xref: /openbmc/linux/fs/aio.c (revision df2634f43f5106947f3735a0b61a6527a4b278cd)
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 #include <linux/kernel.h>
12 #include <linux/init.h>
13 #include <linux/errno.h>
14 #include <linux/time.h>
15 #include <linux/aio_abi.h>
16 #include <linux/module.h>
17 #include <linux/syscalls.h>
18 #include <linux/backing-dev.h>
19 #include <linux/uio.h>
20 
21 #define DEBUG 0
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/slab.h>
30 #include <linux/timer.h>
31 #include <linux/aio.h>
32 #include <linux/highmem.h>
33 #include <linux/workqueue.h>
34 #include <linux/security.h>
35 #include <linux/eventfd.h>
36 #include <linux/blkdev.h>
37 #include <linux/mempool.h>
38 #include <linux/hash.h>
39 #include <linux/compat.h>
40 
41 #include <asm/kmap_types.h>
42 #include <asm/uaccess.h>
43 
44 #if DEBUG > 1
45 #define dprintk		printk
46 #else
47 #define dprintk(x...)	do { ; } while (0)
48 #endif
49 
50 /*------ sysctl variables----*/
51 static DEFINE_SPINLOCK(aio_nr_lock);
52 unsigned long aio_nr;		/* current system wide number of aio requests */
53 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
54 /*----end sysctl variables---*/
55 
56 static struct kmem_cache	*kiocb_cachep;
57 static struct kmem_cache	*kioctx_cachep;
58 
59 static struct workqueue_struct *aio_wq;
60 
61 /* Used for rare fput completion. */
62 static void aio_fput_routine(struct work_struct *);
63 static DECLARE_WORK(fput_work, aio_fput_routine);
64 
65 static DEFINE_SPINLOCK(fput_lock);
66 static LIST_HEAD(fput_head);
67 
68 #define AIO_BATCH_HASH_BITS	3 /* allocated on-stack, so don't go crazy */
69 #define AIO_BATCH_HASH_SIZE	(1 << AIO_BATCH_HASH_BITS)
70 struct aio_batch_entry {
71 	struct hlist_node list;
72 	struct address_space *mapping;
73 };
74 mempool_t *abe_pool;
75 
76 static void aio_kick_handler(struct work_struct *);
77 static void aio_queue_work(struct kioctx *);
78 
79 /* aio_setup
80  *	Creates the slab caches used by the aio routines, panic on
81  *	failure as this is done early during the boot sequence.
82  */
83 static int __init aio_setup(void)
84 {
85 	kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
86 	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
87 
88 	aio_wq = create_workqueue("aio");
89 	abe_pool = mempool_create_kmalloc_pool(1, sizeof(struct aio_batch_entry));
90 	BUG_ON(!aio_wq || !abe_pool);
91 
92 	pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
93 
94 	return 0;
95 }
96 __initcall(aio_setup);
97 
98 static void aio_free_ring(struct kioctx *ctx)
99 {
100 	struct aio_ring_info *info = &ctx->ring_info;
101 	long i;
102 
103 	for (i=0; i<info->nr_pages; i++)
104 		put_page(info->ring_pages[i]);
105 
106 	if (info->mmap_size) {
107 		down_write(&ctx->mm->mmap_sem);
108 		do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
109 		up_write(&ctx->mm->mmap_sem);
110 	}
111 
112 	if (info->ring_pages && info->ring_pages != info->internal_pages)
113 		kfree(info->ring_pages);
114 	info->ring_pages = NULL;
115 	info->nr = 0;
116 }
117 
118 static int aio_setup_ring(struct kioctx *ctx)
119 {
120 	struct aio_ring *ring;
121 	struct aio_ring_info *info = &ctx->ring_info;
122 	unsigned nr_events = ctx->max_reqs;
123 	unsigned long size;
124 	int nr_pages;
125 
126 	/* Compensate for the ring buffer's head/tail overlap entry */
127 	nr_events += 2;	/* 1 is required, 2 for good luck */
128 
129 	size = sizeof(struct aio_ring);
130 	size += sizeof(struct io_event) * nr_events;
131 	nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
132 
133 	if (nr_pages < 0)
134 		return -EINVAL;
135 
136 	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
137 
138 	info->nr = 0;
139 	info->ring_pages = info->internal_pages;
140 	if (nr_pages > AIO_RING_PAGES) {
141 		info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
142 		if (!info->ring_pages)
143 			return -ENOMEM;
144 	}
145 
146 	info->mmap_size = nr_pages * PAGE_SIZE;
147 	dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
148 	down_write(&ctx->mm->mmap_sem);
149 	info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
150 				  PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
151 				  0);
152 	if (IS_ERR((void *)info->mmap_base)) {
153 		up_write(&ctx->mm->mmap_sem);
154 		info->mmap_size = 0;
155 		aio_free_ring(ctx);
156 		return -EAGAIN;
157 	}
158 
159 	dprintk("mmap address: 0x%08lx\n", info->mmap_base);
160 	info->nr_pages = get_user_pages(current, ctx->mm,
161 					info->mmap_base, nr_pages,
162 					1, 0, info->ring_pages, NULL);
163 	up_write(&ctx->mm->mmap_sem);
164 
165 	if (unlikely(info->nr_pages != nr_pages)) {
166 		aio_free_ring(ctx);
167 		return -EAGAIN;
168 	}
169 
170 	ctx->user_id = info->mmap_base;
171 
172 	info->nr = nr_events;		/* trusted copy */
173 
174 	ring = kmap_atomic(info->ring_pages[0], KM_USER0);
175 	ring->nr = nr_events;	/* user copy */
176 	ring->id = ctx->user_id;
177 	ring->head = ring->tail = 0;
178 	ring->magic = AIO_RING_MAGIC;
179 	ring->compat_features = AIO_RING_COMPAT_FEATURES;
180 	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
181 	ring->header_length = sizeof(struct aio_ring);
182 	kunmap_atomic(ring, KM_USER0);
183 
184 	return 0;
185 }
186 
187 
188 /* aio_ring_event: returns a pointer to the event at the given index from
189  * kmap_atomic(, km).  Release the pointer with put_aio_ring_event();
190  */
191 #define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
192 #define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
193 #define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
194 
195 #define aio_ring_event(info, nr, km) ({					\
196 	unsigned pos = (nr) + AIO_EVENTS_OFFSET;			\
197 	struct io_event *__event;					\
198 	__event = kmap_atomic(						\
199 			(info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
200 	__event += pos % AIO_EVENTS_PER_PAGE;				\
201 	__event;							\
202 })
203 
204 #define put_aio_ring_event(event, km) do {	\
205 	struct io_event *__event = (event);	\
206 	(void)__event;				\
207 	kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
208 } while(0)
209 
210 static void ctx_rcu_free(struct rcu_head *head)
211 {
212 	struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
213 	unsigned nr_events = ctx->max_reqs;
214 
215 	kmem_cache_free(kioctx_cachep, ctx);
216 
217 	if (nr_events) {
218 		spin_lock(&aio_nr_lock);
219 		BUG_ON(aio_nr - nr_events > aio_nr);
220 		aio_nr -= nr_events;
221 		spin_unlock(&aio_nr_lock);
222 	}
223 }
224 
225 /* __put_ioctx
226  *	Called when the last user of an aio context has gone away,
227  *	and the struct needs to be freed.
