xref: /linux/fs/aio.c (revision 54a8a2220c936a47840c9a3d74910c5a56fae2ed)
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 
19 #define DEBUG 0
20 
21 #include <linux/sched.h>
22 #include <linux/fs.h>
23 #include <linux/file.h>
24 #include <linux/mm.h>
25 #include <linux/mman.h>
26 #include <linux/slab.h>
27 #include <linux/timer.h>
28 #include <linux/aio.h>
29 #include <linux/highmem.h>
30 #include <linux/workqueue.h>
31 #include <linux/security.h>
32 #include <linux/rcuref.h>
33 
34 #include <asm/kmap_types.h>
35 #include <asm/uaccess.h>
36 #include <asm/mmu_context.h>
37 
38 #if DEBUG > 1
39 #define dprintk		printk
40 #else
41 #define dprintk(x...)	do { ; } while (0)
42 #endif
43 
44 /*------ sysctl variables----*/
45 atomic_t aio_nr = ATOMIC_INIT(0);	/* current system wide number of aio requests */
46 unsigned aio_max_nr = 0x10000;	/* system wide maximum number of aio requests */
47 /*----end sysctl variables---*/
48 
49 static kmem_cache_t	*kiocb_cachep;
50 static kmem_cache_t	*kioctx_cachep;
51 
52 static struct workqueue_struct *aio_wq;
53 
54 /* Used for rare fput completion. */
55 static void aio_fput_routine(void *);
56 static DECLARE_WORK(fput_work, aio_fput_routine, NULL);
57 
58 static DEFINE_SPINLOCK(fput_lock);
59 static LIST_HEAD(fput_head);
60 
61 static void aio_kick_handler(void *);
62 static void aio_queue_work(struct kioctx *);
63 
64 /* aio_setup
65  *	Creates the slab caches used by the aio routines, panic on
66  *	failure as this is done early during the boot sequence.
67  */
68 static int __init aio_setup(void)
69 {
70 	kiocb_cachep = kmem_cache_create("kiocb", sizeof(struct kiocb),
71 				0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
72 	kioctx_cachep = kmem_cache_create("kioctx", sizeof(struct kioctx),
73 				0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
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 = kmalloc(sizeof(struct page *) * nr_pages, GFP_KERNEL);
126 		if (!info->ring_pages)
127 			return -ENOMEM;
128 		memset(info->ring_pages, 0, sizeof(struct page *) * nr_pages);
129 	}
130 
131 	info->mmap_size = nr_pages * PAGE_SIZE;
132 	dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
133 	down_write(&ctx->mm->mmap_sem);
134 	info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
135 				  PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE,
136 				  0);
137 	if (IS_ERR((void *)info->mmap_base)) {
138 		up_write(&ctx->mm->mmap_sem);
139 		printk("mmap err: %ld\n", -info->mmap_base);
140 		info->mmap_size = 0;
141 		aio_free_ring(ctx);
142 		return -EAGAIN;
143 	}
144 
145 	dprintk("mmap address: 0x%08lx\n", info->mmap_base);
146 	info->nr_pages = get_user_pages(current, ctx->mm,
147 					info->mmap_base, nr_pages,
148 					1, 0, info->ring_pages, NULL);
149 	up_write(&ctx->mm->mmap_sem);
150 
151 	if (unlikely(info->nr_pages != nr_pages)) {
152 		aio_free_ring(ctx);
153 		return -EAGAIN;
154 	}
155 
156 	ctx->user_id = info->mmap_base;
157 
158 	info->nr = nr_events;		/* trusted copy */
159 
160 	ring = kmap_atomic(info->ring_pages[0], KM_USER0);
161 	ring->nr = nr_events;	/* user copy */
162 	ring->id = ctx->user_id;
163 	ring->head = ring->tail = 0;
164 	ring->magic = AIO_RING_MAGIC;
165 	ring->compat_features = AIO_RING_COMPAT_FEATURES;
166 	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
167 	ring->header_length = sizeof(struct aio_ring);
168 	kunmap_atomic(ring, KM_USER0);
169 
170 	return 0;
171 }
172 
173 
174 /* aio_ring_event: returns a pointer to the event at the given index from
175  * kmap_atomic(, km).  Release the pointer with put_aio_ring_event();
176  */
177 #define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
178 #define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
179 #define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
180 
181 #define aio_ring_event(info, nr, km) ({					\
182 	unsigned pos = (nr) + AIO_EVENTS_OFFSET;			\
183 	struct io_event *__event;					\
184 	__event = kmap_atomic(						\
185 			(info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
186 	__event += pos % AIO_EVENTS_PER_PAGE;				\
187 	__event;							\
188 })
189 
190 #define put_aio_ring_event(event, km) do {	\
191 	struct io_event *__event = (event);	\
192 	(void)__event;				\
193 	kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
194 } while(0)
195 
196 /* ioctx_alloc
197  *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
198  */
199 static struct kioctx *ioctx_alloc(unsigned nr_events)
200 {
201 	struct mm_struct *mm;
202 	struct kioctx *ctx;
203 
204 	/* Prevent overflows */
205 	if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
206 	    (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
207 		pr_debug("ENOMEM: nr_events too high\n");
208 		return ERR_PTR(-EINVAL);
209 	}
210 
211 	if (nr_events > aio_max_nr)
212 		return ERR_PTR(-EAGAIN);
213 
214 	ctx = kmem_cache_alloc(kioctx_cachep, GFP_KERNEL);
215 	if (!ctx)
216 		return ERR_PTR(-ENOMEM);
217 
218 	memset(ctx, 0, sizeof(*ctx));
219 	ctx->max_reqs = nr_events;
220 	mm = ctx->mm = current->mm;
221 	atomic_inc(&mm->mm_count);
222 
223 	atomic_set(&ctx->users, 1);
224 	spin_lock_init(&ctx->ctx_lock);
225 	spin_lock_init(&ctx->ring_info.ring_lock);
226 	init_waitqueue_head(&ctx->wait);
227 
228 	INIT_LIST_HEAD(&ctx->active_reqs);
229 	INIT_LIST_HEAD(&ctx->run_list);
230 	INIT_WORK(&ctx->wq, aio_kick_handler, ctx);
231 
232 	if (aio_setup_ring(ctx) < 0)
233 		goto out_freectx;
234 
235 	/* limit the number of system wide aios */
236 	atomic_add(ctx->max_reqs, &aio_nr);	/* undone by __put_ioctx */
237 	if (unlikely(atomic_read(&aio_nr) > aio_max_nr))
238 		goto out_cleanup;
239 
240 	/* now link into global list.  kludge.  FIXME */
241 	write_lock(&mm->ioctx_list_lock);
242 	ctx->next = mm->ioctx_list;
243 	mm->ioctx_list = ctx;
244 	write_unlock(&mm->ioctx_list_lock);
245 
246 	dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
247 		ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
248 	return ctx;
249 
250 out_cleanup:
251 	atomic_sub(ctx->max_reqs, &aio_nr);
252 	ctx->max_reqs = 0;	/* prevent __put_ioctx from sub'ing aio_nr */
253 	__put_ioctx(ctx);
254 	return ERR_PTR(-EAGAIN);
255 
256 out_freectx:
257 	mmdrop(mm);
258 	kmem_cache_free(kioctx_cachep, ctx);
259 	ctx = ERR_PTR(-ENOMEM);
260 
261 	dprintk("aio: error allocating ioctx %p\n", ctx);
262 	return ctx;
263 }
264 
265 /* aio_cancel_all
266  *	Cancels all outstanding aio requests on an aio context.  Used
267  *	when the processes owning a context have all exited to encourage
268  *	the rapid destruction of the kioctx.
