• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include <linux/nospec.h>
44 
45 #include <asm/kmap_types.h>
46 #include <asm/uaccess.h>
47 
48 #include "internal.h"
49 
50 #define AIO_RING_MAGIC			0xa10a10a1
51 #define AIO_RING_COMPAT_FEATURES	1
52 #define AIO_RING_INCOMPAT_FEATURES	0
53 struct aio_ring {
54 	unsigned	id;	/* kernel internal index number */
55 	unsigned	nr;	/* number of io_events */
56 	unsigned	head;	/* Written to by userland or under ring_lock
57 				 * mutex by aio_read_events_ring(). */
58 	unsigned	tail;
59 
60 	unsigned	magic;
61 	unsigned	compat_features;
62 	unsigned	incompat_features;
63 	unsigned	header_length;	/* size of aio_ring */
64 
65 
66 	struct io_event		io_events[0];
67 }; /* 128 bytes + ring size */
68 
69 #define AIO_RING_PAGES	8
70 
71 struct kioctx_table {
72 	struct rcu_head		rcu;
73 	unsigned		nr;
74 	struct kioctx __rcu	*table[];
75 };
76 
77 struct kioctx_cpu {
78 	unsigned		reqs_available;
79 };
80 
81 struct ctx_rq_wait {
82 	struct completion comp;
83 	atomic_t count;
84 };
85 
86 struct kioctx {
87 	struct percpu_ref	users;
88 	atomic_t		dead;
89 
90 	struct percpu_ref	reqs;
91 
92 	unsigned long		user_id;
93 
94 	struct __percpu kioctx_cpu *cpu;
95 
96 	/*
97 	 * For percpu reqs_available, number of slots we move to/from global
98 	 * counter at a time:
99 	 */
100 	unsigned		req_batch;
101 	/*
102 	 * This is what userspace passed to io_setup(), it's not used for
103 	 * anything but counting against the global max_reqs quota.
104 	 *
105 	 * The real limit is nr_events - 1, which will be larger (see
106 	 * aio_setup_ring())
107 	 */
108 	unsigned		max_reqs;
109 
110 	/* Size of ringbuffer, in units of struct io_event */
111 	unsigned		nr_events;
112 
113 	unsigned long		mmap_base;
114 	unsigned long		mmap_size;
115 
116 	struct page		**ring_pages;
117 	long			nr_pages;
118 
119 	struct rcu_head		free_rcu;
120 	struct work_struct	free_work;	/* see free_ioctx() */
121 
122 	/*
123 	 * signals when all in-flight requests are done
124 	 */
125 	struct ctx_rq_wait	*rq_wait;
126 
127 	struct {
128 		/*
129 		 * This counts the number of available slots in the ringbuffer,
130 		 * so we avoid overflowing it: it's decremented (if positive)
131 		 * when allocating a kiocb and incremented when the resulting
132 		 * io_event is pulled off the ringbuffer.
133 		 *
134 		 * We batch accesses to it with a percpu version.
135 		 */
136 		atomic_t	reqs_available;
137 	} ____cacheline_aligned_in_smp;
138 
139 	struct {
140 		spinlock_t	ctx_lock;
141 		struct list_head active_reqs;	/* used for cancellation */
142 	} ____cacheline_aligned_in_smp;
143 
144 	struct {
145 		struct mutex	ring_lock;
146 		wait_queue_head_t wait;
147 	} ____cacheline_aligned_in_smp;
148 
149 	struct {
150 		unsigned	tail;
151 		unsigned	completed_events;
152 		spinlock_t	completion_lock;
153 	} ____cacheline_aligned_in_smp;
154 
155 	struct page		*internal_pages[AIO_RING_PAGES];
156 	struct file		*aio_ring_file;
157 
158 	unsigned		id;
159 };
160 
161 /*
162  * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
163  * cancelled or completed (this makes a certain amount of sense because
164  * successful cancellation - io_cancel() - does deliver the completion to
165  * userspace).
166  *
167  * And since most things don't implement kiocb cancellation and we'd really like
168  * kiocb completion to be lockless when possible, we use ki_cancel to
169  * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
170  * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
171  */
172 #define KIOCB_CANCELLED		((void *) (~0ULL))
173 
174 struct aio_kiocb {
175 	struct kiocb		common;
176 
177 	struct kioctx		*ki_ctx;
178 	kiocb_cancel_fn		*ki_cancel;
179 
180 	struct iocb __user	*ki_user_iocb;	/* user's aiocb */
181 	__u64			ki_user_data;	/* user's data for completion */
182 
183 	struct list_head	ki_list;	/* the aio core uses this
184 						 * for cancellation */
185 
186 	/*
187 	 * If the aio_resfd field of the userspace iocb is not zero,
188 	 * this is the underlying eventfd context to deliver events to.
189 	 */
190 	struct eventfd_ctx	*ki_eventfd;
191 };
192 
193 /*------ sysctl variables----*/
194 static DEFINE_SPINLOCK(aio_nr_lock);
195 unsigned long aio_nr;		/* current system wide number of aio requests */
196 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
197 /*----end sysctl variables---*/
198 
199 static struct kmem_cache	*kiocb_cachep;
200 static struct kmem_cache	*kioctx_cachep;
201 
202 static struct vfsmount *aio_mnt;
203 
204 static const struct file_operations aio_ring_fops;
205 static const struct address_space_operations aio_ctx_aops;
206 
aio_private_file(struct kioctx * ctx,loff_t nr_pages)207 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
208 {
209 	struct qstr this = QSTR_INIT("[aio]", 5);
210 	struct file *file;
211 	struct path path;
212 	struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
213 	if (IS_ERR(inode))
214 		return ERR_CAST(inode);
215 
216 	inode->i_mapping->a_ops = &aio_ctx_aops;
217 	inode->i_mapping->private_data = ctx;
218 	inode->i_size = PAGE_SIZE * nr_pages;
219 
220 	path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
221 	if (!path.dentry) {
222 		iput(inode);
223 		return ERR_PTR(-ENOMEM);
224 	}
225 	path.mnt = mntget(aio_mnt);
226 
227 	d_instantiate(path.dentry, inode);
228 	file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
229 	if (IS_ERR(file)) {
230 		path_put(&path);
231 		return file;
232 	}
233 
234 	file->f_flags = O_RDWR;
235 	return file;
236 }
237 
aio_mount(struct file_system_type * fs_type,int flags,const char * dev_name,void * data)238 static struct dentry *aio_mount(struct file_system_type *fs_type,
239 				int flags, const char *dev_name, void *data)
240 {
241 	static const struct dentry_operations ops = {
242 		.d_dname	= simple_dname,
243 	};
244 	struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops,
245 					   AIO_RING_MAGIC);
246 
247 	if (!IS_ERR(root))
248 		root->d_sb->s_iflags |= SB_I_NOEXEC;
249 	return root;
250 }
251 
252 /* aio_setup
253  *	Creates the slab caches used by the aio routines, panic on
254  *	failure as this is done early during the boot sequence.
