xref: /openbmc/linux/fs/aio.c (revision db446a08c23d5475e6b08c87acca79ebb20f283c)
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/anon_inodes.h>
40 #include <linux/migrate.h>
41 #include <linux/ramfs.h>
42 #include <linux/percpu-refcount.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
52 struct aio_ring {
53 	unsigned	id;	/* kernel internal index number */
54 	unsigned	nr;	/* number of io_events */
55 	unsigned	head;
56 	unsigned	tail;
57 
58 	unsigned	magic;
59 	unsigned	compat_features;
60 	unsigned	incompat_features;
61 	unsigned	header_length;	/* size of aio_ring */
62 
63 
64 	struct io_event		io_events[0];
65 }; /* 128 bytes + ring size */
66 
67 #define AIO_RING_PAGES	8
68 
69 struct kioctx_table {
70 	struct rcu_head	rcu;
71 	unsigned	nr;
72 	struct kioctx	*table[];
73 };
74 
75 struct kioctx_cpu {
76 	unsigned		reqs_available;
77 };
78 
79 struct kioctx {
80 	struct percpu_ref	users;
81 	atomic_t		dead;
82 
83 	unsigned long		user_id;
84 
85 	struct __percpu kioctx_cpu *cpu;
86 
87 	/*
88 	 * For percpu reqs_available, number of slots we move to/from global
89 	 * counter at a time:
90 	 */
91 	unsigned		req_batch;
92 	/*
93 	 * This is what userspace passed to io_setup(), it's not used for
94 	 * anything but counting against the global max_reqs quota.
95 	 *
96 	 * The real limit is nr_events - 1, which will be larger (see
97 	 * aio_setup_ring())
98 	 */
99 	unsigned		max_reqs;
100 
101 	/* Size of ringbuffer, in units of struct io_event */
102 	unsigned		nr_events;
103 
104 	unsigned long		mmap_base;
105 	unsigned long		mmap_size;
106 
107 	struct page		**ring_pages;
108 	long			nr_pages;
109 
110 	struct rcu_head		rcu_head;
111 	struct work_struct	free_work;
112 
113 	struct {
114 		/*
115 		 * This counts the number of available slots in the ringbuffer,
116 		 * so we avoid overflowing it: it's decremented (if positive)
117 		 * when allocating a kiocb and incremented when the resulting
118 		 * io_event is pulled off the ringbuffer.
119 		 *
120 		 * We batch accesses to it with a percpu version.
121 		 */
122 		atomic_t	reqs_available;
123 	} ____cacheline_aligned_in_smp;
124 
125 	struct {
126 		spinlock_t	ctx_lock;
127 		struct list_head active_reqs;	/* used for cancellation */
128 	} ____cacheline_aligned_in_smp;
129 
130 	struct {
131 		struct mutex	ring_lock;
132 		wait_queue_head_t wait;
133 	} ____cacheline_aligned_in_smp;
134 
135 	struct {
136 		unsigned	tail;
137 		spinlock_t	completion_lock;
138 	} ____cacheline_aligned_in_smp;
139 
140 	struct page		*internal_pages[AIO_RING_PAGES];
141 	struct file		*aio_ring_file;
142 
143 	unsigned		id;
144 };
145 
146 /*------ sysctl variables----*/
147 static DEFINE_SPINLOCK(aio_nr_lock);
148 unsigned long aio_nr;		/* current system wide number of aio requests */
149 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
150 /*----end sysctl variables---*/
151 
152 static struct kmem_cache	*kiocb_cachep;
153 static struct kmem_cache	*kioctx_cachep;
154 
155 /* aio_setup
156  *	Creates the slab caches used by the aio routines, panic on
157  *	failure as this is done early during the boot sequence.
158  */
159 static int __init aio_setup(void)
160 {
161 	kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
162 	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
163 
164 	pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
165 
166 	return 0;
167 }
168 __initcall(aio_setup);
169 
170 static void aio_free_ring(struct kioctx *ctx)
171 {
172 	int i;
173 	struct file *aio_ring_file = ctx->aio_ring_file;
174 
175 	for (i = 0; i < ctx->nr_pages; i++) {
176 		pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
177 				page_count(ctx->ring_pages[i]));
178 		put_page(ctx->ring_pages[i]);
179 	}
180 
181 	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages)
182 		kfree(ctx->ring_pages);
183 
184 	if (aio_ring_file) {
185 		truncate_setsize(aio_ring_file->f_inode, 0);
186 		pr_debug("pid(%d) i_nlink=%u d_count=%d d_unhashed=%d i_count=%d\n",
187 			current->pid, aio_ring_file->f_inode->i_nlink,
188 			aio_ring_file->f_path.dentry->d_count,
189 			d_unhashed(aio_ring_file->f_path.dentry),
190 			atomic_read(&aio_ring_file->f_inode->i_count));
191 		fput(aio_ring_file);
192 		ctx->aio_ring_file = NULL;
193 	}
194 }
195 
196 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
197 {
198 	vma->vm_ops = &generic_file_vm_ops;
199 	return 0;
200 }
201 
202 static const struct file_operations aio_ring_fops = {
203 	.mmap = aio_ring_mmap,
204 };
205 
206 static int aio_set_page_dirty(struct page *page)
207 {
208 	return 0;
209 }
210 
211 #if IS_ENABLED(CONFIG_MIGRATION)
212 static int aio_migratepage(struct address_space *mapping, struct page *new,
213 			struct page *old, enum migrate_mode mode)
214 {
215 	struct kioctx *ctx = mapping->private_data;
216 	unsigned long flags;
217 	unsigned idx = old->index;
218 	int rc;
219 
220 	/* Writeback must be complete */
221 	BUG_ON(PageWriteback(old));
222 	put_page(old);
223 
224 	rc = migrate_page_move_mapping(mapping, new, old, NULL, mode);
225 	if (rc != MIGRATEPAGE_SUCCESS) {
226 		get_page(old);
