]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - fs/aio.c
UBUNTU: [Config] CONFIG_TYPEC_TCPM=m
[mirror_ubuntu-artful-kernel.git] / fs / aio.c
1 /*
2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
4 *
5 * Implements an efficient asynchronous io interface.
6 *
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 *
9 * See ../COPYING for licensing terms.
10 */
11 #define pr_fmt(fmt) "%s: " fmt, __func__
12
13 #include <linux/kernel.h>
14 #include <linux/init.h>
15 #include <linux/errno.h>
16 #include <linux/time.h>
17 #include <linux/aio_abi.h>
18 #include <linux/export.h>
19 #include <linux/syscalls.h>
20 #include <linux/backing-dev.h>
21 #include <linux/uio.h>
22
23 #include <linux/sched/signal.h>
24 #include <linux/fs.h>
25 #include <linux/file.h>
26 #include <linux/mm.h>
27 #include <linux/mman.h>
28 #include <linux/mmu_context.h>
29 #include <linux/percpu.h>
30 #include <linux/slab.h>
31 #include <linux/timer.h>
32 #include <linux/aio.h>
33 #include <linux/highmem.h>
34 #include <linux/workqueue.h>
35 #include <linux/security.h>
36 #include <linux/eventfd.h>
37 #include <linux/blkdev.h>
38 #include <linux/compat.h>
39 #include <linux/migrate.h>
40 #include <linux/ramfs.h>
41 #include <linux/percpu-refcount.h>
42 #include <linux/mount.h>
43
44 #include <asm/kmap_types.h>
45 #include <linux/uaccess.h>
46
47 #include "internal.h"
48
49 #define AIO_RING_MAGIC 0xa10a10a1
50 #define AIO_RING_COMPAT_FEATURES 1
51 #define AIO_RING_INCOMPAT_FEATURES 0
52 struct aio_ring {
53 unsigned id; /* kernel internal index number */
54 unsigned nr; /* number of io_events */
55 unsigned head; /* Written to by userland or under ring_lock
56 * mutex by aio_read_events_ring(). */
57 unsigned tail;
58
59 unsigned magic;
60 unsigned compat_features;
61 unsigned incompat_features;
62 unsigned header_length; /* size of aio_ring */
63
64
65 struct io_event io_events[0];
66 }; /* 128 bytes + ring size */
67
68 #define AIO_RING_PAGES 8
69
70 struct kioctx_table {
71 struct rcu_head rcu;
72 unsigned nr;
73 struct kioctx *table[];
74 };
75
76 struct kioctx_cpu {
77 unsigned reqs_available;
78 };
79
80 struct ctx_rq_wait {
81 struct completion comp;
82 atomic_t count;
83 };
84
85 struct kioctx {
86 struct percpu_ref users;
87 atomic_t dead;
88
89 struct percpu_ref reqs;
90
91 unsigned long user_id;
92
93 struct __percpu kioctx_cpu *cpu;
94
95 /*
96 * For percpu reqs_available, number of slots we move to/from global
97 * counter at a time:
98 */
99 unsigned req_batch;
100 /*
101 * This is what userspace passed to io_setup(), it's not used for
102 * anything but counting against the global max_reqs quota.
103 *
104 * The real limit is nr_events - 1, which will be larger (see
105 * aio_setup_ring())
106 */
107 unsigned max_reqs;
108
109 /* Size of ringbuffer, in units of struct io_event */
110 unsigned nr_events;
111
112 unsigned long mmap_base;
113 unsigned long mmap_size;
114
115 struct page **ring_pages;
116 long nr_pages;
117
118 struct work_struct free_work;
119
120 /*
121 * signals when all in-flight requests are done
122 */
123 struct ctx_rq_wait *rq_wait;
124
125 struct {
126 /*
127 * This counts the number of available slots in the ringbuffer,
128 * so we avoid overflowing it: it's decremented (if positive)
129 * when allocating a kiocb and incremented when the resulting
130 * io_event is pulled off the ringbuffer.
131 *
132 * We batch accesses to it with a percpu version.
133 */
134 atomic_t reqs_available;
135 } ____cacheline_aligned_in_smp;
136
137 struct {
138 spinlock_t ctx_lock;
139 struct list_head active_reqs; /* used for cancellation */
140 } ____cacheline_aligned_in_smp;
141
142 struct {
143 struct mutex ring_lock;
144 wait_queue_head_t wait;
145 } ____cacheline_aligned_in_smp;
146
147 struct {
148 unsigned tail;
149 unsigned completed_events;
150 spinlock_t completion_lock;
151 } ____cacheline_aligned_in_smp;
152
153 struct page *internal_pages[AIO_RING_PAGES];
154 struct file *aio_ring_file;
155
156 unsigned id;
157 };
158
159 /*
160 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
161 * cancelled or completed (this makes a certain amount of sense because
162 * successful cancellation - io_cancel() - does deliver the completion to
163 * userspace).
164 *
165 * And since most things don't implement kiocb cancellation and we'd really like
166 * kiocb completion to be lockless when possible, we use ki_cancel to
167 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
168 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
169 */
170 #define KIOCB_CANCELLED ((void *) (~0ULL))
171
172 struct aio_kiocb {
173 struct kiocb common;
174
175 struct kioctx *ki_ctx;
176 kiocb_cancel_fn *ki_cancel;
177
178 struct iocb __user *ki_user_iocb; /* user's aiocb */
179 __u64 ki_user_data; /* user's data for completion */
180
181 struct list_head ki_list; /* the aio core uses this
182 * for cancellation */
183
184 /*
185 * If the aio_resfd field of the userspace iocb is not zero,
186 * this is the underlying eventfd context to deliver events to.
187 */
188 struct eventfd_ctx *ki_eventfd;
189 };
190
191 /*------ sysctl variables----*/
192 static DEFINE_SPINLOCK(aio_nr_lock);
193 unsigned long aio_nr; /* current system wide number of aio requests */
194 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
195 /*----end sysctl variables---*/
196
197 static struct kmem_cache *kiocb_cachep;
198 static struct kmem_cache *kioctx_cachep;
199
200 static struct vfsmount *aio_mnt;
201
202 static const struct file_operations aio_ring_fops;
203 static const struct address_space_operations aio_ctx_aops;
204
205 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
206 {
207 struct qstr this = QSTR_INIT("[aio]", 5);
208 struct file *file;
209 struct path path;
210 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
211 if (IS_ERR(inode))
212 return ERR_CAST(inode);
213
214 inode->i_mapping->a_ops = &aio_ctx_aops;
215 inode->i_mapping->private_data = ctx;
216 inode->i_size = PAGE_SIZE * nr_pages;
217
218 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
219 if (!path.dentry) {
220 iput(inode);
221 return ERR_PTR(-ENOMEM);
222 }
223 path.mnt = mntget(aio_mnt);
224
225 d_instantiate(path.dentry, inode);
226 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
227 if (IS_ERR(file)) {
228 path_put(&path);
229 return file;
230 }
231
232 file->f_flags = O_RDWR;
233 return file;
234 }
235
236 static struct dentry *aio_mount(struct file_system_type *fs_type,
237 int flags, const char *dev_name, void *data)
238 {
239 static const struct dentry_operations ops = {
240 .d_dname = simple_dname,
241 };
242 struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops,
243 AIO_RING_MAGIC);
244
245 if (!IS_ERR(root))
246 root->d_sb->s_iflags |= SB_I_NOEXEC;
247 return root;
248 }
249
250 /* aio_setup
251 * Creates the slab caches used by the aio routines, panic on
252 * failure as this is done early during the boot sequence.
