr_events) {
 207		spin_lock(&aio_nr_lock);
 208		BUG_ON(aio_nr - nr_events > aio_nr);
 209		aio_nr -= nr_events;
 210		spin_unlock(&aio_nr_lock);
 211	}
 
 
 
 
 212}
 213
 214/* __put_ioctx
 215 *	Called when the last user of an aio context has gone away,
 216 *	and the struct needs to be freed.
 217 */
 218static void __put_ioctx(struct kioctx *ctx)
 219{
 220	BUG_ON(ctx->reqs_active);
 
 
 221
 222	cancel_delayed_work(&ctx->wq);
 223	cancel_work_sync(&ctx->wq.work);
 224	aio_free_ring(ctx);
 225	mmdrop(ctx->mm);
 226	ctx->mm = NULL;
 227	pr_debug("__put_ioctx: freeing %p\n", ctx);
 228	call_rcu(&ctx->rcu_head, ctx_rcu_free);
 229}
 230
 231static inline void get_ioctx(struct kioctx *kioctx)
 232{
 233	BUG_ON(atomic_read(&kioctx->users) <= 0);
 234	atomic_inc(&kioctx->users);
 
 
 
 
 
 
 235}
 236
 237static inline int try_get_ioctx(struct kioctx *kioctx)
 
 
 
 
 
 238{
 239	return atomic_inc_not_zero(&kioctx->users);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 240}
 241
 242static inline void put_ioctx(struct kioctx *kioctx)
 243{
 244	BUG_ON(atomic_read(&kioctx->users) <= 0);
 245	if (unlikely(atomic_dec_and_test(&kioctx->users)))
 246		__put_ioctx(kioctx);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 247}
 248
 249/* ioctx_alloc
 250 *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
 251 */
 252static struct kioctx *ioctx_alloc(unsigned nr_events)
 253{
 254	struct mm_struct *mm;
 255	struct kioctx *ctx;
 256	int did_sync = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 257
 258	/* Prevent overflows */
 259	if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
 260	    (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
 261		pr_debug("ENOMEM: nr_events too high\n");
 262		return ERR_PTR(-EINVAL);
 263	}
 264
 265	if ((unsigned long)nr_events > aio_max_nr)
 266		return ERR_PTR(-EAGAIN);
 267
 268	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
 269	if (!ctx)
 270		return ERR_PTR(-ENOMEM);
 271
 272	ctx->max_reqs = nr_events;
 273	mm = ctx->mm = current->mm;
 274	atomic_inc(&mm->mm_count);
 275
 276	atomic_set(&ctx->users, 1);
 277	spin_lock_init(&ctx->ctx_lock);
 278	spin_lock_init(&ctx->ring_info.ring_lock);
 
 
 
 
 279	init_waitqueue_head(&ctx->wait);
 280
 281	INIT_LIST_HEAD(&ctx->active_reqs);
 282	INIT_LIST_HEAD(&ctx->run_list);
 283	INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
 284
 285	if (aio_setup_ring(ctx) < 0)
 286		goto out_freectx;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 287
 288	/* limit the number of system wide aios */
 289	do {
 290		spin_lock_bh(&aio_nr_lock);
 291		if (aio_nr + nr_events > aio_max_nr ||
 292		    aio_nr + nr_events < aio_nr)
 293			ctx->max_reqs = 0;
 294		else
 295			aio_nr += ctx->max_reqs;
 296		spin_unlock_bh(&aio_nr_lock);
 297		if (ctx->max_reqs || did_sync)
 298			break;
 299
 300		/* wait for rcu callbacks to have completed before giving up */
 301		synchronize_rcu();
 302		did_sync = 1;
 303		ctx->max_reqs = nr_events;
 304	} while (1);
 305
 306	if (ctx->max_reqs == 0)
 307		goto out_cleanup;
 
 308
 309	/* now link into global list. */
 310	spin_lock(&mm->ioctx_lock);
 311	hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
 312	spin_unlock(&mm->ioctx_lock);
 313
 314	dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
 315		ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
 316	return ctx;
 317
 318out_cleanup:
 319	__put_ioctx(ctx);
 320	return ERR_PTR(-EAGAIN);
 321
 322out_freectx:
 323	mmdrop(mm);
 
 
 
 324	kmem_cache_free(kioctx_cachep, ctx);
 325	ctx = ERR_PTR(-ENOMEM);
 326
 327	dprintk("aio: error allocating ioctx %p\n", ctx);
 328	return ctx;
 329}
 330
 331/* aio_cancel_all
 332 *	Cancels all outstanding aio requests on an aio context.  Used 
 333 *	when the processes owning a context have all exited to encourage 
 334 *	the rapid destruction of the kioctx.
 335 */
 336static void aio_cancel_all(struct kioctx *ctx)
 
 337{
 338	int (*cancel)(struct kiocb *, struct io_event *);
 339	struct io_event res;
 340	spin_lock_irq(&ctx->ctx_lock);
 341	ctx->dead = 1;
 342	while (!list_empty(&ctx->active_reqs)) {
 343		struct list_head *pos = ctx->active_reqs.next;
 344		struct kiocb *iocb = list_kiocb(pos);
 345		list_del_init(&iocb->ki_list);
 346		cancel = iocb->ki_cancel;
 347		kiocbSetCancelled(iocb);
 348		if (cancel) {
 349			iocb->ki_users++;
 350			spin_unlock_irq(&ctx->ctx_lock);
 351			cancel(iocb, &res);
 352			spin_lock_irq(&ctx->ctx_lock);
 353		}
 354	}
 355	spin_unlock_irq(&ctx->ctx_lock);
 356}
 357
 358static void wait_for_all_aios(struct kioctx *ctx)
 359{
 360	struct task_struct *tsk = current;
 361	DECLARE_WAITQUEUE(wait, tsk);
 
 
 
 
 362
 363	spin_lock_irq(&ctx->ctx_lock);
 364	if (!ctx->reqs_active)
 365		goto out;
 366
 367	add_wait_queue(&ctx->wait, &wait);
 368	set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 369	while (ctx->reqs_active) {
 370		spin_unlock_irq(&ctx->ctx_lock);
 371		io_schedule();
 372		set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 373		spin_lock_irq(&ctx->ctx_lock);
 374	}
 375	__set_task_state(tsk, TASK_RUNNING);
 376	remove_wait_queue(&ctx->wait, &wait);
 377
 378out:
 379	spin_unlock_irq(&ctx->ctx_lock);
 
 
 
 
 
 
 
 380}
 381
 382/* wait_on_sync_kiocb:
 383 *	Waits on the given sync kiocb to complete.
 384 */
 385ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
 386{
 387	while (iocb->ki_users) {
 388		set_current_state(TASK_UNINTERRUPTIBLE);
 389		if (!iocb->ki_users)
 390			break;
 391		io_schedule();
 392	}
 393	__set_current_state(TASK_RUNNING);
 394	return iocb->ki_user_data;
 395}
 396EXPORT_SYMBOL(wait_on_sync_kiocb);
 397
 398/* exit_aio: called when the last user of mm goes away.  At this point, 
 399 * there is no way for any new requests to be submited or any of the 
 400 * io_* syscalls to be called on the context.  However, there may be 
 401 * outstanding requests which hold references to the context; as they 
 402 * go away, they will call put_ioctx and release any pinned memory
 403 * associated with the request (held via struct page * references).
 
 404 */
 405void exit_aio(struct mm_struct *mm)
 406{
 
 407	struct kioctx *ctx;
 
 
 
 
 
 408
 409	while (!hlist_empty(&mm->ioctx_list)) {
 410		ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
 411		hlist_del_rcu(&ctx->list);
 
 
 
 
 
 
 
 
 412
 413		aio_cancel_all(ctx);
 414
 415		wait_for_all_aios(ctx);
 416		/*
 417		 * Ensure we don't leave the ctx on the aio_wq
 
 
 
 
 
 418		 */
 419		cancel_work_sync(&ctx->wq.work);
 420
 421		if (1 != atomic_read(&ctx->users))
 422			printk(KERN_DEBUG
 423				"exit_aio:ioctx still alive: %d %d %d\n",
 424				atomic_read(&ctx->users), ctx->dead,
 425				ctx->reqs_active);
 426		put_ioctx(ctx);
 427	}
 428}
 429
 430/* aio_get_req
 431 *	Allocate a slot for an aio request.  Increments the users count
 432 * of the kioctx so that the kioctx stays around until all requests are
 433 * complete.  Returns NULL if no requests are free.
 434 *
 435 * Returns with kiocb->users set to 2.  The io submit code path holds
 436 * an extra reference while submitting the i/o.
 437 * This prevents races between the aio code path referencing the
 438 * req (after submitting it) and aio_complete() freeing the req.
 439 */
 440static struct kiocb *__aio_get_req(struct kioctx *ctx)
 441{
 442	struct kiocb *req = NULL;
 443	struct aio_ring *ring;
 444	int okay = 0;
 445
 446	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
 447	if (unlikely(!req))
 448		return NULL;
 449
 450	req->ki_flags = 0;
 451	req->ki_users = 2;
 452	req->ki_key = 0;
 453	req->ki_ctx = ctx;
 454	req->ki_cancel = NULL;
 455	req->ki_retry = NULL;
 456	req->ki_dtor = NULL;
 457	req->private = NULL;
 458	req->ki_iovec = NULL;
 459	INIT_LIST_HEAD(&req->ki_run_list);
 460	req->ki_eventfd = NULL;
 461
 462	/* Check if the completion queue has enough free space to
 463	 * accept an event from this io.
 464	 */
 465	spin_lock_irq(&ctx->ctx_lock);
 466	ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
 467	if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
 468		list_add(&req->ki_list, &ctx->active_reqs);
 469		ctx->reqs_active++;
 470		okay = 1;
 471	}
 472	kunmap_atomic(ring, KM_USER0);
 473	spin_unlock_irq(&ctx->ctx_lock);
 474
 475	if (!okay) {
 476		kmem_cache_free(kiocb_cachep, req);
 477		req = NULL;
 478	}
 479
 480	return req;
 481}
 482
 483static inline struct kiocb *aio_get_req(struct kioctx *ctx)
 484{
 485	struct kiocb *req;
 486	/* Handle a potential starvation case -- should be exceedingly rare as 
 487	 * requests will be stuck on fput_head only if the aio_fput_routine is 
 488	 * delayed and the requests were the last user of the struct file.
 489	 */
 490	req = __aio_get_req(ctx);
 491	if (unlikely(NULL == req)) {
 492		aio_fput_routine(NULL);
 493		req = __aio_get_req(ctx);
 494	}
 495	return req;
 496}
 497
 498static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
 499{
 500	assert_spin_locked(&ctx->ctx_lock);
 
 501
 502	if (req->ki_eventfd != NULL)
 503		eventfd_ctx_put(req->ki_eventfd);
 504	if (req->ki_dtor)
 505		req->ki_dtor(req);
 506	if (req->ki_iovec != &req->ki_inline_vec)
 507		kfree(req->ki_iovec);
 508	kmem_cache_free(kiocb_cachep, req);
 509	ctx->reqs_active--;
 510
 511	if (unlikely(!ctx->reqs_active && ctx->dead))
 512		wake_up_all(&ctx->wait);
 513}
 514
 515static void aio_fput_routine(struct work_struct *data)
 516{
 517	spin_lock_irq(&fput_lock);
 518	while (likely(!list_empty(&fput_head))) {
 519		struct kiocb *req = list_kiocb(fput_head.next);
 520		struct kioctx *ctx = req->ki_ctx;
 521
 522		list_del(&req->ki_list);
 523		spin_unlock_irq(&fput_lock);
 524
 525		/* Complete the fput(s) */
 526		if (req->ki_filp != NULL)
 527			fput(req->ki_filp);
 528
 529		/* Link the iocb into the context's free list */
 530		spin_lock_irq(&ctx->ctx_lock);
 531		really_put_req(ctx, req);
 532		spin_unlock_irq(&ctx->ctx_lock);
 533
 534		put_ioctx(ctx);
 535		spin_lock_irq(&fput_lock);
 536	}
 537	spin_unlock_irq(&fput_lock);
 
 
 
 
 
 538}
 539
 540/* __aio_put_req
 541 *	Returns true if this put was the last user of the request.
 
