xref: /linux/fs/aio.c (revision d67b569f5f620c0fb95d5212642746b7ba9d29e4)
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 
33 #include <asm/kmap_types.h>
34 #include <asm/uaccess.h>
35 #include <asm/mmu_context.h>
36 
37 #if DEBUG > 1
38 #define dprintk		printk
39 #else
40 #define dprintk(x...)	do { ; } while (0)
41 #endif
42 
43 /*------ sysctl variables----*/
44 atomic_t aio_nr = ATOMIC_INIT(0);	/* current system wide number of aio requests */
45 unsigned aio_max_nr = 0x10000;	/* system wide maximum number of aio requests */
46 /*----end sysctl variables---*/
47 
48 static kmem_cache_t	*kiocb_cachep;
49 static kmem_cache_t	*kioctx_cachep;
50 
51 static struct workqueue_struct *aio_wq;
52 
53 /* Used for rare fput completion. */
54 static void aio_fput_routine(void *);
55 static DECLARE_WORK(fput_work, aio_fput_routine, NULL);
56 
57 static DEFINE_SPINLOCK(fput_lock);
58 static LIST_HEAD(fput_head);
59 
60 static void aio_kick_handler(void *);
61 static void aio_queue_work(struct kioctx *);
62 
63 /* aio_setup
64  *	Creates the slab caches used by the aio routines, panic on
65  *	failure as this is done early during the boot sequence.
66  */
67 static int __init aio_setup(void)
68 {
69 	kiocb_cachep = kmem_cache_create("kiocb", sizeof(struct kiocb),
70 				0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
71 	kioctx_cachep = kmem_cache_create("kioctx", sizeof(struct kioctx),
72 				0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
73 
74 	aio_wq = create_workqueue("aio");
75 
76 	pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
77 
78 	return 0;
79 }
80 
81 static void aio_free_ring(struct kioctx *ctx)
82 {
83 	struct aio_ring_info *info = &ctx->ring_info;
84 	long i;
85 
86 	for (i=0; i<info->nr_pages; i++)
87 		put_page(info->ring_pages[i]);
88 
89 	if (info->mmap_size) {
90 		down_write(&ctx->mm->mmap_sem);
91 		do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
92 		up_write(&ctx->mm->mmap_sem);
93 	}
94 
95 	if (info->ring_pages && info->ring_pages != info->internal_pages)
96 		kfree(info->ring_pages);
97 	info->ring_pages = NULL;
98 	info->nr = 0;
99 }
100 
101 static int aio_setup_ring(struct kioctx *ctx)
102 {
103 	struct aio_ring *ring;
104 	struct aio_ring_info *info = &ctx->ring_info;
105 	unsigned nr_events = ctx->max_reqs;
106 	unsigned long size;
107 	int nr_pages;
108 
109 	/* Compensate for the ring buffer's head/tail overlap entry */
110 	nr_events += 2;	/* 1 is required, 2 for good luck */
111 
112 	size = sizeof(struct aio_ring);
113 	size += sizeof(struct io_event) * nr_events;
114 	nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
115 
116 	if (nr_pages < 0)
117 		return -EINVAL;
118 
119 	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
120 
121 	info->nr = 0;
122 	info->ring_pages = info->internal_pages;
123 	if (nr_pages > AIO_RING_PAGES) {
124 		info->ring_pages = kmalloc(sizeof(struct page *) * nr_pages, GFP_KERNEL);
125 		if (!info->ring_pages)
126 			return -ENOMEM;
127 		memset(info->ring_pages, 0, sizeof(struct page *) * nr_pages);
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_ANON|MAP_PRIVATE,
135 				  0);
136 	if (IS_ERR((void *)info->mmap_base)) {
137 		up_write(&ctx->mm->mmap_sem);
138 		printk("mmap err: %ld\n", -info->mmap_base);
139 		info->mmap_size = 0;
140 		aio_free_ring(ctx);
141 		return -EAGAIN;
142 	}
143 
144 	dprintk("mmap address: 0x%08lx\n", info->mmap_base);
145 	info->nr_pages = get_user_pages(current, ctx->mm,
146 					info->mmap_base, nr_pages,
147 					1, 0, info->ring_pages, NULL);
148 	up_write(&ctx->mm->mmap_sem);
149 
150 	if (unlikely(info->nr_pages != nr_pages)) {
151 		aio_free_ring(ctx);
152 		return -EAGAIN;
153 	}
154 
155 	ctx->user_id = info->mmap_base;
156 
157 	info->nr = nr_events;		/* trusted copy */
158 
159 	ring = kmap_atomic(info->ring_pages[0], KM_USER0);
160 	ring->nr = nr_events;	/* user copy */
161 	ring->id = ctx->user_id;
162 	ring->head = ring->tail = 0;
163 	ring->magic = AIO_RING_MAGIC;
164 	ring->compat_features = AIO_RING_COMPAT_FEATURES;
165 	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
166 	ring->header_length = sizeof(struct aio_ring);
167 	kunmap_atomic(ring, KM_USER0);
168 
169 	return 0;
170 }
171 
172 
173 /* aio_ring_event: returns a pointer to the event at the given index from
174  * kmap_atomic(, km).  Release the pointer with put_aio_ring_event();
175  */
176 #define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
177 #define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
178 #define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
179 
180 #define aio_ring_event(info, nr, km) ({					\
181 	unsigned pos = (nr) + AIO_EVENTS_OFFSET;			\
182 	struct io_event *__event;					\
183 	__event = kmap_atomic(						\
184 			(info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
185 	__event += pos % AIO_EVENTS_PER_PAGE;				\
186 	__event;							\
187 })
188 
189 #define put_aio_ring_event(event, km) do {	\
190 	struct io_event *__event = (event);	\
191 	(void)__event;				\
192 	kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
193 } while(0)
194 
195 /* ioctx_alloc
196  *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
197  */
198 static struct kioctx *ioctx_alloc(unsigned nr_events)
199 {
200 	struct mm_struct *mm;
201 	struct kioctx *ctx;
202 
203 	/* Prevent overflows */
204 	if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
205 	    (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
