]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - fs/aio.c
aio/migratepages: make aio migrate pages sane
[mirror_ubuntu-artful-kernel.git] / fs / aio.c
1 /*
2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
4 *
5 * Implements an efficient asynchronous io interface.
6 *
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 *
9 * See ../COPYING for licensing terms.
10 */
11 #define pr_fmt(fmt) "%s: " fmt, __func__
12
13 #include <linux/kernel.h>
14 #include <linux/init.h>
15 #include <linux/errno.h>
16 #include <linux/time.h>
17 #include <linux/aio_abi.h>
18 #include <linux/export.h>
19 #include <linux/syscalls.h>
20 #include <linux/backing-dev.h>
21 #include <linux/uio.h>
22
23 #include <linux/sched.h>
24 #include <linux/fs.h>
25 #include <linux/file.h>
26 #include <linux/mm.h>
27 #include <linux/mman.h>
28 #include <linux/mmu_context.h>
29 #include <linux/percpu.h>
30 #include <linux/slab.h>
31 #include <linux/timer.h>
32 #include <linux/aio.h>
33 #include <linux/highmem.h>
34 #include <linux/workqueue.h>
35 #include <linux/security.h>
36 #include <linux/eventfd.h>
37 #include <linux/blkdev.h>
38 #include <linux/compat.h>
39 #include <linux/migrate.h>
40 #include <linux/ramfs.h>
41 #include <linux/percpu-refcount.h>
42 #include <linux/mount.h>
43
44 #include <asm/kmap_types.h>
45 #include <asm/uaccess.h>
46
47 #include "internal.h"
48
49 #define AIO_RING_MAGIC 0xa10a10a1
50 #define AIO_RING_COMPAT_FEATURES 1
51 #define AIO_RING_INCOMPAT_FEATURES 0
52 struct aio_ring {
53 unsigned id; /* kernel internal index number */
54 unsigned nr; /* number of io_events */
55 unsigned head;
56 unsigned tail;
57
58 unsigned magic;
59 unsigned compat_features;
60 unsigned incompat_features;
61 unsigned header_length; /* size of aio_ring */
62
63
64 struct io_event io_events[0];
65 }; /* 128 bytes + ring size */
66
67 #define AIO_RING_PAGES 8
68
69 struct kioctx_table {
70 struct rcu_head rcu;
71 unsigned nr;
72 struct kioctx *table[];
73 };
74
75 struct kioctx_cpu {
76 unsigned reqs_available;
77 };
78
79 struct kioctx {
80 struct percpu_ref users;
81 atomic_t dead;
82
83 struct percpu_ref reqs;
84
85 unsigned long user_id;
86
87 struct __percpu kioctx_cpu *cpu;
88
89 /*
90 * For percpu reqs_available, number of slots we move to/from global
91 * counter at a time:
92 */
93 unsigned req_batch;
94 /*
95 * This is what userspace passed to io_setup(), it's not used for
96 * anything but counting against the global max_reqs quota.
97 *
98 * The real limit is nr_events - 1, which will be larger (see
99 * aio_setup_ring())
100 */
101 unsigned max_reqs;
102
103 /* Size of ringbuffer, in units of struct io_event */
104 unsigned nr_events;
105
106 unsigned long mmap_base;
107 unsigned long mmap_size;
108
109 struct page **ring_pages;
110 long nr_pages;
111
112 struct work_struct free_work;
113
114 struct {
115 /*
116 * This counts the number of available slots in the ringbuffer,
117 * so we avoid overflowing it: it's decremented (if positive)
118 * when allocating a kiocb and incremented when the resulting
119 * io_event is pulled off the ringbuffer.
120 *
121 * We batch accesses to it with a percpu version.
122 */
123 atomic_t reqs_available;
124 } ____cacheline_aligned_in_smp;
125
126 struct {
127 spinlock_t ctx_lock;
128 struct list_head active_reqs; /* used for cancellation */
129 } ____cacheline_aligned_in_smp;
130
131 struct {
132 struct mutex ring_lock;
133 wait_queue_head_t wait;
134 } ____cacheline_aligned_in_smp;
135
136 struct {
137 unsigned tail;
138 spinlock_t completion_lock;
139 } ____cacheline_aligned_in_smp;
140
141 struct page *internal_pages[AIO_RING_PAGES];
142 struct file *aio_ring_file;
143
144 unsigned id;
145 };
146
147 /*------ sysctl variables----*/
148 static DEFINE_SPINLOCK(aio_nr_lock);
149 unsigned long aio_nr; /* current system wide number of aio requests */
150 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
151 /*----end sysctl variables---*/
152
153 static struct kmem_cache *kiocb_cachep;
154 static struct kmem_cache *kioctx_cachep;
155
156 static struct vfsmount *aio_mnt;
157
158 static const struct file_operations aio_ring_fops;
159 static const struct address_space_operations aio_ctx_aops;
160
161 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
162 {
163 struct qstr this = QSTR_INIT("[aio]", 5);
164 struct file *file;
165 struct path path;
166 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
167 if (IS_ERR(inode))
168 return ERR_CAST(inode);
169
170 inode->i_mapping->a_ops = &aio_ctx_aops;
171 inode->i_mapping->private_data = ctx;
172 inode->i_size = PAGE_SIZE * nr_pages;
173
174 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
175 if (!path.dentry) {
176 iput(inode);
177 return ERR_PTR(-ENOMEM);
178 }
179 path.mnt = mntget(aio_mnt);
180
181 d_instantiate(path.dentry, inode);
182 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
183 if (IS_ERR(file)) {
184 path_put(&path);
185 return file;
186 }
187
188 file->f_flags = O_RDWR;
189 file->private_data = ctx;
190 return file;
191 }
192
193 static struct dentry *aio_mount(struct file_system_type *fs_type,
194 int flags, const char *dev_name, void *data)
195 {
196 static const struct dentry_operations ops = {
197 .d_dname = simple_dname,
198 };
199 return mount_pseudo(fs_type, "aio:", NULL, &ops, 0xa10a10a1);
200 }
201
202 /* aio_setup
203 * Creates the slab caches used by the aio routines, panic on
204 * failure as this is done early during the boot sequence.
