projects/linux·blob master · refs

2503 lines · 62.7 KiB

1/*
2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <[email protected]>
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/blk_plug.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
56struct 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
80struct kioctx_table {
81 struct rcu_head rcu;
82 unsigned nr;
83 struct kioctx __rcu *table[] __counted_by(nr);
84};
85
86struct kioctx_cpu {
87 unsigned reqs_available;
88};
89
90struct ctx_rq_wait {
91 struct completion comp;
92 atomic_t count;
93};
94
95struct 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 kioctx_cpu __percpu *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 folio **ring_folios;
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 folio *internal_folios[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 */
173struct fsync_iocb {
174 struct file *file;
175 struct work_struct work;
176 bool datasync;
177 struct cred *creds;
178};
179
180struct 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 */
197struct 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
221struct aio_inode_info {
222 struct inode vfs_inode;
223 spinlock_t migrate_lock;
224 struct kioctx *ctx;
225};
226
227static inline struct aio_inode_info *AIO_I(struct inode *inode)
228{
229 return container_of(inode, struct aio_inode_info, vfs_inode);
230}
231
232/*------ sysctl variables----*/
233static DEFINE_SPINLOCK(aio_nr_lock);
234static unsigned long aio_nr; /* current system wide number of aio requests */
235static unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
236/*----end sysctl variables---*/
237#ifdef CONFIG_SYSCTL
238static const struct ctl_table aio_sysctls[] = {
239 {
240 .procname = "aio-nr",
241 .data = &aio_nr,
242 .maxlen = sizeof(aio_nr),
243 .mode = 0444,
244 .proc_handler = proc_doulongvec_minmax,
245 },
246 {
247 .procname = "aio-max-nr",
248 .data = &aio_max_nr,
249 .maxlen = sizeof(aio_max_nr),
250 .mode = 0644,
251 .proc_handler = proc_doulongvec_minmax,
252 },
253};
254
255static void __init aio_sysctl_init(void)
256{
257 register_sysctl_init("fs", aio_sysctls);
258}
259#else
260#define aio_sysctl_init() do { } while (0)
261#endif
262
263static struct kmem_cache *kiocb_cachep;
264static struct kmem_cache *kioctx_cachep;
265static struct kmem_cache *aio_inode_cachep;
266
267static struct vfsmount *aio_mnt;
268
269static const struct file_operations aio_ring_fops;
270static const struct address_space_operations aio_ctx_aops;
271
272static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
273{
274 struct file *file;
275 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
276
277 if (IS_ERR(inode))
278 return ERR_CAST(inode);
279
280 inode->i_mapping->a_ops = &aio_ctx_aops;
281 AIO_I(inode)->ctx = ctx;
282 inode->i_size = PAGE_SIZE * nr_pages;
283
284 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
285 O_RDWR, &aio_ring_fops);
286 if (IS_ERR(file))
287 iput(inode);
288 return file;
289}
290
291static struct inode *aio_alloc_inode(struct super_block *sb)
292{
293 struct aio_inode_info *ai;
294
295 ai = alloc_inode_sb(sb, aio_inode_cachep, GFP_KERNEL);
296 if (!ai)
297 return NULL;
298 ai->ctx = NULL;
299
300 return &ai->vfs_inode;
301}
302
303static void aio_free_inode(struct inode *inode)
304{
305 kmem_cache_free(aio_inode_cachep, AIO_I(inode));
306}
307
308static const struct super_operations aio_super_operations = {
309 .alloc_inode = aio_alloc_inode,
310 .free_inode = aio_free_inode,
311 .statfs = simple_statfs,
312};
313
314static int aio_init_fs_context(struct fs_context *fc)
315{
316 struct pseudo_fs_context *pfc;
317
318 pfc = init_pseudo(fc, AIO_RING_MAGIC);
319 if (!pfc)
320 return -ENOMEM;
321 pfc->ops = &aio_super_operations;
322 return 0;
323}
324
325static void init_once(void *obj)
326{
327 struct aio_inode_info *ai = obj;
328
329 inode_init_once(&ai->vfs_inode);
330 spin_lock_init(&ai->migrate_lock);
331}
332
333/* aio_setup
334 * Creates the slab caches used by the aio routines, panic on
335 * failure as this is done early during the boot sequence.
336 */
337static int __init aio_setup(void)
338{
339 static struct file_system_type aio_fs = {
340 .name = "aio",
341 .init_fs_context = aio_init_fs_context,
342 .kill_sb = kill_anon_super,
343 };
344
345 aio_inode_cachep = kmem_cache_create("aio_inode_cache",
346 sizeof(struct aio_inode_info), 0,
347 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_ACCOUNT),
348 init_once);
349 aio_mnt = kern_mount(&aio_fs);
350 if (IS_ERR(aio_mnt))
351 panic("Failed to create aio fs mount.");
352
353 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
354 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
355 aio_sysctl_init();
356 return 0;
357}
358__initcall(aio_setup);
359
360static void put_aio_ring_file(struct kioctx *ctx)
361{
362 struct file *aio_ring_file = ctx->aio_ring_file;
363
364 if (aio_ring_file) {
365 struct inode *inode = file_inode(aio_ring_file);
366
367 truncate_setsize(inode, 0);
368
369 /* Prevent further access to the kioctx from migratepages */
370 spin_lock(&AIO_I(inode)->migrate_lock);
371 AIO_I(inode)->ctx = NULL;
372 ctx->aio_ring_file = NULL;
373 spin_unlock(&AIO_I(inode)->migrate_lock);
374
375 fput(aio_ring_file);
376 }
377}
378
379static void aio_free_ring(struct kioctx *ctx)
380{
381 int i;
382
383 /* Disconnect the kiotx from the ring file. This prevents future
384 * accesses to the kioctx from page migration.
385 */
386 put_aio_ring_file(ctx);
387
388 for (i = 0; i < ctx->nr_pages; i++) {
389 struct folio *folio = ctx->ring_folios[i];
390
391 if (!folio)
392 continue;
393
394 pr_debug("pid(%d) [%d] folio->count=%d\n", current->pid, i,
395 folio_ref_count(folio));
396 ctx->ring_folios[i] = NULL;
397 folio_put(folio);
398 }
399
400 if (ctx->ring_folios && ctx->ring_folios != ctx->internal_folios) {
401 kfree(ctx->ring_folios);
402 ctx->ring_folios = NULL;
403 }
404}
405
406static int aio_ring_mremap(struct vm_area_struct *vma)
407{
408 struct file *file = vma->vm_file;
409 struct mm_struct *mm = vma->vm_mm;
410 struct kioctx_table *table;
411 int i, res = -EINVAL;
412
413 spin_lock(&mm->ioctx_lock);
414 rcu_read_lock();
415 table = rcu_dereference(mm->ioctx_table);
416 if (!table)
417 goto out_unlock;
418
419 for (i = 0; i < table->nr; i++) {
420 struct kioctx *ctx;
421
422 ctx = rcu_dereference(table->table[i]);
423 if (ctx && ctx->aio_ring_file == file) {
424 if (!atomic_read(&ctx->dead) &&
425 (ctx->mmap_size == (vma->vm_end - vma->vm_start))) {
426 ctx->user_id = ctx->mmap_base = vma->vm_start;
427 res = 0;
428 }
429 break;
430 }
431 }
432
433out_unlock:
434 rcu_read_unlock();
435 spin_unlock(&mm->ioctx_lock);
436 return res;
437}
438
439static const struct vm_operations_struct aio_ring_vm_ops = {
440 .mremap = aio_ring_mremap,
441#if IS_ENABLED(CONFIG_MMU)
442 .fault = filemap_fault,
443 .map_pages = filemap_map_pages,
444 .page_mkwrite = filemap_page_mkwrite,
445#endif
446};
447
448static int aio_ring_mmap_prepare(struct vm_area_desc *desc)
449{
450 vma_desc_set_flags(desc, VMA_DONTEXPAND_BIT, VMA_DONTCOPY_BIT);
451 desc->vm_ops = &aio_ring_vm_ops;
452 return 0;
453}
454
455static const struct file_operations aio_ring_fops = {
456 .mmap_prepare = aio_ring_mmap_prepare,
457};
458
459#if IS_ENABLED(CONFIG_MIGRATION)
460static int aio_migrate_folio(struct address_space *mapping, struct folio *dst,
461 struct folio *src, enum migrate_mode mode)
462{
463 struct kioctx *ctx;
464 struct aio_inode_info *ai = AIO_I(mapping->host);
465 unsigned long flags;
466 pgoff_t idx;
467 int rc = 0;
468
469 /* ai->migrate_lock here protects against the kioctx teardown. */
470 spin_lock(&ai->migrate_lock);
471 ctx = ai->ctx;
472 if (!ctx) {
473 rc = -EINVAL;
474 goto out;
475 }
476
477 /* The ring_lock mutex. The prevents aio_read_events() from writing
478 * to the ring's head, and prevents page migration from mucking in
479 * a partially initialized kiotx.
