1 /*
2  *	An async IO implementation for Linux
3  *	Written by Benjamin LaHaise <bcrl@kvack.org>
4  *
5  *	Implements an efficient asynchronous io interface.
6  *
7  *	Copyright 2000, 2001, 2002 Red Hat, Inc.  All Rights Reserved.
8  *	Copyright 2018 Christoph Hellwig.
9  *
10  *	See ../COPYING for licensing terms.
11  */
12 #define pr_fmt(fmt) "%s: " fmt, __func__
13 
14 #include <linux/kernel.h>
15 #include <linux/init.h>
16 #include <linux/errno.h>
17 #include <linux/time.h>
18 #include <linux/aio_abi.h>
19 #include <linux/export.h>
20 #include <linux/syscalls.h>
21 #include <linux/backing-dev.h>
22 #include <linux/refcount.h>
23 #include <linux/uio.h>
24 
25 #include <linux/sched/signal.h>
26 #include <linux/fs.h>
27 #include <linux/file.h>
28 #include <linux/mm.h>
29 #include <linux/mman.h>
30 #include <linux/percpu.h>
31 #include <linux/slab.h>
32 #include <linux/timer.h>
33 #include <linux/aio.h>
34 #include <linux/highmem.h>
35 #include <linux/workqueue.h>
36 #include <linux/security.h>
37 #include <linux/eventfd.h>
38 #include <linux/blkdev.h>
39 #include <linux/compat.h>
40 #include <linux/migrate.h>
41 #include <linux/ramfs.h>
42 #include <linux/percpu-refcount.h>
43 #include <linux/mount.h>
44 #include <linux/pseudo_fs.h>
45 
46 #include <linux/uaccess.h>
47 #include <linux/nospec.h>
48 
49 #include "internal.h"
50 
51 #define KIOCB_KEY		0
52 
53 #define AIO_RING_MAGIC			0xa10a10a1
54 #define AIO_RING_COMPAT_FEATURES	1
55 #define AIO_RING_INCOMPAT_FEATURES	0
56 struct aio_ring {
57 	unsigned	id;	/* kernel internal index number */
58 	unsigned	nr;	/* number of io_events */
59 	unsigned	head;	/* Written to by userland or under ring_lock
60 				 * mutex by aio_read_events_ring(). */
61 	unsigned	tail;
62 
63 	unsigned	magic;
64 	unsigned	compat_features;
65 	unsigned	incompat_features;
66 	unsigned	header_length;	/* size of aio_ring */
67 
68 
69 	struct io_event		io_events[];
70 }; /* 128 bytes + ring size */
71 
72 /*
73  * Plugging is meant to work with larger batches of IOs. If we don't
74  * have more than the below, then don't bother setting up a plug.
75  */
76 #define AIO_PLUG_THRESHOLD	2
77 
78 #define AIO_RING_PAGES	8
79 
80 struct kioctx_table {
81 	struct rcu_head		rcu;
82 	unsigned		nr;
83 	struct kioctx __rcu	*table[] __counted_by(nr);
84 };
85 
86 struct kioctx_cpu {
87 	unsigned		reqs_available;
88 };
89 
90 struct ctx_rq_wait {
91 	struct completion comp;
92 	atomic_t count;
93 };
94 
95 struct kioctx {
96 	struct percpu_ref	users;
97 	atomic_t		dead;
98 
99 	struct percpu_ref	reqs;
100 
101 	unsigned long		user_id;
102 
103 	struct __percpu kioctx_cpu *cpu;
104 
105 	/*
106 	 * For percpu reqs_available, number of slots we move to/from global
107 	 * counter at a time:
108 	 */
109 	unsigned		req_batch;
110 	/*
111 	 * This is what userspace passed to io_setup(), it's not used for
112 	 * anything but counting against the global max_reqs quota.
113 	 *
114 	 * The real limit is nr_events - 1, which will be larger (see
115 	 * aio_setup_ring())
116 	 */
117 	unsigned		max_reqs;
118 
119 	/* Size of ringbuffer, in units of struct io_event */
120 	unsigned		nr_events;
121 
122 	unsigned long		mmap_base;
123 	unsigned long		mmap_size;
124 
125 	struct page		**ring_pages;
126 	long			nr_pages;
127 
128 	struct rcu_work		free_rwork;	/* see free_ioctx() */
129 
130 	/*
131 	 * signals when all in-flight requests are done
132 	 */
133 	struct ctx_rq_wait	*rq_wait;
134 
135 	struct {
136 		/*
137 		 * This counts the number of available slots in the ringbuffer,
138 		 * so we avoid overflowing it: it's decremented (if positive)
139 		 * when allocating a kiocb and incremented when the resulting
140 		 * io_event is pulled off the ringbuffer.
141 		 *
142 		 * We batch accesses to it with a percpu version.
143 		 */
144 		atomic_t	reqs_available;
145 	} ____cacheline_aligned_in_smp;
146 
147 	struct {
148 		spinlock_t	ctx_lock;
149 		struct list_head active_reqs;	/* used for cancellation */
150 	} ____cacheline_aligned_in_smp;
151 
152 	struct {
153 		struct mutex	ring_lock;
154 		wait_queue_head_t wait;
155 	} ____cacheline_aligned_in_smp;
156 
157 	struct {
158 		unsigned	tail;
159 		unsigned	completed_events;
160 		spinlock_t	completion_lock;
161 	} ____cacheline_aligned_in_smp;
162 
163 	struct page		*internal_pages[AIO_RING_PAGES];
164 	struct file		*aio_ring_file;
165 
166 	unsigned		id;
167 };
168 
169 /*
170  * First field must be the file pointer in all the
171  * iocb unions! See also 'struct kiocb' in <linux/fs.h>
172  */
173 struct fsync_iocb {
174 	struct file		*file;
175 	struct work_struct	work;
176 	bool			datasync;
177 	struct cred		*creds;
178 };
179 
180 struct poll_iocb {
181 	struct file		*file;
182 	struct wait_queue_head	*head;
183 	__poll_t		events;
184 	bool			cancelled;
185 	bool			work_scheduled;
186 	bool			work_need_resched;
187 	struct wait_queue_entry	wait;
188 	struct work_struct	work;
189 };
190 
191 /*
192  * NOTE! Each of the iocb union members has the file pointer
193  * as the first entry in their struct definition. So you can
194  * access the file pointer through any of the sub-structs,
195  * or directly as just 'ki_filp' in this struct.
196  */
197 struct aio_kiocb {
198 	union {
199 		struct file		*ki_filp;
200 		struct kiocb		rw;
201 		struct fsync_iocb	fsync;
202 		struct poll_iocb	poll;
203 	};
204 
205 	struct kioctx		*ki_ctx;
206 	kiocb_cancel_fn		*ki_cancel;
207 
208 	struct io_event		ki_res;
209 
210 	struct list_head	ki_list;	/* the aio core uses this
211 						 * for cancellation */
212 	refcount_t		ki_refcnt;
213 
214 	/*
215 	 * If the aio_resfd field of the userspace iocb is not zero,
216 	 * this is the underlying eventfd context to deliver events to.
217 	 */
218 	struct eventfd_ctx	*ki_eventfd;
219 };
220 
221 /*------ sysctl variables----*/
222 static DEFINE_SPINLOCK(aio_nr_lock);
223 static unsigned long aio_nr;		/* current system wide number of aio requests */
224 static unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
225 /*----end sysctl variables---*/
226 #ifdef CONFIG_SYSCTL
227 static struct ctl_table aio_sysctls[] = {
228 	{
229 		.procname	= "aio-nr",
230 		.data		= &aio_nr,
231 		.maxlen		= sizeof(aio_nr),
232 		.mode		= 0444,
233 		.proc_handler	= proc_doulongvec_minmax,
234 	},
235 	{
236 		.procname	= "aio-max-nr",
237 		.data		= &aio_max_nr,
238 		.maxlen		= sizeof(aio_max_nr),
239 		.mode		= 0644,
240 		.proc_handler	= proc_doulongvec_minmax,
241 	},
242 	{}
243 };
244 
aio_sysctl_init(void)245 static void __init aio_sysctl_init(void)
246 {
247 	register_sysctl_init("fs", aio_sysctls);
248 }
249 #else
250 #define aio_sysctl_init() do { } while (0)
251 #endif
252 
253 static struct kmem_cache	*kiocb_cachep;
254 static struct kmem_cache	*kioctx_cachep;
255 
256 static struct vfsmount *aio_mnt;
257 
258 static const struct file_operations aio_ring_fops;
259 static const struct address_space_operations aio_ctx_aops;
260 
aio_private_file(struct kioctx * ctx,loff_t nr_pages)261 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
262 {
263 	struct file *file;
264 	struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
265 	if (IS_ERR(inode))
266 		return ERR_CAST(inode);
267 
268 	inode->i_mapping->a_ops = &aio_ctx_aops;
269 	inode->i_mapping->private_data = ctx;
270 	inode->i_size = PAGE_SIZE * nr_pages;
271 
272 	file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
273 				O_RDWR, &aio_ring_fops);
274 	if (IS_ERR(file))
275 		iput(inode);
276 	return file;
277 }
278 
aio_init_fs_context(struct fs_context * fc)279 static int aio_init_fs_context(struct fs_context *fc)
280 {
281 	if (!init_pseudo(fc, AIO_RING_MAGIC))
282 		return -ENOMEM;
283 	fc->s_iflags |= SB_I_NOEXEC;
284 	return 0;
285 }
286 
287 /* aio_setup
288  *	Creates the slab caches used by the aio routines, panic on
289  *	failure as this is done early during the boot sequence.
290  */
aio_setup(void)291 static int __init aio_setup(void)
292 {
293 	static struct file_system_type aio_fs = {
294 		.name		= "aio",
295 		.init_fs_context = aio_init_fs_context,
296 		.kill_sb	= kill_anon_super,
297 	};
298 	aio_mnt = kern_mount(&aio_fs);
299 	if (IS_ERR(aio_mnt))
300 		panic("Failed to create aio fs mount.");
301 
302 	kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
303 	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
304 	aio_sysctl_init();
305 	return 0;
306 }
307 __initcall(aio_setup);
308 
put_aio_ring_file(struct kioctx * ctx)309 static void put_aio_ring_file(struct kioctx *ctx)
310 {
311 	struct file *aio_ring_file = ctx->aio_ring_file;
312 	struct address_space *i_mapping;
313 
314 	if (aio_ring_file) {
315 		truncate_setsize(file_inode(aio_ring_file), 0);
316 
317 		/* Prevent further access to the kioctx from migratepages */
318 		i_mapping = aio_ring_file->f_mapping;
319 		spin_lock(&i_mapping->private_lock);
320 		i_mapping->private_data = NULL;
321 		ctx->aio_ring_file = NULL;
322 		spin_unlock(&i_mapping->private_lock);
323 
324 		fput(aio_ring_file);
325 	}
326 }
327 
aio_free_ring(struct kioctx * ctx)328 static void aio_free_ring(struct kioctx *ctx)
329 {
330 	int i;
331 
332 	/* Disconnect the kiotx from the ring file.  This prevents future
333 	 * accesses to the kioctx from page migration.