205 */
206 static int __init aio_setup(void)
207 {
208 static struct file_system_type aio_fs = {
209 .name = "aio",
210 .mount = aio_mount,
211 .kill_sb = kill_anon_super,
212 };
213 aio_mnt = kern_mount(&aio_fs);
214 if (IS_ERR(aio_mnt))
215 panic("Failed to create aio fs mount.");
216
217 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
218 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
219
220 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
221
222 return 0;
223 }
224 __initcall(aio_setup);
225
226 static void put_aio_ring_file(struct kioctx *ctx)
227 {
228 struct file *aio_ring_file = ctx->aio_ring_file;
229 if (aio_ring_file) {
230 truncate_setsize(aio_ring_file->f_inode, 0);
231
232 /* Prevent further access to the kioctx from migratepages */
233 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
234 aio_ring_file->f_inode->i_mapping->private_data = NULL;
235 ctx->aio_ring_file = NULL;
236 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
237
238 fput(aio_ring_file);
239 }
240 }
241
242 static void aio_free_ring(struct kioctx *ctx)
243 {
244 int i;
245
246 for (i = 0; i < ctx->nr_pages; i++) {
247 struct page *page;
248 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
249 page_count(ctx->ring_pages[i]));
250 page = ctx->ring_pages[i];
251 if (!page)
252 continue;
253 ctx->ring_pages[i] = NULL;
254 put_page(page);
255 }
256
257 put_aio_ring_file(ctx);
258
259 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
260 kfree(ctx->ring_pages);
261 ctx->ring_pages = NULL;
262 }
263 }
264
265 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
266 {
267 vma->vm_ops = &generic_file_vm_ops;
268 return 0;
269 }
270
271 static const struct file_operations aio_ring_fops = {
272 .mmap = aio_ring_mmap,
273 };
274
275 static int aio_set_page_dirty(struct page *page)
276 {
277 return 0;
278 }
279
280 #if IS_ENABLED(CONFIG_MIGRATION)
281 static int aio_migratepage(struct address_space *mapping, struct page *new,
282 struct page *old, enum migrate_mode mode)
283 {
284 struct kioctx *ctx;
285 unsigned long flags;
286 int rc;
287
288 rc = 0;
289
290 /* Make sure the old page hasn't already been changed */
291 spin_lock(&mapping->private_lock);
292 ctx = mapping->private_data;
293 if (ctx) {
294 pgoff_t idx;
295 spin_lock_irqsave(&ctx->completion_lock, flags);
296 idx = old->index;
297 if (idx < (pgoff_t)ctx->nr_pages) {
298 if (ctx->ring_pages[idx] != old)
299 rc = -EAGAIN;
300 } else
301 rc = -EINVAL;
302 spin_unlock_irqrestore(&ctx->completion_lock, flags);
303 } else
304 rc = -EINVAL;
305 spin_unlock(&mapping->private_lock);
306
307 if (rc != 0)
308 return rc;
309
310 /* Writeback must be complete */
311 BUG_ON(PageWriteback(old));
312 get_page(new);
313
314 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
315 if (rc != MIGRATEPAGE_SUCCESS) {
316 put_page(new);
317 return rc;
318 }
319
320 /* We can potentially race against kioctx teardown here. Use the
321 * address_space's private data lock to protect the mapping's
322 * private_data.
323 */
324 spin_lock(&mapping->private_lock);
325 ctx = mapping->private_data;
326 if (ctx) {
327 pgoff_t idx;
328 spin_lock_irqsave(&ctx->completion_lock, flags);
329 migrate_page_copy(new, old);
330 idx = old->index;
331 if (idx < (pgoff_t)ctx->nr_pages) {
332 /* And only do the move if things haven't changed */
333 if (ctx->ring_pages[idx] == old)
334 ctx->ring_pages[idx] = new;
335 else
336 rc = -EAGAIN;
337 } else
338 rc = -EINVAL;
339 spin_unlock_irqrestore(&ctx->completion_lock, flags);
340 } else
341 rc = -EBUSY;
342 spin_unlock(&mapping->private_lock);
343
344 if (rc == MIGRATEPAGE_SUCCESS)
345 put_page(old);
346 else
347 put_page(new);
348
349 return rc;
350 }
351 #endif
352
353 static const struct address_space_operations aio_ctx_aops = {
354 .set_page_dirty = aio_set_page_dirty,
355 #if IS_ENABLED(CONFIG_MIGRATION)
356 .migratepage = aio_migratepage,
357 #endif
358 };
359
360 static int aio_setup_ring(struct kioctx *ctx)
361 {
362 struct aio_ring *ring;
363 unsigned nr_events = ctx->max_reqs;
364 struct mm_struct *mm = current->mm;
365 unsigned long size, populate;
366 int nr_pages;
367 int i;
368 struct file *file;
369
370 /* Compensate for the ring buffer's head/tail overlap entry */
371 nr_events += 2; /* 1 is required, 2 for good luck */
372
373 size = sizeof(struct aio_ring);
374 size += sizeof(struct io_event) * nr_events;
375
376 nr_pages = PFN_UP(size);
377 if (nr_pages < 0)
378 return -EINVAL;
379
380 file = aio_private_file(ctx, nr_pages);
381 if (IS_ERR(file)) {
382 ctx->aio_ring_file = NULL;
383 return -EAGAIN;
384 }
385
386 for (i = 0; i < nr_pages; i++) {
387 struct page *page;
388 page = find_or_create_page(file->f_inode->i_mapping,
389 i, GFP_HIGHUSER | __GFP_ZERO);
390 if (!page)
391 break;
392 pr_debug("pid(%d) page[%d]->count=%d\n",
393 current->pid, i, page_count(page));
394 SetPageUptodate(page);
395 SetPageDirty(page);
396 unlock_page(page);
397 }
398 ctx->aio_ring_file = file;
399 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
400 / sizeof(struct io_event);
401
402 ctx->ring_pages = ctx->internal_pages;
403 if (nr_pages > AIO_RING_PAGES) {
404 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
405 GFP_KERNEL);
406 if (!ctx->ring_pages) {
407 put_aio_ring_file(ctx);
408 return -ENOMEM;
409 }
410 }
411
412 ctx->mmap_size = nr_pages * PAGE_SIZE;
413 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
414
415 down_write(&mm->mmap_sem);
416 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
417 PROT_READ | PROT_WRITE,
418 MAP_SHARED | MAP_POPULATE, 0, &populate);
419 if (IS_ERR((void *)ctx->mmap_base)) {
420 up_write(&mm->mmap_sem);
421 ctx->mmap_size = 0;
422 aio_free_ring(ctx);
423 return -EAGAIN;
424 }
425
426 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
427
428 /* We must do this while still holding mmap_sem for write, as we
429 * need to be protected against userspace attempting to mremap()
430 * or munmap() the ring buffer.