269  */
270 static void aio_cancel_all(struct kioctx *ctx)
271 {
272 	int (*cancel)(struct kiocb *, struct io_event *);
273 	struct io_event res;
274 	spin_lock_irq(&ctx->ctx_lock);
275 	ctx->dead = 1;
276 	while (!list_empty(&ctx->active_reqs)) {
277 		struct list_head *pos = ctx->active_reqs.next;
278 		struct kiocb *iocb = list_kiocb(pos);
279 		list_del_init(&iocb->ki_list);
280 		cancel = iocb->ki_cancel;
281 		kiocbSetCancelled(iocb);
282 		if (cancel) {
283 			iocb->ki_users++;
284 			spin_unlock_irq(&ctx->ctx_lock);
285 			cancel(iocb, &res);
286 			spin_lock_irq(&ctx->ctx_lock);
287 		}
288 	}
289 	spin_unlock_irq(&ctx->ctx_lock);
290 }
291 
292 static void wait_for_all_aios(struct kioctx *ctx)
293 {
294 	struct task_struct *tsk = current;
295 	DECLARE_WAITQUEUE(wait, tsk);
296 
297 	if (!ctx->reqs_active)
298 		return;
299 
300 	add_wait_queue(&ctx->wait, &wait);
301 	set_task_state(tsk, TASK_UNINTERRUPTIBLE);
302 	while (ctx->reqs_active) {
303 		schedule();
304 		set_task_state(tsk, TASK_UNINTERRUPTIBLE);
305 	}
306 	__set_task_state(tsk, TASK_RUNNING);
307 	remove_wait_queue(&ctx->wait, &wait);
308 }
309 
310 /* wait_on_sync_kiocb:
311  *	Waits on the given sync kiocb to complete.
312  */
313 ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb)
314 {
315 	while (iocb->ki_users) {
316 		set_current_state(TASK_UNINTERRUPTIBLE);
317 		if (!iocb->ki_users)
318 			break;
319 		schedule();
320 	}
321 	__set_current_state(TASK_RUNNING);
322 	return iocb->ki_user_data;
323 }
324 
325 /* exit_aio: called when the last user of mm goes away.  At this point,
326  * there is no way for any new requests to be submited or any of the
327  * io_* syscalls to be called on the context.  However, there may be
328  * outstanding requests which hold references to the context; as they
329  * go away, they will call put_ioctx and release any pinned memory
330  * associated with the request (held via struct page * references).
331  */
332 void fastcall exit_aio(struct mm_struct *mm)
333 {
334 	struct kioctx *ctx = mm->ioctx_list;
335 	mm->ioctx_list = NULL;
336 	while (ctx) {
337 		struct kioctx *next = ctx->next;
338 		ctx->next = NULL;
339 		aio_cancel_all(ctx);
340 
341 		wait_for_all_aios(ctx);
342 		/*
343 		 * this is an overkill, but ensures we don't leave
344 		 * the ctx on the aio_wq
345 		 */
346 		flush_workqueue(aio_wq);
347 
348 		if (1 != atomic_read(&ctx->users))
349 			printk(KERN_DEBUG
350 				"exit_aio:ioctx still alive: %d %d %d\n",
351 				atomic_read(&ctx->users), ctx->dead,
352 				ctx->reqs_active);
353 		put_ioctx(ctx);
354 		ctx = next;
355 	}
356 }
357 
358 /* __put_ioctx
359  *	Called when the last user of an aio context has gone away,
360  *	and the struct needs to be freed.
361  */
362 void fastcall __put_ioctx(struct kioctx *ctx)
363 {
364 	unsigned nr_events = ctx->max_reqs;
365 
366 	if (unlikely(ctx->reqs_active))
367 		BUG();
368 
369 	cancel_delayed_work(&ctx->wq);
370 	flush_workqueue(aio_wq);
371 	aio_free_ring(ctx);
372 	mmdrop(ctx->mm);
373 	ctx->mm = NULL;
374 	pr_debug("__put_ioctx: freeing %p\n", ctx);
375 	kmem_cache_free(kioctx_cachep, ctx);
376 
377 	atomic_sub(nr_events, &aio_nr);
378 }
379 
380 /* aio_get_req
381  *	Allocate a slot for an aio request.  Increments the users count
382  * of the kioctx so that the kioctx stays around until all requests are
383  * complete.  Returns NULL if no requests are free.
384  *
385  * Returns with kiocb->users set to 2.  The io submit code path holds
386  * an extra reference while submitting the i/o.
387  * This prevents races between the aio code path referencing the
388  * req (after submitting it) and aio_complete() freeing the req.
389  */
390 static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx));
391 static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx)
392 {
393 	struct kiocb *req = NULL;
394 	struct aio_ring *ring;
395 	int okay = 0;
396 
397 	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
398 	if (unlikely(!req))
399 		return NULL;
400 
401 	req->ki_flags = 1 << KIF_LOCKED;
402 	req->ki_users = 2;
403 	req->ki_key = 0;
404 	req->ki_ctx = ctx;
405 	req->ki_cancel = NULL;
406 	req->ki_retry = NULL;
407 	req->ki_dtor = NULL;
408 	req->private = NULL;
409 	INIT_LIST_HEAD(&req->ki_run_list);
410 
411 	/* Check if the completion queue has enough free space to
412 	 * accept an event from this io.
413 	 */
414 	spin_lock_irq(&ctx->ctx_lock);
415 	ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
416 	if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
417 		list_add(&req->ki_list, &ctx->active_reqs);
418 		get_ioctx(ctx);
419 		ctx->reqs_active++;
420 		okay = 1;
421 	}
422 	kunmap_atomic(ring, KM_USER0);
423 	spin_unlock_irq(&ctx->ctx_lock);
424 
425 	if (!okay) {
426 		kmem_cache_free(kiocb_cachep, req);
427 		req = NULL;
428 	}
429 
430 	return req;
431 }
432 
433 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
434 {
435 	struct kiocb *req;
436 	/* Handle a potential starvation case -- should be exceedingly rare as
437 	 * requests will be stuck on fput_head only if the aio_fput_routine is
438 	 * delayed and the requests were the last user of the struct file.
439 	 */
440 	req = __aio_get_req(ctx);
441 	if (unlikely(NULL == req)) {
442 		aio_fput_routine(NULL);
443 		req = __aio_get_req(ctx);
444 	}
445 	return req;
446 }
447 
448 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
449 {
450 	if (req->ki_dtor)
451 		req->ki_dtor(req);
452 	kmem_cache_free(kiocb_cachep, req);
453 	ctx->reqs_active--;
454 
455 	if (unlikely(!ctx->reqs_active && ctx->dead))
456 		wake_up(&ctx->wait);
457 }
458 
459 static void aio_fput_routine(void *data)
460 {
461 	spin_lock_irq(&fput_lock);
462 	while (likely(!list_empty(&fput_head))) {
463 		struct kiocb *req = list_kiocb(fput_head.next);
464 		struct kioctx *ctx = req->ki_ctx;
465 
466 		list_del(&req->ki_list);
467 		spin_unlock_irq(&fput_lock);
468 
469 		/* Complete the fput */
470 		__fput(req->ki_filp);
471 
472 		/* Link the iocb into the context's free list */
473 		spin_lock_irq(&ctx->ctx_lock);
474 		really_put_req(ctx, req);
475 		spin_unlock_irq(&ctx->ctx_lock);
476 
477 		put_ioctx(ctx);
478 		spin_lock_irq(&fput_lock);
479 	}
480 	spin_unlock_irq(&fput_lock);
481 }
482 
483 /* __aio_put_req
484  *	Returns true if this put was the last user of the request.