227 		return rc;
228 	}
229 
230 	get_page(new);
231 
232 	spin_lock_irqsave(&ctx->completion_lock, flags);
233 	migrate_page_copy(new, old);
234 	ctx->ring_pages[idx] = new;
235 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
236 
237 	return rc;
238 }
239 #endif
240 
241 static const struct address_space_operations aio_ctx_aops = {
242 	.set_page_dirty = aio_set_page_dirty,
243 #if IS_ENABLED(CONFIG_MIGRATION)
244 	.migratepage	= aio_migratepage,
245 #endif
246 };
247 
248 static int aio_setup_ring(struct kioctx *ctx)
249 {
250 	struct aio_ring *ring;
251 	unsigned nr_events = ctx->max_reqs;
252 	struct mm_struct *mm = current->mm;
253 	unsigned long size, populate;
254 	int nr_pages;
255 	int i;
256 	struct file *file;
257 
258 	/* Compensate for the ring buffer's head/tail overlap entry */
259 	nr_events += 2;	/* 1 is required, 2 for good luck */
260 
261 	size = sizeof(struct aio_ring);
262 	size += sizeof(struct io_event) * nr_events;
263 
264 	nr_pages = PFN_UP(size);
265 	if (nr_pages < 0)
266 		return -EINVAL;
267 
268 	file = anon_inode_getfile_private("[aio]", &aio_ring_fops, ctx, O_RDWR);
269 	if (IS_ERR(file)) {
270 		ctx->aio_ring_file = NULL;
271 		return -EAGAIN;
272 	}
273 
274 	file->f_inode->i_mapping->a_ops = &aio_ctx_aops;
275 	file->f_inode->i_mapping->private_data = ctx;
276 	file->f_inode->i_size = PAGE_SIZE * (loff_t)nr_pages;
277 
278 	for (i = 0; i < nr_pages; i++) {
279 		struct page *page;
280 		page = find_or_create_page(file->f_inode->i_mapping,
281 					   i, GFP_HIGHUSER | __GFP_ZERO);
282 		if (!page)
283 			break;
284 		pr_debug("pid(%d) page[%d]->count=%d\n",
285 			 current->pid, i, page_count(page));
286 		SetPageUptodate(page);
287 		SetPageDirty(page);
288 		unlock_page(page);
289 	}
290 	ctx->aio_ring_file = file;
291 	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
292 			/ sizeof(struct io_event);
293 
294 	ctx->ring_pages = ctx->internal_pages;
295 	if (nr_pages > AIO_RING_PAGES) {
296 		ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
297 					  GFP_KERNEL);
298 		if (!ctx->ring_pages)
299 			return -ENOMEM;
300 	}
301 
302 	ctx->mmap_size = nr_pages * PAGE_SIZE;
303 	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
304 
305 	down_write(&mm->mmap_sem);
306 	ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
307 				       PROT_READ | PROT_WRITE,
308 				       MAP_SHARED | MAP_POPULATE, 0, &populate);
309 	if (IS_ERR((void *)ctx->mmap_base)) {
310 		up_write(&mm->mmap_sem);
311 		ctx->mmap_size = 0;
312 		aio_free_ring(ctx);
313 		return -EAGAIN;
314 	}
315 	up_write(&mm->mmap_sem);
316 
317 	mm_populate(ctx->mmap_base, populate);
318 
319 	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
320 	ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
321 				       1, 0, ctx->ring_pages, NULL);
322 	for (i = 0; i < ctx->nr_pages; i++)
323 		put_page(ctx->ring_pages[i]);
324 
325 	if (unlikely(ctx->nr_pages != nr_pages)) {
326 		aio_free_ring(ctx);
327 		return -EAGAIN;
328 	}
329 
330 	ctx->user_id = ctx->mmap_base;
331 	ctx->nr_events = nr_events; /* trusted copy */
332 
333 	ring = kmap_atomic(ctx->ring_pages[0]);
334 	ring->nr = nr_events;	/* user copy */
335 	ring->id = ~0U;
336 	ring->head = ring->tail = 0;
337 	ring->magic = AIO_RING_MAGIC;
338 	ring->compat_features = AIO_RING_COMPAT_FEATURES;
339 	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
340 	ring->header_length = sizeof(struct aio_ring);
341 	kunmap_atomic(ring);
342 	flush_dcache_page(ctx->ring_pages[0]);
343 
344 	return 0;
345 }
346 
347 #define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
348 #define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
349 #define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
350 
351 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
352 {
353 	struct kioctx *ctx = req->ki_ctx;
354 	unsigned long flags;
355 
356 	spin_lock_irqsave(&ctx->ctx_lock, flags);
357 
358 	if (!req->ki_list.next)
359 		list_add(&req->ki_list, &ctx->active_reqs);
360 
361 	req->ki_cancel = cancel;
362 
363 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
364 }
365 EXPORT_SYMBOL(kiocb_set_cancel_fn);
366 
367 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
368 {
369 	kiocb_cancel_fn *old, *cancel;
370 
371 	/*
372 	 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
373 	 * actually has a cancel function, hence the cmpxchg()
374 	 */
375 
376 	cancel = ACCESS_ONCE(kiocb->ki_cancel);
377 	do {
378 		if (!cancel || cancel == KIOCB_CANCELLED)
379 			return -EINVAL;
380 
381 		old = cancel;
382 		cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
383 	} while (cancel != old);
384 
385 	return cancel(kiocb);
386 }
387 
388 static void free_ioctx_rcu(struct rcu_head *head)
389 {
390 	struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
391 
392 	free_percpu(ctx->cpu);
393 	kmem_cache_free(kioctx_cachep, ctx);
394 }
395 
396 /*
397  * When this function runs, the kioctx has been removed from the "hash table"
398  * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
399  * now it's safe to cancel any that need to be.