480 */
481 if (!mutex_trylock(&ctx->ring_lock)) {
482 rc = -EAGAIN;
483 goto out;
484 }
485
486 idx = src->index;
487 if (idx < (pgoff_t)ctx->nr_pages) {
488 /* Make sure the old folio hasn't already been changed */
489 if (ctx->ring_folios[idx] != src)
490 rc = -EAGAIN;
491 } else
492 rc = -EINVAL;
493
494 if (rc != 0)
495 goto out_unlock;
496
497 /* Writeback must be complete */
498 BUG_ON(folio_test_writeback(src));
499 folio_get(dst);
500
501 rc = folio_migrate_mapping(mapping, dst, src, 1);
502 if (rc) {
503 folio_put(dst);
504 goto out_unlock;
505 }
506
507 /* Take completion_lock to prevent other writes to the ring buffer
508 * while the old folio is copied to the new. This prevents new
509 * events from being lost.
510 */
511 spin_lock_irqsave(&ctx->completion_lock, flags);
512 folio_copy(dst, src);
513 folio_migrate_flags(dst, src);
514 BUG_ON(ctx->ring_folios[idx] != src);
515 ctx->ring_folios[idx] = dst;
516 spin_unlock_irqrestore(&ctx->completion_lock, flags);
517
518 /* The old folio is no longer accessible. */
519 folio_put(src);
520
521out_unlock:
522 mutex_unlock(&ctx->ring_lock);
523out:
524 spin_unlock(&ai->migrate_lock);
525 return rc;
526}
527#else
528#define aio_migrate_folio NULL
529#endif
530
531static const struct address_space_operations aio_ctx_aops = {
532 .dirty_folio = noop_dirty_folio,
533 .migrate_folio = aio_migrate_folio,
534};
535
536static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
537{
538 struct aio_ring *ring;
539 struct mm_struct *mm = current->mm;
540 unsigned long size, unused;
541 int nr_pages;
542 int i;
543 struct file *file;
544
545 /* Compensate for the ring buffer's head/tail overlap entry */
546 nr_events += 2; /* 1 is required, 2 for good luck */
547
548 size = sizeof(struct aio_ring);
549 size += sizeof(struct io_event) * nr_events;
550
551 nr_pages = PFN_UP(size);
552 if (nr_pages < 0)
553 return -EINVAL;
554
555 file = aio_private_file(ctx, nr_pages);
556 if (IS_ERR(file)) {
557 ctx->aio_ring_file = NULL;
558 return -ENOMEM;
559 }
560
561 ctx->aio_ring_file = file;
562 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
563 / sizeof(struct io_event);
564
565 ctx->ring_folios = ctx->internal_folios;
566 if (nr_pages > AIO_RING_PAGES) {
567 ctx->ring_folios = kzalloc_objs(struct folio *, nr_pages);
568 if (!ctx->ring_folios) {
569 put_aio_ring_file(ctx);
570 return -ENOMEM;
571 }
572 }
573
574 for (i = 0; i < nr_pages; i++) {
575 struct folio *folio;
576
577 folio = __filemap_get_folio(file->f_mapping, i,
578 FGP_LOCK | FGP_ACCESSED | FGP_CREAT,
579 GFP_USER | __GFP_ZERO);
580 if (IS_ERR(folio))
581 break;
582
583 pr_debug("pid(%d) [%d] folio->count=%d\n", current->pid, i,
584 folio_ref_count(folio));
585 folio_end_read(folio, true);
586
587 ctx->ring_folios[i] = folio;
588 }
589 ctx->nr_pages = i;
590
591 if (unlikely(i != nr_pages)) {
592 aio_free_ring(ctx);
593 return -ENOMEM;
594 }
595
596 ctx->mmap_size = nr_pages * PAGE_SIZE;
597 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
598
599 if (mmap_write_lock_killable(mm)) {
600 ctx->mmap_size = 0;
601 aio_free_ring(ctx);
602 return -EINTR;
603 }
604
605 ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
606 PROT_READ | PROT_WRITE,
607 MAP_SHARED, EMPTY_VMA_FLAGS, 0, &unused, NULL);
608 mmap_write_unlock(mm);
609 if (IS_ERR((void *)ctx->mmap_base)) {
610 ctx->mmap_size = 0;
611 aio_free_ring(ctx);
612 return -ENOMEM;
613 }
614
615 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
616
617 ctx->user_id = ctx->mmap_base;
618 ctx->nr_events = nr_events; /* trusted copy */
619
620 ring = folio_address(ctx->ring_folios[0]);
621 ring->nr = nr_events; /* user copy */
622 ring->id = ~0U;
623 ring->head = ring->tail = 0;
624 ring->magic = AIO_RING_MAGIC;
625 ring->compat_features = AIO_RING_COMPAT_FEATURES;
626 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
627 ring->header_length = sizeof(struct aio_ring);
628 flush_dcache_folio(ctx->ring_folios[0]);
629
630 return 0;
631}
632
633#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
634#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
635#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
636
637void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
638{
639 struct aio_kiocb *req;
640 struct kioctx *ctx;
641 unsigned long flags;
642
643 /*
644 * kiocb didn't come from aio or is neither a read nor a write, hence
645 * ignore it.
646 */
647 if (!(iocb->ki_flags & IOCB_AIO_RW))
648 return;
649
650 req = container_of(iocb, struct aio_kiocb, rw);
651
652 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
653 return;
654
655 ctx = req->ki_ctx;
656
657 spin_lock_irqsave(&ctx->ctx_lock, flags);
658 list_add_tail(&req->ki_list, &ctx->active_reqs);
659 req->ki_cancel = cancel;
660 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
661}
662EXPORT_SYMBOL(kiocb_set_cancel_fn);
663
664/*
665 * free_ioctx() should be RCU delayed to synchronize against the RCU
666 * protected lookup_ioctx() and also needs process context to call
667 * aio_free_ring(). Use rcu_work.
668 */
669static void free_ioctx(struct work_struct *work)
670{
671 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
672 free_rwork);
673 pr_debug("freeing %p\n", ctx);
674
675 aio_free_ring(ctx);
676 free_percpu(ctx->cpu);
677 percpu_ref_exit(&ctx->reqs);
678 percpu_ref_exit(&ctx->users);
679 kmem_cache_free(kioctx_cachep, ctx);
680}
681
682static void free_ioctx_reqs(struct percpu_ref *ref)
683{
684 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
685
686 /* At this point we know that there are no any in-flight requests */
687 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
688 complete(&ctx->rq_wait->comp);
689
690 /* Synchronize against RCU protected table->table[] dereferences */
691 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
692 queue_rcu_work(system_percpu_wq, &ctx->free_rwork);
693}
694
695/*
696 * When this function runs, the kioctx has been removed from the "hash table"
697 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
698 * now it's safe to cancel any that need to be.
699 */
700static void free_ioctx_users(struct percpu_ref *ref)
701{
702 struct kioctx *ctx = container_of(ref, struct kioctx, users);
703 struct aio_kiocb *req;
704
705 spin_lock_irq(&ctx->ctx_lock);
706
707 while (!list_empty(&ctx->active_reqs)) {
708 req = list_first_entry(&ctx->active_reqs,
709 struct aio_kiocb, ki_list);
710 req->ki_cancel(&req->rw);
711 list_del_init(&req->ki_list);
712 }
713
714 spin_unlock_irq(&ctx->ctx_lock);
715
716 percpu_ref_kill(&ctx->reqs);
717 percpu_ref_put(&ctx->reqs);
718}
719
720static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
721{
722 unsigned i, new_nr;
723 struct kioctx_table *table, *old;
724 struct aio_ring *ring;
725
726 spin_lock(&mm->ioctx_lock);
727 table = rcu_dereference_raw(mm->ioctx_table);
728
729 while (1) {
730 if (table)
731 for (i = 0; i < table->nr; i++)
732 if (!rcu_access_pointer(table->table[i])) {
733 ctx->id = i;
734 rcu_assign_pointer(table->table[i], ctx);
735 spin_unlock(&mm->ioctx_lock);
736
737 /* While kioctx setup is in progress,
738 * we are protected from page migration
739 * changes ring_folios by ->ring_lock.
740 */
741 ring = folio_address(ctx->ring_folios[0]);
742 ring->id = ctx->id;
743 return 0;
744 }
745
746 new_nr = (table ? table->nr : 1) * 4;
747 spin_unlock(&mm->ioctx_lock);
748
749 table = kzalloc_flex(*table, table, new_nr);
750 if (!table)
751 return -ENOMEM;
752
753 table->nr = new_nr;
754
755 spin_lock(&mm->ioctx_lock);
756 old = rcu_dereference_raw(mm->ioctx_table);
757
758 if (!old) {
759 rcu_assign_pointer(mm->ioctx_table, table);
760 } else if (table->nr > old->nr) {
761 memcpy(table->table, old->table,
762 old->nr * sizeof(struct kioctx *));
763
764 rcu_assign_pointer(mm->ioctx_table, table);
765 kfree_rcu(old, rcu);
766 } else {
767 kfree(table);
768 table = old;
769 }
770 }
771}
772
773static void aio_nr_sub(unsigned nr)
774{
775 spin_lock(&aio_nr_lock);
776 if (WARN_ON(aio_nr - nr > aio_nr))
777 aio_nr = 0;
778 else
779 aio_nr -= nr;
780 spin_unlock(&aio_nr_lock);
781}
782
783/* ioctx_alloc
784 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
785 */
786static struct kioctx *ioctx_alloc(unsigned nr_events)
787{
788 struct mm_struct *mm = current->mm;
789 struct kioctx *ctx;
790 int err = -ENOMEM;
791
792 /*
793 * Store the original nr_events -- what userspace passed to io_setup(),
794 * for counting against the global limit -- before it changes.