206 		pr_debug("ENOMEM: nr_events too high\n");
207 		return ERR_PTR(-EINVAL);
208 	}
209 
210 	if (nr_events > aio_max_nr)
211 		return ERR_PTR(-EAGAIN);
212 
213 	ctx = kmem_cache_alloc(kioctx_cachep, GFP_KERNEL);
214 	if (!ctx)
215 		return ERR_PTR(-ENOMEM);
216 
217 	memset(ctx, 0, sizeof(*ctx));
218 	ctx->max_reqs = nr_events;
219 	mm = ctx->mm = current->mm;
220 	atomic_inc(&mm->mm_count);
221 
222 	atomic_set(&ctx->users, 1);
223 	spin_lock_init(&ctx->ctx_lock);
224 	spin_lock_init(&ctx->ring_info.ring_lock);
225 	init_waitqueue_head(&ctx->wait);
226 
227 	INIT_LIST_HEAD(&ctx->active_reqs);
228 	INIT_LIST_HEAD(&ctx->run_list);
229 	INIT_WORK(&ctx->wq, aio_kick_handler, ctx);
230 
231 	if (aio_setup_ring(ctx) < 0)
232 		goto out_freectx;
233 
234 	/* limit the number of system wide aios */
235 	atomic_add(ctx->max_reqs, &aio_nr);	/* undone by __put_ioctx */
236 	if (unlikely(atomic_read(&aio_nr) > aio_max_nr))
237 		goto out_cleanup;
238 
239 	/* now link into global list.  kludge.  FIXME */
240 	write_lock(&mm->ioctx_list_lock);
241 	ctx->next = mm->ioctx_list;
242 	mm->ioctx_list = ctx;
243 	write_unlock(&mm->ioctx_list_lock);
244 
245 	dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
246 		ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
247 	return ctx;
248 
249 out_cleanup:
250 	atomic_sub(ctx->max_reqs, &aio_nr);
251 	ctx->max_reqs = 0;	/* prevent __put_ioctx from sub'ing aio_nr */
252 	__put_ioctx(ctx);
253 	return ERR_PTR(-EAGAIN);
254 
255 out_freectx:
256 	mmdrop(mm);
257 	kmem_cache_free(kioctx_cachep, ctx);
258 	ctx = ERR_PTR(-ENOMEM);
259 
260 	dprintk("aio: error allocating ioctx %p\n", ctx);
261 	return ctx;
262 }
263 
264 /* aio_cancel_all
265  *	Cancels all outstanding aio requests on an aio context.  Used
266  *	when the processes owning a context have all exited to encourage
267  *	the rapid destruction of the kioctx.
268  */
269 static void aio_cancel_all(struct kioctx *ctx)
270 {
271 	int (*cancel)(struct kiocb *, struct io_event *);
272 	struct io_event res;
273 	spin_lock_irq(&ctx->ctx_lock);
274 	ctx->dead = 1;
275 	while (!list_empty(&ctx->active_reqs)) {
276 		struct list_head *pos = ctx->active_reqs.next;
277 		struct kiocb *iocb = list_kiocb(pos);
278 		list_del_init(&iocb->ki_list);
279 		cancel = iocb->ki_cancel;
280 		kiocbSetCancelled(iocb);
281 		if (cancel) {
282 			iocb->ki_users++;
283 			spin_unlock_irq(&ctx->ctx_lock);
284 			cancel(iocb, &res);
285 			spin_lock_irq(&ctx->ctx_lock);
286 		}
287 	}
288 	spin_unlock_irq(&ctx->ctx_lock);
289 }
290 
291 static void wait_for_all_aios(struct kioctx *ctx)
292 {
293 	struct task_struct *tsk = current;
294 	DECLARE_WAITQUEUE(wait, tsk);
295 
296 	if (!ctx->reqs_active)
297 		return;
298 
299 	add_wait_queue(&ctx->wait, &wait);
300 	set_task_state(tsk, TASK_UNINTERRUPTIBLE);
301 	while (ctx->reqs_active) {
302 		schedule();
303 		set_task_state(tsk, TASK_UNINTERRUPTIBLE);
304 	}
305 	__set_task_state(tsk, TASK_RUNNING);
306 	remove_wait_queue(&ctx->wait, &wait);
307 }
308 
309 /* wait_on_sync_kiocb:
310  *	Waits on the given sync kiocb to complete.
311  */
312 ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb)
313 {
314 	while (iocb->ki_users) {
315 		set_current_state(TASK_UNINTERRUPTIBLE);
316 		if (!iocb->ki_users)
317 			break;
318 		schedule();
319 	}
320 	__set_current_state(TASK_RUNNING);
321 	return iocb->ki_user_data;
322 }
323 
324 /* exit_aio: called when the last user of mm goes away.  At this point,
325  * there is no way for any new requests to be submited or any of the
326  * io_* syscalls to be called on the context.  However, there may be
327  * outstanding requests which hold references to the context; as they
328  * go away, they will call put_ioctx and release any pinned memory
329  * associated with the request (held via struct page * references).
330  */
331 void fastcall exit_aio(struct mm_struct *mm)
332 {
333 	struct kioctx *ctx = mm->ioctx_list;
334 	mm->ioctx_list = NULL;
335 	while (ctx) {
336 		struct kioctx *next = ctx->next;
337 		ctx->next = NULL;
338 		aio_cancel_all(ctx);
339 
340 		wait_for_all_aios(ctx);
341 		/*
342 		 * this is an overkill, but ensures we don't leave
343 		 * the ctx on the aio_wq
344 		 */
345 		flush_workqueue(aio_wq);
346 
347 		if (1 != atomic_read(&ctx->users))
348 			printk(KERN_DEBUG
349 				"exit_aio:ioctx still alive: %d %d %d\n",
350 				atomic_read(&ctx->users), ctx->dead,
351 				ctx->reqs_active);
352 		put_ioctx(ctx);
353 		ctx = next;
354 	}
355 }
356 
357 /* __put_ioctx
358  *	Called when the last user of an aio context has gone away,
359  *	and the struct needs to be freed.
360  */
361 void fastcall __put_ioctx(struct kioctx *ctx)
362 {
363 	unsigned nr_events = ctx->max_reqs;
364 
365 	if (unlikely(ctx->reqs_active))
366 		BUG();
367 
368 	cancel_delayed_work(&ctx->wq);
369 	flush_workqueue(aio_wq);
370 	aio_free_ring(ctx);
371 	mmdrop(ctx->mm);
372 	ctx->mm = NULL;
373 	pr_debug("__put_ioctx: freeing %p\n", ctx);
374 	kmem_cache_free(kioctx_cachep, ctx);
375 
376 	atomic_sub(nr_events, &aio_nr);
377 }
378 
379 /* aio_get_req
380  *	Allocate a slot for an aio request.  Increments the users count
381  * of the kioctx so that the kioctx stays around until all requests are
382  * complete.  Returns NULL if no requests are free.
383  *
384  * Returns with kiocb->users set to 2.  The io submit code path holds
385  * an extra reference while submitting the i/o.