400  */
401 static void free_ioctx(struct work_struct *work)
402 {
403 	struct kioctx *ctx = container_of(work, struct kioctx, free_work);
404 	struct aio_ring *ring;
405 	struct kiocb *req;
406 	unsigned cpu, avail;
407 	DEFINE_WAIT(wait);
408 
409 	spin_lock_irq(&ctx->ctx_lock);
410 
411 	while (!list_empty(&ctx->active_reqs)) {
412 		req = list_first_entry(&ctx->active_reqs,
413 				       struct kiocb, ki_list);
414 
415 		list_del_init(&req->ki_list);
416 		kiocb_cancel(ctx, req);
417 	}
418 
419 	spin_unlock_irq(&ctx->ctx_lock);
420 
421 	for_each_possible_cpu(cpu) {
422 		struct kioctx_cpu *kcpu = per_cpu_ptr(ctx->cpu, cpu);
423 
424 		atomic_add(kcpu->reqs_available, &ctx->reqs_available);
425 		kcpu->reqs_available = 0;
426 	}
427 
428 	while (1) {
429 		prepare_to_wait(&ctx->wait, &wait, TASK_UNINTERRUPTIBLE);
430 
431 		ring = kmap_atomic(ctx->ring_pages[0]);
432 		avail = (ring->head <= ring->tail)
433 			 ? ring->tail - ring->head
434 			 : ctx->nr_events - ring->head + ring->tail;
435 
436 		atomic_add(avail, &ctx->reqs_available);
437 		ring->head = ring->tail;
438 		kunmap_atomic(ring);
439 
440 		if (atomic_read(&ctx->reqs_available) >= ctx->nr_events - 1)
441 			break;
442 
443 		schedule();
444 	}
445 	finish_wait(&ctx->wait, &wait);
446 
447 	WARN_ON(atomic_read(&ctx->reqs_available) > ctx->nr_events - 1);
448 
449 	aio_free_ring(ctx);
450 
451 	pr_debug("freeing %p\n", ctx);
452 
453 	/*
454 	 * Here the call_rcu() is between the wait_event() for reqs_active to
455 	 * hit 0, and freeing the ioctx.
456 	 *
457 	 * aio_complete() decrements reqs_active, but it has to touch the ioctx
458 	 * after to issue a wakeup so we use rcu.
459 	 */
460 	call_rcu(&ctx->rcu_head, free_ioctx_rcu);
461 }
462 
463 static void free_ioctx_ref(struct percpu_ref *ref)
464 {
465 	struct kioctx *ctx = container_of(ref, struct kioctx, users);
466 
467 	INIT_WORK(&ctx->free_work, free_ioctx);
468 	schedule_work(&ctx->free_work);
469 }
470 
471 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
472 {
473 	unsigned i, new_nr;
474 	struct kioctx_table *table, *old;
475 	struct aio_ring *ring;
476 
477 	spin_lock(&mm->ioctx_lock);
478 	table = rcu_dereference(mm->ioctx_table);
479 
480 	while (1) {
481 		if (table)
482 			for (i = 0; i < table->nr; i++)
483 				if (!table->table[i]) {
484 					ctx->id = i;
485 					table->table[i] = ctx;
486 					spin_unlock(&mm->ioctx_lock);
487 
488 					ring = kmap_atomic(ctx->ring_pages[0]);
489 					ring->id = ctx->id;
490 					kunmap_atomic(ring);
491 					return 0;
492 				}
493 
494 		new_nr = (table ? table->nr : 1) * 4;
495 
496 		spin_unlock(&mm->ioctx_lock);
497 
498 		table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
499 				new_nr, GFP_KERNEL);
500 		if (!table)
501 			return -ENOMEM;
502 
503 		table->nr = new_nr;
504 
505 		spin_lock(&mm->ioctx_lock);
506 		old = rcu_dereference(mm->ioctx_table);
507 
508 		if (!old) {
509 			rcu_assign_pointer(mm->ioctx_table, table);
510 		} else if (table->nr > old->nr) {
511 			memcpy(table->table, old->table,
512 			       old->nr * sizeof(struct kioctx *));
513 
514 			rcu_assign_pointer(mm->ioctx_table, table);
515 			kfree_rcu(old, rcu);
516 		} else {
517 			kfree(table);
518 			table = old;
519 		}
520 	}
521 }
522 
523 /* ioctx_alloc
524  *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
525  */
526 static struct kioctx *ioctx_alloc(unsigned nr_events)
527 {
528 	struct mm_struct *mm = current->mm;
529 	struct kioctx *ctx;
530 	int err = -ENOMEM;
531 
532 	/*
533 	 * We keep track of the number of available ringbuffer slots, to prevent
534 	 * overflow (reqs_available), and we also use percpu counters for this.
535 	 *
536 	 * So since up to half the slots might be on other cpu's percpu counters
537 	 * and unavailable, double nr_events so userspace sees what they
538 	 * expected: additionally, we move req_batch slots to/from percpu
539 	 * counters at a time, so make sure that isn't 0:
540 	 */
541 	nr_events = max(nr_events, num_possible_cpus() * 4);
542 	nr_events *= 2;
543 
544 	/* Prevent overflows */
545 	if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
546 	    (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
547 		pr_debug("ENOMEM: nr_events too high\n");
548 		return ERR_PTR(-EINVAL);
549 	}
550 
551 	if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
552 		return ERR_PTR(-EAGAIN);
553 
554 	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
555 	if (!ctx)
556 		return ERR_PTR(-ENOMEM);
557 
558 	ctx->max_reqs = nr_events;
559 
560 	if (percpu_ref_init(&ctx->users, free_ioctx_ref))
561 		goto out_freectx;
562 
563 	spin_lock_init(&ctx->ctx_lock);
564 	spin_lock_init(&ctx->completion_lock);
565 	mutex_init(&ctx->ring_lock);
566 	init_waitqueue_head(&ctx->wait);
567 
568 	INIT_LIST_HEAD(&ctx->active_reqs);
569 
570 	ctx->cpu = alloc_percpu(struct kioctx_cpu);
571 	if (!ctx->cpu)
572 		goto out_freeref;
573 
574 	if (aio_setup_ring(ctx) < 0)
575 		goto out_freepcpu;
576 
577 	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
578 	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
579 	BUG_ON(!ctx->req_batch);
580 
581 	err = ioctx_add_table(ctx, mm);
582 	if (err)
583 		goto out_cleanup_noerr;
584 
585 	/* limit the number of system wide aios */
586 	spin_lock(&aio_nr_lock);
587 	if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
588 	    aio_nr + nr_events < aio_nr) {
589 		spin_unlock(&aio_nr_lock);
590 		goto out_cleanup;
591 	}
592 	aio_nr += ctx->max_reqs;
593 	spin_unlock(&aio_nr_lock);
594 
595 	percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
596 
597 	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
598 		 ctx, ctx->user_id, mm, ctx->nr_events);
599 	return ctx;
600 
601 out_cleanup:
602 	err = -EAGAIN;
603 out_cleanup_noerr:
604 	aio_free_ring(ctx);
605 out_freepcpu:
606 	free_percpu(ctx->cpu);
607 out_freeref:
608 	free_percpu(ctx->users.pcpu_count);
609 out_freectx:
610 	if (ctx->aio_ring_file)
611 		fput(ctx->aio_ring_file);
612 	kmem_cache_free(kioctx_cachep, ctx);
613 	pr_debug("error allocating ioctx %d\n", err);
614 	return ERR_PTR(err);
615 }
616 
617 /* kill_ioctx
618  *	Cancels all outstanding aio requests on an aio context.  Used
619  *	when the processes owning a context have all exited to encourage
620  *	the rapid destruction of the kioctx.