255  */
aio_setup(void)256 static int __init aio_setup(void)
257 {
258 	static struct file_system_type aio_fs = {
259 		.name		= "aio",
260 		.mount		= aio_mount,
261 		.kill_sb	= kill_anon_super,
262 	};
263 	aio_mnt = kern_mount(&aio_fs);
264 	if (IS_ERR(aio_mnt))
265 		panic("Failed to create aio fs mount.");
266 
267 	kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
268 	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
269 
270 	pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
271 
272 	return 0;
273 }
274 __initcall(aio_setup);
275 
put_aio_ring_file(struct kioctx * ctx)276 static void put_aio_ring_file(struct kioctx *ctx)
277 {
278 	struct file *aio_ring_file = ctx->aio_ring_file;
279 	if (aio_ring_file) {
280 		truncate_setsize(aio_ring_file->f_inode, 0);
281 
282 		/* Prevent further access to the kioctx from migratepages */
283 		spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
284 		aio_ring_file->f_inode->i_mapping->private_data = NULL;
285 		ctx->aio_ring_file = NULL;
286 		spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
287 
288 		fput(aio_ring_file);
289 	}
290 }
291 
aio_free_ring(struct kioctx * ctx)292 static void aio_free_ring(struct kioctx *ctx)
293 {
294 	int i;
295 
296 	/* Disconnect the kiotx from the ring file.  This prevents future
297 	 * accesses to the kioctx from page migration.
298 	 */
299 	put_aio_ring_file(ctx);
300 
301 	for (i = 0; i < ctx->nr_pages; i++) {
302 		struct page *page;
303 		pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
304 				page_count(ctx->ring_pages[i]));
305 		page = ctx->ring_pages[i];
306 		if (!page)
307 			continue;
308 		ctx->ring_pages[i] = NULL;
309 		put_page(page);
310 	}
311 
312 	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
313 		kfree(ctx->ring_pages);
314 		ctx->ring_pages = NULL;
315 	}
316 }
317 
aio_ring_mremap(struct vm_area_struct * vma)318 static int aio_ring_mremap(struct vm_area_struct *vma)
319 {
320 	struct file *file = vma->vm_file;
321 	struct mm_struct *mm = vma->vm_mm;
322 	struct kioctx_table *table;
323 	int i, res = -EINVAL;
324 
325 	spin_lock(&mm->ioctx_lock);
326 	rcu_read_lock();
327 	table = rcu_dereference(mm->ioctx_table);
328 	for (i = 0; i < table->nr; i++) {
329 		struct kioctx *ctx;
330 
331 		ctx = rcu_dereference(table->table[i]);
332 		if (ctx && ctx->aio_ring_file == file) {
333 			if (!atomic_read(&ctx->dead)) {
334 				ctx->user_id = ctx->mmap_base = vma->vm_start;
335 				res = 0;
336 			}
337 			break;
338 		}
339 	}
340 
341 	rcu_read_unlock();
342 	spin_unlock(&mm->ioctx_lock);
343 	return res;
344 }
345 
346 static const struct vm_operations_struct aio_ring_vm_ops = {
347 	.mremap		= aio_ring_mremap,
348 #if IS_ENABLED(CONFIG_MMU)
349 	.fault		= filemap_fault,
350 	.map_pages	= filemap_map_pages,
351 	.page_mkwrite	= filemap_page_mkwrite,
352 #endif
353 };
354 
aio_ring_mmap(struct file * file,struct vm_area_struct * vma)355 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
356 {
357 	vma->vm_flags |= VM_DONTEXPAND;
358 	vma->vm_ops = &aio_ring_vm_ops;
359 	return 0;
360 }
361 
362 static const struct file_operations aio_ring_fops = {
363 	.mmap = aio_ring_mmap,
364 };
365 
366 #if IS_ENABLED(CONFIG_MIGRATION)
aio_migratepage(struct address_space * mapping,struct page * new,struct page * old,enum migrate_mode mode)367 static int aio_migratepage(struct address_space *mapping, struct page *new,
368 			struct page *old, enum migrate_mode mode)
369 {
370 	struct kioctx *ctx;
371 	unsigned long flags;
372 	pgoff_t idx;
373 	int rc;
374 
375 	rc = 0;
376 
377 	/* mapping->private_lock here protects against the kioctx teardown.  */
378 	spin_lock(&mapping->private_lock);
379 	ctx = mapping->private_data;
380 	if (!ctx) {
381 		rc = -EINVAL;
382 		goto out;
383 	}
384 
385 	/* The ring_lock mutex.  The prevents aio_read_events() from writing
386 	 * to the ring's head, and prevents page migration from mucking in
387 	 * a partially initialized kiotx.
388 	 */
389 	if (!mutex_trylock(&ctx->ring_lock)) {
390 		rc = -EAGAIN;
391 		goto out;
392 	}
393 
394 	idx = old->index;
395 	if (idx < (pgoff_t)ctx->nr_pages) {
396 		/* Make sure the old page hasn't already been changed */
397 		if (ctx->ring_pages[idx] != old)
398 			rc = -EAGAIN;
399 	} else
400 		rc = -EINVAL;
401 
402 	if (rc != 0)
403 		goto out_unlock;
404 
405 	/* Writeback must be complete */
406 	BUG_ON(PageWriteback(old));
407 	get_page(new);
408 
409 	rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
410 	if (rc != MIGRATEPAGE_SUCCESS) {
411 		put_page(new);
412 		goto out_unlock;
413 	}
414 
415 	/* Take completion_lock to prevent other writes to the ring buffer
416 	 * while the old page is copied to the new.  This prevents new
417 	 * events from being lost.
418 	 */
419 	spin_lock_irqsave(&ctx->completion_lock, flags);
420 	migrate_page_copy(new, old);
421 	BUG_ON(ctx->ring_pages[idx] != old);
422 	ctx->ring_pages[idx] = new;
423 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
424 
425 	/* The old page is no longer accessible. */
426 	put_page(old);
427 
428 out_unlock:
429 	mutex_unlock(&ctx->ring_lock);
430 out:
431 	spin_unlock(&mapping->private_lock);
432 	return rc;
433 }
434 #endif
435 
436 static const struct address_space_operations aio_ctx_aops = {
437 	.set_page_dirty = __set_page_dirty_no_writeback,
438 #if IS_ENABLED(CONFIG_MIGRATION)
439 	.migratepage	= aio_migratepage,
440 #endif
441 };
442 
aio_setup_ring(struct kioctx * ctx)443 static int aio_setup_ring(struct kioctx *ctx)
444 {
445 	struct aio_ring *ring;
446 	unsigned nr_events = ctx->max_reqs;
447 	struct mm_struct *mm = current->mm;
448 	unsigned long size, unused;
449 	int nr_pages;
450 	int i;
451 	struct file *file;
452 
453 	/* Compensate for the ring buffer's head/tail overlap entry */
454 	nr_events += 2;	/* 1 is required, 2 for good luck */
455 
456 	size = sizeof(struct aio_ring);
457 	size += sizeof(struct io_event) * nr_events;
458 
459 	nr_pages = PFN_UP(size);
460 	if (nr_pages < 0)
461 		return -EINVAL;
462 
463 	file = aio_private_file(ctx, nr_pages);
464 	if (IS_ERR(file)) {
465 		ctx->aio_ring_file = NULL;
466 		return -ENOMEM;
467 	}
468 
469 	ctx->aio_ring_file = file;
470 	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
471 			/ sizeof(struct io_event);
472 
473 	ctx->ring_pages = ctx->internal_pages;
474 	if (nr_pages > AIO_RING_PAGES) {
475 		ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
476 					  GFP_KERNEL);
477 		if (!ctx->ring_pages) {
478 			put_aio_ring_file(ctx);
479 			return -ENOMEM;
480 		}
481 	}
482 
483 	for (i = 0; i < nr_pages; i++) {
484 		struct page *page;
485 		page = find_or_create_page(file->f_inode->i_mapping,
486 					   i, GFP_HIGHUSER | __GFP_ZERO);
487 		if (!page)
488 			break;
489 		pr_debug("pid(%d) page[%d]->count=%d\n",
490 			 current->pid, i, page_count(page));
491 		SetPageUptodate(page);
492 		unlock_page(page);
493 