431 */
432 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
433 1, 0, ctx->ring_pages, NULL);
434
435 /* Dropping the reference here is safe as the page cache will hold
436 * onto the pages for us. It is also required so that page migration
437 * can unmap the pages and get the right reference count.
438 */
439 for (i = 0; i < ctx->nr_pages; i++)
440 put_page(ctx->ring_pages[i]);
441
442 up_write(&mm->mmap_sem);
443
444 if (unlikely(ctx->nr_pages != nr_pages)) {
445 aio_free_ring(ctx);
446 return -EAGAIN;
447 }
448
449 ctx->user_id = ctx->mmap_base;
450 ctx->nr_events = nr_events; /* trusted copy */
451
452 ring = kmap_atomic(ctx->ring_pages[0]);
453 ring->nr = nr_events; /* user copy */
454 ring->id = ~0U;
455 ring->head = ring->tail = 0;
456 ring->magic = AIO_RING_MAGIC;
457 ring->compat_features = AIO_RING_COMPAT_FEATURES;
458 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
459 ring->header_length = sizeof(struct aio_ring);
460 kunmap_atomic(ring);
461 flush_dcache_page(ctx->ring_pages[0]);
462
463 return 0;
464 }
465
466 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
467 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
468 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
469
470 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
471 {
472 struct kioctx *ctx = req->ki_ctx;
473 unsigned long flags;
474
475 spin_lock_irqsave(&ctx->ctx_lock, flags);
476
477 if (!req->ki_list.next)
478 list_add(&req->ki_list, &ctx->active_reqs);
479
480 req->ki_cancel = cancel;
481
482 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
483 }
484 EXPORT_SYMBOL(kiocb_set_cancel_fn);
485
486 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
487 {
488 kiocb_cancel_fn *old, *cancel;
489
490 /*
491 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
492 * actually has a cancel function, hence the cmpxchg()
493 */
494
495 cancel = ACCESS_ONCE(kiocb->ki_cancel);
496 do {
497 if (!cancel || cancel == KIOCB_CANCELLED)
498 return -EINVAL;
499
500 old = cancel;
501 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
502 } while (cancel != old);
503
504 return cancel(kiocb);
505 }
506
507 static void free_ioctx(struct work_struct *work)
508 {
509 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
510
511 pr_debug("freeing %p\n", ctx);
512
513 aio_free_ring(ctx);
514 free_percpu(ctx->cpu);
515 kmem_cache_free(kioctx_cachep, ctx);
516 }
517
518 static void free_ioctx_reqs(struct percpu_ref *ref)
519 {
520 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
521
522 INIT_WORK(&ctx->free_work, free_ioctx);
523 schedule_work(&ctx->free_work);
524 }
525
526 /*
527 * When this function runs, the kioctx has been removed from the "hash table"
528 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
529 * now it's safe to cancel any that need to be.
530 */
531 static void free_ioctx_users(struct percpu_ref *ref)
532 {
533 struct kioctx *ctx = container_of(ref, struct kioctx, users);
534 struct kiocb *req;
535
536 spin_lock_irq(&ctx->ctx_lock);
537
538 while (!list_empty(&ctx->active_reqs)) {
539 req = list_first_entry(&ctx->active_reqs,
540 struct kiocb, ki_list);
541
542 list_del_init(&req->ki_list);
543 kiocb_cancel(ctx, req);
544 }
545
546 spin_unlock_irq(&ctx->ctx_lock);
547
548 percpu_ref_kill(&ctx->reqs);
549 percpu_ref_put(&ctx->reqs);
550 }
551
552 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
553 {
554 unsigned i, new_nr;
555 struct kioctx_table *table, *old;
556 struct aio_ring *ring;
557
558 spin_lock(&mm->ioctx_lock);
559 rcu_read_lock();
560 table = rcu_dereference(mm->ioctx_table);
561
562 while (1) {
563 if (table)
564 for (i = 0; i < table->nr; i++)
565 if (!table->table[i]) {
566 ctx->id = i;
567 table->table[i] = ctx;
568 rcu_read_unlock();
569 spin_unlock(&mm->ioctx_lock);
570
571 ring = kmap_atomic(ctx->ring_pages[0]);
572 ring->id = ctx->id;
573 kunmap_atomic(ring);
574 return 0;
575 }
576
577 new_nr = (table ? table->nr : 1) * 4;
578
579 rcu_read_unlock();
580 spin_unlock(&mm->ioctx_lock);
581
582 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
583 new_nr, GFP_KERNEL);
584 if (!table)
585 return -ENOMEM;
586
587 table->nr = new_nr;
588
589 spin_lock(&mm->ioctx_lock);
590 rcu_read_lock();
591 old = rcu_dereference(mm->ioctx_table);
592
593 if (!old) {
594 rcu_assign_pointer(mm->ioctx_table, table);
595 } else if (table->nr > old->nr) {
596 memcpy(table->table, old->table,
597 old->nr * sizeof(struct kioctx *));
598
599 rcu_assign_pointer(mm->ioctx_table, table);
600 kfree_rcu(old, rcu);
601 } else {
602 kfree(table);
603 table = old;
604 }
605 }
606 }
607
608 static void aio_nr_sub(unsigned nr)
609 {
610 spin_lock(&aio_nr_lock);
611 if (WARN_ON(aio_nr - nr > aio_nr))
612 aio_nr = 0;
613 else
614 aio_nr -= nr;
615 spin_unlock(&aio_nr_lock);
616 }
617
618 /* ioctx_alloc
619 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
620 */
621 static struct kioctx *ioctx_alloc(unsigned nr_events)
622 {
623 struct mm_struct *mm = current->mm;
624 struct kioctx *ctx;
625 int err = -ENOMEM;
626
627 /*
628 * We keep track of the number of available ringbuffer slots, to prevent
629 * overflow (reqs_available), and we also use percpu counters for this.
