xref: /openbmc/linux/fs/aio.c (revision 732a675a)
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/uio.h>
19 
20 #define DEBUG 0
21 
22 #include <linux/sched.h>
23 #include <linux/fs.h>
24 #include <linux/file.h>
25 #include <linux/mm.h>
26 #include <linux/mman.h>
27 #include <linux/slab.h>
28 #include <linux/timer.h>
29 #include <linux/aio.h>
30 #include <linux/highmem.h>
31 #include <linux/workqueue.h>
32 #include <linux/security.h>
33 #include <linux/eventfd.h>
34 
35 #include <asm/kmap_types.h>
36 #include <asm/uaccess.h>
37 #include <asm/mmu_context.h>
38 
39 #if DEBUG > 1
40 #define dprintk		printk
41 #else
42 #define dprintk(x...)	do { ; } while (0)
43 #endif
44 
45 /*------ sysctl variables----*/
46 static DEFINE_SPINLOCK(aio_nr_lock);
47 unsigned long aio_nr;		/* current system wide number of aio requests */
48 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
49 /*----end sysctl variables---*/
50 
51 static struct kmem_cache	*kiocb_cachep;
52 static struct kmem_cache	*kioctx_cachep;
53 
54 static struct workqueue_struct *aio_wq;
55 
56 /* Used for rare fput completion. */
57 static void aio_fput_routine(struct work_struct *);
58 static DECLARE_WORK(fput_work, aio_fput_routine);
59 
60 static DEFINE_SPINLOCK(fput_lock);
61 static LIST_HEAD(fput_head);
62 
63 static void aio_kick_handler(struct work_struct *);
64 static void aio_queue_work(struct kioctx *);
65 
66 /* aio_setup
67  *	Creates the slab caches used by the aio routines, panic on
68  *	failure as this is done early during the boot sequence.
69  */
70 static int __init aio_setup(void)
71 {
72 	kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
73 	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
74 
75 	aio_wq = create_workqueue("aio");
76 
77 	pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
78 
79 	return 0;
80 }
81 
82 static void aio_free_ring(struct kioctx *ctx)
83 {
84 	struct aio_ring_info *info = &ctx->ring_info;
85 	long i;
86 
87 	for (i=0; i<info->nr_pages; i++)
88 		put_page(info->ring_pages[i]);
89 
90 	if (info->mmap_size) {
91 		down_write(&ctx->mm->mmap_sem);
92 		do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
93 		up_write(&ctx->mm->mmap_sem);
94 	}
95 
96 	if (info->ring_pages && info->ring_pages != info->internal_pages)
97 		kfree(info->ring_pages);
98 	info->ring_pages = NULL;
99 	info->nr = 0;
100 }
101 
102 static int aio_setup_ring(struct kioctx *ctx)
103 {
104 	struct aio_ring *ring;
105 	struct aio_ring_info *info = &ctx->ring_info;
106 	unsigned nr_events = ctx->max_reqs;
107 	unsigned long size;
108 	int nr_pages;
109 
110 	/* Compensate for the ring buffer's head/tail overlap entry */
111 	nr_events += 2;	/* 1 is required, 2 for good luck */
112 
113 	size = sizeof(struct aio_ring);
114 	size += sizeof(struct io_event) * nr_events;
115 	nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
116 
117 	if (nr_pages < 0)
118 		return -EINVAL;
119 
120 	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
121 
122 	info->nr = 0;
123 	info->ring_pages = info->internal_pages;
124 	if (nr_pages > AIO_RING_PAGES) {
125 		info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
126 		if (!info->ring_pages)
127 			return -ENOMEM;
128 	}
129 
130 	info->mmap_size = nr_pages * PAGE_SIZE;
131 	dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
132 	down_write(&ctx->mm->mmap_sem);
133 	info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
134 				  PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
135 				  0);
136 	if (IS_ERR((void *)info->mmap_base)) {
137 		up_write(&ctx->mm->mmap_sem);
138 		info->mmap_size = 0;
139 		aio_free_ring(ctx);
140 		return -EAGAIN;
141 	}
142 
143 	dprintk("mmap address: 0x%08lx\n", info->mmap_base);
144 	info->nr_pages = get_user_pages(current, ctx->mm,
145 					info->mmap_base, nr_pages,
146 					1, 0, info->ring_pages, NULL);
147 	up_write(&ctx->mm->mmap_sem);
148 
149 	if (unlikely(info->nr_pages != nr_pages)) {
150 		aio_free_ring(ctx);
151 		return -EAGAIN;
152 	}
153 
154 	ctx->user_id = info->mmap_base;
155 
156 	info->nr = nr_events;		/* trusted copy */
157 
158 	ring = kmap_atomic(info->ring_pages[0], KM_USER0);
159 	ring->nr = nr_events;	/* user copy */
160 	ring->id = ctx->user_id;
161 	ring->head = ring->tail = 0;
162 	ring->magic = AIO_RING_MAGIC;
163 	ring->compat_features = AIO_RING_COMPAT_FEATURES;
164 	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
165 	ring->header_length = sizeof(struct aio_ring);
166 	kunmap_atomic(ring, KM_USER0);
167 
168 	return 0;
169 }
170 
171 
172 /* aio_ring_event: returns a pointer to the event at the given index from
173  * kmap_atomic(, km).  Release the pointer with put_aio_ring_event();
174  */
175 #define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
176 #define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
177 #define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
178 
179 #define aio_ring_event(info, nr, km) ({					\
180 	unsigned pos = (nr) + AIO_EVENTS_OFFSET;			\
181 	struct io_event *__event;					\
182 	__event = kmap_atomic(						\
183 			(info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
184 	__event += pos % AIO_EVENTS_PER_PAGE;				\
185 	__event;							\
186 })
187 
188 #define put_aio_ring_event(event, km) do {	\
189 	struct io_event *__event = (event);	\
190 	(void)__event;				\
191 	kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
192 } while(0)
193 
194 
195 /* __put_ioctx
196  *	Called when the last user of an aio context has gone away,
197  *	and the struct needs to be freed.