228  */
229 static void __put_ioctx(struct kioctx *ctx)
230 {
231 	BUG_ON(ctx->reqs_active);
232 
233 	cancel_delayed_work(&ctx->wq);
234 	cancel_work_sync(&ctx->wq.work);
235 	aio_free_ring(ctx);
236 	mmdrop(ctx->mm);
237 	ctx->mm = NULL;
238 	pr_debug("__put_ioctx: freeing %p\n", ctx);
239 	call_rcu(&ctx->rcu_head, ctx_rcu_free);
240 }
241 
242 #define get_ioctx(kioctx) do {						\
243 	BUG_ON(atomic_read(&(kioctx)->users) <= 0);			\
244 	atomic_inc(&(kioctx)->users);					\
245 } while (0)
246 #define put_ioctx(kioctx) do {						\
247 	BUG_ON(atomic_read(&(kioctx)->users) <= 0);			\
248 	if (unlikely(atomic_dec_and_test(&(kioctx)->users))) 		\
249 		__put_ioctx(kioctx);					\
250 } while (0)
251 
252 /* ioctx_alloc
253  *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
254  */
255 static struct kioctx *ioctx_alloc(unsigned nr_events)
256 {
257 	struct mm_struct *mm;
258 	struct kioctx *ctx;
259 	int did_sync = 0;
260 
261 	/* Prevent overflows */
262 	if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
263 	    (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
264 		pr_debug("ENOMEM: nr_events too high\n");
265 		return ERR_PTR(-EINVAL);
266 	}
267 
268 	if ((unsigned long)nr_events > aio_max_nr)
269 		return ERR_PTR(-EAGAIN);
270 
271 	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
272 	if (!ctx)
273 		return ERR_PTR(-ENOMEM);
274 
275 	ctx->max_reqs = nr_events;
276 	mm = ctx->mm = current->mm;
277 	atomic_inc(&mm->mm_count);
278 
279 	atomic_set(&ctx->users, 1);
280 	spin_lock_init(&ctx->ctx_lock);
281 	spin_lock_init(&ctx->ring_info.ring_lock);
282 	init_waitqueue_head(&ctx->wait);
283 
284 	INIT_LIST_HEAD(&ctx->active_reqs);
285 	INIT_LIST_HEAD(&ctx->run_list);
286 	INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
287 
288 	if (aio_setup_ring(ctx) < 0)
289 		goto out_freectx;
290 
291 	/* limit the number of system wide aios */
292 	do {
293 		spin_lock_bh(&aio_nr_lock);
294 		if (aio_nr + nr_events > aio_max_nr ||
295 		    aio_nr + nr_events < aio_nr)
296 			ctx->max_reqs = 0;
297 		else
298 			aio_nr += ctx->max_reqs;
299 		spin_unlock_bh(&aio_nr_lock);
300 		if (ctx->max_reqs || did_sync)
301 			break;
302 
303 		/* wait for rcu callbacks to have completed before giving up */
304 		synchronize_rcu();
305 		did_sync = 1;
306 		ctx->max_reqs = nr_events;
307 	} while (1);
308 
309 	if (ctx->max_reqs == 0)
310 		goto out_cleanup;
311 
312 	/* now link into global list. */
313 	spin_lock(&mm->ioctx_lock);
314 	hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
315 	spin_unlock(&mm->ioctx_lock);
316 
317 	dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
318 		ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
319 	return ctx;
320 
321 out_cleanup:
322 	__put_ioctx(ctx);
323 	return ERR_PTR(-EAGAIN);
324 
325 out_freectx:
326 	mmdrop(mm);
327 	kmem_cache_free(kioctx_cachep, ctx);
328 	ctx = ERR_PTR(-ENOMEM);
329 
330 	dprintk("aio: error allocating ioctx %p\n", ctx);
331 	return ctx;
332 }
333 
334 /* aio_cancel_all
335  *	Cancels all outstanding aio requests on an aio context.  Used
336  *	when the processes owning a context have all exited to encourage
337  *	the rapid destruction of the kioctx.
338  */
339 static void aio_cancel_all(struct kioctx *ctx)
340 {
341 	int (*cancel)(struct kiocb *, struct io_event *);
342 	struct io_event res;
343 	spin_lock_irq(&ctx->ctx_lock);
344 	ctx->dead = 1;
345 	while (!list_empty(&ctx->active_reqs)) {
346 		struct list_head *pos = ctx->active_reqs.next;
347 		struct kiocb *iocb = list_kiocb(pos);
348 		list_del_init(&iocb->ki_list);
349 		cancel = iocb->ki_cancel;
350 		kiocbSetCancelled(iocb);
351 		if (cancel) {
352 			iocb->ki_users++;
353 			spin_unlock_irq(&ctx->ctx_lock);
354 			cancel(iocb, &res);
355 			spin_lock_irq(&ctx->ctx_lock);
356 		}
357 	}
358 	spin_unlock_irq(&ctx->ctx_lock);
359 }
360 
361 static void wait_for_all_aios(struct kioctx *ctx)
362 {
363 	struct task_struct *tsk = current;
364 	DECLARE_WAITQUEUE(wait, tsk);
365 
366 	spin_lock_irq(&ctx->ctx_lock);
367 	if (!ctx->reqs_active)
368 		goto out;
369 
370 	add_wait_queue(&ctx->wait, &wait);
371 	set_task_state(tsk, TASK_UNINTERRUPTIBLE);
372 	while (ctx->reqs_active) {
373 		spin_unlock_irq(&ctx->ctx_lock);
374 		io_schedule();
375 		set_task_state(tsk, TASK_UNINTERRUPTIBLE);
376 		spin_lock_irq(&ctx->ctx_lock);
377 	}
378 	__set_task_state(tsk, TASK_RUNNING);
379 	remove_wait_queue(&ctx->wait, &wait);
380 
381 out:
382 	spin_unlock_irq(&ctx->ctx_lock);
383 }
384 
385 /* wait_on_sync_kiocb:
386  *	Waits on the given sync kiocb to complete.
387  */
388 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
389 {
390 	while (iocb->ki_users) {
391 		set_current_state(TASK_UNINTERRUPTIBLE);
392 		if (!iocb->ki_users)
393 			break;
394 		io_schedule();
395 	}
396 	__set_current_state(TASK_RUNNING);
397 	return iocb->ki_user_data;
398 }
399 EXPORT_SYMBOL(wait_on_sync_kiocb);
400 
401 /* exit_aio: called when the last user of mm goes away.  At this point,
402  * there is no way for any new requests to be submited or any of the
403  * io_* syscalls to be called on the context.  However, there may be
404  * outstanding requests which hold references to the context; as they
405  * go away, they will call put_ioctx and release any pinned memory
406  * associated with the request (held via struct page * references).
407  */
408 void exit_aio(struct mm_struct *mm)
409 {
410 	struct kioctx *ctx;
411 
412 	while (!hlist_empty(&mm->ioctx_list)) {
413 		ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
414 		hlist_del_rcu(&ctx->list);
415 
416 		aio_cancel_all(ctx);
417 
418 		wait_for_all_aios(ctx);
419 		/*
420 		 * Ensure we don't leave the ctx on the aio_wq
421 		 */
422 		cancel_work_sync(&ctx->wq.work);
423 
424 		if (1 != atomic_read(&ctx->users))
425 			printk(KERN_DEBUG
426 				"exit_aio:ioctx still alive: %d %d %d\n",
427 				atomic_read(&ctx->users), ctx->dead,
428 				ctx->reqs_active);
429 		put_ioctx(ctx);
430 	}
431 }
432 
433 /* aio_get_req
434  *	Allocate a slot for an aio request.  Increments the users count
435  * of the kioctx so that the kioctx stays around until all requests are
436  * complete.  Returns NULL if no requests are free.
437  *
438  * Returns with kiocb->users set to 2.  The io submit code path holds
439  * an extra reference while submitting the i/o.
440  * This prevents races between the aio code path referencing the
441  * req (after submitting it) and aio_complete() freeing the req.
442  */
443 static struct kiocb *__aio_get_req(struct kioctx *ctx)
444 {
445 	struct kiocb *req = NULL;
446 	struct aio_ring *ring;
447 	int okay = 0;
448 
449 	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
450 	if (unlikely(!req))
451 		return NULL;
452 
453 	req->ki_flags = 0;
454 	req->ki_users = 2;
455 	req->ki_key = 0;
456 	req->ki_ctx = ctx;
457 	req->ki_cancel = NULL;
458 	req->ki_retry = NULL;
459 	req->ki_dtor = NULL;
460 	req->private = NULL;
461 	req->ki_iovec = NULL;
462 	INIT_LIST_HEAD(&req->ki_run_list);
463 	req->ki_eventfd = NULL;
464 
465 	/* Check if the completion queue has enough free space to
466 	 * accept an event from this io.