630 *
631 * So since up to half the slots might be on other cpu's percpu counters
632 * and unavailable, double nr_events so userspace sees what they
633 * expected: additionally, we move req_batch slots to/from percpu
634 * counters at a time, so make sure that isn't 0:
635 */
636 nr_events = max(nr_events, num_possible_cpus() * 4);
637 nr_events *= 2;
638
639 /* Prevent overflows */
640 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
641 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
642 pr_debug("ENOMEM: nr_events too high\n");
643 return ERR_PTR(-EINVAL);
644 }
645
646 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
647 return ERR_PTR(-EAGAIN);
648
649 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
650 if (!ctx)
651 return ERR_PTR(-ENOMEM);
652
653 ctx->max_reqs = nr_events;
654
655 if (percpu_ref_init(&ctx->users, free_ioctx_users))
656 goto err;
657
658 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs))
659 goto err;
660
661 spin_lock_init(&ctx->ctx_lock);
662 spin_lock_init(&ctx->completion_lock);
663 mutex_init(&ctx->ring_lock);
664 init_waitqueue_head(&ctx->wait);
665
666 INIT_LIST_HEAD(&ctx->active_reqs);
667
668 ctx->cpu = alloc_percpu(struct kioctx_cpu);
669 if (!ctx->cpu)
670 goto err;
671
672 if (aio_setup_ring(ctx) < 0)
673 goto err;
674
675 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
676 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
677 if (ctx->req_batch < 1)
678 ctx->req_batch = 1;
679
680 /* limit the number of system wide aios */
681 spin_lock(&aio_nr_lock);
682 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
683 aio_nr + nr_events < aio_nr) {
684 spin_unlock(&aio_nr_lock);
685 err = -EAGAIN;
686 goto err_ctx;
687 }
688 aio_nr += ctx->max_reqs;
689 spin_unlock(&aio_nr_lock);
690
691 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
692 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
693
694 err = ioctx_add_table(ctx, mm);
695 if (err)
696 goto err_cleanup;
697
698 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
699 ctx, ctx->user_id, mm, ctx->nr_events);
700 return ctx;
701
702 err_cleanup:
703 aio_nr_sub(ctx->max_reqs);
704 err_ctx:
705 aio_free_ring(ctx);
706 err:
707 free_percpu(ctx->cpu);
708 free_percpu(ctx->reqs.pcpu_count);
709 free_percpu(ctx->users.pcpu_count);
710 kmem_cache_free(kioctx_cachep, ctx);
711 pr_debug("error allocating ioctx %d\n", err);
712 return ERR_PTR(err);
713 }
714
715 /* kill_ioctx
716 * Cancels all outstanding aio requests on an aio context. Used
717 * when the processes owning a context have all exited to encourage
718 * the rapid destruction of the kioctx.
719 */
720 static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx)
721 {
722 if (!atomic_xchg(&ctx->dead, 1)) {
723 struct kioctx_table *table;
724
725 spin_lock(&mm->ioctx_lock);
726 rcu_read_lock();
727 table = rcu_dereference(mm->ioctx_table);
728
729 WARN_ON(ctx != table->table[ctx->id]);
730 table->table[ctx->id] = NULL;
731 rcu_read_unlock();
732 spin_unlock(&mm->ioctx_lock);
733
734 /* percpu_ref_kill() will do the necessary call_rcu() */
735 wake_up_all(&ctx->wait);
736
737 /*
738 * It'd be more correct to do this in free_ioctx(), after all
739 * the outstanding kiocbs have finished - but by then io_destroy
740 * has already returned, so io_setup() could potentially return
741 * -EAGAIN with no ioctxs actually in use (as far as userspace
742 * could tell).
743 */
744 aio_nr_sub(ctx->max_reqs);
745
746 if (ctx->mmap_size)
747 vm_munmap(ctx->mmap_base, ctx->mmap_size);
748
749 percpu_ref_kill(&ctx->users);
750 }
751 }
752
753 /* wait_on_sync_kiocb:
754 * Waits on the given sync kiocb to complete.
755 */
756 ssize_t wait_on_sync_kiocb(struct kiocb *req)
757 {
758 while (!req->ki_ctx) {
759 set_current_state(TASK_UNINTERRUPTIBLE);
760 if (req->ki_ctx)
761 break;
762 io_schedule();
763 }
764 __set_current_state(TASK_RUNNING);
765 return req->ki_user_data;
766 }
767 EXPORT_SYMBOL(wait_on_sync_kiocb);
768
769 /*
770 * exit_aio: called when the last user of mm goes away. At this point, there is
771 * no way for any new requests to be submited or any of the io_* syscalls to be
772 * called on the context.
773 *
774 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
775 * them.
776 */
777 void exit_aio(struct mm_struct *mm)
778 {
779 struct kioctx_table *table;
780 struct kioctx *ctx;
781 unsigned i = 0;
782
783 while (1) {
784 rcu_read_lock();
785 table = rcu_dereference(mm->ioctx_table);
786
787 do {
788 if (!table || i >= table->nr) {
789 rcu_read_unlock();
790 rcu_assign_pointer(mm->ioctx_table, NULL);
791 if (table)
792 kfree(table);
793 return;
794 }
795
796 ctx = table->table[i++];
797 } while (!ctx);
798
799 rcu_read_unlock();
800
801 /*
802 * We don't need to bother with munmap() here -
803 * exit_mmap(mm) is coming and it'll unmap everything.