494 		ctx->ring_pages[i] = page;
495 	}
496 	ctx->nr_pages = i;
497 
498 	if (unlikely(i != nr_pages)) {
499 		aio_free_ring(ctx);
500 		return -ENOMEM;
501 	}
502 
503 	ctx->mmap_size = nr_pages * PAGE_SIZE;
504 	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
505 
506 	down_write(&mm->mmap_sem);
507 	ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
508 				       PROT_READ | PROT_WRITE,
509 				       MAP_SHARED, 0, &unused);
510 	up_write(&mm->mmap_sem);
511 	if (IS_ERR((void *)ctx->mmap_base)) {
512 		ctx->mmap_size = 0;
513 		aio_free_ring(ctx);
514 		return -ENOMEM;
515 	}
516 
517 	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
518 
519 	ctx->user_id = ctx->mmap_base;
520 	ctx->nr_events = nr_events; /* trusted copy */
521 
522 	ring = kmap_atomic(ctx->ring_pages[0]);
523 	ring->nr = nr_events;	/* user copy */
524 	ring->id = ~0U;
525 	ring->head = ring->tail = 0;
526 	ring->magic = AIO_RING_MAGIC;
527 	ring->compat_features = AIO_RING_COMPAT_FEATURES;
528 	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
529 	ring->header_length = sizeof(struct aio_ring);
530 	kunmap_atomic(ring);
531 	flush_dcache_page(ctx->ring_pages[0]);
532 
533 	return 0;
534 }
535 
536 #define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
537 #define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
538 #define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
539 
kiocb_set_cancel_fn(struct kiocb * iocb,kiocb_cancel_fn * cancel)540 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
541 {
542 	struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
543 	struct kioctx *ctx = req->ki_ctx;
544 	unsigned long flags;
545 
546 	spin_lock_irqsave(&ctx->ctx_lock, flags);
547 
548 	if (!req->ki_list.next)
549 		list_add(&req->ki_list, &ctx->active_reqs);
550 
551 	req->ki_cancel = cancel;
552 
553 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
554 }
555 EXPORT_SYMBOL(kiocb_set_cancel_fn);
556 
kiocb_cancel(struct aio_kiocb * kiocb)557 static int kiocb_cancel(struct aio_kiocb *kiocb)
558 {
559 	kiocb_cancel_fn *old, *cancel;
560 
561 	/*
562 	 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
563 	 * actually has a cancel function, hence the cmpxchg()
564 	 */
565 
566 	cancel = ACCESS_ONCE(kiocb->ki_cancel);
567 	do {
568 		if (!cancel || cancel == KIOCB_CANCELLED)
569 			return -EINVAL;
570 
571 		old = cancel;
572 		cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
573 	} while (cancel != old);
574 
575 	return cancel(&kiocb->common);
576 }
577 
578 /*
579  * free_ioctx() should be RCU delayed to synchronize against the RCU
580  * protected lookup_ioctx() and also needs process context to call
581  * aio_free_ring(), so the double bouncing through kioctx->free_rcu and
582  * ->free_work.
583  */
free_ioctx(struct work_struct * work)584 static void free_ioctx(struct work_struct *work)
585 {
586 	struct kioctx *ctx = container_of(work, struct kioctx, free_work);
587 
588 	pr_debug("freeing %p\n", ctx);
589 
590 	aio_free_ring(ctx);
591 	free_percpu(ctx->cpu);
592 	percpu_ref_exit(&ctx->reqs);
593 	percpu_ref_exit(&ctx->users);
594 	kmem_cache_free(kioctx_cachep, ctx);
595 }
596 
free_ioctx_rcufn(struct rcu_head * head)597 static void free_ioctx_rcufn(struct rcu_head *head)
598 {
599 	struct kioctx *ctx = container_of(head, struct kioctx, free_rcu);
600 
601 	INIT_WORK(&ctx->free_work, free_ioctx);
602 	schedule_work(&ctx->free_work);
603 }
604 
free_ioctx_reqs(struct percpu_ref * ref)605 static void free_ioctx_reqs(struct percpu_ref *ref)
606 {
607 	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
608 
609 	/* At this point we know that there are no any in-flight requests */
610 	if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
611 		complete(&ctx->rq_wait->comp);
612 
613 	/* Synchronize against RCU protected table->table[] dereferences */
614 	call_rcu(&ctx->free_rcu, free_ioctx_rcufn);
615 }
616 
617 /*
618  * When this function runs, the kioctx has been removed from the "hash table"
619  * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
620  * now it's safe to cancel any that need to be.
621  */
free_ioctx_users(struct percpu_ref * ref)622 static void free_ioctx_users(struct percpu_ref *ref)
623 {
624 	struct kioctx *ctx = container_of(ref, struct kioctx, users);
625 	struct aio_kiocb *req;
626 
627 	spin_lock_irq(&ctx->ctx_lock);
628 
629 	while (!list_empty(&ctx->active_reqs)) {
630 		req = list_first_entry(&ctx->active_reqs,
631 				       struct aio_kiocb, ki_list);
632 		kiocb_cancel(req);
633 		list_del_init(&req->ki_list);
634 	}
635 
636 	spin_unlock_irq(&ctx->ctx_lock);
637 
638 	percpu_ref_kill(&ctx->reqs);
639 	percpu_ref_put(&ctx->reqs);
640 }
641 
ioctx_add_table(struct kioctx * ctx,struct mm_struct * mm)642 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
643 {
644 	unsigned i, new_nr;
645 	struct kioctx_table *table, *old;
646 	struct aio_ring *ring;
647 
648 	spin_lock(&mm->ioctx_lock);
649 	table = rcu_dereference_raw(mm->ioctx_table);
650 
651 	while (1) {
652 		if (table)
653 			for (i = 0; i < table->nr; i++)
654 				if (!rcu_access_pointer(table->table[i])) {
655 					ctx->id = i;
656 					rcu_assign_pointer(table->table[i], ctx);
657 					spin_unlock(&mm->ioctx_lock);
658 
659 					/* While kioctx setup is in progress,
660 					 * we are protected from page migration
661 					 * changes ring_pages by ->ring_lock.
662 					 */
663 					ring = kmap_atomic(ctx->ring_pages[0]);
664 					ring->id = ctx->id;
665 					kunmap_atomic(ring);
666 					return 0;
667 				}
668 
669 		new_nr = (table ? table->nr : 1) * 4;
670 		spin_unlock(&mm->ioctx_lock);
671 
672 		table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
673 				new_nr, GFP_KERNEL);
674 		if (!table)
675 			return -ENOMEM;
676 
677 		table->nr = new_nr;
678 
679 		spin_lock(&mm->ioctx_lock);
680 		old = rcu_dereference_raw(mm->ioctx_table);
681 
682 		if (!old) {
683 			rcu_assign_pointer(mm->ioctx_table, table);
684 		} else if (table->nr > old->nr) {
685 			memcpy(table->table, old->table,
686 			       old->nr * sizeof(struct kioctx *));
687 
688 			rcu_assign_pointer(mm->ioctx_table, table);
689 			kfree_rcu(old, rcu);
690 		} else {
691 			kfree(table);
692 			table = old;
693 		}
694 	}
695 }
696 
aio_nr_sub(unsigned nr)697 static void aio_nr_sub(unsigned nr)
698 {
699 	spin_lock(&aio_nr_lock);
700 	if (WARN_ON(aio_nr - nr > aio_nr))
701 		aio_nr = 0;
702 	else
703 		aio_nr -= nr;
704 	spin_unlock(&aio_nr_lock);
705 }
706 
707 /* ioctx_alloc
708  *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
709  */
ioctx_alloc(unsigned nr_events)710 static struct kioctx *ioctx_alloc(unsigned nr_events)
711 {
712 	struct mm_struct *mm = current->mm;
713 	struct kioctx *ctx;
714 	int err = -ENOMEM;
715 
716 	/*
717 	 * We keep track of the number of available ringbuffer slots, to prevent
718 	 * overflow (reqs_available), and we also use percpu counters for this.