 542 */
 543static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
 544{
 545	dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
 546		req, atomic_long_read(&req->ki_filp->f_count));
 547
 548	assert_spin_locked(&ctx->ctx_lock);
 
 549
 550	req->ki_users--;
 551	BUG_ON(req->ki_users < 0);
 552	if (likely(req->ki_users))
 553		return 0;
 554	list_del(&req->ki_list);		/* remove from active_reqs */
 555	req->ki_cancel = NULL;
 556	req->ki_retry = NULL;
 557
 558	/*
 559	 * Try to optimize the aio and eventfd file* puts, by avoiding to
 560	 * schedule work in case it is not final fput() time. In normal cases,
 561	 * we would not be holding the last reference to the file*, so
 562	 * this function will be executed w/out any aio kthread wakeup.
 563	 */
 564	if (unlikely(!fput_atomic(req->ki_filp))) {
 565		get_ioctx(ctx);
 566		spin_lock(&fput_lock);
 567		list_add(&req->ki_list, &fput_head);
 568		spin_unlock(&fput_lock);
 569		schedule_work(&fput_work);
 570	} else {
 571		req->ki_filp = NULL;
 572		really_put_req(ctx, req);
 573	}
 574	return 1;
 575}
 576
 577/* aio_put_req
 578 *	Returns true if this put was the last user of the kiocb,
 579 *	false if the request is still in use.
 580 */
 581int aio_put_req(struct kiocb *req)
 582{
 583	struct kioctx *ctx = req->ki_ctx;
 584	int ret;
 585	spin_lock_irq(&ctx->ctx_lock);
 586	ret = __aio_put_req(ctx, req);
 587	spin_unlock_irq(&ctx->ctx_lock);
 588	return ret;
 589}
 590EXPORT_SYMBOL(aio_put_req);
 591
 592static struct kioctx *lookup_ioctx(unsigned long ctx_id)
 593{
 
 594	struct mm_struct *mm = current->mm;
 595	struct kioctx *ctx, *ret = NULL;
 596	struct hlist_node *n;
 
 597
 598	rcu_read_lock();
 599
 600	hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
 601		/*
 602		 * RCU protects us against accessing freed memory but
 603		 * we have to be careful not to get a reference when the
 604		 * reference count already dropped to 0 (ctx->dead test
 605		 * is unreliable because of races).
 606		 */
 607		if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
 608			ret = ctx;
 609			break;
 610		}
 611	}
 612
 613	rcu_read_unlock();
 614	return ret;
 615}
 616
 617/*
 618 * Queue up a kiocb to be retried. Assumes that the kiocb
 619 * has already been marked as kicked, and places it on
 620 * the retry run list for the corresponding ioctx, if it
 621 * isn't already queued. Returns 1 if it actually queued
 622 * the kiocb (to tell the caller to activate the work
 623 * queue to process it), or 0, if it found that it was
 624 * already queued.
 625 */
 626static inline int __queue_kicked_iocb(struct kiocb *iocb)
 627{
 628	struct kioctx *ctx = iocb->ki_ctx;
 629
 630	assert_spin_locked(&ctx->ctx_lock);
 631
 632	if (list_empty(&iocb->ki_run_list)) {
 633		list_add_tail(&iocb->ki_run_list,
 634			&ctx->run_list);
 635		return 1;
 636	}
 637	return 0;
 638}
 639
 640/* aio_run_iocb
 641 *	This is the core aio execution routine. It is
 642 *	invoked both for initial i/o submission and
 643 *	subsequent retries via the aio_kick_handler.
 644 *	Expects to be invoked with iocb->ki_ctx->lock
 645 *	already held. The lock is released and reacquired
 646 *	as needed during processing.
 647 *
 648 * Calls the iocb retry method (already setup for the
 649 * iocb on initial submission) for operation specific
 650 * handling, but takes care of most of common retry
 651 * execution details for a given iocb. The retry method
 652 * needs to be non-blocking as far as possible, to avoid
 653 * holding up other iocbs waiting to be serviced by the
 654 * retry kernel thread.
 655 *
 656 * The trickier parts in this code have to do with
 657 * ensuring that only one retry instance is in progress
 658 * for a given iocb at any time. Providing that guarantee
 659 * simplifies the coding of individual aio operations as
 660 * it avoids various potential races.
 661 */
 662static ssize_t aio_run_iocb(struct kiocb *iocb)
 663{
 664	struct kioctx	*ctx = iocb->ki_ctx;
 665	ssize_t (*retry)(struct kiocb *);
 666	ssize_t ret;
 667
 668	if (!(retry = iocb->ki_retry)) {
 669		printk("aio_run_iocb: iocb->ki_retry = NULL\n");
 670		return 0;
 671	}
 672
 673	/*
 674	 * We don't want the next retry iteration for this
 675	 * operation to start until this one has returned and
 676	 * updated the iocb state. However, wait_queue functions
 677	 * can trigger a kick_iocb from interrupt context in the
 678	 * meantime, indicating that data is available for the next
 679	 * iteration. We want to remember that and enable the
 680	 * next retry iteration _after_ we are through with
 681	 * this one.
 682	 *
 683	 * So, in order to be able to register a "kick", but
 684	 * prevent it from being queued now, we clear the kick
 685	 * flag, but make the kick code *think* that the iocb is
 686	 * still on the run list until we are actually done.
 687	 * When we are done with this iteration, we check if
 688	 * the iocb was kicked in the meantime and if so, queue
 689	 * it up afresh.
 690	 */
 691
 692	kiocbClearKicked(iocb);
 693
 694	/*
 695	 * This is so that aio_complete knows it doesn't need to
 696	 * pull the iocb off the run list (We can't just call
 697	 * INIT_LIST_HEAD because we don't want a kick_iocb to
 698	 * queue this on the run list yet)
 699	 */
 700	iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
 701	spin_unlock_irq(&ctx->ctx_lock);
 702
 703	/* Quit retrying if the i/o has been cancelled */
 704	if (kiocbIsCancelled(iocb)) {
 705		ret = -EINTR;
 706		aio_complete(iocb, ret, 0);
 707		/* must not access the iocb after this */
 708		goto out;
 709	}
 710
 711	/*
 712	 * Now we are all set to call the retry method in async
 713	 * context.
 714	 */
 715	ret = retry(iocb);
 716
 717	if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
 718		/*
 719		 * There's no easy way to restart the syscall since other AIO's
 720		 * may be already running. Just fail this IO with EINTR.
 721		 */
 722		if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
 723			     ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
 724			ret = -EINTR;
 725		aio_complete(iocb, ret, 0);
 726	}
 727out:
 728	spin_lock_irq(&ctx->ctx_lock);
 729
 730	if (-EIOCBRETRY == ret) {
 731		/*
 732		 * OK, now that we are done with this iteration
 733		 * and know that there is more left to go,
 734		 * this is where we let go so that a subsequent
 735		 * "kick" can start the next iteration
 736		 */
 737
 738		/* will make __queue_kicked_iocb succeed from here on */
 739		INIT_LIST_HEAD(&iocb->ki_run_list);
 740		/* we must queue the next iteration ourselves, if it
 741		 * has already been kicked */
 742		if (kiocbIsKicked(iocb)) {
 743			__queue_kicked_iocb(iocb);
 744
 745			/*
 746			 * __queue_kicked_iocb will always return 1 here, because
 747			 * iocb->ki_run_list is empty at this point so it should
 748			 * be safe to unconditionally queue the context into the
 749			 * work queue.
 750			 */
 751			aio_queue_work(ctx);
 752		}
 753	}
 754	return ret;
 755}
 756
 757/*
 758 * __aio_run_iocbs:
 759 * 	Process all pending retries queued on the ioctx
 760 * 	run list.
 761 * Assumes it is operating within the aio issuer's mm
 762 * context.
 763 */
 764static int __aio_run_iocbs(struct kioctx *ctx)
 765{
 766	struct kiocb *iocb;
 767	struct list_head run_list;
 768
 769	assert_spin_locked(&ctx->ctx_lock);
 770
 771	list_replace_init(&ctx->run_list, &run_list);
 772	while (!list_empty(&run_list)) {
 773		iocb = list_entry(run_list.next, struct kiocb,
 774			ki_run_list);
 775		list_del(&iocb->ki_run_list);
 776		/*
 777		 * Hold an extra reference while retrying i/o.
 778		 */
 779		iocb->ki_users++;       /* grab extra reference */
 780		aio_run_iocb(iocb);
 781		__aio_put_req(ctx, iocb);
 782 	}
 783	if (!list_empty(&ctx->run_list))
 784		return 1;
 785	return 0;
 786}
 787
 788static void aio_queue_work(struct kioctx * ctx)
 789{
 790	unsigned long timeout;
 791	/*
 792	 * if someone is waiting, get the work started right
 793	 * away, otherwise, use a longer delay
 794	 */
 795	smp_mb();
 796	if (waitqueue_active(&ctx->wait))
 797		timeout = 1;
 798	else
 799		timeout = HZ/10;
 800	queue_delayed_work(aio_wq, &ctx->wq, timeout);
 801}
 802
 803/*
 804 * aio_run_all_iocbs:
 805 *	Process all pending retries queued on the ioctx
 806 *	run list, and keep running them until the list
 807 *	stays empty.
 808 * Assumes it is operating within the aio issuer's mm context.
 809 */
 810static inline void aio_run_all_iocbs(struct kioctx *ctx)
 811{
 812	spin_lock_irq(&ctx->ctx_lock);
 813	while (__aio_run_iocbs(ctx))
 814		;
 815	spin_unlock_irq(&ctx->ctx_lock);
 816}
 817
 818/*
 819 * aio_kick_handler:
 820 * 	Work queue handler triggered to process pending
 821 * 	retries on an ioctx. Takes on the aio issuer's
 822 *	mm context before running the iocbs, so that
 823 *	copy_xxx_user operates on the issuer's address
 824 *      space.
 825 * Run on aiod's context.
 826 */
 827static void aio_kick_handler(struct work_struct *work)
 828{
 829	struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
 830	mm_segment_t oldfs = get_fs();
 831	struct mm_struct *mm;
 832	int requeue;
 833
 834	set_fs(USER_DS);
 835	use_mm(ctx->mm);
 836	spin_lock_irq(&ctx->ctx_lock);
 837	requeue =__aio_run_iocbs(ctx);
 838	mm = ctx->mm;
 839	spin_unlock_irq(&ctx->ctx_lock);
 840 	unuse_mm(mm);
 841	set_fs(oldfs);
 842	/*
 843	 * we're in a worker thread already, don't use queue_delayed_work,
 844	 */
 845	if (requeue)
 846		queue_delayed_work(aio_wq, &ctx->wq, 0);
 847}
 848
 849
 850/*
 851 * Called by kick_iocb to queue the kiocb for retry
 852 * and if required activate the aio work queue to process
 853 * it
 854 */
 855static void try_queue_kicked_iocb(struct kiocb *iocb)
 856{
 857 	struct kioctx	*ctx = iocb->ki_ctx;
 858	unsigned long flags;
 859	int run = 0;
 860
 861	spin_lock_irqsave(&ctx->ctx_lock, flags);
 862	/* set this inside the lock so that we can't race with aio_run_iocb()
 863	 * testing it and putting the iocb on the run list under the lock */
 864	if (!kiocbTryKick(iocb))
 865		run = __queue_kicked_iocb(iocb);
 866	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 867	if (run)
 868		aio_queue_work(ctx);
 869}
 870
 871/*
 872 * kick_iocb:
 873 *      Called typically from a wait queue callback context
 874 *      to trigger a retry of the iocb.
 875 *      The retry is usually executed by aio workqueue
 876 *      threads (See aio_kick_handler).
 877 */
 878void kick_iocb(struct kiocb *iocb)
 879{
 880	/* sync iocbs are easy: they can only ever be executing from a 
 881	 * single context. */
 882	if (is_sync_kiocb(iocb)) {
 883		kiocbSetKicked(iocb);
 884	        wake_up_process(iocb->ki_obj.tsk);
 885		return;
 886	}
 887
 888	try_queue_kicked_iocb(iocb);
 889}
 890EXPORT_SYMBOL(kick_iocb);
 891
 892/* aio_complete
 893 *	Called when the io request on the given iocb is complete.
 894 *	Returns true if this is the last user of the request.  The 
 895 *	only other user of the request can be the cancellation code.
 896 */
 897int aio_complete(struct kiocb *iocb, long res, long res2)
 898{
 899	struct kioctx	*ctx = iocb->ki_ctx;
 900	struct aio_ring_info	*info;
 901	struct aio_ring	*ring;
 902	struct io_event	*event;
 903	unsigned long	flags;
 904	unsigned long	tail;
 905	int		ret;
 906
 907	/*
 908	 * Special case handling for sync iocbs:
 909	 *  - events go directly into the iocb for fast handling
 910	 *  - the sync task with the iocb in its stack holds the single iocb
 911	 *    ref, no other paths have a way to get another ref
 912	 *  - the sync task helpfully left a reference to itself in the iocb
 913	 */
 914	if (is_sync_kiocb(iocb)) {
 915		BUG_ON(iocb->ki_users != 1);
 916		iocb->ki_user_data = res;
 917		iocb->ki_users = 0;
 
 918		wake_up_process(iocb->ki_obj.tsk);
 919		return 1;
 920	}
 921
 922	info = &ctx->ring_info;
 
 923
 924	/* add a completion event to the ring buffer.
 925	 * must be done holding ctx->ctx_lock to prevent
 926	 * other code from messing with the tail
 927	 * pointer since we might be called from irq
 928	 * context.
 929	 */
 930	spin_lock_irqsave(&ctx->ctx_lock, flags);
 931
 932	if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
 933		list_del_init(&iocb->ki_run_list);
 934
 935	/*
 936	 * cancelled requests don't get events, userland was given one
 937	 * when the event got cancelled.
 