804 * Since aio_free_ring() uses non-zero ->mmap_size
805 * as indicator that it needs to unmap the area,
806 * just set it to 0; aio_free_ring() is the only
807 * place that uses ->mmap_size, so it's safe.
808 */
809 ctx->mmap_size = 0;
810
811 kill_ioctx(mm, ctx);
812 }
813 }
814
815 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
816 {
817 struct kioctx_cpu *kcpu;
818
819 preempt_disable();
820 kcpu = this_cpu_ptr(ctx->cpu);
821
822 kcpu->reqs_available += nr;
823 while (kcpu->reqs_available >= ctx->req_batch * 2) {
824 kcpu->reqs_available -= ctx->req_batch;
825 atomic_add(ctx->req_batch, &ctx->reqs_available);
826 }
827
828 preempt_enable();
829 }
830
831 static bool get_reqs_available(struct kioctx *ctx)
832 {
833 struct kioctx_cpu *kcpu;
834 bool ret = false;
835
836 preempt_disable();
837 kcpu = this_cpu_ptr(ctx->cpu);
838
839 if (!kcpu->reqs_available) {
840 int old, avail = atomic_read(&ctx->reqs_available);
841
842 do {
843 if (avail < ctx->req_batch)
844 goto out;
845
846 old = avail;
847 avail = atomic_cmpxchg(&ctx->reqs_available,
848 avail, avail - ctx->req_batch);
849 } while (avail != old);
850
851 kcpu->reqs_available += ctx->req_batch;
852 }
853
854 ret = true;
855 kcpu->reqs_available--;
856 out:
857 preempt_enable();
858 return ret;
859 }
860
861 /* aio_get_req
862 * Allocate a slot for an aio request.
863 * Returns NULL if no requests are free.
864 */
865 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
866 {
867 struct kiocb *req;
868
869 if (!get_reqs_available(ctx))
870 return NULL;
871
872 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
873 if (unlikely(!req))
874 goto out_put;
875
876 percpu_ref_get(&ctx->reqs);
877
878 req->ki_ctx = ctx;
879 return req;
880 out_put:
881 put_reqs_available(ctx, 1);
882 return NULL;
883 }
884
885 static void kiocb_free(struct kiocb *req)
886 {
887 if (req->ki_filp)
888 fput(req->ki_filp);
889 if (req->ki_eventfd != NULL)
890 eventfd_ctx_put(req->ki_eventfd);
891 kmem_cache_free(kiocb_cachep, req);
892 }
893
894 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
895 {
896 struct aio_ring __user *ring = (void __user *)ctx_id;
897 struct mm_struct *mm = current->mm;
898 struct kioctx *ctx, *ret = NULL;
899 struct kioctx_table *table;
900 unsigned id;
901
902 if (get_user(id, &ring->id))
903 return NULL;
904
905 rcu_read_lock();
906 table = rcu_dereference(mm->ioctx_table);
907
908 if (!table || id >= table->nr)
909 goto out;
910
911 ctx = table->table[id];
912 if (ctx && ctx->user_id == ctx_id) {
913 percpu_ref_get(&ctx->users);
914 ret = ctx;
915 }
916 out:
917 rcu_read_unlock();
918 return ret;
919 }
920
921 /* aio_complete
922 * Called when the io request on the given iocb is complete.
923 */
924 void aio_complete(struct kiocb *iocb, long res, long res2)
925 {
926 struct kioctx *ctx = iocb->ki_ctx;
927 struct aio_ring *ring;
928 struct io_event *ev_page, *event;
929 unsigned long flags;
930 unsigned tail, pos;
931
932 /*
933 * Special case handling for sync iocbs:
934 * - events go directly into the iocb for fast handling
935 * - the sync task with the iocb in its stack holds the single iocb
936 * ref, no other paths have a way to get another ref
937 * - the sync task helpfully left a reference to itself in the iocb
938 */
939 if (is_sync_kiocb(iocb)) {
940 iocb->ki_user_data = res;
941 smp_wmb();
942 iocb->ki_ctx = ERR_PTR(-EXDEV);
943 wake_up_process(iocb->ki_obj.tsk);
944 return;
945 }
946
947 if (iocb->ki_list.next) {
948 unsigned long flags;
949
950 spin_lock_irqsave(&ctx->ctx_lock, flags);
951 list_del(&iocb->ki_list);
952 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
953 }
954
955 /*
956 * Add a completion event to the ring buffer. Must be done holding
957 * ctx->completion_lock to prevent other code from messing with the tail
958 * pointer since we might be called from irq context.
959 */
960 spin_lock_irqsave(&ctx->completion_lock, flags);
961
962 tail = ctx->tail;
963 pos = tail + AIO_EVENTS_OFFSET;
964
965 if (++tail >= ctx->nr_events)
966 tail = 0;
967
968 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
969 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
970
971 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
972 event->data = iocb->ki_user_data;
973 event->res = res;
974 event->res2 = res2;
975
976 kunmap_atomic(ev_page);
977 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
978
979 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
980 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
981 res, res2);
982
983 /* after flagging the request as done, we
984 * must never even look at it again
985 */
986 smp_wmb(); /* make event visible before updating tail */
987
988 ctx->tail = tail;
989
990 ring = kmap_atomic(ctx->ring_pages[0]);
991 ring->tail = tail;
992 kunmap_atomic(ring);
993 flush_dcache_page(ctx->ring_pages[0]);
994
995 spin_unlock_irqrestore(&ctx->completion_lock, flags);
996
997 pr_debug("added to ring %p at [%u]\n", iocb, tail);
998
999 /*
1000 * Check if the user asked us to deliver the result through an
1001 * eventfd. The eventfd_signal() function is safe to be called
1002 * from IRQ context.
1003 */
1004 if (iocb->ki_eventfd != NULL)
1005 eventfd_signal(iocb->ki_eventfd, 1);
1006
1007 /* everything turned out well, dispose of the aiocb. */
1008 kiocb_free(iocb);
1009
1010 /*
1011 * We have to order our ring_info tail store above and test
1012 * of the wait list below outside the wait lock. This is
1013 * like in wake_up_bit() where clearing a bit has to be
1014 * ordered with the unlocked test.