485  */
486 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
487 {
488 	dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
489 		req, atomic_read(&req->ki_filp->f_count));
490 
491 	req->ki_users --;
492 	if (unlikely(req->ki_users < 0))
493 		BUG();
494 	if (likely(req->ki_users))
495 		return 0;
496 	list_del(&req->ki_list);		/* remove from active_reqs */
497 	req->ki_cancel = NULL;
498 	req->ki_retry = NULL;
499 
500 	/* Must be done under the lock to serialise against cancellation.
501 	 * Call this aio_fput as it duplicates fput via the fput_work.
502 	 */
503 	if (unlikely(rcuref_dec_and_test(&req->ki_filp->f_count))) {
504 		get_ioctx(ctx);
505 		spin_lock(&fput_lock);
506 		list_add(&req->ki_list, &fput_head);
507 		spin_unlock(&fput_lock);
508 		queue_work(aio_wq, &fput_work);
509 	} else
510 		really_put_req(ctx, req);
511 	return 1;
512 }
513 
514 /* aio_put_req
515  *	Returns true if this put was the last user of the kiocb,
516  *	false if the request is still in use.
517  */
518 int fastcall aio_put_req(struct kiocb *req)
519 {
520 	struct kioctx *ctx = req->ki_ctx;
521 	int ret;
522 	spin_lock_irq(&ctx->ctx_lock);
523 	ret = __aio_put_req(ctx, req);
524 	spin_unlock_irq(&ctx->ctx_lock);
525 	if (ret)
526 		put_ioctx(ctx);
527 	return ret;
528 }
529 
530 /*	Lookup an ioctx id.  ioctx_list is lockless for reads.
531  *	FIXME: this is O(n) and is only suitable for development.
532  */
533 struct kioctx *lookup_ioctx(unsigned long ctx_id)
534 {
535 	struct kioctx *ioctx;
536 	struct mm_struct *mm;
537 
538 	mm = current->mm;
539 	read_lock(&mm->ioctx_list_lock);
540 	for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
541 		if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
542 			get_ioctx(ioctx);
543 			break;
544 		}
545 	read_unlock(&mm->ioctx_list_lock);
546 
547 	return ioctx;
548 }
549 
550 static int lock_kiocb_action(void *param)
551 {
552 	schedule();
553 	return 0;
554 }
555 
556 static inline void lock_kiocb(struct kiocb *iocb)
557 {
558 	wait_on_bit_lock(&iocb->ki_flags, KIF_LOCKED, lock_kiocb_action,
559 			 TASK_UNINTERRUPTIBLE);
560 }
561 
562 static inline void unlock_kiocb(struct kiocb *iocb)
563 {
564 	kiocbClearLocked(iocb);
565 	smp_mb__after_clear_bit();
566 	wake_up_bit(&iocb->ki_flags, KIF_LOCKED);
567 }
568 
569 /*
570  * use_mm
571  *	Makes the calling kernel thread take on the specified
572  *	mm context.
573  *	Called by the retry thread execute retries within the
574  *	iocb issuer's mm context, so that copy_from/to_user
575  *	operations work seamlessly for aio.
576  *	(Note: this routine is intended to be called only
577  *	from a kernel thread context)
578  */
579 static void use_mm(struct mm_struct *mm)
580 {
581 	struct mm_struct *active_mm;
582 	struct task_struct *tsk = current;
583 
584 	task_lock(tsk);
585 	tsk->flags |= PF_BORROWED_MM;
586 	active_mm = tsk->active_mm;
587 	atomic_inc(&mm->mm_count);
588 	tsk->mm = mm;
589 	tsk->active_mm = mm;
590 	/*
591 	 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise
592 	 * it won't work. Update it accordingly if you change it here
593 	 */
594 	activate_mm(active_mm, mm);
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  * Comments: Called with ctx->ctx_lock held. This nests
609  * task_lock instead ctx_lock.
610  */
611 static void unuse_mm(struct mm_struct *mm)
612 {
613 	struct task_struct *tsk = current;
614 
615 	task_lock(tsk);
616 	tsk->flags &= ~PF_BORROWED_MM;
617 	tsk->mm = NULL;
618 	/* active_mm is still 'mm' */
619 	enter_lazy_tlb(mm, tsk);
620 	task_unlock(tsk);
621 }
622 
623 /*
624  * Queue up a kiocb to be retried. Assumes that the kiocb
625  * has already been marked as kicked, and places it on
626  * the retry run list for the corresponding ioctx, if it
627  * isn't already queued. Returns 1 if it actually queued
628  * the kiocb (to tell the caller to activate the work
629  * queue to process it), or 0, if it found that it was
630  * already queued.
631  *
632  * Should be called with the spin lock iocb->ki_ctx->ctx_lock
633  * held
634  */
635 static inline int __queue_kicked_iocb(struct kiocb *iocb)
636 {
637 	struct kioctx *ctx = iocb->ki_ctx;
638 
639 	if (list_empty(&iocb->ki_run_list)) {
640 		list_add_tail(&iocb->ki_run_list,
641 			&ctx->run_list);
642 		return 1;
643 	}
644 	return 0;
645 }
646 
647 /* aio_run_iocb
648  *	This is the core aio execution routine. It is
649  *	invoked both for initial i/o submission and
650  *	subsequent retries via the aio_kick_handler.
651  *	Expects to be invoked with iocb->ki_ctx->lock
652  *	already held. The lock is released and reaquired
653  *	as needed during processing.
654  *
655  * Calls the iocb retry method (already setup for the
656  * iocb on initial submission) for operation specific
657  * handling, but takes care of most of common retry
658  * execution details for a given iocb. The retry method
659  * needs to be non-blocking as far as possible, to avoid
660  * holding up other iocbs waiting to be serviced by the
661  * retry kernel thread.
662  *
663  * The trickier parts in this code have to do with
664  * ensuring that only one retry instance is in progress
665  * for a given iocb at any time. Providing that guarantee
666  * simplifies the coding of individual aio operations as
667  * it avoids various potential races.
668  */
669 static ssize_t aio_run_iocb(struct kiocb *iocb)
670 {
671 	struct kioctx	*ctx = iocb->ki_ctx;
672 	ssize_t (*retry)(struct kiocb *);
673 	ssize_t ret;
674 
675 	if (iocb->ki_retried++ > 1024*1024) {
676 		printk("Maximal retry count.  Bytes done %Zd\n",
677 			iocb->ki_nbytes - iocb->ki_left);
678 		return -EAGAIN;
679 	}
680 
681 	if (!(iocb->ki_retried & 0xff)) {
682 		pr_debug("%ld retry: %d of %d\n", iocb->ki_retried,
683 			iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
684 	}
685 
686 	if (!(retry = iocb->ki_retry)) {
687 		printk("aio_run_iocb: iocb->ki_retry = NULL\n");
688 		return 0;
689 	}
690 
691 	/*
692 	 * We don't want the next retry iteration for this
693 	 * operation to start until this one has returned and
694 	 * updated the iocb state. However, wait_queue functions
695 	 * can trigger a kick_iocb from interrupt context in the
696 	 * meantime, indicating that data is available for the next
697 	 * iteration. We want to remember that and enable the
698 	 * next retry iteration _after_ we are through with
699 	 * this one.