 938	 */
 939	if (kiocbIsCancelled(iocb))
 940		goto put_rq;
 941
 942	ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
 
 943
 944	tail = info->tail;
 945	event = aio_ring_event(info, tail, KM_IRQ0);
 946	if (++tail >= info->nr)
 947		tail = 0;
 948
 
 
 
 949	event->obj = (u64)(unsigned long)iocb->ki_obj.user;
 950	event->data = iocb->ki_user_data;
 951	event->res = res;
 952	event->res2 = res2;
 953
 954	dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
 955		ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
 956		res, res2);
 
 
 
 957
 958	/* after flagging the request as done, we
 959	 * must never even look at it again
 960	 */
 961	smp_wmb();	/* make event visible before updating tail */
 962
 963	info->tail = tail;
 
 
 964	ring->tail = tail;
 
 
 965
 966	put_aio_ring_event(event, KM_IRQ0);
 967	kunmap_atomic(ring, KM_IRQ1);
 968
 969	pr_debug("added to ring %p at [%lu]\n", iocb, tail);
 970
 971	/*
 972	 * Check if the user asked us to deliver the result through an
 973	 * eventfd. The eventfd_signal() function is safe to be called
 974	 * from IRQ context.
 975	 */
 976	if (iocb->ki_eventfd != NULL)
 977		eventfd_signal(iocb->ki_eventfd, 1);
 978
 979put_rq:
 980	/* everything turned out well, dispose of the aiocb. */
 981	ret = __aio_put_req(ctx, iocb);
 982
 983	/*
 984	 * We have to order our ring_info tail store above and test
 985	 * of the wait list below outside the wait lock.  This is
 986	 * like in wake_up_bit() where clearing a bit has to be
 987	 * ordered with the unlocked test.
 988	 */
 989	smp_mb();
 990
 991	if (waitqueue_active(&ctx->wait))
 992		wake_up(&ctx->wait);
 993
 994	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 995	return ret;
 996}
 997EXPORT_SYMBOL(aio_complete);
 998
 999/* aio_read_evt
1000 *	Pull an event off of the ioctx's event ring.  Returns the number of 
1001 *	events fetched (0 or 1 ;-)
1002 *	FIXME: make this use cmpxchg.
1003 *	TODO: make the ringbuffer user mmap()able (requires FIXME).
1004 */
1005static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
 
1006{
1007	struct aio_ring_info *info = &ioctx->ring_info;
1008	struct aio_ring *ring;
1009	unsigned long head;
1010	int ret = 0;
1011
1012	ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1013	dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1014		 (unsigned long)ring->head, (unsigned long)ring->tail,
1015		 (unsigned long)ring->nr);
1016
1017	if (ring->head == ring->tail)
1018		goto out;
1019
1020	spin_lock(&info->ring_lock);
 
 
 
 
1021
1022	head = ring->head % info->nr;
1023	if (head != ring->tail) {
1024		struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1025		*ent = *evp;
1026		head = (head + 1) % info->nr;
1027		smp_mb(); /* finish reading the event before updatng the head */
1028		ring->head = head;
1029		ret = 1;
1030		put_aio_ring_event(evp, KM_USER1);
1031	}
1032	spin_unlock(&info->ring_lock);
1033
1034out:
1035	kunmap_atomic(ring, KM_USER0);
1036	dprintk("leaving aio_read_evt: %d  h%lu t%lu\n", ret,
1037		 (unsigned long)ring->head, (unsigned long)ring->tail);
1038	return ret;
1039}
1040
1041struct aio_timeout {
1042	struct timer_list	timer;
1043	int			timed_out;
1044	struct task_struct	*p;
1045};
1046
1047static void timeout_func(unsigned long data)
1048{
1049	struct aio_timeout *to = (struct aio_timeout *)data;
1050
1051	to->timed_out = 1;
1052	wake_up_process(to->p);
1053}
1054
1055static inline void init_timeout(struct aio_timeout *to)
1056{
1057	setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1058	to->timed_out = 0;
1059	to->p = current;
1060}
1061
1062static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1063			       const struct timespec *ts)
1064{
1065	to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1066	if (time_after(to->timer.expires, jiffies))
1067		add_timer(&to->timer);
1068	else
1069		to->timed_out = 1;
1070}
1071
1072static inline void clear_timeout(struct aio_timeout *to)
1073{
1074	del_singleshot_timer_sync(&to->timer);
1075}
1076
1077static int read_events(struct kioctx *ctx,
1078			long min_nr, long nr,
1079			struct io_event __user *event,
1080			struct timespec __user *timeout)
1081{
1082	long			start_jiffies = jiffies;
1083	struct task_struct	*tsk = current;
1084	DECLARE_WAITQUEUE(wait, tsk);
1085	int			ret;
1086	int			i = 0;
1087	struct io_event		ent;
1088	struct aio_timeout	to;
1089	int			retry = 0;
1090
1091	/* needed to zero any padding within an entry (there shouldn't be 
1092	 * any, but C is fun!
1093	 */
1094	memset(&ent, 0, sizeof(ent));
1095retry:
1096	ret = 0;
1097	while (likely(i < nr)) {
1098		ret = aio_read_evt(ctx, &ent);
1099		if (unlikely(ret <= 0))
1100			break;
1101
1102		dprintk("read event: %Lx %Lx %Lx %Lx\n",
1103			ent.data, ent.obj, ent.res, ent.res2);
1104
1105		/* Could we split the check in two? */
1106		ret = -EFAULT;
1107		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1108			dprintk("aio: lost an event due to EFAULT.\n");
1109			break;
1110		}
1111		ret = 0;
1112
1113		/* Good, event copied to userland, update counts. */
1114		event ++;
1115		i ++;
1116	}
1117
1118	if (min_nr <= i)
1119		return i;
1120	if (ret)
1121		return ret;
1122
1123	/* End fast path */
1124
1125	/* racey check, but it gets redone */
1126	if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1127		retry = 1;
1128		aio_run_all_iocbs(ctx);
1129		goto retry;
1130	}
1131
1132	init_timeout(&to);
1133	if (timeout) {
1134		struct timespec	ts;
1135		ret = -EFAULT;
1136		if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1137			goto out;
1138
1139		set_timeout(start_jiffies, &to, &ts);
1140	}
1141
1142	while (likely(i < nr)) {
1143		add_wait_queue_exclusive(&ctx->wait, &wait);
1144		do {
1145			set_task_state(tsk, TASK_INTERRUPTIBLE);
1146			ret = aio_read_evt(ctx, &ent);
1147			if (ret)
1148				break;
1149			if (min_nr <= i)
1150				break;
1151			if (unlikely(ctx->dead)) {
1152				ret = -EINVAL;
1153				break;
1154			}
1155			if (to.timed_out)	/* Only check after read evt */
1156				break;
1157			/* Try to only show up in io wait if there are ops
1158			 *  in flight */
1159			if (ctx->reqs_active)
1160				io_schedule();
1161			else
1162				schedule();
1163			if (signal_pending(tsk)) {
1164				ret = -EINTR;
1165				break;
1166			}
1167			/*ret = aio_read_evt(ctx, &ent);*/
1168		} while (1) ;
1169
1170		set_task_state(tsk, TASK_RUNNING);
1171		remove_wait_queue(&ctx->wait, &wait);
 
 
1172
1173		if (unlikely(ret <= 0))
1174			break;
1175
1176		ret = -EFAULT;
1177		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1178			dprintk("aio: lost an event due to EFAULT.\n");
1179			break;
1180		}
1181
1182		/* Good, event copied to userland, update counts. */
1183		event ++;
1184		i ++;
1185	}
1186
1187	if (timeout)
1188		clear_timeout(&to);
1189out:
1190	destroy_timer_on_stack(&to.timer);
1191	return i ? i : ret;
1192}
1193
1194/* Take an ioctx and remove it from the list of ioctx's.  Protects 
1195 * against races with itself via ->dead.
1196 */
1197static void io_destroy(struct kioctx *ioctx)
1198{
1199	struct mm_struct *mm = current->mm;
1200	int was_dead;
1201
1202	/* delete the entry from the list is someone else hasn't already */
1203	spin_lock(&mm->ioctx_lock);
1204	was_dead = ioctx->dead;
1205	ioctx->dead = 1;
1206	hlist_del_rcu(&ioctx->list);
1207	spin_unlock(&mm->ioctx_lock);
1208
1209	dprintk("aio_release(%p)\n", ioctx);
1210	if (likely(!was_dead))
1211		put_ioctx(ioctx);	/* twice for the list */
1212
1213	aio_cancel_all(ioctx);
1214	wait_for_all_aios(ioctx);
1215
1216	/*
1217	 * Wake up any waiters.  The setting of ctx->dead must be seen
1218	 * by other CPUs at this point.  Right now, we rely on the
1219	 * locking done by the above calls to ensure this consistency.
 
 
 
 
 
 
 
 
 
1220	 */
1221	wake_up_all(&ioctx->wait);
1222	put_ioctx(ioctx);	/* once for the lookup */
 
 
 
 
 
1223}
1224
1225/* sys_io_setup:
1226 *	Create an aio_context capable of receiving at least nr_events.
1227 *	ctxp must not point to an aio_context that already exists, and
1228 *	must be initialized to 0 prior to the call.  On successful
1229 *	creation of the aio_context, *ctxp is filled in with the resulting 
1230 *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1231 *	if the specified nr_events exceeds internal limits.  May fail 
1232 *	with -EAGAIN if the specified nr_events exceeds the user's limit 
1233 *	of available events.  May fail with -ENOMEM if insufficient kernel
1234 *	resources are available.  May fail with -EFAULT if an invalid
1235 *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1236 *	implemented.
1237 */
1238SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1239{
1240	struct kioctx *ioctx = NULL;
1241	unsigned long ctx;
1242	long ret;
1243
1244	ret = get_user(ctx, ctxp);
1245	if (unlikely(ret))
1246		goto out;
1247
1248	ret = -EINVAL;
1249	if (unlikely(ctx || nr_events == 0)) {
1250		pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1251		         ctx, nr_events);
1252		goto out;
1253	}
1254
1255	ioctx = ioctx_alloc(nr_events);
1256	ret = PTR_ERR(ioctx);
1257	if (!IS_ERR(ioctx)) {
1258		ret = put_user(ioctx->user_id, ctxp);
1259		if (!ret)
1260			return 0;
1261
1262		get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1263		io_destroy(ioctx);
1264	}
1265
1266out:
1267	return ret;
1268}
1269
1270/* sys_io_destroy:
1271 *	Destroy the aio_context specified.  May cancel any outstanding 
1272 *	AIOs and block on completion.  Will fail with -ENOSYS if not
1273 *	implemented.  May fail with -EINVAL if the context pointed to
1274 *	is invalid.
1275 */
1276SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1277{
1278	struct kioctx *ioctx = lookup_ioctx(ctx);
1279	if (likely(NULL != ioctx)) {
1280		io_destroy(ioctx);
1281		return 0;
1282	}
1283	pr_debug("EINVAL: io_destroy: invalid context id\n");
1284	return -EINVAL;
1285}
1286
1287static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1288{
1289	struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1290
1291	BUG_ON(ret <= 0);
 
1292
1293	while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1294		ssize_t this = min((ssize_t)iov->iov_len, ret);
1295		iov->iov_base += this;
1296		iov->iov_len -= this;
1297		iocb->ki_left -= this;
1298		ret -= this;
1299		if (iov->iov_len == 0) {
1300			iocb->ki_cur_seg++;
1301			iov++;
1302		}
1303	}
1304
1305	/* the caller should not have done more io than what fit in
1306	 * the remaining iovecs */
1307	BUG_ON(ret > 0 && iocb->ki_left == 0);
1308}
1309
1310static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1311{
1312	struct file *file = iocb->ki_filp;
1313	struct address_space *mapping = file->f_mapping;
1314	struct inode *inode = mapping->host;
1315	ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1316			 unsigned long, loff_t);
1317	ssize_t ret = 0;
1318	unsigned short opcode;
1319
1320	if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1321		(iocb->ki_opcode == IOCB_CMD_PREAD)) {
1322		rw_op = file->f_op->aio_read;
1323		opcode = IOCB_CMD_PREADV;
1324	} else {
1325		rw_op = file->f_op->aio_write;
1326		opcode = IOCB_CMD_PWRITEV;
1327	}
1328
1329	/* This matches the pread()/pwrite() logic */
1330	if (iocb->ki_pos < 0)
1331		return -EINVAL;
1332
1333	do {
1334		ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1335			    iocb->ki_nr_segs - iocb->ki_cur_seg,
1336			    iocb->ki_pos);
1337		if (ret > 0)
1338			aio_advance_iovec(iocb, ret);
1339
1340	/* retry all partial writes.  retry partial reads as long as its a
1341	 * regular file. */
1342	} while (ret > 0 && iocb->ki_left > 0 &&
1343		 (opcode == IOCB_CMD_PWRITEV ||
1344		  (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1345
1346	/* This means we must have transferred all that we could */
1347	/* No need to retry anymore */
1348	if ((ret == 0) || (iocb->ki_left == 0))
1349		ret = iocb->ki_nbytes - iocb->ki_left;
1350
1351	/* If we managed to write some out we return that, rather than
1352	 * the eventual error. */
1353	if (opcode == IOCB_CMD_PWRITEV
1354	    && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1355	    && iocb->ki_nbytes - iocb->ki_left)
1356		ret = iocb->ki_nbytes - iocb->ki_left;
1357
1358	return ret;
1359}
1360
1361static ssize_t aio_fdsync(struct kiocb *iocb)
1362{
1363	struct file *file = iocb->ki_filp;
1364	ssize_t ret = -EINVAL;
1365
1366	if (file->f_op->aio_fsync)
1367		ret = file->f_op->aio_fsync(iocb, 1);
1368	return ret;
1369}
1370
1371static ssize_t aio_fsync(struct kiocb *iocb)
1372{
1373	struct file *file = iocb->ki_filp;
1374	ssize_t ret = -EINVAL;
1375
1376	if (file->f_op->aio_fsync)
1377		ret = file->f_op->aio_fsync(iocb, 0);
1378	return ret;
1379}
1380
1381static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
 