621  */
622 static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx)
623 {
624 	if (!atomic_xchg(&ctx->dead, 1)) {
625 		struct kioctx_table *table;
626 
627 		spin_lock(&mm->ioctx_lock);
628 		table = rcu_dereference(mm->ioctx_table);
629 
630 		WARN_ON(ctx != table->table[ctx->id]);
631 		table->table[ctx->id] = NULL;
632 		spin_unlock(&mm->ioctx_lock);
633 
634 		/* percpu_ref_kill() will do the necessary call_rcu() */
635 		wake_up_all(&ctx->wait);
636 
637 		/*
638 		 * It'd be more correct to do this in free_ioctx(), after all
639 		 * the outstanding kiocbs have finished - but by then io_destroy
640 		 * has already returned, so io_setup() could potentially return
641 		 * -EAGAIN with no ioctxs actually in use (as far as userspace
642 		 *  could tell).
643 		 */
644 		spin_lock(&aio_nr_lock);
645 		BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
646 		aio_nr -= ctx->max_reqs;
647 		spin_unlock(&aio_nr_lock);
648 
649 		if (ctx->mmap_size)
650 			vm_munmap(ctx->mmap_base, ctx->mmap_size);
651 
652 		percpu_ref_kill(&ctx->users);
653 	}
654 }
655 
656 /* wait_on_sync_kiocb:
657  *	Waits on the given sync kiocb to complete.
658  */
659 ssize_t wait_on_sync_kiocb(struct kiocb *req)
660 {
661 	while (!req->ki_ctx) {
662 		set_current_state(TASK_UNINTERRUPTIBLE);
663 		if (req->ki_ctx)
664 			break;
665 		io_schedule();
666 	}
667 	__set_current_state(TASK_RUNNING);
668 	return req->ki_user_data;
669 }
670 EXPORT_SYMBOL(wait_on_sync_kiocb);
671 
672 /*
673  * exit_aio: called when the last user of mm goes away.  At this point, there is
674  * no way for any new requests to be submited or any of the io_* syscalls to be
675  * called on the context.
676  *
677  * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
678  * them.
679  */
680 void exit_aio(struct mm_struct *mm)
681 {
682 	struct kioctx_table *table;
683 	struct kioctx *ctx;
684 	unsigned i = 0;
685 
686 	while (1) {
687 		rcu_read_lock();
688 		table = rcu_dereference(mm->ioctx_table);
689 
690 		do {
691 			if (!table || i >= table->nr) {
692 				rcu_read_unlock();
693 				rcu_assign_pointer(mm->ioctx_table, NULL);
694 				if (table)
695 					kfree(table);
696 				return;
697 			}
698 
699 			ctx = table->table[i++];
700 		} while (!ctx);
701 
702 		rcu_read_unlock();
703 
704 		/*
705 		 * We don't need to bother with munmap() here -
706 		 * exit_mmap(mm) is coming and it'll unmap everything.
707 		 * Since aio_free_ring() uses non-zero ->mmap_size
708 		 * as indicator that it needs to unmap the area,
709 		 * just set it to 0; aio_free_ring() is the only
710 		 * place that uses ->mmap_size, so it's safe.
711 		 */
712 		ctx->mmap_size = 0;
713 
714 		kill_ioctx(mm, ctx);
715 	}
716 }
717 
718 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
719 {
720 	struct kioctx_cpu *kcpu;
721 
722 	preempt_disable();
723 	kcpu = this_cpu_ptr(ctx->cpu);
724 
725 	kcpu->reqs_available += nr;
726 	while (kcpu->reqs_available >= ctx->req_batch * 2) {
727 		kcpu->reqs_available -= ctx->req_batch;
728 		atomic_add(ctx->req_batch, &ctx->reqs_available);
729 	}
730 
731 	preempt_enable();
732 }
733 
734 static bool get_reqs_available(struct kioctx *ctx)
735 {
736 	struct kioctx_cpu *kcpu;
737 	bool ret = false;
738 
739 	preempt_disable();
740 	kcpu = this_cpu_ptr(ctx->cpu);
741 
742 	if (!kcpu->reqs_available) {
743 		int old, avail = atomic_read(&ctx->reqs_available);
744 
745 		do {
746 			if (avail < ctx->req_batch)
747 				goto out;
748 
749 			old = avail;
750 			avail = atomic_cmpxchg(&ctx->reqs_available,
751 					       avail, avail - ctx->req_batch);
752 		} while (avail != old);
753 
754 		kcpu->reqs_available += ctx->req_batch;
755 	}
756 
757 	ret = true;
758 	kcpu->reqs_available--;
759 out:
760 	preempt_enable();
761 	return ret;
762 }
763 
764 /* aio_get_req
765  *	Allocate a slot for an aio request.
766  * Returns NULL if no requests are free.
