1
2
3
4
5
6
7
8
9
10
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/mmu_context.h>
31#include <linux/percpu.h>
32#include <linux/slab.h>
33#include <linux/timer.h>
34#include <linux/aio.h>
35#include <linux/highmem.h>
36#include <linux/workqueue.h>
37#include <linux/security.h>
38#include <linux/eventfd.h>
39#include <linux/blkdev.h>
40#include <linux/compat.h>
41#include <linux/migrate.h>
42#include <linux/ramfs.h>
43#include <linux/percpu-refcount.h>
44#include <linux/mount.h>
45
46#include <asm/kmap_types.h>
47#include <linux/uaccess.h>
48#include <linux/nospec.h>
49
50#include "internal.h"
51
52#define KIOCB_KEY 0
53
54#define AIO_RING_MAGIC 0xa10a10a1
55#define AIO_RING_COMPAT_FEATURES 1
56#define AIO_RING_INCOMPAT_FEATURES 0
57struct aio_ring {
58 unsigned id;
59 unsigned nr;
60 unsigned head;
61
62 unsigned tail;
63
64 unsigned magic;
65 unsigned compat_features;
66 unsigned incompat_features;
67 unsigned header_length;
68
69
70 struct io_event io_events[0];
71};
72
73
74
75
76
77#define AIO_PLUG_THRESHOLD 2
78
79#define AIO_RING_PAGES 8
80
81struct kioctx_table {
82 struct rcu_head rcu;
83 unsigned nr;
84 struct kioctx __rcu *table[];
85};
86
87struct kioctx_cpu {
88 unsigned reqs_available;
89};
90
91struct ctx_rq_wait {
92 struct completion comp;
93 atomic_t count;
94};
95
96struct kioctx {
97 struct percpu_ref users;
98 atomic_t dead;
99
100 struct percpu_ref reqs;
101
102 unsigned long user_id;
103
104 struct __percpu kioctx_cpu *cpu;
105
106
107
108
109
110 unsigned req_batch;
111
112
113
114
115
116
117
118 unsigned max_reqs;
119
120
121 unsigned nr_events;
122
123 unsigned long mmap_base;
124 unsigned long mmap_size;
125
126 struct page **ring_pages;
127 long nr_pages;
128
129 struct rcu_work free_rwork;
130
131
132
133
134 struct ctx_rq_wait *rq_wait;
135
136 struct {
137
138
139
140
141
142
143
144
145 atomic_t reqs_available;
146 } ____cacheline_aligned_in_smp;
147
148 struct {
149 spinlock_t ctx_lock;
150 struct list_head active_reqs;
151 } ____cacheline_aligned_in_smp;
152
153 struct {
154 struct mutex ring_lock;
155 wait_queue_head_t wait;
156 } ____cacheline_aligned_in_smp;
157
158 struct {
159 unsigned tail;
160 unsigned completed_events;
161 spinlock_t completion_lock;
162 } ____cacheline_aligned_in_smp;
163
164 struct page *internal_pages[AIO_RING_PAGES];
165 struct file *aio_ring_file;
166
167 unsigned id;
168};
169
170
171
172
173
174struct fsync_iocb {
175 struct file *file;
176 struct work_struct work;
177 bool datasync;
178};
179
180struct poll_iocb {
181 struct file *file;
182 struct wait_queue_head *head;
183 __poll_t events;
184 bool done;
185 bool cancelled;
186 struct wait_queue_entry wait;
187 struct work_struct work;
188};
189
190
191
192
193
194
195
196struct aio_kiocb {
197 union {
198 struct file *ki_filp;
199 struct kiocb rw;
200 struct fsync_iocb fsync;
201 struct poll_iocb poll;
202 };
203
204 struct kioctx *ki_ctx;
205 kiocb_cancel_fn *ki_cancel;
206
207 struct io_event ki_res;
208
209 struct list_head ki_list;
210
211 refcount_t ki_refcnt;
212
213
214
215
216
217 struct eventfd_ctx *ki_eventfd;
218};
219
220
221static DEFINE_SPINLOCK(aio_nr_lock);
222unsigned long aio_nr;
223unsigned long aio_max_nr = 0x10000;
224
225
226static struct kmem_cache *kiocb_cachep;
227static struct kmem_cache *kioctx_cachep;
228
229static struct vfsmount *aio_mnt;
230
231static const struct file_operations aio_ring_fops;
232static const struct address_space_operations aio_ctx_aops;
233
234static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
235{
236 struct file *file;
237 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
238 if (IS_ERR(inode))
239 return ERR_CAST(inode);
240
241 inode->i_mapping->a_ops = &aio_ctx_aops;
242 inode->i_mapping->private_data = ctx;
243 inode->i_size = PAGE_SIZE * nr_pages;
244
245 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
246 O_RDWR, &aio_ring_fops);
247 if (IS_ERR(file))
248 iput(inode);
249 return file;
250}
251
252static struct dentry *aio_mount(struct file_system_type *fs_type,
253 int flags, const char *dev_name, void *data)
254{
255 struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, NULL,
256 AIO_RING_MAGIC);
257
258 if (!IS_ERR(root))
259 root->d_sb->s_iflags |= SB_I_NOEXEC;
260 return root;
261}
262
263
264
265
266
267static int __init aio_setup(void)
268{
269 static struct file_system_type aio_fs = {
270 .name = "aio",
271 .mount = aio_mount,
272 .kill_sb = kill_anon_super,
273 };
274 aio_mnt = kern_mount(&aio_fs);
275 if (IS_ERR(aio_mnt))
276 panic("Failed to create aio fs mount.");
277
278 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
279 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
280 return 0;
281}
282__initcall(aio_setup);
283
284static void put_aio_ring_file(struct kioctx *ctx)
285{
286 struct file *aio_ring_file = ctx->aio_ring_file;
287 struct address_space *i_mapping;
288
289 if (aio_ring_file) {
290 truncate_setsize(file_inode(aio_ring_file), 0);
291
292
293 i_mapping = aio_ring_file->f_mapping;
294 spin_lock(&i_mapping->private_lock);
295 i_mapping->private_data = NULL;
296 ctx->aio_ring_file = NULL;
297 spin_unlock(&i_mapping->private_lock);
298
299 fput(aio_ring_file);
300 }
301}
302
303static void aio_free_ring(struct kioctx *ctx)
304{
305 int i;
306
307
308
309
310 put_aio_ring_file(ctx);
311
312 for (i = 0; i < ctx->nr_pages; i++) {
313 struct page *page;
314 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
315 page_count(ctx->ring_pages[i]));
316 page = ctx->ring_pages[i];
317 if (!page)
318 continue;
319 ctx->ring_pages[i] = NULL;
320 put_page(page);
321 }
322
323 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
324 kfree(ctx->ring_pages);
325 ctx->ring_pages = NULL;
326 }
327}
328
329static int aio_ring_mremap(struct vm_area_struct *vma)
330{
331 struct file *file = vma->vm_file;
332 struct mm_struct *mm = vma->vm_mm;
333 struct kioctx_table *table;
334 int i, res = -EINVAL;
335
336 spin_lock(&mm->ioctx_lock);
337 rcu_read_lock();
338 table = rcu_dereference(mm->ioctx_table);