 
 
 
1382{
1383	ssize_t ret;
1384
 
 
1385#ifdef CONFIG_COMPAT
1386	if (compat)
1387		ret = compat_rw_copy_check_uvector(type,
1388				(struct compat_iovec __user *)kiocb->ki_buf,
1389				kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1390				&kiocb->ki_iovec);
1391	else
1392#endif
1393		ret = rw_copy_check_uvector(type,
1394				(struct iovec __user *)kiocb->ki_buf,
1395				kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1396				&kiocb->ki_iovec);
1397	if (ret < 0)
1398		goto out;
1399
1400	kiocb->ki_nr_segs = kiocb->ki_nbytes;
1401	kiocb->ki_cur_seg = 0;
1402	/* ki_nbytes/left now reflect bytes instead of segs */
1403	kiocb->ki_nbytes = ret;
1404	kiocb->ki_left = ret;
1405
1406	ret = 0;
1407out:
1408	return ret;
1409}
1410
1411static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
 
 
 
1412{
1413	kiocb->ki_iovec = &kiocb->ki_inline_vec;
1414	kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1415	kiocb->ki_iovec->iov_len = kiocb->ki_left;
1416	kiocb->ki_nr_segs = 1;
1417	kiocb->ki_cur_seg = 0;
 
1418	return 0;
1419}
1420
1421/*
1422 * aio_setup_iocb:
1423 *	Performs the initial checks and aio retry method
1424 *	setup for the kiocb at the time of io submission.
1425 */
1426static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
 
1427{
1428	struct file *file = kiocb->ki_filp;
1429	ssize_t ret = 0;
 
 
 
 
 
1430
1431	switch (kiocb->ki_opcode) {
1432	case IOCB_CMD_PREAD:
1433		ret = -EBADF;
1434		if (unlikely(!(file->f_mode & FMODE_READ)))
1435			break;
1436		ret = -EFAULT;
1437		if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1438			kiocb->ki_left)))
1439			break;
1440		ret = security_file_permission(file, MAY_READ);
1441		if (unlikely(ret))
1442			break;
1443		ret = aio_setup_single_vector(kiocb);
1444		if (ret)
1445			break;
1446		ret = -EINVAL;
1447		if (file->f_op->aio_read)
1448			kiocb->ki_retry = aio_rw_vect_retry;
1449		break;
1450	case IOCB_CMD_PWRITE:
1451		ret = -EBADF;
1452		if (unlikely(!(file->f_mode & FMODE_WRITE)))
1453			break;
1454		ret = -EFAULT;
1455		if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1456			kiocb->ki_left)))
1457			break;
1458		ret = security_file_permission(file, MAY_WRITE);
1459		if (unlikely(ret))
1460			break;
1461		ret = aio_setup_single_vector(kiocb);
1462		if (ret)
1463			break;
1464		ret = -EINVAL;
1465		if (file->f_op->aio_write)
1466			kiocb->ki_retry = aio_rw_vect_retry;
1467		break;
1468	case IOCB_CMD_PREADV:
1469		ret = -EBADF;
1470		if (unlikely(!(file->f_mode & FMODE_READ)))
1471			break;
1472		ret = security_file_permission(file, MAY_READ);
1473		if (unlikely(ret))
1474			break;
1475		ret = aio_setup_vectored_rw(READ, kiocb, compat);
1476		if (ret)
1477			break;
1478		ret = -EINVAL;
1479		if (file->f_op->aio_read)
1480			kiocb->ki_retry = aio_rw_vect_retry;
1481		break;
1482	case IOCB_CMD_PWRITEV:
1483		ret = -EBADF;
1484		if (unlikely(!(file->f_mode & FMODE_WRITE)))
1485			break;
1486		ret = security_file_permission(file, MAY_WRITE);
1487		if (unlikely(ret))
1488			break;
1489		ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1490		if (ret)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1491			break;
1492		ret = -EINVAL;
1493		if (file->f_op->aio_write)
1494			kiocb->ki_retry = aio_rw_vect_retry;
 
 
 
 
 
 
1495		break;
 
1496	case IOCB_CMD_FDSYNC:
1497		ret = -EINVAL;
1498		if (file->f_op->aio_fsync)
1499			kiocb->ki_retry = aio_fdsync;
 
1500		break;
 
1501	case IOCB_CMD_FSYNC:
1502		ret = -EINVAL;
1503		if (file->f_op->aio_fsync)
1504			kiocb->ki_retry = aio_fsync;
 
1505		break;
 
1506	default:
1507		dprintk("EINVAL: io_submit: no operation provided\n");
1508		ret = -EINVAL;
1509	}
1510
1511	if (!kiocb->ki_retry)
1512		return ret;
 
 
 
 
 
 
 
 
 
 
 
 
1513
1514	return 0;
1515}
1516
1517static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1518			 struct iocb *iocb, bool compat)
1519{
1520	struct kiocb *req;
1521	struct file *file;
1522	ssize_t ret;
1523
1524	/* enforce forwards compatibility on users */
1525	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1526		pr_debug("EINVAL: io_submit: reserve field set\n");
1527		return -EINVAL;
1528	}
1529
1530	/* prevent overflows */
1531	if (unlikely(
1532	    (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1533	    (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1534	    ((ssize_t)iocb->aio_nbytes < 0)
1535	   )) {
1536		pr_debug("EINVAL: io_submit: overflow check\n");
1537		return -EINVAL;
1538	}
1539
1540	file = fget(iocb->aio_fildes);
1541	if (unlikely(!file))
1542		return -EBADF;
1543
1544	req = aio_get_req(ctx);		/* returns with 2 references to req */
1545	if (unlikely(!req)) {
1546		fput(file);
1547		return -EAGAIN;
 
 
 
 
 
1548	}
1549	req->ki_filp = file;
1550	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1551		/*
1552		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1553		 * instance of the file* now. The file descriptor must be
1554		 * an eventfd() fd, and will be signaled for each completed
1555		 * event using the eventfd_signal() function.
1556		 */
1557		req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1558		if (IS_ERR(req->ki_eventfd)) {
1559			ret = PTR_ERR(req->ki_eventfd);
1560			req->ki_eventfd = NULL;
1561			goto out_put_req;
1562		}
1563	}
1564
1565	ret = put_user(req->ki_key, &user_iocb->aio_key);
1566	if (unlikely(ret)) {
1567		dprintk("EFAULT: aio_key\n");
1568		goto out_put_req;
1569	}
1570
1571	req->ki_obj.user = user_iocb;
1572	req->ki_user_data = iocb->aio_data;
1573	req->ki_pos = iocb->aio_offset;
 
1574
1575	req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1576	req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1577	req->ki_opcode = iocb->aio_lio_opcode;
1578
1579	ret = aio_setup_iocb(req, compat);
1580
1581	if (ret)
1582		goto out_put_req;
1583
1584	spin_lock_irq(&ctx->ctx_lock);
1585	/*
1586	 * We could have raced with io_destroy() and are currently holding a
1587	 * reference to ctx which should be destroyed. We cannot submit IO
1588	 * since ctx gets freed as soon as io_submit() puts its reference.  The
1589	 * check here is reliable: io_destroy() sets ctx->dead before waiting
1590	 * for outstanding IO and the barrier between these two is realized by
1591	 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock.  Analogously we
1592	 * increment ctx->reqs_active before checking for ctx->dead and the
1593	 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1594	 * don't see ctx->dead set here, io_destroy() waits for our IO to
1595	 * finish.
1596	 */
1597	if (ctx->dead) {
1598		spin_unlock_irq(&ctx->ctx_lock);
1599		ret = -EINVAL;
1600		goto out_put_req;
1601	}
1602	aio_run_iocb(req);
1603	if (!list_empty(&ctx->run_list)) {
1604		/* drain the run list */
1605		while (__aio_run_iocbs(ctx))
1606			;
1607	}
1608	spin_unlock_irq(&ctx->ctx_lock);
1609
1610	aio_put_req(req);	/* drop extra ref to req */
1611	return 0;
1612
1613out_put_req:
1614	aio_put_req(req);	/* drop extra ref to req */
1615	aio_put_req(req);	/* drop i/o ref to req */
 
1616	return ret;
1617}
1618
1619long do_io_submit(aio_context_t ctx_id, long nr,
1620		  struct iocb __user *__user *iocbpp, bool compat)
1621{
1622	struct kioctx *ctx;
1623	long ret = 0;
1624	int i;
1625	struct blk_plug plug;
1626
1627	if (unlikely(nr < 0))
1628		return -EINVAL;
1629
1630	if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1631		nr = LONG_MAX/sizeof(*iocbpp);
1632
1633	if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1634		return -EFAULT;
1635
1636	ctx = lookup_ioctx(ctx_id);
1637	if (unlikely(!ctx)) {
1638		pr_debug("EINVAL: io_submit: invalid context id\n");
1639		return -EINVAL;
1640	}
1641
1642	blk_start_plug(&plug);
1643
1644	/*
1645	 * AKPM: should this return a partial result if some of the IOs were
1646	 * successfully submitted?
1647	 */
1648	for (i=0; i<nr; i++) {
1649		struct iocb __user *user_iocb;
1650		struct iocb tmp;
1651
1652		if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1653			ret = -EFAULT;
1654			break;
1655		}
1656
1657		if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1658			ret = -EFAULT;
1659			break;
1660		}
1661
1662		ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1663		if (ret)
1664			break;
1665	}
1666	blk_finish_plug(&plug);
1667
1668	put_ioctx(ctx);
1669	return i ? i : ret;
1670}
1671
1672/* sys_io_submit:
1673 *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1674 *	the number of iocbs queued.  May return -EINVAL if the aio_context
1675 *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
1676 *	*iocbpp[0] is not properly initialized, if the operation specified
1677 *	is invalid for the file descriptor in the iocb.  May fail with
1678 *	-EFAULT if any of the data structures point to invalid data.  May
1679 *	fail with -EBADF if the file descriptor specified in the first
1680 *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
1681 *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1682 *	fail with -ENOSYS if not implemented.
1683 */
1684SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1685		struct iocb __user * __user *, iocbpp)
1686{
1687	return do_io_submit(ctx_id, nr, iocbpp, 0);
1688}
1689
1690/* lookup_kiocb
1691 *	Finds a given iocb for cancellation.
1692 */
1693static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1694				  u32 key)
1695{
1696	struct list_head *pos;
1697
1698	assert_spin_locked(&ctx->ctx_lock);
1699
 
 
 