767  */
768 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
769 {
770 	struct kiocb *req;
771 
772 	if (!get_reqs_available(ctx))
773 		return NULL;
774 
775 	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
776 	if (unlikely(!req))
777 		goto out_put;
778 
779 	req->ki_ctx = ctx;
780 	return req;
781 out_put:
782 	put_reqs_available(ctx, 1);
783 	return NULL;
784 }
785 
786 static void kiocb_free(struct kiocb *req)
787 {
788 	if (req->ki_filp)
789 		fput(req->ki_filp);
790 	if (req->ki_eventfd != NULL)
791 		eventfd_ctx_put(req->ki_eventfd);
792 	kmem_cache_free(kiocb_cachep, req);
793 }
794 
795 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
796 {
797 	struct aio_ring __user *ring  = (void __user *)ctx_id;
798 	struct mm_struct *mm = current->mm;
799 	struct kioctx *ctx, *ret = NULL;
800 	struct kioctx_table *table;
801 	unsigned id;
802 
803 	if (get_user(id, &ring->id))
804 		return NULL;
805 
806 	rcu_read_lock();
807 	table = rcu_dereference(mm->ioctx_table);
808 
809 	if (!table || id >= table->nr)
810 		goto out;
811 
812 	ctx = table->table[id];
813 	if (ctx->user_id == ctx_id) {
814 		percpu_ref_get(&ctx->users);
815 		ret = ctx;
816 	}
817 out:
818 	rcu_read_unlock();
819 	return ret;
820 }
821 
822 /* aio_complete
823  *	Called when the io request on the given iocb is complete.
824  */
825 void aio_complete(struct kiocb *iocb, long res, long res2)
826 {
827 	struct kioctx	*ctx = iocb->ki_ctx;
828 	struct aio_ring	*ring;
829 	struct io_event	*ev_page, *event;
830 	unsigned long	flags;
831 	unsigned tail, pos;
832 
833 	/*
834 	 * Special case handling for sync iocbs:
835 	 *  - events go directly into the iocb for fast handling
836 	 *  - the sync task with the iocb in its stack holds the single iocb
837 	 *    ref, no other paths have a way to get another ref
838 	 *  - the sync task helpfully left a reference to itself in the iocb
839 	 */
840 	if (is_sync_kiocb(iocb)) {
841 		iocb->ki_user_data = res;
842 		smp_wmb();
843 		iocb->ki_ctx = ERR_PTR(-EXDEV);
844 		wake_up_process(iocb->ki_obj.tsk);
845 		return;
846 	}
847 
848 	/*
849 	 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
850 	 * need to issue a wakeup after incrementing reqs_available.
851 	 */
852 	rcu_read_lock();
853 
854 	if (iocb->ki_list.next) {
855 		unsigned long flags;
856 
857 		spin_lock_irqsave(&ctx->ctx_lock, flags);
858 		list_del(&iocb->ki_list);
859 		spin_unlock_irqrestore(&ctx->ctx_lock, flags);
860 	}
861 
862 	/*
863 	 * Add a completion event to the ring buffer. Must be done holding
864 	 * ctx->completion_lock to prevent other code from messing with the tail
865 	 * pointer since we might be called from irq context.
866 	 */
867 	spin_lock_irqsave(&ctx->completion_lock, flags);
868 
869 	tail = ctx->tail;
870 	pos = tail + AIO_EVENTS_OFFSET;
871 
872 	if (++tail >= ctx->nr_events)
873 		tail = 0;
874 
875 	ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
876 	event = ev_page + pos % AIO_EVENTS_PER_PAGE;
877 
878 	event->obj = (u64)(unsigned long)iocb->ki_obj.user;
879 	event->data = iocb->ki_user_data;
880 	event->res = res;
881 	event->res2 = res2;
882 
883 	kunmap_atomic(ev_page);
884 	flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
885 
886 	pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
887 		 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
888 		 res, res2);
889 
890 	/* after flagging the request as done, we
891 	 * must never even look at it again
892 	 */
893 	smp_wmb();	/* make event visible before updating tail */
894 
895 	ctx->tail = tail;
896 
897 	ring = kmap_atomic(ctx->ring_pages[0]);
898 	ring->tail = tail;
899 	kunmap_atomic(ring);
900 	flush_dcache_page(ctx->ring_pages[0]);
901 
902 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
903 
904 	pr_debug("added to ring %p at [%u]\n", iocb, tail);
905 
906 	/*
907 	 * Check if the user asked us to deliver the result through an
908 	 * eventfd. The eventfd_signal() function is safe to be called
909 	 * from IRQ context.
910 	 */
911 	if (iocb->ki_eventfd != NULL)
912 		eventfd_signal(iocb->ki_eventfd, 1);
913 
914 	/* everything turned out well, dispose of the aiocb. */
915 	kiocb_free(iocb);
916 
917 	/*
918 	 * We have to order our ring_info tail store above and test
919 	 * of the wait list below outside the wait lock.  This is
920 	 * like in wake_up_bit() where clearing a bit has to be
921 	 * ordered with the unlocked test.