467 	 */
468 	spin_lock_irq(&ctx->ctx_lock);
469 	ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
470 	if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
471 		list_add(&req->ki_list, &ctx->active_reqs);
472 		ctx->reqs_active++;
473 		okay = 1;
474 	}
475 	kunmap_atomic(ring, KM_USER0);
476 	spin_unlock_irq(&ctx->ctx_lock);
477 
478 	if (!okay) {
479 		kmem_cache_free(kiocb_cachep, req);
480 		req = NULL;
481 	}
482 
483 	return req;
484 }
485 
486 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
487 {
488 	struct kiocb *req;
489 	/* Handle a potential starvation case -- should be exceedingly rare as
490 	 * requests will be stuck on fput_head only if the aio_fput_routine is
491 	 * delayed and the requests were the last user of the struct file.
492 	 */
493 	req = __aio_get_req(ctx);
494 	if (unlikely(NULL == req)) {
495 		aio_fput_routine(NULL);
496 		req = __aio_get_req(ctx);
497 	}
498 	return req;
499 }
500 
501 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
502 {
503 	assert_spin_locked(&ctx->ctx_lock);
504 
505 	if (req->ki_eventfd != NULL)
506 		eventfd_ctx_put(req->ki_eventfd);
507 	if (req->ki_dtor)
508 		req->ki_dtor(req);
509 	if (req->ki_iovec != &req->ki_inline_vec)
510 		kfree(req->ki_iovec);
511 	kmem_cache_free(kiocb_cachep, req);
512 	ctx->reqs_active--;
513 
514 	if (unlikely(!ctx->reqs_active && ctx->dead))
515 		wake_up(&ctx->wait);
516 }
517 
518 static void aio_fput_routine(struct work_struct *data)
519 {
520 	spin_lock_irq(&fput_lock);
521 	while (likely(!list_empty(&fput_head))) {
522 		struct kiocb *req = list_kiocb(fput_head.next);
523 		struct kioctx *ctx = req->ki_ctx;
524 
525 		list_del(&req->ki_list);
526 		spin_unlock_irq(&fput_lock);
527 
528 		/* Complete the fput(s) */
529 		if (req->ki_filp != NULL)
530 			fput(req->ki_filp);
531 
532 		/* Link the iocb into the context's free list */
533 		spin_lock_irq(&ctx->ctx_lock);
534 		really_put_req(ctx, req);
535 		spin_unlock_irq(&ctx->ctx_lock);
536 
537 		put_ioctx(ctx);
538 		spin_lock_irq(&fput_lock);
539 	}
540 	spin_unlock_irq(&fput_lock);
541 }
542 
543 /* __aio_put_req
544  *	Returns true if this put was the last user of the request.
545  */
546 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
547 {
548 	dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
549 		req, atomic_long_read(&req->ki_filp->f_count));
550 
551 	assert_spin_locked(&ctx->ctx_lock);
552 
553 	req->ki_users--;
554 	BUG_ON(req->ki_users < 0);
555 	if (likely(req->ki_users))
556 		return 0;
557 	list_del(&req->ki_list);		/* remove from active_reqs */
558 	req->ki_cancel = NULL;
559 	req->ki_retry = NULL;
560 
561 	/*
562 	 * Try to optimize the aio and eventfd file* puts, by avoiding to
563 	 * schedule work in case it is not final fput() time. In normal cases,
564 	 * we would not be holding the last reference to the file*, so
565 	 * this function will be executed w/out any aio kthread wakeup.
566 	 */
567 	if (unlikely(!fput_atomic(req->ki_filp))) {
568 		get_ioctx(ctx);
569 		spin_lock(&fput_lock);
570 		list_add(&req->ki_list, &fput_head);
571 		spin_unlock(&fput_lock);
572 		queue_work(aio_wq, &fput_work);
573 	} else {
574 		req->ki_filp = NULL;
575 		really_put_req(ctx, req);
576 	}
577 	return 1;
578 }
579 
580 /* aio_put_req
581  *	Returns true if this put was the last user of the kiocb,
582  *	false if the request is still in use.
583  */
584 int aio_put_req(struct kiocb *req)
585 {
586 	struct kioctx *ctx = req->ki_ctx;
587 	int ret;
588 	spin_lock_irq(&ctx->ctx_lock);
589 	ret = __aio_put_req(ctx, req);
590 	spin_unlock_irq(&ctx->ctx_lock);
591 	return ret;
592 }
593 EXPORT_SYMBOL(aio_put_req);
594 
595 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
596 {
597 	struct mm_struct *mm = current->mm;
598 	struct kioctx *ctx, *ret = NULL;
599 	struct hlist_node *n;
600 
601 	rcu_read_lock();
602 
603 	hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
604 		if (ctx->user_id == ctx_id && !ctx->dead) {
605 			get_ioctx(ctx);
606 			ret = ctx;
607 			break;
608 		}
609 	}
610 
611 	rcu_read_unlock();
612 	return ret;
613 }
614 
615 /*
616  * Queue up a kiocb to be retried. Assumes that the kiocb
617  * has already been marked as kicked, and places it on
618  * the retry run list for the corresponding ioctx, if it
619  * isn't already queued. Returns 1 if it actually queued
620  * the kiocb (to tell the caller to activate the work
621  * queue to process it), or 0, if it found that it was
622  * already queued.
623  */
624 static inline int __queue_kicked_iocb(struct kiocb *iocb)
625 {
626 	struct kioctx *ctx = iocb->ki_ctx;
627 
628 	assert_spin_locked(&ctx->ctx_lock);
629 
630 	if (list_empty(&iocb->ki_run_list)) {
631 		list_add_tail(&iocb->ki_run_list,
632 			&ctx->run_list);
633 		return 1;
634 	}
635 	return 0;
636 }
637 
638 /* aio_run_iocb
639  *	This is the core aio execution routine. It is
640  *	invoked both for initial i/o submission and
641  *	subsequent retries via the aio_kick_handler.
642  *	Expects to be invoked with iocb->ki_ctx->lock
643  *	already held. The lock is released and reacquired
644  *	as needed during processing.
645  *
646  * Calls the iocb retry method (already setup for the
647  * iocb on initial submission) for operation specific
648  * handling, but takes care of most of common retry
649  * execution details for a given iocb. The retry method
650  * needs to be non-blocking as far as possible, to avoid
651  * holding up other iocbs waiting to be serviced by the
652  * retry kernel thread.
653  *
654  * The trickier parts in this code have to do with
655  * ensuring that only one retry instance is in progress
656  * for a given iocb at any time. Providing that guarantee
657  * simplifies the coding of individual aio operations as
658  * it avoids various potential races.
659  */
660 static ssize_t aio_run_iocb(struct kiocb *iocb)
661 {
662 	struct kioctx	*ctx = iocb->ki_ctx;
663 	ssize_t (*retry)(struct kiocb *);
664 	ssize_t ret;
665 
666 	if (!(retry = iocb->ki_retry)) {
667 		printk("aio_run_iocb: iocb->ki_retry = NULL\n");
668 		return 0;
669 	}
670 
671 	/*
672 	 * We don't want the next retry iteration for this
673 	 * operation to start until this one has returned and
674 	 * updated the iocb state. However, wait_queue functions
675 	 * can trigger a kick_iocb from interrupt context in the
676 	 * meantime, indicating that data is available for the next
677 	 * iteration. We want to remember that and enable the
678 	 * next retry iteration _after_ we are through with
679 	 * this one.
680 	 *
681 	 * So, in order to be able to register a "kick", but
682 	 * prevent it from being queued now, we clear the kick
683 	 * flag, but make the kick code *think* that the iocb is
684 	 * still on the run list until we are actually done.
685 	 * When we are done with this iteration, we check if
686 	 * the iocb was kicked in the meantime and if so, queue
687 	 * it up afresh.