253 */
254 static int __init aio_setup(void)
255 {
256 static struct file_system_type aio_fs = {
257 .name = "aio",
258 .mount = aio_mount,
259 .kill_sb = kill_anon_super,
260 };
261 aio_mnt = kern_mount(&aio_fs);
262 if (IS_ERR(aio_mnt))
263 panic("Failed to create aio fs mount.");
264
265 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
266 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
267
268 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
269
270 return 0;
271 }
272 __initcall(aio_setup);
273
274 static void put_aio_ring_file(struct kioctx *ctx)
275 {
276 struct file *aio_ring_file = ctx->aio_ring_file;
277 struct address_space *i_mapping;
278
279 if (aio_ring_file) {
280 truncate_setsize(file_inode(aio_ring_file), 0);
281
282 /* Prevent further access to the kioctx from migratepages */
283 i_mapping = aio_ring_file->f_mapping;
284 spin_lock(&i_mapping->private_lock);
285 i_mapping->private_data = NULL;
286 ctx->aio_ring_file = NULL;
287 spin_unlock(&i_mapping->private_lock);
288
289 fput(aio_ring_file);
290 }
291 }
292
293 static void aio_free_ring(struct kioctx *ctx)
294 {
295 int i;
296
297 /* Disconnect the kiotx from the ring file. This prevents future
298 * accesses to the kioctx from page migration.
299 */
300 put_aio_ring_file(ctx);
301
302 for (i = 0; i < ctx->nr_pages; i++) {
303 struct page *page;
304 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
305 page_count(ctx->ring_pages[i]));
306 page = ctx->ring_pages[i];
307 if (!page)
308 continue;
309 ctx->ring_pages[i] = NULL;
310 put_page(page);
311 }
312
313 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
314 kfree(ctx->ring_pages);
315 ctx->ring_pages = NULL;
316 }
317 }
318
319 static int aio_ring_mremap(struct vm_area_struct *vma)
320 {
321 struct file *file = vma->vm_file;
322 struct mm_struct *mm = vma->vm_mm;
323 struct kioctx_table *table;
324 int i, res = -EINVAL;
325
326 spin_lock(&mm->ioctx_lock);
327 rcu_read_lock();
328 table = rcu_dereference(mm->ioctx_table);
329 for (i = 0; i < table->nr; i++) {
330 struct kioctx *ctx;
331
332 ctx = table->table[i];
333 if (ctx && ctx->aio_ring_file == file) {
334 if (!atomic_read(&ctx->dead)) {
335 ctx->user_id = ctx->mmap_base = vma->vm_start;
336 res = 0;
337 }
338 break;
339 }
340 }
341
342 rcu_read_unlock();
343 spin_unlock(&mm->ioctx_lock);
344 return res;
345 }
346
347 static const struct vm_operations_struct aio_ring_vm_ops = {
348 .mremap = aio_ring_mremap,
349 #if IS_ENABLED(CONFIG_MMU)
350 .fault = filemap_fault,
351 .map_pages = filemap_map_pages,
352 .page_mkwrite = filemap_page_mkwrite,
353 #endif
354 };
355
356 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
357 {
358 vma->vm_flags |= VM_DONTEXPAND;
359 vma->vm_ops = &aio_ring_vm_ops;
360 return 0;
361 }
362
363 static const struct file_operations aio_ring_fops = {
364 .mmap = aio_ring_mmap,
365 };
366
367 #if IS_ENABLED(CONFIG_MIGRATION)
368 static int aio_migratepage(struct address_space *mapping, struct page *new,
369 struct page *old, enum migrate_mode mode)
370 {
371 struct kioctx *ctx;
372 unsigned long flags;
373 pgoff_t idx;
374 int rc;
375
376 rc = 0;
377
378 /* mapping->private_lock here protects against the kioctx teardown. */
379 spin_lock(&mapping->private_lock);
380 ctx = mapping->private_data;
381 if (!ctx) {
382 rc = -EINVAL;
383 goto out;
384 }
385
386 /* The ring_lock mutex. The prevents aio_read_events() from writing
387 * to the ring's head, and prevents page migration from mucking in
388 * a partially initialized kiotx.
389 */
390 if (!mutex_trylock(&ctx->ring_lock)) {
391 rc = -EAGAIN;
392 goto out;
393 }
394
395 idx = old->index;
396 if (idx < (pgoff_t)ctx->nr_pages) {
397 /* Make sure the old page hasn't already been changed */
398 if (ctx->ring_pages[idx] != old)
399 rc = -EAGAIN;
400 } else
401 rc = -EINVAL;
402
403 if (rc != 0)
404 goto out_unlock;
405
406 /* Writeback must be complete */
407 BUG_ON(PageWriteback(old));
408 get_page(new);
409
410 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
411 if (rc != MIGRATEPAGE_SUCCESS) {
412 put_page(new);
413 goto out_unlock;
414 }
415
416 /* Take completion_lock to prevent other writes to the ring buffer
417 * while the old page is copied to the new. This prevents new
418 * events from being lost.