922 	 */
923 	smp_mb();
924 
925 	if (waitqueue_active(&ctx->wait))
926 		wake_up(&ctx->wait);
927 
928 	rcu_read_unlock();
929 }
930 EXPORT_SYMBOL(aio_complete);
931 
932 /* aio_read_events
933  *	Pull an event off of the ioctx's event ring.  Returns the number of
934  *	events fetched
935  */
936 static long aio_read_events_ring(struct kioctx *ctx,
937 				 struct io_event __user *event, long nr)
938 {
939 	struct aio_ring *ring;
940 	unsigned head, tail, pos;
941 	long ret = 0;
942 	int copy_ret;
943 
944 	mutex_lock(&ctx->ring_lock);
945 
946 	ring = kmap_atomic(ctx->ring_pages[0]);
947 	head = ring->head;
948 	tail = ring->tail;
949 	kunmap_atomic(ring);
950 
951 	pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
952 
953 	if (head == tail)
954 		goto out;
955 
956 	while (ret < nr) {
957 		long avail;
958 		struct io_event *ev;
959 		struct page *page;
960 
961 		avail = (head <= tail ?  tail : ctx->nr_events) - head;
962 		if (head == tail)
963 			break;
964 
965 		avail = min(avail, nr - ret);
966 		avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
967 			    ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
968 
969 		pos = head + AIO_EVENTS_OFFSET;
970 		page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
971 		pos %= AIO_EVENTS_PER_PAGE;
972 
973 		ev = kmap(page);
974 		copy_ret = copy_to_user(event + ret, ev + pos,
975 					sizeof(*ev) * avail);
976 		kunmap(page);
977 
978 		if (unlikely(copy_ret)) {
979 			ret = -EFAULT;
980 			goto out;
981 		}
982 
983 		ret += avail;
984 		head += avail;
985 		head %= ctx->nr_events;
986 	}
987 
988 	ring = kmap_atomic(ctx->ring_pages[0]);
989 	ring->head = head;
990 	kunmap_atomic(ring);
991 	flush_dcache_page(ctx->ring_pages[0]);
992 
993 	pr_debug("%li  h%u t%u\n", ret, head, tail);
994 
995 	put_reqs_available(ctx, ret);
996 out:
997 	mutex_unlock(&ctx->ring_lock);
998 
999 	return ret;
1000 }
1001 
1002 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1003 			    struct io_event __user *event, long *i)
1004 {
1005 	long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1006 
1007 	if (ret > 0)
1008 		*i += ret;
1009 
1010 	if (unlikely(atomic_read(&ctx->dead)))
1011 		ret = -EINVAL;
1012 
1013 	if (!*i)
1014 		*i = ret;
1015 
1016 	return ret < 0 || *i >= min_nr;
1017 }
1018 
1019 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1020 			struct io_event __user *event,
1021 			struct timespec __user *timeout)
1022 {
1023 	ktime_t until = { .tv64 = KTIME_MAX };
1024 	long ret = 0;
1025 
1026 	if (timeout) {
1027 		struct timespec	ts;
1028 
1029 		if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1030 			return -EFAULT;
1031 
1032 		until = timespec_to_ktime(ts);
1033 	}
1034 
1035 	/*
1036 	 * Note that aio_read_events() is being called as the conditional - i.e.
1037 	 * we're calling it after prepare_to_wait() has set task state to
1038 	 * TASK_INTERRUPTIBLE.
1039 	 *
1040 	 * But aio_read_events() can block, and if it blocks it's going to flip
1041 	 * the task state back to TASK_RUNNING.
1042 	 *
1043 	 * This should be ok, provided it doesn't flip the state back to
1044 	 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1045 	 * will only happen if the mutex_lock() call blocks, and we then find
1046 	 * the ringbuffer empty. So in practice we should be ok, but it's
1047 	 * something to be aware of when touching this code.
1048 	 */
1049 	wait_event_interruptible_hrtimeout(ctx->wait,
1050 			aio_read_events(ctx, min_nr, nr, event, &ret), until);
1051 
1052 	if (!ret && signal_pending(current))
1053 		ret = -EINTR;
1054 
1055 	return ret;
1056 }
1057 
1058 /* sys_io_setup:
1059  *	Create an aio_context capable of receiving at least nr_events.
1060  *	ctxp must not point to an aio_context that already exists, and
1061  *	must be initialized to 0 prior to the call.  On successful
1062  *	creation of the aio_context, *ctxp is filled in with the resulting
1063  *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1064  *	if the specified nr_events exceeds internal limits.  May fail
1065  *	with -EAGAIN if the specified nr_events exceeds the user's limit
1066  *	of available events.  May fail with -ENOMEM if insufficient kernel
1067  *	resources are available.  May fail with -EFAULT if an invalid
1068  *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1069  *	implemented.
1070  */
1071 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1072 {
1073 	struct kioctx *ioctx = NULL;
1074 	unsigned long ctx;
1075 	long ret;
1076 
1077 	ret = get_user(ctx, ctxp);
1078 	if (unlikely(ret))
1079 		goto out;
1080 
1081 	ret = -EINVAL;
1082 	if (unlikely(ctx || nr_events == 0)) {
1083 		pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1084 		         ctx, nr_events);
1085 		goto out;
1086 	}
1087 
1088 	ioctx = ioctx_alloc(nr_events);
1089 	ret = PTR_ERR(ioctx);
1090 	if (!IS_ERR(ioctx)) {
1091 		ret = put_user(ioctx->user_id, ctxp);
1092 		if (ret)
1093 			kill_ioctx(current->mm, ioctx);
1094 		percpu_ref_put(&ioctx->users);
1095 	}
1096 
1097 out:
1098 	return ret;
1099 }
1100 
1101 /* sys_io_destroy:
1102  *	Destroy the aio_context specified.  May cancel any outstanding
1103  *	AIOs and block on completion.  Will fail with -ENOSYS if not
1104  *	implemented.  May fail with -EINVAL if the context pointed to
1105  *	is invalid.
1106  */
1107 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1108 {
1109 	struct kioctx *ioctx = lookup_ioctx(ctx);
1110 	if (likely(NULL != ioctx)) {
1111 		kill_ioctx(current->mm, ioctx);
1112 		percpu_ref_put(&ioctx->users);
1113 		return 0;
1114 	}
1115 	pr_debug("EINVAL: io_destroy: invalid context id\n");
1116 	return -EINVAL;
1117 }
1118 
1119 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1120 			    unsigned long, loff_t);
1121 
1122 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1123 				     int rw, char __user *buf,
1124 				     unsigned long *nr_segs,
1125 				     struct iovec **iovec,
1126 				     bool compat)
1127 {
1128 	ssize_t ret;
1129 
1130 	*nr_segs = kiocb->ki_nbytes;
1131 
1132 #ifdef CONFIG_COMPAT
1133 	if (compat)
1134 		ret = compat_rw_copy_check_uvector(rw,
1135 				(struct compat_iovec __user *)buf,
1136 				*nr_segs, 1, *iovec, iovec);
1137 	else
1138 #endif
1139 		ret = rw_copy_check_uvector(rw,
1140 				(struct iovec __user *)buf,
1141 				*nr_segs, 1, *iovec, iovec);
1142 	if (ret < 0)
1143 		return ret;
1144 
1145 	/* ki_nbytes now reflect bytes instead of segs */
1146 	kiocb->ki_nbytes = ret;
1147 	return 0;
1148 }
1149 
1150 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1151 				       int rw, char __user *buf,
1152 				       unsigned long *nr_segs,
1153 				       struct iovec *iovec)
1154 {
1155 	if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1156 		return -EFAULT;
1157 
1158 	iovec->iov_base = buf;
1159 	iovec->iov_len = kiocb->ki_nbytes;
1160 	*nr_segs = 1;
1161 	return 0;
1162 }
1163 
1164 /*
1165  * aio_setup_iocb:
1166  *	Performs the initial checks and aio retry method
1167  *	setup for the kiocb at the time of io submission.