719 	 *
720 	 * So since up to half the slots might be on other cpu's percpu counters
721 	 * and unavailable, double nr_events so userspace sees what they
722 	 * expected: additionally, we move req_batch slots to/from percpu
723 	 * counters at a time, so make sure that isn't 0:
724 	 */
725 	nr_events = max(nr_events, num_possible_cpus() * 4);
726 	nr_events *= 2;
727 
728 	/* Prevent overflows */
729 	if (nr_events > (0x10000000U / sizeof(struct io_event))) {
730 		pr_debug("ENOMEM: nr_events too high\n");
731 		return ERR_PTR(-EINVAL);
732 	}
733 
734 	if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
735 		return ERR_PTR(-EAGAIN);
736 
737 	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
738 	if (!ctx)
739 		return ERR_PTR(-ENOMEM);
740 
741 	ctx->max_reqs = nr_events;
742 
743 	spin_lock_init(&ctx->ctx_lock);
744 	spin_lock_init(&ctx->completion_lock);
745 	mutex_init(&ctx->ring_lock);
746 	/* Protect against page migration throughout kiotx setup by keeping
747 	 * the ring_lock mutex held until setup is complete. */
748 	mutex_lock(&ctx->ring_lock);
749 	init_waitqueue_head(&ctx->wait);
750 
751 	INIT_LIST_HEAD(&ctx->active_reqs);
752 
753 	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
754 		goto err;
755 
756 	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
757 		goto err;
758 
759 	ctx->cpu = alloc_percpu(struct kioctx_cpu);
760 	if (!ctx->cpu)
761 		goto err;
762 
763 	err = aio_setup_ring(ctx);
764 	if (err < 0)
765 		goto err;
766 
767 	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
768 	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
769 	if (ctx->req_batch < 1)
770 		ctx->req_batch = 1;
771 
772 	/* limit the number of system wide aios */
773 	spin_lock(&aio_nr_lock);
774 	if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
775 	    aio_nr + nr_events < aio_nr) {
776 		spin_unlock(&aio_nr_lock);
777 		err = -EAGAIN;
778 		goto err_ctx;
779 	}
780 	aio_nr += ctx->max_reqs;
781 	spin_unlock(&aio_nr_lock);
782 
783 	percpu_ref_get(&ctx->users);	/* io_setup() will drop this ref */
784 	percpu_ref_get(&ctx->reqs);	/* free_ioctx_users() will drop this */
785 
786 	err = ioctx_add_table(ctx, mm);
787 	if (err)
788 		goto err_cleanup;
789 
790 	/* Release the ring_lock mutex now that all setup is complete. */
791 	mutex_unlock(&ctx->ring_lock);
792 
793 	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
794 		 ctx, ctx->user_id, mm, ctx->nr_events);
795 	return ctx;
796 
797 err_cleanup:
798 	aio_nr_sub(ctx->max_reqs);
799 err_ctx:
800 	atomic_set(&ctx->dead, 1);
801 	if (ctx->mmap_size)
802 		vm_munmap(ctx->mmap_base, ctx->mmap_size);
803 	aio_free_ring(ctx);
804 err:
805 	mutex_unlock(&ctx->ring_lock);
806 	free_percpu(ctx->cpu);
807 	percpu_ref_exit(&ctx->reqs);
808 	percpu_ref_exit(&ctx->users);
809 	kmem_cache_free(kioctx_cachep, ctx);
810 	pr_debug("error allocating ioctx %d\n", err);
811 	return ERR_PTR(err);
812 }
813 
814 /* kill_ioctx
815  *	Cancels all outstanding aio requests on an aio context.  Used
816  *	when the processes owning a context have all exited to encourage
817  *	the rapid destruction of the kioctx.
818  */
kill_ioctx(struct mm_struct * mm,struct kioctx * ctx,struct ctx_rq_wait * wait)819 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
820 		      struct ctx_rq_wait *wait)
821 {
822 	struct kioctx_table *table;
823 
824 	spin_lock(&mm->ioctx_lock);
825 	if (atomic_xchg(&ctx->dead, 1)) {
826 		spin_unlock(&mm->ioctx_lock);
827 		return -EINVAL;
828 	}
829 
830 	table = rcu_dereference_raw(mm->ioctx_table);
831 	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
832 	RCU_INIT_POINTER(table->table[ctx->id], NULL);
833 	spin_unlock(&mm->ioctx_lock);
834 
835 	/* free_ioctx_reqs() will do the necessary RCU synchronization */
836 	wake_up_all(&ctx->wait);
837 
838 	/*
839 	 * It'd be more correct to do this in free_ioctx(), after all
840 	 * the outstanding kiocbs have finished - but by then io_destroy
841 	 * has already returned, so io_setup() could potentially return
842 	 * -EAGAIN with no ioctxs actually in use (as far as userspace
843 	 *  could tell).
844 	 */
845 	aio_nr_sub(ctx->max_reqs);
846 
847 	if (ctx->mmap_size)
848 		vm_munmap(ctx->mmap_base, ctx->mmap_size);
849 
850 	ctx->rq_wait = wait;
851 	percpu_ref_kill(&ctx->users);
852 	return 0;
853 }
854 
855 /*
856  * exit_aio: called when the last user of mm goes away.  At this point, there is
857  * no way for any new requests to be submited or any of the io_* syscalls to be
858  * called on the context.
859  *
860  * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
861  * them.
862  */
exit_aio(struct mm_struct * mm)863 void exit_aio(struct mm_struct *mm)
864 {
865 	struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
866 	struct ctx_rq_wait wait;
867 	int i, skipped;
868 
869 	if (!table)
870 		return;
871 
872 	atomic_set(&wait.count, table->nr);
873 	init_completion(&wait.comp);
874 
875 	skipped = 0;
876 	for (i = 0; i < table->nr; ++i) {
877 		struct kioctx *ctx =
878 			rcu_dereference_protected(table->table[i], true);
879 
880 		if (!ctx) {
881 			skipped++;
882 			continue;
883 		}
884 
885 		/*
886 		 * We don't need to bother with munmap() here - exit_mmap(mm)
887 		 * is coming and it'll unmap everything. And we simply can't,
888 		 * this is not necessarily our ->mm.
889 		 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
890 		 * that it needs to unmap the area, just set it to 0.