198  */
199 static void __put_ioctx(struct kioctx *ctx)
200 {
201 	unsigned nr_events = ctx->max_reqs;
202 
203 	BUG_ON(ctx->reqs_active);
204 
205 	cancel_delayed_work(&ctx->wq);
206 	cancel_work_sync(&ctx->wq.work);
207 	aio_free_ring(ctx);
208 	mmdrop(ctx->mm);
209 	ctx->mm = NULL;
210 	pr_debug("__put_ioctx: freeing %p\n", ctx);
211 	kmem_cache_free(kioctx_cachep, ctx);
212 
213 	if (nr_events) {
214 		spin_lock(&aio_nr_lock);
215 		BUG_ON(aio_nr - nr_events > aio_nr);
216 		aio_nr -= nr_events;
217 		spin_unlock(&aio_nr_lock);
218 	}
219 }
220 
221 #define get_ioctx(kioctx) do {						\
222 	BUG_ON(atomic_read(&(kioctx)->users) <= 0);			\
223 	atomic_inc(&(kioctx)->users);					\
224 } while (0)
225 #define put_ioctx(kioctx) do {						\
226 	BUG_ON(atomic_read(&(kioctx)->users) <= 0);			\
227 	if (unlikely(atomic_dec_and_test(&(kioctx)->users))) 		\
228 		__put_ioctx(kioctx);					\
229 } while (0)
230 
231 /* ioctx_alloc
232  *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
233  */
234 static struct kioctx *ioctx_alloc(unsigned nr_events)
235 {
236 	struct mm_struct *mm;
237 	struct kioctx *ctx;
238 
239 	/* Prevent overflows */
240 	if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
241 	    (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
242 		pr_debug("ENOMEM: nr_events too high\n");
243 		return ERR_PTR(-EINVAL);
244 	}
245 
246 	if ((unsigned long)nr_events > aio_max_nr)
247 		return ERR_PTR(-EAGAIN);
248 
249 	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
250 	if (!ctx)
251 		return ERR_PTR(-ENOMEM);
252 
253 	ctx->max_reqs = nr_events;
254 	mm = ctx->mm = current->mm;
255 	atomic_inc(&mm->mm_count);
256 
257 	atomic_set(&ctx->users, 1);
258 	spin_lock_init(&ctx->ctx_lock);
259 	spin_lock_init(&ctx->ring_info.ring_lock);
260 	init_waitqueue_head(&ctx->wait);
261 
262 	INIT_LIST_HEAD(&ctx->active_reqs);
263 	INIT_LIST_HEAD(&ctx->run_list);
264 	INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
265 
266 	if (aio_setup_ring(ctx) < 0)
267 		goto out_freectx;
268 
269 	/* limit the number of system wide aios */
270 	spin_lock(&aio_nr_lock);
271 	if (aio_nr + ctx->max_reqs > aio_max_nr ||
272 	    aio_nr + ctx->max_reqs < aio_nr)
273 		ctx->max_reqs = 0;
274 	else
275 		aio_nr += ctx->max_reqs;
276 	spin_unlock(&aio_nr_lock);
277 	if (ctx->max_reqs == 0)
278 		goto out_cleanup;
279 
280 	/* now link into global list. */
281 	write_lock(&mm->ioctx_list_lock);
282 	ctx->next = mm->ioctx_list;
283 	mm->ioctx_list = ctx;
284 	write_unlock(&mm->ioctx_list_lock);
285 
286 	dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
287 		ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
288 	return ctx;
289 
290 out_cleanup:
291 	__put_ioctx(ctx);
292 	return ERR_PTR(-EAGAIN);
293 
294 out_freectx:
295 	mmdrop(mm);
296 	kmem_cache_free(kioctx_cachep, ctx);
297 	ctx = ERR_PTR(-ENOMEM);
298 
299 	dprintk("aio: error allocating ioctx %p\n", ctx);
300 	return ctx;
301 }
302 
303 /* aio_cancel_all
304  *	Cancels all outstanding aio requests on an aio context.  Used
305  *	when the processes owning a context have all exited to encourage
306  *	the rapid destruction of the kioctx.
307  */
308 static void aio_cancel_all(struct kioctx *ctx)
309 {
310 	int (*cancel)(struct kiocb *, struct io_event *);
311 	struct io_event res;
312 	spin_lock_irq(&ctx->ctx_lock);
313 	ctx->dead = 1;
314 	while (!list_empty(&ctx->active_reqs)) {
315 		struct list_head *pos = ctx->active_reqs.next;
316 		struct kiocb *iocb = list_kiocb(pos);
317 		list_del_init(&iocb->ki_list);
318 		cancel = iocb->ki_cancel;
319 		kiocbSetCancelled(iocb);
320 		if (cancel) {
321 			iocb->ki_users++;
322 			spin_unlock_irq(&ctx->ctx_lock);
323 			cancel(iocb, &res);
324 			spin_lock_irq(&ctx->ctx_lock);
325 		}
326 	}
327 	spin_unlock_irq(&ctx->ctx_lock);
328 }
329 
330 static void wait_for_all_aios(struct kioctx *ctx)
331 {
332 	struct task_struct *tsk = current;
333 	DECLARE_WAITQUEUE(wait, tsk);
334 
335 	spin_lock_irq(&ctx->ctx_lock);
336 	if (!ctx->reqs_active)
337 		goto out;
338 
339 	add_wait_queue(&ctx->wait, &wait);
340 	set_task_state(tsk, TASK_UNINTERRUPTIBLE);
341 	while (ctx->reqs_active) {
342 		spin_unlock_irq(&ctx->ctx_lock);
343 		io_schedule();
344 		set_task_state(tsk, TASK_UNINTERRUPTIBLE);
345 		spin_lock_irq(&ctx->ctx_lock);
346 	}
347 	__set_task_state(tsk, TASK_RUNNING);
348 	remove_wait_queue(&ctx->wait, &wait);
349 
350 out:
351 	spin_unlock_irq(&ctx->ctx_lock);
352 }
353 
354 /* wait_on_sync_kiocb:
355  *	Waits on the given sync kiocb to complete.
356  */
357 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
358 {
359 	while (iocb->ki_users) {
360 		set_current_state(TASK_UNINTERRUPTIBLE);
361 		if (!iocb->ki_users)
362 			break;
363 		io_schedule();
364 	}
365 	__set_current_state(TASK_RUNNING);
366 	return iocb->ki_user_data;
367 }
368 
369 /* exit_aio: called when the last user of mm goes away.  At this point,
370  * there is no way for any new requests to be submited or any of the
371  * io_* syscalls to be called on the context.  However, there may be
372  * outstanding requests which hold references to the context; as they
373  * go away, they will call put_ioctx and release any pinned memory
374  * associated with the request (held via struct page * references).
375  */
376 void exit_aio(struct mm_struct *mm)
377 {
378 	struct kioctx *ctx = mm->ioctx_list;
379 	mm->ioctx_list = NULL;
380 	while (ctx) {
381 		struct kioctx *next = ctx->next;
382 		ctx->next = NULL;
383 		aio_cancel_all(ctx);
384 
385 		wait_for_all_aios(ctx);
386 		/*
387 		 * Ensure we don't leave the ctx on the aio_wq
388 		 */
389 		cancel_work_sync(&ctx->wq.work);
390 
391 		if (1 != atomic_read(&ctx->users))
392 			printk(KERN_DEBUG
393 				"exit_aio:ioctx still alive: %d %d %d\n",
394 				atomic_read(&ctx->users), ctx->dead,
395 				ctx->reqs_active);
396 		put_ioctx(ctx);
397 		ctx = next;
398 	}
399 }
400 
401 /* aio_get_req
402  *	Allocate a slot for an aio request.  Increments the users count
403  * of the kioctx so that the kioctx stays around until all requests are
404  * complete.  Returns NULL if no requests are free.
405  *
406  * Returns with kiocb->users set to 2.  The io submit code path holds
407  * an extra reference while submitting the i/o.
408  * This prevents races between the aio code path referencing the
409  * req (after submitting it) and aio_complete() freeing the req.
410  */
411 static struct kiocb *__aio_get_req(struct kioctx *ctx)
412 {
413 	struct kiocb *req = NULL;
414 	struct aio_ring *ring;
415 	int okay = 0;
416 
417 	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
418 	if (unlikely(!req))
419 		return NULL;
420 
421 	req->ki_flags = 0;
422 	req->ki_users = 2;
423 	req->ki_key = 0;
424 	req->ki_ctx = ctx;
425 	req->ki_cancel = NULL;
426 	req->ki_retry = NULL;
427 	req->ki_dtor = NULL;
428 	req->private = NULL;
429 	req->ki_iovec = NULL;
430 	INIT_LIST_HEAD(&req->ki_run_list);
431 	req->ki_eventfd = ERR_PTR(-EINVAL);
432 
433 	/* Check if the completion queue has enough free space to
434 	 * accept an event from this io.
435 	 */
436 	spin_lock_irq(&ctx->ctx_lock);
437 	ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
438 	if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
439 		list_add(&req->ki_list, &ctx->active_reqs);
440 		ctx->reqs_active++;
441 		okay = 1;
442 	}
443 	kunmap_atomic(ring, KM_USER0);
444 	spin_unlock_irq(&ctx->ctx_lock);
445 
446 	if (!okay) {
447 		kmem_cache_free(kiocb_cachep, req);
448 		req = NULL;
449 	}
450 
451 	return req;
452 }
453 
454 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
455 {
456 	struct kiocb *req;
457 	/* Handle a potential starvation case -- should be exceedingly rare as
458 	 * requests will be stuck on fput_head only if the aio_fput_routine is
459 	 * delayed and the requests were the last user of the struct file.