1168  */
1169 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1170 			    char __user *buf, bool compat)
1171 {
1172 	struct file *file = req->ki_filp;
1173 	ssize_t ret;
1174 	unsigned long nr_segs;
1175 	int rw;
1176 	fmode_t mode;
1177 	aio_rw_op *rw_op;
1178 	struct iovec inline_vec, *iovec = &inline_vec;
1179 
1180 	switch (opcode) {
1181 	case IOCB_CMD_PREAD:
1182 	case IOCB_CMD_PREADV:
1183 		mode	= FMODE_READ;
1184 		rw	= READ;
1185 		rw_op	= file->f_op->aio_read;
1186 		goto rw_common;
1187 
1188 	case IOCB_CMD_PWRITE:
1189 	case IOCB_CMD_PWRITEV:
1190 		mode	= FMODE_WRITE;
1191 		rw	= WRITE;
1192 		rw_op	= file->f_op->aio_write;
1193 		goto rw_common;
1194 rw_common:
1195 		if (unlikely(!(file->f_mode & mode)))
1196 			return -EBADF;
1197 
1198 		if (!rw_op)
1199 			return -EINVAL;
1200 
1201 		ret = (opcode == IOCB_CMD_PREADV ||
1202 		       opcode == IOCB_CMD_PWRITEV)
1203 			? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1204 						&iovec, compat)
1205 			: aio_setup_single_vector(req, rw, buf, &nr_segs,
1206 						  iovec);
1207 		if (ret)
1208 			return ret;
1209 
1210 		ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1211 		if (ret < 0) {
1212 			if (iovec != &inline_vec)
1213 				kfree(iovec);
1214 			return ret;
1215 		}
1216 
1217 		req->ki_nbytes = ret;
1218 
1219 		/* XXX: move/kill - rw_verify_area()? */
1220 		/* This matches the pread()/pwrite() logic */
1221 		if (req->ki_pos < 0) {
1222 			ret = -EINVAL;
1223 			break;
1224 		}
1225 
1226 		if (rw == WRITE)
1227 			file_start_write(file);
1228 
1229 		ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1230 
1231 		if (rw == WRITE)
1232 			file_end_write(file);
1233 		break;
1234 
1235 	case IOCB_CMD_FDSYNC:
1236 		if (!file->f_op->aio_fsync)
1237 			return -EINVAL;
1238 
1239 		ret = file->f_op->aio_fsync(req, 1);
1240 		break;
1241 
1242 	case IOCB_CMD_FSYNC:
1243 		if (!file->f_op->aio_fsync)
1244 			return -EINVAL;
1245 
1246 		ret = file->f_op->aio_fsync(req, 0);
1247 		break;
1248 
1249 	default:
1250 		pr_debug("EINVAL: no operation provided\n");
1251 		return -EINVAL;
1252 	}
1253 
1254 	if (iovec != &inline_vec)
1255 		kfree(iovec);
1256 
1257 	if (ret != -EIOCBQUEUED) {
1258 		/*
1259 		 * There's no easy way to restart the syscall since other AIO's
1260 		 * may be already running. Just fail this IO with EINTR.
1261 		 */
1262 		if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1263 			     ret == -ERESTARTNOHAND ||
1264 			     ret == -ERESTART_RESTARTBLOCK))
1265 			ret = -EINTR;
1266 		aio_complete(req, ret, 0);
1267 	}
1268 
1269 	return 0;
1270 }
1271 
1272 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1273 			 struct iocb *iocb, bool compat)
1274 {
1275 	struct kiocb *req;
1276 	ssize_t ret;
1277 
1278 	/* enforce forwards compatibility on users */
1279 	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1280 		pr_debug("EINVAL: reserve field set\n");
1281 		return -EINVAL;
1282 	}
1283 
1284 	/* prevent overflows */
1285 	if (unlikely(
1286 	    (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1287 	    (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1288 	    ((ssize_t)iocb->aio_nbytes < 0)
1289 	   )) {
1290 		pr_debug("EINVAL: io_submit: overflow check\n");
1291 		return -EINVAL;
1292 	}
1293 
1294 	req = aio_get_req(ctx);
1295 	if (unlikely(!req))
1296 		return -EAGAIN;
1297 
1298 	req->ki_filp = fget(iocb->aio_fildes);
1299 	if (unlikely(!req->ki_filp)) {
1300 		ret = -EBADF;
1301 		goto out_put_req;
1302 	}
1303 
1304 	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1305 		/*
1306 		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1307 		 * instance of the file* now. The file descriptor must be
1308 		 * an eventfd() fd, and will be signaled for each completed
1309 		 * event using the eventfd_signal() function.