891 		 */
892 		ctx->mmap_size = 0;
893 		kill_ioctx(mm, ctx, &wait);
894 	}
895 
896 	if (!atomic_sub_and_test(skipped, &wait.count)) {
897 		/* Wait until all IO for the context are done. */
898 		wait_for_completion(&wait.comp);
899 	}
900 
901 	RCU_INIT_POINTER(mm->ioctx_table, NULL);
902 	kfree(table);
903 }
904 
put_reqs_available(struct kioctx * ctx,unsigned nr)905 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
906 {
907 	struct kioctx_cpu *kcpu;
908 	unsigned long flags;
909 
910 	local_irq_save(flags);
911 	kcpu = this_cpu_ptr(ctx->cpu);
912 	kcpu->reqs_available += nr;
913 
914 	while (kcpu->reqs_available >= ctx->req_batch * 2) {
915 		kcpu->reqs_available -= ctx->req_batch;
916 		atomic_add(ctx->req_batch, &ctx->reqs_available);
917 	}
918 
919 	local_irq_restore(flags);
920 }
921 
get_reqs_available(struct kioctx * ctx)922 static bool get_reqs_available(struct kioctx *ctx)
923 {
924 	struct kioctx_cpu *kcpu;
925 	bool ret = false;
926 	unsigned long flags;
927 
928 	local_irq_save(flags);
929 	kcpu = this_cpu_ptr(ctx->cpu);
930 	if (!kcpu->reqs_available) {
931 		int old, avail = atomic_read(&ctx->reqs_available);
932 
933 		do {
934 			if (avail < ctx->req_batch)
935 				goto out;
936 
937 			old = avail;
938 			avail = atomic_cmpxchg(&ctx->reqs_available,
939 					       avail, avail - ctx->req_batch);
940 		} while (avail != old);
941 
942 		kcpu->reqs_available += ctx->req_batch;
943 	}
944 
945 	ret = true;
946 	kcpu->reqs_available--;
947 out:
948 	local_irq_restore(flags);
949 	return ret;
950 }
951 
952 /* refill_reqs_available
953  *	Updates the reqs_available reference counts used for tracking the
954  *	number of free slots in the completion ring.  This can be called
955  *	from aio_complete() (to optimistically update reqs_available) or
956  *	from aio_get_req() (the we're out of events case).  It must be
957  *	called holding ctx->completion_lock.
958  */
refill_reqs_available(struct kioctx * ctx,unsigned head,unsigned tail)959 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
960                                   unsigned tail)
961 {
962 	unsigned events_in_ring, completed;
963 
964 	/* Clamp head since userland can write to it. */
965 	head %= ctx->nr_events;
966 	if (head <= tail)
967 		events_in_ring = tail - head;
968 	else
969 		events_in_ring = ctx->nr_events - (head - tail);
970 
971 	completed = ctx->completed_events;
972 	if (events_in_ring < completed)
973 		completed -= events_in_ring;
974 	else
975 		completed = 0;
976 
977 	if (!completed)
978 		return;
979 
980 	ctx->completed_events -= completed;
981 	put_reqs_available(ctx, completed);
982 }
983 
984 /* user_refill_reqs_available
985  *	Called to refill reqs_available when aio_get_req() encounters an
986  *	out of space in the completion ring.
987  */
user_refill_reqs_available(struct kioctx * ctx)988 static void user_refill_reqs_available(struct kioctx *ctx)
989 {
990 	spin_lock_irq(&ctx->completion_lock);
991 	if (ctx->completed_events) {
992 		struct aio_ring *ring;
993 		unsigned head;
994 
995 		/* Access of ring->head may race with aio_read_events_ring()
996 		 * here, but that's okay since whether we read the old version
997 		 * or the new version, and either will be valid.  The important
998 		 * part is that head cannot pass tail since we prevent
999 		 * aio_complete() from updating tail by holding
1000 		 * ctx->completion_lock.  Even if head is invalid, the check
1001 		 * against ctx->completed_events below will make sure we do the
1002 		 * safe/right thing.
1003 		 */
1004 		ring = kmap_atomic(ctx->ring_pages[0]);
1005 		head = ring->head;
1006 		kunmap_atomic(ring);
1007 
1008 		refill_reqs_available(ctx, head, ctx->tail);
1009 	}
1010 
1011 	spin_unlock_irq(&ctx->completion_lock);
1012 }
1013 
1014 /* aio_get_req
1015  *	Allocate a slot for an aio request.
1016  * Returns NULL if no requests are free.
1017  */
aio_get_req(struct kioctx * ctx)1018 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1019 {
1020 	struct aio_kiocb *req;
1021 
1022 	if (!get_reqs_available(ctx)) {
1023 		user_refill_reqs_available(ctx);
1024 		if (!get_reqs_available(ctx))
1025 			return NULL;
1026 	}
1027 
1028 	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1029 	if (unlikely(!req))
1030 		goto out_put;
1031 
1032 	percpu_ref_get(&ctx->reqs);
1033 
1034 	req->ki_ctx = ctx;
1035 	return req;
1036 out_put:
1037 	put_reqs_available(ctx, 1);
1038 	return NULL;
1039 }
1040 
kiocb_free(struct aio_kiocb * req)1041 static void kiocb_free(struct aio_kiocb *req)
1042 {
1043 	if (req->common.ki_filp)
1044 		fput(req->common.ki_filp);
1045 	if (req->ki_eventfd != NULL)
1046 		eventfd_ctx_put(req->ki_eventfd);
1047 	kmem_cache_free(kiocb_cachep, req);
1048 }
1049 
lookup_ioctx(unsigned long ctx_id)1050 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1051 {
1052 	struct aio_ring __user *ring  = (void __user *)ctx_id;
1053 	struct mm_struct *mm = current->mm;
1054 	struct kioctx *ctx, *ret = NULL;
1055 	struct kioctx_table *table;
1056 	unsigned id;
1057 
1058 	if (get_user(id, &ring->id))
1059 		return NULL;
1060 
1061 	rcu_read_lock();
1062 	table = rcu_dereference(mm->ioctx_table);
1063 
1064 	if (!table || id >= table->nr)
1065 		goto out;
1066 
1067 	id = array_index_nospec(id, table->nr);
1068 	ctx = rcu_dereference(table->table[id]);
1069 	if (ctx && ctx->user_id == ctx_id) {
1070 		if (percpu_ref_tryget_live(&ctx->users))
1071 			ret = ctx;
1072 	}
1073 out:
1074 	rcu_read_unlock();
1075 	return ret;
1076 }
1077 
1078 /* aio_complete
1079  *	Called when the io request on the given iocb is complete.
1080  */
aio_complete(struct kiocb * kiocb,long res,long res2)1081 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1082 {
1083 	struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1084 	struct kioctx	*ctx = iocb->ki_ctx;
1085 	struct aio_ring	*ring;
1086 	struct io_event	*ev_page, *event;
1087 	unsigned tail, pos, head;
1088 	unsigned long	flags;
1089 
1090 	/*
1091 	 * Special case handling for sync iocbs:
1092 	 *  - events go directly into the iocb for fast handling
1093 	 *  - the sync task with the iocb in its stack holds the single iocb
1094 	 *    ref, no other paths have a way to get another ref
1095 	 *  - the sync task helpfully left a reference to itself in the iocb
1096 	 */
1097 	BUG_ON(is_sync_kiocb(kiocb));
1098 
1099 	if (iocb->ki_list.next) {
1100 		unsigned long flags;
1101 
1102 		spin_lock_irqsave(&ctx->ctx_lock, flags);
1103 		list_del(&iocb->ki_list);
1104 		spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1105 	}
1106 
1107 	/*
1108 	 * Add a completion event to the ring buffer. Must be done holding
1109 	 * ctx->completion_lock to prevent other code from messing with the tail
1110 	 * pointer since we might be called from irq context.
1111 	 */
1112 	spin_lock_irqsave(&ctx->completion_lock, flags);
1113 
1114 	tail = ctx->tail;
1115 	pos = tail + AIO_EVENTS_OFFSET;
1116 
1117 	if (++tail >= ctx->nr_events)
1118 		tail = 0;
1119 
1120 	ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1121 	event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1122 
1123 	event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1124 	event->data = iocb->ki_user_data;
1125 	event->res = res;
1126 	event->res2 = res2;
1127 
1128 	kunmap_atomic(ev_page);
1129 	flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1130 
1131 	pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1132 		 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1133 		 res, res2);
1134 
1135 	/* after flagging the request as done, we
1136 	 * must never even look at it again
1137 	 */
1138 	smp_wmb();	/* make event visible before updating tail */
1139 
1140 	ctx->tail = tail;
1141 
1142 	ring = kmap_atomic(ctx->ring_pages[0]);
1143 	head = ring->head;
1144 	ring->tail = tail;
1145 	kunmap_atomic(ring);
1146 	flush_dcache_page(ctx->ring_pages[0]);
1147 
1148 	ctx->completed_events++;
1149 	if (ctx->completed_events > 1)
1150 		refill_reqs_available(ctx, head, tail);
1151 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
1152 
1153 	pr_debug("added to ring %p at [%u]\n", iocb, tail);
1154 
1155 	/*
1156 	 * Check if the user asked us to deliver the result through an
1157 	 * eventfd. The eventfd_signal() function is safe to be called
1158 	 * from IRQ context.