460 	 */
461 	req = __aio_get_req(ctx);
462 	if (unlikely(NULL == req)) {
463 		aio_fput_routine(NULL);
464 		req = __aio_get_req(ctx);
465 	}
466 	return req;
467 }
468 
469 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
470 {
471 	assert_spin_locked(&ctx->ctx_lock);
472 
473 	if (!IS_ERR(req->ki_eventfd))
474 		fput(req->ki_eventfd);
475 	if (req->ki_dtor)
476 		req->ki_dtor(req);
477 	if (req->ki_iovec != &req->ki_inline_vec)
478 		kfree(req->ki_iovec);
479 	kmem_cache_free(kiocb_cachep, req);
480 	ctx->reqs_active--;
481 
482 	if (unlikely(!ctx->reqs_active && ctx->dead))
483 		wake_up(&ctx->wait);
484 }
485 
486 static void aio_fput_routine(struct work_struct *data)
487 {
488 	spin_lock_irq(&fput_lock);
489 	while (likely(!list_empty(&fput_head))) {
490 		struct kiocb *req = list_kiocb(fput_head.next);
491 		struct kioctx *ctx = req->ki_ctx;
492 
493 		list_del(&req->ki_list);
494 		spin_unlock_irq(&fput_lock);
495 
496 		/* Complete the fput */
497 		__fput(req->ki_filp);
498 
499 		/* Link the iocb into the context's free list */
500 		spin_lock_irq(&ctx->ctx_lock);
501 		really_put_req(ctx, req);
502 		spin_unlock_irq(&ctx->ctx_lock);
503 
504 		put_ioctx(ctx);
505 		spin_lock_irq(&fput_lock);
506 	}
507 	spin_unlock_irq(&fput_lock);
508 }
509 
510 /* __aio_put_req
511  *	Returns true if this put was the last user of the request.
512  */
513 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
514 {
515 	dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
516 		req, atomic_read(&req->ki_filp->f_count));
517 
518 	assert_spin_locked(&ctx->ctx_lock);
519 
520 	req->ki_users --;
521 	BUG_ON(req->ki_users < 0);
522 	if (likely(req->ki_users))
523 		return 0;
524 	list_del(&req->ki_list);		/* remove from active_reqs */
525 	req->ki_cancel = NULL;
526 	req->ki_retry = NULL;
527 
528 	/* Must be done under the lock to serialise against cancellation.
529 	 * Call this aio_fput as it duplicates fput via the fput_work.
530 	 */
531 	if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) {
532 		get_ioctx(ctx);
533 		spin_lock(&fput_lock);
534 		list_add(&req->ki_list, &fput_head);
535 		spin_unlock(&fput_lock);
536 		queue_work(aio_wq, &fput_work);
537 	} else
538 		really_put_req(ctx, req);
539 	return 1;
540 }
541 
542 /* aio_put_req
543  *	Returns true if this put was the last user of the kiocb,
544  *	false if the request is still in use.
545  */
546 int aio_put_req(struct kiocb *req)
547 {
548 	struct kioctx *ctx = req->ki_ctx;
549 	int ret;
550 	spin_lock_irq(&ctx->ctx_lock);
551 	ret = __aio_put_req(ctx, req);
552 	spin_unlock_irq(&ctx->ctx_lock);
553 	return ret;
554 }
555 
556 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
557 {
558 	struct kioctx *ioctx;
559 	struct mm_struct *mm;
560 
561 	mm = current->mm;
562 	read_lock(&mm->ioctx_list_lock);
563 	for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
564 		if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
565 			get_ioctx(ioctx);
566 			break;
567 		}
568 	read_unlock(&mm->ioctx_list_lock);
569 
570 	return ioctx;
571 }
572 
573 /*
574  * use_mm
575  *	Makes the calling kernel thread take on the specified
576  *	mm context.
577  *	Called by the retry thread execute retries within the
578  *	iocb issuer's mm context, so that copy_from/to_user
579  *	operations work seamlessly for aio.
580  *	(Note: this routine is intended to be called only
581  *	from a kernel thread context)
582  */
583 static void use_mm(struct mm_struct *mm)
584 {
585 	struct mm_struct *active_mm;
586 	struct task_struct *tsk = current;
587 
588 	task_lock(tsk);
589 	tsk->flags |= PF_BORROWED_MM;
590 	active_mm = tsk->active_mm;
591 	atomic_inc(&mm->mm_count);
592 	tsk->mm = mm;
593 	tsk->active_mm = mm;
594 	switch_mm(active_mm, mm, tsk);
595 	task_unlock(tsk);
596 
597 	mmdrop(active_mm);
598 }
599 
600 /*
601  * unuse_mm
602  *	Reverses the effect of use_mm, i.e. releases the
603  *	specified mm context which was earlier taken on
604  *	by the calling kernel thread
605  *	(Note: this routine is intended to be called only
606  *	from a kernel thread context)
607  */
608 static void unuse_mm(struct mm_struct *mm)
609 {
610 	struct task_struct *tsk = current;
611 
612 	task_lock(tsk);
613 	tsk->flags &= ~PF_BORROWED_MM;
614 	tsk->mm = NULL;
615 	/* active_mm is still 'mm' */
616 	enter_lazy_tlb(mm, tsk);
617 	task_unlock(tsk);
618 }
619 
620 /*
621  * Queue up a kiocb to be retried. Assumes that the kiocb
622  * has already been marked as kicked, and places it on
623  * the retry run list for the corresponding ioctx, if it
624  * isn't already queued. Returns 1 if it actually queued
625  * the kiocb (to tell the caller to activate the work
626  * queue to process it), or 0, if it found that it was
627  * already queued.
628  */
629 static inline int __queue_kicked_iocb(struct kiocb *iocb)
630 {
631 	struct kioctx *ctx = iocb->ki_ctx;
632 
633 	assert_spin_locked(&ctx->ctx_lock);
634 
635 	if (list_empty(&iocb->ki_run_list)) {
636 		list_add_tail(&iocb->ki_run_list,
637 			&ctx->run_list);
638 		return 1;
639 	}
640 	return 0;
641 }
642 
643 /* aio_run_iocb
644  *	This is the core aio execution routine. It is
645  *	invoked both for initial i/o submission and
646  *	subsequent retries via the aio_kick_handler.
647  *	Expects to be invoked with iocb->ki_ctx->lock
648  *	already held. The lock is released and reacquired
649  *	as needed during processing.
650  *
651  * Calls the iocb retry method (already setup for the
652  * iocb on initial submission) for operation specific
653  * handling, but takes care of most of common retry
654  * execution details for a given iocb. The retry method
655  * needs to be non-blocking as far as possible, to avoid
656  * holding up other iocbs waiting to be serviced by the
657  * retry kernel thread.
658  *
659  * The trickier parts in this code have to do with
660  * ensuring that only one retry instance is in progress
661  * for a given iocb at any time. Providing that guarantee
662  * simplifies the coding of individual aio operations as
663  * it avoids various potential races.
664  */
665 static ssize_t aio_run_iocb(struct kiocb *iocb)
666 {
667 	struct kioctx	*ctx = iocb->ki_ctx;
668 	ssize_t (*retry)(struct kiocb *);
669 	ssize_t ret;
670 
671 	if (!(retry = iocb->ki_retry)) {
672 		printk("aio_run_iocb: iocb->ki_retry = NULL\n");
673 		return 0;
674 	}
675 
676 	/*
677 	 * We don't want the next retry iteration for this
678 	 * operation to start until this one has returned and
679 	 * updated the iocb state. However, wait_queue functions
680 	 * can trigger a kick_iocb from interrupt context in the
681 	 * meantime, indicating that data is available for the next
682 	 * iteration. We want to remember that and enable the
683 	 * next retry iteration _after_ we are through with
684 	 * this one.