688 	 */
689 
690 	kiocbClearKicked(iocb);
691 
692 	/*
693 	 * This is so that aio_complete knows it doesn't need to
694 	 * pull the iocb off the run list (We can't just call
695 	 * INIT_LIST_HEAD because we don't want a kick_iocb to
696 	 * queue this on the run list yet)
697 	 */
698 	iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
699 	spin_unlock_irq(&ctx->ctx_lock);
700 
701 	/* Quit retrying if the i/o has been cancelled */
702 	if (kiocbIsCancelled(iocb)) {
703 		ret = -EINTR;
704 		aio_complete(iocb, ret, 0);
705 		/* must not access the iocb after this */
706 		goto out;
707 	}
708 
709 	/*
710 	 * Now we are all set to call the retry method in async
711 	 * context.
712 	 */
713 	ret = retry(iocb);
714 
715 	if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
716 		/*
717 		 * There's no easy way to restart the syscall since other AIO's
718 		 * may be already running. Just fail this IO with EINTR.
719 		 */
720 		if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
721 			     ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
722 			ret = -EINTR;
723 		aio_complete(iocb, ret, 0);
724 	}
725 out:
726 	spin_lock_irq(&ctx->ctx_lock);
727 
728 	if (-EIOCBRETRY == ret) {
729 		/*
730 		 * OK, now that we are done with this iteration
731 		 * and know that there is more left to go,
732 		 * this is where we let go so that a subsequent
733 		 * "kick" can start the next iteration
734 		 */
735 
736 		/* will make __queue_kicked_iocb succeed from here on */
737 		INIT_LIST_HEAD(&iocb->ki_run_list);
738 		/* we must queue the next iteration ourselves, if it
739 		 * has already been kicked */
740 		if (kiocbIsKicked(iocb)) {
741 			__queue_kicked_iocb(iocb);
742 
743 			/*
744 			 * __queue_kicked_iocb will always return 1 here, because
745 			 * iocb->ki_run_list is empty at this point so it should
746 			 * be safe to unconditionally queue the context into the
747 			 * work queue.
748 			 */
749 			aio_queue_work(ctx);
750 		}
751 	}
752 	return ret;
753 }
754 
755 /*
756  * __aio_run_iocbs:
757  * 	Process all pending retries queued on the ioctx
758  * 	run list.
759  * Assumes it is operating within the aio issuer's mm
760  * context.
761  */
762 static int __aio_run_iocbs(struct kioctx *ctx)
763 {
764 	struct kiocb *iocb;
765 	struct list_head run_list;
766 
767 	assert_spin_locked(&ctx->ctx_lock);
768 
769 	list_replace_init(&ctx->run_list, &run_list);
770 	while (!list_empty(&run_list)) {
771 		iocb = list_entry(run_list.next, struct kiocb,
772 			ki_run_list);
773 		list_del(&iocb->ki_run_list);
774 		/*
775 		 * Hold an extra reference while retrying i/o.
776 		 */
777 		iocb->ki_users++;       /* grab extra reference */
778 		aio_run_iocb(iocb);
779 		__aio_put_req(ctx, iocb);
780  	}
781 	if (!list_empty(&ctx->run_list))
782 		return 1;
783 	return 0;
784 }
785 
786 static void aio_queue_work(struct kioctx * ctx)
787 {
788 	unsigned long timeout;
789 	/*
790 	 * if someone is waiting, get the work started right
791 	 * away, otherwise, use a longer delay
792 	 */
793 	smp_mb();
794 	if (waitqueue_active(&ctx->wait))
795 		timeout = 1;
796 	else
797 		timeout = HZ/10;
798 	queue_delayed_work(aio_wq, &ctx->wq, timeout);
799 }
800 
801 /*
802  * aio_run_all_iocbs:
803  *	Process all pending retries queued on the ioctx
804  *	run list, and keep running them until the list
805  *	stays empty.
806  * Assumes it is operating within the aio issuer's mm context.
807  */
808 static inline void aio_run_all_iocbs(struct kioctx *ctx)
809 {
810 	spin_lock_irq(&ctx->ctx_lock);
811 	while (__aio_run_iocbs(ctx))
812 		;
813 	spin_unlock_irq(&ctx->ctx_lock);
814 }
815 
816 /*
817  * aio_kick_handler:
818  * 	Work queue handler triggered to process pending
819  * 	retries on an ioctx. Takes on the aio issuer's
820  *	mm context before running the iocbs, so that
821  *	copy_xxx_user operates on the issuer's address
822  *      space.
823  * Run on aiod's context.
824  */
825 static void aio_kick_handler(struct work_struct *work)
826 {
827 	struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
828 	mm_segment_t oldfs = get_fs();
829 	struct mm_struct *mm;
830 	int requeue;
831 
832 	set_fs(USER_DS);
833 	use_mm(ctx->mm);
834 	spin_lock_irq(&ctx->ctx_lock);
835 	requeue =__aio_run_iocbs(ctx);
836 	mm = ctx->mm;
837 	spin_unlock_irq(&ctx->ctx_lock);
838  	unuse_mm(mm);
839 	set_fs(oldfs);
840 	/*
841 	 * we're in a worker thread already, don't use queue_delayed_work,
842 	 */
843 	if (requeue)
844 		queue_delayed_work(aio_wq, &ctx->wq, 0);
845 }
846 
847 
848 /*
849  * Called by kick_iocb to queue the kiocb for retry
850  * and if required activate the aio work queue to process
851  * it
852  */
853 static void try_queue_kicked_iocb(struct kiocb *iocb)
854 {
855  	struct kioctx	*ctx = iocb->ki_ctx;
856 	unsigned long flags;
857 	int run = 0;
858 
859 	spin_lock_irqsave(&ctx->ctx_lock, flags);
860 	/* set this inside the lock so that we can't race with aio_run_iocb()
861 	 * testing it and putting the iocb on the run list under the lock */
862 	if (!kiocbTryKick(iocb))
863 		run = __queue_kicked_iocb(iocb);
864 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
865 	if (run)
866 		aio_queue_work(ctx);
867 }
868 
869 /*
870  * kick_iocb:
871  *      Called typically from a wait queue callback context
872  *      to trigger a retry of the iocb.
873  *      The retry is usually executed by aio workqueue
874  *      threads (See aio_kick_handler).
875  */
876 void kick_iocb(struct kiocb *iocb)
877 {
878 	/* sync iocbs are easy: they can only ever be executing from a
879 	 * single context. */
880 	if (is_sync_kiocb(iocb)) {
881 		kiocbSetKicked(iocb);
882 	        wake_up_process(iocb->ki_obj.tsk);
883 		return;
884 	}
885 
886 	try_queue_kicked_iocb(iocb);
887 }
888 EXPORT_SYMBOL(kick_iocb);
889 
890 /* aio_complete
891  *	Called when the io request on the given iocb is complete.
892  *	Returns true if this is the last user of the request.  The
893  *	only other user of the request can be the cancellation code.
894  */
895 int aio_complete(struct kiocb *iocb, long res, long res2)
896 {
897 	struct kioctx	*ctx = iocb->ki_ctx;
898 	struct aio_ring_info	*info;
899 	struct aio_ring	*ring;
900 	struct io_event	*event;
901 	unsigned long	flags;
902 	unsigned long	tail;
903 	int		ret;
904 
905 	/*
906 	 * Special case handling for sync iocbs:
907 	 *  - events go directly into the iocb for fast handling
908 	 *  - the sync task with the iocb in its stack holds the single iocb
909 	 *    ref, no other paths have a way to get another ref
910 	 *  - the sync task helpfully left a reference to itself in the iocb
911 	 */
912 	if (is_sync_kiocb(iocb)) {
913 		BUG_ON(iocb->ki_users != 1);
914 		iocb->ki_user_data = res;
915 		iocb->ki_users = 0;
916 		wake_up_process(iocb->ki_obj.tsk);
917 		return 1;
918 	}
919 
920 	info = &ctx->ring_info;
921 
922 	/* add a completion event to the ring buffer.