334 	 */
335 	put_aio_ring_file(ctx);
336 
337 	for (i = 0; i < ctx->nr_pages; i++) {
338 		struct page *page;
339 		pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
340 				page_count(ctx->ring_pages[i]));
341 		page = ctx->ring_pages[i];
342 		if (!page)
343 			continue;
344 		ctx->ring_pages[i] = NULL;
345 		put_page(page);
346 	}
347 
348 	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
349 		kfree(ctx->ring_pages);
350 		ctx->ring_pages = NULL;
351 	}
352 }
353 
aio_ring_mremap(struct vm_area_struct * vma)354 static int aio_ring_mremap(struct vm_area_struct *vma)
355 {
356 	struct file *file = vma->vm_file;
357 	struct mm_struct *mm = vma->vm_mm;
358 	struct kioctx_table *table;
359 	int i, res = -EINVAL;
360 
361 	spin_lock(&mm->ioctx_lock);
362 	rcu_read_lock();
363 	table = rcu_dereference(mm->ioctx_table);
364 	if (!table)
365 		goto out_unlock;
366 
367 	for (i = 0; i < table->nr; i++) {
368 		struct kioctx *ctx;
369 
370 		ctx = rcu_dereference(table->table[i]);
371 		if (ctx && ctx->aio_ring_file == file) {
372 			if (!atomic_read(&ctx->dead)) {
373 				ctx->user_id = ctx->mmap_base = vma->vm_start;
374 				res = 0;
375 			}
376 			break;
377 		}
378 	}
379 
380 out_unlock:
381 	rcu_read_unlock();
382 	spin_unlock(&mm->ioctx_lock);
383 	return res;
384 }
385 
386 static const struct vm_operations_struct aio_ring_vm_ops = {
387 	.mremap		= aio_ring_mremap,
388 #if IS_ENABLED(CONFIG_MMU)
389 	.fault		= filemap_fault,
390 	.map_pages	= filemap_map_pages,
391 	.page_mkwrite	= filemap_page_mkwrite,
392 #endif
393 };
394 
aio_ring_mmap(struct file * file,struct vm_area_struct * vma)395 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
396 {
397 	vm_flags_set(vma, VM_DONTEXPAND);
398 	vma->vm_ops = &aio_ring_vm_ops;
399 	return 0;
400 }
401 
402 static const struct file_operations aio_ring_fops = {
403 	.mmap = aio_ring_mmap,
404 };
405 
406 #if IS_ENABLED(CONFIG_MIGRATION)
aio_migrate_folio(struct address_space * mapping,struct folio * dst,struct folio * src,enum migrate_mode mode)407 static int aio_migrate_folio(struct address_space *mapping, struct folio *dst,
408 			struct folio *src, enum migrate_mode mode)
409 {
410 	struct kioctx *ctx;
411 	unsigned long flags;
412 	pgoff_t idx;
413 	int rc;
414 
415 	/*
416 	 * We cannot support the _NO_COPY case here, because copy needs to
417 	 * happen under the ctx->completion_lock. That does not work with the
418 	 * migration workflow of MIGRATE_SYNC_NO_COPY.
419 	 */
420 	if (mode == MIGRATE_SYNC_NO_COPY)
421 		return -EINVAL;
422 
423 	rc = 0;
424 
425 	/* mapping->private_lock here protects against the kioctx teardown.  */
426 	spin_lock(&mapping->private_lock);
427 	ctx = mapping->private_data;
428 	if (!ctx) {
429 		rc = -EINVAL;
430 		goto out;
431 	}
432 
433 	/* The ring_lock mutex.  The prevents aio_read_events() from writing
434 	 * to the ring's head, and prevents page migration from mucking in
435 	 * a partially initialized kiotx.
436 	 */
437 	if (!mutex_trylock(&ctx->ring_lock)) {
438 		rc = -EAGAIN;
439 		goto out;
440 	}
441 
442 	idx = src->index;
443 	if (idx < (pgoff_t)ctx->nr_pages) {
444 		/* Make sure the old folio hasn't already been changed */
445 		if (ctx->ring_pages[idx] != &src->page)
446 			rc = -EAGAIN;
447 	} else
448 		rc = -EINVAL;
449 
450 	if (rc != 0)
451 		goto out_unlock;
452 
453 	/* Writeback must be complete */
454 	BUG_ON(folio_test_writeback(src));
455 	folio_get(dst);
456 
457 	rc = folio_migrate_mapping(mapping, dst, src, 1);
458 	if (rc != MIGRATEPAGE_SUCCESS) {
459 		folio_put(dst);
460 		goto out_unlock;
461 	}
462 
463 	/* Take completion_lock to prevent other writes to the ring buffer
464 	 * while the old folio is copied to the new.  This prevents new
465 	 * events from being lost.
466 	 */
467 	spin_lock_irqsave(&ctx->completion_lock, flags);
468 	folio_migrate_copy(dst, src);
469 	BUG_ON(ctx->ring_pages[idx] != &src->page);
470 	ctx->ring_pages[idx] = &dst->page;
471 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
472 
473 	/* The old folio is no longer accessible. */
474 	folio_put(src);
475 
476 out_unlock:
477 	mutex_unlock(&ctx->ring_lock);
478 out:
479 	spin_unlock(&mapping->private_lock);
480 	return rc;
481 }
482 #else
483 #define aio_migrate_folio NULL
484 #endif
485 
486 static const struct address_space_operations aio_ctx_aops = {
487 	.dirty_folio	= noop_dirty_folio,
488 	.migrate_folio	= aio_migrate_folio,
489 };
490 
aio_setup_ring(struct kioctx * ctx,unsigned int nr_events)491 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
492 {
493 	struct aio_ring *ring;
494 	struct mm_struct *mm = current->mm;
495 	unsigned long size, unused;
496 	int nr_pages;
497 	int i;
498 	struct file *file;
499 
500 	/* Compensate for the ring buffer's head/tail overlap entry */
501 	nr_events += 2;	/* 1 is required, 2 for good luck */
502 
503 	size = sizeof(struct aio_ring);
504 	size += sizeof(struct io_event) * nr_events;
505 
506 	nr_pages = PFN_UP(size);
507 	if (nr_pages < 0)
508 		return -EINVAL;
509 
510 	file = aio_private_file(ctx, nr_pages);
511 	if (IS_ERR(file)) {
512 		ctx->aio_ring_file = NULL;
513 		return -ENOMEM;
514 	}
515 
516 	ctx->aio_ring_file = file;
517 	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
518 			/ sizeof(struct io_event);
519 
520 	ctx->ring_pages = ctx->internal_pages;
521 	if (nr_pages > AIO_RING_PAGES) {
522 		ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
523 					  GFP_KERNEL);
524 		if (!ctx->ring_pages) {
525 			put_aio_ring_file(ctx);
526 			return -ENOMEM;
527 		}
528 	}
529 
530 	for (i = 0; i < nr_pages; i++) {
531 		struct page *page;
532 		page = find_or_create_page(file->f_mapping,
533 					   i, GFP_USER | __GFP_ZERO);
534 		if (!page)
535 			break;
536 		pr_debug("pid(%d) page[%d]->count=%d\n",
537 			 current->pid, i, page_count(page));
538 		SetPageUptodate(page);
539 		unlock_page(page);
540 
541 		ctx->ring_pages[i] = page;
542 	}
543 	ctx->nr_pages = i;
544 
545 	if (unlikely(i != nr_pages)) {
546 		aio_free_ring(ctx);
547 		return -ENOMEM;
548 	}
549 
550 	ctx->mmap_size = nr_pages * PAGE_SIZE;
551 	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
552 
553 	if (mmap_write_lock_killable(mm)) {
554 		ctx->mmap_size = 0;
555 		aio_free_ring(ctx);
556 		return -EINTR;
557 	}
558 
559 	ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
560 				 PROT_READ | PROT_WRITE,
561 				 MAP_SHARED, 0, 0, &unused, NULL);
562 	mmap_write_unlock(mm);
563 	if (IS_ERR((void *)ctx->mmap_base)) {
564 		ctx->mmap_size = 0;
565 		aio_free_ring(ctx);
566 		return -ENOMEM;
567 	}
568 
569 	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
570 
571 	ctx->user_id = ctx->mmap_base;
572 	ctx->nr_events = nr_events; /* trusted copy */
573 
574 	ring = page_address(ctx->ring_pages[0]);
575 	ring->nr = nr_events;	/* user copy */
576 	ring->id = ~0U;
577 	ring->head = ring->tail = 0;
578 	ring->magic = AIO_RING_MAGIC;
579 	ring->compat_features = AIO_RING_COMPAT_FEATURES;
580 	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
581 	ring->header_length = sizeof(struct aio_ring);
582 	flush_dcache_page(ctx->ring_pages[0]);
583 
584 	return 0;
585 }
586 
587 #define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
588 #define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
589 #define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
590 
kiocb_set_cancel_fn(struct kiocb * iocb,kiocb_cancel_fn * cancel)591 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
592 {
593 	struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
594 	struct kioctx *ctx = req->ki_ctx;
595 	unsigned long flags;
596 
597 	/*
598 	 * kiocb didn't come from aio or is neither a read nor a write, hence
599 	 * ignore it.
600 	 */
601 	if (!(iocb->ki_flags & IOCB_AIO_RW))
602 		return;
603 
604 	if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
605 		return;
606 
607 	spin_lock_irqsave(&ctx->ctx_lock, flags);
608 	list_add_tail(&req->ki_list, &ctx->active_reqs);
609 	req->ki_cancel = cancel;
610 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
611 }
612 EXPORT_SYMBOL(kiocb_set_cancel_fn);
613 
614 /*
615  * free_ioctx() should be RCU delayed to synchronize against the RCU
616  * protected lookup_ioctx() and also needs process context to call
617  * aio_free_ring().  Use rcu_work.