339 for (i = 0; i < table->nr; i++) {
340 struct kioctx *ctx;
341
342 ctx = rcu_dereference(table->table[i]);
343 if (ctx && ctx->aio_ring_file == file) {
344 if (!atomic_read(&ctx->dead)) {
345 ctx->user_id = ctx->mmap_base = vma->vm_start;
346 res = 0;
347 }
348 break;
349 }
350 }
351
352 rcu_read_unlock();
353 spin_unlock(&mm->ioctx_lock);
354 return res;
355}
356
357static const struct vm_operations_struct aio_ring_vm_ops = {
358 .mremap = aio_ring_mremap,
359#if IS_ENABLED(CONFIG_MMU)
360 .fault = filemap_fault,
361 .map_pages = filemap_map_pages,
362 .page_mkwrite = filemap_page_mkwrite,
363#endif
364};
365
366static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
367{
368 vma->vm_flags |= VM_DONTEXPAND;
369 vma->vm_ops = &aio_ring_vm_ops;
370 return 0;
371}
372
373static const struct file_operations aio_ring_fops = {
374 .mmap = aio_ring_mmap,
375};
376
377#if IS_ENABLED(CONFIG_MIGRATION)
378static int aio_migratepage(struct address_space *mapping, struct page *new,
379 struct page *old, enum migrate_mode mode)
380{
381 struct kioctx *ctx;
382 unsigned long flags;
383 pgoff_t idx;
384 int rc;
385
386
387
388
389
390
391 if (mode == MIGRATE_SYNC_NO_COPY)
392 return -EINVAL;
393
394 rc = 0;
395
396
397 spin_lock(&mapping->private_lock);
398 ctx = mapping->private_data;
399 if (!ctx) {
400 rc = -EINVAL;
401 goto out;
402 }
403
404
405
406
407
408 if (!mutex_trylock(&ctx->ring_lock)) {
409 rc = -EAGAIN;
410 goto out;
411 }
412
413 idx = old->index;
414 if (idx < (pgoff_t)ctx->nr_pages) {
415
416 if (ctx->ring_pages[idx] != old)
417 rc = -EAGAIN;
418 } else
419 rc = -EINVAL;
420
421 if (rc != 0)
422 goto out_unlock;
423
424
425 BUG_ON(PageWriteback(old));
426 get_page(new);
427
428 rc = migrate_page_move_mapping(mapping, new, old, mode, 1);
429 if (rc != MIGRATEPAGE_SUCCESS) {
430 put_page(new);
431 goto out_unlock;
432 }
433
434
435
436
437
438 spin_lock_irqsave(&ctx->completion_lock, flags);
439 migrate_page_copy(new, old);
440 BUG_ON(ctx->ring_pages[idx] != old);
441 ctx->ring_pages[idx] = new;
442 spin_unlock_irqrestore(&ctx->completion_lock, flags);
443
444
445 put_page(old);
446
447out_unlock:
448 mutex_unlock(&ctx->ring_lock);
449out:
450 spin_unlock(&mapping->private_lock);
451 return rc;
452}
453#endif
454
455static const struct address_space_operations aio_ctx_aops = {
456 .set_page_dirty = __set_page_dirty_no_writeback,
457#if IS_ENABLED(CONFIG_MIGRATION)
458 .migratepage = aio_migratepage,
459#endif
460};
461
462static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
463{
464 struct aio_ring *ring;
465 struct mm_struct *mm = current->mm;
466 unsigned long size, unused;
467 int nr_pages;
468 int i;
469 struct file *file;
470
471
472 nr_events += 2;
473
474 size = sizeof(struct aio_ring);
475 size += sizeof(struct io_event) * nr_events;
476
477 nr_pages = PFN_UP(size);
478 if (nr_pages < 0)
479 return -EINVAL;
480
481 file = aio_private_file(ctx, nr_pages);
482 if (IS_ERR(file)) {
483 ctx->aio_ring_file = NULL;
484 return -ENOMEM;
485 }
486
487 ctx->aio_ring_file = file;
488 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
489 / sizeof(struct io_event);
490
491 ctx->ring_pages = ctx->internal_pages;
492 if (nr_pages > AIO_RING_PAGES) {
493 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
494 GFP_KERNEL);
495 if (!ctx->ring_pages) {
496 put_aio_ring_file(ctx);
497 return -ENOMEM;
498 }
499 }
500
501 for (i = 0; i < nr_pages; i++) {
502 struct page *page;
503 page = find_or_create_page(file->f_mapping,
504 i, GFP_HIGHUSER | __GFP_ZERO);
505 if (!page)
506 break;
507 pr_debug("pid(%d) page[%d]->count=%d\n",
508 current->pid, i, page_count(page));
509 SetPageUptodate(page);
510 unlock_page(page);
511
512 ctx->ring_pages[i] = page;
513 }
514 ctx->nr_pages = i;
515
516 if (unlikely(i != nr_pages)) {
517 aio_free_ring(ctx);
518 return -ENOMEM;
519 }
520
521 ctx->mmap_size = nr_pages * PAGE_SIZE;
522 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
523
524 if (down_write_killable(&mm->mmap_sem)) {
525 ctx->mmap_size = 0;
526 aio_free_ring(ctx);
527 return -EINTR;
528 }
529
530 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
531 PROT_READ | PROT_WRITE,
532 MAP_SHARED, 0, &unused, NULL);
533 up_write(&mm->mmap_sem);
534 if (IS_ERR((void *)ctx->mmap_base)) {
535 ctx->mmap_size = 0;
536 aio_free_ring(ctx);
537 return -ENOMEM;
538 }
539
540 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
541
542 ctx->user_id = ctx->mmap_base;
543 ctx->nr_events = nr_events;
544
545 ring = kmap_atomic(ctx->ring_pages[0]);
546 ring->nr = nr_events;
547 ring->id = ~0U;
548 ring->head = ring->tail = 0;
549 ring->magic = AIO_RING_MAGIC;
550 ring->compat_features = AIO_RING_COMPAT_FEATURES;
551 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
552 ring->header_length = sizeof(struct aio_ring);
553 kunmap_atomic(ring);
554 flush_dcache_page(ctx->ring_pages[0]);
555
556 return 0;
557}
558
559#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
560#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
561#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
562
563void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
564{
565 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
566 struct kioctx *ctx = req->ki_ctx;
567 unsigned long flags;
568
569 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
570 return;
571
572 spin_lock_irqsave(&ctx->ctx_lock, flags);
573 list_add_tail(&req->ki_list, &ctx->active_reqs);
574 req->ki_cancel = cancel;
575 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
576}
577EXPORT_SYMBOL(kiocb_set_cancel_fn);
578
579
580
581
582
583
584static void free_ioctx(struct work_struct *work)
585{
586 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
587 free_rwork);
588 pr_debug("freeing %p\n", ctx);
589
590 aio_free_ring(ctx);
591 free_percpu(ctx->cpu);
592 percpu_ref_exit(&ctx->reqs);
593 percpu_ref_exit(&ctx->users);
594 kmem_cache_free(kioctx_cachep, ctx);
595}
596
597static void free_ioctx_reqs(struct percpu_ref *ref)
598{
599 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
600
601