923 	 * must be done holding ctx->ctx_lock to prevent
924 	 * other code from messing with the tail
925 	 * pointer since we might be called from irq
926 	 * context.
927 	 */
928 	spin_lock_irqsave(&ctx->ctx_lock, flags);
929 
930 	if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
931 		list_del_init(&iocb->ki_run_list);
932 
933 	/*
934 	 * cancelled requests don't get events, userland was given one
935 	 * when the event got cancelled.
936 	 */
937 	if (kiocbIsCancelled(iocb))
938 		goto put_rq;
939 
940 	ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
941 
942 	tail = info->tail;
943 	event = aio_ring_event(info, tail, KM_IRQ0);
944 	if (++tail >= info->nr)
945 		tail = 0;
946 
947 	event->obj = (u64)(unsigned long)iocb->ki_obj.user;
948 	event->data = iocb->ki_user_data;
949 	event->res = res;
950 	event->res2 = res2;
951 
952 	dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
953 		ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
954 		res, res2);
955 
956 	/* after flagging the request as done, we
957 	 * must never even look at it again
958 	 */
959 	smp_wmb();	/* make event visible before updating tail */
960 
961 	info->tail = tail;
962 	ring->tail = tail;
963 
964 	put_aio_ring_event(event, KM_IRQ0);
965 	kunmap_atomic(ring, KM_IRQ1);
966 
967 	pr_debug("added to ring %p at [%lu]\n", iocb, tail);
968 
969 	/*
970 	 * Check if the user asked us to deliver the result through an
971 	 * eventfd. The eventfd_signal() function is safe to be called
972 	 * from IRQ context.
973 	 */
974 	if (iocb->ki_eventfd != NULL)
975 		eventfd_signal(iocb->ki_eventfd, 1);
976 
977 put_rq:
978 	/* everything turned out well, dispose of the aiocb. */
979 	ret = __aio_put_req(ctx, iocb);
980 
981 	/*
982 	 * We have to order our ring_info tail store above and test
983 	 * of the wait list below outside the wait lock.  This is
984 	 * like in wake_up_bit() where clearing a bit has to be
985 	 * ordered with the unlocked test.
986 	 */
987 	smp_mb();
988 
989 	if (waitqueue_active(&ctx->wait))
990 		wake_up(&ctx->wait);
991 
992 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
993 	return ret;
994 }
995 EXPORT_SYMBOL(aio_complete);
996 
997 /* aio_read_evt
998  *	Pull an event off of the ioctx's event ring.  Returns the number of
999  *	events fetched (0 or 1 ;-)
1000  *	FIXME: make this use cmpxchg.
1001  *	TODO: make the ringbuffer user mmap()able (requires FIXME).
1002  */
1003 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1004 {
1005 	struct aio_ring_info *info = &ioctx->ring_info;
1006 	struct aio_ring *ring;
1007 	unsigned long head;
1008 	int ret = 0;
1009 
1010 	ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1011 	dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1012 		 (unsigned long)ring->head, (unsigned long)ring->tail,
1013 		 (unsigned long)ring->nr);
1014 
1015 	if (ring->head == ring->tail)
1016 		goto out;
1017 
1018 	spin_lock(&info->ring_lock);
1019 
1020 	head = ring->head % info->nr;
1021 	if (head != ring->tail) {
1022 		struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1023 		*ent = *evp;
1024 		head = (head + 1) % info->nr;
1025 		smp_mb(); /* finish reading the event before updatng the head */
1026 		ring->head = head;
1027 		ret = 1;
1028 		put_aio_ring_event(evp, KM_USER1);
1029 	}
1030 	spin_unlock(&info->ring_lock);
1031 
1032 out:
1033 	kunmap_atomic(ring, KM_USER0);
1034 	dprintk("leaving aio_read_evt: %d  h%lu t%lu\n", ret,
1035 		 (unsigned long)ring->head, (unsigned long)ring->tail);
1036 	return ret;
1037 }
1038 
1039 struct aio_timeout {
1040 	struct timer_list	timer;
1041 	int			timed_out;
1042 	struct task_struct	*p;
1043 };
1044 
1045 static void timeout_func(unsigned long data)
1046 {
1047 	struct aio_timeout *to = (struct aio_timeout *)data;
1048 
1049 	to->timed_out = 1;
1050 	wake_up_process(to->p);
1051 }
1052 
1053 static inline void init_timeout(struct aio_timeout *to)
1054 {
1055 	setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1056 	to->timed_out = 0;
1057 	to->p = current;
1058 }
1059 
1060 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1061 			       const struct timespec *ts)
1062 {
1063 	to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1064 	if (time_after(to->timer.expires, jiffies))
1065 		add_timer(&to->timer);
1066 	else
1067 		to->timed_out = 1;
1068 }
1069 
1070 static inline void clear_timeout(struct aio_timeout *to)
1071 {
1072 	del_singleshot_timer_sync(&to->timer);
1073 }
1074 
1075 static int read_events(struct kioctx *ctx,
1076 			long min_nr, long nr,
1077 			struct io_event __user *event,
1078 			struct timespec __user *timeout)
1079 {
1080 	long			start_jiffies = jiffies;
1081 	struct task_struct	*tsk = current;
1082 	DECLARE_WAITQUEUE(wait, tsk);
1083 	int			ret;
1084 	int			i = 0;
1085 	struct io_event		ent;
1086 	struct aio_timeout	to;
1087 	int			retry = 0;
1088 
1089 	/* needed to zero any padding within an entry (there shouldn't be
1090 	 * any, but C is fun!
1091 	 */
1092 	memset(&ent, 0, sizeof(ent));
1093 retry:
1094 	ret = 0;
1095 	while (likely(i < nr)) {
1096 		ret = aio_read_evt(ctx, &ent);
1097 		if (unlikely(ret <= 0))
1098 			break;
1099 
1100 		dprintk("read event: %Lx %Lx %Lx %Lx\n",
1101 			ent.data, ent.obj, ent.res, ent.res2);
1102 
1103 		/* Could we split the check in two? */
1104 		ret = -EFAULT;
1105 		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1106 			dprintk("aio: lost an event due to EFAULT.\n");
1107 			break;
1108 		}
1109 		ret = 0;
1110 
1111 		/* Good, event copied to userland, update counts. */
1112 		event ++;
1113 		i ++;
1114 	}
1115 
1116 	if (min_nr <= i)
1117 		return i;
1118 	if (ret)
1119 		return ret;
1120 
1121 	/* End fast path */
1122 
1123 	/* racey check, but it gets redone */
1124 	if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1125 		retry = 1;
1126 		aio_run_all_iocbs(ctx);
1127 		goto retry;
1128 	}
1129 
1130 	init_timeout(&to);
1131 	if (timeout) {
1132 		struct timespec	ts;
1133 		ret = -EFAULT;
1134 		if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1135 			goto out;
1136 
1137 		set_timeout(start_jiffies, &to, &ts);
1138 	}
1139 
1140 	while (likely(i < nr)) {
1141 		add_wait_queue_exclusive(&ctx->wait, &wait);
1142 		do {
1143 			set_task_state(tsk, TASK_INTERRUPTIBLE);
1144 			ret = aio_read_evt(ctx, &ent);
1145 			if (ret)
1146 				break;
1147 			if (min_nr <= i)
1148 				break;
1149 			if (unlikely(ctx->dead)) {
1150 				ret = -EINVAL;
1151 				break;
1152 			}
1153 			if (to.timed_out)	/* Only check after read evt */
1154 				break;
1155 			/* Try to only show up in io wait if there are ops
1156 			 *  in flight */
1157 			if (ctx->reqs_active)
1158 				io_schedule();
1159 			else
1160 				schedule();
1161 			if (signal_pending(tsk)) {
1162 				ret = -EINTR;
1163 				break;
1164 			}
1165 			/*ret = aio_read_evt(ctx, &ent);*/
1166 		} while (1) ;
1167 
1168 		set_task_state(tsk, TASK_RUNNING);
1169 		remove_wait_queue(&ctx->wait, &wait);
1170 
1171 		if (unlikely(ret <= 0))
1172 			break;
1173 
1174 		ret = -EFAULT;
1175 		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1176 			dprintk("aio: lost an event due to EFAULT.\n");
1177 			break;
1178 		}
1179 
1180 		/* Good, event copied to userland, update counts. */
1181 		event ++;
1182 		i ++;
1183 	}
1184 
1185 	if (timeout)
1186 		clear_timeout(&to);
1187 out:
1188 	destroy_timer_on_stack(&to.timer);
1189 	return i ? i : ret;
1190 }
1191 
1192 /* Take an ioctx and remove it from the list of ioctx's.  Protects
1193  * against races with itself via ->dead.