795 */
796 unsigned int max_reqs = nr_events;
797
798 /*
799 * We keep track of the number of available ringbuffer slots, to prevent
800 * overflow (reqs_available), and we also use percpu counters for this.
801 *
802 * So since up to half the slots might be on other cpu's percpu counters
803 * and unavailable, double nr_events so userspace sees what they
804 * expected: additionally, we move req_batch slots to/from percpu
805 * counters at a time, so make sure that isn't 0:
806 */
807 nr_events = max(nr_events, num_possible_cpus() * 4);
808 nr_events *= 2;
809
810 /* Prevent overflows */
811 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
812 pr_debug("ENOMEM: nr_events too high\n");
813 return ERR_PTR(-EINVAL);
814 }
815
816 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
817 return ERR_PTR(-EAGAIN);
818
819 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
820 if (!ctx)
821 return ERR_PTR(-ENOMEM);
822
823 ctx->max_reqs = max_reqs;
824
825 spin_lock_init(&ctx->ctx_lock);
826 spin_lock_init(&ctx->completion_lock);
827 mutex_init(&ctx->ring_lock);
828 /* Protect against page migration throughout kiotx setup by keeping
829 * the ring_lock mutex held until setup is complete. */
830 mutex_lock(&ctx->ring_lock);
831 init_waitqueue_head(&ctx->wait);
832
833 INIT_LIST_HEAD(&ctx->active_reqs);
834
835 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
836 goto err;
837
838 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
839 goto err;
840
841 ctx->cpu = alloc_percpu(struct kioctx_cpu);
842 if (!ctx->cpu)
843 goto err;
844
845 err = aio_setup_ring(ctx, nr_events);
846 if (err < 0)
847 goto err;
848
849 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
850 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
851 if (ctx->req_batch < 1)
852 ctx->req_batch = 1;
853
854 /* limit the number of system wide aios */
855 spin_lock(&aio_nr_lock);
856 if (aio_nr + ctx->max_reqs > aio_max_nr ||
857 aio_nr + ctx->max_reqs < aio_nr) {
858 spin_unlock(&aio_nr_lock);
859 err = -EAGAIN;
860 goto err_ctx;
861 }
862 aio_nr += ctx->max_reqs;
863 spin_unlock(&aio_nr_lock);
864
865 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
866 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
867
868 err = ioctx_add_table(ctx, mm);
869 if (err)
870 goto err_cleanup;
871
872 /* Release the ring_lock mutex now that all setup is complete. */
873 mutex_unlock(&ctx->ring_lock);
874
875 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
876 ctx, ctx->user_id, mm, ctx->nr_events);
877 return ctx;
878
879err_cleanup:
880 aio_nr_sub(ctx->max_reqs);
881err_ctx:
882 atomic_set(&ctx->dead, 1);
883 if (ctx->mmap_size)
884 vm_munmap(ctx->mmap_base, ctx->mmap_size);
885 aio_free_ring(ctx);
886err:
887 mutex_unlock(&ctx->ring_lock);
888 free_percpu(ctx->cpu);
889 percpu_ref_exit(&ctx->reqs);
890 percpu_ref_exit(&ctx->users);
891 kmem_cache_free(kioctx_cachep, ctx);
892 pr_debug("error allocating ioctx %d\n", err);
893 return ERR_PTR(err);
894}
895
896/* kill_ioctx
897 * Cancels all outstanding aio requests on an aio context. Used
898 * when the processes owning a context have all exited to encourage
899 * the rapid destruction of the kioctx.
900 */
901static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
902 struct ctx_rq_wait *wait)
903{
904 struct kioctx_table *table;
905
906 spin_lock(&mm->ioctx_lock);
907 if (atomic_xchg(&ctx->dead, 1)) {
908 spin_unlock(&mm->ioctx_lock);
909 return -EINVAL;
910 }
911
912 table = rcu_dereference_raw(mm->ioctx_table);
913 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
914 RCU_INIT_POINTER(table->table[ctx->id], NULL);
915 spin_unlock(&mm->ioctx_lock);
916
917 /* free_ioctx_reqs() will do the necessary RCU synchronization */
918 wake_up_all(&ctx->wait);
919
920 /*
921 * It'd be more correct to do this in free_ioctx(), after all
922 * the outstanding kiocbs have finished - but by then io_destroy
923 * has already returned, so io_setup() could potentially return
924 * -EAGAIN with no ioctxs actually in use (as far as userspace
925 * could tell).
926 */
927 aio_nr_sub(ctx->max_reqs);
928
929 if (ctx->mmap_size)
930 vm_munmap(ctx->mmap_base, ctx->mmap_size);
931
932 ctx->rq_wait = wait;
933 percpu_ref_kill(&ctx->users);
934 return 0;
935}
936
937/*
938 * exit_aio: called when the last user of mm goes away. At this point, there is
939 * no way for any new requests to be submitted or any of the io_* syscalls to be
940 * called on the context.
941 *
942 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
943 * them.
944 */
945void exit_aio(struct mm_struct *mm)
946{
947 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
948 struct ctx_rq_wait wait;
949 int i, skipped;
950
951 if (!table)
952 return;
953
954 atomic_set(&wait.count, table->nr);
955 init_completion(&wait.comp);
956
957 skipped = 0;
958 for (i = 0; i < table->nr; ++i) {
959 struct kioctx *ctx =
960 rcu_dereference_protected(table->table[i], true);
961
962 if (!ctx) {
963 skipped++;
964 continue;
965 }
966
967 /*
968 * We don't need to bother with munmap() here - exit_mmap(mm)
969 * is coming and it'll unmap everything. And we simply can't,
970 * this is not necessarily our ->mm.
971 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
972 * that it needs to unmap the area, just set it to 0.
973 */
974 ctx->mmap_size = 0;
975 kill_ioctx(mm, ctx, &wait);
976 }
977
978 if (!atomic_sub_and_test(skipped, &wait.count)) {
979 /* Wait until all IO for the context are done. */
980 wait_for_completion(&wait.comp);
981 }
982
983 RCU_INIT_POINTER(mm->ioctx_table, NULL);
984 kfree(table);
985}
986
987static void put_reqs_available(struct kioctx *ctx, unsigned nr)
988{
989 struct kioctx_cpu *kcpu;
990 unsigned long flags;
991
992 local_irq_save(flags);
993 kcpu = this_cpu_ptr(ctx->cpu);
994 kcpu->reqs_available += nr;
995
996 while (kcpu->reqs_available >= ctx->req_batch * 2) {
997 kcpu->reqs_available -= ctx->req_batch;
998 atomic_add(ctx->req_batch, &ctx->reqs_available);
999 }
1000
1001 local_irq_restore(flags);
1002}
1003
1004static bool __get_reqs_available(struct kioctx *ctx)
1005{
1006 struct kioctx_cpu *kcpu;
1007 bool ret = false;
1008 unsigned long flags;
1009
1010 local_irq_save(flags);
1011 kcpu = this_cpu_ptr(ctx->cpu);
1012 if (!kcpu->reqs_available) {
1013 int avail = atomic_read(&ctx->reqs_available);
1014
1015 do {
1016 if (avail < ctx->req_batch)
1017 goto out;
1018 } while (!atomic_try_cmpxchg(&ctx->reqs_available,
1019 &avail, avail - ctx->req_batch));
1020
1021 kcpu->reqs_available += ctx->req_batch;
1022 }
1023
1024 ret = true;
1025 kcpu->reqs_available--;
1026out:
1027 local_irq_restore(flags);
1028 return ret;
1029}
1030
1031/* refill_reqs_available
1032 * Updates the reqs_available reference counts used for tracking the
1033 * number of free slots in the completion ring. This can be called
1034 * from aio_complete() (to optimistically update reqs_available) or
1035 * from aio_get_req() (the we're out of events case). It must be
1036 * called holding ctx->completion_lock.
1037 */
1038static void refill_reqs_available(struct kioctx *ctx, unsigned head,
1039 unsigned tail)
1040{
1041 unsigned events_in_ring, completed;
1042
1043 /* Clamp head since userland can write to it. */
1044 head %= ctx->nr_events;
1045 if (head <= tail)
1046 events_in_ring = tail - head;
1047 else
1048 events_in_ring = ctx->nr_events - (head - tail);
1049
1050 completed = ctx->completed_events;
1051 if (events_in_ring < completed)
1052 completed -= events_in_ring;
1053 else
1054 completed = 0;
1055
1056 if (!completed)
1057 return;
1058
1059 ctx->completed_events -= completed;
1060 put_reqs_available(ctx, completed);
1061}
1062
1063/* user_refill_reqs_available
1064 * Called to refill reqs_available when aio_get_req() encounters an
1065 * out of space in the completion ring.