419 */
420 spin_lock_irqsave(&ctx->completion_lock, flags);
421 migrate_page_copy(new, old);
422 BUG_ON(ctx->ring_pages[idx] != old);
423 ctx->ring_pages[idx] = new;
424 spin_unlock_irqrestore(&ctx->completion_lock, flags);
425
426 /* The old page is no longer accessible. */
427 put_page(old);
428
429 out_unlock:
430 mutex_unlock(&ctx->ring_lock);
431 out:
432 spin_unlock(&mapping->private_lock);
433 return rc;
434 }
435 #endif
436
437 static const struct address_space_operations aio_ctx_aops = {
438 .set_page_dirty = __set_page_dirty_no_writeback,
439 #if IS_ENABLED(CONFIG_MIGRATION)
440 .migratepage = aio_migratepage,
441 #endif
442 };
443
444 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
445 {
446 struct aio_ring *ring;
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_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 if (down_write_killable(&mm->mmap_sem)) {
507 ctx->mmap_size = 0;
508 aio_free_ring(ctx);
509 return -EINTR;
510 }
511
512 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
513 PROT_READ | PROT_WRITE,
514 MAP_SHARED, 0, &unused, NULL);
515 up_write(&mm->mmap_sem);
516 if (IS_ERR((void *)ctx->mmap_base)) {
517 ctx->mmap_size = 0;
518 aio_free_ring(ctx);
519 return -ENOMEM;
520 }
521
522 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
523
524 ctx->user_id = ctx->mmap_base;
525 ctx->nr_events = nr_events; /* trusted copy */
526
527 ring = kmap_atomic(ctx->ring_pages[0]);
528 ring->nr = nr_events; /* user copy */
529 ring->id = ~0U;
530 ring->head = ring->tail = 0;
531 ring->magic = AIO_RING_MAGIC;
532 ring->compat_features = AIO_RING_COMPAT_FEATURES;
533 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
534 ring->header_length = sizeof(struct aio_ring);
535 kunmap_atomic(ring);
536 flush_dcache_page(ctx->ring_pages[0]);
537
538 return 0;
539 }
540
541 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
542 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
543 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
544
545 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
546 {
547 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
548 struct kioctx *ctx = req->ki_ctx;
549 unsigned long flags;
550
551 spin_lock_irqsave(&ctx->ctx_lock, flags);
552
553 if (!req->ki_list.next)
554 list_add(&req->ki_list, &ctx->active_reqs);
555
556 req->ki_cancel = cancel;
557
558 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
559 }
560 EXPORT_SYMBOL(kiocb_set_cancel_fn);
561
562 static int kiocb_cancel(struct aio_kiocb *kiocb)
563 {
564 kiocb_cancel_fn *old, *cancel;
565
566 /*
567 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
568 * actually has a cancel function, hence the cmpxchg()
569 */
570
571 cancel = ACCESS_ONCE(kiocb->ki_cancel);
572 do {
573 if (!cancel || cancel == KIOCB_CANCELLED)
574 return -EINVAL;
575
576 old = cancel;
577 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
578 } while (cancel != old);
579
580 return cancel(&kiocb->common);
581 }
582
583 static void free_ioctx(struct work_struct *work)
584 {
585 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
586
587 pr_debug("freeing %p\n", ctx);
588
589 aio_free_ring(ctx);
590 free_percpu(ctx->cpu);
591 percpu_ref_exit(&ctx->reqs);
592 percpu_ref_exit(&ctx->users);
593 kmem_cache_free(kioctx_cachep, ctx);
594 }
595
596 static void free_ioctx_reqs(struct percpu_ref *ref)
597 {
598 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
599
600 /* At this point we know that there are no any in-flight requests */
601 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
602 complete(&ctx->rq_wait->comp);
603
604 INIT_WORK(&ctx->free_work, free_ioctx);
605 schedule_work(&ctx->free_work);
606 }
607
608 /*
609 * When this function runs, the kioctx has been removed from the "hash table"
610 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
611 * now it's safe to cancel any that need to be.
612 */
613 static void free_ioctx_users(struct percpu_ref *ref)
614 {
615 struct kioctx *ctx = container_of(ref, struct kioctx, users);
616 struct aio_kiocb *req;
617
618 spin_lock_irq(&ctx->ctx_lock);
619
620 while (!list_empty(&ctx->active_reqs)) {
621 req = list_first_entry(&ctx->active_reqs,
622 struct aio_kiocb, ki_list);
623
624 list_del_init(&req->ki_list);
625 kiocb_cancel(req);
626 }
627
628 spin_unlock_irq(&ctx->ctx_lock);
629
630 percpu_ref_kill(&ctx->reqs);
631 percpu_ref_put(&ctx->reqs);
632 }
633
634 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
635 {
636 unsigned i, new_nr;
637 struct kioctx_table *table, *old;
638 struct aio_ring *ring;
639
640 spin_lock(&mm->ioctx_lock);
641 table = rcu_dereference_raw(mm->ioctx_table);
642
643 while (1) {
644 if (table)
645 for (i = 0; i < table->nr; i++)
646 if (!table->table[i]) {
647 ctx->id = i;
648 table->table[i] = ctx;
649 spin_unlock(&mm->ioctx_lock);
650
651 /* While kioctx setup is in progress,
652 * we are protected from page migration
653 * changes ring_pages by ->ring_lock.
654 */
655 ring = kmap_atomic(ctx->ring_pages[0]);
656 ring->id = ctx->id;
657 kunmap_atomic(ring);
658 return 0;
659 }
660
661 new_nr = (table ? table->nr : 1) * 4;
662 spin_unlock(&mm->ioctx_lock);
663
664 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
665 new_nr, GFP_KERNEL);
666 if (!table)
667 return -ENOMEM;
668
669 table->nr = new_nr;
670
671 spin_lock(&mm->ioctx_lock);
672 old = rcu_dereference_raw(mm->ioctx_table);
673
674 if (!old) {
675 rcu_assign_pointer(mm->ioctx_table, table);
676 } else if (table->nr > old->nr) {
677 memcpy(table->table, old->table,
678 old->nr * sizeof(struct kioctx *));
679
680 rcu_assign_pointer(mm->ioctx_table, table);
681 kfree_rcu(old, rcu);
682 } else {
683 kfree(table);
684 table = old;
685 }
686 }
687 }
688
689 static void aio_nr_sub(unsigned nr)
690 {
691 spin_lock(&aio_nr_lock);
692 if (WARN_ON(aio_nr - nr > aio_nr))
693 aio_nr = 0;
694 else
695 aio_nr -= nr;
696 spin_unlock(&aio_nr_lock);
697 }
698
699 /* ioctx_alloc
700 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
701 */
702 static struct kioctx *ioctx_alloc(unsigned nr_events)
703 {
704 struct mm_struct *mm = current->mm;
705 struct kioctx *ctx;
706 int err = -ENOMEM;
707
708 /*
709 * Store the original nr_events -- what userspace passed to io_setup(),
710 * for counting against the global limit -- before it changes.
711 */
712 unsigned int max_reqs = nr_events;
713
714 /*
715 * We keep track of the number of available ringbuffer slots, to prevent
716 * overflow (reqs_available), and we also use percpu counters for this.