386  * This prevents races between the aio code path referencing the
387  * req (after submitting it) and aio_complete() freeing the req.
388  */
389 static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx));
390 static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx)
391 {
392 	struct kiocb *req = NULL;
393 	struct aio_ring *ring;
394 	int okay = 0;
395 
396 	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
397 	if (unlikely(!req))
398 		return NULL;
399 
400 	req->ki_flags = 1 << KIF_LOCKED;
401 	req->ki_users = 2;
402 	req->ki_key = 0;
403 	req->ki_ctx = ctx;
404 	req->ki_cancel = NULL;
405 	req->ki_retry = NULL;
406 	req->ki_dtor = NULL;
407 	req->private = NULL;
408 	INIT_LIST_HEAD(&req->ki_run_list);
409 
410 	/* Check if the completion queue has enough free space to
411 	 * accept an event from this io.
412 	 */
413 	spin_lock_irq(&ctx->ctx_lock);
414 	ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
415 	if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
416 		list_add(&req->ki_list, &ctx->active_reqs);
417 		get_ioctx(ctx);
418 		ctx->reqs_active++;
419 		okay = 1;
420 	}
421 	kunmap_atomic(ring, KM_USER0);
422 	spin_unlock_irq(&ctx->ctx_lock);
423 
424 	if (!okay) {
425 		kmem_cache_free(kiocb_cachep, req);
426 		req = NULL;
427 	}
428 
429 	return req;
430 }
431 
432 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
433 {
434 	struct kiocb *req;
435 	/* Handle a potential starvation case -- should be exceedingly rare as
436 	 * requests will be stuck on fput_head only if the aio_fput_routine is
437 	 * delayed and the requests were the last user of the struct file.
438 	 */
439 	req = __aio_get_req(ctx);
440 	if (unlikely(NULL == req)) {
441 		aio_fput_routine(NULL);
442 		req = __aio_get_req(ctx);
443 	}
444 	return req;
445 }
446 
447 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
448 {
449 	if (req->ki_dtor)
450 		req->ki_dtor(req);
451 	kmem_cache_free(kiocb_cachep, req);
452 	ctx->reqs_active--;
453 
454 	if (unlikely(!ctx->reqs_active && ctx->dead))
455 		wake_up(&ctx->wait);
456 }
457 
458 static void aio_fput_routine(void *data)
459 {
460 	spin_lock_irq(&fput_lock);
461 	while (likely(!list_empty(&fput_head))) {
462 		struct kiocb *req = list_kiocb(fput_head.next);
463 		struct kioctx *ctx = req->ki_ctx;
464 
465 		list_del(&req->ki_list);
466 		spin_unlock_irq(&fput_lock);
467 
468 		/* Complete the fput */
469 		__fput(req->ki_filp);
470 
471 		/* Link the iocb into the context's free list */
472 		spin_lock_irq(&ctx->ctx_lock);
473 		really_put_req(ctx, req);
474 		spin_unlock_irq(&ctx->ctx_lock);
475 
476 		put_ioctx(ctx);
477 		spin_lock_irq(&fput_lock);
478 	}
479 	spin_unlock_irq(&fput_lock);
480 }
481 
482 /* __aio_put_req
483  *	Returns true if this put was the last user of the request.
484  */
485 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
486 {
487 	dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
488 		req, atomic_read(&req->ki_filp->f_count));
489 
490 	req->ki_users --;
491 	if (unlikely(req->ki_users < 0))
492 		BUG();
493 	if (likely(req->ki_users))
494 		return 0;
495 	list_del(&req->ki_list);		/* remove from active_reqs */
496 	req->ki_cancel = NULL;
497 	req->ki_retry = NULL;
498 
499 	/* Must be done under the lock to serialise against cancellation.
500 	 * Call this aio_fput as it duplicates fput via the fput_work.
501 	 */
502 	if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) {
503 		get_ioctx(ctx);
504 		spin_lock(&fput_lock);
505 		list_add(&req->ki_list, &fput_head);
506 		spin_unlock(&fput_lock);
507 		queue_work(aio_wq, &fput_work);
508 	} else
509 		really_put_req(ctx, req);
510 	return 1;
511 }
512 
513 /* aio_put_req
514  *	Returns true if this put was the last user of the kiocb,
515  *	false if the request is still in use.
516  */
517 int fastcall aio_put_req(struct kiocb *req)
518 {
519 	struct kioctx *ctx = req->ki_ctx;
520 	int ret;
521 	spin_lock_irq(&ctx->ctx_lock);
522 	ret = __aio_put_req(ctx, req);
523 	spin_unlock_irq(&ctx->ctx_lock);
524 	if (ret)
525 		put_ioctx(ctx);
526 	return ret;
527 }
528 
529 /*	Lookup an ioctx id.  ioctx_list is lockless for reads.
530  *	FIXME: this is O(n) and is only suitable for development.
531  */
532 struct kioctx *lookup_ioctx(unsigned long ctx_id)
533 {
534 	struct kioctx *ioctx;
535 	struct mm_struct *mm;
536 
537 	mm = current->mm;
538 	read_lock(&mm->ioctx_list_lock);
539 	for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
540 		if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
541 			get_ioctx(ioctx);
542 			break;
543 		}
544 	read_unlock(&mm->ioctx_list_lock);
545 
546 	return ioctx;
547 }
548 
549 /*
550  * use_mm
551  *	Makes the calling kernel thread take on the specified
552  *	mm context.
553  *	Called by the retry thread execute retries within the
554  *	iocb issuer's mm context, so that copy_from/to_user
555  *	operations work seamlessly for aio.
556  *	(Note: this routine is intended to be called only
557  *	from a kernel thread context)
558  */
559 static void use_mm(struct mm_struct *mm)
560 {
561 	struct mm_struct *active_mm;
562 	struct task_struct *tsk = current;
563 
564 	task_lock(tsk);
565 	tsk->flags |= PF_BORROWED_MM;
566 	active_mm = tsk->active_mm;
567 	atomic_inc(&mm->mm_count);
568 	tsk->mm = mm;
569 	tsk->active_mm = mm;
570 	activate_mm(active_mm, mm);
571 	task_unlock(tsk);
572 
573 	mmdrop(active_mm);
574 }
575 
576 /*
577  * unuse_mm
578  *	Reverses the effect of use_mm, i.e. releases the
579  *	specified mm context which was earlier taken on
580  *	by the calling kernel thread
581  *	(Note: this routine is intended to be called only
582  *	from a kernel thread context)
583  *
584  * Comments: Called with ctx->ctx_lock held. This nests
585  * task_lock instead ctx_lock.