602 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
603 complete(&ctx->rq_wait->comp);
604
605
606 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
607 queue_rcu_work(system_wq, &ctx->free_rwork);
608}
609
610
611
612
613
614
615static void free_ioctx_users(struct percpu_ref *ref)
616{
617 struct kioctx *ctx = container_of(ref, struct kioctx, users);
618 struct aio_kiocb *req;
619
620 spin_lock_irq(&ctx->ctx_lock);
621
622 while (!list_empty(&ctx->active_reqs)) {
623 req = list_first_entry(&ctx->active_reqs,
624 struct aio_kiocb, ki_list);
625 req->ki_cancel(&req->rw);
626 list_del_init(&req->ki_list);
627 }
628
629 spin_unlock_irq(&ctx->ctx_lock);
630
631 percpu_ref_kill(&ctx->reqs);
632 percpu_ref_put(&ctx->reqs);
633}
634
635static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
636{
637 unsigned i, new_nr;
638 struct kioctx_table *table, *old;
639 struct aio_ring *ring;
640
641 spin_lock(&mm->ioctx_lock);
642 table = rcu_dereference_raw(mm->ioctx_table);
643
644 while (1) {
645 if (table)
646 for (i = 0; i < table->nr; i++)
647 if (!rcu_access_pointer(table->table[i])) {
648 ctx->id = i;
649 rcu_assign_pointer(table->table[i], ctx);
650 spin_unlock(&mm->ioctx_lock);
651
652
653
654
655
656 ring = kmap_atomic(ctx->ring_pages[0]);
657 ring->id = ctx->id;
658 kunmap_atomic(ring);
659 return 0;
660 }
661
662 new_nr = (table ? table->nr : 1) * 4;
663 spin_unlock(&mm->ioctx_lock);
664
665 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
666 new_nr, GFP_KERNEL);
667 if (!table)
668 return -ENOMEM;
669
670 table->nr = new_nr;
671
672 spin_lock(&mm->ioctx_lock);
673 old = rcu_dereference_raw(mm->ioctx_table);
674
675 if (!old) {
676 rcu_assign_pointer(mm->ioctx_table, table);
677 } else if (table->nr > old->nr) {
678 memcpy(table->table, old->table,
679 old->nr * sizeof(struct kioctx *));
680
681 rcu_assign_pointer(mm->ioctx_table, table);
682 kfree_rcu(old, rcu);
683 } else {
684 kfree(table);
685 table = old;
686 }
687 }
688}
689
690static void aio_nr_sub(unsigned nr)
691{
692 spin_lock(&aio_nr_lock);
693 if (WARN_ON(aio_nr - nr > aio_nr))
694 aio_nr = 0;
695 else
696 aio_nr -= nr;
697 spin_unlock(&aio_nr_lock);
698}
699
700
701
702
703static struct kioctx *ioctx_alloc(unsigned nr_events)
704{
705 struct mm_struct *mm = current->mm;
706 struct kioctx *ctx;
707 int err = -ENOMEM;
708
709
710
711
712
713 unsigned int max_reqs = nr_events;
714
715
716
717
718
719
720
721
722
723
724 nr_events = max(nr_events, num_possible_cpus() * 4);
725 nr_events *= 2;
726
727
728 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
729 pr_debug("ENOMEM: nr_events too high\n");
730 return ERR_PTR(-EINVAL);
731 }
732
733 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
734 return ERR_PTR(-EAGAIN);
735
736 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
737 if (!ctx)
738 return ERR_PTR(-ENOMEM);
739
740 ctx->max_reqs = max_reqs;
741
742 spin_lock_init(&ctx->ctx_lock);
743 spin_lock_init(&ctx->completion_lock);
744 mutex_init(&ctx->ring_lock);
745
746
747 mutex_lock(&ctx->ring_lock);
748 init_waitqueue_head(&ctx->wait);
749
750 INIT_LIST_HEAD(&ctx->active_reqs);
751
752 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
753 goto err;
754
755 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
756 goto err;
757
758 ctx->cpu = alloc_percpu(struct kioctx_cpu);
759 if (!ctx->cpu)
760 goto err;
761
762 err = aio_setup_ring(ctx, nr_events);
763 if (err < 0)
764 goto err;
765
766 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
767 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
768 if (ctx->req_batch < 1)
769 ctx->req_batch = 1;
770
771
772 spin_lock(&aio_nr_lock);
773 if (aio_nr + ctx->max_reqs > aio_max_nr ||
774 aio_nr + ctx->max_reqs < aio_nr) {
775 spin_unlock(&aio_nr_lock);
776 err = -EAGAIN;
777 goto err_ctx;
778 }
779 aio_nr += ctx->max_reqs;
780 spin_unlock(&aio_nr_lock);
781
782 percpu_ref_get(&ctx->users);
783 percpu_ref_get(&ctx->reqs);
784
785 err = ioctx_add_table(ctx, mm);
786 if (err)
787 goto err_cleanup;
788
789
790 mutex_unlock(&ctx->ring_lock);
791
792 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
793 ctx, ctx->user_id, mm, ctx->nr_events);
794 return ctx;
795
796err_cleanup:
797 aio_nr_sub(ctx->max_reqs);
798err_ctx:
799 atomic_set(&ctx->dead, 1);
800 if (ctx->mmap_size)
801 vm_munmap(ctx->mmap_base, ctx->mmap_size);
802 aio_free_ring(ctx);
803err:
804 mutex_unlock(&ctx->ring_lock);
805 free_percpu(ctx->cpu);
806 percpu_ref_exit(&ctx->reqs);
807 percpu_ref_exit(&ctx->users);
808 kmem_cache_free(kioctx_cachep, ctx);
809 pr_debug("error allocating ioctx %d\n", err);
810 return ERR_PTR(err);
811}
812
813
814
815
816
817
818static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
819 struct ctx_rq_wait *wait)
820{
821 struct kioctx_table *table;
822
823 spin_lock(&mm->ioctx_lock);
824 if (atomic_xchg(&ctx->dead, 1)) {
825 spin_unlock(&mm->ioctx_lock);
826 return -EINVAL;
827 }
828
829 table = rcu_dereference_raw(mm->ioctx_table);
830 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
831 RCU_INIT_POINTER(table->table[ctx->id], NULL);
832 spin_unlock(&mm->ioctx_lock);
833
834
835 wake_up_all(&ctx->wait);
836
837
838
839
840
841
842
843
844 aio_nr_sub(ctx->max_reqs);
845
846 if (ctx->mmap_size)
847 vm_munmap(ctx->mmap_base, ctx->mmap_size);
848
849 ctx->rq_wait = wait;
850 percpu_ref_kill(&ctx->users);
851 return 0;
852}
853
854
855
856
857
858
859
860
861
862void exit_aio(struct mm_struct *mm)
863{
864 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
865 struct ctx_rq_wait wait;
866 int i, skipped;
867
868 if (!table)
869 return;
870
871 atomic_set(&wait.count, table->nr);
872 init_completion(&wait.comp);
873
874 skipped = 0;
875 for (i = 0; i < table->nr; ++i) {
876 struct kioctx *ctx =
877 rcu_dereference_protected(table->table[i], true);
878
879 if (!ctx) {
880 skipped++;
881 continue;
882 }
883
884
885
886
887
888
889
890
891 ctx->mmap_size = 0;
892 kill_ioctx(mm, ctx, &wait);
893 }
894
895 if (!atomic_sub_and_test(skipped, &wait.count)) {
896
897 wait_for_completion(&wait.comp);
898 }