1015 */
1016 smp_mb();
1017
1018 if (waitqueue_active(&ctx->wait))
1019 wake_up(&ctx->wait);
1020
1021 percpu_ref_put(&ctx->reqs);
1022 }
1023 EXPORT_SYMBOL(aio_complete);
1024
1025 /* aio_read_events
1026 * Pull an event off of the ioctx's event ring. Returns the number of
1027 * events fetched
1028 */
1029 static long aio_read_events_ring(struct kioctx *ctx,
1030 struct io_event __user *event, long nr)
1031 {
1032 struct aio_ring *ring;
1033 unsigned head, tail, pos;
1034 long ret = 0;
1035 int copy_ret;
1036
1037 mutex_lock(&ctx->ring_lock);
1038
1039 ring = kmap_atomic(ctx->ring_pages[0]);
1040 head = ring->head;
1041 tail = ring->tail;
1042 kunmap_atomic(ring);
1043
1044 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1045
1046 if (head == tail)
1047 goto out;
1048
1049 while (ret < nr) {
1050 long avail;
1051 struct io_event *ev;
1052 struct page *page;
1053
1054 avail = (head <= tail ? tail : ctx->nr_events) - head;
1055 if (head == tail)
1056 break;
1057
1058 avail = min(avail, nr - ret);
1059 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1060 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1061
1062 pos = head + AIO_EVENTS_OFFSET;
1063 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1064 pos %= AIO_EVENTS_PER_PAGE;
1065
1066 ev = kmap(page);
1067 copy_ret = copy_to_user(event + ret, ev + pos,
1068 sizeof(*ev) * avail);
1069 kunmap(page);
1070
1071 if (unlikely(copy_ret)) {
1072 ret = -EFAULT;
1073 goto out;
1074 }
1075
1076 ret += avail;
1077 head += avail;
1078 head %= ctx->nr_events;
1079 }
1080
1081 ring = kmap_atomic(ctx->ring_pages[0]);
1082 ring->head = head;
1083 kunmap_atomic(ring);
1084 flush_dcache_page(ctx->ring_pages[0]);
1085
1086 pr_debug("%li h%u t%u\n", ret, head, tail);
1087
1088 put_reqs_available(ctx, ret);
1089 out:
1090 mutex_unlock(&ctx->ring_lock);
1091
1092 return ret;
1093 }
1094
1095 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1096 struct io_event __user *event, long *i)
1097 {
1098 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1099
1100 if (ret > 0)
1101 *i += ret;
1102
1103 if (unlikely(atomic_read(&ctx->dead)))
1104 ret = -EINVAL;
1105
1106 if (!*i)
1107 *i = ret;
1108
1109 return ret < 0 || *i >= min_nr;
1110 }
1111
1112 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1113 struct io_event __user *event,
1114 struct timespec __user *timeout)
1115 {
1116 ktime_t until = { .tv64 = KTIME_MAX };
1117 long ret = 0;
1118
1119 if (timeout) {
1120 struct timespec ts;
1121
1122 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1123 return -EFAULT;
1124
1125 until = timespec_to_ktime(ts);
1126 }
1127
1128 /*
1129 * Note that aio_read_events() is being called as the conditional - i.e.
1130 * we're calling it after prepare_to_wait() has set task state to
1131 * TASK_INTERRUPTIBLE.
1132 *
1133 * But aio_read_events() can block, and if it blocks it's going to flip
1134 * the task state back to TASK_RUNNING.
1135 *
1136 * This should be ok, provided it doesn't flip the state back to
1137 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1138 * will only happen if the mutex_lock() call blocks, and we then find
1139 * the ringbuffer empty. So in practice we should be ok, but it's
1140 * something to be aware of when touching this code.
1141 */
1142 wait_event_interruptible_hrtimeout(ctx->wait,
1143 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1144
1145 if (!ret && signal_pending(current))
1146 ret = -EINTR;
1147
1148 return ret;
1149 }
1150
1151 /* sys_io_setup:
1152 * Create an aio_context capable of receiving at least nr_events.
1153 * ctxp must not point to an aio_context that already exists, and
1154 * must be initialized to 0 prior to the call. On successful
1155 * creation of the aio_context, *ctxp is filled in with the resulting
1156 * handle. May fail with -EINVAL if *ctxp is not initialized,
1157 * if the specified nr_events exceeds internal limits. May fail
1158 * with -EAGAIN if the specified nr_events exceeds the user's limit
1159 * of available events. May fail with -ENOMEM if insufficient kernel
1160 * resources are available. May fail with -EFAULT if an invalid
1161 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1162 * implemented.
1163 */
1164 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1165 {
1166 struct kioctx *ioctx = NULL;
1167 unsigned long ctx;
1168 long ret;
1169
1170 ret = get_user(ctx, ctxp);
1171 if (unlikely(ret))
1172 goto out;
1173
1174 ret = -EINVAL;
1175 if (unlikely(ctx || nr_events == 0)) {
1176 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1177 ctx, nr_events);
1178 goto out;
1179 }
1180
1181 ioctx = ioctx_alloc(nr_events);
1182 ret = PTR_ERR(ioctx);
1183 if (!IS_ERR(ioctx)) {
1184 ret = put_user(ioctx->user_id, ctxp);
1185 if (ret)
1186 kill_ioctx(current->mm, ioctx);
1187 percpu_ref_put(&ioctx->users);
1188 }
1189
1190 out:
1191 return ret;
1192 }
1193
1194 /* sys_io_destroy:
1195 * Destroy the aio_context specified. May cancel any outstanding
1196 * AIOs and block on completion. Will fail with -ENOSYS if not
1197 * implemented. May fail with -EINVAL if the context pointed to
1198 * is invalid.