700 	 *
701 	 * So, in order to be able to register a "kick", but
702 	 * prevent it from being queued now, we clear the kick
703 	 * flag, but make the kick code *think* that the iocb is
704 	 * still on the run list until we are actually done.
705 	 * When we are done with this iteration, we check if
706 	 * the iocb was kicked in the meantime and if so, queue
707 	 * it up afresh.
708 	 */
709 
710 	kiocbClearKicked(iocb);
711 
712 	/*
713 	 * This is so that aio_complete knows it doesn't need to
714 	 * pull the iocb off the run list (We can't just call
715 	 * INIT_LIST_HEAD because we don't want a kick_iocb to
716 	 * queue this on the run list yet)
717 	 */
718 	iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
719 	spin_unlock_irq(&ctx->ctx_lock);
720 
721 	/* Quit retrying if the i/o has been cancelled */
722 	if (kiocbIsCancelled(iocb)) {
723 		ret = -EINTR;
724 		aio_complete(iocb, ret, 0);
725 		/* must not access the iocb after this */
726 		goto out;
727 	}
728 
729 	/*
730 	 * Now we are all set to call the retry method in async
731 	 * context. By setting this thread's io_wait context
732 	 * to point to the wait queue entry inside the currently
733 	 * running iocb for the duration of the retry, we ensure
734 	 * that async notification wakeups are queued by the
735 	 * operation instead of blocking waits, and when notified,
736 	 * cause the iocb to be kicked for continuation (through
737 	 * the aio_wake_function callback).
738 	 */
739 	BUG_ON(current->io_wait != NULL);
740 	current->io_wait = &iocb->ki_wait;
741 	ret = retry(iocb);
742 	current->io_wait = NULL;
743 
744 	if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
745 		BUG_ON(!list_empty(&iocb->ki_wait.task_list));
746 		aio_complete(iocb, ret, 0);
747 	}
748 out:
749 	spin_lock_irq(&ctx->ctx_lock);
750 
751 	if (-EIOCBRETRY == ret) {
752 		/*
753 		 * OK, now that we are done with this iteration
754 		 * and know that there is more left to go,
755 		 * this is where we let go so that a subsequent
756 		 * "kick" can start the next iteration
757 		 */
758 
759 		/* will make __queue_kicked_iocb succeed from here on */
760 		INIT_LIST_HEAD(&iocb->ki_run_list);
761 		/* we must queue the next iteration ourselves, if it
762 		 * has already been kicked */
763 		if (kiocbIsKicked(iocb)) {
764 			__queue_kicked_iocb(iocb);
765 
766 			/*
767 			 * __queue_kicked_iocb will always return 1 here, because
768 			 * iocb->ki_run_list is empty at this point so it should
769 			 * be safe to unconditionally queue the context into the
770 			 * work queue.
771 			 */
772 			aio_queue_work(ctx);
773 		}
774 	}
775 	return ret;
776 }
777 
778 /*
779  * __aio_run_iocbs:
780  * 	Process all pending retries queued on the ioctx
781  * 	run list.
782  * Assumes it is operating within the aio issuer's mm
783  * context. Expects to be called with ctx->ctx_lock held
784  */
785 static int __aio_run_iocbs(struct kioctx *ctx)
786 {
787 	struct kiocb *iocb;
788 	LIST_HEAD(run_list);
789 
790 	list_splice_init(&ctx->run_list, &run_list);
791 	while (!list_empty(&run_list)) {
792 		iocb = list_entry(run_list.next, struct kiocb,
793 			ki_run_list);
794 		list_del(&iocb->ki_run_list);
795 		/*
796 		 * Hold an extra reference while retrying i/o.
797 		 */
798 		iocb->ki_users++;       /* grab extra reference */
799 		lock_kiocb(iocb);
800 		aio_run_iocb(iocb);
801 		unlock_kiocb(iocb);
802 		if (__aio_put_req(ctx, iocb))  /* drop extra ref */
803 			put_ioctx(ctx);
804  	}
805 	if (!list_empty(&ctx->run_list))
806 		return 1;
807 	return 0;
808 }
809 
810 static void aio_queue_work(struct kioctx * ctx)
811 {
812 	unsigned long timeout;
813 	/*
814 	 * if someone is waiting, get the work started right
815 	 * away, otherwise, use a longer delay
816 	 */
817 	smp_mb();
818 	if (waitqueue_active(&ctx->wait))
819 		timeout = 1;
820 	else
821 		timeout = HZ/10;
822 	queue_delayed_work(aio_wq, &ctx->wq, timeout);
823 }
824 
825 
826 /*
827  * aio_run_iocbs:
828  * 	Process all pending retries queued on the ioctx
829  * 	run list.
830  * Assumes it is operating within the aio issuer's mm
831  * context.
832  */
833 static inline void aio_run_iocbs(struct kioctx *ctx)
834 {
835 	int requeue;
836 
837 	spin_lock_irq(&ctx->ctx_lock);
838 
839 	requeue = __aio_run_iocbs(ctx);
840 	spin_unlock_irq(&ctx->ctx_lock);
841 	if (requeue)
842 		aio_queue_work(ctx);
843 }
844 
845 /*
846  * just like aio_run_iocbs, but keeps running them until
847  * the list stays empty
848  */
849 static inline void aio_run_all_iocbs(struct kioctx *ctx)
850 {
851 	spin_lock_irq(&ctx->ctx_lock);
852 	while (__aio_run_iocbs(ctx))
853 		;
854 	spin_unlock_irq(&ctx->ctx_lock);
855 }
856 
857 /*
858  * aio_kick_handler:
859  * 	Work queue handler triggered to process pending
860  * 	retries on an ioctx. Takes on the aio issuer's
861  *	mm context before running the iocbs, so that
862  *	copy_xxx_user operates on the issuer's address
863  *      space.
864  * Run on aiod's context.
865  */
866 static void aio_kick_handler(void *data)
867 {
868 	struct kioctx *ctx = data;
869 	mm_segment_t oldfs = get_fs();
870 	int requeue;
871 
872 	set_fs(USER_DS);
873 	use_mm(ctx->mm);
874 	spin_lock_irq(&ctx->ctx_lock);
875 	requeue =__aio_run_iocbs(ctx);
876  	unuse_mm(ctx->mm);
877 	spin_unlock_irq(&ctx->ctx_lock);
878 	set_fs(oldfs);
879 	/*
880 	 * we're in a worker thread already, don't use queue_delayed_work,
881 	 */
882 	if (requeue)
883 		queue_work(aio_wq, &ctx->wq);
884 }
885 
886 
887 /*
888  * Called by kick_iocb to queue the kiocb for retry
889  * and if required activate the aio work queue to process
890  * it
891  */
892 static void try_queue_kicked_iocb(struct kiocb *iocb)
893 {
894  	struct kioctx	*ctx = iocb->ki_ctx;
895 	unsigned long flags;
896 	int run = 0;
897 
898 	/* We're supposed to be the only path putting the iocb back on the run
899 	 * list.  If we find that the iocb is *back* on a wait queue already
900 	 * than retry has happened before we could queue the iocb.  This also
901 	 * means that the retry could have completed and freed our iocb, no
902 	 * good. */
903 	BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
904 
905 	spin_lock_irqsave(&ctx->ctx_lock, flags);
906 	/* set this inside the lock so that we can't race with aio_run_iocb()
907 	 * testing it and putting the iocb on the run list under the lock */
908 	if (!kiocbTryKick(iocb))
909 		run = __queue_kicked_iocb(iocb);
910 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
911 	if (run)
912 		aio_queue_work(ctx);
913 }
914 
915 /*
916  * kick_iocb:
917  *      Called typically from a wait queue callback context
918  *      (aio_wake_function) to trigger a retry of the iocb.