1066 */
1067static void user_refill_reqs_available(struct kioctx *ctx)
1068{
1069 spin_lock_irq(&ctx->completion_lock);
1070 if (ctx->completed_events) {
1071 struct aio_ring *ring;
1072 unsigned head;
1073
1074 /* Access of ring->head may race with aio_read_events_ring()
1075 * here, but that's okay since whether we read the old version
1076 * or the new version, and either will be valid. The important
1077 * part is that head cannot pass tail since we prevent
1078 * aio_complete() from updating tail by holding
1079 * ctx->completion_lock. Even if head is invalid, the check
1080 * against ctx->completed_events below will make sure we do the
1081 * safe/right thing.
1082 */
1083 ring = folio_address(ctx->ring_folios[0]);
1084 head = ring->head;
1085
1086 refill_reqs_available(ctx, head, ctx->tail);
1087 }
1088
1089 spin_unlock_irq(&ctx->completion_lock);
1090}
1091
1092static bool get_reqs_available(struct kioctx *ctx)
1093{
1094 if (__get_reqs_available(ctx))
1095 return true;
1096 user_refill_reqs_available(ctx);
1097 return __get_reqs_available(ctx);
1098}
1099
1100/* aio_get_req
1101 * Allocate a slot for an aio request.
1102 * Returns NULL if no requests are free.
1103 *
1104 * The refcount is initialized to 2 - one for the async op completion,
1105 * one for the synchronous code that does this.
1106 */
1107static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1108{
1109 struct aio_kiocb *req;
1110
1111 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1112 if (unlikely(!req))
1113 return NULL;
1114
1115 if (unlikely(!get_reqs_available(ctx))) {
1116 kmem_cache_free(kiocb_cachep, req);
1117 return NULL;
1118 }
1119
1120 percpu_ref_get(&ctx->reqs);
1121 req->ki_ctx = ctx;
1122 INIT_LIST_HEAD(&req->ki_list);
1123 refcount_set(&req->ki_refcnt, 2);
1124 req->ki_eventfd = NULL;
1125 return req;
1126}
1127
1128static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1129{
1130 struct aio_ring __user *ring = (void __user *)ctx_id;
1131 struct mm_struct *mm = current->mm;
1132 struct kioctx *ctx, *ret = NULL;
1133 struct kioctx_table *table;
1134 unsigned id;
1135
1136 if (get_user(id, &ring->id))
1137 return NULL;
1138
1139 rcu_read_lock();
1140 table = rcu_dereference(mm->ioctx_table);
1141
1142 if (!table || id >= table->nr)
1143 goto out;
1144
1145 id = array_index_nospec(id, table->nr);
1146 ctx = rcu_dereference(table->table[id]);
1147 if (ctx && ctx->user_id == ctx_id) {
1148 if (percpu_ref_tryget_live(&ctx->users))
1149 ret = ctx;
1150 }
1151out:
1152 rcu_read_unlock();
1153 return ret;
1154}
1155
1156static inline void iocb_destroy(struct aio_kiocb *iocb)
1157{
1158 if (iocb->ki_eventfd)
1159 eventfd_ctx_put(iocb->ki_eventfd);
1160 if (iocb->ki_filp)
1161 fput(iocb->ki_filp);
1162 percpu_ref_put(&iocb->ki_ctx->reqs);
1163 kmem_cache_free(kiocb_cachep, iocb);
1164}
1165
1166struct aio_waiter {
1167 struct wait_queue_entry w;
1168 size_t min_nr;
1169};
1170
1171/* aio_complete
1172 * Called when the io request on the given iocb is complete.
1173 */
1174static void aio_complete(struct aio_kiocb *iocb)
1175{
1176 struct kioctx *ctx = iocb->ki_ctx;
1177 struct aio_ring *ring;
1178 struct io_event *ev_page, *event;
1179 unsigned tail, pos, head, avail;
1180 unsigned long flags;
1181
1182 /*
1183 * Add a completion event to the ring buffer. Must be done holding
1184 * ctx->completion_lock to prevent other code from messing with the tail
1185 * pointer since we might be called from irq context.
1186 */
1187 spin_lock_irqsave(&ctx->completion_lock, flags);
1188
1189 tail = ctx->tail;
1190 pos = tail + AIO_EVENTS_OFFSET;
1191
1192 if (++tail >= ctx->nr_events)
1193 tail = 0;
1194
1195 ev_page = folio_address(ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE]);
1196 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1197
1198 *event = iocb->ki_res;
1199
1200 flush_dcache_folio(ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE]);
1201
1202 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1203 (void __user *)(unsigned long)iocb->ki_res.obj,
1204 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1205
1206 /* after flagging the request as done, we
1207 * must never even look at it again
1208 */
1209 smp_wmb(); /* make event visible before updating tail */
1210
1211 ctx->tail = tail;
1212
1213 ring = folio_address(ctx->ring_folios[0]);
1214 head = ring->head;
1215 ring->tail = tail;
1216 flush_dcache_folio(ctx->ring_folios[0]);
1217
1218 ctx->completed_events++;
1219 if (ctx->completed_events > 1)
1220 refill_reqs_available(ctx, head, tail);
1221
1222 avail = tail > head
1223 ? tail - head
1224 : tail + ctx->nr_events - head;
1225 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1226
1227 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1228
1229 /*
1230 * Check if the user asked us to deliver the result through an
1231 * eventfd. The eventfd_signal() function is safe to be called
1232 * from IRQ context.
1233 */
1234 if (iocb->ki_eventfd)
1235 eventfd_signal(iocb->ki_eventfd);
1236
1237 /*
1238 * We have to order our ring_info tail store above and test
1239 * of the wait list below outside the wait lock. This is
1240 * like in wake_up_bit() where clearing a bit has to be
1241 * ordered with the unlocked test.
1242 */
1243 smp_mb();
1244
1245 if (waitqueue_active(&ctx->wait)) {
1246 struct aio_waiter *curr, *next;
1247 unsigned long flags;
1248
1249 spin_lock_irqsave(&ctx->wait.lock, flags);
1250 list_for_each_entry_safe(curr, next, &ctx->wait.head, w.entry)
1251 if (avail >= curr->min_nr) {
1252 wake_up_process(curr->w.private);
1253 list_del_init_careful(&curr->w.entry);
1254 }
1255 spin_unlock_irqrestore(&ctx->wait.lock, flags);
1256 }
1257}
1258
1259static inline void iocb_put(struct aio_kiocb *iocb)
1260{
1261 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1262 aio_complete(iocb);
1263 iocb_destroy(iocb);
1264 }
1265}
1266
1267/* aio_read_events_ring
1268 * Pull an event off of the ioctx's event ring. Returns the number of
1269 * events fetched
1270 */
1271static long aio_read_events_ring(struct kioctx *ctx,
1272 struct io_event __user *event, long nr)
1273{
1274 struct aio_ring *ring;
1275 unsigned head, tail, pos;
1276 long ret = 0;
1277 int copy_ret;
1278
1279 /*
1280 * The mutex can block and wake us up and that will cause
1281 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1282 * and repeat. This should be rare enough that it doesn't cause
1283 * performance issues. See the comment in read_events() for more detail.
1284 */
1285 sched_annotate_sleep();
1286 mutex_lock(&ctx->ring_lock);
1287
1288 /* Access to ->ring_folios here is protected by ctx->ring_lock. */
1289 ring = folio_address(ctx->ring_folios[0]);
1290 head = ring->head;
1291 tail = ring->tail;
1292
1293 /*
1294 * Ensure that once we've read the current tail pointer, that
1295 * we also see the events that were stored up to the tail.
1296 */
1297 smp_rmb();
1298
1299 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1300
1301 if (head == tail)
1302 goto out;
1303
1304 head %= ctx->nr_events;
1305 tail %= ctx->nr_events;
1306
1307 while (ret < nr) {
1308 long avail;
1309 struct io_event *ev;
1310 struct folio *folio;
1311
1312 avail = (head <= tail ? tail : ctx->nr_events) - head;
1313 if (head == tail)
1314 break;
1315
1316 pos = head + AIO_EVENTS_OFFSET;
1317 folio = ctx->ring_folios[pos / AIO_EVENTS_PER_PAGE];
1318 pos %= AIO_EVENTS_PER_PAGE;
1319
1320 avail = min(avail, nr - ret);
1321 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1322
1323 ev = folio_address(folio);
1324 copy_ret = copy_to_user(event + ret, ev + pos,
1325 sizeof(*ev) * avail);
1326
1327 if (unlikely(copy_ret)) {
1328 ret = -EFAULT;
1329 goto out;
1330 }
1331
1332 ret += avail;
1333 head += avail;
1334 head %= ctx->nr_events;
1335 }
1336
1337 ring = folio_address(ctx->ring_folios[0]);
1338 ring->head = head;
1339 flush_dcache_folio(ctx->ring_folios[0]);
1340
1341 pr_debug("%li h%u t%u\n", ret, head, tail);
1342out:
1343 mutex_unlock(&ctx->ring_lock);
1344
1345 return ret;
1346}
1347
1348static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1349 struct io_event __user *event, long *i)
1350{
1351 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1352
1353 if (ret > 0)
1354 *i += ret;
1355
1356 if (unlikely(atomic_read(&ctx->dead)))
1357 ret = -EINVAL;
1358
1359 if (!*i)
1360 *i = ret;
1361
1362 return ret < 0 || *i >= min_nr;
1363}
1364
1365static long read_events(struct kioctx *ctx, long min_nr, long nr,
1366 struct io_event __user *event,
1367 ktime_t until)
1368{
1369 struct hrtimer_sleeper t;
1370 struct aio_waiter w;
1371 long ret = 0, ret2 = 0;
1372
1373 /*
1374 * Note that aio_read_events() is being called as the conditional - i.e.