586  */
587 static void unuse_mm(struct mm_struct *mm)
588 {
589 	struct task_struct *tsk = current;
590 
591 	task_lock(tsk);
592 	tsk->flags &= ~PF_BORROWED_MM;
593 	tsk->mm = NULL;
594 	/* active_mm is still 'mm' */
595 	enter_lazy_tlb(mm, tsk);
596 	task_unlock(tsk);
597 }
598 
599 /*
600  * Queue up a kiocb to be retried. Assumes that the kiocb
601  * has already been marked as kicked, and places it on
602  * the retry run list for the corresponding ioctx, if it
603  * isn't already queued. Returns 1 if it actually queued
604  * the kiocb (to tell the caller to activate the work
605  * queue to process it), or 0, if it found that it was
606  * already queued.
607  *
608  * Should be called with the spin lock iocb->ki_ctx->ctx_lock
609  * held
610  */
611 static inline int __queue_kicked_iocb(struct kiocb *iocb)
612 {
613 	struct kioctx *ctx = iocb->ki_ctx;
614 
615 	if (list_empty(&iocb->ki_run_list)) {
616 		list_add_tail(&iocb->ki_run_list,
617 			&ctx->run_list);
618 		return 1;
619 	}
620 	return 0;
621 }
622 
623 /* aio_run_iocb
624  *	This is the core aio execution routine. It is
625  *	invoked both for initial i/o submission and
626  *	subsequent retries via the aio_kick_handler.
627  *	Expects to be invoked with iocb->ki_ctx->lock
628  *	already held. The lock is released and reaquired
629  *	as needed during processing.
630  *
631  * Calls the iocb retry method (already setup for the
632  * iocb on initial submission) for operation specific
633  * handling, but takes care of most of common retry
634  * execution details for a given iocb. The retry method
635  * needs to be non-blocking as far as possible, to avoid
636  * holding up other iocbs waiting to be serviced by the
637  * retry kernel thread.
638  *
639  * The trickier parts in this code have to do with
640  * ensuring that only one retry instance is in progress
641  * for a given iocb at any time. Providing that guarantee
642  * simplifies the coding of individual aio operations as
643  * it avoids various potential races.
644  */
645 static ssize_t aio_run_iocb(struct kiocb *iocb)
646 {
647 	struct kioctx	*ctx = iocb->ki_ctx;
648 	ssize_t (*retry)(struct kiocb *);
649 	ssize_t ret;
650 
651 	if (iocb->ki_retried++ > 1024*1024) {
652 		printk("Maximal retry count.  Bytes done %Zd\n",
653 			iocb->ki_nbytes - iocb->ki_left);
654 		return -EAGAIN;
655 	}
656 
657 	if (!(iocb->ki_retried & 0xff)) {
658 		pr_debug("%ld retry: %d of %d\n", iocb->ki_retried,
659 			iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
660 	}
661 
662 	if (!(retry = iocb->ki_retry)) {
663 		printk("aio_run_iocb: iocb->ki_retry = NULL\n");
664 		return 0;
665 	}
666 
667 	/*
668 	 * We don't want the next retry iteration for this
669 	 * operation to start until this one has returned and
670 	 * updated the iocb state. However, wait_queue functions
671 	 * can trigger a kick_iocb from interrupt context in the
672 	 * meantime, indicating that data is available for the next
673 	 * iteration. We want to remember that and enable the
674 	 * next retry iteration _after_ we are through with
675 	 * this one.
676 	 *
677 	 * So, in order to be able to register a "kick", but
678 	 * prevent it from being queued now, we clear the kick
679 	 * flag, but make the kick code *think* that the iocb is
680 	 * still on the run list until we are actually done.
681 	 * When we are done with this iteration, we check if
682 	 * the iocb was kicked in the meantime and if so, queue
683 	 * it up afresh.
684 	 */
685 
686 	kiocbClearKicked(iocb);
687 
688 	/*
689 	 * This is so that aio_complete knows it doesn't need to
690 	 * pull the iocb off the run list (We can't just call
691 	 * INIT_LIST_HEAD because we don't want a kick_iocb to
692 	 * queue this on the run list yet)
693 	 */
694 	iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
695 	spin_unlock_irq(&ctx->ctx_lock);
696 
697 	/* Quit retrying if the i/o has been cancelled */
698 	if (kiocbIsCancelled(iocb)) {
699 		ret = -EINTR;
700 		aio_complete(iocb, ret, 0);
701 		/* must not access the iocb after this */
702 		goto out;
703 	}
704 
705 	/*
706 	 * Now we are all set to call the retry method in async
707 	 * context. By setting this thread's io_wait context
708 	 * to point to the wait queue entry inside the currently
709 	 * running iocb for the duration of the retry, we ensure
710 	 * that async notification wakeups are queued by the
711 	 * operation instead of blocking waits, and when notified,
712 	 * cause the iocb to be kicked for continuation (through
713 	 * the aio_wake_function callback).
714 	 */
715 	BUG_ON(current->io_wait != NULL);
716 	current->io_wait = &iocb->ki_wait;
717 	ret = retry(iocb);
718 	current->io_wait = NULL;
719 
720 	if (-EIOCBRETRY != ret) {
721  		if (-EIOCBQUEUED != ret) {
722 			BUG_ON(!list_empty(&iocb->ki_wait.task_list));
723 			aio_complete(iocb, ret, 0);
724 			/* must not access the iocb after this */
725 		}
726 	} else {
727 		/*
728 		 * Issue an additional retry to avoid waiting forever if
729 		 * no waits were queued (e.g. in case of a short read).
730 		 */
731 		if (list_empty(&iocb->ki_wait.task_list))
732 			kiocbSetKicked(iocb);
733 	}
734 out:
735 	spin_lock_irq(&ctx->ctx_lock);
736 
737 	if (-EIOCBRETRY == ret) {
738 		/*
739 		 * OK, now that we are done with this iteration
740 		 * and know that there is more left to go,
741 		 * this is where we let go so that a subsequent
742 		 * "kick" can start the next iteration
743 		 */
744 
745 		/* will make __queue_kicked_iocb succeed from here on */
746 		INIT_LIST_HEAD(&iocb->ki_run_list);
747 		/* we must queue the next iteration ourselves, if it
748 		 * has already been kicked */
749 		if (kiocbIsKicked(iocb)) {
750 			__queue_kicked_iocb(iocb);
751 
752 			/*
753 			 * __queue_kicked_iocb will always return 1 here, because
754 			 * iocb->ki_run_list is empty at this point so it should
755 			 * be safe to unconditionally queue the context into the
756 			 * work queue.