685 	 *
686 	 * So, in order to be able to register a "kick", but
687 	 * prevent it from being queued now, we clear the kick
688 	 * flag, but make the kick code *think* that the iocb is
689 	 * still on the run list until we are actually done.
690 	 * When we are done with this iteration, we check if
691 	 * the iocb was kicked in the meantime and if so, queue
692 	 * it up afresh.
693 	 */
694 
695 	kiocbClearKicked(iocb);
696 
697 	/*
698 	 * This is so that aio_complete knows it doesn't need to
699 	 * pull the iocb off the run list (We can't just call
700 	 * INIT_LIST_HEAD because we don't want a kick_iocb to
701 	 * queue this on the run list yet)
702 	 */
703 	iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
704 	spin_unlock_irq(&ctx->ctx_lock);
705 
706 	/* Quit retrying if the i/o has been cancelled */
707 	if (kiocbIsCancelled(iocb)) {
708 		ret = -EINTR;
709 		aio_complete(iocb, ret, 0);
710 		/* must not access the iocb after this */
711 		goto out;
712 	}
713 
714 	/*
715 	 * Now we are all set to call the retry method in async
716 	 * context.
717 	 */
718 	ret = retry(iocb);
719 
720 	if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
721 		BUG_ON(!list_empty(&iocb->ki_wait.task_list));
722 		aio_complete(iocb, ret, 0);
723 	}
724 out:
725 	spin_lock_irq(&ctx->ctx_lock);
726 
727 	if (-EIOCBRETRY == ret) {
728 		/*
729 		 * OK, now that we are done with this iteration
730 		 * and know that there is more left to go,
731 		 * this is where we let go so that a subsequent
732 		 * "kick" can start the next iteration
733 		 */
734 
735 		/* will make __queue_kicked_iocb succeed from here on */
736 		INIT_LIST_HEAD(&iocb->ki_run_list);
737 		/* we must queue the next iteration ourselves, if it
738 		 * has already been kicked */
739 		if (kiocbIsKicked(iocb)) {
740 			__queue_kicked_iocb(iocb);
741 
742 			/*
743 			 * __queue_kicked_iocb will always return 1 here, because
744 			 * iocb->ki_run_list is empty at this point so it should
745 			 * be safe to unconditionally queue the context into the
746 			 * work queue.
747 			 */
748 			aio_queue_work(ctx);
749 		}
750 	}
751 	return ret;
752 }
753 
754 /*
755  * __aio_run_iocbs:
756  * 	Process all pending retries queued on the ioctx
757  * 	run list.
758  * Assumes it is operating within the aio issuer's mm
759  * context.
760  */
761 static int __aio_run_iocbs(struct kioctx *ctx)
762 {
763 	struct kiocb *iocb;
764 	struct list_head run_list;
765 
766 	assert_spin_locked(&ctx->ctx_lock);
767 
768 	list_replace_init(&ctx->run_list, &run_list);
769 	while (!list_empty(&run_list)) {
770 		iocb = list_entry(run_list.next, struct kiocb,
771 			ki_run_list);
772 		list_del(&iocb->ki_run_list);
773 		/*
774 		 * Hold an extra reference while retrying i/o.
775 		 */
776 		iocb->ki_users++;       /* grab extra reference */
777 		aio_run_iocb(iocb);
778 		__aio_put_req(ctx, iocb);
779  	}
780 	if (!list_empty(&ctx->run_list))
781 		return 1;
782 	return 0;
783 }
784 
785 static void aio_queue_work(struct kioctx * ctx)
786 {
787 	unsigned long timeout;
788 	/*
789 	 * if someone is waiting, get the work started right
790 	 * away, otherwise, use a longer delay
791 	 */
792 	smp_mb();
793 	if (waitqueue_active(&ctx->wait))
794 		timeout = 1;
795 	else
796 		timeout = HZ/10;
797 	queue_delayed_work(aio_wq, &ctx->wq, timeout);
798 }
799 
800 
801 /*
802  * aio_run_iocbs:
803  * 	Process all pending retries queued on the ioctx
804  * 	run list.
805  * Assumes it is operating within the aio issuer's mm
806  * context.
807  */
808 static inline void aio_run_iocbs(struct kioctx *ctx)
809 {
810 	int requeue;
811 
812 	spin_lock_irq(&ctx->ctx_lock);
813 
814 	requeue = __aio_run_iocbs(ctx);
815 	spin_unlock_irq(&ctx->ctx_lock);
816 	if (requeue)
817 		aio_queue_work(ctx);
818 }
819 
820 /*
821  * just like aio_run_iocbs, but keeps running them until
822  * the list stays empty
823  */
824 static inline void aio_run_all_iocbs(struct kioctx *ctx)
825 {
826 	spin_lock_irq(&ctx->ctx_lock);
827 	while (__aio_run_iocbs(ctx))
828 		;
829 	spin_unlock_irq(&ctx->ctx_lock);
830 }
831 
832 /*
833  * aio_kick_handler:
834  * 	Work queue handler triggered to process pending
835  * 	retries on an ioctx. Takes on the aio issuer's
836  *	mm context before running the iocbs, so that
837  *	copy_xxx_user operates on the issuer's address
838  *      space.
839  * Run on aiod's context.
840  */
841 static void aio_kick_handler(struct work_struct *work)
842 {
843 	struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
844 	mm_segment_t oldfs = get_fs();
845 	struct mm_struct *mm;
846 	int requeue;
847 
848 	set_fs(USER_DS);
849 	use_mm(ctx->mm);
850 	spin_lock_irq(&ctx->ctx_lock);
851 	requeue =__aio_run_iocbs(ctx);
852 	mm = ctx->mm;
853 	spin_unlock_irq(&ctx->ctx_lock);
854  	unuse_mm(mm);
855 	set_fs(oldfs);
856 	/*
857 	 * we're in a worker thread already, don't use queue_delayed_work,
858 	 */
859 	if (requeue)
860 		queue_delayed_work(aio_wq, &ctx->wq, 0);
861 }
862 
863 
864 /*
865  * Called by kick_iocb to queue the kiocb for retry
866  * and if required activate the aio work queue to process
867  * it
868  */
869 static void try_queue_kicked_iocb(struct kiocb *iocb)
870 {
871  	struct kioctx	*ctx = iocb->ki_ctx;
872 	unsigned long flags;
873 	int run = 0;
874 
875 	/* We're supposed to be the only path putting the iocb back on the run
876 	 * list.  If we find that the iocb is *back* on a wait queue already
877 	 * than retry has happened before we could queue the iocb.  This also
878 	 * means that the retry could have completed and freed our iocb, no
879 	 * good. */
880 	BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
881 
882 	spin_lock_irqsave(&ctx->ctx_lock, flags);
883 	/* set this inside the lock so that we can't race with aio_run_iocb()
884 	 * testing it and putting the iocb on the run list under the lock */
885 	if (!kiocbTryKick(iocb))
886 		run = __queue_kicked_iocb(iocb);
887 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
888 	if (run)
889 		aio_queue_work(ctx);
890 }
891 
892 /*
893  * kick_iocb:
894  *      Called typically from a wait queue callback context
895  *      (aio_wake_function) to trigger a retry of the iocb.
896  *      The retry is usually executed by aio workqueue
897  *      threads (See aio_kick_handler).
898  */
899 void kick_iocb(struct kiocb *iocb)
900 {
901 	/* sync iocbs are easy: they can only ever be executing from a
902 	 * single context. */
903 	if (is_sync_kiocb(iocb)) {
904 		kiocbSetKicked(iocb);
905 	        wake_up_process(iocb->ki_obj.tsk);
906 		return;
907 	}
908 
909 	try_queue_kicked_iocb(iocb);
910 }
911 EXPORT_SYMBOL(kick_iocb);
912 
913 /* aio_complete
914  *	Called when the io request on the given iocb is complete.