1375 * we're calling it after prepare_to_wait() has set task state to
1376 * TASK_INTERRUPTIBLE.
1377 *
1378 * But aio_read_events() can block, and if it blocks it's going to flip
1379 * the task state back to TASK_RUNNING.
1380 *
1381 * This should be ok, provided it doesn't flip the state back to
1382 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1383 * will only happen if the mutex_lock() call blocks, and we then find
1384 * the ringbuffer empty. So in practice we should be ok, but it's
1385 * something to be aware of when touching this code.
1386 */
1387 aio_read_events(ctx, min_nr, nr, event, &ret);
1388 if (until == 0 || ret < 0 || ret >= min_nr)
1389 return ret;
1390
1391 hrtimer_setup_sleeper_on_stack(&t, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1392 if (until != KTIME_MAX) {
1393 hrtimer_set_expires_range_ns(&t.timer, until, current->timer_slack_ns);
1394 hrtimer_sleeper_start_expires(&t, HRTIMER_MODE_REL);
1395 }
1396
1397 init_wait(&w.w);
1398
1399 while (1) {
1400 unsigned long nr_got = ret;
1401
1402 w.min_nr = min_nr - ret;
1403
1404 ret2 = prepare_to_wait_event(&ctx->wait, &w.w, TASK_INTERRUPTIBLE);
1405 if (!ret2 && !t.task)
1406 ret2 = -ETIME;
1407
1408 if (aio_read_events(ctx, min_nr, nr, event, &ret) || ret2)
1409 break;
1410
1411 if (nr_got == ret)
1412 schedule();
1413 }
1414
1415 finish_wait(&ctx->wait, &w.w);
1416 hrtimer_cancel(&t.timer);
1417 destroy_hrtimer_on_stack(&t.timer);
1418
1419 return ret;
1420}
1421
1422/* sys_io_setup:
1423 * Create an aio_context capable of receiving at least nr_events.
1424 * ctxp must not point to an aio_context that already exists, and
1425 * must be initialized to 0 prior to the call. On successful
1426 * creation of the aio_context, *ctxp is filled in with the resulting
1427 * handle. May fail with -EINVAL if *ctxp is not initialized,
1428 * if the specified nr_events exceeds internal limits. May fail
1429 * with -EAGAIN if the specified nr_events exceeds the user's limit
1430 * of available events. May fail with -ENOMEM if insufficient kernel
1431 * resources are available. May fail with -EFAULT if an invalid
1432 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1433 * implemented.
1434 */
1435SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1436{
1437 struct kioctx *ioctx = NULL;
1438 unsigned long ctx;
1439 long ret;
1440
1441 ret = get_user(ctx, ctxp);
1442 if (unlikely(ret))
1443 goto out;
1444
1445 ret = -EINVAL;
1446 if (unlikely(ctx || nr_events == 0)) {
1447 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1448 ctx, nr_events);
1449 goto out;
1450 }
1451
1452 ioctx = ioctx_alloc(nr_events);
1453 ret = PTR_ERR(ioctx);
1454 if (!IS_ERR(ioctx)) {
1455 ret = put_user(ioctx->user_id, ctxp);
1456 if (ret)
1457 kill_ioctx(current->mm, ioctx, NULL);
1458 percpu_ref_put(&ioctx->users);
1459 }
1460
1461out:
1462 return ret;
1463}
1464
1465#ifdef CONFIG_COMPAT
1466COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1467{
1468 struct kioctx *ioctx = NULL;
1469 unsigned long ctx;
1470 long ret;
1471
1472 ret = get_user(ctx, ctx32p);
1473 if (unlikely(ret))
1474 goto out;
1475
1476 ret = -EINVAL;
1477 if (unlikely(ctx || nr_events == 0)) {
1478 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1479 ctx, nr_events);
1480 goto out;
1481 }
1482
1483 ioctx = ioctx_alloc(nr_events);
1484 ret = PTR_ERR(ioctx);
1485 if (!IS_ERR(ioctx)) {
1486 /* truncating is ok because it's a user address */
1487 ret = put_user((u32)ioctx->user_id, ctx32p);
1488 if (ret)
1489 kill_ioctx(current->mm, ioctx, NULL);
1490 percpu_ref_put(&ioctx->users);
1491 }
1492
1493out:
1494 return ret;
1495}
1496#endif
1497
1498/* sys_io_destroy:
1499 * Destroy the aio_context specified. May cancel any outstanding
1500 * AIOs and block on completion. Will fail with -ENOSYS if not
1501 * implemented. May fail with -EINVAL if the context pointed to
1502 * is invalid.
1503 */
1504SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1505{
1506 struct kioctx *ioctx = lookup_ioctx(ctx);
1507 if (likely(NULL != ioctx)) {
1508 struct ctx_rq_wait wait;
1509 int ret;
1510
1511 init_completion(&wait.comp);
1512 atomic_set(&wait.count, 1);
1513
1514 /* Pass requests_done to kill_ioctx() where it can be set
1515 * in a thread-safe way. If we try to set it here then we have
1516 * a race condition if two io_destroy() called simultaneously.
1517 */
1518 ret = kill_ioctx(current->mm, ioctx, &wait);
1519 percpu_ref_put(&ioctx->users);
1520
1521 /* Wait until all IO for the context are done. Otherwise kernel
1522 * keep using user-space buffers even if user thinks the context
1523 * is destroyed.
1524 */
1525 if (!ret)
1526 wait_for_completion(&wait.comp);
1527
1528 return ret;
1529 }
1530 pr_debug("EINVAL: invalid context id\n");
1531 return -EINVAL;
1532}
1533
1534static void aio_remove_iocb(struct aio_kiocb *iocb)
1535{
1536 struct kioctx *ctx = iocb->ki_ctx;
1537 unsigned long flags;
1538
1539 spin_lock_irqsave(&ctx->ctx_lock, flags);
1540 list_del(&iocb->ki_list);
1541 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1542}
1543
1544static void aio_complete_rw(struct kiocb *kiocb, long res)
1545{
1546 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1547
1548 if (!list_empty_careful(&iocb->ki_list))
1549 aio_remove_iocb(iocb);
1550
1551 if (kiocb->ki_flags & IOCB_WRITE) {
1552 struct inode *inode = file_inode(kiocb->ki_filp);
1553
1554 if (S_ISREG(inode->i_mode))
1555 kiocb_end_write(kiocb);
1556 }
1557
1558 iocb->ki_res.res = res;
1559 iocb->ki_res.res2 = 0;
1560 iocb_put(iocb);
1561}
1562
1563static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb, int rw_type)
1564{
1565 int ret;
1566
1567 req->ki_write_stream = 0;
1568 req->ki_complete = aio_complete_rw;
1569 req->private = NULL;
1570 req->ki_pos = iocb->aio_offset;
1571 req->ki_flags = req->ki_filp->f_iocb_flags | IOCB_AIO_RW;
1572 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1573 req->ki_flags |= IOCB_EVENTFD;
1574 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1575 /*
1576 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1577 * aio_reqprio is interpreted as an I/O scheduling
1578 * class and priority.
1579 */
1580 ret = ioprio_check_cap(iocb->aio_reqprio);
1581 if (ret) {
1582 pr_debug("aio ioprio check cap error: %d\n", ret);
1583 return ret;
1584 }
1585
1586 req->ki_ioprio = iocb->aio_reqprio;
1587 } else
1588 req->ki_ioprio = get_current_ioprio();
1589
1590 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags, rw_type);
1591 if (unlikely(ret))
1592 return ret;
1593
1594 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1595 return 0;
1596}
1597
1598static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1599 struct iovec **iovec, bool vectored, bool compat,
1600 struct iov_iter *iter)
1601{
1602 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1603 size_t len = iocb->aio_nbytes;
1604
1605 if (!vectored) {
1606 ssize_t ret = import_ubuf(rw, buf, len, iter);
1607 *iovec = NULL;
1608 return ret;
1609 }
1610
1611 return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1612}
1613
1614static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1615{
1616 switch (ret) {
1617 case -EIOCBQUEUED:
1618 break;
1619 case -ERESTARTSYS:
1620 case -ERESTARTNOINTR:
1621 case -ERESTARTNOHAND:
1622 case -ERESTART_RESTARTBLOCK:
1623 /*
1624 * There's no easy way to restart the syscall since other AIO's
1625 * may be already running. Just fail this IO with EINTR.