1199 */
1200 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1201 {
1202 struct kioctx *ioctx = lookup_ioctx(ctx);
1203 if (likely(NULL != ioctx)) {
1204 kill_ioctx(current->mm, ioctx);
1205 percpu_ref_put(&ioctx->users);
1206 return 0;
1207 }
1208 pr_debug("EINVAL: io_destroy: invalid context id\n");
1209 return -EINVAL;
1210 }
1211
1212 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1213 unsigned long, loff_t);
1214
1215 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1216 int rw, char __user *buf,
1217 unsigned long *nr_segs,
1218 struct iovec **iovec,
1219 bool compat)
1220 {
1221 ssize_t ret;
1222
1223 *nr_segs = kiocb->ki_nbytes;
1224
1225 #ifdef CONFIG_COMPAT
1226 if (compat)
1227 ret = compat_rw_copy_check_uvector(rw,
1228 (struct compat_iovec __user *)buf,
1229 *nr_segs, 1, *iovec, iovec);
1230 else
1231 #endif
1232 ret = rw_copy_check_uvector(rw,
1233 (struct iovec __user *)buf,
1234 *nr_segs, 1, *iovec, iovec);
1235 if (ret < 0)
1236 return ret;
1237
1238 /* ki_nbytes now reflect bytes instead of segs */
1239 kiocb->ki_nbytes = ret;
1240 return 0;
1241 }
1242
1243 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1244 int rw, char __user *buf,
1245 unsigned long *nr_segs,
1246 struct iovec *iovec)
1247 {
1248 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1249 return -EFAULT;
1250
1251 iovec->iov_base = buf;
1252 iovec->iov_len = kiocb->ki_nbytes;
1253 *nr_segs = 1;
1254 return 0;
1255 }
1256
1257 /*
1258 * aio_setup_iocb:
1259 * Performs the initial checks and aio retry method
1260 * setup for the kiocb at the time of io submission.
1261 */
1262 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1263 char __user *buf, bool compat)
1264 {
1265 struct file *file = req->ki_filp;
1266 ssize_t ret;
1267 unsigned long nr_segs;
1268 int rw;
1269 fmode_t mode;
1270 aio_rw_op *rw_op;
1271 struct iovec inline_vec, *iovec = &inline_vec;
1272
1273 switch (opcode) {
1274 case IOCB_CMD_PREAD:
1275 case IOCB_CMD_PREADV:
1276 mode = FMODE_READ;
1277 rw = READ;
1278 rw_op = file->f_op->aio_read;
1279 goto rw_common;
1280
1281 case IOCB_CMD_PWRITE:
1282 case IOCB_CMD_PWRITEV:
1283 mode = FMODE_WRITE;
1284 rw = WRITE;
1285 rw_op = file->f_op->aio_write;
1286 goto rw_common;
1287 rw_common:
1288 if (unlikely(!(file->f_mode & mode)))
1289 return -EBADF;
1290
1291 if (!rw_op)
1292 return -EINVAL;
1293
1294 ret = (opcode == IOCB_CMD_PREADV ||
1295 opcode == IOCB_CMD_PWRITEV)
1296 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1297 &iovec, compat)
1298 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1299 iovec);
1300 if (ret)
1301 return ret;
1302
1303 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1304 if (ret < 0) {
1305 if (iovec != &inline_vec)
1306 kfree(iovec);
1307 return ret;
1308 }
1309
1310 req->ki_nbytes = ret;
1311
1312 /* XXX: move/kill - rw_verify_area()? */
1313 /* This matches the pread()/pwrite() logic */
1314 if (req->ki_pos < 0) {
1315 ret = -EINVAL;
1316 break;
1317 }
1318
1319 if (rw == WRITE)
1320 file_start_write(file);
1321
1322 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1323
1324 if (rw == WRITE)
1325 file_end_write(file);
1326 break;
1327
1328 case IOCB_CMD_FDSYNC:
1329 if (!file->f_op->aio_fsync)
1330 return -EINVAL;
1331
1332 ret = file->f_op->aio_fsync(req, 1);
1333 break;
1334
1335 case IOCB_CMD_FSYNC:
1336 if (!file->f_op->aio_fsync)
1337 return -EINVAL;
1338
1339 ret = file->f_op->aio_fsync(req, 0);
1340 break;
1341
1342 default:
1343 pr_debug("EINVAL: no operation provided\n");
1344 return -EINVAL;
1345 }
1346
1347 if (iovec != &inline_vec)
1348 kfree(iovec);
1349
1350 if (ret != -EIOCBQUEUED) {
1351 /*
1352 * There's no easy way to restart the syscall since other AIO's
1353 * may be already running. Just fail this IO with EINTR.
1354 */
1355 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1356 ret == -ERESTARTNOHAND ||
1357 ret == -ERESTART_RESTARTBLOCK))
1358 ret = -EINTR;
1359 aio_complete(req, ret, 0);
1360 }
1361
1362 return 0;
1363 }
1364
1365 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1366 struct iocb *iocb, bool compat)
1367 {
1368 struct kiocb *req;
1369 ssize_t ret;
1370
1371 /* enforce forwards compatibility on users */
1372 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1373 pr_debug("EINVAL: reserve field set\n");
1374 return -EINVAL;
1375 }
1376
1377 /* prevent overflows */
1378 if (unlikely(
1379 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1380 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1381 ((ssize_t)iocb->aio_nbytes < 0)
1382 )) {
1383 pr_debug("EINVAL: io_submit: overflow check\n");
1384 return -EINVAL;
1385 }
1386
1387 req = aio_get_req(ctx);
1388 if (unlikely(!req))
1389 return -EAGAIN;
1390
1391 req->ki_filp = fget(iocb->aio_fildes);
1392 if (unlikely(!req->ki_filp)) {
1393 ret = -EBADF;
1394 goto out_put_req;
1395 }
1396
1397 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1398 /*
1399 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1400 * instance of the file* now. The file descriptor must be
1401 * an eventfd() fd, and will be signaled for each completed
1402 * event using the eventfd_signal() function.
