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