919  *      The retry is usually executed by aio workqueue
920  *      threads (See aio_kick_handler).
921  */
922 void fastcall kick_iocb(struct kiocb *iocb)
923 {
924 	/* sync iocbs are easy: they can only ever be executing from a
925 	 * single context. */
926 	if (is_sync_kiocb(iocb)) {
927 		kiocbSetKicked(iocb);
928 	        wake_up_process(iocb->ki_obj.tsk);
929 		return;
930 	}
931 
932 	try_queue_kicked_iocb(iocb);
933 }
934 EXPORT_SYMBOL(kick_iocb);
935 
936 /* aio_complete
937  *	Called when the io request on the given iocb is complete.
938  *	Returns true if this is the last user of the request.  The
939  *	only other user of the request can be the cancellation code.
940  */
941 int fastcall aio_complete(struct kiocb *iocb, long res, long res2)
942 {
943 	struct kioctx	*ctx = iocb->ki_ctx;
944 	struct aio_ring_info	*info;
945 	struct aio_ring	*ring;
946 	struct io_event	*event;
947 	unsigned long	flags;
948 	unsigned long	tail;
949 	int		ret;
950 
951 	/* Special case handling for sync iocbs: events go directly
952 	 * into the iocb for fast handling.  Note that this will not
953 	 * work if we allow sync kiocbs to be cancelled. in which
954 	 * case the usage count checks will have to move under ctx_lock
955 	 * for all cases.
956 	 */
957 	if (is_sync_kiocb(iocb)) {
958 		int ret;
959 
960 		iocb->ki_user_data = res;
961 		if (iocb->ki_users == 1) {
962 			iocb->ki_users = 0;
963 			ret = 1;
964 		} else {
965 			spin_lock_irq(&ctx->ctx_lock);
966 			iocb->ki_users--;
967 			ret = (0 == iocb->ki_users);
968 			spin_unlock_irq(&ctx->ctx_lock);
969 		}
970 		/* sync iocbs put the task here for us */
971 		wake_up_process(iocb->ki_obj.tsk);
972 		return ret;
973 	}
974 
975 	info = &ctx->ring_info;
976 
977 	/* add a completion event to the ring buffer.
978 	 * must be done holding ctx->ctx_lock to prevent
979 	 * other code from messing with the tail
980 	 * pointer since we might be called from irq
981 	 * context.
982 	 */
983 	spin_lock_irqsave(&ctx->ctx_lock, flags);
984 
985 	if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
986 		list_del_init(&iocb->ki_run_list);
987 
988 	/*
989 	 * cancelled requests don't get events, userland was given one
990 	 * when the event got cancelled.
991 	 */
992 	if (kiocbIsCancelled(iocb))
993 		goto put_rq;
994 
995 	ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
996 
997 	tail = info->tail;
998 	event = aio_ring_event(info, tail, KM_IRQ0);
999 	if (++tail >= info->nr)
1000 		tail = 0;
1001 
1002 	event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1003 	event->data = iocb->ki_user_data;
1004 	event->res = res;
1005 	event->res2 = res2;
1006 
1007 	dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
1008 		ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1009 		res, res2);
1010 
1011 	/* after flagging the request as done, we
1012 	 * must never even look at it again
1013 	 */
1014 	smp_wmb();	/* make event visible before updating tail */
1015 
1016 	info->tail = tail;
1017 	ring->tail = tail;
1018 
1019 	put_aio_ring_event(event, KM_IRQ0);
1020 	kunmap_atomic(ring, KM_IRQ1);
1021 
1022 	pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1023 
1024 	pr_debug("%ld retries: %d of %d\n", iocb->ki_retried,
1025 		iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
1026 put_rq:
1027 	/* everything turned out well, dispose of the aiocb. */
1028 	ret = __aio_put_req(ctx, iocb);
1029 
1030 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1031 
1032 	if (waitqueue_active(&ctx->wait))
1033 		wake_up(&ctx->wait);
1034 
1035 	if (ret)
1036 		put_ioctx(ctx);
1037 
1038 	return ret;
1039 }
1040 
1041 /* aio_read_evt
1042  *	Pull an event off of the ioctx's event ring.  Returns the number of
1043  *	events fetched (0 or 1 ;-)
1044  *	FIXME: make this use cmpxchg.
1045  *	TODO: make the ringbuffer user mmap()able (requires FIXME).
1046  */
1047 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1048 {
1049 	struct aio_ring_info *info = &ioctx->ring_info;
1050 	struct aio_ring *ring;
1051 	unsigned long head;
1052 	int ret = 0;
1053 
1054 	ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1055 	dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1056 		 (unsigned long)ring->head, (unsigned long)ring->tail,
1057 		 (unsigned long)ring->nr);
1058 
1059 	if (ring->head == ring->tail)
1060 		goto out;
1061 
1062 	spin_lock(&info->ring_lock);
1063 
1064 	head = ring->head % info->nr;
1065 	if (head != ring->tail) {
1066 		struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1067 		*ent = *evp;
1068 		head = (head + 1) % info->nr;
1069 		smp_mb(); /* finish reading the event before updatng the head */
1070 		ring->head = head;
1071 		ret = 1;
1072 		put_aio_ring_event(evp, KM_USER1);
1073 	}
1074 	spin_unlock(&info->ring_lock);
1075 
1076 out:
1077 	kunmap_atomic(ring, KM_USER0);
1078 	dprintk("leaving aio_read_evt: %d  h%lu t%lu\n", ret,
1079 		 (unsigned long)ring->head, (unsigned long)ring->tail);
1080 	return ret;
1081 }
1082 
1083 struct aio_timeout {
1084 	struct timer_list	timer;
1085 	int			timed_out;
1086 	struct task_struct	*p;
1087 };
1088 
1089 static void timeout_func(unsigned long data)
1090 {
1091 	struct aio_timeout *to = (struct aio_timeout *)data;
1092 
1093 	to->timed_out = 1;
1094 	wake_up_process(to->p);
1095 }
1096 
1097 static inline void init_timeout(struct aio_timeout *to)
1098 {
1099 	init_timer(&to->timer);
1100 	to->timer.data = (unsigned long)to;
1101 	to->timer.function = timeout_func;
1102 	to->timed_out = 0;
1103 	to->p = current;
1104 }
1105 
1106 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1107 			       const struct timespec *ts)
1108 {
1109 	to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1110 	if (time_after(to->timer.expires, jiffies))
1111 		add_timer(&to->timer);
1112 	else
1113 		to->timed_out = 1;
1114 }
1115 
1116 static inline void clear_timeout(struct aio_timeout *to)
1117 {
1118 	del_singleshot_timer_sync(&to->timer);
1119 }
1120 
1121 static int read_events(struct kioctx *ctx,
1122 			long min_nr, long nr,
1123 			struct io_event __user *event,
1124 			struct timespec __user *timeout)
1125 {
1126 	long			start_jiffies = jiffies;
1127 	struct task_struct	*tsk = current;
1128 	DECLARE_WAITQUEUE(wait, tsk);
1129 	int			ret;
1130 	int			i = 0;
1131 	struct io_event		ent;
1132 	struct aio_timeout	to;
1133 	int			retry = 0;
1134 
1135 	/* needed to zero any padding within an entry (there shouldn't be
1136 	 * any, but C is fun!