1700	/* TODO: use a hash or array, this sucks. */
1701	list_for_each(pos, &ctx->active_reqs) {
1702		struct kiocb *kiocb = list_kiocb(pos);
1703		if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1704			return kiocb;
1705	}
1706	return NULL;
1707}
1708
1709/* sys_io_cancel:
1710 *	Attempts to cancel an iocb previously passed to io_submit.  If
1711 *	the operation is successfully cancelled, the resulting event is
1712 *	copied into the memory pointed to by result without being placed
1713 *	into the completion queue and 0 is returned.  May fail with
1714 *	-EFAULT if any of the data structures pointed to are invalid.
1715 *	May fail with -EINVAL if aio_context specified by ctx_id is
1716 *	invalid.  May fail with -EAGAIN if the iocb specified was not
1717 *	cancelled.  Will fail with -ENOSYS if not implemented.
1718 */
1719SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1720		struct io_event __user *, result)
1721{
1722	int (*cancel)(struct kiocb *iocb, struct io_event *res);
1723	struct kioctx *ctx;
1724	struct kiocb *kiocb;
1725	u32 key;
1726	int ret;
1727
1728	ret = get_user(key, &iocb->aio_key);
1729	if (unlikely(ret))
1730		return -EFAULT;
1731
1732	ctx = lookup_ioctx(ctx_id);
1733	if (unlikely(!ctx))
1734		return -EINVAL;
1735
1736	spin_lock_irq(&ctx->ctx_lock);
1737	ret = -EAGAIN;
1738	kiocb = lookup_kiocb(ctx, iocb, key);
1739	if (kiocb && kiocb->ki_cancel) {
1740		cancel = kiocb->ki_cancel;
1741		kiocb->ki_users ++;
1742		kiocbSetCancelled(kiocb);
1743	} else
1744		cancel = NULL;
1745	spin_unlock_irq(&ctx->ctx_lock);
1746
1747	if (NULL != cancel) {
1748		struct io_event tmp;
1749		pr_debug("calling cancel\n");
1750		memset(&tmp, 0, sizeof(tmp));
1751		tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1752		tmp.data = kiocb->ki_user_data;
1753		ret = cancel(kiocb, &tmp);
1754		if (!ret) {
1755			/* Cancellation succeeded -- copy the result
1756			 * into the user's buffer.
1757			 */
1758			if (copy_to_user(result, &tmp, sizeof(tmp)))
1759				ret = -EFAULT;
1760		}
1761	} else
1762		ret = -EINVAL;
1763
1764	put_ioctx(ctx);
1765
1766	return ret;
1767}
1768
1769/* io_getevents:
1770 *	Attempts to read at least min_nr events and up to nr events from
1771 *	the completion queue for the aio_context specified by ctx_id. If
1772 *	it succeeds, the number of read events is returned. May fail with
1773 *	-EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1774 *	out of range, if timeout is out of range.  May fail with -EFAULT
1775 *	if any of the memory specified is invalid.  May return 0 or
1776 *	< min_nr if the timeout specified by timeout has elapsed
1777 *	before sufficient events are available, where timeout == NULL
1778 *	specifies an infinite timeout. Note that the timeout pointed to by
1779 *	timeout is relative and will be updated if not NULL and the
1780 *	operation blocks. Will fail with -ENOSYS if not implemented.
1781 */
1782SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1783		long, min_nr,
1784		long, nr,
1785		struct io_event __user *, events,
1786		struct timespec __user *, timeout)
1787{
1788	struct kioctx *ioctx = lookup_ioctx(ctx_id);
1789	long ret = -EINVAL;
1790
1791	if (likely(ioctx)) {
1792		if (likely(min_nr <= nr && min_nr >= 0))
1793			ret = read_events(ioctx, min_nr, nr, events, timeout);
1794		put_ioctx(ioctx);
1795	}
1796
1797	asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1798	return ret;
1799}
v3.15
   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.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 <asm/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	*table[];
  74};
  75
  76struct kioctx_cpu {
  77	unsigned		reqs_available;
  78};
  79
  80struct kioctx {
  81	struct percpu_ref	users;
  82	atomic_t		dead;
  83
  84	struct percpu_ref	reqs;
  85
  86	unsigned long		user_id;
  87
  88	struct __percpu kioctx_cpu *cpu;
  89
  90	/*
  91	 * For percpu reqs_available, number of slots we move to/from global
  92	 * counter at a time:
  93	 */
  94	unsigned		req_batch;
  95	/*
  96	 * This is what userspace passed to io_setup(), it's not used for
  97	 * anything but counting against the global max_reqs quota.
  98	 *
  99	 * The real limit is nr_events - 1, which will be larger (see
 100	 * aio_setup_ring())
 101	 */
 102	unsigned		max_reqs;
 103
 104	/* Size of ringbuffer, in units of struct io_event */
 105	unsigned		nr_events;
 106
 107	unsigned long		mmap_base;
 108	unsigned long		mmap_size;
 109
 110	struct page		**ring_pages;
 111	long			nr_pages;
 112
 113	struct work_struct	free_work;
 114
 115	/*
 116	 * signals when all in-flight requests are done
 117	 */
 118	struct completion *requests_done;
 119
 120	struct {
 121		/*
 122		 * This counts the number of available slots in the ringbuffer,
 123		 * so we avoid overflowing it: it's decremented (if positive)
 124		 * when allocating a kiocb and incremented when the resulting
 125		 * io_event is pulled off the ringbuffer.
 126		 *
 127		 * We batch accesses to it with a percpu version.
 128		 */
 129		atomic_t	reqs_available;
 130	} ____cacheline_aligned_in_smp;
 131
 132	struct {
 133		spinlock_t	ctx_lock;
 134		struct list_head active_reqs;	/* used for cancellation */
 135	} ____cacheline_aligned_in_smp;
 136
 137	struct {
 138		struct mutex	ring_lock;
 139		wait_queue_head_t wait;
 140	} ____cacheline_aligned_in_smp;
 141
 142	struct {
 143		unsigned	tail;
 144		spinlock_t	completion_lock;
 145	} ____cacheline_aligned_in_smp;
 146
 147	struct page		*internal_pages[AIO_RING_PAGES];
 148	struct file		*aio_ring_file;
 149
 150	unsigned		id;
 151};
 152
 153/*------ sysctl variables----*/
 154static DEFINE_SPINLOCK(aio_nr_lock);
 155unsigned long aio_nr;		/* current system wide number of aio requests */
 156unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
 157/*----end sysctl variables---*/
 158
 159static struct kmem_cache	*kiocb_cachep;
 160static struct kmem_cache	*kioctx_cachep;
 161
 162static struct vfsmount *aio_mnt;
 163
 164static const struct file_operations aio_ring_fops;
 165static const struct address_space_operations aio_ctx_aops;
 166
 167static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
 168{
 169	struct qstr this = QSTR_INIT("[aio]", 5);
 170	struct file *file;
 171	struct path path;
 172	struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
 173	if (IS_ERR(inode))
 174		return ERR_CAST(inode);
 175
 176	inode->i_mapping->a_ops = &aio_ctx_aops;
 177	inode->i_mapping->private_data = ctx;
 178	inode->i_size = PAGE_SIZE * nr_pages;
 179
 180	path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
 181	if (!path.dentry) {
 182		iput(inode);
 183		return ERR_PTR(-ENOMEM);
 184	}
 185	path.mnt = mntget(aio_mnt);
 186
 187	d_instantiate(path.dentry, inode);
 188	file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
 189	if (IS_ERR(file)) {
 190		path_put(&path);
 191		return file;
 192	}
 193
 194	file->f_flags = O_RDWR;
 195	file->private_data = ctx;
 196	return file;
 197}
 198
 199static struct dentry *aio_mount(struct file_system_type *fs_type,
 200				int flags, const char *dev_name, void *data)
 201{
 202	static const struct dentry_operations ops = {
 203		.d_dname	= simple_dname,
 204	};
 205	return mount_pseudo(fs_type, "aio:", NULL, &ops, 0xa10a10a1);
 206}
 207
 208/* aio_setup
 209 *	Creates the slab caches used by the aio routines, panic on
 210 *	failure as this is done early during the boot sequence.
 211 */
 212static int __init aio_setup(void)
 213{
 214	static struct file_system_type aio_fs = {
 215		.name		= "aio",
 216		.mount		= aio_mount,
 217		.kill_sb	= kill_anon_super,
 218	};
 219	aio_mnt = kern_mount(&aio_fs);
 220	if (IS_ERR(aio_mnt))
 221		panic("Failed to create aio fs mount.");
 222
 223	kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
 224	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
 225
 226	pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
 
 
 
 227
 228	return 0;
 229}
 230__initcall(aio_setup);
 231
 232static void put_aio_ring_file(struct kioctx *ctx)
 233{
 234	struct file *aio_ring_file = ctx->aio_ring_file;
 235	if (aio_ring_file) {
 236		truncate_setsize(aio_ring_file->f_inode, 0);
 237
 238		/* Prevent further access to the kioctx from migratepages */
 239		spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
 240		aio_ring_file->f_inode->i_mapping->private_data = NULL;
 241		ctx->aio_ring_file = NULL;
 242		spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
 243
 244		fput(aio_ring_file);
 245	}
 246}
 247
 248static void aio_free_ring(struct kioctx *ctx)
 249{
 250	int i;
 
 251
 252	/* Disconnect the kiotx from the ring file.  This prevents future
 253	 * accesses to the kioctx from page migration.
 254	 */
 255	put_aio_ring_file(ctx);
 256
 257	for (i = 0; i < ctx->nr_pages; i++) {
 258		struct page *page;
 259		pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
 260				page_count(ctx->ring_pages[i]));
 261		page = ctx->ring_pages[i];
 262		if (!page)
 263			continue;
 264		ctx->ring_pages[i] = NULL;
 265		put_page(page);
 266	}
 267
 268	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
 269		kfree(ctx->ring_pages);
 270		ctx->ring_pages = NULL;
 271	}
 
 
 
 
 
 
 272}
 273
 274static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
 275{
 276	vma->vm_ops = &generic_file_vm_ops;
 277	return 0;
 278}
 279
 280static const struct file_operations aio_ring_fops = {
 281	.mmap = aio_ring_mmap,
 282};
 283
 284static int aio_set_page_dirty(struct page *page)
 285{
 286	return 0;
 287}
 288
 289#if IS_ENABLED(CONFIG_MIGRATION)
 290static int aio_migratepage(struct address_space *mapping, struct page *new,
 291			struct page *old, enum migrate_mode mode)
 292{
 293	struct kioctx *ctx;
 294	unsigned long flags;
 295	pgoff_t idx;
 296	int rc;
 297
 298	rc = 0;
 299
 300	/* mapping->private_lock here protects against the kioctx teardown.  */
 301	spin_lock(&mapping->private_lock);
 302	ctx = mapping->private_data;
 303	if (!ctx) {
 304		rc = -EINVAL;
 305		goto out;
 306	}
 307
 308	/* The ring_lock mutex.  The prevents aio_read_events() from writing
 309	 * to the ring's head, and prevents page migration from mucking in
 310	 * a partially initialized kiotx.
 311	 */
 312	if (!mutex_trylock(&ctx->ring_lock)) {
 313		rc = -EAGAIN;
 314		goto out;
 315	}
 316
 317	idx = old->index;
 318	if (idx < (pgoff_t)ctx->nr_pages) {
 319		/* Make sure the old page hasn't already been changed */
 320		if (ctx->ring_pages[idx] != old)
 321			rc = -EAGAIN;
 322	} else
 323		rc = -EINVAL;
 324
 325	if (rc != 0)
 326		goto out_unlock;
 327
 328	/* Writeback must be complete */
 329	BUG_ON(PageWriteback(old));
 330	get_page(new);
 331
 332	rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
 333	if (rc != MIGRATEPAGE_SUCCESS) {
 334		put_page(new);
 335		goto out_unlock;
 336	}
 337
 338	/* Take completion_lock to prevent other writes to the ring buffer
 339	 * while the old page is copied to the new.  This prevents new
 340	 * events from being lost.
 341	 */
 342	spin_lock_irqsave(&ctx->completion_lock, flags);
 343	migrate_page_copy(new, old);
 344	BUG_ON(ctx->ring_pages[idx] != old);
 345	ctx->ring_pages[idx] = new;
 346	spin_unlock_irqrestore(&ctx->completion_lock, flags);
 347
 348	/* The old page is no longer accessible. */
 349	put_page(old);
 350
 351out_unlock:
 352	mutex_unlock(&ctx->ring_lock);
 353out:
 354	spin_unlock(&mapping->private_lock);
 355	return rc;
 356}
 357#endif
 358
 359static const struct address_space_operations aio_ctx_aops = {
 360	.set_page_dirty = aio_set_page_dirty,
 361#if IS_ENABLED(CONFIG_MIGRATION)
 362	.migratepage	= aio_migratepage,
 363#endif
 364};
 365
 366static int aio_setup_ring(struct kioctx *ctx)
 367{
 368	struct aio_ring *ring;
 
 369	unsigned nr_events = ctx->max_reqs;
 370	struct mm_struct *mm = current->mm;
 371	unsigned long size, unused;
 372	int nr_pages;
 373	int i;
 374	struct file *file;
 375
 376	/* Compensate for the ring buffer's head/tail overlap entry */
 377	nr_events += 2;	/* 1 is required, 2 for good luck */
 378
 379	size = sizeof(struct aio_ring);
 380	size += sizeof(struct io_event) * nr_events;
 
 381
 382	nr_pages = PFN_UP(size);
 383	if (nr_pages < 0)
 384		return -EINVAL;
 385
 386	file = aio_private_file(ctx, nr_pages);
 387	if (IS_ERR(file)) {
 388		ctx->aio_ring_file = NULL;
 389		return -ENOMEM;
 390	}
 391
 392	ctx->aio_ring_file = file;
 393	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
 394			/ sizeof(struct io_event);
 395
 396	ctx->ring_pages = ctx->internal_pages;
 
 397	if (nr_pages > AIO_RING_PAGES) {
 398		ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
 399					  GFP_KERNEL);
 400		if (!ctx->ring_pages) {
 401			put_aio_ring_file(ctx);
 402			return -ENOMEM;
 403		}
 404	}
 405
 406	for (i = 0; i < nr_pages; i++) {
 407		struct page *page;
 408		page = find_or_create_page(file->f_inode->i_mapping,
 409					   i, GFP_HIGHUSER | __GFP_ZERO);
 410		if (!page)
 411			break;
 412		pr_debug("pid(%d) page[%d]->count=%d\n",
 413			 current->pid, i, page_count(page));
 414		SetPageUptodate(page);
 415		SetPageDirty(page);
 416		unlock_page(page);
 417
 418		ctx->ring_pages[i] = page;
 419	}
 420	ctx->nr_pages = i;
 421
 422	if (unlikely(i != nr_pages)) {
 423		aio_free_ring(ctx);
 424		return -ENOMEM;
 425	}
 426
 427	ctx->mmap_size = nr_pages * PAGE_SIZE;
 428	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
 
 
 