618  */
free_ioctx(struct work_struct * work)619 static void free_ioctx(struct work_struct *work)
620 {
621 	struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
622 					  free_rwork);
623 	pr_debug("freeing %p\n", ctx);
624 
625 	aio_free_ring(ctx);
626 	free_percpu(ctx->cpu);
627 	percpu_ref_exit(&ctx->reqs);
628 	percpu_ref_exit(&ctx->users);
629 	kmem_cache_free(kioctx_cachep, ctx);
630 }
631 
free_ioctx_reqs(struct percpu_ref * ref)632 static void free_ioctx_reqs(struct percpu_ref *ref)
633 {
634 	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
635 
636 	/* At this point we know that there are no any in-flight requests */
637 	if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
638 		complete(&ctx->rq_wait->comp);
639 
640 	/* Synchronize against RCU protected table->table[] dereferences */
641 	INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
642 	queue_rcu_work(system_wq, &ctx->free_rwork);
643 }
644 
645 /*
646  * When this function runs, the kioctx has been removed from the "hash table"
647  * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
648  * now it's safe to cancel any that need to be.
649  */
free_ioctx_users(struct percpu_ref * ref)650 static void free_ioctx_users(struct percpu_ref *ref)
651 {
652 	struct kioctx *ctx = container_of(ref, struct kioctx, users);
653 	struct aio_kiocb *req;
654 
655 	spin_lock_irq(&ctx->ctx_lock);
656 
657 	while (!list_empty(&ctx->active_reqs)) {
658 		req = list_first_entry(&ctx->active_reqs,
659 				       struct aio_kiocb, ki_list);
660 		req->ki_cancel(&req->rw);
661 		list_del_init(&req->ki_list);
662 	}
663 
664 	spin_unlock_irq(&ctx->ctx_lock);
665 
666 	percpu_ref_kill(&ctx->reqs);
667 	percpu_ref_put(&ctx->reqs);
668 }
669 
ioctx_add_table(struct kioctx * ctx,struct mm_struct * mm)670 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
671 {
672 	unsigned i, new_nr;
673 	struct kioctx_table *table, *old;
674 	struct aio_ring *ring;
675 
676 	spin_lock(&mm->ioctx_lock);
677 	table = rcu_dereference_raw(mm->ioctx_table);
678 
679 	while (1) {
680 		if (table)
681 			for (i = 0; i < table->nr; i++)
682 				if (!rcu_access_pointer(table->table[i])) {
683 					ctx->id = i;
684 					rcu_assign_pointer(table->table[i], ctx);
685 					spin_unlock(&mm->ioctx_lock);
686 
687 					/* While kioctx setup is in progress,
688 					 * we are protected from page migration
689 					 * changes ring_pages by ->ring_lock.
690 					 */
691 					ring = page_address(ctx->ring_pages[0]);
692 					ring->id = ctx->id;
693 					return 0;
694 				}
695 
696 		new_nr = (table ? table->nr : 1) * 4;
697 		spin_unlock(&mm->ioctx_lock);
698 
699 		table = kzalloc(struct_size(table, table, new_nr), GFP_KERNEL);
700 		if (!table)
701 			return -ENOMEM;
702 
703 		table->nr = new_nr;
704 
705 		spin_lock(&mm->ioctx_lock);
706 		old = rcu_dereference_raw(mm->ioctx_table);
707 
708 		if (!old) {
709 			rcu_assign_pointer(mm->ioctx_table, table);
710 		} else if (table->nr > old->nr) {
711 			memcpy(table->table, old->table,
712 			       old->nr * sizeof(struct kioctx *));
713 
714 			rcu_assign_pointer(mm->ioctx_table, table);
715 			kfree_rcu(old, rcu);
716 		} else {
717 			kfree(table);
718 			table = old;
719 		}
720 	}
721 }
722 
aio_nr_sub(unsigned nr)723 static void aio_nr_sub(unsigned nr)
724 {
725 	spin_lock(&aio_nr_lock);
726 	if (WARN_ON(aio_nr - nr > aio_nr))
727 		aio_nr = 0;
728 	else
729 		aio_nr -= nr;
730 	spin_unlock(&aio_nr_lock);
731 }
732 
733 /* ioctx_alloc
734  *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
735  */
ioctx_alloc(unsigned nr_events)736 static struct kioctx *ioctx_alloc(unsigned nr_events)
737 {
738 	struct mm_struct *mm = current->mm;
739 	struct kioctx *ctx;
740 	int err = -ENOMEM;
741 
742 	/*
743 	 * Store the original nr_events -- what userspace passed to io_setup(),
744 	 * for counting against the global limit -- before it changes.
745 	 */
746 	unsigned int max_reqs = nr_events;
747 
748 	/*
749 	 * We keep track of the number of available ringbuffer slots, to prevent
750 	 * overflow (reqs_available), and we also use percpu counters for this.
751 	 *
752 	 * So since up to half the slots might be on other cpu's percpu counters
753 	 * and unavailable, double nr_events so userspace sees what they
754 	 * expected: additionally, we move req_batch slots to/from percpu
755 	 * counters at a time, so make sure that isn't 0:
756 	 */
757 	nr_events = max(nr_events, num_possible_cpus() * 4);
758 	nr_events *= 2;
759 
760 	/* Prevent overflows */
761 	if (nr_events > (0x10000000U / sizeof(struct io_event))) {
762 		pr_debug("ENOMEM: nr_events too high\n");
763 		return ERR_PTR(-EINVAL);
764 	}
765 
766 	if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
767 		return ERR_PTR(-EAGAIN);
768 
769 	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
770 	if (!ctx)
771 		return ERR_PTR(-ENOMEM);
772 
773 	ctx->max_reqs = max_reqs;
774 
775 	spin_lock_init(&ctx->ctx_lock);
776 	spin_lock_init(&ctx->completion_lock);
777 	mutex_init(&ctx->ring_lock);
778 	/* Protect against page migration throughout kiotx setup by keeping
779 	 * the ring_lock mutex held until setup is complete. */
780 	mutex_lock(&ctx->ring_lock);
781 	init_waitqueue_head(&ctx->wait);
782 
783 	INIT_LIST_HEAD(&ctx->active_reqs);
784 
785 	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
786 		goto err;
787 
788 	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
789 		goto err;
790 
791 	ctx->cpu = alloc_percpu(struct kioctx_cpu);
792 	if (!ctx->cpu)
793 		goto err;
794 
795 	err = aio_setup_ring(ctx, nr_events);
796 	if (err < 0)
797 		goto err;
798 
799 	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
800 	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
801 	if (ctx->req_batch < 1)
802 		ctx->req_batch = 1;
803 
804 	/* limit the number of system wide aios */
805 	spin_lock(&aio_nr_lock);
806 	if (aio_nr + ctx->max_reqs > aio_max_nr ||
807 	    aio_nr + ctx->max_reqs < aio_nr) {
808 		spin_unlock(&aio_nr_lock);
809 		err = -EAGAIN;
810 		goto err_ctx;
811 	}
812 	aio_nr += ctx->max_reqs;
813 	spin_unlock(&aio_nr_lock);
814 
815 	percpu_ref_get(&ctx->users);	/* io_setup() will drop this ref */
816 	percpu_ref_get(&ctx->reqs);	/* free_ioctx_users() will drop this */
817 
818 	err = ioctx_add_table(ctx, mm);
819 	if (err)
820 		goto err_cleanup;
821 
822 	/* Release the ring_lock mutex now that all setup is complete. */
823 	mutex_unlock(&ctx->ring_lock);
824 
825 	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
826 		 ctx, ctx->user_id, mm, ctx->nr_events);
827 	return ctx;
828 
829 err_cleanup:
830 	aio_nr_sub(ctx->max_reqs);
831 err_ctx:
832 	atomic_set(&ctx->dead, 1);
833 	if (ctx->mmap_size)
834 		vm_munmap(ctx->mmap_base, ctx->mmap_size);
835 	aio_free_ring(ctx);
836 err:
837 	mutex_unlock(&ctx->ring_lock);
838 	free_percpu(ctx->cpu);
839 	percpu_ref_exit(&ctx->reqs);
840 	percpu_ref_exit(&ctx->users);
841 	kmem_cache_free(kioctx_cachep, ctx);
842 	pr_debug("error allocating ioctx %d\n", err);
843 	return ERR_PTR(err);
844 }
845 
846 /* kill_ioctx
847  *	Cancels all outstanding aio requests on an aio context.  Used
848  *	when the processes owning a context have all exited to encourage
849  *	the rapid destruction of the kioctx.
850  */
kill_ioctx(struct mm_struct * mm,struct kioctx * ctx,struct ctx_rq_wait * wait)851 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
852 		      struct ctx_rq_wait *wait)
853 {
854 	struct kioctx_table *table;
855 
856 	spin_lock(&mm->ioctx_lock);
857 	if (atomic_xchg(&ctx->dead, 1)) {
858 		spin_unlock(&mm->ioctx_lock);
859 		return -EINVAL;
860 	}
861 
862 	table = rcu_dereference_raw(mm->ioctx_table);
863 	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
864 	RCU_INIT_POINTER(table->table[ctx->id], NULL);
865 	spin_unlock(&mm->ioctx_lock);
866 
867 	/* free_ioctx_reqs() will do the necessary RCU synchronization */
868 	wake_up_all(&ctx->wait);
869 
870 	/*
871 	 * It'd be more correct to do this in free_ioctx(), after all
872 	 * the outstanding kiocbs have finished - but by then io_destroy
873 	 * has already returned, so io_setup() could potentially return
874 	 * -EAGAIN with no ioctxs actually in use (as far as userspace
875 	 *  could tell).
876 	 */
877 	aio_nr_sub(ctx->max_reqs);
878 
879 	if (ctx->mmap_size)
880 		vm_munmap(ctx->mmap_base, ctx->mmap_size);
881 
882 	ctx->rq_wait = wait;
883 	percpu_ref_kill(&ctx->users);
884 	return 0;
885 }
886 
887 /*
888  * exit_aio: called when the last user of mm goes away.  At this point, there is
889  * no way for any new requests to be submited or any of the io_* syscalls to be
890  * called on the context.
891  *
892  * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
893  * them.
894  */
exit_aio(struct mm_struct * mm)895 void exit_aio(struct mm_struct *mm)
896 {
897 	struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
898 	struct ctx_rq_wait wait;
899 	int i, skipped;
900 
901 	if (!table)
902 		return;
903 
904 	atomic_set(&wait.count, table->nr);
905 	init_completion(&wait.comp);
906 
907 	skipped = 0;
908 	for (i = 0; i < table->nr; ++i) {
909 		struct kioctx *ctx =
910 			rcu_dereference_protected(table->table[i], true);
911 
912 		if (!ctx) {
913 			skipped++;
914 			continue;
915 		}
916 
917 		/*
918 		 * We don't need to bother with munmap() here - exit_mmap(mm)
919 		 * is coming and it'll unmap everything. And we simply can't,
920 		 * this is not necessarily our ->mm.
921 		 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
922 		 * that it needs to unmap the area, just set it to 0.