1626 */
1627 ret = -EINTR;
1628 fallthrough;
1629 default:
1630 req->ki_complete(req, ret);
1631 }
1632}
1633
1634static int aio_read(struct kiocb *req, const struct iocb *iocb,
1635 bool vectored, bool compat)
1636{
1637 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1638 struct iov_iter iter;
1639 struct file *file;
1640 int ret;
1641
1642 ret = aio_prep_rw(req, iocb, READ);
1643 if (ret)
1644 return ret;
1645 file = req->ki_filp;
1646 if (unlikely(!(file->f_mode & FMODE_READ)))
1647 return -EBADF;
1648 if (unlikely(!file->f_op->read_iter))
1649 return -EINVAL;
1650
1651 ret = aio_setup_rw(ITER_DEST, iocb, &iovec, vectored, compat, &iter);
1652 if (ret < 0)
1653 return ret;
1654 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1655 if (!ret)
1656 aio_rw_done(req, file->f_op->read_iter(req, &iter));
1657 kfree(iovec);
1658 return ret;
1659}
1660
1661static int aio_write(struct kiocb *req, const struct iocb *iocb,
1662 bool vectored, bool compat)
1663{
1664 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1665 struct iov_iter iter;
1666 struct file *file;
1667 int ret;
1668
1669 ret = aio_prep_rw(req, iocb, WRITE);
1670 if (ret)
1671 return ret;
1672 file = req->ki_filp;
1673
1674 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1675 return -EBADF;
1676 if (unlikely(!file->f_op->write_iter))
1677 return -EINVAL;
1678
1679 ret = aio_setup_rw(ITER_SOURCE, iocb, &iovec, vectored, compat, &iter);
1680 if (ret < 0)
1681 return ret;
1682 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1683 if (!ret) {
1684 if (S_ISREG(file_inode(file)->i_mode))
1685 kiocb_start_write(req);
1686 req->ki_flags |= IOCB_WRITE;
1687 aio_rw_done(req, file->f_op->write_iter(req, &iter));
1688 }
1689 kfree(iovec);
1690 return ret;
1691}
1692
1693static void aio_fsync_work(struct work_struct *work)
1694{
1695 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1696
1697 scoped_with_creds(iocb->fsync.creds)
1698 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1699
1700 put_cred(iocb->fsync.creds);
1701 iocb_put(iocb);
1702}
1703
1704static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1705 bool datasync)
1706{
1707 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1708 iocb->aio_rw_flags))
1709 return -EINVAL;
1710
1711 if (unlikely(!req->file->f_op->fsync))
1712 return -EINVAL;
1713
1714 req->creds = prepare_creds();
1715 if (!req->creds)
1716 return -ENOMEM;
1717
1718 req->datasync = datasync;
1719 INIT_WORK(&req->work, aio_fsync_work);
1720 schedule_work(&req->work);
1721 return 0;
1722}
1723
1724static void aio_poll_put_work(struct work_struct *work)
1725{
1726 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1727 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1728
1729 iocb_put(iocb);
1730}
1731
1732/*
1733 * Safely lock the waitqueue which the request is on, synchronizing with the
1734 * case where the ->poll() provider decides to free its waitqueue early.
1735 *
1736 * Returns true on success, meaning that req->head->lock was locked, req->wait
1737 * is on req->head, and an RCU read lock was taken. Returns false if the
1738 * request was already removed from its waitqueue (which might no longer exist).
1739 */
1740static bool poll_iocb_lock_wq(struct poll_iocb *req)
1741{
1742 wait_queue_head_t *head;
1743
1744 /*
1745 * While we hold the waitqueue lock and the waitqueue is nonempty,
1746 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1747 * lock in the first place can race with the waitqueue being freed.
1748 *
1749 * We solve this as eventpoll does: by taking advantage of the fact that
1750 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1751 * we enter rcu_read_lock() and see that the pointer to the queue is
1752 * non-NULL, we can then lock it without the memory being freed out from
1753 * under us, then check whether the request is still on the queue.
1754 *
1755 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1756 * case the caller deletes the entry from the queue, leaving it empty.
1757 * In that case, only RCU prevents the queue memory from being freed.
1758 */
1759 rcu_read_lock();
1760 head = smp_load_acquire(&req->head);
1761 if (head) {
1762 spin_lock(&head->lock);
1763 if (!list_empty(&req->wait.entry))
1764 return true;
1765 spin_unlock(&head->lock);
1766 }
1767 rcu_read_unlock();
1768 return false;
1769}
1770
1771static void poll_iocb_unlock_wq(struct poll_iocb *req)
1772{
1773 spin_unlock(&req->head->lock);
1774 rcu_read_unlock();
1775}
1776
1777static void aio_poll_complete_work(struct work_struct *work)
1778{
1779 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1780 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1781 struct poll_table_struct pt = { ._key = req->events };
1782 struct kioctx *ctx = iocb->ki_ctx;
1783 __poll_t mask = 0;
1784
1785 if (!READ_ONCE(req->cancelled))
1786 mask = vfs_poll(req->file, &pt) & req->events;
1787
1788 /*
1789 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1790 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1791 * synchronize with them. In the cancellation case the list_del_init
1792 * itself is not actually needed, but harmless so we keep it in to
1793 * avoid further branches in the fast path.
1794 */
1795 spin_lock_irq(&ctx->ctx_lock);
1796 if (poll_iocb_lock_wq(req)) {
1797 if (!mask && !READ_ONCE(req->cancelled)) {
1798 /*
1799 * The request isn't actually ready to be completed yet.
1800 * Reschedule completion if another wakeup came in.
1801 */
1802 if (req->work_need_resched) {
1803 schedule_work(&req->work);
1804 req->work_need_resched = false;
1805 } else {
1806 req->work_scheduled = false;
1807 }
1808 poll_iocb_unlock_wq(req);
1809 spin_unlock_irq(&ctx->ctx_lock);
1810 return;
1811 }
1812 list_del_init(&req->wait.entry);
1813 poll_iocb_unlock_wq(req);
1814 } /* else, POLLFREE has freed the waitqueue, so we must complete */
1815 list_del_init(&iocb->ki_list);
1816 iocb->ki_res.res = mangle_poll(mask);
1817 spin_unlock_irq(&ctx->ctx_lock);
1818
1819 iocb_put(iocb);
1820}
1821
1822/* assumes we are called with irqs disabled */
1823static int aio_poll_cancel(struct kiocb *iocb)
1824{
1825 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1826 struct poll_iocb *req = &aiocb->poll;
1827
1828 if (poll_iocb_lock_wq(req)) {
1829 WRITE_ONCE(req->cancelled, true);
1830 if (!req->work_scheduled) {
1831 schedule_work(&aiocb->poll.work);
1832 req->work_scheduled = true;
1833 }
1834 poll_iocb_unlock_wq(req);
1835 } /* else, the request was force-cancelled by POLLFREE already */
1836
1837 return 0;
1838}
1839
1840static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1841 void *key)
1842{
1843 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1844 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1845 __poll_t mask = key_to_poll(key);
1846 unsigned long flags;
1847
1848 /* for instances that support it check for an event match first: */
1849 if (mask && !(mask & req->events))
1850 return 0;
1851
1852 /*
1853 * Complete the request inline if possible. This requires that three
1854 * conditions be met:
1855 * 1. An event mask must have been passed. If a plain wakeup was done
1856 * instead, then mask == 0 and we have to call vfs_poll() to get
1857 * the events, so inline completion isn't possible.
1858 * 2. The completion work must not have already been scheduled.
1859 * 3. ctx_lock must not be busy. We have to use trylock because we
1860 * already hold the waitqueue lock, so this inverts the normal
1861 * locking order. Use irqsave/irqrestore because not all
1862 * filesystems (e.g. fuse) call this function with IRQs disabled,
1863 * yet IRQs have to be disabled before ctx_lock is obtained.
1864 */
1865 if (mask && !req->work_scheduled &&
1866 spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1867 struct kioctx *ctx = iocb->ki_ctx;
1868
1869 list_del_init(&req->wait.entry);
1870 list_del(&iocb->ki_list);
1871 iocb->ki_res.res = mangle_poll(mask);
1872 /*
1873 * We hold an arbitrary provider waitqueue lock here. Signaling a
1874 * result eventfd can feed back through epoll and try to take the same
1875 * lock again. Defer all eventfd-backed poll completions.
1876 */
1877 if (iocb->ki_eventfd) {
1878 iocb = NULL;
1879 INIT_WORK(&req->work, aio_poll_put_work);
1880 schedule_work(&req->work);
1881 }
1882 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1883 if (iocb)
1884 iocb_put(iocb);
1885 } else {
1886 /*
1887 * Schedule the completion work if needed. If it was already
1888 * scheduled, record that another wakeup came in.
1889 *
1890 * Don't remove the request from the waitqueue here, as it might
1891 * not actually be complete yet (we won't know until vfs_poll()
1892 * is called), and we must not miss any wakeups. POLLFREE is an
1893 * exception to this; see below.