899
900 RCU_INIT_POINTER(mm->ioctx_table, NULL);
901 kfree(table);
902}
903
904static void put_reqs_available(struct kioctx *ctx, unsigned nr)
905{
906 struct kioctx_cpu *kcpu;
907 unsigned long flags;
908
909 local_irq_save(flags);
910 kcpu = this_cpu_ptr(ctx->cpu);
911 kcpu->reqs_available += nr;
912
913 while (kcpu->reqs_available >= ctx->req_batch * 2) {
914 kcpu->reqs_available -= ctx->req_batch;
915 atomic_add(ctx->req_batch, &ctx->reqs_available);
916 }
917
918 local_irq_restore(flags);
919}
920
921static bool __get_reqs_available(struct kioctx *ctx)
922{
923 struct kioctx_cpu *kcpu;
924 bool ret = false;
925 unsigned long flags;
926
927 local_irq_save(flags);
928 kcpu = this_cpu_ptr(ctx->cpu);
929 if (!kcpu->reqs_available) {
930 int old, avail = atomic_read(&ctx->reqs_available);
931
932 do {
933 if (avail < ctx->req_batch)
934 goto out;
935
936 old = avail;
937 avail = atomic_cmpxchg(&ctx->reqs_available,
938 avail, avail - ctx->req_batch);
939 } while (avail != old);
940
941 kcpu->reqs_available += ctx->req_batch;
942 }
943
944 ret = true;
945 kcpu->reqs_available--;
946out:
947 local_irq_restore(flags);
948 return ret;
949}
950
951
952
953
954
955
956
957
958static void refill_reqs_available(struct kioctx *ctx, unsigned head,
959 unsigned tail)
960{
961 unsigned events_in_ring, completed;
962
963
964 head %= ctx->nr_events;
965 if (head <= tail)
966 events_in_ring = tail - head;
967 else
968 events_in_ring = ctx->nr_events - (head - tail);
969
970 completed = ctx->completed_events;
971 if (events_in_ring < completed)
972 completed -= events_in_ring;
973 else
974 completed = 0;
975
976 if (!completed)
977 return;
978
979 ctx->completed_events -= completed;
980 put_reqs_available(ctx, completed);
981}
982
983
984
985
986
987static void user_refill_reqs_available(struct kioctx *ctx)
988{
989 spin_lock_irq(&ctx->completion_lock);
990 if (ctx->completed_events) {
991 struct aio_ring *ring;
992 unsigned head;
993
994
995
996
997
998
999
1000
1001
1002
1003 ring = kmap_atomic(ctx->ring_pages[0]);
1004 head = ring->head;
1005 kunmap_atomic(ring);
1006
1007 refill_reqs_available(ctx, head, ctx->tail);
1008 }
1009
1010 spin_unlock_irq(&ctx->completion_lock);
1011}
1012
1013static bool get_reqs_available(struct kioctx *ctx)
1014{
1015 if (__get_reqs_available(ctx))
1016 return true;
1017 user_refill_reqs_available(ctx);
1018 return __get_reqs_available(ctx);
1019}
1020
1021
1022
1023
1024
1025
1026
1027
1028static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1029{
1030 struct aio_kiocb *req;
1031
1032 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1033 if (unlikely(!req))
1034 return NULL;
1035
1036 if (unlikely(!get_reqs_available(ctx))) {
1037 kmem_cache_free(kiocb_cachep, req);
1038 return NULL;
1039 }
1040
1041 percpu_ref_get(&ctx->reqs);
1042 req->ki_ctx = ctx;
1043 INIT_LIST_HEAD(&req->ki_list);
1044 refcount_set(&req->ki_refcnt, 2);
1045 req->ki_eventfd = NULL;
1046 return req;
1047}
1048
1049static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1050{
1051 struct aio_ring __user *ring = (void __user *)ctx_id;
1052 struct mm_struct *mm = current->mm;
1053 struct kioctx *ctx, *ret = NULL;
1054 struct kioctx_table *table;
1055 unsigned id;
1056
1057 if (get_user(id, &ring->id))
1058 return NULL;
1059
1060 rcu_read_lock();
1061 table = rcu_dereference(mm->ioctx_table);
1062
1063 if (!table || id >= table->nr)
1064 goto out;
1065
1066 id = array_index_nospec(id, table->nr);
1067 ctx = rcu_dereference(table->table[id]);
1068 if (ctx && ctx->user_id == ctx_id) {
1069 if (percpu_ref_tryget_live(&ctx->users))
1070 ret = ctx;
1071 }
1072out:
1073 rcu_read_unlock();
1074 return ret;
1075}
1076
1077static inline void iocb_destroy(struct aio_kiocb *iocb)
1078{
1079 if (iocb->ki_eventfd)
1080 eventfd_ctx_put(iocb->ki_eventfd);
1081 if (iocb->ki_filp)
1082 fput(iocb->ki_filp);
1083 percpu_ref_put(&iocb->ki_ctx->reqs);
1084 kmem_cache_free(kiocb_cachep, iocb);
1085}
1086
1087
1088
1089
1090static void aio_complete(struct aio_kiocb *iocb)
1091{
1092 struct kioctx *ctx = iocb->ki_ctx;
1093 struct aio_ring *ring;
1094 struct io_event *ev_page, *event;
1095 unsigned tail, pos, head;
1096 unsigned long flags;
1097
1098
1099
1100
1101
1102
1103 spin_lock_irqsave(&ctx->completion_lock, flags);
1104
1105 tail = ctx->tail;
1106 pos = tail + AIO_EVENTS_OFFSET;
1107
1108 if (++tail >= ctx->nr_events)
1109 tail = 0;
1110
1111 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1112 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1113
1114 *event = iocb->ki_res;
1115
1116 kunmap_atomic(ev_page);
1117 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1118
1119 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1120 (void __user *)(unsigned long)iocb->ki_res.obj,
1121 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1122
1123
1124
1125
1126 smp_wmb();
1127
1128 ctx->tail = tail;
1129
1130 ring = kmap_atomic(ctx->ring_pages[0]);
1131 head = ring->head;
1132 ring->tail = tail;
1133 kunmap_atomic(ring);
1134 flush_dcache_page(ctx->ring_pages[0]);
1135
1136 ctx->completed_events++;
1137 if (ctx->completed_events > 1)
1138 refill_reqs_available(ctx, head, tail);
1139 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1140
1141 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1142
1143
1144
1145
1146
1147
1148 if (iocb->ki_eventfd)
1149 eventfd_signal(iocb->ki_eventfd, 1);
1150
1151
1152
1153
1154
1155
1156
1157 smp_mb();
1158
1159 if (waitqueue_active(&ctx->wait))
1160 wake_up(&ctx->wait);
1161}
1162
1163static inline void iocb_put(struct aio_kiocb *iocb)
1164{
1165 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1166 aio_complete(iocb);
1167 iocb_destroy(iocb);
1168 }
1169}
1170
1171
1172
1173
1174
1175static long aio_read_events_ring(struct kioctx *ctx,
1176 struct io_event __user *event, long nr)
1177{
1178 struct aio_ring *ring;
1179 unsigned head, tail, pos;
1180 long ret = 0;
1181 int copy_ret;
1182
1183
1184
1185
1186
1187
1188
1189 sched_annotate_sleep();
1190 mutex_lock(&ctx->ring_lock);
1191
1192