915  *	Returns true if this is the last user of the request.  The
916  *	only other user of the request can be the cancellation code.
917  */
918 int aio_complete(struct kiocb *iocb, long res, long res2)
919 {
920 	struct kioctx	*ctx = iocb->ki_ctx;
921 	struct aio_ring_info	*info;
922 	struct aio_ring	*ring;
923 	struct io_event	*event;
924 	unsigned long	flags;
925 	unsigned long	tail;
926 	int		ret;
927 
928 	/*
929 	 * Special case handling for sync iocbs:
930 	 *  - events go directly into the iocb for fast handling
931 	 *  - the sync task with the iocb in its stack holds the single iocb
932 	 *    ref, no other paths have a way to get another ref
933 	 *  - the sync task helpfully left a reference to itself in the iocb
934 	 */
935 	if (is_sync_kiocb(iocb)) {
936 		BUG_ON(iocb->ki_users != 1);
937 		iocb->ki_user_data = res;
938 		iocb->ki_users = 0;
939 		wake_up_process(iocb->ki_obj.tsk);
940 		return 1;
941 	}
942 
943 	info = &ctx->ring_info;
944 
945 	/* add a completion event to the ring buffer.
946 	 * must be done holding ctx->ctx_lock to prevent
947 	 * other code from messing with the tail
948 	 * pointer since we might be called from irq
949 	 * context.
950 	 */
951 	spin_lock_irqsave(&ctx->ctx_lock, flags);
952 
953 	if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
954 		list_del_init(&iocb->ki_run_list);
955 
956 	/*
957 	 * cancelled requests don't get events, userland was given one
958 	 * when the event got cancelled.
959 	 */
960 	if (kiocbIsCancelled(iocb))
961 		goto put_rq;
962 
963 	ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
964 
965 	tail = info->tail;
966 	event = aio_ring_event(info, tail, KM_IRQ0);
967 	if (++tail >= info->nr)
968 		tail = 0;
969 
970 	event->obj = (u64)(unsigned long)iocb->ki_obj.user;
971 	event->data = iocb->ki_user_data;
972 	event->res = res;
973 	event->res2 = res2;
974 
975 	dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
976 		ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
977 		res, res2);
978 
979 	/* after flagging the request as done, we
980 	 * must never even look at it again
981 	 */
982 	smp_wmb();	/* make event visible before updating tail */
983 
984 	info->tail = tail;
985 	ring->tail = tail;
986 
987 	put_aio_ring_event(event, KM_IRQ0);
988 	kunmap_atomic(ring, KM_IRQ1);
989 
990 	pr_debug("added to ring %p at [%lu]\n", iocb, tail);
991 
992 	/*
993 	 * Check if the user asked us to deliver the result through an
994 	 * eventfd. The eventfd_signal() function is safe to be called
995 	 * from IRQ context.
996 	 */
997 	if (!IS_ERR(iocb->ki_eventfd))
998 		eventfd_signal(iocb->ki_eventfd, 1);
999 
1000 put_rq:
1001 	/* everything turned out well, dispose of the aiocb. */
1002 	ret = __aio_put_req(ctx, iocb);
1003 
1004 	/*
1005 	 * We have to order our ring_info tail store above and test
1006 	 * of the wait list below outside the wait lock.  This is
1007 	 * like in wake_up_bit() where clearing a bit has to be
1008 	 * ordered with the unlocked test.
1009 	 */
1010 	smp_mb();
1011 
1012 	if (waitqueue_active(&ctx->wait))
1013 		wake_up(&ctx->wait);
1014 
1015 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1016 	return ret;
1017 }
1018 
1019 /* aio_read_evt
1020  *	Pull an event off of the ioctx's event ring.  Returns the number of
1021  *	events fetched (0 or 1 ;-)
1022  *	FIXME: make this use cmpxchg.
1023  *	TODO: make the ringbuffer user mmap()able (requires FIXME).
1024  */
1025 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1026 {
1027 	struct aio_ring_info *info = &ioctx->ring_info;
1028 	struct aio_ring *ring;
1029 	unsigned long head;
1030 	int ret = 0;
1031 
1032 	ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1033 	dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1034 		 (unsigned long)ring->head, (unsigned long)ring->tail,
1035 		 (unsigned long)ring->nr);
1036 
1037 	if (ring->head == ring->tail)
1038 		goto out;
1039 
1040 	spin_lock(&info->ring_lock);
1041 
1042 	head = ring->head % info->nr;
1043 	if (head != ring->tail) {
1044 		struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1045 		*ent = *evp;
1046 		head = (head + 1) % info->nr;
1047 		smp_mb(); /* finish reading the event before updatng the head */
1048 		ring->head = head;
1049 		ret = 1;
1050 		put_aio_ring_event(evp, KM_USER1);
1051 	}
1052 	spin_unlock(&info->ring_lock);
1053 
1054 out:
1055 	kunmap_atomic(ring, KM_USER0);
1056 	dprintk("leaving aio_read_evt: %d  h%lu t%lu\n", ret,
1057 		 (unsigned long)ring->head, (unsigned long)ring->tail);
1058 	return ret;
1059 }
1060 
1061 struct aio_timeout {
1062 	struct timer_list	timer;
1063 	int			timed_out;
1064 	struct task_struct	*p;
1065 };
1066 
1067 static void timeout_func(unsigned long data)
1068 {
1069 	struct aio_timeout *to = (struct aio_timeout *)data;
1070 
1071 	to->timed_out = 1;
1072 	wake_up_process(to->p);
1073 }
1074 
1075 static inline void init_timeout(struct aio_timeout *to)
1076 {
1077 	setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1078 	to->timed_out = 0;
1079 	to->p = current;
1080 }
1081 
1082 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1083 			       const struct timespec *ts)
1084 {
1085 	to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1086 	if (time_after(to->timer.expires, jiffies))
1087 		add_timer(&to->timer);
1088 	else
1089 		to->timed_out = 1;
1090 }
1091 
1092 static inline void clear_timeout(struct aio_timeout *to)
1093 {
1094 	del_singleshot_timer_sync(&to->timer);
1095 }
1096 
1097 static int read_events(struct kioctx *ctx,
1098 			long min_nr, long nr,
1099 			struct io_event __user *event,
1100 			struct timespec __user *timeout)
1101 {
1102 	long			start_jiffies = jiffies;
1103 	struct task_struct	*tsk = current;
1104 	DECLARE_WAITQUEUE(wait, tsk);
1105 	int			ret;
1106 	int			i = 0;
1107 	struct io_event		ent;
1108 	struct aio_timeout	to;
1109 	int			retry = 0;
1110 
1111 	/* needed to zero any padding within an entry (there shouldn't be
1112 	 * any, but C is fun!