1194  */
1195 static void io_destroy(struct kioctx *ioctx)
1196 {
1197 	struct mm_struct *mm = current->mm;
1198 	int was_dead;
1199 
1200 	/* delete the entry from the list is someone else hasn't already */
1201 	spin_lock(&mm->ioctx_lock);
1202 	was_dead = ioctx->dead;
1203 	ioctx->dead = 1;
1204 	hlist_del_rcu(&ioctx->list);
1205 	spin_unlock(&mm->ioctx_lock);
1206 
1207 	dprintk("aio_release(%p)\n", ioctx);
1208 	if (likely(!was_dead))
1209 		put_ioctx(ioctx);	/* twice for the list */
1210 
1211 	aio_cancel_all(ioctx);
1212 	wait_for_all_aios(ioctx);
1213 
1214 	/*
1215 	 * Wake up any waiters.  The setting of ctx->dead must be seen
1216 	 * by other CPUs at this point.  Right now, we rely on the
1217 	 * locking done by the above calls to ensure this consistency.
1218 	 */
1219 	wake_up(&ioctx->wait);
1220 	put_ioctx(ioctx);	/* once for the lookup */
1221 }
1222 
1223 /* sys_io_setup:
1224  *	Create an aio_context capable of receiving at least nr_events.
1225  *	ctxp must not point to an aio_context that already exists, and
1226  *	must be initialized to 0 prior to the call.  On successful
1227  *	creation of the aio_context, *ctxp is filled in with the resulting
1228  *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1229  *	if the specified nr_events exceeds internal limits.  May fail
1230  *	with -EAGAIN if the specified nr_events exceeds the user's limit
1231  *	of available events.  May fail with -ENOMEM if insufficient kernel
1232  *	resources are available.  May fail with -EFAULT if an invalid
1233  *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1234  *	implemented.
1235  */
1236 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1237 {
1238 	struct kioctx *ioctx = NULL;
1239 	unsigned long ctx;
1240 	long ret;
1241 
1242 	ret = get_user(ctx, ctxp);
1243 	if (unlikely(ret))
1244 		goto out;
1245 
1246 	ret = -EINVAL;
1247 	if (unlikely(ctx || nr_events == 0)) {
1248 		pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1249 		         ctx, nr_events);
1250 		goto out;
1251 	}
1252 
1253 	ioctx = ioctx_alloc(nr_events);
1254 	ret = PTR_ERR(ioctx);
1255 	if (!IS_ERR(ioctx)) {
1256 		ret = put_user(ioctx->user_id, ctxp);
1257 		if (!ret)
1258 			return 0;
1259 
1260 		get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1261 		io_destroy(ioctx);
1262 	}
1263 
1264 out:
1265 	return ret;
1266 }
1267 
1268 /* sys_io_destroy:
1269  *	Destroy the aio_context specified.  May cancel any outstanding
1270  *	AIOs and block on completion.  Will fail with -ENOSYS if not
1271  *	implemented.  May fail with -EINVAL if the context pointed to
1272  *	is invalid.
1273  */
1274 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1275 {
1276 	struct kioctx *ioctx = lookup_ioctx(ctx);
1277 	if (likely(NULL != ioctx)) {
1278 		io_destroy(ioctx);
1279 		return 0;
1280 	}
1281 	pr_debug("EINVAL: io_destroy: invalid context id\n");
1282 	return -EINVAL;
1283 }
1284 
1285 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1286 {
1287 	struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1288 
1289 	BUG_ON(ret <= 0);
1290 
1291 	while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1292 		ssize_t this = min((ssize_t)iov->iov_len, ret);
1293 		iov->iov_base += this;
1294 		iov->iov_len -= this;
1295 		iocb->ki_left -= this;
1296 		ret -= this;
1297 		if (iov->iov_len == 0) {
1298 			iocb->ki_cur_seg++;
1299 			iov++;
1300 		}
1301 	}
1302 
1303 	/* the caller should not have done more io than what fit in
1304 	 * the remaining iovecs */
1305 	BUG_ON(ret > 0 && iocb->ki_left == 0);
1306 }
1307 
1308 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1309 {
1310 	struct file *file = iocb->ki_filp;
1311 	struct address_space *mapping = file->f_mapping;
1312 	struct inode *inode = mapping->host;
1313 	ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1314 			 unsigned long, loff_t);
1315 	ssize_t ret = 0;
1316 	unsigned short opcode;
1317 
1318 	if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1319 		(iocb->ki_opcode == IOCB_CMD_PREAD)) {
1320 		rw_op = file->f_op->aio_read;
1321 		opcode = IOCB_CMD_PREADV;
1322 	} else {
1323 		rw_op = file->f_op->aio_write;
1324 		opcode = IOCB_CMD_PWRITEV;
1325 	}
1326 
1327 	/* This matches the pread()/pwrite() logic */
1328 	if (iocb->ki_pos < 0)
1329 		return -EINVAL;
1330 
1331 	do {
1332 		ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1333 			    iocb->ki_nr_segs - iocb->ki_cur_seg,
1334 			    iocb->ki_pos);
1335 		if (ret > 0)
1336 			aio_advance_iovec(iocb, ret);
1337 
1338 	/* retry all partial writes.  retry partial reads as long as its a
1339 	 * regular file. */
1340 	} while (ret > 0 && iocb->ki_left > 0 &&
1341 		 (opcode == IOCB_CMD_PWRITEV ||
1342 		  (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1343 
1344 	/* This means we must have transferred all that we could */
1345 	/* No need to retry anymore */
1346 	if ((ret == 0) || (iocb->ki_left == 0))
1347 		ret = iocb->ki_nbytes - iocb->ki_left;
1348 
1349 	/* If we managed to write some out we return that, rather than
1350 	 * the eventual error. */
1351 	if (opcode == IOCB_CMD_PWRITEV
1352 	    && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1353 	    && iocb->ki_nbytes - iocb->ki_left)
1354 		ret = iocb->ki_nbytes - iocb->ki_left;
1355 
1356 	return ret;
1357 }
1358 
1359 static ssize_t aio_fdsync(struct kiocb *iocb)
1360 {
1361 	struct file *file = iocb->ki_filp;
1362 	ssize_t ret = -EINVAL;
1363 
1364 	if (file->f_op->aio_fsync)
1365 		ret = file->f_op->aio_fsync(iocb, 1);
1366 	return ret;
1367 }
1368 
1369 static ssize_t aio_fsync(struct kiocb *iocb)
1370 {
1371 	struct file *file = iocb->ki_filp;
1372 	ssize_t ret = -EINVAL;
1373 
1374 	if (file->f_op->aio_fsync)
1375 		ret = file->f_op->aio_fsync(iocb, 0);
1376 	return ret;
1377 }
1378 
1379 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1380 {
1381 	ssize_t ret;
1382 
1383 #ifdef CONFIG_COMPAT
1384 	if (compat)
1385 		ret = compat_rw_copy_check_uvector(type,
1386 				(struct compat_iovec __user *)kiocb->ki_buf,
1387 				kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1388 				&kiocb->ki_iovec);
1389 	else
1390 #endif
1391 		ret = rw_copy_check_uvector(type,
1392 				(struct iovec __user *)kiocb->ki_buf,
1393 				kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1394 				&kiocb->ki_iovec);
1395 	if (ret < 0)
1396 		goto out;
1397 
1398 	kiocb->ki_nr_segs = kiocb->ki_nbytes;
1399 	kiocb->ki_cur_seg = 0;
1400 	/* ki_nbytes/left now reflect bytes instead of segs */
1401 	kiocb->ki_nbytes = ret;
1402 	kiocb->ki_left = ret;
1403 
1404 	ret = 0;
1405 out:
1406 	return ret;
1407 }
1408 
1409 static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1410 {
1411 	kiocb->ki_iovec = &kiocb->ki_inline_vec;
1412 	kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1413 	kiocb->ki_iovec->iov_len = kiocb->ki_left;
1414 	kiocb->ki_nr_segs = 1;
1415 	kiocb->ki_cur_seg = 0;
1416 	return 0;
1417 }
1418 
1419 /*
1420  * aio_setup_iocb:
1421  *	Performs the initial checks and aio retry method
1422  *	setup for the kiocb at the time of io submission.