1137 	 */
1138 	memset(&ent, 0, sizeof(ent));
1139 retry:
1140 	ret = 0;
1141 	while (likely(i < nr)) {
1142 		ret = aio_read_evt(ctx, &ent);
1143 		if (unlikely(ret <= 0))
1144 			break;
1145 
1146 		dprintk("read event: %Lx %Lx %Lx %Lx\n",
1147 			ent.data, ent.obj, ent.res, ent.res2);
1148 
1149 		/* Could we split the check in two? */
1150 		ret = -EFAULT;
1151 		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1152 			dprintk("aio: lost an event due to EFAULT.\n");
1153 			break;
1154 		}
1155 		ret = 0;
1156 
1157 		/* Good, event copied to userland, update counts. */
1158 		event ++;
1159 		i ++;
1160 	}
1161 
1162 	if (min_nr <= i)
1163 		return i;
1164 	if (ret)
1165 		return ret;
1166 
1167 	/* End fast path */
1168 
1169 	/* racey check, but it gets redone */
1170 	if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1171 		retry = 1;
1172 		aio_run_all_iocbs(ctx);
1173 		goto retry;
1174 	}
1175 
1176 	init_timeout(&to);
1177 	if (timeout) {
1178 		struct timespec	ts;
1179 		ret = -EFAULT;
1180 		if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1181 			goto out;
1182 
1183 		set_timeout(start_jiffies, &to, &ts);
1184 	}
1185 
1186 	while (likely(i < nr)) {
1187 		add_wait_queue_exclusive(&ctx->wait, &wait);
1188 		do {
1189 			set_task_state(tsk, TASK_INTERRUPTIBLE);
1190 			ret = aio_read_evt(ctx, &ent);
1191 			if (ret)
1192 				break;
1193 			if (min_nr <= i)
1194 				break;
1195 			ret = 0;
1196 			if (to.timed_out)	/* Only check after read evt */
1197 				break;
1198 			schedule();
1199 			if (signal_pending(tsk)) {
1200 				ret = -EINTR;
1201 				break;
1202 			}
1203 			/*ret = aio_read_evt(ctx, &ent);*/
1204 		} while (1) ;
1205 
1206 		set_task_state(tsk, TASK_RUNNING);
1207 		remove_wait_queue(&ctx->wait, &wait);
1208 
1209 		if (unlikely(ret <= 0))
1210 			break;
1211 
1212 		ret = -EFAULT;
1213 		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1214 			dprintk("aio: lost an event due to EFAULT.\n");
1215 			break;
1216 		}
1217 
1218 		/* Good, event copied to userland, update counts. */
1219 		event ++;
1220 		i ++;
1221 	}
1222 
1223 	if (timeout)
1224 		clear_timeout(&to);
1225 out:
1226 	return i ? i : ret;
1227 }
1228 
1229 /* Take an ioctx and remove it from the list of ioctx's.  Protects
1230  * against races with itself via ->dead.
1231  */
1232 static void io_destroy(struct kioctx *ioctx)
1233 {
1234 	struct mm_struct *mm = current->mm;
1235 	struct kioctx **tmp;
1236 	int was_dead;
1237 
1238 	/* delete the entry from the list is someone else hasn't already */
1239 	write_lock(&mm->ioctx_list_lock);
1240 	was_dead = ioctx->dead;
1241 	ioctx->dead = 1;
1242 	for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1243 	     tmp = &(*tmp)->next)
1244 		;
1245 	if (*tmp)
1246 		*tmp = ioctx->next;
1247 	write_unlock(&mm->ioctx_list_lock);
1248 
1249 	dprintk("aio_release(%p)\n", ioctx);
1250 	if (likely(!was_dead))
1251 		put_ioctx(ioctx);	/* twice for the list */
1252 
1253 	aio_cancel_all(ioctx);
1254 	wait_for_all_aios(ioctx);
1255 	put_ioctx(ioctx);	/* once for the lookup */
1256 }
1257 
1258 /* sys_io_setup:
1259  *	Create an aio_context capable of receiving at least nr_events.
1260  *	ctxp must not point to an aio_context that already exists, and
1261  *	must be initialized to 0 prior to the call.  On successful
1262  *	creation of the aio_context, *ctxp is filled in with the resulting
1263  *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1264  *	if the specified nr_events exceeds internal limits.  May fail
1265  *	with -EAGAIN if the specified nr_events exceeds the user's limit
1266  *	of available events.  May fail with -ENOMEM if insufficient kernel
1267  *	resources are available.  May fail with -EFAULT if an invalid
1268  *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1269  *	implemented.
1270  */
1271 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1272 {
1273 	struct kioctx *ioctx = NULL;
1274 	unsigned long ctx;
1275 	long ret;
1276 
1277 	ret = get_user(ctx, ctxp);
1278 	if (unlikely(ret))
1279 		goto out;
1280 
1281 	ret = -EINVAL;
1282 	if (unlikely(ctx || (int)nr_events <= 0)) {
1283 		pr_debug("EINVAL: io_setup: ctx or nr_events > max\n");
1284 		goto out;
1285 	}
1286 
1287 	ioctx = ioctx_alloc(nr_events);
1288 	ret = PTR_ERR(ioctx);
1289 	if (!IS_ERR(ioctx)) {
1290 		ret = put_user(ioctx->user_id, ctxp);
1291 		if (!ret)
1292 			return 0;
1293 
1294 		get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1295 		io_destroy(ioctx);
1296 	}
1297 
1298 out:
1299 	return ret;
1300 }
1301 
1302 /* sys_io_destroy:
1303  *	Destroy the aio_context specified.  May cancel any outstanding
1304  *	AIOs and block on completion.  Will fail with -ENOSYS if not
1305  *	implemented.  May fail with -EFAULT if the context pointed to
1306  *	is invalid.
1307  */
1308 asmlinkage long sys_io_destroy(aio_context_t ctx)
1309 {
1310 	struct kioctx *ioctx = lookup_ioctx(ctx);
1311 	if (likely(NULL != ioctx)) {
1312 		io_destroy(ioctx);
1313 		return 0;
1314 	}
1315 	pr_debug("EINVAL: io_destroy: invalid context id\n");
1316 	return -EINVAL;
1317 }
1318 
1319 /*
1320  * aio_p{read,write} are the default  ki_retry methods for
1321  * IO_CMD_P{READ,WRITE}.  They maintains kiocb retry state around potentially
1322  * multiple calls to f_op->aio_read().  They loop around partial progress
1323  * instead of returning -EIOCBRETRY because they don't have the means to call
1324  * kick_iocb().