1113 	 */
1114 	memset(&ent, 0, sizeof(ent));
1115 retry:
1116 	ret = 0;
1117 	while (likely(i < nr)) {
1118 		ret = aio_read_evt(ctx, &ent);
1119 		if (unlikely(ret <= 0))
1120 			break;
1121 
1122 		dprintk("read event: %Lx %Lx %Lx %Lx\n",
1123 			ent.data, ent.obj, ent.res, ent.res2);
1124 
1125 		/* Could we split the check in two? */
1126 		ret = -EFAULT;
1127 		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1128 			dprintk("aio: lost an event due to EFAULT.\n");
1129 			break;
1130 		}
1131 		ret = 0;
1132 
1133 		/* Good, event copied to userland, update counts. */
1134 		event ++;
1135 		i ++;
1136 	}
1137 
1138 	if (min_nr <= i)
1139 		return i;
1140 	if (ret)
1141 		return ret;
1142 
1143 	/* End fast path */
1144 
1145 	/* racey check, but it gets redone */
1146 	if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1147 		retry = 1;
1148 		aio_run_all_iocbs(ctx);
1149 		goto retry;
1150 	}
1151 
1152 	init_timeout(&to);
1153 	if (timeout) {
1154 		struct timespec	ts;
1155 		ret = -EFAULT;
1156 		if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1157 			goto out;
1158 
1159 		set_timeout(start_jiffies, &to, &ts);
1160 	}
1161 
1162 	while (likely(i < nr)) {
1163 		add_wait_queue_exclusive(&ctx->wait, &wait);
1164 		do {
1165 			set_task_state(tsk, TASK_INTERRUPTIBLE);
1166 			ret = aio_read_evt(ctx, &ent);
1167 			if (ret)
1168 				break;
1169 			if (min_nr <= i)
1170 				break;
1171 			if (unlikely(ctx->dead)) {
1172 				ret = -EINVAL;
1173 				break;
1174 			}
1175 			if (to.timed_out)	/* Only check after read evt */
1176 				break;
1177 			/* Try to only show up in io wait if there are ops
1178 			 *  in flight */
1179 			if (ctx->reqs_active)
1180 				io_schedule();
1181 			else
1182 				schedule();
1183 			if (signal_pending(tsk)) {
1184 				ret = -EINTR;
1185 				break;
1186 			}
1187 			/*ret = aio_read_evt(ctx, &ent);*/
1188 		} while (1) ;
1189 
1190 		set_task_state(tsk, TASK_RUNNING);
1191 		remove_wait_queue(&ctx->wait, &wait);
1192 
1193 		if (unlikely(ret <= 0))
1194 			break;
1195 
1196 		ret = -EFAULT;
1197 		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1198 			dprintk("aio: lost an event due to EFAULT.\n");
1199 			break;
1200 		}
1201 
1202 		/* Good, event copied to userland, update counts. */
1203 		event ++;
1204 		i ++;
1205 	}
1206 
1207 	if (timeout)
1208 		clear_timeout(&to);
1209 out:
1210 	destroy_timer_on_stack(&to.timer);
1211 	return i ? i : ret;
1212 }
1213 
1214 /* Take an ioctx and remove it from the list of ioctx's.  Protects
1215  * against races with itself via ->dead.
1216  */
1217 static void io_destroy(struct kioctx *ioctx)
1218 {
1219 	struct mm_struct *mm = current->mm;
1220 	struct kioctx **tmp;
1221 	int was_dead;
1222 
1223 	/* delete the entry from the list is someone else hasn't already */
1224 	write_lock(&mm->ioctx_list_lock);
1225 	was_dead = ioctx->dead;
1226 	ioctx->dead = 1;
1227 	for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1228 	     tmp = &(*tmp)->next)
1229 		;
1230 	if (*tmp)
1231 		*tmp = ioctx->next;
1232 	write_unlock(&mm->ioctx_list_lock);
1233 
1234 	dprintk("aio_release(%p)\n", ioctx);
1235 	if (likely(!was_dead))
1236 		put_ioctx(ioctx);	/* twice for the list */
1237 
1238 	aio_cancel_all(ioctx);
1239 	wait_for_all_aios(ioctx);
1240 
1241 	/*
1242 	 * Wake up any waiters.  The setting of ctx->dead must be seen
1243 	 * by other CPUs at this point.  Right now, we rely on the
1244 	 * locking done by the above calls to ensure this consistency.
1245 	 */
1246 	wake_up(&ioctx->wait);
1247 	put_ioctx(ioctx);	/* once for the lookup */
1248 }
1249 
1250 /* sys_io_setup:
1251  *	Create an aio_context capable of receiving at least nr_events.
1252  *	ctxp must not point to an aio_context that already exists, and
1253  *	must be initialized to 0 prior to the call.  On successful
1254  *	creation of the aio_context, *ctxp is filled in with the resulting
1255  *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1256  *	if the specified nr_events exceeds internal limits.  May fail
1257  *	with -EAGAIN if the specified nr_events exceeds the user's limit
1258  *	of available events.  May fail with -ENOMEM if insufficient kernel
1259  *	resources are available.  May fail with -EFAULT if an invalid
1260  *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1261  *	implemented.
1262  */
1263 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1264 {
1265 	struct kioctx *ioctx = NULL;
1266 	unsigned long ctx;
1267 	long ret;
1268 
1269 	ret = get_user(ctx, ctxp);
1270 	if (unlikely(ret))
1271 		goto out;
1272 
1273 	ret = -EINVAL;
1274 	if (unlikely(ctx || nr_events == 0)) {
1275 		pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1276 		         ctx, nr_events);
1277 		goto out;
1278 	}
1279 
1280 	ioctx = ioctx_alloc(nr_events);
1281 	ret = PTR_ERR(ioctx);
1282 	if (!IS_ERR(ioctx)) {
1283 		ret = put_user(ioctx->user_id, ctxp);
1284 		if (!ret)
1285 			return 0;
1286 
1287 		get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1288 		io_destroy(ioctx);
1289 	}
1290 
1291 out:
1292 	return ret;
1293 }
1294 
1295 /* sys_io_destroy:
1296  *	Destroy the aio_context specified.  May cancel any outstanding
1297  *	AIOs and block on completion.  Will fail with -ENOSYS if not
1298  *	implemented.  May fail with -EFAULT if the context pointed to
1299  *	is invalid.
1300  */
1301 asmlinkage long sys_io_destroy(aio_context_t ctx)
1302 {
1303 	struct kioctx *ioctx = lookup_ioctx(ctx);
1304 	if (likely(NULL != ioctx)) {
1305 		io_destroy(ioctx);
1306 		return 0;
1307 	}
1308 	pr_debug("EINVAL: io_destroy: invalid context id\n");
1309 	return -EINVAL;
1310 }
1311 
1312 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1313 {
1314 	struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1315 
1316 	BUG_ON(ret <= 0);
1317 
1318 	while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1319 		ssize_t this = min((ssize_t)iov->iov_len, ret);
1320 		iov->iov_base += this;
1321 		iov->iov_len -= this;
1322 		iocb->ki_left -= this;
1323 		ret -= this;
1324 		if (iov->iov_len == 0) {
1325 			iocb->ki_cur_seg++;
1326 			iov++;
1327 		}
1328 	}
1329 
1330 	/* the caller should not have done more io than what fit in
1331 	 * the remaining iovecs */
1332 	BUG_ON(ret > 0 && iocb->ki_left == 0);
1333 }
1334 
1335 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1336 {
1337 	struct file *file = iocb->ki_filp;
1338 	struct address_space *mapping = file->f_mapping;
1339 	struct inode *inode = mapping->host;
1340 	ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1341 			 unsigned long, loff_t);
1342 	ssize_t ret = 0;
1343 	unsigned short opcode;
1344 
1345 	if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1346 		(iocb->ki_opcode == IOCB_CMD_PREAD)) {
1347 		rw_op = file->f_op->aio_read;
1348 		opcode = IOCB_CMD_PREADV;
1349 	} else {
1350 		rw_op = file->f_op->aio_write;
1351 		opcode = IOCB_CMD_PWRITEV;
1352 	}
1353 
1354 	/* This matches the pread()/pwrite() logic */
1355 	if (iocb->ki_pos < 0)
1356 		return -EINVAL;