717 *
718 * So since up to half the slots might be on other cpu's percpu counters
719 * and unavailable, double nr_events so userspace sees what they
720 * expected: additionally, we move req_batch slots to/from percpu
721 * counters at a time, so make sure that isn't 0:
722 */
723 nr_events = max(nr_events, num_possible_cpus() * 4);
724 nr_events *= 2;
725
726 /* Prevent overflows */
727 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
728 pr_debug("ENOMEM: nr_events too high\n");
729 return ERR_PTR(-EINVAL);
730 }
731
732 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
733 return ERR_PTR(-EAGAIN);
734
735 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
736 if (!ctx)
737 return ERR_PTR(-ENOMEM);
738
739 ctx->max_reqs = max_reqs;
740
741 spin_lock_init(&ctx->ctx_lock);
742 spin_lock_init(&ctx->completion_lock);
743 mutex_init(&ctx->ring_lock);
744 /* Protect against page migration throughout kiotx setup by keeping
745 * the ring_lock mutex held until setup is complete. */
746 mutex_lock(&ctx->ring_lock);
747 init_waitqueue_head(&ctx->wait);
748
749 INIT_LIST_HEAD(&ctx->active_reqs);
750
751 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
752 goto err;
753
754 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
755 goto err;
756
757 ctx->cpu = alloc_percpu(struct kioctx_cpu);
758 if (!ctx->cpu)
759 goto err;
760
761 err = aio_setup_ring(ctx, nr_events);
762 if (err < 0)
763 goto err;
764
765 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
766 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
767 if (ctx->req_batch < 1)
768 ctx->req_batch = 1;
769
770 /* limit the number of system wide aios */
771 spin_lock(&aio_nr_lock);
772 if (aio_nr + ctx->max_reqs > aio_max_nr ||
773 aio_nr + ctx->max_reqs < aio_nr) {
774 spin_unlock(&aio_nr_lock);
775 err = -EAGAIN;
776 goto err_ctx;
777 }
778 aio_nr += ctx->max_reqs;
779 spin_unlock(&aio_nr_lock);
780
781 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
782 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
783
784 err = ioctx_add_table(ctx, mm);
785 if (err)
786 goto err_cleanup;
787
788 /* Release the ring_lock mutex now that all setup is complete. */
789 mutex_unlock(&ctx->ring_lock);
790
791 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
792 ctx, ctx->user_id, mm, ctx->nr_events);
793 return ctx;
794
795 err_cleanup:
796 aio_nr_sub(ctx->max_reqs);
797 err_ctx:
798 atomic_set(&ctx->dead, 1);
799 if (ctx->mmap_size)
800 vm_munmap(ctx->mmap_base, ctx->mmap_size);
801 aio_free_ring(ctx);
802 err:
803 mutex_unlock(&ctx->ring_lock);
804 free_percpu(ctx->cpu);
805 percpu_ref_exit(&ctx->reqs);
806 percpu_ref_exit(&ctx->users);
807 kmem_cache_free(kioctx_cachep, ctx);
808 pr_debug("error allocating ioctx %d\n", err);
809 return ERR_PTR(err);
810 }
811
812 /* kill_ioctx
813 * Cancels all outstanding aio requests on an aio context. Used
814 * when the processes owning a context have all exited to encourage
815 * the rapid destruction of the kioctx.
816 */
817 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
818 struct ctx_rq_wait *wait)
819 {
820 struct kioctx_table *table;
821
822 spin_lock(&mm->ioctx_lock);
823 if (atomic_xchg(&ctx->dead, 1)) {
824 spin_unlock(&mm->ioctx_lock);
825 return -EINVAL;
826 }
827
828 table = rcu_dereference_raw(mm->ioctx_table);
829 WARN_ON(ctx != table->table[ctx->id]);
830 table->table[ctx->id] = NULL;
831 spin_unlock(&mm->ioctx_lock);
832
833 /* percpu_ref_kill() will do the necessary call_rcu() */
834 wake_up_all(&ctx->wait);
835
836 /*
837 * It'd be more correct to do this in free_ioctx(), after all
838 * the outstanding kiocbs have finished - but by then io_destroy
839 * has already returned, so io_setup() could potentially return
840 * -EAGAIN with no ioctxs actually in use (as far as userspace
841 * could tell).
842 */
843 aio_nr_sub(ctx->max_reqs);
844
845 if (ctx->mmap_size)
846 vm_munmap(ctx->mmap_base, ctx->mmap_size);
847
848 ctx->rq_wait = wait;
849 percpu_ref_kill(&ctx->users);
850 return 0;
851 }
852
853 /*
854 * exit_aio: called when the last user of mm goes away. At this point, there is
855 * no way for any new requests to be submited or any of the io_* syscalls to be
856 * called on the context.
857 *
858 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
859 * them.
860 */
861 void exit_aio(struct mm_struct *mm)
862 {
863 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
864 struct ctx_rq_wait wait;
865 int i, skipped;
866
867 if (!table)
868 return;
869
870 atomic_set(&wait.count, table->nr);
871 init_completion(&wait.comp);
872
873 skipped = 0;
874 for (i = 0; i < table->nr; ++i) {
875 struct kioctx *ctx = table->table[i];
876
877 if (!ctx) {
878 skipped++;
879 continue;
880 }
881
882 /*
883 * We don't need to bother with munmap() here - exit_mmap(mm)
884 * is coming and it'll unmap everything. And we simply can't,
885 * this is not necessarily our ->mm.
886 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
887 * that it needs to unmap the area, just set it to 0.
888 */
889 ctx->mmap_size = 0;
890 kill_ioctx(mm, ctx, &wait);
891 }
892
893 if (!atomic_sub_and_test(skipped, &wait.count)) {
894 /* Wait until all IO for the context are done. */
895 wait_for_completion(&wait.comp);
896 }
897
898 RCU_INIT_POINTER(mm->ioctx_table, NULL);
899 kfree(table);
900 }
901
902 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
903 {
904 struct kioctx_cpu *kcpu;
905 unsigned long flags;
906
907 local_irq_save(flags);
908 kcpu = this_cpu_ptr(ctx->cpu);
909 kcpu->reqs_available += nr;
910
911 while (kcpu->reqs_available >= ctx->req_batch * 2) {
912 kcpu->reqs_available -= ctx->req_batch;
913 atomic_add(ctx->req_batch, &ctx->reqs_available);
914 }
915
916 local_irq_restore(flags);
917 }
918
919 static bool get_reqs_available(struct kioctx *ctx)
920 {
921 struct kioctx_cpu *kcpu;
922 bool ret = false;
923 unsigned long flags;
924
925 local_irq_save(flags);
926 kcpu = this_cpu_ptr(ctx->cpu);
927 if (!kcpu->reqs_available) {
928 int old, avail = atomic_read(&ctx->reqs_available);
929
930 do {
931 if (avail < ctx->req_batch)
932 goto out;
933
934 old = avail;
935 avail = atomic_cmpxchg(&ctx->reqs_available,
936 avail, avail - ctx->req_batch);
937 } while (avail != old);
938
939 kcpu->reqs_available += ctx->req_batch;
940 }
941
942 ret = true;
943 kcpu->reqs_available--;
944 out:
945 local_irq_restore(flags);
946 return ret;
947 }
948
949 /* refill_reqs_available
950 * Updates the reqs_available reference counts used for tracking the
951 * number of free slots in the completion ring. This can be called
952 * from aio_complete() (to optimistically update reqs_available) or
953 * from aio_get_req() (the we're out of events case). It must be
954 * called holding ctx->completion_lock.
955 */
956 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
957 unsigned tail)
958 {
959 unsigned events_in_ring, completed;
960
961 /* Clamp head since userland can write to it. */
962 head %= ctx->nr_events;
963 if (head <= tail)
964 events_in_ring = tail - head;
965 else
966 events_in_ring = ctx->nr_events - (head - tail);
967
968 completed = ctx->completed_events;
969 if (events_in_ring < completed)
970 completed -= events_in_ring;
971 else
972 completed = 0;
973
974 if (!completed)
975 return;
976
977 ctx->completed_events -= completed;
978 put_reqs_available(ctx, completed);
979 }
980
981 /* user_refill_reqs_available
982 * Called to refill reqs_available when aio_get_req() encounters an
983 * out of space in the completion ring.