1193 ring = kmap_atomic(ctx->ring_pages[0]);
1194 head = ring->head;
1195 tail = ring->tail;
1196 kunmap_atomic(ring);
1197
1198
1199
1200
1201
1202 smp_rmb();
1203
1204 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1205
1206 if (head == tail)
1207 goto out;
1208
1209 head %= ctx->nr_events;
1210 tail %= ctx->nr_events;
1211
1212 while (ret < nr) {
1213 long avail;
1214 struct io_event *ev;
1215 struct page *page;
1216
1217 avail = (head <= tail ? tail : ctx->nr_events) - head;
1218 if (head == tail)
1219 break;
1220
1221 pos = head + AIO_EVENTS_OFFSET;
1222 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1223 pos %= AIO_EVENTS_PER_PAGE;
1224
1225 avail = min(avail, nr - ret);
1226 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1227
1228 ev = kmap(page);
1229 copy_ret = copy_to_user(event + ret, ev + pos,
1230 sizeof(*ev) * avail);
1231 kunmap(page);
1232
1233 if (unlikely(copy_ret)) {
1234 ret = -EFAULT;
1235 goto out;
1236 }
1237
1238 ret += avail;
1239 head += avail;
1240 head %= ctx->nr_events;
1241 }
1242
1243 ring = kmap_atomic(ctx->ring_pages[0]);
1244 ring->head = head;
1245 kunmap_atomic(ring);
1246 flush_dcache_page(ctx->ring_pages[0]);
1247
1248 pr_debug("%li h%u t%u\n", ret, head, tail);
1249out:
1250 mutex_unlock(&ctx->ring_lock);
1251
1252 return ret;
1253}
1254
1255static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1256 struct io_event __user *event, long *i)
1257{
1258 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1259
1260 if (ret > 0)
1261 *i += ret;
1262
1263 if (unlikely(atomic_read(&ctx->dead)))
1264 ret = -EINVAL;
1265
1266 if (!*i)
1267 *i = ret;
1268
1269 return ret < 0 || *i >= min_nr;
1270}
1271
1272static long read_events(struct kioctx *ctx, long min_nr, long nr,
1273 struct io_event __user *event,
1274 ktime_t until)
1275{
1276 long ret = 0;
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292 if (until == 0)
1293 aio_read_events(ctx, min_nr, nr, event, &ret);
1294 else
1295 wait_event_interruptible_hrtimeout(ctx->wait,
1296 aio_read_events(ctx, min_nr, nr, event, &ret),
1297 until);
1298 return ret;
1299}
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1315{
1316 struct kioctx *ioctx = NULL;
1317 unsigned long ctx;
1318 long ret;
1319
1320 ret = get_user(ctx, ctxp);
1321 if (unlikely(ret))
1322 goto out;
1323
1324 ret = -EINVAL;
1325 if (unlikely(ctx || nr_events == 0)) {
1326 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1327 ctx, nr_events);
1328 goto out;
1329 }
1330
1331 ioctx = ioctx_alloc(nr_events);
1332 ret = PTR_ERR(ioctx);
1333 if (!IS_ERR(ioctx)) {
1334 ret = put_user(ioctx->user_id, ctxp);
1335 if (ret)
1336 kill_ioctx(current->mm, ioctx, NULL);
1337 percpu_ref_put(&ioctx->users);
1338 }
1339
1340out:
1341 return ret;
1342}
1343
1344#ifdef CONFIG_COMPAT
1345COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1346{
1347 struct kioctx *ioctx = NULL;
1348 unsigned long ctx;
1349 long ret;
1350
1351 ret = get_user(ctx, ctx32p);
1352 if (unlikely(ret))
1353 goto out;
1354
1355 ret = -EINVAL;
1356 if (unlikely(ctx || nr_events == 0)) {
1357 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1358 ctx, nr_events);
1359 goto out;
1360 }
1361
1362 ioctx = ioctx_alloc(nr_events);
1363 ret = PTR_ERR(ioctx);
1364 if (!IS_ERR(ioctx)) {
1365
1366 ret = put_user((u32)ioctx->user_id, ctx32p);
1367 if (ret)
1368 kill_ioctx(current->mm, ioctx, NULL);
1369 percpu_ref_put(&ioctx->users);
1370 }
1371
1372out:
1373 return ret;
1374}
1375#endif
1376
1377
1378
1379
1380
1381
1382
1383SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1384{
1385 struct kioctx *ioctx = lookup_ioctx(ctx);
1386 if (likely(NULL != ioctx)) {
1387 struct ctx_rq_wait wait;
1388 int ret;
1389
1390 init_completion(&wait.comp);
1391 atomic_set(&wait.count, 1);
1392
1393
1394
1395
1396
1397 ret = kill_ioctx(current->mm, ioctx, &wait);
1398 percpu_ref_put(&ioctx->users);
1399
1400
1401
1402
1403
1404 if (!ret)
1405 wait_for_completion(&wait.comp);
1406
1407 return ret;
1408 }
1409 pr_debug("EINVAL: invalid context id\n");
1410 return -EINVAL;
1411}
1412
1413static void aio_remove_iocb(struct aio_kiocb *iocb)
1414{
1415 struct kioctx *ctx = iocb->ki_ctx;
1416 unsigned long flags;
1417
1418 spin_lock_irqsave(&ctx->ctx_lock, flags);
1419 list_del(&iocb->ki_list);
1420 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1421}
1422
1423static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1424{
1425 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1426
1427 if (!list_empty_careful(&iocb->ki_list))
1428 aio_remove_iocb(iocb);
1429
1430 if (kiocb->ki_flags & IOCB_WRITE) {
1431 struct inode *inode = file_inode(kiocb->ki_filp);
1432
1433
1434
1435
1436
1437 if (S_ISREG(inode->i_mode))
1438 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1439 file_end_write(kiocb->ki_filp);
1440 }
1441
1442 iocb->ki_res.res = res;
1443 iocb->ki_res.res2 = res2;
1444 iocb_put(iocb);
1445}
1446
1447static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1448{
1449 int ret;
1450
1451 req->ki_complete = aio_complete_rw;
1452 req->private = NULL;
1453 req->ki_pos = iocb->aio_offset;
1454 req->ki_flags = iocb_flags(req->ki_filp);
1455 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1456 req->ki_flags |= IOCB_EVENTFD;
1457 req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1458 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1459
1460
1461
1462
1463
1464 ret = ioprio_check_cap(iocb->aio_reqprio);
1465 if (ret) {
1466 pr_debug("aio ioprio check cap error: %d\n", ret);
1467 return ret;
1468 }
1469
1470 req->ki_ioprio = iocb->aio_reqprio;
1471 } else
1472 req->ki_ioprio = get_current_ioprio();
1473
1474 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1475 if (unlikely(ret))
1476 return ret;
1477
1478 req->ki_flags &= ~IOCB_HIPRI;
1479 return 0;
1480}
1481
1482static int aio_setup_rw(int rw, const struct iocb *iocb, struct iovec **iovec,
1483 bool vectored, bool compat, struct iov_iter *iter)
1484{
1485 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1486 size_t len = iocb->aio_nbytes;
1487
1488 if (!vectored) {