1357 
1358 	do {
1359 		ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1360 			    iocb->ki_nr_segs - iocb->ki_cur_seg,
1361 			    iocb->ki_pos);
1362 		if (ret > 0)
1363 			aio_advance_iovec(iocb, ret);
1364 
1365 	/* retry all partial writes.  retry partial reads as long as its a
1366 	 * regular file. */
1367 	} while (ret > 0 && iocb->ki_left > 0 &&
1368 		 (opcode == IOCB_CMD_PWRITEV ||
1369 		  (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1370 
1371 	/* This means we must have transferred all that we could */
1372 	/* No need to retry anymore */
1373 	if ((ret == 0) || (iocb->ki_left == 0))
1374 		ret = iocb->ki_nbytes - iocb->ki_left;
1375 
1376 	/* If we managed to write some out we return that, rather than
1377 	 * the eventual error. */
1378 	if (opcode == IOCB_CMD_PWRITEV
1379 	    && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1380 	    && iocb->ki_nbytes - iocb->ki_left)
1381 		ret = iocb->ki_nbytes - iocb->ki_left;
1382 
1383 	return ret;
1384 }
1385 
1386 static ssize_t aio_fdsync(struct kiocb *iocb)
1387 {
1388 	struct file *file = iocb->ki_filp;
1389 	ssize_t ret = -EINVAL;
1390 
1391 	if (file->f_op->aio_fsync)
1392 		ret = file->f_op->aio_fsync(iocb, 1);
1393 	return ret;
1394 }
1395 
1396 static ssize_t aio_fsync(struct kiocb *iocb)
1397 {
1398 	struct file *file = iocb->ki_filp;
1399 	ssize_t ret = -EINVAL;
1400 
1401 	if (file->f_op->aio_fsync)
1402 		ret = file->f_op->aio_fsync(iocb, 0);
1403 	return ret;
1404 }
1405 
1406 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb)
1407 {
1408 	ssize_t ret;
1409 
1410 	ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf,
1411 				    kiocb->ki_nbytes, 1,
1412 				    &kiocb->ki_inline_vec, &kiocb->ki_iovec);
1413 	if (ret < 0)
1414 		goto out;
1415 
1416 	kiocb->ki_nr_segs = kiocb->ki_nbytes;
1417 	kiocb->ki_cur_seg = 0;
1418 	/* ki_nbytes/left now reflect bytes instead of segs */
1419 	kiocb->ki_nbytes = ret;
1420 	kiocb->ki_left = ret;
1421 
1422 	ret = 0;
1423 out:
1424 	return ret;
1425 }
1426 
1427 static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1428 {
1429 	kiocb->ki_iovec = &kiocb->ki_inline_vec;
1430 	kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1431 	kiocb->ki_iovec->iov_len = kiocb->ki_left;
1432 	kiocb->ki_nr_segs = 1;
1433 	kiocb->ki_cur_seg = 0;
1434 	return 0;
1435 }
1436 
1437 /*
1438  * aio_setup_iocb:
1439  *	Performs the initial checks and aio retry method
1440  *	setup for the kiocb at the time of io submission.
1441  */
1442 static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1443 {
1444 	struct file *file = kiocb->ki_filp;
1445 	ssize_t ret = 0;
1446 
1447 	switch (kiocb->ki_opcode) {
1448 	case IOCB_CMD_PREAD:
1449 		ret = -EBADF;
1450 		if (unlikely(!(file->f_mode & FMODE_READ)))
1451 			break;
1452 		ret = -EFAULT;
1453 		if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1454 			kiocb->ki_left)))
1455 			break;
1456 		ret = security_file_permission(file, MAY_READ);
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_read)
1464 			kiocb->ki_retry = aio_rw_vect_retry;
1465 		break;
1466 	case IOCB_CMD_PWRITE:
1467 		ret = -EBADF;
1468 		if (unlikely(!(file->f_mode & FMODE_WRITE)))
1469 			break;
1470 		ret = -EFAULT;
1471 		if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1472 			kiocb->ki_left)))
1473 			break;
1474 		ret = security_file_permission(file, MAY_WRITE);
1475 		if (unlikely(ret))
1476 			break;
1477 		ret = aio_setup_single_vector(kiocb);
1478 		if (ret)
1479 			break;
1480 		ret = -EINVAL;
1481 		if (file->f_op->aio_write)
1482 			kiocb->ki_retry = aio_rw_vect_retry;
1483 		break;
1484 	case IOCB_CMD_PREADV:
1485 		ret = -EBADF;
1486 		if (unlikely(!(file->f_mode & FMODE_READ)))
1487 			break;
1488 		ret = security_file_permission(file, MAY_READ);
1489 		if (unlikely(ret))
1490 			break;
1491 		ret = aio_setup_vectored_rw(READ, kiocb);
1492 		if (ret)
1493 			break;
1494 		ret = -EINVAL;
1495 		if (file->f_op->aio_read)
1496 			kiocb->ki_retry = aio_rw_vect_retry;
1497 		break;
1498 	case IOCB_CMD_PWRITEV:
1499 		ret = -EBADF;
1500 		if (unlikely(!(file->f_mode & FMODE_WRITE)))
1501 			break;
1502 		ret = security_file_permission(file, MAY_WRITE);
1503 		if (unlikely(ret))
1504 			break;
1505 		ret = aio_setup_vectored_rw(WRITE, kiocb);
1506 		if (ret)
1507 			break;
1508 		ret = -EINVAL;
1509 		if (file->f_op->aio_write)
1510 			kiocb->ki_retry = aio_rw_vect_retry;
1511 		break;
1512 	case IOCB_CMD_FDSYNC:
1513 		ret = -EINVAL;
1514 		if (file->f_op->aio_fsync)
1515 			kiocb->ki_retry = aio_fdsync;
1516 		break;
1517 	case IOCB_CMD_FSYNC:
1518 		ret = -EINVAL;
1519 		if (file->f_op->aio_fsync)
1520 			kiocb->ki_retry = aio_fsync;
1521 		break;
1522 	default:
1523 		dprintk("EINVAL: io_submit: no operation provided\n");
1524 		ret = -EINVAL;
1525 	}
1526 
1527 	if (!kiocb->ki_retry)
1528 		return ret;
1529 
1530 	return 0;
1531 }
1532 
1533 /*
1534  * aio_wake_function:
1535  * 	wait queue callback function for aio notification,
1536  * 	Simply triggers a retry of the operation via kick_iocb.
1537  *
1538  * 	This callback is specified in the wait queue entry in
1539  *	a kiocb.
1540  *
1541  * Note:
1542  * This routine is executed with the wait queue lock held.
1543  * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1544  * the ioctx lock inside the wait queue lock. This is safe
1545  * because this callback isn't used for wait queues which
1546  * are nested inside ioctx lock (i.e. ctx->wait)
1547  */
1548 static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1549 			     int sync, void *key)
1550 {
1551 	struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1552 
1553 	list_del_init(&wait->task_list);
1554 	kick_iocb(iocb);
1555 	return 1;
1556 }
1557 
1558 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1559 			 struct iocb *iocb)
1560 {
1561 	struct kiocb *req;
1562 	struct file *file;
1563 	ssize_t ret;
1564 
1565 	/* enforce forwards compatibility on users */
1566 	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1567 		pr_debug("EINVAL: io_submit: reserve field set\n");
1568 		return -EINVAL;
1569 	}
1570 
1571 	/* prevent overflows */
1572 	if (unlikely(
1573 	    (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1574 	    (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1575 	    ((ssize_t)iocb->aio_nbytes < 0)
1576 	   )) {
1577 		pr_debug("EINVAL: io_submit: overflow check\n");
1578 		return -EINVAL;
1579 	}
1580 
1581 	file = fget(iocb->aio_fildes);
1582 	if (unlikely(!file))
1583 		return -EBADF;
1584 
1585 	req = aio_get_req(ctx);		/* returns with 2 references to req */
1586 	if (unlikely(!req)) {
1587 		fput(file);
1588 		return -EAGAIN;
1589 	}
1590 	req->ki_filp = file;
1591 	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1592 		/*
1593 		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1594 		 * instance of the file* now. The file descriptor must be
1595 		 * an eventfd() fd, and will be signaled for each completed
1596 		 * event using the eventfd_signal() function.