 429
 430	down_write(&mm->mmap_sem);
 431	ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
 432				       PROT_READ | PROT_WRITE,
 433				       MAP_SHARED, 0, &unused);
 434	up_write(&mm->mmap_sem);
 435	if (IS_ERR((void *)ctx->mmap_base)) {
 436		ctx->mmap_size = 0;
 437		aio_free_ring(ctx);
 438		return -ENOMEM;
 439	}
 440
 441	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
 442
 443	ctx->user_id = ctx->mmap_base;
 444	ctx->nr_events = nr_events; /* trusted copy */
 445
 446	ring = kmap_atomic(ctx->ring_pages[0]);
 447	ring->nr = nr_events;	/* user copy */
 448	ring->id = ~0U;
 449	ring->head = ring->tail = 0;
 450	ring->magic = AIO_RING_MAGIC;
 451	ring->compat_features = AIO_RING_COMPAT_FEATURES;
 452	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
 453	ring->header_length = sizeof(struct aio_ring);
 454	kunmap_atomic(ring);
 455	flush_dcache_page(ctx->ring_pages[0]);
 456
 457	return 0;
 458}
 459
 
 
 
 
 460#define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
 461#define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
 462#define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
 463
 464void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
 465{
 466	struct kioctx *ctx = req->ki_ctx;
 467	unsigned long flags;
 468
 469	spin_lock_irqsave(&ctx->ctx_lock, flags);
 470
 471	if (!req->ki_list.next)
 472		list_add(&req->ki_list, &ctx->active_reqs);
 
 
 
 
 
 473
 474	req->ki_cancel = cancel;
 475
 476	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 477}
 478EXPORT_SYMBOL(kiocb_set_cancel_fn);
 479
 480static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
 481{
 482	kiocb_cancel_fn *old, *cancel;
 
 483
 484	/*
 485	 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
 486	 * actually has a cancel function, hence the cmpxchg()
 487	 */
 488
 489	cancel = ACCESS_ONCE(kiocb->ki_cancel);
 490	do {
 491		if (!cancel || cancel == KIOCB_CANCELLED)
 492			return -EINVAL;
 493
 494		old = cancel;
 495		cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
 496	} while (cancel != old);
 497
 498	return cancel(kiocb);
 499}
 500
 501static void free_ioctx(struct work_struct *work)
 
 
 
 
 502{
 503	struct kioctx *ctx = container_of(work, struct kioctx, free_work);
 504
 505	pr_debug("freeing %p\n", ctx);
 506
 
 
 507	aio_free_ring(ctx);
 508	free_percpu(ctx->cpu);
 509	kmem_cache_free(kioctx_cachep, ctx);
 
 
 510}
 511
 512static void free_ioctx_reqs(struct percpu_ref *ref)
 513{
 514	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
 515
 516	/* At this point we know that there are no any in-flight requests */
 517	if (ctx->requests_done)
 518		complete(ctx->requests_done);
 519
 520	INIT_WORK(&ctx->free_work, free_ioctx);
 521	schedule_work(&ctx->free_work);
 522}
 523
 524/*
 525 * When this function runs, the kioctx has been removed from the "hash table"
 526 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
 527 * now it's safe to cancel any that need to be.
 528 */
 529static void free_ioctx_users(struct percpu_ref *ref)
 530{
 531	struct kioctx *ctx = container_of(ref, struct kioctx, users);
 532	struct kiocb *req;
 533
 534	spin_lock_irq(&ctx->ctx_lock);
 535
 536	while (!list_empty(&ctx->active_reqs)) {
 537		req = list_first_entry(&ctx->active_reqs,
 538				       struct kiocb, ki_list);
 539
 540		list_del_init(&req->ki_list);
 541		kiocb_cancel(ctx, req);
 542	}
 543
 544	spin_unlock_irq(&ctx->ctx_lock);
 545
 546	percpu_ref_kill(&ctx->reqs);
 547	percpu_ref_put(&ctx->reqs);
 548}
 549
 550static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
 551{
 552	unsigned i, new_nr;
 553	struct kioctx_table *table, *old;
 554	struct aio_ring *ring;
 555
 556	spin_lock(&mm->ioctx_lock);
 557	rcu_read_lock();
 558	table = rcu_dereference(mm->ioctx_table);
 559
 560	while (1) {
 561		if (table)
 562			for (i = 0; i < table->nr; i++)
 563				if (!table->table[i]) {
 564					ctx->id = i;
 565					table->table[i] = ctx;
 566					rcu_read_unlock();
 567					spin_unlock(&mm->ioctx_lock);
 568
 569					/* While kioctx setup is in progress,
 570					 * we are protected from page migration
 571					 * changes ring_pages by ->ring_lock.
 572					 */
 573					ring = kmap_atomic(ctx->ring_pages[0]);
 574					ring->id = ctx->id;
 575					kunmap_atomic(ring);
 576					return 0;
 577				}
 578
 579		new_nr = (table ? table->nr : 1) * 4;
 580
 581		rcu_read_unlock();
 582		spin_unlock(&mm->ioctx_lock);
 583
 584		table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
 585				new_nr, GFP_KERNEL);
 586		if (!table)
 587			return -ENOMEM;
 588
 589		table->nr = new_nr;
 590
 591		spin_lock(&mm->ioctx_lock);
 592		rcu_read_lock();
 593		old = rcu_dereference(mm->ioctx_table);
 594
 595		if (!old) {
 596			rcu_assign_pointer(mm->ioctx_table, table);
 597		} else if (table->nr > old->nr) {
 598			memcpy(table->table, old->table,
 599			       old->nr * sizeof(struct kioctx *));
 600
 601			rcu_assign_pointer(mm->ioctx_table, table);
 602			kfree_rcu(old, rcu);
 603		} else {
 604			kfree(table);
 605			table = old;
 606		}
 607	}
 608}
 609
 610static void aio_nr_sub(unsigned nr)
 611{
 612	spin_lock(&aio_nr_lock);
 613	if (WARN_ON(aio_nr - nr > aio_nr))
 614		aio_nr = 0;
 615	else
 616		aio_nr -= nr;
 617	spin_unlock(&aio_nr_lock);
 618}
 619
 620/* ioctx_alloc
 621 *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
 622 */
 623static struct kioctx *ioctx_alloc(unsigned nr_events)
 624{
 625	struct mm_struct *mm = current->mm;
 626	struct kioctx *ctx;
 627	int err = -ENOMEM;
 628
 629	/*
 630	 * We keep track of the number of available ringbuffer slots, to prevent
 631	 * overflow (reqs_available), and we also use percpu counters for this.
 632	 *
 633	 * So since up to half the slots might be on other cpu's percpu counters
 634	 * and unavailable, double nr_events so userspace sees what they
 635	 * expected: additionally, we move req_batch slots to/from percpu
 636	 * counters at a time, so make sure that isn't 0:
 637	 */
 638	nr_events = max(nr_events, num_possible_cpus() * 4);
 639	nr_events *= 2;
 640
 641	/* Prevent overflows */
 642	if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
 643	    (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
 644		pr_debug("ENOMEM: nr_events too high\n");
 645		return ERR_PTR(-EINVAL);
 646	}
 647
 648	if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
 649		return ERR_PTR(-EAGAIN);
 650
 651	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
 652	if (!ctx)
 653		return ERR_PTR(-ENOMEM);
 654
 655	ctx->max_reqs = nr_events;
 
 
 656
 
 657	spin_lock_init(&ctx->ctx_lock);
 658	spin_lock_init(&ctx->completion_lock);
 659	mutex_init(&ctx->ring_lock);
 660	/* Protect against page migration throughout kiotx setup by keeping
 661	 * the ring_lock mutex held until setup is complete. */
 662	mutex_lock(&ctx->ring_lock);
 663	init_waitqueue_head(&ctx->wait);
 664
 665	INIT_LIST_HEAD(&ctx->active_reqs);
 
 
 666
 667	if (percpu_ref_init(&ctx->users, free_ioctx_users))
 668		goto err;
 669
 670	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs))
 671		goto err;
 672
 673	ctx->cpu = alloc_percpu(struct kioctx_cpu);
 674	if (!ctx->cpu)
 675		goto err;
 676
 677	err = aio_setup_ring(ctx);
 678	if (err < 0)
 679		goto err;
 680
 681	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
 682	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
 683	if (ctx->req_batch < 1)
 684		ctx->req_batch = 1;
 685
 686	/* limit the number of system wide aios */
 687	spin_lock(&aio_nr_lock);
 688	if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
 689	    aio_nr + nr_events < aio_nr) {
 690		spin_unlock(&aio_nr_lock);
 691		err = -EAGAIN;
 692		goto err_ctx;
 693	}
 694	aio_nr += ctx->max_reqs;
 695	spin_unlock(&aio_nr_lock);
 
 696
 697	percpu_ref_get(&ctx->users);	/* io_setup() will drop this ref */
 698	percpu_ref_get(&ctx->reqs);	/* free_ioctx_users() will drop this */
 
 
 
 699
 700	err = ioctx_add_table(ctx, mm);
 701	if (err)
 702		goto err_cleanup;
 703
 704	/* Release the ring_lock mutex now that all setup is complete. */
 705	mutex_unlock(&ctx->ring_lock);
 
 
 706
 707	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
 708		 ctx, ctx->user_id, mm, ctx->nr_events);
 709	return ctx;
 710
 711err_cleanup:
 712	aio_nr_sub(ctx->max_reqs);
 713err_ctx:
 714	aio_free_ring(ctx);
 715err:
 716	mutex_unlock(&ctx->ring_lock);
 717	free_percpu(ctx->cpu);
 718	free_percpu(ctx->reqs.pcpu_count);
 719	free_percpu(ctx->users.pcpu_count);
 720	kmem_cache_free(kioctx_cachep, ctx);
 721	pr_debug("error allocating ioctx %d\n", err);
 722	return ERR_PTR(err);
 
 
 723}
 724
 725/* kill_ioctx
 726 *	Cancels all outstanding aio requests on an aio context.  Used
 727 *	when the processes owning a context have all exited to encourage
 728 *	the rapid destruction of the kioctx.
 729 */
 730static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
 731		struct completion *requests_done)
 732{
 733	if (!atomic_xchg(&ctx->dead, 1)) {
 734		struct kioctx_table *table;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 735
 736		spin_lock(&mm->ioctx_lock);
 737		rcu_read_lock();
 738		table = rcu_dereference(mm->ioctx_table);
 739
 740		WARN_ON(ctx != table->table[ctx->id]);
 741		table->table[ctx->id] = NULL;
 742		rcu_read_unlock();
 743		spin_unlock(&mm->ioctx_lock);
 744
 745		/* percpu_ref_kill() will do the necessary call_rcu() */
 746		wake_up_all(&ctx->wait);
 
 747
 748		/*
 749		 * It'd be more correct to do this in free_ioctx(), after all
 750		 * the outstanding kiocbs have finished - but by then io_destroy
 751		 * has already returned, so io_setup() could potentially return
 752		 * -EAGAIN with no ioctxs actually in use (as far as userspace
 753		 *  could tell).
 754		 */
 755		aio_nr_sub(ctx->max_reqs);
 
 
 756
 757		if (ctx->mmap_size)
 758			vm_munmap(ctx->mmap_base, ctx->mmap_size);
 759
 760		ctx->requests_done = requests_done;
 761		percpu_ref_kill(&ctx->users);
 762	} else {
 763		if (requests_done)
 764			complete(requests_done);
 765	}
 766}
 767
 768/* wait_on_sync_kiocb:
 769 *	Waits on the given sync kiocb to complete.
 770 */
 771ssize_t wait_on_sync_kiocb(struct kiocb *req)
 772{
 773	while (!req->ki_ctx) {
 774		set_current_state(TASK_UNINTERRUPTIBLE);
 775		if (req->ki_ctx)
 776			break;
 777		io_schedule();
 778	}
 779	__set_current_state(TASK_RUNNING);
 780	return req->ki_user_data;
 781}
 782EXPORT_SYMBOL(wait_on_sync_kiocb);
 783
 784/*
 785 * exit_aio: called when the last user of mm goes away.  At this point, there is
 786 * no way for any new requests to be submited or any of the io_* syscalls to be
 787 * called on the context.
 788 *
 789 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
 790 * them.
 791 */
 792void exit_aio(struct mm_struct *mm)
 793{
 794	struct kioctx_table *table;
 795	struct kioctx *ctx;
 796	unsigned i = 0;
 797
 798	while (1) {
 799		rcu_read_lock();
 800		table = rcu_dereference(mm->ioctx_table);
 801
 802		do {
 803			if (!table || i >= table->nr) {
 804				rcu_read_unlock();
 805				rcu_assign_pointer(mm->ioctx_table, NULL);
 806				if (table)
 807					kfree(table);
 808				return;
 809			}
 810
 811			ctx = table->table[i++];
 812		} while (!ctx);
 813
 814		rcu_read_unlock();
 815
 
 816		/*
 817		 * We don't need to bother with munmap() here -
 818		 * exit_mmap(mm) is coming and it'll unmap everything.
 819		 * Since aio_free_ring() uses non-zero ->mmap_size
 820		 * as indicator that it needs to unmap the area,
 821		 * just set it to 0; aio_free_ring() is the only
 822		 * place that uses ->mmap_size, so it's safe.
 823		 */
 824		ctx->mmap_size = 0;
 825
 826		kill_ioctx(mm, ctx, NULL);
 
 
 
 
 
 827	}
 828}
 829
 830static void put_reqs_available(struct kioctx *ctx, unsigned nr)
 
 
 
 
 
 
 
 
 
 
 831{
 832	struct kioctx_cpu *kcpu;
 
 
 
 
 
 
 833
 834	preempt_disable();
 835	kcpu = this_cpu_ptr(ctx->cpu);
 