1423  */
1424 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1425 {
1426 	struct file *file = kiocb->ki_filp;
1427 	ssize_t ret = 0;
1428 
1429 	switch (kiocb->ki_opcode) {
1430 	case IOCB_CMD_PREAD:
1431 		ret = -EBADF;
1432 		if (unlikely(!(file->f_mode & FMODE_READ)))
1433 			break;
1434 		ret = -EFAULT;
1435 		if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1436 			kiocb->ki_left)))
1437 			break;
1438 		ret = security_file_permission(file, MAY_READ);
1439 		if (unlikely(ret))
1440 			break;
1441 		ret = aio_setup_single_vector(kiocb);
1442 		if (ret)
1443 			break;
1444 		ret = -EINVAL;
1445 		if (file->f_op->aio_read)
1446 			kiocb->ki_retry = aio_rw_vect_retry;
1447 		break;
1448 	case IOCB_CMD_PWRITE:
1449 		ret = -EBADF;
1450 		if (unlikely(!(file->f_mode & FMODE_WRITE)))
1451 			break;
1452 		ret = -EFAULT;
1453 		if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1454 			kiocb->ki_left)))
1455 			break;
1456 		ret = security_file_permission(file, MAY_WRITE);
1457 		if (unlikely(ret))
1458 			break;
1459 		ret = aio_setup_single_vector(kiocb);
1460 		if (ret)
1461 			break;
1462 		ret = -EINVAL;
1463 		if (file->f_op->aio_write)
1464 			kiocb->ki_retry = aio_rw_vect_retry;
1465 		break;
1466 	case IOCB_CMD_PREADV:
1467 		ret = -EBADF;
1468 		if (unlikely(!(file->f_mode & FMODE_READ)))
1469 			break;
1470 		ret = security_file_permission(file, MAY_READ);
1471 		if (unlikely(ret))
1472 			break;
1473 		ret = aio_setup_vectored_rw(READ, kiocb, compat);
1474 		if (ret)
1475 			break;
1476 		ret = -EINVAL;
1477 		if (file->f_op->aio_read)
1478 			kiocb->ki_retry = aio_rw_vect_retry;
1479 		break;
1480 	case IOCB_CMD_PWRITEV:
1481 		ret = -EBADF;
1482 		if (unlikely(!(file->f_mode & FMODE_WRITE)))
1483 			break;
1484 		ret = security_file_permission(file, MAY_WRITE);
1485 		if (unlikely(ret))
1486 			break;
1487 		ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1488 		if (ret)
1489 			break;
1490 		ret = -EINVAL;
1491 		if (file->f_op->aio_write)
1492 			kiocb->ki_retry = aio_rw_vect_retry;
1493 		break;
1494 	case IOCB_CMD_FDSYNC:
1495 		ret = -EINVAL;
1496 		if (file->f_op->aio_fsync)
1497 			kiocb->ki_retry = aio_fdsync;
1498 		break;
1499 	case IOCB_CMD_FSYNC:
1500 		ret = -EINVAL;
1501 		if (file->f_op->aio_fsync)
1502 			kiocb->ki_retry = aio_fsync;
1503 		break;
1504 	default:
1505 		dprintk("EINVAL: io_submit: no operation provided\n");
1506 		ret = -EINVAL;
1507 	}
1508 
1509 	if (!kiocb->ki_retry)
1510 		return ret;
1511 
1512 	return 0;
1513 }
1514 
1515 static void aio_batch_add(struct address_space *mapping,
1516 			  struct hlist_head *batch_hash)
1517 {
1518 	struct aio_batch_entry *abe;
1519 	struct hlist_node *pos;
1520 	unsigned bucket;
1521 
1522 	bucket = hash_ptr(mapping, AIO_BATCH_HASH_BITS);
1523 	hlist_for_each_entry(abe, pos, &batch_hash[bucket], list) {
1524 		if (abe->mapping == mapping)
1525 			return;
1526 	}
1527 
1528 	abe = mempool_alloc(abe_pool, GFP_KERNEL);
1529 
1530 	/*
1531 	 * we should be using igrab here, but
1532 	 * we don't want to hammer on the global
1533 	 * inode spinlock just to take an extra
1534 	 * reference on a file that we must already
1535 	 * have a reference to.
1536 	 *
1537 	 * When we're called, we always have a reference
1538 	 * on the file, so we must always have a reference
1539 	 * on the inode, so ihold() is safe here.
1540 	 */
1541 	ihold(mapping->host);
1542 	abe->mapping = mapping;
1543 	hlist_add_head(&abe->list, &batch_hash[bucket]);
1544 	return;
1545 }
1546 
1547 static void aio_batch_free(struct hlist_head *batch_hash)
1548 {
1549 	struct aio_batch_entry *abe;
1550 	struct hlist_node *pos, *n;
1551 	int i;
1552 
1553 	for (i = 0; i < AIO_BATCH_HASH_SIZE; i++) {
1554 		hlist_for_each_entry_safe(abe, pos, n, &batch_hash[i], list) {
1555 			blk_run_address_space(abe->mapping);
1556 			iput(abe->mapping->host);
1557 			hlist_del(&abe->list);
1558 			mempool_free(abe, abe_pool);
1559 		}
1560 	}
1561 }
1562 
1563 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1564 			 struct iocb *iocb, struct hlist_head *batch_hash,
1565 			 bool compat)
1566 {
1567 	struct kiocb *req;
1568 	struct file *file;
1569 	ssize_t ret;
1570 
1571 	/* enforce forwards compatibility on users */
1572 	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1573 		pr_debug("EINVAL: io_submit: reserve field set\n");
1574 		return -EINVAL;
1575 	}
1576 
1577 	/* prevent overflows */
1578 	if (unlikely(
1579 	    (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1580 	    (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1581 	    ((ssize_t)iocb->aio_nbytes < 0)
1582 	   )) {
1583 		pr_debug("EINVAL: io_submit: overflow check\n");
1584 		return -EINVAL;
1585 	}
1586 
1587 	file = fget(iocb->aio_fildes);
1588 	if (unlikely(!file))
1589 		return -EBADF;
1590 
1591 	req = aio_get_req(ctx);		/* returns with 2 references to req */
1592 	if (unlikely(!req)) {
1593 		fput(file);
1594 		return -EAGAIN;
1595 	}
1596 	req->ki_filp = file;
1597 	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1598 		/*
1599 		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1600 		 * instance of the file* now. The file descriptor must be
1601 		 * an eventfd() fd, and will be signaled for each completed
1602 		 * event using the eventfd_signal() function.