923 		 */
924 		ctx->mmap_size = 0;
925 		kill_ioctx(mm, ctx, &wait);
926 	}
927 
928 	if (!atomic_sub_and_test(skipped, &wait.count)) {
929 		/* Wait until all IO for the context are done. */
930 		wait_for_completion(&wait.comp);
931 	}
932 
933 	RCU_INIT_POINTER(mm->ioctx_table, NULL);
934 	kfree(table);
935 }
936 
put_reqs_available(struct kioctx * ctx,unsigned nr)937 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
938 {
939 	struct kioctx_cpu *kcpu;
940 	unsigned long flags;
941 
942 	local_irq_save(flags);
943 	kcpu = this_cpu_ptr(ctx->cpu);
944 	kcpu->reqs_available += nr;
945 
946 	while (kcpu->reqs_available >= ctx->req_batch * 2) {
947 		kcpu->reqs_available -= ctx->req_batch;
948 		atomic_add(ctx->req_batch, &ctx->reqs_available);
949 	}
950 
951 	local_irq_restore(flags);
952 }
953 
__get_reqs_available(struct kioctx * ctx)954 static bool __get_reqs_available(struct kioctx *ctx)
955 {
956 	struct kioctx_cpu *kcpu;
957 	bool ret = false;
958 	unsigned long flags;
959 
960 	local_irq_save(flags);
961 	kcpu = this_cpu_ptr(ctx->cpu);
962 	if (!kcpu->reqs_available) {
963 		int avail = atomic_read(&ctx->reqs_available);
964 
965 		do {
966 			if (avail < ctx->req_batch)
967 				goto out;
968 		} while (!atomic_try_cmpxchg(&ctx->reqs_available,
969 					     &avail, avail - ctx->req_batch));
970 
971 		kcpu->reqs_available += ctx->req_batch;
972 	}
973 
974 	ret = true;
975 	kcpu->reqs_available--;
976 out:
977 	local_irq_restore(flags);
978 	return ret;
979 }
980 
981 /* refill_reqs_available
982  *	Updates the reqs_available reference counts used for tracking the
983  *	number of free slots in the completion ring.  This can be called
984  *	from aio_complete() (to optimistically update reqs_available) or
985  *	from aio_get_req() (the we're out of events case).  It must be
986  *	called holding ctx->completion_lock.
987  */
refill_reqs_available(struct kioctx * ctx,unsigned head,unsigned tail)988 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
989                                   unsigned tail)
990 {
991 	unsigned events_in_ring, completed;
992 
993 	/* Clamp head since userland can write to it. */
994 	head %= ctx->nr_events;
995 	if (head <= tail)
996 		events_in_ring = tail - head;
997 	else
998 		events_in_ring = ctx->nr_events - (head - tail);
999 
1000 	completed = ctx->completed_events;
1001 	if (events_in_ring < completed)
1002 		completed -= events_in_ring;
1003 	else
1004 		completed = 0;
1005 
1006 	if (!completed)
1007 		return;
1008 
1009 	ctx->completed_events -= completed;
1010 	put_reqs_available(ctx, completed);
1011 }
1012 
1013 /* user_refill_reqs_available
1014  *	Called to refill reqs_available when aio_get_req() encounters an
1015  *	out of space in the completion ring.
1016  */
user_refill_reqs_available(struct kioctx * ctx)1017 static void user_refill_reqs_available(struct kioctx *ctx)
1018 {
1019 	spin_lock_irq(&ctx->completion_lock);
1020 	if (ctx->completed_events) {
1021 		struct aio_ring *ring;
1022 		unsigned head;
1023 
1024 		/* Access of ring->head may race with aio_read_events_ring()
1025 		 * here, but that's okay since whether we read the old version
1026 		 * or the new version, and either will be valid.  The important
1027 		 * part is that head cannot pass tail since we prevent
1028 		 * aio_complete() from updating tail by holding
1029 		 * ctx->completion_lock.  Even if head is invalid, the check
1030 		 * against ctx->completed_events below will make sure we do the
1031 		 * safe/right thing.
1032 		 */
1033 		ring = page_address(ctx->ring_pages[0]);
1034 		head = ring->head;
1035 
1036 		refill_reqs_available(ctx, head, ctx->tail);
1037 	}
1038 
1039 	spin_unlock_irq(&ctx->completion_lock);
1040 }
1041 
get_reqs_available(struct kioctx * ctx)1042 static bool get_reqs_available(struct kioctx *ctx)
1043 {
1044 	if (__get_reqs_available(ctx))
1045 		return true;
1046 	user_refill_reqs_available(ctx);
1047 	return __get_reqs_available(ctx);
1048 }
1049 
1050 /* aio_get_req
1051  *	Allocate a slot for an aio request.
1052  * Returns NULL if no requests are free.
1053  *
1054  * The refcount is initialized to 2 - one for the async op completion,
1055  * one for the synchronous code that does this.
1056  */
aio_get_req(struct kioctx * ctx)1057 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1058 {
1059 	struct aio_kiocb *req;
1060 
1061 	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1062 	if (unlikely(!req))
1063 		return NULL;
1064 
1065 	if (unlikely(!get_reqs_available(ctx))) {
1066 		kmem_cache_free(kiocb_cachep, req);
1067 		return NULL;
1068 	}
1069 
1070 	percpu_ref_get(&ctx->reqs);
1071 	req->ki_ctx = ctx;
1072 	INIT_LIST_HEAD(&req->ki_list);
1073 	refcount_set(&req->ki_refcnt, 2);
1074 	req->ki_eventfd = NULL;
1075 	return req;
1076 }
1077 
lookup_ioctx(unsigned long ctx_id)1078 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1079 {
1080 	struct aio_ring __user *ring  = (void __user *)ctx_id;
1081 	struct mm_struct *mm = current->mm;
1082 	struct kioctx *ctx, *ret = NULL;
1083 	struct kioctx_table *table;
1084 	unsigned id;
1085 
1086 	if (get_user(id, &ring->id))
1087 		return NULL;
1088 
1089 	rcu_read_lock();
1090 	table = rcu_dereference(mm->ioctx_table);
1091 
1092 	if (!table || id >= table->nr)
1093 		goto out;
1094 
1095 	id = array_index_nospec(id, table->nr);
1096 	ctx = rcu_dereference(table->table[id]);
1097 	if (ctx && ctx->user_id == ctx_id) {
1098 		if (percpu_ref_tryget_live(&ctx->users))
1099 			ret = ctx;
1100 	}
1101 out:
1102 	rcu_read_unlock();
1103 	return ret;
1104 }
1105 
iocb_destroy(struct aio_kiocb * iocb)1106 static inline void iocb_destroy(struct aio_kiocb *iocb)
1107 {
1108 	if (iocb->ki_eventfd)
1109 		eventfd_ctx_put(iocb->ki_eventfd);
1110 	if (iocb->ki_filp)
1111 		fput(iocb->ki_filp);
1112 	percpu_ref_put(&iocb->ki_ctx->reqs);
1113 	kmem_cache_free(kiocb_cachep, iocb);
1114 }
1115 
1116 /* aio_complete
1117  *	Called when the io request on the given iocb is complete.
1118  */
aio_complete(struct aio_kiocb * iocb)1119 static void aio_complete(struct aio_kiocb *iocb)
1120 {
1121 	struct kioctx	*ctx = iocb->ki_ctx;
1122 	struct aio_ring	*ring;
1123 	struct io_event	*ev_page, *event;
1124 	unsigned tail, pos, head;
1125 	unsigned long	flags;
1126 
1127 	/*
1128 	 * Add a completion event to the ring buffer. Must be done holding
1129 	 * ctx->completion_lock to prevent other code from messing with the tail
1130 	 * pointer since we might be called from irq context.
1131 	 */
1132 	spin_lock_irqsave(&ctx->completion_lock, flags);
1133 
1134 	tail = ctx->tail;
1135 	pos = tail + AIO_EVENTS_OFFSET;
1136 
1137 	if (++tail >= ctx->nr_events)
1138 		tail = 0;
1139 
1140 	ev_page = page_address(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1141 	event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1142 
1143 	*event = iocb->ki_res;
1144 
1145 	flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1146 
1147 	pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1148 		 (void __user *)(unsigned long)iocb->ki_res.obj,
1149 		 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1150 
1151 	/* after flagging the request as done, we
1152 	 * must never even look at it again
1153 	 */
1154 	smp_wmb();	/* make event visible before updating tail */
1155 
1156 	ctx->tail = tail;
1157 
1158 	ring = page_address(ctx->ring_pages[0]);
1159 	head = ring->head;
1160 	ring->tail = tail;
1161 	flush_dcache_page(ctx->ring_pages[0]);
1162 
1163 	ctx->completed_events++;
1164 	if (ctx->completed_events > 1)
1165 		refill_reqs_available(ctx, head, tail);
1166 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
1167 
1168 	pr_debug("added to ring %p at [%u]\n", iocb, tail);
1169 
1170 	/*
1171 	 * Check if the user asked us to deliver the result through an
1172 	 * eventfd. The eventfd_signal() function is safe to be called
1173 	 * from IRQ context.
1174 	 */
1175 	if (iocb->ki_eventfd)
1176 		eventfd_signal(iocb->ki_eventfd, 1);
1177 
1178 	/*
1179 	 * We have to order our ring_info tail store above and test
1180 	 * of the wait list below outside the wait lock.  This is
1181 	 * like in wake_up_bit() where clearing a bit has to be
1182 	 * ordered with the unlocked test.
1183 	 */
1184 	smp_mb();
1185 
1186 	if (waitqueue_active(&ctx->wait))
1187 		wake_up(&ctx->wait);
1188 }
1189 
iocb_put(struct aio_kiocb * iocb)1190 static inline void iocb_put(struct aio_kiocb *iocb)
1191 {
1192 	if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1193 		aio_complete(iocb);
1194 		iocb_destroy(iocb);
1195 	}
1196 }
1197 
1198 /* aio_read_events_ring
1199  *	Pull an event off of the ioctx's event ring.  Returns the number of
1200  *	events fetched
1201  */
aio_read_events_ring(struct kioctx * ctx,struct io_event __user * event,long nr)1202 static long aio_read_events_ring(struct kioctx *ctx,
1203 				 struct io_event __user *event, long nr)
1204 {
1205 	struct aio_ring *ring;
1206 	unsigned head, tail, pos;
1207 	long ret = 0;
1208 	int copy_ret;
1209 
1210 	/*
1211 	 * The mutex can block and wake us up and that will cause
1212 	 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1213 	 * and repeat. This should be rare enough that it doesn't cause
1214 	 * peformance issues. See the comment in read_events() for more detail.