1310 		 */
1311 		req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1312 		if (IS_ERR(req->ki_eventfd)) {
1313 			ret = PTR_ERR(req->ki_eventfd);
1314 			req->ki_eventfd = NULL;
1315 			goto out_put_req;
1316 		}
1317 	}
1318 
1319 	ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1320 	if (unlikely(ret)) {
1321 		pr_debug("EFAULT: aio_key\n");
1322 		goto out_put_req;
1323 	}
1324 
1325 	req->ki_obj.user = user_iocb;
1326 	req->ki_user_data = iocb->aio_data;
1327 	req->ki_pos = iocb->aio_offset;
1328 	req->ki_nbytes = iocb->aio_nbytes;
1329 
1330 	ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1331 			   (char __user *)(unsigned long)iocb->aio_buf,
1332 			   compat);
1333 	if (ret)
1334 		goto out_put_req;
1335 
1336 	return 0;
1337 out_put_req:
1338 	put_reqs_available(ctx, 1);
1339 	kiocb_free(req);
1340 	return ret;
1341 }
1342 
1343 long do_io_submit(aio_context_t ctx_id, long nr,
1344 		  struct iocb __user *__user *iocbpp, bool compat)
1345 {
1346 	struct kioctx *ctx;
1347 	long ret = 0;
1348 	int i = 0;
1349 	struct blk_plug plug;
1350 
1351 	if (unlikely(nr < 0))
1352 		return -EINVAL;
1353 
1354 	if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1355 		nr = LONG_MAX/sizeof(*iocbpp);
1356 
1357 	if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1358 		return -EFAULT;
1359 
1360 	ctx = lookup_ioctx(ctx_id);
1361 	if (unlikely(!ctx)) {
1362 		pr_debug("EINVAL: invalid context id\n");
1363 		return -EINVAL;
1364 	}
1365 
1366 	blk_start_plug(&plug);
1367 
1368 	/*
1369 	 * AKPM: should this return a partial result if some of the IOs were
1370 	 * successfully submitted?
1371 	 */
1372 	for (i=0; i<nr; i++) {
1373 		struct iocb __user *user_iocb;
1374 		struct iocb tmp;
1375 
1376 		if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1377 			ret = -EFAULT;
1378 			break;
1379 		}
1380 
1381 		if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1382 			ret = -EFAULT;
1383 			break;
1384 		}
1385 
1386 		ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1387 		if (ret)
1388 			break;
1389 	}
1390 	blk_finish_plug(&plug);
1391 
1392 	percpu_ref_put(&ctx->users);
1393 	return i ? i : ret;
1394 }
1395 
1396 /* sys_io_submit:
1397  *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1398  *	the number of iocbs queued.  May return -EINVAL if the aio_context
1399  *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
1400  *	*iocbpp[0] is not properly initialized, if the operation specified
1401  *	is invalid for the file descriptor in the iocb.  May fail with
1402  *	-EFAULT if any of the data structures point to invalid data.  May
1403  *	fail with -EBADF if the file descriptor specified in the first
1404  *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
1405  *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1406  *	fail with -ENOSYS if not implemented.
1407  */
1408 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1409 		struct iocb __user * __user *, iocbpp)
1410 {
1411 	return do_io_submit(ctx_id, nr, iocbpp, 0);
1412 }
1413 
1414 /* lookup_kiocb
1415  *	Finds a given iocb for cancellation.
1416  */
1417 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1418 				  u32 key)
1419 {
1420 	struct list_head *pos;
1421 
1422 	assert_spin_locked(&ctx->ctx_lock);
1423 
1424 	if (key != KIOCB_KEY)
1425 		return NULL;
1426 
1427 	/* TODO: use a hash or array, this sucks. */
1428 	list_for_each(pos, &ctx->active_reqs) {
1429 		struct kiocb *kiocb = list_kiocb(pos);
1430 		if (kiocb->ki_obj.user == iocb)
1431 			return kiocb;
1432 	}
1433 	return NULL;
1434 }
1435 
1436 /* sys_io_cancel:
1437  *	Attempts to cancel an iocb previously passed to io_submit.  If
1438  *	the operation is successfully cancelled, the resulting event is
1439  *	copied into the memory pointed to by result without being placed
1440  *	into the completion queue and 0 is returned.  May fail with
1441  *	-EFAULT if any of the data structures pointed to are invalid.
1442  *	May fail with -EINVAL if aio_context specified by ctx_id is
1443  *	invalid.  May fail with -EAGAIN if the iocb specified was not
1444  *	cancelled.  Will fail with -ENOSYS if not implemented.
1445  */
1446 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1447 		struct io_event __user *, result)
1448 {
1449 	struct kioctx *ctx;
1450 	struct kiocb *kiocb;
1451 	u32 key;
1452 	int ret;
1453 
1454 	ret = get_user(key, &iocb->aio_key);
1455 	if (unlikely(ret))
1456 		return -EFAULT;
1457 
1458 	ctx = lookup_ioctx(ctx_id);
1459 	if (unlikely(!ctx))
1460 		return -EINVAL;
1461 
1462 	spin_lock_irq(&ctx->ctx_lock);
1463 
1464 	kiocb = lookup_kiocb(ctx, iocb, key);
1465 	if (kiocb)
1466 		ret = kiocb_cancel(ctx, kiocb);
1467 	else
1468 		ret = -EINVAL;
1469 
1470 	spin_unlock_irq(&ctx->ctx_lock);
1471 
1472 	if (!ret) {
1473 		/*
1474 		 * The result argument is no longer used - the io_event is
1475 		 * always delivered via the ring buffer. -EINPROGRESS indicates
1476 		 * cancellation is progress:
1477 		 */
1478 		ret = -EINPROGRESS;
1479 	}
1480 
1481 	percpu_ref_put(&ctx->users);
1482 
1483 	return ret;
1484 }
1485 
1486 /* io_getevents:
1487  *	Attempts to read at least min_nr events and up to nr events from
1488  *	the completion queue for the aio_context specified by ctx_id. If
1489  *	it succeeds, the number of read events is returned. May fail with
1490  *	-EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1491  *	out of range, if timeout is out of range.  May fail with -EFAULT
1492  *	if any of the memory specified is invalid.  May return 0 or
1493  *	< min_nr if the timeout specified by timeout has elapsed
1494  *	before sufficient events are available, where timeout == NULL
1495  *	specifies an infinite timeout. Note that the timeout pointed to by
1496  *	timeout is relative.  Will fail with -ENOSYS if not implemented.
1497  */
1498 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1499 		long, min_nr,
1500 		long, nr,
1501 		struct io_event __user *, events,
1502 		struct timespec __user *, timeout)
1503 {
1504 	struct kioctx *ioctx = lookup_ioctx(ctx_id);
1505 	long ret = -EINVAL;
1506 
1507 	if (likely(ioctx)) {
1508 		if (likely(min_nr <= nr && min_nr >= 0))
1509 			ret = read_events(ioctx, min_nr, nr, events, timeout);
1510 		percpu_ref_put(&ioctx->users);
1511 	}
1512 	return ret;
1513 }
1514