984 */
985 static void user_refill_reqs_available(struct kioctx *ctx)
986 {
987 spin_lock_irq(&ctx->completion_lock);
988 if (ctx->completed_events) {
989 struct aio_ring *ring;
990 unsigned head;
991
992 /* Access of ring->head may race with aio_read_events_ring()
993 * here, but that's okay since whether we read the old version
994 * or the new version, and either will be valid. The important
995 * part is that head cannot pass tail since we prevent
996 * aio_complete() from updating tail by holding
997 * ctx->completion_lock. Even if head is invalid, the check
998 * against ctx->completed_events below will make sure we do the
999 * safe/right thing.
1000 */
1001 ring = kmap_atomic(ctx->ring_pages[0]);
1002 head = ring->head;
1003 kunmap_atomic(ring);
1004
1005 refill_reqs_available(ctx, head, ctx->tail);
1006 }
1007
1008 spin_unlock_irq(&ctx->completion_lock);
1009 }
1010
1011 /* aio_get_req
1012 * Allocate a slot for an aio request.
1013 * Returns NULL if no requests are free.
1014 */
1015 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1016 {
1017 struct aio_kiocb *req;
1018
1019 if (!get_reqs_available(ctx)) {
1020 user_refill_reqs_available(ctx);
1021 if (!get_reqs_available(ctx))
1022 return NULL;
1023 }
1024
1025 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1026 if (unlikely(!req))
1027 goto out_put;
1028
1029 percpu_ref_get(&ctx->reqs);
1030
1031 req->ki_ctx = ctx;
1032 return req;
1033 out_put:
1034 put_reqs_available(ctx, 1);
1035 return NULL;
1036 }
1037
1038 static void kiocb_free(struct aio_kiocb *req)
1039 {
1040 if (req->common.ki_filp)
1041 fput(req->common.ki_filp);
1042 if (req->ki_eventfd != NULL)
1043 eventfd_ctx_put(req->ki_eventfd);
1044 kmem_cache_free(kiocb_cachep, req);
1045 }
1046
1047 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1048 {
1049 struct aio_ring __user *ring = (void __user *)ctx_id;
1050 struct mm_struct *mm = current->mm;
1051 struct kioctx *ctx, *ret = NULL;
1052 struct kioctx_table *table;
1053 unsigned id;
1054
1055 if (get_user(id, &ring->id))
1056 return NULL;
1057
1058 rcu_read_lock();
1059 table = rcu_dereference(mm->ioctx_table);
1060
1061 if (!table || id >= table->nr)
1062 goto out;
1063
1064 ctx = table->table[id];
1065 if (ctx && ctx->user_id == ctx_id) {
1066 percpu_ref_get(&ctx->users);
1067 ret = ctx;
1068 }
1069 out:
1070 rcu_read_unlock();
1071 return ret;
1072 }
1073
1074 /* aio_complete
1075 * Called when the io request on the given iocb is complete.
1076 */
1077 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1078 {
1079 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1080 struct kioctx *ctx = iocb->ki_ctx;
1081 struct aio_ring *ring;
1082 struct io_event *ev_page, *event;
1083 unsigned tail, pos, head;
1084 unsigned long flags;
1085
1086 if (kiocb->ki_flags & IOCB_WRITE) {
1087 struct file *file = kiocb->ki_filp;
1088
1089 /*
1090 * Tell lockdep we inherited freeze protection from submission
1091 * thread.
1092 */
1093 if (S_ISREG(file_inode(file)->i_mode))
1094 __sb_writers_acquired(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1095 file_end_write(file);
1096 }
1097
1098 /*
1099 * Special case handling for sync iocbs:
1100 * - events go directly into the iocb for fast handling
1101 * - the sync task with the iocb in its stack holds the single iocb
1102 * ref, no other paths have a way to get another ref
1103 * - the sync task helpfully left a reference to itself in the iocb
1104 */
1105 BUG_ON(is_sync_kiocb(kiocb));
1106
1107 if (iocb->ki_list.next) {
1108 unsigned long flags;
1109
1110 spin_lock_irqsave(&ctx->ctx_lock, flags);
1111 list_del(&iocb->ki_list);
1112 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1113 }
1114
1115 /*
1116 * Add a completion event to the ring buffer. Must be done holding
1117 * ctx->completion_lock to prevent other code from messing with the tail
1118 * pointer since we might be called from irq context.
1119 */
1120 spin_lock_irqsave(&ctx->completion_lock, flags);
1121
1122 tail = ctx->tail;
1123 pos = tail + AIO_EVENTS_OFFSET;
1124
1125 if (++tail >= ctx->nr_events)
1126 tail = 0;
1127
1128 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1129 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1130
1131 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1132 event->data = iocb->ki_user_data;
1133 event->res = res;
1134 event->res2 = res2;
1135
1136 kunmap_atomic(ev_page);
1137 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1138
1139 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1140 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1141 res, res2);
1142
1143 /* after flagging the request as done, we
1144 * must never even look at it again
1145 */
1146 smp_wmb(); /* make event visible before updating tail */
1147
1148 ctx->tail = tail;
1149
1150 ring = kmap_atomic(ctx->ring_pages[0]);
1151 head = ring->head;
1152 ring->tail = tail;
1153 kunmap_atomic(ring);
1154 flush_dcache_page(ctx->ring_pages[0]);
1155
1156 ctx->completed_events++;
1157 if (ctx->completed_events > 1)
1158 refill_reqs_available(ctx, head, tail);
1159 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1160
1161 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1162
1163 /*
1164 * Check if the user asked us to deliver the result through an
1165 * eventfd. The eventfd_signal() function is safe to be called
1166 * from IRQ context.
1167 */
1168 if (iocb->ki_eventfd != NULL)
1169 eventfd_signal(iocb->ki_eventfd, 1);
1170
1171 /* everything turned out well, dispose of the aiocb. */
1172 kiocb_free(iocb);
1173
1174 /*
1175 * We have to order our ring_info tail store above and test
1176 * of the wait list below outside the wait lock. This is
1177 * like in wake_up_bit() where clearing a bit has to be
1178 * ordered with the unlocked test.
1179 */
1180 smp_mb();
1181
1182 if (waitqueue_active(&ctx->wait))
1183 wake_up(&ctx->wait);
1184
1185 percpu_ref_put(&ctx->reqs);
1186 }
1187
1188 /* aio_read_events_ring
1189 * Pull an event off of the ioctx's event ring. Returns the number of
1190 * events fetched
1191 */
1192 static long aio_read_events_ring(struct kioctx *ctx,
1193 struct io_event __user *event, long nr)
1194 {
1195 struct aio_ring *ring;
1196 unsigned head, tail, pos;
1197 long ret = 0;
1198 int copy_ret;
1199
1200 /*
1201 * The mutex can block and wake us up and that will cause
1202 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1203 * and repeat. This should be rare enough that it doesn't cause
1204 * peformance issues. See the comment in read_events() for more detail.