1489 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1490 *iovec = NULL;
1491 return ret;
1492 }
1493#ifdef CONFIG_COMPAT
1494 if (compat)
1495 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1496 iter);
1497#endif
1498 return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1499}
1500
1501static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1502{
1503 switch (ret) {
1504 case -EIOCBQUEUED:
1505 break;
1506 case -ERESTARTSYS:
1507 case -ERESTARTNOINTR:
1508 case -ERESTARTNOHAND:
1509 case -ERESTART_RESTARTBLOCK:
1510
1511
1512
1513
1514 ret = -EINTR;
1515
1516 default:
1517 req->ki_complete(req, ret, 0);
1518 }
1519}
1520
1521static int aio_read(struct kiocb *req, const struct iocb *iocb,
1522 bool vectored, bool compat)
1523{
1524 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1525 struct iov_iter iter;
1526 struct file *file;
1527 int ret;
1528
1529 ret = aio_prep_rw(req, iocb);
1530 if (ret)
1531 return ret;
1532 file = req->ki_filp;
1533 if (unlikely(!(file->f_mode & FMODE_READ)))
1534 return -EBADF;
1535 ret = -EINVAL;
1536 if (unlikely(!file->f_op->read_iter))
1537 return -EINVAL;
1538
1539 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1540 if (ret)
1541 return ret;
1542 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1543 if (!ret)
1544 aio_rw_done(req, call_read_iter(file, req, &iter));
1545 kfree(iovec);
1546 return ret;
1547}
1548
1549static int aio_write(struct kiocb *req, const struct iocb *iocb,
1550 bool vectored, bool compat)
1551{
1552 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1553 struct iov_iter iter;
1554 struct file *file;
1555 int ret;
1556
1557 ret = aio_prep_rw(req, iocb);
1558 if (ret)
1559 return ret;
1560 file = req->ki_filp;
1561
1562 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1563 return -EBADF;
1564 if (unlikely(!file->f_op->write_iter))
1565 return -EINVAL;
1566
1567 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1568 if (ret)
1569 return ret;
1570 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1571 if (!ret) {
1572
1573
1574
1575
1576
1577
1578
1579 if (S_ISREG(file_inode(file)->i_mode)) {
1580 __sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true);
1581 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1582 }
1583 req->ki_flags |= IOCB_WRITE;
1584 aio_rw_done(req, call_write_iter(file, req, &iter));
1585 }
1586 kfree(iovec);
1587 return ret;
1588}
1589
1590static void aio_fsync_work(struct work_struct *work)
1591{
1592 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1593
1594 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1595 iocb_put(iocb);
1596}
1597
1598static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1599 bool datasync)
1600{
1601 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1602 iocb->aio_rw_flags))
1603 return -EINVAL;
1604
1605 if (unlikely(!req->file->f_op->fsync))
1606 return -EINVAL;
1607
1608 req->datasync = datasync;
1609 INIT_WORK(&req->work, aio_fsync_work);
1610 schedule_work(&req->work);
1611 return 0;
1612}
1613
1614static void aio_poll_complete_work(struct work_struct *work)
1615{
1616 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1617 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1618 struct poll_table_struct pt = { ._key = req->events };
1619 struct kioctx *ctx = iocb->ki_ctx;
1620 __poll_t mask = 0;
1621
1622 if (!READ_ONCE(req->cancelled))
1623 mask = vfs_poll(req->file, &pt) & req->events;
1624
1625
1626
1627
1628
1629
1630
1631
1632 spin_lock_irq(&ctx->ctx_lock);
1633 if (!mask && !READ_ONCE(req->cancelled)) {
1634 add_wait_queue(req->head, &req->wait);
1635 spin_unlock_irq(&ctx->ctx_lock);
1636 return;
1637 }
1638 list_del_init(&iocb->ki_list);
1639 iocb->ki_res.res = mangle_poll(mask);
1640 req->done = true;
1641 spin_unlock_irq(&ctx->ctx_lock);
1642
1643 iocb_put(iocb);
1644}
1645
1646
1647static int aio_poll_cancel(struct kiocb *iocb)
1648{
1649 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1650 struct poll_iocb *req = &aiocb->poll;
1651
1652 spin_lock(&req->head->lock);
1653 WRITE_ONCE(req->cancelled, true);
1654 if (!list_empty(&req->wait.entry)) {
1655 list_del_init(&req->wait.entry);
1656 schedule_work(&aiocb->poll.work);
1657 }
1658 spin_unlock(&req->head->lock);
1659
1660 return 0;
1661}
1662
1663static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1664 void *key)
1665{
1666 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1667 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1668 __poll_t mask = key_to_poll(key);
1669 unsigned long flags;
1670
1671
1672 if (mask && !(mask & req->events))
1673 return 0;
1674
1675 list_del_init(&req->wait.entry);
1676
1677 if (mask && spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1678
1679
1680
1681
1682
1683
1684 list_del(&iocb->ki_list);
1685 iocb->ki_res.res = mangle_poll(mask);
1686 req->done = true;
1687 spin_unlock_irqrestore(&iocb->ki_ctx->ctx_lock, flags);
1688 iocb_put(iocb);
1689 } else {
1690 schedule_work(&req->work);
1691 }
1692 return 1;
1693}
1694
1695struct aio_poll_table {
1696 struct poll_table_struct pt;
1697 struct aio_kiocb *iocb;
1698 int error;
1699};
1700
1701static void
1702aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1703 struct poll_table_struct *p)
1704{
1705 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1706
1707
1708 if (unlikely(pt->iocb->poll.head)) {
1709 pt->error = -EINVAL;
1710 return;
1711 }
1712
1713 pt->error = 0;
1714 pt->iocb->poll.head = head;
1715 add_wait_queue(head, &pt->iocb->poll.wait);
1716}
1717
1718static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1719{
1720 struct kioctx *ctx = aiocb->ki_ctx;
1721 struct poll_iocb *req = &aiocb->poll;
1722 struct aio_poll_table apt;
1723 bool cancel = false;
1724 __poll_t mask;
1725
1726
1727 if ((u16)iocb->aio_buf != iocb->aio_buf)
1728 return -EINVAL;
1729
1730 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1731 return -EINVAL;
1732
1733 INIT_WORK(&req->work, aio_poll_complete_work);
1734 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1735
1736 req->head = NULL;