757 			 */
758 			aio_queue_work(ctx);
759 		}
760 	}
761 	return ret;
762 }
763 
764 /*
765  * __aio_run_iocbs:
766  * 	Process all pending retries queued on the ioctx
767  * 	run list.
768  * Assumes it is operating within the aio issuer's mm
769  * context. Expects to be called with ctx->ctx_lock held
770  */
771 static int __aio_run_iocbs(struct kioctx *ctx)
772 {
773 	struct kiocb *iocb;
774 	LIST_HEAD(run_list);
775 
776 	list_splice_init(&ctx->run_list, &run_list);
777 	while (!list_empty(&run_list)) {
778 		iocb = list_entry(run_list.next, struct kiocb,
779 			ki_run_list);
780 		list_del(&iocb->ki_run_list);
781 		/*
782 		 * Hold an extra reference while retrying i/o.
783 		 */
784 		iocb->ki_users++;       /* grab extra reference */
785 		aio_run_iocb(iocb);
786 		if (__aio_put_req(ctx, iocb))  /* drop extra ref */
787 			put_ioctx(ctx);
788  	}
789 	if (!list_empty(&ctx->run_list))
790 		return 1;
791 	return 0;
792 }
793 
794 static void aio_queue_work(struct kioctx * ctx)
795 {
796 	unsigned long timeout;
797 	/*
798 	 * if someone is waiting, get the work started right
799 	 * away, otherwise, use a longer delay
800 	 */
801 	smp_mb();
802 	if (waitqueue_active(&ctx->wait))
803 		timeout = 1;
804 	else
805 		timeout = HZ/10;
806 	queue_delayed_work(aio_wq, &ctx->wq, timeout);
807 }
808 
809 
810 /*
811  * aio_run_iocbs:
812  * 	Process all pending retries queued on the ioctx
813  * 	run list.
814  * Assumes it is operating within the aio issuer's mm
815  * context.
816  */
817 static inline void aio_run_iocbs(struct kioctx *ctx)
818 {
819 	int requeue;
820 
821 	spin_lock_irq(&ctx->ctx_lock);
822 
823 	requeue = __aio_run_iocbs(ctx);
824 	spin_unlock_irq(&ctx->ctx_lock);
825 	if (requeue)
826 		aio_queue_work(ctx);
827 }
828 
829 /*
830  * just like aio_run_iocbs, but keeps running them until
831  * the list stays empty
832  */
833 static inline void aio_run_all_iocbs(struct kioctx *ctx)
834 {
835 	spin_lock_irq(&ctx->ctx_lock);
836 	while (__aio_run_iocbs(ctx))
837 		;
838 	spin_unlock_irq(&ctx->ctx_lock);
839 }
840 
841 /*
842  * aio_kick_handler:
843  * 	Work queue handler triggered to process pending
844  * 	retries on an ioctx. Takes on the aio issuer's
845  *	mm context before running the iocbs, so that
846  *	copy_xxx_user operates on the issuer's address
847  *      space.
848  * Run on aiod's context.
849  */
850 static void aio_kick_handler(void *data)
851 {
852 	struct kioctx *ctx = data;
853 	mm_segment_t oldfs = get_fs();
854 	int requeue;
855 
856 	set_fs(USER_DS);
857 	use_mm(ctx->mm);
858 	spin_lock_irq(&ctx->ctx_lock);
859 	requeue =__aio_run_iocbs(ctx);
860  	unuse_mm(ctx->mm);
861 	spin_unlock_irq(&ctx->ctx_lock);
862 	set_fs(oldfs);
863 	/*
864 	 * we're in a worker thread already, don't use queue_delayed_work,
865 	 */
866 	if (requeue)
867 		queue_work(aio_wq, &ctx->wq);
868 }
869 
870 
871 /*
872  * Called by kick_iocb to queue the kiocb for retry
873  * and if required activate the aio work queue to process
874  * it
875  */
876 static void queue_kicked_iocb(struct kiocb *iocb)
877 {
878  	struct kioctx	*ctx = iocb->ki_ctx;
879 	unsigned long flags;
880 	int run = 0;
881 
882 	WARN_ON((!list_empty(&iocb->ki_wait.task_list)));
883 
884 	spin_lock_irqsave(&ctx->ctx_lock, flags);
885 	run = __queue_kicked_iocb(iocb);
886 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
887 	if (run)
888 		aio_queue_work(ctx);
889 }
890 
891 /*
892  * kick_iocb:
893  *      Called typically from a wait queue callback context
894  *      (aio_wake_function) to trigger a retry of the iocb.
895  *      The retry is usually executed by aio workqueue
896  *      threads (See aio_kick_handler).
897  */
898 void fastcall kick_iocb(struct kiocb *iocb)
899 {
900 	/* sync iocbs are easy: they can only ever be executing from a
901 	 * single context. */
902 	if (is_sync_kiocb(iocb)) {
903 		kiocbSetKicked(iocb);
904 	        wake_up_process(iocb->ki_obj.tsk);
905 		return;
906 	}
907 
908 	/* If its already kicked we shouldn't queue it again */
909 	if (!kiocbTryKick(iocb)) {
910 		queue_kicked_iocb(iocb);
911 	}
912 }
913 EXPORT_SYMBOL(kick_iocb);
914 
915 /* aio_complete
916  *	Called when the io request on the given iocb is complete.
917  *	Returns true if this is the last user of the request.  The
918  *	only other user of the request can be the cancellation code.
919  */
920 int fastcall aio_complete(struct kiocb *iocb, long res, long res2)
921 {
922 	struct kioctx	*ctx = iocb->ki_ctx;
923 	struct aio_ring_info	*info;
924 	struct aio_ring	*ring;
925 	struct io_event	*event;
926 	unsigned long	flags;
927 	unsigned long	tail;
928 	int		ret;
929 
930 	/* Special case handling for sync iocbs: events go directly
931 	 * into the iocb for fast handling.  Note that this will not
932 	 * work if we allow sync kiocbs to be cancelled. in which
933 	 * case the usage count checks will have to move under ctx_lock
934 	 * for all cases.
935 	 */
936 	if (is_sync_kiocb(iocb)) {
937 		int ret;
938 
939 		iocb->ki_user_data = res;
940 		if (iocb->ki_users == 1) {
941 			iocb->ki_users = 0;
942 			ret = 1;
943 		} else {
944 			spin_lock_irq(&ctx->ctx_lock);
945 			iocb->ki_users--;
946 			ret = (0 == iocb->ki_users);
947 			spin_unlock_irq(&ctx->ctx_lock);
948 		}
949 		/* sync iocbs put the task here for us */
950 		wake_up_process(iocb->ki_obj.tsk);
951 		return ret;
952 	}
953 
954 	info = &ctx->ring_info;
955 
956 	/* add a completion event to the ring buffer.