1215 	 */
1216 	sched_annotate_sleep();
1217 	mutex_lock(&ctx->ring_lock);
1218 
1219 	/* Access to ->ring_pages here is protected by ctx->ring_lock. */
1220 	ring = page_address(ctx->ring_pages[0]);
1221 	head = ring->head;
1222 	tail = ring->tail;
1223 
1224 	/*
1225 	 * Ensure that once we've read the current tail pointer, that
1226 	 * we also see the events that were stored up to the tail.
1227 	 */
1228 	smp_rmb();
1229 
1230 	pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1231 
1232 	if (head == tail)
1233 		goto out;
1234 
1235 	head %= ctx->nr_events;
1236 	tail %= ctx->nr_events;
1237 
1238 	while (ret < nr) {
1239 		long avail;
1240 		struct io_event *ev;
1241 		struct page *page;
1242 
1243 		avail = (head <= tail ?  tail : ctx->nr_events) - head;
1244 		if (head == tail)
1245 			break;
1246 
1247 		pos = head + AIO_EVENTS_OFFSET;
1248 		page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1249 		pos %= AIO_EVENTS_PER_PAGE;
1250 
1251 		avail = min(avail, nr - ret);
1252 		avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1253 
1254 		ev = page_address(page);
1255 		copy_ret = copy_to_user(event + ret, ev + pos,
1256 					sizeof(*ev) * avail);
1257 
1258 		if (unlikely(copy_ret)) {
1259 			ret = -EFAULT;
1260 			goto out;
1261 		}
1262 
1263 		ret += avail;
1264 		head += avail;
1265 		head %= ctx->nr_events;
1266 	}
1267 
1268 	ring = page_address(ctx->ring_pages[0]);
1269 	ring->head = head;
1270 	flush_dcache_page(ctx->ring_pages[0]);
1271 
1272 	pr_debug("%li  h%u t%u\n", ret, head, tail);
1273 out:
1274 	mutex_unlock(&ctx->ring_lock);
1275 
1276 	return ret;
1277 }
1278 
aio_read_events(struct kioctx * ctx,long min_nr,long nr,struct io_event __user * event,long * i)1279 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1280 			    struct io_event __user *event, long *i)
1281 {
1282 	long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1283 
1284 	if (ret > 0)
1285 		*i += ret;
1286 
1287 	if (unlikely(atomic_read(&ctx->dead)))
1288 		ret = -EINVAL;
1289 
1290 	if (!*i)
1291 		*i = ret;
1292 
1293 	return ret < 0 || *i >= min_nr;
1294 }
1295 
read_events(struct kioctx * ctx,long min_nr,long nr,struct io_event __user * event,ktime_t until)1296 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1297 			struct io_event __user *event,
1298 			ktime_t until)
1299 {
1300 	long ret = 0;
1301 
1302 	/*
1303 	 * Note that aio_read_events() is being called as the conditional - i.e.
1304 	 * we're calling it after prepare_to_wait() has set task state to
1305 	 * TASK_INTERRUPTIBLE.
1306 	 *
1307 	 * But aio_read_events() can block, and if it blocks it's going to flip
1308 	 * the task state back to TASK_RUNNING.
1309 	 *
1310 	 * This should be ok, provided it doesn't flip the state back to
1311 	 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1312 	 * will only happen if the mutex_lock() call blocks, and we then find
1313 	 * the ringbuffer empty. So in practice we should be ok, but it's
1314 	 * something to be aware of when touching this code.
1315 	 */
1316 	if (until == 0)
1317 		aio_read_events(ctx, min_nr, nr, event, &ret);
1318 	else
1319 		wait_event_interruptible_hrtimeout(ctx->wait,
1320 				aio_read_events(ctx, min_nr, nr, event, &ret),
1321 				until);
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 #ifdef CONFIG_COMPAT
COMPAT_SYSCALL_DEFINE2(io_setup,unsigned,nr_events,u32 __user *,ctx32p)1369 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1370 {
1371 	struct kioctx *ioctx = NULL;
1372 	unsigned long ctx;
1373 	long ret;
1374 
1375 	ret = get_user(ctx, ctx32p);
1376 	if (unlikely(ret))
1377 		goto out;
1378 
1379 	ret = -EINVAL;
1380 	if (unlikely(ctx || nr_events == 0)) {
1381 		pr_debug("EINVAL: ctx %lu nr_events %u\n",
1382 		         ctx, nr_events);
1383 		goto out;
1384 	}
1385 
1386 	ioctx = ioctx_alloc(nr_events);
1387 	ret = PTR_ERR(ioctx);
1388 	if (!IS_ERR(ioctx)) {
1389 		/* truncating is ok because it's a user address */
1390 		ret = put_user((u32)ioctx->user_id, ctx32p);
1391 		if (ret)
1392 			kill_ioctx(current->mm, ioctx, NULL);
1393 		percpu_ref_put(&ioctx->users);
1394 	}
1395 
1396 out:
1397 	return ret;
1398 }
1399 #endif
1400 
1401 /* sys_io_destroy:
1402  *	Destroy the aio_context specified.  May cancel any outstanding
1403  *	AIOs and block on completion.  Will fail with -ENOSYS if not
1404  *	implemented.  May fail with -EINVAL if the context pointed to
1405  *	is invalid.
1406  */
SYSCALL_DEFINE1(io_destroy,aio_context_t,ctx)1407 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1408 {
1409 	struct kioctx *ioctx = lookup_ioctx(ctx);
1410 	if (likely(NULL != ioctx)) {
1411 		struct ctx_rq_wait wait;
1412 		int ret;
1413 
1414 		init_completion(&wait.comp);
1415 		atomic_set(&wait.count, 1);
1416 
1417 		/* Pass requests_done to kill_ioctx() where it can be set
1418 		 * in a thread-safe way. If we try to set it here then we have
1419 		 * a race condition if two io_destroy() called simultaneously.
1420 		 */
1421 		ret = kill_ioctx(current->mm, ioctx, &wait);
1422 		percpu_ref_put(&ioctx->users);
1423 
1424 		/* Wait until all IO for the context are done. Otherwise kernel
1425 		 * keep using user-space buffers even if user thinks the context
1426 		 * is destroyed.
1427 		 */
1428 		if (!ret)
1429 			wait_for_completion(&wait.comp);
1430 
1431 		return ret;
1432 	}
1433 	pr_debug("EINVAL: invalid context id\n");
1434 	return -EINVAL;
1435 }
1436 
aio_remove_iocb(struct aio_kiocb * iocb)1437 static void aio_remove_iocb(struct aio_kiocb *iocb)
1438 {
1439 	struct kioctx *ctx = iocb->ki_ctx;
1440 	unsigned long flags;
1441 
1442 	spin_lock_irqsave(&ctx->ctx_lock, flags);
1443 	list_del(&iocb->ki_list);
1444 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1445 }
1446 
aio_complete_rw(struct kiocb * kiocb,long res)1447 static void aio_complete_rw(struct kiocb *kiocb, long res)
1448 {
1449 	struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1450 
1451 	if (!list_empty_careful(&iocb->ki_list))
1452 		aio_remove_iocb(iocb);
1453 
1454 	if (kiocb->ki_flags & IOCB_WRITE) {
1455 		struct inode *inode = file_inode(kiocb->ki_filp);
1456 
1457 		if (S_ISREG(inode->i_mode))
1458 			kiocb_end_write(kiocb);
1459 	}
1460 
1461 	iocb->ki_res.res = res;
1462 	iocb->ki_res.res2 = 0;
1463 	iocb_put(iocb);
1464 }
1465 
aio_prep_rw(struct kiocb * req,const struct iocb * iocb)1466 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1467 {
1468 	int ret;
1469 
1470 	req->ki_complete = aio_complete_rw;
1471 	req->private = NULL;
1472 	req->ki_pos = iocb->aio_offset;
1473 	req->ki_flags = req->ki_filp->f_iocb_flags | IOCB_AIO_RW;
1474 	if (iocb->aio_flags & IOCB_FLAG_RESFD)
1475 		req->ki_flags |= IOCB_EVENTFD;
1476 	if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1477 		/*
1478 		 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1479 		 * aio_reqprio is interpreted as an I/O scheduling
1480 		 * class and priority.
1481 		 */
1482 		ret = ioprio_check_cap(iocb->aio_reqprio);
1483 		if (ret) {
1484 			pr_debug("aio ioprio check cap error: %d\n", ret);
1485 			return ret;
1486 		}
1487 
1488 		req->ki_ioprio = iocb->aio_reqprio;
1489 	} else
1490 		req->ki_ioprio = get_current_ioprio();
1491 
1492 	ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1493 	if (unlikely(ret))
1494 		return ret;
1495 
1496 	req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1497 	return 0;
1498 }
1499 
aio_setup_rw(int rw,const struct iocb * iocb,struct iovec ** iovec,bool vectored,bool compat,struct iov_iter * iter)1500 static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1501 		struct iovec **iovec, bool vectored, bool compat,
1502 		struct iov_iter *iter)
1503 {
1504 	void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1505 	size_t len = iocb->aio_nbytes;
1506 
1507 	if (!vectored) {
1508 		ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1509 		*iovec = NULL;
1510 		return ret;
1511 	}
1512 
1513 	return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1514 }
1515 
aio_rw_done(struct kiocb * req,ssize_t ret)1516 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1517 {
1518 	switch (ret) {
1519 	case -EIOCBQUEUED:
1520 		break;
1521 	case -ERESTARTSYS:
1522 	case -ERESTARTNOINTR:
1523 	case -ERESTARTNOHAND:
1524 	case -ERESTART_RESTARTBLOCK:
1525 		/*
1526 		 * There's no easy way to restart the syscall since other AIO's
1527 		 * may be already running. Just fail this IO with EINTR.