 
 
 
 
 
 
 
 
 836
 837	kcpu->reqs_available += nr;
 838	while (kcpu->reqs_available >= ctx->req_batch * 2) {
 839		kcpu->reqs_available -= ctx->req_batch;
 840		atomic_add(ctx->req_batch, &ctx->reqs_available);
 
 
 
 
 
 841	}
 
 
 
 
 
 
 
 
 
 
 842
 843	preempt_enable();
 
 
 
 
 
 
 
 
 
 
 
 
 844}
 845
 846static bool get_reqs_available(struct kioctx *ctx)
 847{
 848	struct kioctx_cpu *kcpu;
 849	bool ret = false;
 850
 851	preempt_disable();
 852	kcpu = this_cpu_ptr(ctx->cpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 853
 854	if (!kcpu->reqs_available) {
 855		int old, avail = atomic_read(&ctx->reqs_available);
 856
 857		do {
 858			if (avail < ctx->req_batch)
 859				goto out;
 860
 861			old = avail;
 862			avail = atomic_cmpxchg(&ctx->reqs_available,
 863					       avail, avail - ctx->req_batch);
 864		} while (avail != old);
 865
 866		kcpu->reqs_available += ctx->req_batch;
 
 867	}
 868
 869	ret = true;
 870	kcpu->reqs_available--;
 871out:
 872	preempt_enable();
 873	return ret;
 874}
 875
 876/* aio_get_req
 877 *	Allocate a slot for an aio request.
 878 * Returns NULL if no requests are free.
 879 */
 880static inline struct kiocb *aio_get_req(struct kioctx *ctx)
 881{
 882	struct kiocb *req;
 
 883
 884	if (!get_reqs_available(ctx))
 885		return NULL;
 886
 887	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
 888	if (unlikely(!req))
 889		goto out_put;
 
 
 
 
 890
 891	percpu_ref_get(&ctx->reqs);
 892
 893	req->ki_ctx = ctx;
 894	return req;
 895out_put:
 896	put_reqs_available(ctx, 1);
 897	return NULL;
 
 
 
 
 
 
 
 
 
 
 898}
 899
 900static void kiocb_free(struct kiocb *req)
 
 
 
 
 901{
 902	if (req->ki_filp)
 903		fput(req->ki_filp);
 904	if (req->ki_eventfd != NULL)
 905		eventfd_ctx_put(req->ki_eventfd);
 906	kmem_cache_free(kiocb_cachep, req);
 
 907}
 
 908
 909static struct kioctx *lookup_ioctx(unsigned long ctx_id)
 910{
 911	struct aio_ring __user *ring  = (void __user *)ctx_id;
 912	struct mm_struct *mm = current->mm;
 913	struct kioctx *ctx, *ret = NULL;
 914	struct kioctx_table *table;
 915	unsigned id;
 916
 917	if (get_user(id, &ring->id))
 918		return NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 919
 920	rcu_read_lock();
 921	table = rcu_dereference(mm->ioctx_table);
 
 
 
 
 
 
 922
 923	if (!table || id >= table->nr)
 
 
 
 
 924		goto out;
 
 925
 926	ctx = table->table[id];
 927	if (ctx && ctx->user_id == ctx_id) {
 928		percpu_ref_get(&ctx->users);
 929		ret = ctx;
 
 
 
 
 
 
 
 
 
 
 
 930	}
 931out:
 932	rcu_read_unlock();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 933	return ret;
 934}
 935
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 936/* aio_complete
 937 *	Called when the io request on the given iocb is complete.
 
 
 938 */
 939void aio_complete(struct kiocb *iocb, long res, long res2)
 940{
 941	struct kioctx	*ctx = iocb->ki_ctx;
 
 942	struct aio_ring	*ring;
 943	struct io_event	*ev_page, *event;
 944	unsigned long	flags;
 945	unsigned tail, pos;
 
 946
 947	/*
 948	 * Special case handling for sync iocbs:
 949	 *  - events go directly into the iocb for fast handling
 950	 *  - the sync task with the iocb in its stack holds the single iocb
 951	 *    ref, no other paths have a way to get another ref
 952	 *  - the sync task helpfully left a reference to itself in the iocb
 953	 */
 954	if (is_sync_kiocb(iocb)) {
 
 955		iocb->ki_user_data = res;
 956		smp_wmb();
 957		iocb->ki_ctx = ERR_PTR(-EXDEV);
 958		wake_up_process(iocb->ki_obj.tsk);
 959		return;
 960	}
 961
 962	if (iocb->ki_list.next) {
 963		unsigned long flags;
 964
 965		spin_lock_irqsave(&ctx->ctx_lock, flags);
 966		list_del(&iocb->ki_list);
 967		spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 968	}
 
 
 
 
 
 
 969
 970	/*
 971	 * Add a completion event to the ring buffer. Must be done holding
 972	 * ctx->completion_lock to prevent other code from messing with the tail
 973	 * pointer since we might be called from irq context.
 974	 */
 975	spin_lock_irqsave(&ctx->completion_lock, flags);
 
 976
 977	tail = ctx->tail;
 978	pos = tail + AIO_EVENTS_OFFSET;
 979
 980	if (++tail >= ctx->nr_events)
 
 
 981		tail = 0;
 982
 983	ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
 984	event = ev_page + pos % AIO_EVENTS_PER_PAGE;
 985
 986	event->obj = (u64)(unsigned long)iocb->ki_obj.user;
 987	event->data = iocb->ki_user_data;
 988	event->res = res;
 989	event->res2 = res2;
 990
 991	kunmap_atomic(ev_page);
 992	flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
 993
 994	pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
 995		 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
 996		 res, res2);
 997
 998	/* after flagging the request as done, we
 999	 * must never even look at it again
1000	 */
1001	smp_wmb();	/* make event visible before updating tail */
1002
1003	ctx->tail = tail;
1004
1005	ring = kmap_atomic(ctx->ring_pages[0]);
1006	ring->tail = tail;
1007	kunmap_atomic(ring);
1008	flush_dcache_page(ctx->ring_pages[0]);
1009
1010	spin_unlock_irqrestore(&ctx->completion_lock, flags);
 
1011
1012	pr_debug("added to ring %p at [%u]\n", iocb, tail);
1013
1014	/*
1015	 * Check if the user asked us to deliver the result through an
1016	 * eventfd. The eventfd_signal() function is safe to be called
1017	 * from IRQ context.
1018	 */
1019	if (iocb->ki_eventfd != NULL)
1020		eventfd_signal(iocb->ki_eventfd, 1);
1021
 
1022	/* everything turned out well, dispose of the aiocb. */
1023	kiocb_free(iocb);
1024
1025	/*
1026	 * We have to order our ring_info tail store above and test
1027	 * of the wait list below outside the wait lock.  This is
1028	 * like in wake_up_bit() where clearing a bit has to be
1029	 * ordered with the unlocked test.
1030	 */
1031	smp_mb();
1032
1033	if (waitqueue_active(&ctx->wait))
1034		wake_up(&ctx->wait);
1035
1036	percpu_ref_put(&ctx->reqs);
 
1037}
1038EXPORT_SYMBOL(aio_complete);
1039
1040/* aio_read_events
1041 *	Pull an event off of the ioctx's event ring.  Returns the number of
1042 *	events fetched
 
 
1043 */
1044static long aio_read_events_ring(struct kioctx *ctx,
1045				 struct io_event __user *event, long nr)
1046{
 
1047	struct aio_ring *ring;
1048	unsigned head, tail, pos;
1049	long ret = 0;
1050	int copy_ret;
 
 
 
 
1051
1052	mutex_lock(&ctx->ring_lock);
 
1053
1054	/* Access to ->ring_pages here is protected by ctx->ring_lock. */
1055	ring = kmap_atomic(ctx->ring_pages[0]);
1056	head = ring->head;
1057	tail = ring->tail;
1058	kunmap_atomic(ring);
1059
1060	pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
 
 
 
 
 
 
 
 
 
 
1061
1062	if (head == tail)
1063		goto out;
 
 
 
 
1064
1065	while (ret < nr) {
1066		long avail;
1067		struct io_event *ev;
1068		struct page *page;
 
1069
1070		avail = (head <= tail ?  tail : ctx->nr_events) - head;
1071		if (head == tail)
1072			break;
 
 
 
 
1073
1074		avail = min(avail, nr - ret);
1075		avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1076			    ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
 
 
 
1077
1078		pos = head + AIO_EVENTS_OFFSET;
1079		page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1080		pos %= AIO_EVENTS_PER_PAGE;
 
 
 
 
 
 
1081
1082		ev = kmap(page);
1083		copy_ret = copy_to_user(event + ret, ev + pos,
1084					sizeof(*ev) * avail);
1085		kunmap(page);
1086
1087		if (unlikely(copy_ret)) {
1088			ret = -EFAULT;
1089			goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1090		}
 
1091
1092		ret += avail;
1093		head += avail;
1094		head %= ctx->nr_events;
1095	}
1096
1097	ring = kmap_atomic(ctx->ring_pages[0]);
1098	ring->head = head;
1099	kunmap_atomic(ring);
1100	flush_dcache_page(ctx->ring_pages[0]);
 
 
1101
1102	pr_debug("%li  h%u t%u\n", ret, head, tail);
 
 
 
 
 
1103
1104	put_reqs_available(ctx, ret);
1105out:
1106	mutex_unlock(&ctx->ring_lock);
 
 
 
 
 
 
1107
1108	return ret;
1109}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1110
1111static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1112			    struct io_event __user *event, long *i)
1113{
1114	long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1115
1116	if (ret > 0)
1117		*i += ret;
1118
1119	if (unlikely(atomic_read(&ctx->dead)))
1120		ret = -EINVAL;
 
 
 
1121
1122	if (!*i)
1123		*i = ret;
 
 
1124
1125	return ret < 0 || *i >= min_nr;
 
 
 
 
1126}
1127
1128static long read_events(struct kioctx *ctx, long min_nr, long nr,
1129			struct io_event __user *event,
1130			struct timespec __user *timeout)
 
1131{
1132	ktime_t until = { .tv64 = KTIME_MAX };
1133	long ret = 0;
1134
1135	if (timeout) {
1136		struct timespec	ts;
1137
1138		if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1139			return -EFAULT;
 
 
 
 
 
1140
1141		until = timespec_to_ktime(ts);
1142	}
1143
1144	/*
1145	 * Note that aio_read_events() is being called as the conditional - i.e.
1146	 * we're calling it after prepare_to_wait() has set task state to
1147	 * TASK_INTERRUPTIBLE.
1148	 *
1149	 * But aio_read_events() can block, and if it blocks it's going to flip
1150	 * the task state back to TASK_RUNNING.
1151	 *
1152	 * This should be ok, provided it doesn't flip the state back to
1153	 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1154	 * will only happen if the mutex_lock() call blocks, and we then find
1155	 * the ringbuffer empty. So in practice we should be ok, but it's
1156	 * something to be aware of when touching this code.
1157	 */
1158	wait_event_interruptible_hrtimeout(ctx->wait,
1159			aio_read_events(ctx, min_nr, nr, event, &ret), until);
1160
1161	if (!ret && signal_pending(current))
1162		ret = -EINTR;
1163
1164	return ret;
1165}
1166
1167/* sys_io_setup:
1168 *	Create an aio_context capable of receiving at least nr_events.
1169 *	ctxp must not point to an aio_context that already exists, and
1170 *	must be initialized to 0 prior to the call.  On successful
1171 *	creation of the aio_context, *ctxp is filled in with the resulting 
1172 *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1173 *	if the specified nr_events exceeds internal limits.  May fail 
1174 *	with -EAGAIN if the specified nr_events exceeds the user's limit 
1175 *	of available events.  May fail with -ENOMEM if insufficient kernel
1176 *	resources are available.  May fail with -EFAULT if an invalid
1177 *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1178 *	implemented.
1179 */
1180SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1181{
1182	struct kioctx *ioctx = NULL;
1183	unsigned long ctx;
1184	long ret;
1185
1186	ret = get_user(ctx, ctxp);
1187	if (unlikely(ret))
1188		goto out;
1189
1190	ret = -EINVAL;
1191	if (unlikely(ctx || nr_events == 0)) {
1192		pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1193		         ctx, nr_events);
1194		goto out;
1195	}
1196
1197	ioctx = ioctx_alloc(nr_events);
1198	ret = PTR_ERR(ioctx);
1199	if (!IS_ERR(ioctx)) {
1200		ret = put_user(ioctx->user_id, ctxp);
1201		if (ret)
1202			kill_ioctx(current->mm, ioctx, NULL);
1203		percpu_ref_put(&ioctx->users);
 
 
1204	}
1205
1206out:
1207	return ret;
1208}
1209
1210/* sys_io_destroy:
1211 *	Destroy the aio_context specified.  May cancel any outstanding 
1212 *	AIOs and block on completion.  Will fail with -ENOSYS if not
1213 *	implemented.  May fail with -EINVAL if the context pointed to
1214 *	is invalid.
1215 */
1216SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1217{
1218	struct kioctx *ioctx = lookup_ioctx(ctx);
1219	if (likely(NULL != ioctx)) {
1220		struct completion requests_done =
1221			COMPLETION_INITIALIZER_ONSTACK(requests_done);
 