1159 	 */
1160 	if (iocb->ki_eventfd != NULL)
1161 		eventfd_signal(iocb->ki_eventfd, 1);
1162 
1163 	/* everything turned out well, dispose of the aiocb. */
1164 	kiocb_free(iocb);
1165 
1166 	/*
1167 	 * We have to order our ring_info tail store above and test
1168 	 * of the wait list below outside the wait lock.  This is
1169 	 * like in wake_up_bit() where clearing a bit has to be
1170 	 * ordered with the unlocked test.
1171 	 */
1172 	smp_mb();
1173 
1174 	if (waitqueue_active(&ctx->wait))
1175 		wake_up(&ctx->wait);
1176 
1177 	percpu_ref_put(&ctx->reqs);
1178 }
1179 
1180 /* aio_read_events_ring
1181  *	Pull an event off of the ioctx's event ring.  Returns the number of
1182  *	events fetched
1183  */
aio_read_events_ring(struct kioctx * ctx,struct io_event __user * event,long nr)1184 static long aio_read_events_ring(struct kioctx *ctx,
1185 				 struct io_event __user *event, long nr)
1186 {
1187 	struct aio_ring *ring;
1188 	unsigned head, tail, pos;
1189 	long ret = 0;
1190 	int copy_ret;
1191 
1192 	/*
1193 	 * The mutex can block and wake us up and that will cause
1194 	 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1195 	 * and repeat. This should be rare enough that it doesn't cause
1196 	 * peformance issues. See the comment in read_events() for more detail.
1197 	 */
1198 	sched_annotate_sleep();
1199 	mutex_lock(&ctx->ring_lock);
1200 
1201 	/* Access to ->ring_pages here is protected by ctx->ring_lock. */
1202 	ring = kmap_atomic(ctx->ring_pages[0]);
1203 	head = ring->head;
1204 	tail = ring->tail;
1205 	kunmap_atomic(ring);
1206 
1207 	/*
1208 	 * Ensure that once we've read the current tail pointer, that
1209 	 * we also see the events that were stored up to the tail.
1210 	 */
1211 	smp_rmb();
1212 
1213 	pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1214 
1215 	if (head == tail)
1216 		goto out;
1217 
1218 	head %= ctx->nr_events;
1219 	tail %= ctx->nr_events;
1220 
1221 	while (ret < nr) {
1222 		long avail;
1223 		struct io_event *ev;
1224 		struct page *page;
1225 
1226 		avail = (head <= tail ?  tail : ctx->nr_events) - head;
1227 		if (head == tail)
1228 			break;
1229 
1230 		avail = min(avail, nr - ret);
1231 		avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1232 			    ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1233 
1234 		pos = head + AIO_EVENTS_OFFSET;
1235 		page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1236 		pos %= AIO_EVENTS_PER_PAGE;
1237 
1238 		ev = kmap(page);
1239 		copy_ret = copy_to_user(event + ret, ev + pos,
1240 					sizeof(*ev) * avail);
1241 		kunmap(page);
1242 
1243 		if (unlikely(copy_ret)) {
1244 			ret = -EFAULT;
1245 			goto out;
1246 		}
1247 
1248 		ret += avail;
1249 		head += avail;
1250 		head %= ctx->nr_events;
1251 	}
1252 
1253 	ring = kmap_atomic(ctx->ring_pages[0]);
1254 	ring->head = head;
1255 	kunmap_atomic(ring);
1256 	flush_dcache_page(ctx->ring_pages[0]);
1257 
1258 	pr_debug("%li  h%u t%u\n", ret, head, tail);
1259 out:
1260 	mutex_unlock(&ctx->ring_lock);
1261 
1262 	return ret;
1263 }
1264 
aio_read_events(struct kioctx * ctx,long min_nr,long nr,struct io_event __user * event,long * i)1265 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1266 			    struct io_event __user *event, long *i)
1267 {
1268 	long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1269 
1270 	if (ret > 0)
1271 		*i += ret;
1272 
1273 	if (unlikely(atomic_read(&ctx->dead)))
1274 		ret = -EINVAL;
1275 
1276 	if (!*i)
1277 		*i = ret;
1278 
1279 	return ret < 0 || *i >= min_nr;
1280 }
1281 
read_events(struct kioctx * ctx,long min_nr,long nr,struct io_event __user * event,struct timespec __user * timeout)1282 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1283 			struct io_event __user *event,
1284 			struct timespec __user *timeout)
1285 {
1286 	ktime_t until = { .tv64 = KTIME_MAX };
1287 	long ret = 0;
1288 
1289 	if (timeout) {
1290 		struct timespec	ts;
1291 
1292 		if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1293 			return -EFAULT;
1294 
1295 		until = timespec_to_ktime(ts);
1296 	}
1297 
1298 	/*
1299 	 * Note that aio_read_events() is being called as the conditional - i.e.
1300 	 * we're calling it after prepare_to_wait() has set task state to
1301 	 * TASK_INTERRUPTIBLE.
1302 	 *
1303 	 * But aio_read_events() can block, and if it blocks it's going to flip
1304 	 * the task state back to TASK_RUNNING.
1305 	 *
1306 	 * This should be ok, provided it doesn't flip the state back to
1307 	 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1308 	 * will only happen if the mutex_lock() call blocks, and we then find
1309 	 * the ringbuffer empty. So in practice we should be ok, but it's
1310 	 * something to be aware of when touching this code.
1311 	 */
1312 	if (until.tv64 == 0)
1313 		aio_read_events(ctx, min_nr, nr, event, &ret);
1314 	else
1315 		wait_event_interruptible_hrtimeout(ctx->wait,
1316 				aio_read_events(ctx, min_nr, nr, event, &ret),
1317 				until);
1318 
1319 	if (!ret && signal_pending(current))
1320 		ret = -EINTR;
1321 
1322 	return ret;
1323 }
1324 
1325 /* sys_io_setup:
1326  *	Create an aio_context capable of receiving at least nr_events.
1327  *	ctxp must not point to an aio_context that already exists, and
1328  *	must be initialized to 0 prior to the call.  On successful
1329  *	creation of the aio_context, *ctxp is filled in with the resulting
1330  *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1331  *	if the specified nr_events exceeds internal limits.  May fail
1332  *	with -EAGAIN if the specified nr_events exceeds the user's limit
1333  *	of available events.  May fail with -ENOMEM if insufficient kernel
1334  *	resources are available.  May fail with -EFAULT if an invalid
1335  *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1336  *	implemented.