957 	 * must be done holding ctx->ctx_lock to prevent
958 	 * other code from messing with the tail
959 	 * pointer since we might be called from irq
960 	 * context.
961 	 */
962 	spin_lock_irqsave(&ctx->ctx_lock, flags);
963 
964 	if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
965 		list_del_init(&iocb->ki_run_list);
966 
967 	/*
968 	 * cancelled requests don't get events, userland was given one
969 	 * when the event got cancelled.
970 	 */
971 	if (kiocbIsCancelled(iocb))
972 		goto put_rq;
973 
974 	ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
975 
976 	tail = info->tail;
977 	event = aio_ring_event(info, tail, KM_IRQ0);
978 	if (++tail >= info->nr)
979 		tail = 0;
980 
981 	event->obj = (u64)(unsigned long)iocb->ki_obj.user;
982 	event->data = iocb->ki_user_data;
983 	event->res = res;
984 	event->res2 = res2;
985 
986 	dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
987 		ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
988 		res, res2);
989 
990 	/* after flagging the request as done, we
991 	 * must never even look at it again
992 	 */
993 	smp_wmb();	/* make event visible before updating tail */
994 
995 	info->tail = tail;
996 	ring->tail = tail;
997 
998 	put_aio_ring_event(event, KM_IRQ0);
999 	kunmap_atomic(ring, KM_IRQ1);
1000 
1001 	pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1002 
1003 	pr_debug("%ld retries: %d of %d\n", iocb->ki_retried,
1004 		iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
1005 put_rq:
1006 	/* everything turned out well, dispose of the aiocb. */
1007 	ret = __aio_put_req(ctx, iocb);
1008 
1009 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1010 
1011 	if (waitqueue_active(&ctx->wait))
1012 		wake_up(&ctx->wait);
1013 
1014 	if (ret)
1015 		put_ioctx(ctx);
1016 
1017 	return ret;
1018 }
1019 
1020 /* aio_read_evt
1021  *	Pull an event off of the ioctx's event ring.  Returns the number of
1022  *	events fetched (0 or 1 ;-)
1023  *	FIXME: make this use cmpxchg.
1024  *	TODO: make the ringbuffer user mmap()able (requires FIXME).
1025  */
1026 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1027 {
1028 	struct aio_ring_info *info = &ioctx->ring_info;
1029 	struct aio_ring *ring;
1030 	unsigned long head;
1031 	int ret = 0;
1032 
1033 	ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1034 	dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1035 		 (unsigned long)ring->head, (unsigned long)ring->tail,
1036 		 (unsigned long)ring->nr);
1037 
1038 	if (ring->head == ring->tail)
1039 		goto out;
1040 
1041 	spin_lock(&info->ring_lock);
1042 
1043 	head = ring->head % info->nr;
1044 	if (head != ring->tail) {
1045 		struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1046 		*ent = *evp;
1047 		head = (head + 1) % info->nr;
1048 		smp_mb(); /* finish reading the event before updatng the head */
1049 		ring->head = head;
1050 		ret = 1;
1051 		put_aio_ring_event(evp, KM_USER1);
1052 	}
1053 	spin_unlock(&info->ring_lock);
1054 
1055 out:
1056 	kunmap_atomic(ring, KM_USER0);
1057 	dprintk("leaving aio_read_evt: %d  h%lu t%lu\n", ret,
1058 		 (unsigned long)ring->head, (unsigned long)ring->tail);
1059 	return ret;
1060 }
1061 
1062 struct aio_timeout {
1063 	struct timer_list	timer;
1064 	int			timed_out;
1065 	struct task_struct	*p;
1066 };
1067 
1068 static void timeout_func(unsigned long data)
1069 {
1070 	struct aio_timeout *to = (struct aio_timeout *)data;
1071 
1072 	to->timed_out = 1;
1073 	wake_up_process(to->p);
1074 }
1075 
1076 static inline void init_timeout(struct aio_timeout *to)
1077 {
1078 	init_timer(&to->timer);
1079 	to->timer.data = (unsigned long)to;
1080 	to->timer.function = timeout_func;
1081 	to->timed_out = 0;
1082 	to->p = current;
1083 }
1084 
1085 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1086 			       const struct timespec *ts)
1087 {
1088 	to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1089 	if (time_after(to->timer.expires, jiffies))
1090 		add_timer(&to->timer);
1091 	else
1092 		to->timed_out = 1;
1093 }
1094 
1095 static inline void clear_timeout(struct aio_timeout *to)
1096 {
1097 	del_singleshot_timer_sync(&to->timer);
1098 }
1099 
1100 static int read_events(struct kioctx *ctx,
1101 			long min_nr, long nr,
1102 			struct io_event __user *event,
1103 			struct timespec __user *timeout)
1104 {
1105 	long			start_jiffies = jiffies;
1106 	struct task_struct	*tsk = current;
1107 	DECLARE_WAITQUEUE(wait, tsk);
1108 	int			ret;
1109 	int			i = 0;
1110 	struct io_event		ent;
1111 	struct aio_timeout	to;
1112 	int			retry = 0;
1113 
1114 	/* needed to zero any padding within an entry (there shouldn't be
1115 	 * any, but C is fun!