1325  */
1326 static ssize_t aio_pread(struct kiocb *iocb)
1327 {
1328 	struct file *file = iocb->ki_filp;
1329 	struct address_space *mapping = file->f_mapping;
1330 	struct inode *inode = mapping->host;
1331 	ssize_t ret = 0;
1332 
1333 	do {
1334 		ret = file->f_op->aio_read(iocb, iocb->ki_buf,
1335 			iocb->ki_left, iocb->ki_pos);
1336 		/*
1337 		 * Can't just depend on iocb->ki_left to determine
1338 		 * whether we are done. This may have been a short read.
1339 		 */
1340 		if (ret > 0) {
1341 			iocb->ki_buf += ret;
1342 			iocb->ki_left -= ret;
1343 		}
1344 
1345 		/*
1346 		 * For pipes and sockets we return once we have some data; for
1347 		 * regular files we retry till we complete the entire read or
1348 		 * find that we can't read any more data (e.g short reads).
1349 		 */
1350 	} while (ret > 0 && iocb->ki_left > 0 &&
1351 		 !S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode));
1352 
1353 	/* This means we must have transferred all that we could */
1354 	/* No need to retry anymore */
1355 	if ((ret == 0) || (iocb->ki_left == 0))
1356 		ret = iocb->ki_nbytes - iocb->ki_left;
1357 
1358 	return ret;
1359 }
1360 
1361 /* see aio_pread() */
1362 static ssize_t aio_pwrite(struct kiocb *iocb)
1363 {
1364 	struct file *file = iocb->ki_filp;
1365 	ssize_t ret = 0;
1366 
1367 	do {
1368 		ret = file->f_op->aio_write(iocb, iocb->ki_buf,
1369 			iocb->ki_left, iocb->ki_pos);
1370 		if (ret > 0) {
1371 			iocb->ki_buf += ret;
1372 			iocb->ki_left -= ret;
1373 		}
1374 	} while (ret > 0 && iocb->ki_left > 0);
1375 
1376 	if ((ret == 0) || (iocb->ki_left == 0))
1377 		ret = iocb->ki_nbytes - iocb->ki_left;
1378 
1379 	return ret;
1380 }
1381 
1382 static ssize_t aio_fdsync(struct kiocb *iocb)
1383 {
1384 	struct file *file = iocb->ki_filp;
1385 	ssize_t ret = -EINVAL;
1386 
1387 	if (file->f_op->aio_fsync)
1388 		ret = file->f_op->aio_fsync(iocb, 1);
1389 	return ret;
1390 }
1391 
1392 static ssize_t aio_fsync(struct kiocb *iocb)
1393 {
1394 	struct file *file = iocb->ki_filp;
1395 	ssize_t ret = -EINVAL;
1396 
1397 	if (file->f_op->aio_fsync)
1398 		ret = file->f_op->aio_fsync(iocb, 0);
1399 	return ret;
1400 }
1401 
1402 /*
1403  * aio_setup_iocb:
1404  *	Performs the initial checks and aio retry method
1405  *	setup for the kiocb at the time of io submission.
1406  */
1407 static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1408 {
1409 	struct file *file = kiocb->ki_filp;
1410 	ssize_t ret = 0;
1411 
1412 	switch (kiocb->ki_opcode) {
1413 	case IOCB_CMD_PREAD:
1414 		ret = -EBADF;
1415 		if (unlikely(!(file->f_mode & FMODE_READ)))
1416 			break;
1417 		ret = -EFAULT;
1418 		if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1419 			kiocb->ki_left)))
1420 			break;
1421 		ret = -EINVAL;
1422 		if (file->f_op->aio_read)
1423 			kiocb->ki_retry = aio_pread;
1424 		break;
1425 	case IOCB_CMD_PWRITE:
1426 		ret = -EBADF;
1427 		if (unlikely(!(file->f_mode & FMODE_WRITE)))
1428 			break;
1429 		ret = -EFAULT;
1430 		if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1431 			kiocb->ki_left)))
1432 			break;
1433 		ret = -EINVAL;
1434 		if (file->f_op->aio_write)
1435 			kiocb->ki_retry = aio_pwrite;
1436 		break;
1437 	case IOCB_CMD_FDSYNC:
1438 		ret = -EINVAL;
1439 		if (file->f_op->aio_fsync)
1440 			kiocb->ki_retry = aio_fdsync;
1441 		break;
1442 	case IOCB_CMD_FSYNC:
1443 		ret = -EINVAL;
1444 		if (file->f_op->aio_fsync)
1445 			kiocb->ki_retry = aio_fsync;
1446 		break;
1447 	default:
1448 		dprintk("EINVAL: io_submit: no operation provided\n");
1449 		ret = -EINVAL;
1450 	}
1451 
1452 	if (!kiocb->ki_retry)
1453 		return ret;
1454 
1455 	return 0;
1456 }
1457 
1458 /*
1459  * aio_wake_function:
1460  * 	wait queue callback function for aio notification,
1461  * 	Simply triggers a retry of the operation via kick_iocb.
1462  *
1463  * 	This callback is specified in the wait queue entry in
1464  *	a kiocb	(current->io_wait points to this wait queue
1465  *	entry when an aio operation executes; it is used
1466  * 	instead of a synchronous wait when an i/o blocking
1467  *	condition is encountered during aio).
1468  *
1469  * Note:
1470  * This routine is executed with the wait queue lock held.
1471  * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1472  * the ioctx lock inside the wait queue lock. This is safe
1473  * because this callback isn't used for wait queues which
1474  * are nested inside ioctx lock (i.e. ctx->wait)
1475  */
1476 static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1477 			     int sync, void *key)
1478 {
1479 	struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1480 
1481 	list_del_init(&wait->task_list);
1482 	kick_iocb(iocb);
1483 	return 1;
1484 }
1485 
1486 int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1487 			 struct iocb *iocb)
1488 {
1489 	struct kiocb *req;
1490 	struct file *file;
1491 	ssize_t ret;
1492 
1493 	/* enforce forwards compatibility on users */
1494 	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
1495 		     iocb->aio_reserved3)) {
1496 		pr_debug("EINVAL: io_submit: reserve field set\n");
1497 		return -EINVAL;
1498 	}
1499 
1500 	/* prevent overflows */
1501 	if (unlikely(
1502 	    (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1503 	    (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1504 	    ((ssize_t)iocb->aio_nbytes < 0)
1505 	   )) {
1506 		pr_debug("EINVAL: io_submit: overflow check\n");
1507 		return -EINVAL;
1508 	}
1509 
1510 	file = fget(iocb->aio_fildes);
1511 	if (unlikely(!file))
1512 		return -EBADF;
1513 
1514 	req = aio_get_req(ctx);		/* returns with 2 references to req */
1515 	if (unlikely(!req)) {
1516 		fput(file);
1517 		return -EAGAIN;
1518 	}
1519 
1520 	req->ki_filp = file;
1521 	ret = put_user(req->ki_key, &user_iocb->aio_key);
1522 	if (unlikely(ret)) {
1523 		dprintk("EFAULT: aio_key\n");
1524 		goto out_put_req;
1525 	}
1526 
1527 	req->ki_obj.user = user_iocb;
1528 	req->ki_user_data = iocb->aio_data;
1529 	req->ki_pos = iocb->aio_offset;
1530 
1531 	req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1532 	req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1533 	req->ki_opcode = iocb->aio_lio_opcode;
1534 	init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1535 	INIT_LIST_HEAD(&req->ki_wait.task_list);
1536 	req->ki_retried = 0;
1537 
1538 	ret = aio_setup_iocb(req);
1539 
1540 	if (ret)
1541 		goto out_put_req;
1542 
1543 	spin_lock_irq(&ctx->ctx_lock);
1544 	aio_run_iocb(req);
1545 	unlock_kiocb(req);
1546 	if (!list_empty(&ctx->run_list)) {
1547 		/* drain the run list */
1548 		while (__aio_run_iocbs(ctx))
1549 			;
1550 	}
1551 	spin_unlock_irq(&ctx->ctx_lock);
1552 	aio_put_req(req);	/* drop extra ref to req */
1553 	return 0;
1554 
1555 out_put_req:
1556 	aio_put_req(req);	/* drop extra ref to req */
1557 	aio_put_req(req);	/* drop i/o ref to req */
1558 	return ret;
1559 }
1560 
1561 /* sys_io_submit:
1562  *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1563  *	the number of iocbs queued.  May return -EINVAL if the aio_context
1564  *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
1565  *	*iocbpp[0] is not properly initialized, if the operation specified
1566  *	is invalid for the file descriptor in the iocb.  May fail with
1567  *	-EFAULT if any of the data structures point to invalid data.  May
1568  *	fail with -EBADF if the file descriptor specified in the first
1569  *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
1570  *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1571  *	fail with -ENOSYS if not implemented.