1337  */
SYSCALL_DEFINE2(io_setup,unsigned,nr_events,aio_context_t __user *,ctxp)1338 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1339 {
1340 	struct kioctx *ioctx = NULL;
1341 	unsigned long ctx;
1342 	long ret;
1343 
1344 	ret = get_user(ctx, ctxp);
1345 	if (unlikely(ret))
1346 		goto out;
1347 
1348 	ret = -EINVAL;
1349 	if (unlikely(ctx || nr_events == 0)) {
1350 		pr_debug("EINVAL: ctx %lu nr_events %u\n",
1351 		         ctx, nr_events);
1352 		goto out;
1353 	}
1354 
1355 	ioctx = ioctx_alloc(nr_events);
1356 	ret = PTR_ERR(ioctx);
1357 	if (!IS_ERR(ioctx)) {
1358 		ret = put_user(ioctx->user_id, ctxp);
1359 		if (ret)
1360 			kill_ioctx(current->mm, ioctx, NULL);
1361 		percpu_ref_put(&ioctx->users);
1362 	}
1363 
1364 out:
1365 	return ret;
1366 }
1367 
1368 /* sys_io_destroy:
1369  *	Destroy the aio_context specified.  May cancel any outstanding
1370  *	AIOs and block on completion.  Will fail with -ENOSYS if not
1371  *	implemented.  May fail with -EINVAL if the context pointed to
1372  *	is invalid.
1373  */
SYSCALL_DEFINE1(io_destroy,aio_context_t,ctx)1374 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1375 {
1376 	struct kioctx *ioctx = lookup_ioctx(ctx);
1377 	if (likely(NULL != ioctx)) {
1378 		struct ctx_rq_wait wait;
1379 		int ret;
1380 
1381 		init_completion(&wait.comp);
1382 		atomic_set(&wait.count, 1);
1383 
1384 		/* Pass requests_done to kill_ioctx() where it can be set
1385 		 * in a thread-safe way. If we try to set it here then we have
1386 		 * a race condition if two io_destroy() called simultaneously.
1387 		 */
1388 		ret = kill_ioctx(current->mm, ioctx, &wait);
1389 		percpu_ref_put(&ioctx->users);
1390 
1391 		/* Wait until all IO for the context are done. Otherwise kernel
1392 		 * keep using user-space buffers even if user thinks the context
1393 		 * is destroyed.
1394 		 */
1395 		if (!ret)
1396 			wait_for_completion(&wait.comp);
1397 
1398 		return ret;
1399 	}
1400 	pr_debug("EINVAL: invalid context id\n");
1401 	return -EINVAL;
1402 }
1403 
1404 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1405 
aio_setup_vectored_rw(int rw,char __user * buf,size_t len,struct iovec ** iovec,bool compat,struct iov_iter * iter)1406 static int aio_setup_vectored_rw(int rw, char __user *buf, size_t len,
1407 				 struct iovec **iovec,
1408 				 bool compat,
1409 				 struct iov_iter *iter)
1410 {
1411 #ifdef CONFIG_COMPAT
1412 	if (compat)
1413 		return compat_import_iovec(rw,
1414 				(struct compat_iovec __user *)buf,
1415 				len, UIO_FASTIOV, iovec, iter);
1416 #endif
1417 	return import_iovec(rw, (struct iovec __user *)buf,
1418 				len, UIO_FASTIOV, iovec, iter);
1419 }
1420 
1421 /*
1422  * aio_run_iocb:
1423  *	Performs the initial checks and io submission.
1424  */
aio_run_iocb(struct kiocb * req,unsigned opcode,char __user * buf,size_t len,bool compat)1425 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1426 			    char __user *buf, size_t len, bool compat)
1427 {
1428 	struct file *file = req->ki_filp;
1429 	ssize_t ret;
1430 	int rw;
1431 	fmode_t mode;
1432 	rw_iter_op *iter_op;
1433 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1434 	struct iov_iter iter;
1435 
1436 	switch (opcode) {
1437 	case IOCB_CMD_PREAD:
1438 	case IOCB_CMD_PREADV:
1439 		mode	= FMODE_READ;
1440 		rw	= READ;
1441 		iter_op	= file->f_op->read_iter;
1442 		goto rw_common;
1443 
1444 	case IOCB_CMD_PWRITE:
1445 	case IOCB_CMD_PWRITEV:
1446 		mode	= FMODE_WRITE;
1447 		rw	= WRITE;
1448 		iter_op	= file->f_op->write_iter;
1449 		goto rw_common;
1450 rw_common:
1451 		if (unlikely(!(file->f_mode & mode)))
1452 			return -EBADF;
1453 
1454 		if (!iter_op)
1455 			return -EINVAL;
1456 
1457 		if (opcode == IOCB_CMD_PREADV || opcode == IOCB_CMD_PWRITEV)
1458 			ret = aio_setup_vectored_rw(rw, buf, len,
1459 						&iovec, compat, &iter);
1460 		else {
1461 			ret = import_single_range(rw, buf, len, iovec, &iter);
1462 			iovec = NULL;
1463 		}
1464 		if (!ret)
1465 			ret = rw_verify_area(rw, file, &req->ki_pos,
1466 					     iov_iter_count(&iter));
1467 		if (ret < 0) {
1468 			kfree(iovec);
1469 			return ret;
1470 		}
1471 
1472 		len = ret;
1473 
1474 		if (rw == WRITE)
1475 			file_start_write(file);
1476 
1477 		ret = iter_op(req, &iter);
1478 
1479 		if (rw == WRITE)
1480 			file_end_write(file);
1481 		kfree(iovec);
1482 		break;
1483 
1484 	case IOCB_CMD_FDSYNC:
1485 		if (!file->f_op->aio_fsync)
1486 			return -EINVAL;
1487 
1488 		ret = file->f_op->aio_fsync(req, 1);
1489 		break;
1490 
1491 	case IOCB_CMD_FSYNC:
1492 		if (!file->f_op->aio_fsync)
1493 			return -EINVAL;
1494 
1495 		ret = file->f_op->aio_fsync(req, 0);
1496 		break;
1497 
1498 	default:
1499 		pr_debug("EINVAL: no operation provided\n");
1500 		return -EINVAL;
1501 	}
1502 
1503 	if (ret != -EIOCBQUEUED) {
1504 		/*
1505 		 * There's no easy way to restart the syscall since other AIO's
1506 		 * may be already running. Just fail this IO with EINTR.
1507 		 */
1508 		if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1509 			     ret == -ERESTARTNOHAND ||
1510 			     ret == -ERESTART_RESTARTBLOCK))
1511 			ret = -EINTR;
1512 		aio_complete(req, ret, 0);
1513 	}
1514 
1515 	return 0;
1516 }
1517 
io_submit_one(struct kioctx * ctx,struct iocb __user * user_iocb,struct iocb * iocb,bool compat)1518 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1519 			 struct iocb *iocb, bool compat)
1520 {
1521 	struct aio_kiocb *req;
1522 	ssize_t ret;
1523 
1524 	/* enforce forwards compatibility on users */
1525 	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1526 		pr_debug("EINVAL: 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: overflow check\n");
1537 		return -EINVAL;
1538 	}
1539 
1540 	req = aio_get_req(ctx);
1541 	if (unlikely(!req))
1542 		return -EAGAIN;
1543 
1544 	req->common.ki_filp = fget(iocb->aio_fildes);
1545 	if (unlikely(!req->common.ki_filp)) {
1546 		ret = -EBADF;
1547 		goto out_put_req;
1548 	}
1549 	req->common.ki_pos = iocb->aio_offset;
1550 	req->common.ki_complete = aio_complete;
1551 	req->common.ki_flags = iocb_flags(req->common.ki_filp);
1552 
1553 	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1554 		/*
1555 		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1556 		 * instance of the file* now. The file descriptor must be
1557 		 * an eventfd() fd, and will be signaled for each completed
1558 		 * event using the eventfd_signal() function.