1528 		 */
1529 		ret = -EINTR;
1530 		fallthrough;
1531 	default:
1532 		req->ki_complete(req, ret);
1533 	}
1534 }
1535 
aio_read(struct kiocb * req,const struct iocb * iocb,bool vectored,bool compat)1536 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1537 			bool vectored, bool compat)
1538 {
1539 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1540 	struct iov_iter iter;
1541 	struct file *file;
1542 	int ret;
1543 
1544 	ret = aio_prep_rw(req, iocb);
1545 	if (ret)
1546 		return ret;
1547 	file = req->ki_filp;
1548 	if (unlikely(!(file->f_mode & FMODE_READ)))
1549 		return -EBADF;
1550 	if (unlikely(!file->f_op->read_iter))
1551 		return -EINVAL;
1552 
1553 	ret = aio_setup_rw(ITER_DEST, iocb, &iovec, vectored, compat, &iter);
1554 	if (ret < 0)
1555 		return ret;
1556 	ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1557 	if (!ret)
1558 		aio_rw_done(req, call_read_iter(file, req, &iter));
1559 	kfree(iovec);
1560 	return ret;
1561 }
1562 
aio_write(struct kiocb * req,const struct iocb * iocb,bool vectored,bool compat)1563 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1564 			 bool vectored, bool compat)
1565 {
1566 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1567 	struct iov_iter iter;
1568 	struct file *file;
1569 	int ret;
1570 
1571 	ret = aio_prep_rw(req, iocb);
1572 	if (ret)
1573 		return ret;
1574 	file = req->ki_filp;
1575 
1576 	if (unlikely(!(file->f_mode & FMODE_WRITE)))
1577 		return -EBADF;
1578 	if (unlikely(!file->f_op->write_iter))
1579 		return -EINVAL;
1580 
1581 	ret = aio_setup_rw(ITER_SOURCE, iocb, &iovec, vectored, compat, &iter);
1582 	if (ret < 0)
1583 		return ret;
1584 	ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1585 	if (!ret) {
1586 		if (S_ISREG(file_inode(file)->i_mode))
1587 			kiocb_start_write(req);
1588 		req->ki_flags |= IOCB_WRITE;
1589 		aio_rw_done(req, call_write_iter(file, req, &iter));
1590 	}
1591 	kfree(iovec);
1592 	return ret;
1593 }
1594 
aio_fsync_work(struct work_struct * work)1595 static void aio_fsync_work(struct work_struct *work)
1596 {
1597 	struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1598 	const struct cred *old_cred = override_creds(iocb->fsync.creds);
1599 
1600 	iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1601 	revert_creds(old_cred);
1602 	put_cred(iocb->fsync.creds);
1603 	iocb_put(iocb);
1604 }
1605 
aio_fsync(struct fsync_iocb * req,const struct iocb * iocb,bool datasync)1606 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1607 		     bool datasync)
1608 {
1609 	if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1610 			iocb->aio_rw_flags))
1611 		return -EINVAL;
1612 
1613 	if (unlikely(!req->file->f_op->fsync))
1614 		return -EINVAL;
1615 
1616 	req->creds = prepare_creds();
1617 	if (!req->creds)
1618 		return -ENOMEM;
1619 
1620 	req->datasync = datasync;
1621 	INIT_WORK(&req->work, aio_fsync_work);
1622 	schedule_work(&req->work);
1623 	return 0;
1624 }
1625 
aio_poll_put_work(struct work_struct * work)1626 static void aio_poll_put_work(struct work_struct *work)
1627 {
1628 	struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1629 	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1630 
1631 	iocb_put(iocb);
1632 }
1633 
1634 /*
1635  * Safely lock the waitqueue which the request is on, synchronizing with the
1636  * case where the ->poll() provider decides to free its waitqueue early.
1637  *
1638  * Returns true on success, meaning that req->head->lock was locked, req->wait
1639  * is on req->head, and an RCU read lock was taken.  Returns false if the
1640  * request was already removed from its waitqueue (which might no longer exist).
1641  */
poll_iocb_lock_wq(struct poll_iocb * req)1642 static bool poll_iocb_lock_wq(struct poll_iocb *req)
1643 {
1644 	wait_queue_head_t *head;
1645 
1646 	/*
1647 	 * While we hold the waitqueue lock and the waitqueue is nonempty,
1648 	 * wake_up_pollfree() will wait for us.  However, taking the waitqueue
1649 	 * lock in the first place can race with the waitqueue being freed.
1650 	 *
1651 	 * We solve this as eventpoll does: by taking advantage of the fact that
1652 	 * all users of wake_up_pollfree() will RCU-delay the actual free.  If
1653 	 * we enter rcu_read_lock() and see that the pointer to the queue is
1654 	 * non-NULL, we can then lock it without the memory being freed out from
1655 	 * under us, then check whether the request is still on the queue.
1656 	 *
1657 	 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1658 	 * case the caller deletes the entry from the queue, leaving it empty.
1659 	 * In that case, only RCU prevents the queue memory from being freed.
1660 	 */
1661 	rcu_read_lock();
1662 	head = smp_load_acquire(&req->head);
1663 	if (head) {
1664 		spin_lock(&head->lock);
1665 		if (!list_empty(&req->wait.entry))
1666 			return true;
1667 		spin_unlock(&head->lock);
1668 	}
1669 	rcu_read_unlock();
1670 	return false;
1671 }
1672 
poll_iocb_unlock_wq(struct poll_iocb * req)1673 static void poll_iocb_unlock_wq(struct poll_iocb *req)
1674 {
1675 	spin_unlock(&req->head->lock);
1676 	rcu_read_unlock();
1677 }
1678 
aio_poll_complete_work(struct work_struct * work)1679 static void aio_poll_complete_work(struct work_struct *work)
1680 {
1681 	struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1682 	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1683 	struct poll_table_struct pt = { ._key = req->events };
1684 	struct kioctx *ctx = iocb->ki_ctx;
1685 	__poll_t mask = 0;
1686 
1687 	if (!READ_ONCE(req->cancelled))
1688 		mask = vfs_poll(req->file, &pt) & req->events;
1689 
1690 	/*
1691 	 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1692 	 * calling ->ki_cancel.  We need the ctx_lock roundtrip here to
1693 	 * synchronize with them.  In the cancellation case the list_del_init
1694 	 * itself is not actually needed, but harmless so we keep it in to
1695 	 * avoid further branches in the fast path.
1696 	 */
1697 	spin_lock_irq(&ctx->ctx_lock);
1698 	if (poll_iocb_lock_wq(req)) {
1699 		if (!mask && !READ_ONCE(req->cancelled)) {
1700 			/*
1701 			 * The request isn't actually ready to be completed yet.
1702 			 * Reschedule completion if another wakeup came in.
1703 			 */
1704 			if (req->work_need_resched) {
1705 				schedule_work(&req->work);
1706 				req->work_need_resched = false;
1707 			} else {
1708 				req->work_scheduled = false;
1709 			}
1710 			poll_iocb_unlock_wq(req);
1711 			spin_unlock_irq(&ctx->ctx_lock);
1712 			return;
1713 		}
1714 		list_del_init(&req->wait.entry);
1715 		poll_iocb_unlock_wq(req);
1716 	} /* else, POLLFREE has freed the waitqueue, so we must complete */
1717 	list_del_init(&iocb->ki_list);
1718 	iocb->ki_res.res = mangle_poll(mask);
1719 	spin_unlock_irq(&ctx->ctx_lock);
1720 
1721 	iocb_put(iocb);
1722 }
1723 
1724 /* assumes we are called with irqs disabled */
aio_poll_cancel(struct kiocb * iocb)1725 static int aio_poll_cancel(struct kiocb *iocb)
1726 {
1727 	struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1728 	struct poll_iocb *req = &aiocb->poll;
1729 
1730 	if (poll_iocb_lock_wq(req)) {
1731 		WRITE_ONCE(req->cancelled, true);
1732 		if (!req->work_scheduled) {
1733 			schedule_work(&aiocb->poll.work);
1734 			req->work_scheduled = true;
1735 		}
1736 		poll_iocb_unlock_wq(req);
1737 	} /* else, the request was force-cancelled by POLLFREE already */
1738 
1739 	return 0;
1740 }
1741 
aio_poll_wake(struct wait_queue_entry * wait,unsigned mode,int sync,void * key)1742 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1743 		void *key)
1744 {
1745 	struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1746 	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1747 	__poll_t mask = key_to_poll(key);
1748 	unsigned long flags;
1749 
1750 	/* for instances that support it check for an event match first: */
1751 	if (mask && !(mask & req->events))
1752 		return 0;
1753 
1754 	/*
1755 	 * Complete the request inline if possible.  This requires that three
1756 	 * conditions be met:
1757 	 *   1. An event mask must have been passed.  If a plain wakeup was done
1758 	 *	instead, then mask == 0 and we have to call vfs_poll() to get
1759 	 *	the events, so inline completion isn't possible.
1760 	 *   2. The completion work must not have already been scheduled.
1761 	 *   3. ctx_lock must not be busy.  We have to use trylock because we
1762 	 *	already hold the waitqueue lock, so this inverts the normal
1763 	 *	locking order.  Use irqsave/irqrestore because not all
1764 	 *	filesystems (e.g. fuse) call this function with IRQs disabled,
1765 	 *	yet IRQs have to be disabled before ctx_lock is obtained.
1766 	 */
1767 	if (mask && !req->work_scheduled &&
1768 	    spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1769 		struct kioctx *ctx = iocb->ki_ctx;
1770 
1771 		list_del_init(&req->wait.entry);
1772 		list_del(&iocb->ki_list);
1773 		iocb->ki_res.res = mangle_poll(mask);
1774 		if (iocb->ki_eventfd && !eventfd_signal_allowed()) {
1775 			iocb = NULL;
1776 			INIT_WORK(&req->work, aio_poll_put_work);
1777 			schedule_work(&req->work);
1778 		}
1779 		spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1780 		if (iocb)
1781 			iocb_put(iocb);
1782 	} else {
1783 		/*
1784 		 * Schedule the completion work if needed.  If it was already
1785 		 * scheduled, record that another wakeup came in.
1786 		 *
1787 		 * Don't remove the request from the waitqueue here, as it might
1788 		 * not actually be complete yet (we won't know until vfs_poll()
1789 		 * is called), and we must not miss any wakeups.  POLLFREE is an
1790 		 * exception to this; see below.
1791 		 */
1792 		if (req->work_scheduled) {
1793 			req->work_need_resched = true;
1794 		} else {
1795 			schedule_work(&req->work);
1796 			req->work_scheduled = true;
1797 		}
1798 
1799 		/*
1800 		 * If the waitqueue is being freed early but we can't complete
1801 		 * the request inline, we have to tear down the request as best
1802 		 * we can.  That means immediately removing the request from its
1803 		 * waitqueue and preventing all further accesses to the
1804 		 * waitqueue via the request.  We also need to schedule the
1805 		 * completion work (done above).  Also mark the request as
1806 		 * cancelled, to potentially skip an unneeded call to ->poll().
1807 		 */
1808 		if (mask & POLLFREE) {
1809 			WRITE_ONCE(req->cancelled, true);
1810 			list_del_init(&req->wait.entry);
1811 
1812 			/*
1813 			 * Careful: this *must* be the last step, since as soon
1814 			 * as req->head is NULL'ed out, the request can be
1815 			 * completed and freed, since aio_poll_complete_work()
1816 			 * will no longer need to take the waitqueue lock.