1116 	 */
1117 	memset(&ent, 0, sizeof(ent));
1118 retry:
1119 	ret = 0;
1120 	while (likely(i < nr)) {
1121 		ret = aio_read_evt(ctx, &ent);
1122 		if (unlikely(ret <= 0))
1123 			break;
1124 
1125 		dprintk("read event: %Lx %Lx %Lx %Lx\n",
1126 			ent.data, ent.obj, ent.res, ent.res2);
1127 
1128 		/* Could we split the check in two? */
1129 		ret = -EFAULT;
1130 		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1131 			dprintk("aio: lost an event due to EFAULT.\n");
1132 			break;
1133 		}
1134 		ret = 0;
1135 
1136 		/* Good, event copied to userland, update counts. */
1137 		event ++;
1138 		i ++;
1139 	}
1140 
1141 	if (min_nr <= i)
1142 		return i;
1143 	if (ret)
1144 		return ret;
1145 
1146 	/* End fast path */
1147 
1148 	/* racey check, but it gets redone */
1149 	if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1150 		retry = 1;
1151 		aio_run_all_iocbs(ctx);
1152 		goto retry;
1153 	}
1154 
1155 	init_timeout(&to);
1156 	if (timeout) {
1157 		struct timespec	ts;
1158 		ret = -EFAULT;
1159 		if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1160 			goto out;
1161 
1162 		set_timeout(start_jiffies, &to, &ts);
1163 	}
1164 
1165 	while (likely(i < nr)) {
1166 		add_wait_queue_exclusive(&ctx->wait, &wait);
1167 		do {
1168 			set_task_state(tsk, TASK_INTERRUPTIBLE);
1169 			ret = aio_read_evt(ctx, &ent);
1170 			if (ret)
1171 				break;
1172 			if (min_nr <= i)
1173 				break;
1174 			ret = 0;
1175 			if (to.timed_out)	/* Only check after read evt */
1176 				break;
1177 			schedule();
1178 			if (signal_pending(tsk)) {
1179 				ret = -EINTR;
1180 				break;
1181 			}
1182 			/*ret = aio_read_evt(ctx, &ent);*/
1183 		} while (1) ;
1184 
1185 		set_task_state(tsk, TASK_RUNNING);
1186 		remove_wait_queue(&ctx->wait, &wait);
1187 
1188 		if (unlikely(ret <= 0))
1189 			break;
1190 
1191 		ret = -EFAULT;
1192 		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1193 			dprintk("aio: lost an event due to EFAULT.\n");
1194 			break;
1195 		}
1196 
1197 		/* Good, event copied to userland, update counts. */
1198 		event ++;
1199 		i ++;
1200 	}
1201 
1202 	if (timeout)
1203 		clear_timeout(&to);
1204 out:
1205 	return i ? i : ret;
1206 }
1207 
1208 /* Take an ioctx and remove it from the list of ioctx's.  Protects
1209  * against races with itself via ->dead.
1210  */
1211 static void io_destroy(struct kioctx *ioctx)
1212 {
1213 	struct mm_struct *mm = current->mm;
1214 	struct kioctx **tmp;
1215 	int was_dead;
1216 
1217 	/* delete the entry from the list is someone else hasn't already */
1218 	write_lock(&mm->ioctx_list_lock);
1219 	was_dead = ioctx->dead;
1220 	ioctx->dead = 1;
1221 	for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1222 	     tmp = &(*tmp)->next)
1223 		;
1224 	if (*tmp)
1225 		*tmp = ioctx->next;
1226 	write_unlock(&mm->ioctx_list_lock);
1227 
1228 	dprintk("aio_release(%p)\n", ioctx);
1229 	if (likely(!was_dead))
1230 		put_ioctx(ioctx);	/* twice for the list */
1231 
1232 	aio_cancel_all(ioctx);
1233 	wait_for_all_aios(ioctx);
1234 	put_ioctx(ioctx);	/* once for the lookup */
1235 }
1236 
1237 /* sys_io_setup:
1238  *	Create an aio_context capable of receiving at least nr_events.
1239  *	ctxp must not point to an aio_context that already exists, and
1240  *	must be initialized to 0 prior to the call.  On successful
1241  *	creation of the aio_context, *ctxp is filled in with the resulting
1242  *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1243  *	if the specified nr_events exceeds internal limits.  May fail
1244  *	with -EAGAIN if the specified nr_events exceeds the user's limit
1245  *	of available events.  May fail with -ENOMEM if insufficient kernel
1246  *	resources are available.  May fail with -EFAULT if an invalid
1247  *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1248  *	implemented.
1249  */
1250 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1251 {
1252 	struct kioctx *ioctx = NULL;
1253 	unsigned long ctx;
1254 	long ret;
1255 
1256 	ret = get_user(ctx, ctxp);
1257 	if (unlikely(ret))
1258 		goto out;
1259 
1260 	ret = -EINVAL;
1261 	if (unlikely(ctx || (int)nr_events <= 0)) {
1262 		pr_debug("EINVAL: io_setup: ctx or nr_events > max\n");
1263 		goto out;
1264 	}
1265 
1266 	ioctx = ioctx_alloc(nr_events);
1267 	ret = PTR_ERR(ioctx);
1268 	if (!IS_ERR(ioctx)) {
1269 		ret = put_user(ioctx->user_id, ctxp);
1270 		if (!ret)
1271 			return 0;
1272 
1273 		get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1274 		io_destroy(ioctx);
1275 	}
1276 
1277 out:
1278 	return ret;
1279 }
1280 
1281 /* sys_io_destroy:
1282  *	Destroy the aio_context specified.  May cancel any outstanding
1283  *	AIOs and block on completion.  Will fail with -ENOSYS if not
1284  *	implemented.  May fail with -EFAULT if the context pointed to
1285  *	is invalid.
1286  */
1287 asmlinkage long sys_io_destroy(aio_context_t ctx)
1288 {
1289 	struct kioctx *ioctx = lookup_ioctx(ctx);
1290 	if (likely(NULL != ioctx)) {
1291 		io_destroy(ioctx);
1292 		return 0;
1293 	}
1294 	pr_debug("EINVAL: io_destroy: invalid context id\n");
1295 	return -EINVAL;
1296 }
1297 
1298 /*
1299  * Default retry method for aio_read (also used for first time submit)
1300  * Responsible for updating iocb state as retries progress
1301  */
1302 static ssize_t aio_pread(struct kiocb *iocb)
1303 {
1304 	struct file *file = iocb->ki_filp;
1305 	struct address_space *mapping = file->f_mapping;
1306 	struct inode *inode = mapping->host;
1307 	ssize_t ret = 0;
1308 
1309 	ret = file->f_op->aio_read(iocb, iocb->ki_buf,
1310 		iocb->ki_left, iocb->ki_pos);
1311 
1312 	/*
1313 	 * Can't just depend on iocb->ki_left to determine
1314 	 * whether we are done. This may have been a short read.
1315 	 */
1316 	if (ret > 0) {
1317 		iocb->ki_buf += ret;
1318 		iocb->ki_left -= ret;
1319 		/*
1320 		 * For pipes and sockets we return once we have
1321 		 * some data; for regular files we retry till we
1322 		 * complete the entire read or find that we can't
1323 		 * read any more data (e.g short reads).