1894 */
1895 if (req->work_scheduled) {
1896 req->work_need_resched = true;
1897 } else {
1898 schedule_work(&req->work);
1899 req->work_scheduled = true;
1900 }
1901
1902 /*
1903 * If the waitqueue is being freed early but we can't complete
1904 * the request inline, we have to tear down the request as best
1905 * we can. That means immediately removing the request from its
1906 * waitqueue and preventing all further accesses to the
1907 * waitqueue via the request. We also need to schedule the
1908 * completion work (done above). Also mark the request as
1909 * cancelled, to potentially skip an unneeded call to ->poll().
1910 */
1911 if (mask & POLLFREE) {
1912 WRITE_ONCE(req->cancelled, true);
1913 list_del_init(&req->wait.entry);
1914
1915 /*
1916 * Careful: this *must* be the last step, since as soon
1917 * as req->head is NULL'ed out, the request can be
1918 * completed and freed, since aio_poll_complete_work()
1919 * will no longer need to take the waitqueue lock.
1920 */
1921 smp_store_release(&req->head, NULL);
1922 }
1923 }
1924 return 1;
1925}
1926
1927struct aio_poll_table {
1928 struct poll_table_struct pt;
1929 struct aio_kiocb *iocb;
1930 bool queued;
1931 int error;
1932};
1933
1934static void
1935aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1936 struct poll_table_struct *p)
1937{
1938 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1939
1940 /* multiple wait queues per file are not supported */
1941 if (unlikely(pt->queued)) {
1942 pt->error = -EINVAL;
1943 return;
1944 }
1945
1946 pt->queued = true;
1947 pt->error = 0;
1948 pt->iocb->poll.head = head;
1949 add_wait_queue(head, &pt->iocb->poll.wait);
1950}
1951
1952static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1953{
1954 struct kioctx *ctx = aiocb->ki_ctx;
1955 struct poll_iocb *req = &aiocb->poll;
1956 struct aio_poll_table apt;
1957 bool cancel = false;
1958 __poll_t mask;
1959
1960 /* reject any unknown events outside the normal event mask. */
1961 if ((u16)iocb->aio_buf != iocb->aio_buf)
1962 return -EINVAL;
1963 /* reject fields that are not defined for poll */
1964 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1965 return -EINVAL;
1966
1967 INIT_WORK(&req->work, aio_poll_complete_work);
1968 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1969
1970 req->head = NULL;
1971 req->cancelled = false;
1972 req->work_scheduled = false;
1973 req->work_need_resched = false;
1974
1975 apt.pt._qproc = aio_poll_queue_proc;
1976 apt.pt._key = req->events;
1977 apt.iocb = aiocb;
1978 apt.queued = false;
1979 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1980
1981 /* initialized the list so that we can do list_empty checks */
1982 INIT_LIST_HEAD(&req->wait.entry);
1983 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1984
1985 mask = vfs_poll(req->file, &apt.pt) & req->events;
1986 spin_lock_irq(&ctx->ctx_lock);
1987 if (likely(apt.queued)) {
1988 bool on_queue = poll_iocb_lock_wq(req);
1989
1990 if (!on_queue || req->work_scheduled) {
1991 /*
1992 * aio_poll_wake() already either scheduled the async
1993 * completion work, or completed the request inline.
1994 */
1995 if (apt.error) /* unsupported case: multiple queues */
1996 cancel = true;
1997 apt.error = 0;
1998 mask = 0;
1999 }
2000 if (mask || apt.error) {
2001 /* Steal to complete synchronously. */
2002 list_del_init(&req->wait.entry);
2003 } else if (cancel) {
2004 /* Cancel if possible (may be too late though). */
2005 WRITE_ONCE(req->cancelled, true);
2006 } else if (on_queue) {
2007 /*
2008 * Actually waiting for an event, so add the request to
2009 * active_reqs so that it can be cancelled if needed.
2010 */
2011 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
2012 aiocb->ki_cancel = aio_poll_cancel;
2013 }
2014 if (on_queue)
2015 poll_iocb_unlock_wq(req);
2016 }
2017 if (mask) { /* no async, we'd stolen it */
2018 aiocb->ki_res.res = mangle_poll(mask);
2019 apt.error = 0;
2020 }
2021 spin_unlock_irq(&ctx->ctx_lock);
2022 if (mask)
2023 iocb_put(aiocb);
2024 return apt.error;
2025}
2026
2027static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
2028 struct iocb __user *user_iocb, struct aio_kiocb *req,
2029 bool compat)
2030{
2031 req->ki_filp = fget(iocb->aio_fildes);
2032 if (unlikely(!req->ki_filp))
2033 return -EBADF;
2034
2035 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
2036 struct eventfd_ctx *eventfd;
2037 /*
2038 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
2039 * instance of the file* now. The file descriptor must be
2040 * an eventfd() fd, and will be signaled for each completed
2041 * event using the eventfd_signal() function.
2042 */
2043 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
2044 if (IS_ERR(eventfd))
2045 return PTR_ERR(eventfd);
2046
2047 req->ki_eventfd = eventfd;
2048 }
2049
2050 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
2051 pr_debug("EFAULT: aio_key\n");
2052 return -EFAULT;
2053 }
2054
2055 req->ki_res.obj = (u64)(unsigned long)user_iocb;
2056 req->ki_res.data = iocb->aio_data;
2057 req->ki_res.res = 0;
2058 req->ki_res.res2 = 0;
2059
2060 switch (iocb->aio_lio_opcode) {
2061 case IOCB_CMD_PREAD:
2062 return aio_read(&req->rw, iocb, false, compat);
2063 case IOCB_CMD_PWRITE:
2064 return aio_write(&req->rw, iocb, false, compat);
2065 case IOCB_CMD_PREADV:
2066 return aio_read(&req->rw, iocb, true, compat);
2067 case IOCB_CMD_PWRITEV:
2068 return aio_write(&req->rw, iocb, true, compat);
2069 case IOCB_CMD_FSYNC:
2070 return aio_fsync(&req->fsync, iocb, false);
2071 case IOCB_CMD_FDSYNC:
2072 return aio_fsync(&req->fsync, iocb, true);
2073 case IOCB_CMD_POLL:
2074 return aio_poll(req, iocb);
2075 default:
2076 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
2077 return -EINVAL;
2078 }
2079}
2080
2081static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
2082 bool compat)
2083{
2084 struct aio_kiocb *req;
2085 struct iocb iocb;
2086 int err;
2087
2088 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
2089 return -EFAULT;
2090
2091 /* enforce forwards compatibility on users */
2092 if (unlikely(iocb.aio_reserved2)) {
2093 pr_debug("EINVAL: reserve field set\n");
2094 return -EINVAL;
2095 }
2096
2097 /* prevent overflows */
2098 if (unlikely(
2099 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
2100 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
2101 ((ssize_t)iocb.aio_nbytes < 0)
2102 )) {
2103 pr_debug("EINVAL: overflow check\n");
2104 return -EINVAL;
2105 }
2106
2107 req = aio_get_req(ctx);
2108 if (unlikely(!req))
2109 return -EAGAIN;
2110
2111 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2112
2113 /* Done with the synchronous reference */
2114 iocb_put(req);
2115
2116 /*
2117 * If err is 0, we'd either done aio_complete() ourselves or have
2118 * arranged for that to be done asynchronously. Anything non-zero
2119 * means that we need to destroy req ourselves.
2120 */
2121 if (unlikely(err)) {
2122 iocb_destroy(req);
2123 put_reqs_available(ctx, 1);
2124 }
2125 return err;
2126}
2127
2128/* sys_io_submit:
2129 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2130 * the number of iocbs queued. May return -EINVAL if the aio_context
2131 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2132 * *iocbpp[0] is not properly initialized, if the operation specified
2133 * is invalid for the file descriptor in the iocb. May fail with
2134 * -EFAULT if any of the data structures point to invalid data. May
2135 * fail with -EBADF if the file descriptor specified in the first
2136 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2137 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2138 * fail with -ENOSYS if not implemented.