1603 		 */
1604 		req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1605 		if (IS_ERR(req->ki_eventfd)) {
1606 			ret = PTR_ERR(req->ki_eventfd);
1607 			req->ki_eventfd = NULL;
1608 			goto out_put_req;
1609 		}
1610 	}
1611 
1612 	ret = put_user(req->ki_key, &user_iocb->aio_key);
1613 	if (unlikely(ret)) {
1614 		dprintk("EFAULT: aio_key\n");
1615 		goto out_put_req;
1616 	}
1617 
1618 	req->ki_obj.user = user_iocb;
1619 	req->ki_user_data = iocb->aio_data;
1620 	req->ki_pos = iocb->aio_offset;
1621 
1622 	req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1623 	req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1624 	req->ki_opcode = iocb->aio_lio_opcode;
1625 
1626 	ret = aio_setup_iocb(req, compat);
1627 
1628 	if (ret)
1629 		goto out_put_req;
1630 
1631 	spin_lock_irq(&ctx->ctx_lock);
1632 	aio_run_iocb(req);
1633 	if (!list_empty(&ctx->run_list)) {
1634 		/* drain the run list */
1635 		while (__aio_run_iocbs(ctx))
1636 			;
1637 	}
1638 	spin_unlock_irq(&ctx->ctx_lock);
1639 	if (req->ki_opcode == IOCB_CMD_PREAD ||
1640 	    req->ki_opcode == IOCB_CMD_PREADV ||
1641 	    req->ki_opcode == IOCB_CMD_PWRITE ||
1642 	    req->ki_opcode == IOCB_CMD_PWRITEV)
1643 		aio_batch_add(file->f_mapping, batch_hash);
1644 
1645 	aio_put_req(req);	/* drop extra ref to req */
1646 	return 0;
1647 
1648 out_put_req:
1649 	aio_put_req(req);	/* drop extra ref to req */
1650 	aio_put_req(req);	/* drop i/o ref to req */
1651 	return ret;
1652 }
1653 
1654 long do_io_submit(aio_context_t ctx_id, long nr,
1655 		  struct iocb __user *__user *iocbpp, bool compat)
1656 {
1657 	struct kioctx *ctx;
1658 	long ret = 0;
1659 	int i;
1660 	struct hlist_head batch_hash[AIO_BATCH_HASH_SIZE] = { { 0, }, };
1661 
1662 	if (unlikely(nr < 0))
1663 		return -EINVAL;
1664 
1665 	if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1666 		nr = LONG_MAX/sizeof(*iocbpp);
1667 
1668 	if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1669 		return -EFAULT;
1670 
1671 	ctx = lookup_ioctx(ctx_id);
1672 	if (unlikely(!ctx)) {
1673 		pr_debug("EINVAL: io_submit: invalid context id\n");
1674 		return -EINVAL;
1675 	}
1676 
1677 	/*
1678 	 * AKPM: should this return a partial result if some of the IOs were
1679 	 * successfully submitted?
1680 	 */
1681 	for (i=0; i<nr; i++) {
1682 		struct iocb __user *user_iocb;
1683 		struct iocb tmp;
1684 
1685 		if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1686 			ret = -EFAULT;
1687 			break;
1688 		}
1689 
1690 		if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1691 			ret = -EFAULT;
1692 			break;
1693 		}
1694 
1695 		ret = io_submit_one(ctx, user_iocb, &tmp, batch_hash, compat);
1696 		if (ret)
1697 			break;
1698 	}
1699 	aio_batch_free(batch_hash);
1700 
1701 	put_ioctx(ctx);
1702 	return i ? i : ret;
1703 }
1704 
1705 /* sys_io_submit:
1706  *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1707  *	the number of iocbs queued.  May return -EINVAL if the aio_context
1708  *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
1709  *	*iocbpp[0] is not properly initialized, if the operation specified
1710  *	is invalid for the file descriptor in the iocb.  May fail with
1711  *	-EFAULT if any of the data structures point to invalid data.  May
1712  *	fail with -EBADF if the file descriptor specified in the first
1713  *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
1714  *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1715  *	fail with -ENOSYS if not implemented.
1716  */
1717 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1718 		struct iocb __user * __user *, iocbpp)
1719 {
1720 	return do_io_submit(ctx_id, nr, iocbpp, 0);
1721 }
1722 
1723 /* lookup_kiocb
1724  *	Finds a given iocb for cancellation.
1725  */
1726 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1727 				  u32 key)
1728 {
1729 	struct list_head *pos;
1730 
1731 	assert_spin_locked(&ctx->ctx_lock);
1732 
1733 	/* TODO: use a hash or array, this sucks. */
1734 	list_for_each(pos, &ctx->active_reqs) {
1735 		struct kiocb *kiocb = list_kiocb(pos);
1736 		if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1737 			return kiocb;
1738 	}
1739 	return NULL;
1740 }
1741 
1742 /* sys_io_cancel:
1743  *	Attempts to cancel an iocb previously passed to io_submit.  If
1744  *	the operation is successfully cancelled, the resulting event is
1745  *	copied into the memory pointed to by result without being placed
1746  *	into the completion queue and 0 is returned.  May fail with
1747  *	-EFAULT if any of the data structures pointed to are invalid.
1748  *	May fail with -EINVAL if aio_context specified by ctx_id is
1749  *	invalid.  May fail with -EAGAIN if the iocb specified was not
1750  *	cancelled.  Will fail with -ENOSYS if not implemented.
1751  */
1752 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1753 		struct io_event __user *, result)
1754 {
1755 	int (*cancel)(struct kiocb *iocb, struct io_event *res);
1756 	struct kioctx *ctx;
1757 	struct kiocb *kiocb;
1758 	u32 key;
1759 	int ret;
1760 
1761 	ret = get_user(key, &iocb->aio_key);
1762 	if (unlikely(ret))
1763 		return -EFAULT;
1764 
1765 	ctx = lookup_ioctx(ctx_id);
1766 	if (unlikely(!ctx))
1767 		return -EINVAL;
1768 
1769 	spin_lock_irq(&ctx->ctx_lock);
1770 	ret = -EAGAIN;
1771 	kiocb = lookup_kiocb(ctx, iocb, key);
1772 	if (kiocb && kiocb->ki_cancel) {
1773 		cancel = kiocb->ki_cancel;
1774 		kiocb->ki_users ++;
1775 		kiocbSetCancelled(kiocb);
1776 	} else
1777 		cancel = NULL;
1778 	spin_unlock_irq(&ctx->ctx_lock);
1779 
1780 	if (NULL != cancel) {
1781 		struct io_event tmp;
1782 		pr_debug("calling cancel\n");
1783 		memset(&tmp, 0, sizeof(tmp));
1784 		tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1785 		tmp.data = kiocb->ki_user_data;
1786 		ret = cancel(kiocb, &tmp);
1787 		if (!ret) {
1788 			/* Cancellation succeeded -- copy the result
1789 			 * into the user's buffer.
1790 			 */
1791 			if (copy_to_user(result, &tmp, sizeof(tmp)))
1792 				ret = -EFAULT;
1793 		}
1794 	} else
1795 		ret = -EINVAL;
1796 
1797 	put_ioctx(ctx);
1798 
1799 	return ret;
1800 }
1801 
1802 /* io_getevents:
1803  *	Attempts to read at least min_nr events and up to nr events from
1804  *	the completion queue for the aio_context specified by ctx_id. If
1805  *	it succeeds, the number of read events is returned. May fail with
1806  *	-EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1807  *	out of range, if timeout is out of range.  May fail with -EFAULT
1808  *	if any of the memory specified is invalid.  May return 0 or
1809  *	< min_nr if the timeout specified by timeout has elapsed
1810  *	before sufficient events are available, where timeout == NULL
1811  *	specifies an infinite timeout. Note that the timeout pointed to by
1812  *	timeout is relative and will be updated if not NULL and the
1813  *	operation blocks. Will fail with -ENOSYS if not implemented.
1814  */
1815 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1816 		long, min_nr,
1817 		long, nr,
1818 		struct io_event __user *, events,
1819 		struct timespec __user *, timeout)
1820 {
1821 	struct kioctx *ioctx = lookup_ioctx(ctx_id);
1822 	long ret = -EINVAL;
1823 
1824 	if (likely(ioctx)) {
1825 		if (likely(min_nr <= nr && min_nr >= 0))
1826 			ret = read_events(ioctx, min_nr, nr, events, timeout);
1827 		put_ioctx(ioctx);
1828 	}
1829 
1830 	asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1831 	return ret;
1832 }
1833