1205 */
1206 sched_annotate_sleep();
1207 mutex_lock(&ctx->ring_lock);
1208
1209 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1210 ring = kmap_atomic(ctx->ring_pages[0]);
1211 head = ring->head;
1212 tail = ring->tail;
1213 kunmap_atomic(ring);
1214
1215 /*
1216 * Ensure that once we've read the current tail pointer, that
1217 * we also see the events that were stored up to the tail.
1218 */
1219 smp_rmb();
1220
1221 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1222
1223 if (head == tail)
1224 goto out;
1225
1226 head %= ctx->nr_events;
1227 tail %= ctx->nr_events;
1228
1229 while (ret < nr) {
1230 long avail;
1231 struct io_event *ev;
1232 struct page *page;
1233
1234 avail = (head <= tail ? tail : ctx->nr_events) - head;
1235 if (head == tail)
1236 break;
1237
1238 avail = min(avail, nr - ret);
1239 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1240 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1241
1242 pos = head + AIO_EVENTS_OFFSET;
1243 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1244 pos %= AIO_EVENTS_PER_PAGE;
1245
1246 ev = kmap(page);
1247 copy_ret = copy_to_user(event + ret, ev + pos,
1248 sizeof(*ev) * avail);
1249 kunmap(page);
1250
1251 if (unlikely(copy_ret)) {
1252 ret = -EFAULT;
1253 goto out;
1254 }
1255
1256 ret += avail;
1257 head += avail;
1258 head %= ctx->nr_events;
1259 }
1260
1261 ring = kmap_atomic(ctx->ring_pages[0]);
1262 ring->head = head;
1263 kunmap_atomic(ring);
1264 flush_dcache_page(ctx->ring_pages[0]);
1265
1266 pr_debug("%li h%u t%u\n", ret, head, tail);
1267 out:
1268 mutex_unlock(&ctx->ring_lock);
1269
1270 return ret;
1271 }
1272
1273 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1274 struct io_event __user *event, long *i)
1275 {
1276 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1277
1278 if (ret > 0)
1279 *i += ret;
1280
1281 if (unlikely(atomic_read(&ctx->dead)))
1282 ret = -EINVAL;
1283
1284 if (!*i)
1285 *i = ret;
1286
1287 return ret < 0 || *i >= min_nr;
1288 }
1289
1290 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1291 struct io_event __user *event,
1292 struct timespec __user *timeout)
1293 {
1294 ktime_t until = KTIME_MAX;
1295 long ret = 0;
1296
1297 if (timeout) {
1298 struct timespec ts;
1299
1300 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1301 return -EFAULT;
1302
1303 until = timespec_to_ktime(ts);
1304 }
1305
1306 /*
1307 * Note that aio_read_events() is being called as the conditional - i.e.
1308 * we're calling it after prepare_to_wait() has set task state to
1309 * TASK_INTERRUPTIBLE.
1310 *
1311 * But aio_read_events() can block, and if it blocks it's going to flip
1312 * the task state back to TASK_RUNNING.
1313 *
1314 * This should be ok, provided it doesn't flip the state back to
1315 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1316 * will only happen if the mutex_lock() call blocks, and we then find
1317 * the ringbuffer empty. So in practice we should be ok, but it's
1318 * something to be aware of when touching this code.
1319 */
1320 if (until == 0)
1321 aio_read_events(ctx, min_nr, nr, event, &ret);
1322 else
1323 wait_event_interruptible_hrtimeout(ctx->wait,
1324 aio_read_events(ctx, min_nr, nr, event, &ret),
1325 until);
1326
1327 if (!ret && signal_pending(current))
1328 ret = -EINTR;
1329
1330 return ret;
1331 }
1332
1333 /* sys_io_setup:
1334 * Create an aio_context capable of receiving at least nr_events.
1335 * ctxp must not point to an aio_context that already exists, and
1336 * must be initialized to 0 prior to the call. On successful
1337 * creation of the aio_context, *ctxp is filled in with the resulting
1338 * handle. May fail with -EINVAL if *ctxp is not initialized,
1339 * if the specified nr_events exceeds internal limits. May fail
1340 * with -EAGAIN if the specified nr_events exceeds the user's limit
1341 * of available events. May fail with -ENOMEM if insufficient kernel
1342 * resources are available. May fail with -EFAULT if an invalid
1343 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1344 * implemented.
1345 */
1346 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1347 {
1348 struct kioctx *ioctx = NULL;
1349 unsigned long ctx;
1350 long ret;
1351
1352 ret = get_user(ctx, ctxp);
1353 if (unlikely(ret))
1354 goto out;
1355
1356 ret = -EINVAL;
1357 if (unlikely(ctx || nr_events == 0)) {
1358 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1359 ctx, nr_events);
1360 goto out;
1361 }
1362
1363 ioctx = ioctx_alloc(nr_events);
1364 ret = PTR_ERR(ioctx);
1365 if (!IS_ERR(ioctx)) {
1366 ret = put_user(ioctx->user_id, ctxp);
1367 if (ret)
1368 kill_ioctx(current->mm, ioctx, NULL);
1369 percpu_ref_put(&ioctx->users);
1370 }
1371
1372 out:
1373 return ret;
1374 }
1375
1376 #ifdef CONFIG_COMPAT
1377 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1378 {
1379 struct kioctx *ioctx = NULL;
1380 unsigned long ctx;
1381 long ret;
1382
1383 ret = get_user(ctx, ctx32p);
1384 if (unlikely(ret))
1385 goto out;
1386
1387 ret = -EINVAL;
1388 if (unlikely(ctx || nr_events == 0)) {
1389 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1390 ctx, nr_events);
1391 goto out;
1392 }
1393
1394 ioctx = ioctx_alloc(nr_events);
1395 ret = PTR_ERR(ioctx);
1396 if (!IS_ERR(ioctx)) {
1397 /* truncating is ok because it's a user address */
1398 ret = put_user((u32)ioctx->user_id, ctx32p);
1399 if (ret)
1400 kill_ioctx(current->mm, ioctx, NULL);
1401 percpu_ref_put(&ioctx->users);
1402 }
1403
1404 out:
1405 return ret;
1406 }
1407 #endif
1408
1409 /* sys_io_destroy:
1410 * Destroy the aio_context specified. May cancel any outstanding
1411 * AIOs and block on completion. Will fail with -ENOSYS if not
1412 * implemented. May fail with -EINVAL if the context pointed to
1413 * is invalid.
1414 */
1415 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1416 {
1417 struct kioctx *ioctx = lookup_ioctx(ctx);
1418 if (likely(NULL != ioctx)) {
1419 struct ctx_rq_wait wait;
1420 int ret;
1421
1422 init_completion(&wait.comp);
1423 atomic_set(&wait.count, 1);
1424
1425 /* Pass requests_done to kill_ioctx() where it can be set
1426 * in a thread-safe way. If we try to set it here then we have
1427 * a race condition if two io_destroy() called simultaneously.
1428 */
1429 ret = kill_ioctx(current->mm, ioctx, &wait);
1430 percpu_ref_put(&ioctx->users);
1431
1432 /* Wait until all IO for the context are done. Otherwise kernel
1433 * keep using user-space buffers even if user thinks the context
1434 * is destroyed.