1817 			 */
1818 			smp_store_release(&req->head, NULL);
1819 		}
1820 	}
1821 	return 1;
1822 }
1823 
1824 struct aio_poll_table {
1825 	struct poll_table_struct	pt;
1826 	struct aio_kiocb		*iocb;
1827 	bool				queued;
1828 	int				error;
1829 };
1830 
1831 static void
aio_poll_queue_proc(struct file * file,struct wait_queue_head * head,struct poll_table_struct * p)1832 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1833 		struct poll_table_struct *p)
1834 {
1835 	struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1836 
1837 	/* multiple wait queues per file are not supported */
1838 	if (unlikely(pt->queued)) {
1839 		pt->error = -EINVAL;
1840 		return;
1841 	}
1842 
1843 	pt->queued = true;
1844 	pt->error = 0;
1845 	pt->iocb->poll.head = head;
1846 	add_wait_queue(head, &pt->iocb->poll.wait);
1847 }
1848 
aio_poll(struct aio_kiocb * aiocb,const struct iocb * iocb)1849 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1850 {
1851 	struct kioctx *ctx = aiocb->ki_ctx;
1852 	struct poll_iocb *req = &aiocb->poll;
1853 	struct aio_poll_table apt;
1854 	bool cancel = false;
1855 	__poll_t mask;
1856 
1857 	/* reject any unknown events outside the normal event mask. */
1858 	if ((u16)iocb->aio_buf != iocb->aio_buf)
1859 		return -EINVAL;
1860 	/* reject fields that are not defined for poll */
1861 	if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1862 		return -EINVAL;
1863 
1864 	INIT_WORK(&req->work, aio_poll_complete_work);
1865 	req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1866 
1867 	req->head = NULL;
1868 	req->cancelled = false;
1869 	req->work_scheduled = false;
1870 	req->work_need_resched = false;
1871 
1872 	apt.pt._qproc = aio_poll_queue_proc;
1873 	apt.pt._key = req->events;
1874 	apt.iocb = aiocb;
1875 	apt.queued = false;
1876 	apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1877 
1878 	/* initialized the list so that we can do list_empty checks */
1879 	INIT_LIST_HEAD(&req->wait.entry);
1880 	init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1881 
1882 	mask = vfs_poll(req->file, &apt.pt) & req->events;
1883 	spin_lock_irq(&ctx->ctx_lock);
1884 	if (likely(apt.queued)) {
1885 		bool on_queue = poll_iocb_lock_wq(req);
1886 
1887 		if (!on_queue || req->work_scheduled) {
1888 			/*
1889 			 * aio_poll_wake() already either scheduled the async
1890 			 * completion work, or completed the request inline.
1891 			 */
1892 			if (apt.error) /* unsupported case: multiple queues */
1893 				cancel = true;
1894 			apt.error = 0;
1895 			mask = 0;
1896 		}
1897 		if (mask || apt.error) {
1898 			/* Steal to complete synchronously. */
1899 			list_del_init(&req->wait.entry);
1900 		} else if (cancel) {
1901 			/* Cancel if possible (may be too late though). */
1902 			WRITE_ONCE(req->cancelled, true);
1903 		} else if (on_queue) {
1904 			/*
1905 			 * Actually waiting for an event, so add the request to
1906 			 * active_reqs so that it can be cancelled if needed.
1907 			 */
1908 			list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1909 			aiocb->ki_cancel = aio_poll_cancel;
1910 		}
1911 		if (on_queue)
1912 			poll_iocb_unlock_wq(req);
1913 	}
1914 	if (mask) { /* no async, we'd stolen it */
1915 		aiocb->ki_res.res = mangle_poll(mask);
1916 		apt.error = 0;
1917 	}
1918 	spin_unlock_irq(&ctx->ctx_lock);
1919 	if (mask)
1920 		iocb_put(aiocb);
1921 	return apt.error;
1922 }
1923 
__io_submit_one(struct kioctx * ctx,const struct iocb * iocb,struct iocb __user * user_iocb,struct aio_kiocb * req,bool compat)1924 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1925 			   struct iocb __user *user_iocb, struct aio_kiocb *req,
1926 			   bool compat)
1927 {
1928 	req->ki_filp = fget(iocb->aio_fildes);
1929 	if (unlikely(!req->ki_filp))
1930 		return -EBADF;
1931 
1932 	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1933 		struct eventfd_ctx *eventfd;
1934 		/*
1935 		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1936 		 * instance of the file* now. The file descriptor must be
1937 		 * an eventfd() fd, and will be signaled for each completed
1938 		 * event using the eventfd_signal() function.
1939 		 */
1940 		eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1941 		if (IS_ERR(eventfd))
1942 			return PTR_ERR(eventfd);
1943 
1944 		req->ki_eventfd = eventfd;
1945 	}
1946 
1947 	if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1948 		pr_debug("EFAULT: aio_key\n");
1949 		return -EFAULT;
1950 	}
1951 
1952 	req->ki_res.obj = (u64)(unsigned long)user_iocb;
1953 	req->ki_res.data = iocb->aio_data;
1954 	req->ki_res.res = 0;
1955 	req->ki_res.res2 = 0;
1956 
1957 	switch (iocb->aio_lio_opcode) {
1958 	case IOCB_CMD_PREAD:
1959 		return aio_read(&req->rw, iocb, false, compat);
1960 	case IOCB_CMD_PWRITE:
1961 		return aio_write(&req->rw, iocb, false, compat);
1962 	case IOCB_CMD_PREADV:
1963 		return aio_read(&req->rw, iocb, true, compat);
1964 	case IOCB_CMD_PWRITEV:
1965 		return aio_write(&req->rw, iocb, true, compat);
1966 	case IOCB_CMD_FSYNC:
1967 		return aio_fsync(&req->fsync, iocb, false);
1968 	case IOCB_CMD_FDSYNC:
1969 		return aio_fsync(&req->fsync, iocb, true);
1970 	case IOCB_CMD_POLL:
1971 		return aio_poll(req, iocb);
1972 	default:
1973 		pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1974 		return -EINVAL;
1975 	}
1976 }
1977 
io_submit_one(struct kioctx * ctx,struct iocb __user * user_iocb,bool compat)1978 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1979 			 bool compat)
1980 {
1981 	struct aio_kiocb *req;
1982 	struct iocb iocb;
1983 	int err;
1984 
1985 	if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1986 		return -EFAULT;
1987 
1988 	/* enforce forwards compatibility on users */
1989 	if (unlikely(iocb.aio_reserved2)) {
1990 		pr_debug("EINVAL: reserve field set\n");
1991 		return -EINVAL;
1992 	}
1993 
1994 	/* prevent overflows */
1995 	if (unlikely(
1996 	    (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
1997 	    (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
1998 	    ((ssize_t)iocb.aio_nbytes < 0)
1999 	   )) {
2000 		pr_debug("EINVAL: overflow check\n");
2001 		return -EINVAL;
2002 	}
2003 
2004 	req = aio_get_req(ctx);
2005 	if (unlikely(!req))
2006 		return -EAGAIN;
2007 
2008 	err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2009 
2010 	/* Done with the synchronous reference */
2011 	iocb_put(req);
2012 
2013 	/*
2014 	 * If err is 0, we'd either done aio_complete() ourselves or have
2015 	 * arranged for that to be done asynchronously.  Anything non-zero
2016 	 * means that we need to destroy req ourselves.
2017 	 */
2018 	if (unlikely(err)) {
2019 		iocb_destroy(req);
2020 		put_reqs_available(ctx, 1);
2021 	}
2022 	return err;
2023 }
2024 
2025 /* sys_io_submit:
2026  *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
2027  *	the number of iocbs queued.  May return -EINVAL if the aio_context
2028  *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
2029  *	*iocbpp[0] is not properly initialized, if the operation specified
2030  *	is invalid for the file descriptor in the iocb.  May fail with
2031  *	-EFAULT if any of the data structures point to invalid data.  May
2032  *	fail with -EBADF if the file descriptor specified in the first
2033  *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
2034  *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
2035  *	fail with -ENOSYS if not implemented.
2036  */
SYSCALL_DEFINE3(io_submit,aio_context_t,ctx_id,long,nr,struct iocb __user * __user *,iocbpp)2037 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2038 		struct iocb __user * __user *, iocbpp)
2039 {
2040 	struct kioctx *ctx;
2041 	long ret = 0;
2042 	int i = 0;
2043 	struct blk_plug plug;
2044 
2045 	if (unlikely(nr < 0))
2046 		return -EINVAL;
2047 
2048 	ctx = lookup_ioctx(ctx_id);
2049 	if (unlikely(!ctx)) {
2050 		pr_debug("EINVAL: invalid context id\n");
2051 		return -EINVAL;
2052 	}
2053 
2054 	if (nr > ctx->nr_events)
2055 		nr = ctx->nr_events;
2056 
2057 	if (nr > AIO_PLUG_THRESHOLD)
2058 		blk_start_plug(&plug);
2059 	for (i = 0; i < nr; i++) {
2060 		struct iocb __user *user_iocb;
2061 
2062 		if (unlikely(get_user(user_iocb, iocbpp + i))) {
2063 			ret = -EFAULT;
2064 			break;
2065 		}
2066 
2067 		ret = io_submit_one(ctx, user_iocb, false);
2068 		if (ret)
2069 			break;
2070 	}
2071 	if (nr > AIO_PLUG_THRESHOLD)
2072 		blk_finish_plug(&plug);
2073 
2074 	percpu_ref_put(&ctx->users);
2075 	return i ? i : ret;
2076 }
2077 
2078 #ifdef CONFIG_COMPAT
COMPAT_SYSCALL_DEFINE3(io_submit,compat_aio_context_t,ctx_id,int,nr,compat_uptr_t __user *,iocbpp)2079 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2080 		       int, nr, compat_uptr_t __user *, iocbpp)
2081 {
2082 	struct kioctx *ctx;
2083 	long ret = 0;
2084 	int i = 0;
2085 	struct blk_plug plug;
2086 
2087 	if (unlikely(nr < 0))
2088 		return -EINVAL;
2089 
2090 	ctx = lookup_ioctx(ctx_id);
2091 	if (unlikely(!ctx)) {
2092 		pr_debug("EINVAL: invalid context id\n");
2093 		return -EINVAL;
2094 	}
2095 
2096 	if (nr > ctx->nr_events)
2097 		nr = ctx->nr_events;
2098 
2099 	if (nr > AIO_PLUG_THRESHOLD)
2100 		blk_start_plug(&plug);
2101 	for (i = 0; i < nr; i++) {
2102 		compat_uptr_t user_iocb;
2103 
2104 		if (unlikely(get_user(user_iocb, iocbpp + i))) {
2105 			ret = -EFAULT;
2106 			break;
2107 		}
2108 
2109 		ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2110 		if (ret)
2111 			break;
2112 	}
2113 	if (nr > AIO_PLUG_THRESHOLD)
2114 		blk_finish_plug(&plug);
2115 
2116 	percpu_ref_put(&ctx->users);
2117 	return i ? i : ret;
2118 }
2119 #endif
2120 
2121 /* sys_io_cancel:
2122  *	Attempts to cancel an iocb previously passed to io_submit.  If
2123  *	the operation is successfully cancelled, the resulting event is
2124  *	copied into the memory pointed to by result without being placed
2125  *	into the completion queue and 0 is returned.  May fail with
2126  *	-EFAULT if any of the data structures pointed to are invalid.