1737 req->done = false;
1738 req->cancelled = false;
1739
1740 apt.pt._qproc = aio_poll_queue_proc;
1741 apt.pt._key = req->events;
1742 apt.iocb = aiocb;
1743 apt.error = -EINVAL;
1744
1745
1746 INIT_LIST_HEAD(&req->wait.entry);
1747 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1748
1749 mask = vfs_poll(req->file, &apt.pt) & req->events;
1750 spin_lock_irq(&ctx->ctx_lock);
1751 if (likely(req->head)) {
1752 spin_lock(&req->head->lock);
1753 if (unlikely(list_empty(&req->wait.entry))) {
1754 if (apt.error)
1755 cancel = true;
1756 apt.error = 0;
1757 mask = 0;
1758 }
1759 if (mask || apt.error) {
1760 list_del_init(&req->wait.entry);
1761 } else if (cancel) {
1762 WRITE_ONCE(req->cancelled, true);
1763 } else if (!req->done) {
1764 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1765 aiocb->ki_cancel = aio_poll_cancel;
1766 }
1767 spin_unlock(&req->head->lock);
1768 }
1769 if (mask) {
1770 aiocb->ki_res.res = mangle_poll(mask);
1771 apt.error = 0;
1772 }
1773 spin_unlock_irq(&ctx->ctx_lock);
1774 if (mask)
1775 iocb_put(aiocb);
1776 return apt.error;
1777}
1778
1779static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1780 struct iocb __user *user_iocb, struct aio_kiocb *req,
1781 bool compat)
1782{
1783 req->ki_filp = fget(iocb->aio_fildes);
1784 if (unlikely(!req->ki_filp))
1785 return -EBADF;
1786
1787 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1788 struct eventfd_ctx *eventfd;
1789
1790
1791
1792
1793
1794
1795 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1796 if (IS_ERR(eventfd))
1797 return PTR_ERR(eventfd);
1798
1799 req->ki_eventfd = eventfd;
1800 }
1801
1802 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1803 pr_debug("EFAULT: aio_key\n");
1804 return -EFAULT;
1805 }
1806
1807 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1808 req->ki_res.data = iocb->aio_data;
1809 req->ki_res.res = 0;
1810 req->ki_res.res2 = 0;
1811
1812 switch (iocb->aio_lio_opcode) {
1813 case IOCB_CMD_PREAD:
1814 return aio_read(&req->rw, iocb, false, compat);
1815 case IOCB_CMD_PWRITE:
1816 return aio_write(&req->rw, iocb, false, compat);
1817 case IOCB_CMD_PREADV:
1818 return aio_read(&req->rw, iocb, true, compat);
1819 case IOCB_CMD_PWRITEV:
1820 return aio_write(&req->rw, iocb, true, compat);
1821 case IOCB_CMD_FSYNC:
1822 return aio_fsync(&req->fsync, iocb, false);
1823 case IOCB_CMD_FDSYNC:
1824 return aio_fsync(&req->fsync, iocb, true);
1825 case IOCB_CMD_POLL:
1826 return aio_poll(req, iocb);
1827 default:
1828 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1829 return -EINVAL;
1830 }
1831}
1832
1833static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1834 bool compat)
1835{
1836 struct aio_kiocb *req;
1837 struct iocb iocb;
1838 int err;
1839
1840 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1841 return -EFAULT;
1842
1843
1844 if (unlikely(iocb.aio_reserved2)) {
1845 pr_debug("EINVAL: reserve field set\n");
1846 return -EINVAL;
1847 }
1848
1849
1850 if (unlikely(
1851 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
1852 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
1853 ((ssize_t)iocb.aio_nbytes < 0)
1854 )) {
1855 pr_debug("EINVAL: overflow check\n");
1856 return -EINVAL;
1857 }
1858
1859 req = aio_get_req(ctx);
1860 if (unlikely(!req))
1861 return -EAGAIN;
1862
1863 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
1864
1865
1866 iocb_put(req);
1867
1868
1869
1870
1871
1872
1873 if (unlikely(err)) {
1874 iocb_destroy(req);
1875 put_reqs_available(ctx, 1);
1876 }
1877 return err;
1878}
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1893 struct iocb __user * __user *, iocbpp)
1894{
1895 struct kioctx *ctx;
1896 long ret = 0;
1897 int i = 0;
1898 struct blk_plug plug;
1899
1900 if (unlikely(nr < 0))
1901 return -EINVAL;
1902
1903 ctx = lookup_ioctx(ctx_id);
1904 if (unlikely(!ctx)) {
1905 pr_debug("EINVAL: invalid context id\n");
1906 return -EINVAL;
1907 }
1908
1909 if (nr > ctx->nr_events)
1910 nr = ctx->nr_events;
1911
1912 if (nr > AIO_PLUG_THRESHOLD)
1913 blk_start_plug(&plug);
1914 for (i = 0; i < nr; i++) {
1915 struct iocb __user *user_iocb;
1916
1917 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1918 ret = -EFAULT;
1919 break;
1920 }
1921
1922 ret = io_submit_one(ctx, user_iocb, false);
1923 if (ret)
1924 break;
1925 }
1926 if (nr > AIO_PLUG_THRESHOLD)
1927 blk_finish_plug(&plug);
1928
1929 percpu_ref_put(&ctx->users);
1930 return i ? i : ret;
1931}
1932
1933#ifdef CONFIG_COMPAT
1934COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1935 int, nr, compat_uptr_t __user *, iocbpp)
1936{
1937 struct kioctx *ctx;
1938 long ret = 0;
1939 int i = 0;
1940 struct blk_plug plug;
1941
1942 if (unlikely(nr < 0))
1943 return -EINVAL;
1944
1945 ctx = lookup_ioctx(ctx_id);
1946 if (unlikely(!ctx)) {
1947 pr_debug("EINVAL: invalid context id\n");
1948 return -EINVAL;
1949 }
1950
1951 if (nr > ctx->nr_events)
1952 nr = ctx->nr_events;
1953
1954 if (nr > AIO_PLUG_THRESHOLD)
1955 blk_start_plug(&plug);
1956 for (i = 0; i < nr; i++) {
1957 compat_uptr_t user_iocb;
1958
1959 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1960 ret = -EFAULT;
1961 break;
1962 }
1963
1964 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
1965 if (ret)
1966 break;
1967 }
1968 if (nr > AIO_PLUG_THRESHOLD)
1969 blk_finish_plug(&plug);
1970
1971 percpu_ref_put(&ctx->users);
1972 return i ? i : ret;
1973}
1974#endif
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1987 struct io_event __user *, result)
1988{
1989 struct kioctx *ctx;
1990 struct aio_kiocb *kiocb;
1991 int ret = -EINVAL;
1992 u32 key;
1993 u64 obj = (u64)(unsigned long)iocb;
1994
1995 if (unlikely(get_user(key, &iocb->aio_key)))
1996 return -EFAULT;
1997 if (unlikely(key != KIOCB_KEY))
1998 return -EINVAL;
1999
2000 ctx = lookup_ioctx(ctx_id);
2001 if (unlikely(!ctx))
2002 return -EINVAL;
2003
2004 spin_lock_irq(&ctx->ctx_lock);
2005
2006 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2007 if (kiocb->ki_res.obj == obj) {
2008 ret = kiocb->ki_cancel(&kiocb->rw);
2009 list_del_init(&kiocb->ki_list);