1403 */
1404 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1405 if (IS_ERR(req->ki_eventfd)) {
1406 ret = PTR_ERR(req->ki_eventfd);
1407 req->ki_eventfd = NULL;
1408 goto out_put_req;
1409 }
1410 }
1411
1412 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1413 if (unlikely(ret)) {
1414 pr_debug("EFAULT: aio_key\n");
1415 goto out_put_req;
1416 }
1417
1418 req->ki_obj.user = user_iocb;
1419 req->ki_user_data = iocb->aio_data;
1420 req->ki_pos = iocb->aio_offset;
1421 req->ki_nbytes = iocb->aio_nbytes;
1422
1423 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1424 (char __user *)(unsigned long)iocb->aio_buf,
1425 compat);
1426 if (ret)
1427 goto out_put_req;
1428
1429 return 0;
1430 out_put_req:
1431 put_reqs_available(ctx, 1);
1432 percpu_ref_put(&ctx->reqs);
1433 kiocb_free(req);
1434 return ret;
1435 }
1436
1437 long do_io_submit(aio_context_t ctx_id, long nr,
1438 struct iocb __user *__user *iocbpp, bool compat)
1439 {
1440 struct kioctx *ctx;
1441 long ret = 0;
1442 int i = 0;
1443 struct blk_plug plug;
1444
1445 if (unlikely(nr < 0))
1446 return -EINVAL;
1447
1448 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1449 nr = LONG_MAX/sizeof(*iocbpp);
1450
1451 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1452 return -EFAULT;
1453
1454 ctx = lookup_ioctx(ctx_id);
1455 if (unlikely(!ctx)) {
1456 pr_debug("EINVAL: invalid context id\n");
1457 return -EINVAL;
1458 }
1459
1460 blk_start_plug(&plug);
1461
1462 /*
1463 * AKPM: should this return a partial result if some of the IOs were
1464 * successfully submitted?
1465 */
1466 for (i=0; i<nr; i++) {
1467 struct iocb __user *user_iocb;
1468 struct iocb tmp;
1469
1470 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1471 ret = -EFAULT;
1472 break;
1473 }
1474
1475 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1476 ret = -EFAULT;
1477 break;
1478 }
1479
1480 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1481 if (ret)
1482 break;
1483 }
1484 blk_finish_plug(&plug);
1485
1486 percpu_ref_put(&ctx->users);
1487 return i ? i : ret;
1488 }
1489
1490 /* sys_io_submit:
1491 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1492 * the number of iocbs queued. May return -EINVAL if the aio_context
1493 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1494 * *iocbpp[0] is not properly initialized, if the operation specified
1495 * is invalid for the file descriptor in the iocb. May fail with
1496 * -EFAULT if any of the data structures point to invalid data. May
1497 * fail with -EBADF if the file descriptor specified in the first
1498 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1499 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1500 * fail with -ENOSYS if not implemented.
1501 */
1502 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1503 struct iocb __user * __user *, iocbpp)
1504 {
1505 return do_io_submit(ctx_id, nr, iocbpp, 0);
1506 }
1507
1508 /* lookup_kiocb
1509 * Finds a given iocb for cancellation.
1510 */
1511 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1512 u32 key)
1513 {
1514 struct list_head *pos;
1515
1516 assert_spin_locked(&ctx->ctx_lock);
1517
1518 if (key != KIOCB_KEY)
1519 return NULL;
1520
1521 /* TODO: use a hash or array, this sucks. */
1522 list_for_each(pos, &ctx->active_reqs) {
1523 struct kiocb *kiocb = list_kiocb(pos);
1524 if (kiocb->ki_obj.user == iocb)
1525 return kiocb;
1526 }
1527 return NULL;
1528 }
1529
1530 /* sys_io_cancel:
1531 * Attempts to cancel an iocb previously passed to io_submit. If
1532 * the operation is successfully cancelled, the resulting event is
1533 * copied into the memory pointed to by result without being placed
1534 * into the completion queue and 0 is returned. May fail with
1535 * -EFAULT if any of the data structures pointed to are invalid.
1536 * May fail with -EINVAL if aio_context specified by ctx_id is
1537 * invalid. May fail with -EAGAIN if the iocb specified was not
1538 * cancelled. Will fail with -ENOSYS if not implemented.
1539 */
1540 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1541 struct io_event __user *, result)
1542 {
1543 struct kioctx *ctx;
1544 struct kiocb *kiocb;
1545 u32 key;
1546 int ret;
1547
1548 ret = get_user(key, &iocb->aio_key);
1549 if (unlikely(ret))
1550 return -EFAULT;
1551
1552 ctx = lookup_ioctx(ctx_id);
1553 if (unlikely(!ctx))
1554 return -EINVAL;
1555
1556 spin_lock_irq(&ctx->ctx_lock);
1557
1558 kiocb = lookup_kiocb(ctx, iocb, key);
1559 if (kiocb)
1560 ret = kiocb_cancel(ctx, kiocb);
1561 else
1562 ret = -EINVAL;
1563
1564 spin_unlock_irq(&ctx->ctx_lock);
1565
1566 if (!ret) {
1567 /*
1568 * The result argument is no longer used - the io_event is
1569 * always delivered via the ring buffer. -EINPROGRESS indicates
1570 * cancellation is progress:
1571 */
1572 ret = -EINPROGRESS;
1573 }
1574
1575 percpu_ref_put(&ctx->users);
1576
1577 return ret;
1578 }
1579
1580 /* io_getevents:
1581 * Attempts to read at least min_nr events and up to nr events from
1582 * the completion queue for the aio_context specified by ctx_id. If
1583 * it succeeds, the number of read events is returned. May fail with
1584 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1585 * out of range, if timeout is out of range. May fail with -EFAULT
1586 * if any of the memory specified is invalid. May return 0 or
1587 * < min_nr if the timeout specified by timeout has elapsed
1588 * before sufficient events are available, where timeout == NULL
1589 * specifies an infinite timeout. Note that the timeout pointed to by
1590 * timeout is relative. Will fail with -ENOSYS if not implemented.
1591 */
1592 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1593 long, min_nr,
1594 long, nr,
1595 struct io_event __user *, events,
1596 struct timespec __user *, timeout)
1597 {
1598 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1599 long ret = -EINVAL;
1600
1601 if (likely(ioctx)) {
1602 if (likely(min_nr <= nr && min_nr >= 0))
1603 ret = read_events(ioctx, min_nr, nr, events, timeout);
1604 percpu_ref_put(&ioctx->users);
1605 }
1606 return ret;
1607 }