1435 */
1436 if (!ret)
1437 wait_for_completion(&wait.comp);
1438
1439 return ret;
1440 }
1441 pr_debug("EINVAL: invalid context id\n");
1442 return -EINVAL;
1443 }
1444
1445 static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec,
1446 bool vectored, bool compat, struct iov_iter *iter)
1447 {
1448 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1449 size_t len = iocb->aio_nbytes;
1450
1451 if (!vectored) {
1452 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1453 *iovec = NULL;
1454 return ret;
1455 }
1456 #ifdef CONFIG_COMPAT
1457 if (compat)
1458 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1459 iter);
1460 #endif
1461 return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1462 }
1463
1464 static inline ssize_t aio_ret(struct kiocb *req, ssize_t ret)
1465 {
1466 switch (ret) {
1467 case -EIOCBQUEUED:
1468 return ret;
1469 case -ERESTARTSYS:
1470 case -ERESTARTNOINTR:
1471 case -ERESTARTNOHAND:
1472 case -ERESTART_RESTARTBLOCK:
1473 /*
1474 * There's no easy way to restart the syscall since other AIO's
1475 * may be already running. Just fail this IO with EINTR.
1476 */
1477 ret = -EINTR;
1478 /*FALLTHRU*/
1479 default:
1480 aio_complete(req, ret, 0);
1481 return 0;
1482 }
1483 }
1484
1485 static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored,
1486 bool compat)
1487 {
1488 struct file *file = req->ki_filp;
1489 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1490 struct iov_iter iter;
1491 ssize_t ret;
1492
1493 if (unlikely(!(file->f_mode & FMODE_READ)))
1494 return -EBADF;
1495 if (unlikely(!file->f_op->read_iter))
1496 return -EINVAL;
1497
1498 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1499 if (ret)
1500 return ret;
1501 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1502 if (!ret)
1503 ret = aio_ret(req, call_read_iter(file, req, &iter));
1504 kfree(iovec);
1505 return ret;
1506 }
1507
1508 static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored,
1509 bool compat)
1510 {
1511 struct file *file = req->ki_filp;
1512 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1513 struct iov_iter iter;
1514 ssize_t ret;
1515
1516 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1517 return -EBADF;
1518 if (unlikely(!file->f_op->write_iter))
1519 return -EINVAL;
1520
1521 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1522 if (ret)
1523 return ret;
1524 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1525 if (!ret) {
1526 req->ki_flags |= IOCB_WRITE;
1527 file_start_write(file);
1528 ret = aio_ret(req, call_write_iter(file, req, &iter));
1529 /*
1530 * We release freeze protection in aio_complete(). Fool lockdep
1531 * by telling it the lock got released so that it doesn't
1532 * complain about held lock when we return to userspace.
1533 */
1534 if (S_ISREG(file_inode(file)->i_mode))
1535 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1536 }
1537 kfree(iovec);
1538 return ret;
1539 }
1540
1541 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1542 struct iocb *iocb, bool compat)
1543 {
1544 struct aio_kiocb *req;
1545 struct file *file;
1546 ssize_t ret;
1547
1548 /* enforce forwards compatibility on users */
1549 if (unlikely(iocb->aio_reserved2)) {
1550 pr_debug("EINVAL: reserve field set\n");
1551 return -EINVAL;
1552 }
1553
1554 /* prevent overflows */
1555 if (unlikely(
1556 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1557 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1558 ((ssize_t)iocb->aio_nbytes < 0)
1559 )) {
1560 pr_debug("EINVAL: overflow check\n");
1561 return -EINVAL;
1562 }
1563
1564 req = aio_get_req(ctx);
1565 if (unlikely(!req))
1566 return -EAGAIN;
1567
1568 req->common.ki_filp = file = fget(iocb->aio_fildes);
1569 if (unlikely(!req->common.ki_filp)) {
1570 ret = -EBADF;
1571 goto out_put_req;
1572 }
1573 req->common.ki_pos = iocb->aio_offset;
1574 req->common.ki_complete = aio_complete;
1575 req->common.ki_flags = iocb_flags(req->common.ki_filp);
1576 req->common.ki_hint = file_write_hint(file);
1577
1578 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1579 /*
1580 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1581 * instance of the file* now. The file descriptor must be
1582 * an eventfd() fd, and will be signaled for each completed
1583 * event using the eventfd_signal() function.
1584 */
1585 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1586 if (IS_ERR(req->ki_eventfd)) {
1587 ret = PTR_ERR(req->ki_eventfd);
1588 req->ki_eventfd = NULL;
1589 goto out_put_req;
1590 }
1591
1592 req->common.ki_flags |= IOCB_EVENTFD;
1593 }
1594
1595 ret = kiocb_set_rw_flags(&req->common, iocb->aio_rw_flags);
1596 if (unlikely(ret)) {
1597 pr_debug("EINVAL: aio_rw_flags\n");
1598 goto out_put_req;
1599 }
1600
1601 if ((req->common.ki_flags & IOCB_NOWAIT) &&
1602 !(req->common.ki_flags & IOCB_DIRECT)) {
1603 ret = -EOPNOTSUPP;
1604 goto out_put_req;
1605 }
1606
1607 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1608 if (unlikely(ret)) {
1609 pr_debug("EFAULT: aio_key\n");
1610 goto out_put_req;
1611 }
1612
1613 req->ki_user_iocb = user_iocb;
1614 req->ki_user_data = iocb->aio_data;
1615
1616 get_file(file);
1617 switch (iocb->aio_lio_opcode) {
1618 case IOCB_CMD_PREAD:
1619 ret = aio_read(&req->common, iocb, false, compat);
1620 break;
1621 case IOCB_CMD_PWRITE:
1622 ret = aio_write(&req->common, iocb, false, compat);
1623 break;
1624 case IOCB_CMD_PREADV:
1625 ret = aio_read(&req->common, iocb, true, compat);
1626 break;
1627 case IOCB_CMD_PWRITEV:
1628 ret = aio_write(&req->common, iocb, true, compat);
1629 break;
1630 default:
1631 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1632 ret = -EINVAL;
1633 break;
1634 }
1635 fput(file);
1636
1637 if (ret && ret != -EIOCBQUEUED)
1638 goto out_put_req;
1639 return 0;
1640 out_put_req:
1641 put_reqs_available(ctx, 1);
1642 percpu_ref_put(&ctx->reqs);
1643 kiocb_free(req);
1644 return ret;
1645 }
1646
1647 static long do_io_submit(aio_context_t ctx_id, long nr,
1648 struct iocb __user *__user *iocbpp, bool compat)
1649 {
1650 struct kioctx *ctx;
1651 long ret = 0;
1652 int i = 0;
1653 struct blk_plug plug;
1654
1655 if (unlikely(nr < 0))
1656 return -EINVAL;
1657
1658 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1659 nr = LONG_MAX/sizeof(*iocbpp);
1660
1661 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1662 return -EFAULT;
1663
1664 ctx = lookup_ioctx(ctx_id);
1665 if (unlikely(!ctx)) {
1666 pr_debug("EINVAL: invalid context id\n");
1667 return -EINVAL;
1668 }
1669
1670 blk_start_plug(&plug);
1671
1672 /*
1673 * AKPM: should this return a partial result if some of the IOs were
1674 * successfully submitted?