1597 		 */
1598 		req->ki_eventfd = eventfd_fget((int) iocb->aio_resfd);
1599 		if (IS_ERR(req->ki_eventfd)) {
1600 			ret = PTR_ERR(req->ki_eventfd);
1601 			goto out_put_req;
1602 		}
1603 	}
1604 
1605 	ret = put_user(req->ki_key, &user_iocb->aio_key);
1606 	if (unlikely(ret)) {
1607 		dprintk("EFAULT: aio_key\n");
1608 		goto out_put_req;
1609 	}
1610 
1611 	req->ki_obj.user = user_iocb;
1612 	req->ki_user_data = iocb->aio_data;
1613 	req->ki_pos = iocb->aio_offset;
1614 
1615 	req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1616 	req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1617 	req->ki_opcode = iocb->aio_lio_opcode;
1618 	init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1619 	INIT_LIST_HEAD(&req->ki_wait.task_list);
1620 
1621 	ret = aio_setup_iocb(req);
1622 
1623 	if (ret)
1624 		goto out_put_req;
1625 
1626 	spin_lock_irq(&ctx->ctx_lock);
1627 	aio_run_iocb(req);
1628 	if (!list_empty(&ctx->run_list)) {
1629 		/* drain the run list */
1630 		while (__aio_run_iocbs(ctx))
1631 			;
1632 	}
1633 	spin_unlock_irq(&ctx->ctx_lock);
1634 	aio_put_req(req);	/* drop extra ref to req */
1635 	return 0;
1636 
1637 out_put_req:
1638 	aio_put_req(req);	/* drop extra ref to req */
1639 	aio_put_req(req);	/* drop i/o ref to req */
1640 	return ret;
1641 }
1642 
1643 /* sys_io_submit:
1644  *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1645  *	the number of iocbs queued.  May return -EINVAL if the aio_context
1646  *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
1647  *	*iocbpp[0] is not properly initialized, if the operation specified
1648  *	is invalid for the file descriptor in the iocb.  May fail with
1649  *	-EFAULT if any of the data structures point to invalid data.  May
1650  *	fail with -EBADF if the file descriptor specified in the first
1651  *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
1652  *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1653  *	fail with -ENOSYS if not implemented.
1654  */
1655 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1656 			      struct iocb __user * __user *iocbpp)
1657 {
1658 	struct kioctx *ctx;
1659 	long ret = 0;
1660 	int i;
1661 
1662 	if (unlikely(nr < 0))
1663 		return -EINVAL;
1664 
1665 	if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1666 		return -EFAULT;
1667 
1668 	ctx = lookup_ioctx(ctx_id);
1669 	if (unlikely(!ctx)) {
1670 		pr_debug("EINVAL: io_submit: invalid context id\n");
1671 		return -EINVAL;
1672 	}
1673 
1674 	/*
1675 	 * AKPM: should this return a partial result if some of the IOs were
1676 	 * successfully submitted?
1677 	 */
1678 	for (i=0; i<nr; i++) {
1679 		struct iocb __user *user_iocb;
1680 		struct iocb tmp;
1681 
1682 		if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1683 			ret = -EFAULT;
1684 			break;
1685 		}
1686 
1687 		if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1688 			ret = -EFAULT;
1689 			break;
1690 		}
1691 
1692 		ret = io_submit_one(ctx, user_iocb, &tmp);
1693 		if (ret)
1694 			break;
1695 	}
1696 
1697 	put_ioctx(ctx);
1698 	return i ? i : ret;
1699 }
1700 
1701 /* lookup_kiocb
1702  *	Finds a given iocb for cancellation.
1703  */
1704 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1705 				  u32 key)
1706 {
1707 	struct list_head *pos;
1708 
1709 	assert_spin_locked(&ctx->ctx_lock);
1710 
1711 	/* TODO: use a hash or array, this sucks. */
1712 	list_for_each(pos, &ctx->active_reqs) {
1713 		struct kiocb *kiocb = list_kiocb(pos);
1714 		if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1715 			return kiocb;
1716 	}
1717 	return NULL;
1718 }
1719 
1720 /* sys_io_cancel:
1721  *	Attempts to cancel an iocb previously passed to io_submit.  If
1722  *	the operation is successfully cancelled, the resulting event is
1723  *	copied into the memory pointed to by result without being placed
1724  *	into the completion queue and 0 is returned.  May fail with
1725  *	-EFAULT if any of the data structures pointed to are invalid.
1726  *	May fail with -EINVAL if aio_context specified by ctx_id is
1727  *	invalid.  May fail with -EAGAIN if the iocb specified was not
1728  *	cancelled.  Will fail with -ENOSYS if not implemented.
1729  */
1730 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1731 			      struct io_event __user *result)
1732 {
1733 	int (*cancel)(struct kiocb *iocb, struct io_event *res);
1734 	struct kioctx *ctx;
1735 	struct kiocb *kiocb;
1736 	u32 key;
1737 	int ret;
1738 
1739 	ret = get_user(key, &iocb->aio_key);
1740 	if (unlikely(ret))
1741 		return -EFAULT;
1742 
1743 	ctx = lookup_ioctx(ctx_id);
1744 	if (unlikely(!ctx))
1745 		return -EINVAL;
1746 
1747 	spin_lock_irq(&ctx->ctx_lock);
1748 	ret = -EAGAIN;
1749 	kiocb = lookup_kiocb(ctx, iocb, key);
1750 	if (kiocb && kiocb->ki_cancel) {
1751 		cancel = kiocb->ki_cancel;
1752 		kiocb->ki_users ++;
1753 		kiocbSetCancelled(kiocb);
1754 	} else
1755 		cancel = NULL;
1756 	spin_unlock_irq(&ctx->ctx_lock);
1757 
1758 	if (NULL != cancel) {
1759 		struct io_event tmp;
1760 		pr_debug("calling cancel\n");
1761 		memset(&tmp, 0, sizeof(tmp));
1762 		tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1763 		tmp.data = kiocb->ki_user_data;
1764 		ret = cancel(kiocb, &tmp);
1765 		if (!ret) {
1766 			/* Cancellation succeeded -- copy the result
1767 			 * into the user's buffer.
1768 			 */
1769 			if (copy_to_user(result, &tmp, sizeof(tmp)))
1770 				ret = -EFAULT;
1771 		}
1772 	} else
1773 		ret = -EINVAL;
1774 
1775 	put_ioctx(ctx);
1776 
1777 	return ret;
1778 }
1779 
1780 /* io_getevents:
1781  *	Attempts to read at least min_nr events and up to nr events from
1782  *	the completion queue for the aio_context specified by ctx_id.  May
1783  *	fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1784  *	if nr is out of range, if when is out of range.  May fail with
1785  *	-EFAULT if any of the memory specified to is invalid.  May return
1786  *	0 or < min_nr if no events are available and the timeout specified
1787  *	by when	has elapsed, where when == NULL specifies an infinite
1788  *	timeout.  Note that the timeout pointed to by when is relative and
1789  *	will be updated if not NULL and the operation blocks.  Will fail
1790  *	with -ENOSYS if not implemented.
1791  */
1792 asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1793 				 long min_nr,
1794 				 long nr,
1795 				 struct io_event __user *events,
1796 				 struct timespec __user *timeout)
1797 {
1798 	struct kioctx *ioctx = lookup_ioctx(ctx_id);
1799 	long ret = -EINVAL;
1800 
1801 	if (likely(ioctx)) {
1802 		if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1803 			ret = read_events(ioctx, min_nr, nr, events, timeout);
1804 		put_ioctx(ioctx);
1805 	}
1806 
1807 	asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1808 	return ret;
1809 }
1810 
1811 __initcall(aio_setup);
1812 
1813 EXPORT_SYMBOL(aio_complete);
1814 EXPORT_SYMBOL(aio_put_req);
1815 EXPORT_SYMBOL(wait_on_sync_kiocb);
1816