1559 		 */
1560 		req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1561 		if (IS_ERR(req->ki_eventfd)) {
1562 			ret = PTR_ERR(req->ki_eventfd);
1563 			req->ki_eventfd = NULL;
1564 			goto out_put_req;
1565 		}
1566 
1567 		req->common.ki_flags |= IOCB_EVENTFD;
1568 	}
1569 
1570 	ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1571 	if (unlikely(ret)) {
1572 		pr_debug("EFAULT: aio_key\n");
1573 		goto out_put_req;
1574 	}
1575 
1576 	req->ki_user_iocb = user_iocb;
1577 	req->ki_user_data = iocb->aio_data;
1578 
1579 	ret = aio_run_iocb(&req->common, iocb->aio_lio_opcode,
1580 			   (char __user *)(unsigned long)iocb->aio_buf,
1581 			   iocb->aio_nbytes,
1582 			   compat);
1583 	if (ret)
1584 		goto out_put_req;
1585 
1586 	return 0;
1587 out_put_req:
1588 	put_reqs_available(ctx, 1);
1589 	percpu_ref_put(&ctx->reqs);
1590 	kiocb_free(req);
1591 	return ret;
1592 }
1593 
do_io_submit(aio_context_t ctx_id,long nr,struct iocb __user * __user * iocbpp,bool compat)1594 long do_io_submit(aio_context_t ctx_id, long nr,
1595 		  struct iocb __user *__user *iocbpp, bool compat)
1596 {
1597 	struct kioctx *ctx;
1598 	long ret = 0;
1599 	int i = 0;
1600 	struct blk_plug plug;
1601 
1602 	if (unlikely(nr < 0))
1603 		return -EINVAL;
1604 
1605 	if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1606 		nr = LONG_MAX/sizeof(*iocbpp);
1607 
1608 	if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1609 		return -EFAULT;
1610 
1611 	ctx = lookup_ioctx(ctx_id);
1612 	if (unlikely(!ctx)) {
1613 		pr_debug("EINVAL: invalid context id\n");
1614 		return -EINVAL;
1615 	}
1616 
1617 	blk_start_plug(&plug);
1618 
1619 	/*
1620 	 * AKPM: should this return a partial result if some of the IOs were
1621 	 * successfully submitted?
1622 	 */
1623 	for (i=0; i<nr; i++) {
1624 		struct iocb __user *user_iocb;
1625 		struct iocb tmp;
1626 
1627 		if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1628 			ret = -EFAULT;
1629 			break;
1630 		}
1631 
1632 		if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1633 			ret = -EFAULT;
1634 			break;
1635 		}
1636 
1637 		ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1638 		if (ret)
1639 			break;
1640 	}
1641 	blk_finish_plug(&plug);
1642 
1643 	percpu_ref_put(&ctx->users);
1644 	return i ? i : ret;
1645 }
1646 
1647 /* sys_io_submit:
1648  *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1649  *	the number of iocbs queued.  May return -EINVAL if the aio_context
1650  *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
1651  *	*iocbpp[0] is not properly initialized, if the operation specified
1652  *	is invalid for the file descriptor in the iocb.  May fail with
1653  *	-EFAULT if any of the data structures point to invalid data.  May
1654  *	fail with -EBADF if the file descriptor specified in the first
1655  *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
1656  *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1657  *	fail with -ENOSYS if not implemented.
1658  */
SYSCALL_DEFINE3(io_submit,aio_context_t,ctx_id,long,nr,struct iocb __user * __user *,iocbpp)1659 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1660 		struct iocb __user * __user *, iocbpp)
1661 {
1662 	return do_io_submit(ctx_id, nr, iocbpp, 0);
1663 }
1664 
1665 /* lookup_kiocb
1666  *	Finds a given iocb for cancellation.
1667  */
1668 static struct aio_kiocb *
lookup_kiocb(struct kioctx * ctx,struct iocb __user * iocb,u32 key)1669 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1670 {
1671 	struct aio_kiocb *kiocb;
1672 
1673 	assert_spin_locked(&ctx->ctx_lock);
1674 
1675 	if (key != KIOCB_KEY)
1676 		return NULL;
1677 
1678 	/* TODO: use a hash or array, this sucks. */
1679 	list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1680 		if (kiocb->ki_user_iocb == iocb)
1681 			return kiocb;
1682 	}
1683 	return NULL;
1684 }
1685 
1686 /* sys_io_cancel:
1687  *	Attempts to cancel an iocb previously passed to io_submit.  If
1688  *	the operation is successfully cancelled, the resulting event is
1689  *	copied into the memory pointed to by result without being placed
1690  *	into the completion queue and 0 is returned.  May fail with
1691  *	-EFAULT if any of the data structures pointed to are invalid.
1692  *	May fail with -EINVAL if aio_context specified by ctx_id is
1693  *	invalid.  May fail with -EAGAIN if the iocb specified was not
1694  *	cancelled.  Will fail with -ENOSYS if not implemented.
1695  */
SYSCALL_DEFINE3(io_cancel,aio_context_t,ctx_id,struct iocb __user *,iocb,struct io_event __user *,result)1696 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1697 		struct io_event __user *, result)
1698 {
1699 	struct kioctx *ctx;
1700 	struct aio_kiocb *kiocb;
1701 	u32 key;
1702 	int ret;
1703 
1704 	ret = get_user(key, &iocb->aio_key);
1705 	if (unlikely(ret))
1706 		return -EFAULT;
1707 
1708 	ctx = lookup_ioctx(ctx_id);
1709 	if (unlikely(!ctx))
1710 		return -EINVAL;
1711 
1712 	spin_lock_irq(&ctx->ctx_lock);
1713 
1714 	kiocb = lookup_kiocb(ctx, iocb, key);
1715 	if (kiocb)
1716 		ret = kiocb_cancel(kiocb);
1717 	else
1718 		ret = -EINVAL;
1719 
1720 	spin_unlock_irq(&ctx->ctx_lock);
1721 
1722 	if (!ret) {
1723 		/*
1724 		 * The result argument is no longer used - the io_event is
1725 		 * always delivered via the ring buffer. -EINPROGRESS indicates
1726 		 * cancellation is progress:
1727 		 */
1728 		ret = -EINPROGRESS;
1729 	}
1730 
1731 	percpu_ref_put(&ctx->users);
1732 
1733 	return ret;
1734 }
1735 
1736 /* io_getevents:
1737  *	Attempts to read at least min_nr events and up to nr events from
1738  *	the completion queue for the aio_context specified by ctx_id. If
1739  *	it succeeds, the number of read events is returned. May fail with
1740  *	-EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1741  *	out of range, if timeout is out of range.  May fail with -EFAULT
1742  *	if any of the memory specified is invalid.  May return 0 or
1743  *	< min_nr if the timeout specified by timeout has elapsed
1744  *	before sufficient events are available, where timeout == NULL
1745  *	specifies an infinite timeout. Note that the timeout pointed to by
1746  *	timeout is relative.  Will fail with -ENOSYS if not implemented.
1747  */
SYSCALL_DEFINE5(io_getevents,aio_context_t,ctx_id,long,min_nr,long,nr,struct io_event __user *,events,struct timespec __user *,timeout)1748 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1749 		long, min_nr,
1750 		long, nr,
1751 		struct io_event __user *, events,
1752 		struct timespec __user *, timeout)
1753 {
1754 	struct kioctx *ioctx = lookup_ioctx(ctx_id);
1755 	long ret = -EINVAL;
1756 
1757 	if (likely(ioctx)) {
1758 		if (likely(min_nr <= nr && min_nr >= 0))
1759 			ret = read_events(ioctx, min_nr, nr, events, timeout);
1760 		percpu_ref_put(&ioctx->users);
1761 	}
1762 	return ret;
1763 }
1764