1572  */
1573 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1574 			      struct iocb __user * __user *iocbpp)
1575 {
1576 	struct kioctx *ctx;
1577 	long ret = 0;
1578 	int i;
1579 
1580 	if (unlikely(nr < 0))
1581 		return -EINVAL;
1582 
1583 	if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1584 		return -EFAULT;
1585 
1586 	ctx = lookup_ioctx(ctx_id);
1587 	if (unlikely(!ctx)) {
1588 		pr_debug("EINVAL: io_submit: invalid context id\n");
1589 		return -EINVAL;
1590 	}
1591 
1592 	/*
1593 	 * AKPM: should this return a partial result if some of the IOs were
1594 	 * successfully submitted?
1595 	 */
1596 	for (i=0; i<nr; i++) {
1597 		struct iocb __user *user_iocb;
1598 		struct iocb tmp;
1599 
1600 		if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1601 			ret = -EFAULT;
1602 			break;
1603 		}
1604 
1605 		if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1606 			ret = -EFAULT;
1607 			break;
1608 		}
1609 
1610 		ret = io_submit_one(ctx, user_iocb, &tmp);
1611 		if (ret)
1612 			break;
1613 	}
1614 
1615 	put_ioctx(ctx);
1616 	return i ? i : ret;
1617 }
1618 
1619 /* lookup_kiocb
1620  *	Finds a given iocb for cancellation.
1621  *	MUST be called with ctx->ctx_lock held.
1622  */
1623 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1624 				  u32 key)
1625 {
1626 	struct list_head *pos;
1627 	/* TODO: use a hash or array, this sucks. */
1628 	list_for_each(pos, &ctx->active_reqs) {
1629 		struct kiocb *kiocb = list_kiocb(pos);
1630 		if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1631 			return kiocb;
1632 	}
1633 	return NULL;
1634 }
1635 
1636 /* sys_io_cancel:
1637  *	Attempts to cancel an iocb previously passed to io_submit.  If
1638  *	the operation is successfully cancelled, the resulting event is
1639  *	copied into the memory pointed to by result without being placed
1640  *	into the completion queue and 0 is returned.  May fail with
1641  *	-EFAULT if any of the data structures pointed to are invalid.
1642  *	May fail with -EINVAL if aio_context specified by ctx_id is
1643  *	invalid.  May fail with -EAGAIN if the iocb specified was not
1644  *	cancelled.  Will fail with -ENOSYS if not implemented.
1645  */
1646 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1647 			      struct io_event __user *result)
1648 {
1649 	int (*cancel)(struct kiocb *iocb, struct io_event *res);
1650 	struct kioctx *ctx;
1651 	struct kiocb *kiocb;
1652 	u32 key;
1653 	int ret;
1654 
1655 	ret = get_user(key, &iocb->aio_key);
1656 	if (unlikely(ret))
1657 		return -EFAULT;
1658 
1659 	ctx = lookup_ioctx(ctx_id);
1660 	if (unlikely(!ctx))
1661 		return -EINVAL;
1662 
1663 	spin_lock_irq(&ctx->ctx_lock);
1664 	ret = -EAGAIN;
1665 	kiocb = lookup_kiocb(ctx, iocb, key);
1666 	if (kiocb && kiocb->ki_cancel) {
1667 		cancel = kiocb->ki_cancel;
1668 		kiocb->ki_users ++;
1669 		kiocbSetCancelled(kiocb);
1670 	} else
1671 		cancel = NULL;
1672 	spin_unlock_irq(&ctx->ctx_lock);
1673 
1674 	if (NULL != cancel) {
1675 		struct io_event tmp;
1676 		pr_debug("calling cancel\n");
1677 		lock_kiocb(kiocb);
1678 		memset(&tmp, 0, sizeof(tmp));
1679 		tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1680 		tmp.data = kiocb->ki_user_data;
1681 		ret = cancel(kiocb, &tmp);
1682 		if (!ret) {
1683 			/* Cancellation succeeded -- copy the result
1684 			 * into the user's buffer.
1685 			 */
1686 			if (copy_to_user(result, &tmp, sizeof(tmp)))
1687 				ret = -EFAULT;
1688 		}
1689 		unlock_kiocb(kiocb);
1690 	} else
1691 		ret = -EINVAL;
1692 
1693 	put_ioctx(ctx);
1694 
1695 	return ret;
1696 }
1697 
1698 /* io_getevents:
1699  *	Attempts to read at least min_nr events and up to nr events from
1700  *	the completion queue for the aio_context specified by ctx_id.  May
1701  *	fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1702  *	if nr is out of range, if when is out of range.  May fail with
1703  *	-EFAULT if any of the memory specified to is invalid.  May return
1704  *	0 or < min_nr if no events are available and the timeout specified
1705  *	by when	has elapsed, where when == NULL specifies an infinite
1706  *	timeout.  Note that the timeout pointed to by when is relative and
1707  *	will be updated if not NULL and the operation blocks.  Will fail
1708  *	with -ENOSYS if not implemented.
1709  */
1710 asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1711 				 long min_nr,
1712 				 long nr,
1713 				 struct io_event __user *events,
1714 				 struct timespec __user *timeout)
1715 {
1716 	struct kioctx *ioctx = lookup_ioctx(ctx_id);
1717 	long ret = -EINVAL;
1718 
1719 	if (likely(ioctx)) {
1720 		if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1721 			ret = read_events(ioctx, min_nr, nr, events, timeout);
1722 		put_ioctx(ioctx);
1723 	}
1724 
1725 	return ret;
1726 }
1727 
1728 __initcall(aio_setup);
1729 
1730 EXPORT_SYMBOL(aio_complete);
1731 EXPORT_SYMBOL(aio_put_req);
1732 EXPORT_SYMBOL(wait_on_sync_kiocb);
1733