2127  *	May fail with -EINVAL if aio_context specified by ctx_id is
2128  *	invalid.  May fail with -EAGAIN if the iocb specified was not
2129  *	cancelled.  Will fail with -ENOSYS if not implemented.
2130  */
SYSCALL_DEFINE3(io_cancel,aio_context_t,ctx_id,struct iocb __user *,iocb,struct io_event __user *,result)2131 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2132 		struct io_event __user *, result)
2133 {
2134 	struct kioctx *ctx;
2135 	struct aio_kiocb *kiocb;
2136 	int ret = -EINVAL;
2137 	u32 key;
2138 	u64 obj = (u64)(unsigned long)iocb;
2139 
2140 	if (unlikely(get_user(key, &iocb->aio_key)))
2141 		return -EFAULT;
2142 	if (unlikely(key != KIOCB_KEY))
2143 		return -EINVAL;
2144 
2145 	ctx = lookup_ioctx(ctx_id);
2146 	if (unlikely(!ctx))
2147 		return -EINVAL;
2148 
2149 	spin_lock_irq(&ctx->ctx_lock);
2150 	/* TODO: use a hash or array, this sucks. */
2151 	list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2152 		if (kiocb->ki_res.obj == obj) {
2153 			ret = kiocb->ki_cancel(&kiocb->rw);
2154 			list_del_init(&kiocb->ki_list);
2155 			break;
2156 		}
2157 	}
2158 	spin_unlock_irq(&ctx->ctx_lock);
2159 
2160 	if (!ret) {
2161 		/*
2162 		 * The result argument is no longer used - the io_event is
2163 		 * always delivered via the ring buffer. -EINPROGRESS indicates
2164 		 * cancellation is progress:
2165 		 */
2166 		ret = -EINPROGRESS;
2167 	}
2168 
2169 	percpu_ref_put(&ctx->users);
2170 
2171 	return ret;
2172 }
2173 
do_io_getevents(aio_context_t ctx_id,long min_nr,long nr,struct io_event __user * events,struct timespec64 * ts)2174 static long do_io_getevents(aio_context_t ctx_id,
2175 		long min_nr,
2176 		long nr,
2177 		struct io_event __user *events,
2178 		struct timespec64 *ts)
2179 {
2180 	ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2181 	struct kioctx *ioctx = lookup_ioctx(ctx_id);
2182 	long ret = -EINVAL;
2183 
2184 	if (likely(ioctx)) {
2185 		if (likely(min_nr <= nr && min_nr >= 0))
2186 			ret = read_events(ioctx, min_nr, nr, events, until);
2187 		percpu_ref_put(&ioctx->users);
2188 	}
2189 
2190 	return ret;
2191 }
2192 
2193 /* io_getevents:
2194  *	Attempts to read at least min_nr events and up to nr events from
2195  *	the completion queue for the aio_context specified by ctx_id. If
2196  *	it succeeds, the number of read events is returned. May fail with
2197  *	-EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2198  *	out of range, if timeout is out of range.  May fail with -EFAULT
2199  *	if any of the memory specified is invalid.  May return 0 or
2200  *	< min_nr if the timeout specified by timeout has elapsed
2201  *	before sufficient events are available, where timeout == NULL
2202  *	specifies an infinite timeout. Note that the timeout pointed to by
2203  *	timeout is relative.  Will fail with -ENOSYS if not implemented.
2204  */
2205 #ifdef CONFIG_64BIT
2206 
SYSCALL_DEFINE5(io_getevents,aio_context_t,ctx_id,long,min_nr,long,nr,struct io_event __user *,events,struct __kernel_timespec __user *,timeout)2207 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2208 		long, min_nr,
2209 		long, nr,
2210 		struct io_event __user *, events,
2211 		struct __kernel_timespec __user *, timeout)
2212 {
2213 	struct timespec64	ts;
2214 	int			ret;
2215 
2216 	if (timeout && unlikely(get_timespec64(&ts, timeout)))
2217 		return -EFAULT;
2218 
2219 	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2220 	if (!ret && signal_pending(current))
2221 		ret = -EINTR;
2222 	return ret;
2223 }
2224 
2225 #endif
2226 
2227 struct __aio_sigset {
2228 	const sigset_t __user	*sigmask;
2229 	size_t		sigsetsize;
2230 };
2231 
SYSCALL_DEFINE6(io_pgetevents,aio_context_t,ctx_id,long,min_nr,long,nr,struct io_event __user *,events,struct __kernel_timespec __user *,timeout,const struct __aio_sigset __user *,usig)2232 SYSCALL_DEFINE6(io_pgetevents,
2233 		aio_context_t, ctx_id,
2234 		long, min_nr,
2235 		long, nr,
2236 		struct io_event __user *, events,
2237 		struct __kernel_timespec __user *, timeout,
2238 		const struct __aio_sigset __user *, usig)
2239 {
2240 	struct __aio_sigset	ksig = { NULL, };
2241 	struct timespec64	ts;
2242 	bool interrupted;
2243 	int ret;
2244 
2245 	if (timeout && unlikely(get_timespec64(&ts, timeout)))
2246 		return -EFAULT;
2247 
2248 	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2249 		return -EFAULT;
2250 
2251 	ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2252 	if (ret)
2253 		return ret;
2254 
2255 	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2256 
2257 	interrupted = signal_pending(current);
2258 	restore_saved_sigmask_unless(interrupted);
2259 	if (interrupted && !ret)
2260 		ret = -ERESTARTNOHAND;
2261 
2262 	return ret;
2263 }
2264 
2265 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2266 
SYSCALL_DEFINE6(io_pgetevents_time32,aio_context_t,ctx_id,long,min_nr,long,nr,struct io_event __user *,events,struct old_timespec32 __user *,timeout,const struct __aio_sigset __user *,usig)2267 SYSCALL_DEFINE6(io_pgetevents_time32,
2268 		aio_context_t, ctx_id,
2269 		long, min_nr,
2270 		long, nr,
2271 		struct io_event __user *, events,
2272 		struct old_timespec32 __user *, timeout,
2273 		const struct __aio_sigset __user *, usig)
2274 {
2275 	struct __aio_sigset	ksig = { NULL, };
2276 	struct timespec64	ts;
2277 	bool interrupted;
2278 	int ret;
2279 
2280 	if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2281 		return -EFAULT;
2282 
2283 	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2284 		return -EFAULT;
2285 
2286 
2287 	ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2288 	if (ret)
2289 		return ret;
2290 
2291 	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2292 
2293 	interrupted = signal_pending(current);
2294 	restore_saved_sigmask_unless(interrupted);
2295 	if (interrupted && !ret)
2296 		ret = -ERESTARTNOHAND;
2297 
2298 	return ret;
2299 }
2300 
2301 #endif
2302 
2303 #if defined(CONFIG_COMPAT_32BIT_TIME)
2304 
SYSCALL_DEFINE5(io_getevents_time32,__u32,ctx_id,__s32,min_nr,__s32,nr,struct io_event __user *,events,struct old_timespec32 __user *,timeout)2305 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2306 		__s32, min_nr,
2307 		__s32, nr,
2308 		struct io_event __user *, events,
2309 		struct old_timespec32 __user *, timeout)
2310 {
2311 	struct timespec64 t;
2312 	int ret;
2313 
2314 	if (timeout && get_old_timespec32(&t, timeout))
2315 		return -EFAULT;
2316 
2317 	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2318 	if (!ret && signal_pending(current))
2319 		ret = -EINTR;
2320 	return ret;
2321 }
2322 
2323 #endif
2324 
2325 #ifdef CONFIG_COMPAT
2326 
2327 struct __compat_aio_sigset {
2328 	compat_uptr_t		sigmask;
2329 	compat_size_t		sigsetsize;
2330 };
2331 
2332 #if defined(CONFIG_COMPAT_32BIT_TIME)
2333 
COMPAT_SYSCALL_DEFINE6(io_pgetevents,compat_aio_context_t,ctx_id,compat_long_t,min_nr,compat_long_t,nr,struct io_event __user *,events,struct old_timespec32 __user *,timeout,const struct __compat_aio_sigset __user *,usig)2334 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2335 		compat_aio_context_t, ctx_id,
2336 		compat_long_t, min_nr,
2337 		compat_long_t, nr,
2338 		struct io_event __user *, events,
2339 		struct old_timespec32 __user *, timeout,
2340 		const struct __compat_aio_sigset __user *, usig)
2341 {
2342 	struct __compat_aio_sigset ksig = { 0, };
2343 	struct timespec64 t;
2344 	bool interrupted;
2345 	int ret;
2346 
2347 	if (timeout && get_old_timespec32(&t, timeout))
2348 		return -EFAULT;
2349 
2350 	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2351 		return -EFAULT;
2352 
2353 	ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2354 	if (ret)
2355 		return ret;
2356 
2357 	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2358 
2359 	interrupted = signal_pending(current);
2360 	restore_saved_sigmask_unless(interrupted);
2361 	if (interrupted && !ret)
2362 		ret = -ERESTARTNOHAND;
2363 
2364 	return ret;
2365 }
2366 
2367 #endif
2368 
COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,compat_aio_context_t,ctx_id,compat_long_t,min_nr,compat_long_t,nr,struct io_event __user *,events,struct __kernel_timespec __user *,timeout,const struct __compat_aio_sigset __user *,usig)2369 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2370 		compat_aio_context_t, ctx_id,
2371 		compat_long_t, min_nr,
2372 		compat_long_t, nr,
2373 		struct io_event __user *, events,
2374 		struct __kernel_timespec __user *, timeout,
2375 		const struct __compat_aio_sigset __user *, usig)
2376 {
2377 	struct __compat_aio_sigset ksig = { 0, };
2378 	struct timespec64 t;
2379 	bool interrupted;
2380 	int ret;
2381 
2382 	if (timeout && get_timespec64(&t, timeout))
2383 		return -EFAULT;
2384 
2385 	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2386 		return -EFAULT;
2387 
2388 	ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2389 	if (ret)
2390 		return ret;
2391 
2392 	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2393 
2394 	interrupted = signal_pending(current);
2395 	restore_saved_sigmask_unless(interrupted);
2396 	if (interrupted && !ret)
2397 		ret = -ERESTARTNOHAND;
2398 
2399 	return ret;
2400 }
2401 #endif
2402