 
 
 
1222
1223		/* Pass requests_done to kill_ioctx() where it can be set
1224		 * in a thread-safe way. If we try to set it here then we have
1225		 * a race condition if two io_destroy() called simultaneously.
1226		 */
1227		kill_ioctx(current->mm, ioctx, &requests_done);
1228		percpu_ref_put(&ioctx->users);
1229
1230		/* Wait until all IO for the context are done. Otherwise kernel
1231		 * keep using user-space buffers even if user thinks the context
1232		 * is destroyed.
1233		 */
1234		wait_for_completion(&requests_done);
 
 
 
 
 
 
1235
1236		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1237	}
1238	pr_debug("EINVAL: io_destroy: invalid context id\n");
1239	return -EINVAL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1240}
1241
1242typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1243			    unsigned long, loff_t);
 
 
 
 
 
 
 
1244
1245static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1246				     int rw, char __user *buf,
1247				     unsigned long *nr_segs,
1248				     struct iovec **iovec,
1249				     bool compat)
1250{
1251	ssize_t ret;
1252
1253	*nr_segs = kiocb->ki_nbytes;
1254
1255#ifdef CONFIG_COMPAT
1256	if (compat)
1257		ret = compat_rw_copy_check_uvector(rw,
1258				(struct compat_iovec __user *)buf,
1259				*nr_segs, 1, *iovec, iovec);
 
1260	else
1261#endif
1262		ret = rw_copy_check_uvector(rw,
1263				(struct iovec __user *)buf,
1264				*nr_segs, 1, *iovec, iovec);
 
1265	if (ret < 0)
1266		return ret;
1267
1268	/* ki_nbytes now reflect bytes instead of segs */
 
 
1269	kiocb->ki_nbytes = ret;
1270	return 0;
 
 
 
 
1271}
1272
1273static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1274				       int rw, char __user *buf,
1275				       unsigned long *nr_segs,
1276				       struct iovec *iovec)
1277{
1278	if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1279		return -EFAULT;
1280
1281	iovec->iov_base = buf;
1282	iovec->iov_len = kiocb->ki_nbytes;
1283	*nr_segs = 1;
1284	return 0;
1285}
1286
1287/*
1288 * aio_setup_iocb:
1289 *	Performs the initial checks and aio retry method
1290 *	setup for the kiocb at the time of io submission.
1291 */
1292static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1293			    char __user *buf, bool compat)
1294{
1295	struct file *file = req->ki_filp;
1296	ssize_t ret;
1297	unsigned long nr_segs;
1298	int rw;
1299	fmode_t mode;
1300	aio_rw_op *rw_op;
1301	struct iovec inline_vec, *iovec = &inline_vec;
1302
1303	switch (opcode) {
1304	case IOCB_CMD_PREAD:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1305	case IOCB_CMD_PREADV:
1306		mode	= FMODE_READ;
1307		rw	= READ;
1308		rw_op	= file->f_op->aio_read;
1309		goto rw_common;
1310
1311	case IOCB_CMD_PWRITE:
 
 
 
 
 
 
 
1312	case IOCB_CMD_PWRITEV:
1313		mode	= FMODE_WRITE;
1314		rw	= WRITE;
1315		rw_op	= file->f_op->aio_write;
1316		goto rw_common;
1317rw_common:
1318		if (unlikely(!(file->f_mode & mode)))
1319			return -EBADF;
1320
1321		if (!rw_op)
1322			return -EINVAL;
1323
1324		ret = (opcode == IOCB_CMD_PREADV ||
1325		       opcode == IOCB_CMD_PWRITEV)
1326			? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1327						&iovec, compat)
1328			: aio_setup_single_vector(req, rw, buf, &nr_segs,
1329						  iovec);
1330		if (!ret)
1331			ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1332		if (ret < 0) {
1333			if (iovec != &inline_vec)
1334				kfree(iovec);
1335			return ret;
1336		}
1337
1338		req->ki_nbytes = ret;
1339
1340		/* XXX: move/kill - rw_verify_area()? */
1341		/* This matches the pread()/pwrite() logic */
1342		if (req->ki_pos < 0) {
1343			ret = -EINVAL;
1344			break;
1345		}
1346
1347		if (rw == WRITE)
1348			file_start_write(file);
1349
1350		ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1351
1352		if (rw == WRITE)
1353			file_end_write(file);
1354		break;
1355
1356	case IOCB_CMD_FDSYNC:
1357		if (!file->f_op->aio_fsync)
1358			return -EINVAL;
1359
1360		ret = file->f_op->aio_fsync(req, 1);
1361		break;
1362
1363	case IOCB_CMD_FSYNC:
1364		if (!file->f_op->aio_fsync)
1365			return -EINVAL;
1366
1367		ret = file->f_op->aio_fsync(req, 0);
1368		break;
1369
1370	default:
1371		pr_debug("EINVAL: no operation provided\n");
1372		return -EINVAL;
1373	}
1374
1375	if (iovec != &inline_vec)
1376		kfree(iovec);
1377
1378	if (ret != -EIOCBQUEUED) {
1379		/*
1380		 * There's no easy way to restart the syscall since other AIO's
1381		 * may be already running. Just fail this IO with EINTR.
1382		 */
1383		if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1384			     ret == -ERESTARTNOHAND ||
1385			     ret == -ERESTART_RESTARTBLOCK))
1386			ret = -EINTR;
1387		aio_complete(req, ret, 0);
1388	}
1389
1390	return 0;
1391}
1392
1393static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1394			 struct iocb *iocb, bool compat)
1395{
1396	struct kiocb *req;
 
1397	ssize_t ret;
1398
1399	/* enforce forwards compatibility on users */
1400	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1401		pr_debug("EINVAL: reserve field set\n");
1402		return -EINVAL;
1403	}
1404
1405	/* prevent overflows */
1406	if (unlikely(
1407	    (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1408	    (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1409	    ((ssize_t)iocb->aio_nbytes < 0)
1410	   )) {
1411		pr_debug("EINVAL: io_submit: overflow check\n");
1412		return -EINVAL;
1413	}
1414
1415	req = aio_get_req(ctx);
1416	if (unlikely(!req))
 
 
 
 
 
1417		return -EAGAIN;
1418
1419	req->ki_filp = fget(iocb->aio_fildes);
1420	if (unlikely(!req->ki_filp)) {
1421		ret = -EBADF;
1422		goto out_put_req;
1423	}
1424
1425	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1426		/*
1427		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1428		 * instance of the file* now. The file descriptor must be
1429		 * an eventfd() fd, and will be signaled for each completed
1430		 * event using the eventfd_signal() function.
1431		 */
1432		req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1433		if (IS_ERR(req->ki_eventfd)) {
1434			ret = PTR_ERR(req->ki_eventfd);
1435			req->ki_eventfd = NULL;
1436			goto out_put_req;
1437		}
1438	}
1439
1440	ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1441	if (unlikely(ret)) {
1442		pr_debug("EFAULT: aio_key\n");
1443		goto out_put_req;
1444	}
1445
1446	req->ki_obj.user = user_iocb;
1447	req->ki_user_data = iocb->aio_data;
1448	req->ki_pos = iocb->aio_offset;
1449	req->ki_nbytes = iocb->aio_nbytes;
1450
1451	ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1452			   (char __user *)(unsigned long)iocb->aio_buf,
1453			   compat);
 
 
 
1454	if (ret)
1455		goto out_put_req;
1456
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1457	return 0;
 
1458out_put_req:
1459	put_reqs_available(ctx, 1);
1460	percpu_ref_put(&ctx->reqs);
1461	kiocb_free(req);
1462	return ret;
1463}
1464
1465long do_io_submit(aio_context_t ctx_id, long nr,
1466		  struct iocb __user *__user *iocbpp, bool compat)
1467{
1468	struct kioctx *ctx;
1469	long ret = 0;
1470	int i = 0;
1471	struct blk_plug plug;
1472
1473	if (unlikely(nr < 0))
1474		return -EINVAL;
1475
1476	if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1477		nr = LONG_MAX/sizeof(*iocbpp);
1478
1479	if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1480		return -EFAULT;
1481
1482	ctx = lookup_ioctx(ctx_id);
1483	if (unlikely(!ctx)) {
1484		pr_debug("EINVAL: invalid context id\n");
1485		return -EINVAL;
1486	}
1487
1488	blk_start_plug(&plug);
1489
1490	/*
1491	 * AKPM: should this return a partial result if some of the IOs were
1492	 * successfully submitted?
1493	 */
1494	for (i=0; i<nr; i++) {
1495		struct iocb __user *user_iocb;
1496		struct iocb tmp;
1497
1498		if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1499			ret = -EFAULT;
1500			break;
1501		}
1502
1503		if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1504			ret = -EFAULT;
1505			break;
1506		}
1507
1508		ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1509		if (ret)
1510			break;
1511	}
1512	blk_finish_plug(&plug);
1513
1514	percpu_ref_put(&ctx->users);
1515	return i ? i : ret;
1516}
1517
1518/* sys_io_submit:
1519 *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1520 *	the number of iocbs queued.  May return -EINVAL if the aio_context
1521 *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
1522 *	*iocbpp[0] is not properly initialized, if the operation specified
1523 *	is invalid for the file descriptor in the iocb.  May fail with
1524 *	-EFAULT if any of the data structures point to invalid data.  May
1525 *	fail with -EBADF if the file descriptor specified in the first
1526 *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
1527 *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1528 *	fail with -ENOSYS if not implemented.
1529 */
1530SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1531		struct iocb __user * __user *, iocbpp)
1532{
1533	return do_io_submit(ctx_id, nr, iocbpp, 0);
1534}
1535
1536/* lookup_kiocb
1537 *	Finds a given iocb for cancellation.
1538 */
1539static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1540				  u32 key)
1541{
1542	struct list_head *pos;
1543
1544	assert_spin_locked(&ctx->ctx_lock);
1545
1546	if (key != KIOCB_KEY)
1547		return NULL;
1548
1549	/* TODO: use a hash or array, this sucks. */
1550	list_for_each(pos, &ctx->active_reqs) {
1551		struct kiocb *kiocb = list_kiocb(pos);
1552		if (kiocb->ki_obj.user == iocb)
1553			return kiocb;
1554	}
1555	return NULL;
1556}
1557
1558/* sys_io_cancel:
1559 *	Attempts to cancel an iocb previously passed to io_submit.  If
1560 *	the operation is successfully cancelled, the resulting event is
1561 *	copied into the memory pointed to by result without being placed
1562 *	into the completion queue and 0 is returned.  May fail with
1563 *	-EFAULT if any of the data structures pointed to are invalid.
1564 *	May fail with -EINVAL if aio_context specified by ctx_id is
1565 *	invalid.  May fail with -EAGAIN if the iocb specified was not
1566 *	cancelled.  Will fail with -ENOSYS if not implemented.
1567 */
1568SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1569		struct io_event __user *, result)
1570{
 
1571	struct kioctx *ctx;
1572	struct kiocb *kiocb;
1573	u32 key;
1574	int ret;
1575
1576	ret = get_user(key, &iocb->aio_key);
1577	if (unlikely(ret))
1578		return -EFAULT;
1579
1580	ctx = lookup_ioctx(ctx_id);
1581	if (unlikely(!ctx))
1582		return -EINVAL;
1583
1584	spin_lock_irq(&ctx->ctx_lock);
1585
1586	kiocb = lookup_kiocb(ctx, iocb, key);
1587	if (kiocb)
1588		ret = kiocb_cancel(ctx, kiocb);
1589	else
1590		ret = -EINVAL;
1591
 
1592	spin_unlock_irq(&ctx->ctx_lock);
1593
1594	if (!ret) {
1595		/*
1596		 * The result argument is no longer used - the io_event is
1597		 * always delivered via the ring buffer. -EINPROGRESS indicates
1598		 * cancellation is progress:
1599		 */
1600		ret = -EINPROGRESS;
1601	}
 
 
 
 
 
 
 
 
1602
1603	percpu_ref_put(&ctx->users);
1604
1605	return ret;
1606}
1607
1608/* io_getevents:
1609 *	Attempts to read at least min_nr events and up to nr events from
1610 *	the completion queue for the aio_context specified by ctx_id. If
1611 *	it succeeds, the number of read events is returned. May fail with
1612 *	-EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1613 *	out of range, if timeout is out of range.  May fail with -EFAULT
1614 *	if any of the memory specified is invalid.  May return 0 or
1615 *	< min_nr if the timeout specified by timeout has elapsed
1616 *	before sufficient events are available, where timeout == NULL
1617 *	specifies an infinite timeout. Note that the timeout pointed to by
1618 *	timeout is relative.  Will fail with -ENOSYS if not implemented.
 
1619 */
1620SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1621		long, min_nr,
1622		long, nr,
1623		struct io_event __user *, events,
1624		struct timespec __user *, timeout)
1625{
1626	struct kioctx *ioctx = lookup_ioctx(ctx_id);
1627	long ret = -EINVAL;
1628
1629	if (likely(ioctx)) {
1630		if (likely(min_nr <= nr && min_nr >= 0))
1631			ret = read_events(ioctx, min_nr, nr, events, timeout);
1632		percpu_ref_put(&ioctx->users);
1633	}
 
 
1634	return ret;
1635}