2010 break;
2011 }
2012 }
2013 spin_unlock_irq(&ctx->ctx_lock);
2014
2015 if (!ret) {
2016
2017
2018
2019
2020
2021 ret = -EINPROGRESS;
2022 }
2023
2024 percpu_ref_put(&ctx->users);
2025
2026 return ret;
2027}
2028
2029static long do_io_getevents(aio_context_t ctx_id,
2030 long min_nr,
2031 long nr,
2032 struct io_event __user *events,
2033 struct timespec64 *ts)
2034{
2035 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2036 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2037 long ret = -EINVAL;
2038
2039 if (likely(ioctx)) {
2040 if (likely(min_nr <= nr && min_nr >= 0))
2041 ret = read_events(ioctx, min_nr, nr, events, until);
2042 percpu_ref_put(&ioctx->users);
2043 }
2044
2045 return ret;
2046}
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060#if !defined(CONFIG_64BIT_TIME) || defined(CONFIG_64BIT)
2061
2062SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2063 long, min_nr,
2064 long, nr,
2065 struct io_event __user *, events,
2066 struct __kernel_timespec __user *, timeout)
2067{
2068 struct timespec64 ts;
2069 int ret;
2070
2071 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2072 return -EFAULT;
2073
2074 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2075 if (!ret && signal_pending(current))
2076 ret = -EINTR;
2077 return ret;
2078}
2079
2080#endif
2081
2082struct __aio_sigset {
2083 const sigset_t __user *sigmask;
2084 size_t sigsetsize;
2085};
2086
2087SYSCALL_DEFINE6(io_pgetevents,
2088 aio_context_t, ctx_id,
2089 long, min_nr,
2090 long, nr,
2091 struct io_event __user *, events,
2092 struct __kernel_timespec __user *, timeout,
2093 const struct __aio_sigset __user *, usig)
2094{
2095 struct __aio_sigset ksig = { NULL, };
2096 sigset_t ksigmask, sigsaved;
2097 struct timespec64 ts;
2098 bool interrupted;
2099 int ret;
2100
2101 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2102 return -EFAULT;
2103
2104 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2105 return -EFAULT;
2106
2107 ret = set_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
2108 if (ret)
2109 return ret;
2110
2111 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2112
2113 interrupted = signal_pending(current);
2114 restore_user_sigmask(ksig.sigmask, &sigsaved, interrupted);
2115 if (interrupted && !ret)
2116 ret = -ERESTARTNOHAND;
2117
2118 return ret;
2119}
2120
2121#if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2122
2123SYSCALL_DEFINE6(io_pgetevents_time32,
2124 aio_context_t, ctx_id,
2125 long, min_nr,
2126 long, nr,
2127 struct io_event __user *, events,
2128 struct old_timespec32 __user *, timeout,
2129 const struct __aio_sigset __user *, usig)
2130{
2131 struct __aio_sigset ksig = { NULL, };
2132 sigset_t ksigmask, sigsaved;
2133 struct timespec64 ts;
2134 bool interrupted;
2135 int ret;
2136
2137 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2138 return -EFAULT;
2139
2140 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2141 return -EFAULT;
2142
2143
2144 ret = set_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
2145 if (ret)
2146 return ret;
2147
2148 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2149
2150 interrupted = signal_pending(current);
2151 restore_user_sigmask(ksig.sigmask, &sigsaved, interrupted);
2152 if (interrupted && !ret)
2153 ret = -ERESTARTNOHAND;
2154
2155 return ret;
2156}
2157
2158#endif
2159
2160#if defined(CONFIG_COMPAT_32BIT_TIME)
2161
2162SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2163 __s32, min_nr,
2164 __s32, nr,
2165 struct io_event __user *, events,
2166 struct old_timespec32 __user *, timeout)
2167{
2168 struct timespec64 t;
2169 int ret;
2170
2171 if (timeout && get_old_timespec32(&t, timeout))
2172 return -EFAULT;
2173
2174 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2175 if (!ret && signal_pending(current))
2176 ret = -EINTR;
2177 return ret;
2178}
2179
2180#endif
2181
2182#ifdef CONFIG_COMPAT
2183
2184struct __compat_aio_sigset {
2185 compat_sigset_t __user *sigmask;
2186 compat_size_t sigsetsize;
2187};
2188
2189#if defined(CONFIG_COMPAT_32BIT_TIME)
2190
2191COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2192 compat_aio_context_t, ctx_id,
2193 compat_long_t, min_nr,
2194 compat_long_t, nr,
2195 struct io_event __user *, events,
2196 struct old_timespec32 __user *, timeout,
2197 const struct __compat_aio_sigset __user *, usig)
2198{
2199 struct __compat_aio_sigset ksig = { NULL, };
2200 sigset_t ksigmask, sigsaved;
2201 struct timespec64 t;
2202 bool interrupted;
2203 int ret;
2204
2205 if (timeout && get_old_timespec32(&t, timeout))
2206 return -EFAULT;
2207
2208 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2209 return -EFAULT;
2210
2211 ret = set_compat_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
2212 if (ret)
2213 return ret;
2214
2215 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2216
2217 interrupted = signal_pending(current);
2218 restore_user_sigmask(ksig.sigmask, &sigsaved, interrupted);
2219 if (interrupted && !ret)
2220 ret = -ERESTARTNOHAND;
2221
2222 return ret;
2223}
2224
2225#endif
2226
2227COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2228 compat_aio_context_t, ctx_id,
2229 compat_long_t, min_nr,
2230 compat_long_t, nr,
2231 struct io_event __user *, events,
2232 struct __kernel_timespec __user *, timeout,
2233 const struct __compat_aio_sigset __user *, usig)
2234{
2235 struct __compat_aio_sigset ksig = { NULL, };
2236 sigset_t ksigmask, sigsaved;
2237 struct timespec64 t;
2238 bool interrupted;
2239 int ret;
2240
2241 if (timeout && get_timespec64(&t, timeout))
2242 return -EFAULT;
2243
2244 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2245 return -EFAULT;
2246
2247 ret = set_compat_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
2248 if (ret)
2249 return ret;
2250
2251 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2252
2253 interrupted = signal_pending(current);
2254 restore_user_sigmask(ksig.sigmask, &sigsaved, interrupted);
2255 if (interrupted && !ret)
2256 ret = -ERESTARTNOHAND;
2257
2258 return ret;
2259}
2260#endif
2261