2139 */
2140SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2141 struct iocb __user * __user *, iocbpp)
2142{
2143 struct kioctx *ctx;
2144 long ret = 0;
2145 int i = 0;
2146 struct blk_plug plug;
2147
2148 if (unlikely(nr < 0))
2149 return -EINVAL;
2150
2151 ctx = lookup_ioctx(ctx_id);
2152 if (unlikely(!ctx)) {
2153 pr_debug("EINVAL: invalid context id\n");
2154 return -EINVAL;
2155 }
2156
2157 if (nr > ctx->nr_events)
2158 nr = ctx->nr_events;
2159
2160 if (nr > AIO_PLUG_THRESHOLD)
2161 blk_start_plug(&plug);
2162 for (i = 0; i < nr; i++) {
2163 struct iocb __user *user_iocb;
2164
2165 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2166 ret = -EFAULT;
2167 break;
2168 }
2169
2170 ret = io_submit_one(ctx, user_iocb, false);
2171 if (ret)
2172 break;
2173 }
2174 if (nr > AIO_PLUG_THRESHOLD)
2175 blk_finish_plug(&plug);
2176
2177 percpu_ref_put(&ctx->users);
2178 return i ? i : ret;
2179}
2180
2181#ifdef CONFIG_COMPAT
2182COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2183 int, nr, compat_uptr_t __user *, iocbpp)
2184{
2185 struct kioctx *ctx;
2186 long ret = 0;
2187 int i = 0;
2188 struct blk_plug plug;
2189
2190 if (unlikely(nr < 0))
2191 return -EINVAL;
2192
2193 ctx = lookup_ioctx(ctx_id);
2194 if (unlikely(!ctx)) {
2195 pr_debug("EINVAL: invalid context id\n");
2196 return -EINVAL;
2197 }
2198
2199 if (nr > ctx->nr_events)
2200 nr = ctx->nr_events;
2201
2202 if (nr > AIO_PLUG_THRESHOLD)
2203 blk_start_plug(&plug);
2204 for (i = 0; i < nr; i++) {
2205 compat_uptr_t user_iocb;
2206
2207 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2208 ret = -EFAULT;
2209 break;
2210 }
2211
2212 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2213 if (ret)
2214 break;
2215 }
2216 if (nr > AIO_PLUG_THRESHOLD)
2217 blk_finish_plug(&plug);
2218
2219 percpu_ref_put(&ctx->users);
2220 return i ? i : ret;
2221}
2222#endif
2223
2224/* sys_io_cancel:
2225 * Attempts to cancel an iocb previously passed to io_submit. If
2226 * the operation is successfully cancelled, the resulting event is
2227 * copied into the memory pointed to by result without being placed
2228 * into the completion queue and 0 is returned. May fail with
2229 * -EFAULT if any of the data structures pointed to are invalid.
2230 * May fail with -EINVAL if aio_context specified by ctx_id is
2231 * invalid. May fail with -EAGAIN if the iocb specified was not
2232 * cancelled. Will fail with -ENOSYS if not implemented.
2233 */
2234SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2235 struct io_event __user *, result)
2236{
2237 struct kioctx *ctx;
2238 struct aio_kiocb *kiocb;
2239 int ret = -EINVAL;
2240 u32 key;
2241 u64 obj = (u64)(unsigned long)iocb;
2242
2243 if (unlikely(get_user(key, &iocb->aio_key)))
2244 return -EFAULT;
2245 if (unlikely(key != KIOCB_KEY))
2246 return -EINVAL;
2247
2248 ctx = lookup_ioctx(ctx_id);
2249 if (unlikely(!ctx))
2250 return -EINVAL;
2251
2252 spin_lock_irq(&ctx->ctx_lock);
2253 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2254 if (kiocb->ki_res.obj == obj) {
2255 ret = kiocb->ki_cancel(&kiocb->rw);
2256 list_del_init(&kiocb->ki_list);
2257 break;
2258 }
2259 }
2260 spin_unlock_irq(&ctx->ctx_lock);
2261
2262 if (!ret) {
2263 /*
2264 * The result argument is no longer used - the io_event is
2265 * always delivered via the ring buffer. -EINPROGRESS indicates
2266 * cancellation is progress:
2267 */
2268 ret = -EINPROGRESS;
2269 }
2270
2271 percpu_ref_put(&ctx->users);
2272
2273 return ret;
2274}
2275
2276static long do_io_getevents(aio_context_t ctx_id,
2277 long min_nr,
2278 long nr,
2279 struct io_event __user *events,
2280 struct timespec64 *ts)
2281{
2282 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2283 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2284 long ret = -EINVAL;
2285
2286 if (likely(ioctx)) {
2287 if (likely(min_nr <= nr && min_nr >= 0))
2288 ret = read_events(ioctx, min_nr, nr, events, until);
2289 percpu_ref_put(&ioctx->users);
2290 }
2291
2292 return ret;
2293}
2294
2295/* io_getevents:
2296 * Attempts to read at least min_nr events and up to nr events from
2297 * the completion queue for the aio_context specified by ctx_id. If
2298 * it succeeds, the number of read events is returned. May fail with
2299 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2300 * out of range, if timeout is out of range. May fail with -EFAULT
2301 * if any of the memory specified is invalid. May return 0 or
2302 * < min_nr if the timeout specified by timeout has elapsed
2303 * before sufficient events are available, where timeout == NULL
2304 * specifies an infinite timeout. Note that the timeout pointed to by
2305 * timeout is relative. Will fail with -ENOSYS if not implemented.
2306 */
2307#ifdef CONFIG_64BIT
2308
2309SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2310 long, min_nr,
2311 long, nr,
2312 struct io_event __user *, events,
2313 struct __kernel_timespec __user *, timeout)
2314{
2315 struct timespec64 ts;
2316 int ret;
2317
2318 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2319 return -EFAULT;
2320
2321 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2322 if (!ret && signal_pending(current))
2323 ret = -EINTR;
2324 return ret;
2325}
2326
2327#endif
2328
2329struct __aio_sigset {
2330 const sigset_t __user *sigmask;
2331 size_t sigsetsize;
2332};
2333
2334SYSCALL_DEFINE6(io_pgetevents,
2335 aio_context_t, ctx_id,
2336 long, min_nr,
2337 long, nr,
2338 struct io_event __user *, events,
2339 struct __kernel_timespec __user *, timeout,
2340 const struct __aio_sigset __user *, usig)
2341{
2342 struct __aio_sigset ksig = { NULL, };
2343 struct timespec64 ts;
2344 bool interrupted;
2345 int ret;
2346
2347 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2348 return -EFAULT;
2349
2350 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2351 return -EFAULT;
2352
2353 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2354 if (ret)
2355 return ret;
2356
2357 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : 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#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2368
2369SYSCALL_DEFINE6(io_pgetevents_time32,
2370 aio_context_t, ctx_id,
2371 long, min_nr,
2372 long, nr,
2373 struct io_event __user *, events,
2374 struct old_timespec32 __user *, timeout,
2375 const struct __aio_sigset __user *, usig)
2376{
2377 struct __aio_sigset ksig = { NULL, };
2378 struct timespec64 ts;
2379 bool interrupted;
2380 int ret;
2381
2382 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2383 return -EFAULT;
2384
2385 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2386 return -EFAULT;
2387
2388
2389 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2390 if (ret)
2391 return ret;
2392
2393 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2394
2395 interrupted = signal_pending(current);
2396 restore_saved_sigmask_unless(interrupted);
2397 if (interrupted && !ret)
2398 ret = -ERESTARTNOHAND;
2399
2400 return ret;
2401}
2402
2403#endif
2404
2405#if defined(CONFIG_COMPAT_32BIT_TIME)
2406
2407SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2408 __s32, min_nr,
2409 __s32, nr,
2410 struct io_event __user *, events,
2411 struct old_timespec32 __user *, timeout)
2412{
2413 struct timespec64 t;
2414 int ret;
2415
2416 if (timeout && get_old_timespec32(&t, timeout))
2417 return -EFAULT;
2418
2419 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2420 if (!ret && signal_pending(current))
2421 ret = -EINTR;
2422 return ret;
2423}
2424
2425#endif
2426
2427#ifdef CONFIG_COMPAT
2428
2429struct __compat_aio_sigset {
2430 compat_uptr_t sigmask;
2431 compat_size_t sigsetsize;
2432};
2433
2434#if defined(CONFIG_COMPAT_32BIT_TIME)
2435
2436COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2437 compat_aio_context_t, ctx_id,
2438 compat_long_t, min_nr,
2439 compat_long_t, nr,
2440 struct io_event __user *, events,
2441 struct old_timespec32 __user *, timeout,
2442 const struct __compat_aio_sigset __user *, usig)
2443{
2444 struct __compat_aio_sigset ksig = { 0, };
2445 struct timespec64 t;
2446 bool interrupted;
2447 int ret;
2448
2449 if (timeout && get_old_timespec32(&t, timeout))
2450 return -EFAULT;
2451
2452 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2453 return -EFAULT;
2454
2455 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2456 if (ret)
2457 return ret;
2458
2459 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2460
2461 interrupted = signal_pending(current);
2462 restore_saved_sigmask_unless(interrupted);
2463 if (interrupted && !ret)
2464 ret = -ERESTARTNOHAND;
2465
2466 return ret;
2467}
2468
2469#endif
2470
2471COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2472 compat_aio_context_t, ctx_id,
2473 compat_long_t, min_nr,
2474 compat_long_t, nr,
2475 struct io_event __user *, events,
2476 struct __kernel_timespec __user *, timeout,
2477 const struct __compat_aio_sigset __user *, usig)
2478{
2479 struct __compat_aio_sigset ksig = { 0, };
2480 struct timespec64 t;
2481 bool interrupted;
2482 int ret;
2483
2484 if (timeout && get_timespec64(&t, timeout))
2485 return -EFAULT;
2486
2487 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2488 return -EFAULT;
2489
2490 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2491 if (ret)
2492 return ret;
2493
2494 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2495
2496 interrupted = signal_pending(current);
2497 restore_saved_sigmask_unless(interrupted);
2498 if (interrupted && !ret)
2499 ret = -ERESTARTNOHAND;
2500
2501 return ret;
2502}
2503#endif