1675 */
1676 for (i=0; i<nr; i++) {
1677 struct iocb __user *user_iocb;
1678 struct iocb tmp;
1679
1680 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1681 ret = -EFAULT;
1682 break;
1683 }
1684
1685 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1686 ret = -EFAULT;
1687 break;
1688 }
1689
1690 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1691 if (ret)
1692 break;
1693 }
1694 blk_finish_plug(&plug);
1695
1696 percpu_ref_put(&ctx->users);
1697 return i ? i : ret;
1698 }
1699
1700 /* sys_io_submit:
1701 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1702 * the number of iocbs queued. May return -EINVAL if the aio_context
1703 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1704 * *iocbpp[0] is not properly initialized, if the operation specified
1705 * is invalid for the file descriptor in the iocb. May fail with
1706 * -EFAULT if any of the data structures point to invalid data. May
1707 * fail with -EBADF if the file descriptor specified in the first
1708 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1709 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1710 * fail with -ENOSYS if not implemented.
1711 */
1712 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1713 struct iocb __user * __user *, iocbpp)
1714 {
1715 return do_io_submit(ctx_id, nr, iocbpp, 0);
1716 }
1717
1718 #ifdef CONFIG_COMPAT
1719 static inline long
1720 copy_iocb(long nr, u32 __user *ptr32, struct iocb __user * __user *ptr64)
1721 {
1722 compat_uptr_t uptr;
1723 int i;
1724
1725 for (i = 0; i < nr; ++i) {
1726 if (get_user(uptr, ptr32 + i))
1727 return -EFAULT;
1728 if (put_user(compat_ptr(uptr), ptr64 + i))
1729 return -EFAULT;
1730 }
1731 return 0;
1732 }
1733
1734 #define MAX_AIO_SUBMITS (PAGE_SIZE/sizeof(struct iocb *))
1735
1736 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1737 int, nr, u32 __user *, iocb)
1738 {
1739 struct iocb __user * __user *iocb64;
1740 long ret;
1741
1742 if (unlikely(nr < 0))
1743 return -EINVAL;
1744
1745 if (nr > MAX_AIO_SUBMITS)
1746 nr = MAX_AIO_SUBMITS;
1747
1748 iocb64 = compat_alloc_user_space(nr * sizeof(*iocb64));
1749 ret = copy_iocb(nr, iocb, iocb64);
1750 if (!ret)
1751 ret = do_io_submit(ctx_id, nr, iocb64, 1);
1752 return ret;
1753 }
1754 #endif
1755
1756 /* lookup_kiocb
1757 * Finds a given iocb for cancellation.
1758 */
1759 static struct aio_kiocb *
1760 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1761 {
1762 struct aio_kiocb *kiocb;
1763
1764 assert_spin_locked(&ctx->ctx_lock);
1765
1766 if (key != KIOCB_KEY)
1767 return NULL;
1768
1769 /* TODO: use a hash or array, this sucks. */
1770 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1771 if (kiocb->ki_user_iocb == iocb)
1772 return kiocb;
1773 }
1774 return NULL;
1775 }
1776
1777 /* sys_io_cancel:
1778 * Attempts to cancel an iocb previously passed to io_submit. If
1779 * the operation is successfully cancelled, the resulting event is
1780 * copied into the memory pointed to by result without being placed
1781 * into the completion queue and 0 is returned. May fail with
1782 * -EFAULT if any of the data structures pointed to are invalid.
1783 * May fail with -EINVAL if aio_context specified by ctx_id is
1784 * invalid. May fail with -EAGAIN if the iocb specified was not
1785 * cancelled. Will fail with -ENOSYS if not implemented.
1786 */
1787 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1788 struct io_event __user *, result)
1789 {
1790 struct kioctx *ctx;
1791 struct aio_kiocb *kiocb;
1792 u32 key;
1793 int ret;
1794
1795 ret = get_user(key, &iocb->aio_key);
1796 if (unlikely(ret))
1797 return -EFAULT;
1798
1799 ctx = lookup_ioctx(ctx_id);
1800 if (unlikely(!ctx))
1801 return -EINVAL;
1802
1803 spin_lock_irq(&ctx->ctx_lock);
1804
1805 kiocb = lookup_kiocb(ctx, iocb, key);
1806 if (kiocb)
1807 ret = kiocb_cancel(kiocb);
1808 else
1809 ret = -EINVAL;
1810
1811 spin_unlock_irq(&ctx->ctx_lock);
1812
1813 if (!ret) {
1814 /*
1815 * The result argument is no longer used - the io_event is
1816 * always delivered via the ring buffer. -EINPROGRESS indicates
1817 * cancellation is progress:
1818 */
1819 ret = -EINPROGRESS;
1820 }
1821
1822 percpu_ref_put(&ctx->users);
1823
1824 return ret;
1825 }
1826
1827 /* io_getevents:
1828 * Attempts to read at least min_nr events and up to nr events from
1829 * the completion queue for the aio_context specified by ctx_id. If
1830 * it succeeds, the number of read events is returned. May fail with
1831 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1832 * out of range, if timeout is out of range. May fail with -EFAULT
1833 * if any of the memory specified is invalid. May return 0 or
1834 * < min_nr if the timeout specified by timeout has elapsed
1835 * before sufficient events are available, where timeout == NULL
1836 * specifies an infinite timeout. Note that the timeout pointed to by
1837 * timeout is relative. Will fail with -ENOSYS if not implemented.
1838 */
1839 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1840 long, min_nr,
1841 long, nr,
1842 struct io_event __user *, events,
1843 struct timespec __user *, timeout)
1844 {
1845 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1846 long ret = -EINVAL;
1847
1848 if (likely(ioctx)) {
1849 if (likely(min_nr <= nr && min_nr >= 0))
1850 ret = read_events(ioctx, min_nr, nr, events, timeout);
1851 percpu_ref_put(&ioctx->users);
1852 }
1853 return ret;
1854 }
1855
1856 #ifdef CONFIG_COMPAT
1857 COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id,
1858 compat_long_t, min_nr,
1859 compat_long_t, nr,
1860 struct io_event __user *, events,
1861 struct compat_timespec __user *, timeout)
1862 {
1863 struct timespec t;
1864 struct timespec __user *ut = NULL;
1865
1866 if (timeout) {
1867 if (compat_get_timespec(&t, timeout))
1868 return -EFAULT;
1869
1870 ut = compat_alloc_user_space(sizeof(*ut));
1871 if (copy_to_user(ut, &t, sizeof(t)))
1872 return -EFAULT;
1873 }
1874 return sys_io_getevents(ctx_id, min_nr, nr, events, ut);
1875 }
1876 #endif