1324 		 */
1325 		if (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))
1326 			ret = -EIOCBRETRY;
1327 	}
1328 
1329 	/* This means we must have transferred all that we could */
1330 	/* No need to retry anymore */
1331 	if ((ret == 0) || (iocb->ki_left == 0))
1332 		ret = iocb->ki_nbytes - iocb->ki_left;
1333 
1334 	return ret;
1335 }
1336 
1337 /*
1338  * Default retry method for aio_write (also used for first time submit)
1339  * Responsible for updating iocb state as retries progress
1340  */
1341 static ssize_t aio_pwrite(struct kiocb *iocb)
1342 {
1343 	struct file *file = iocb->ki_filp;
1344 	ssize_t ret = 0;
1345 
1346 	ret = file->f_op->aio_write(iocb, iocb->ki_buf,
1347 		iocb->ki_left, iocb->ki_pos);
1348 
1349 	if (ret > 0) {
1350 		iocb->ki_buf += ret;
1351 		iocb->ki_left -= ret;
1352 
1353 		ret = -EIOCBRETRY;
1354 	}
1355 
1356 	/* This means we must have transferred all that we could */
1357 	/* No need to retry anymore */
1358 	if ((ret == 0) || (iocb->ki_left == 0))
1359 		ret = iocb->ki_nbytes - iocb->ki_left;
1360 
1361 	return ret;
1362 }
1363 
1364 static ssize_t aio_fdsync(struct kiocb *iocb)
1365 {
1366 	struct file *file = iocb->ki_filp;
1367 	ssize_t ret = -EINVAL;
1368 
1369 	if (file->f_op->aio_fsync)
1370 		ret = file->f_op->aio_fsync(iocb, 1);
1371 	return ret;
1372 }
1373 
1374 static ssize_t aio_fsync(struct kiocb *iocb)
1375 {
1376 	struct file *file = iocb->ki_filp;
1377 	ssize_t ret = -EINVAL;
1378 
1379 	if (file->f_op->aio_fsync)
1380 		ret = file->f_op->aio_fsync(iocb, 0);
1381 	return ret;
1382 }
1383 
1384 /*
1385  * aio_setup_iocb:
1386  *	Performs the initial checks and aio retry method
1387  *	setup for the kiocb at the time of io submission.
1388  */
1389 static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1390 {
1391 	struct file *file = kiocb->ki_filp;
1392 	ssize_t ret = 0;
1393 
1394 	switch (kiocb->ki_opcode) {
1395 	case IOCB_CMD_PREAD:
1396 		ret = -EBADF;
1397 		if (unlikely(!(file->f_mode & FMODE_READ)))
1398 			break;
1399 		ret = -EFAULT;
1400 		if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1401 			kiocb->ki_left)))
1402 			break;
1403 		ret = -EINVAL;
1404 		if (file->f_op->aio_read)
1405 			kiocb->ki_retry = aio_pread;
1406 		break;
1407 	case IOCB_CMD_PWRITE:
1408 		ret = -EBADF;
1409 		if (unlikely(!(file->f_mode & FMODE_WRITE)))
1410 			break;
1411 		ret = -EFAULT;
1412 		if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1413 			kiocb->ki_left)))
1414 			break;
1415 		ret = -EINVAL;
1416 		if (file->f_op->aio_write)
1417 			kiocb->ki_retry = aio_pwrite;
1418 		break;
1419 	case IOCB_CMD_FDSYNC:
1420 		ret = -EINVAL;
1421 		if (file->f_op->aio_fsync)
1422 			kiocb->ki_retry = aio_fdsync;
1423 		break;
1424 	case IOCB_CMD_FSYNC:
1425 		ret = -EINVAL;
1426 		if (file->f_op->aio_fsync)
1427 			kiocb->ki_retry = aio_fsync;
1428 		break;
1429 	default:
1430 		dprintk("EINVAL: io_submit: no operation provided\n");
1431 		ret = -EINVAL;
1432 	}
1433 
1434 	if (!kiocb->ki_retry)
1435 		return ret;
1436 
1437 	return 0;
1438 }
1439 
1440 /*
1441  * aio_wake_function:
1442  * 	wait queue callback function for aio notification,
1443  * 	Simply triggers a retry of the operation via kick_iocb.
1444  *
1445  * 	This callback is specified in the wait queue entry in
1446  *	a kiocb	(current->io_wait points to this wait queue
1447  *	entry when an aio operation executes; it is used
1448  * 	instead of a synchronous wait when an i/o blocking
1449  *	condition is encountered during aio).
1450  *
1451  * Note:
1452  * This routine is executed with the wait queue lock held.
1453  * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1454  * the ioctx lock inside the wait queue lock. This is safe
1455  * because this callback isn't used for wait queues which
1456  * are nested inside ioctx lock (i.e. ctx->wait)
1457  */
1458 static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1459 			     int sync, void *key)
1460 {
1461 	struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1462 
1463 	list_del_init(&wait->task_list);
1464 	kick_iocb(iocb);
1465 	return 1;
1466 }
1467 
1468 int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1469 			 struct iocb *iocb)
1470 {
1471 	struct kiocb *req;
1472 	struct file *file;
1473 	ssize_t ret;
1474 
1475 	/* enforce forwards compatibility on users */
1476 	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
1477 		     iocb->aio_reserved3)) {
1478 		pr_debug("EINVAL: io_submit: reserve field set\n");
1479 		return -EINVAL;
1480 	}
1481 
1482 	/* prevent overflows */
1483 	if (unlikely(
1484 	    (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1485 	    (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1486 	    ((ssize_t)iocb->aio_nbytes < 0)
1487 	   )) {
1488 		pr_debug("EINVAL: io_submit: overflow check\n");
1489 		return -EINVAL;
1490 	}
1491 
1492 	file = fget(iocb->aio_fildes);
1493 	if (unlikely(!file))
1494 		return -EBADF;
1495 
1496 	req = aio_get_req(ctx);		/* returns with 2 references to req */
1497 	if (unlikely(!req)) {
1498 		fput(file);
1499 		return -EAGAIN;
1500 	}
1501 
1502 	req->ki_filp = file;
1503 	ret = put_user(req->ki_key, &user_iocb->aio_key);
1504 	if (unlikely(ret)) {
1505 		dprintk("EFAULT: aio_key\n");
1506 		goto out_put_req;
1507 	}
1508 
1509 	req->ki_obj.user = user_iocb;
1510 	req->ki_user_data = iocb->aio_data;
1511 	req->ki_pos = iocb->aio_offset;
1512 
1513 	req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1514 	req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1515 	req->ki_opcode = iocb->aio_lio_opcode;
1516 	init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1517 	INIT_LIST_HEAD(&req->ki_wait.task_list);
1518 	req->ki_retried = 0;
1519 
1520 	ret = aio_setup_iocb(req);
1521 
1522 	if (ret)
1523 		goto out_put_req;
1524 
1525 	spin_lock_irq(&ctx->ctx_lock);
1526 	if (likely(list_empty(&ctx->run_list))) {
1527 		aio_run_iocb(req);
1528 	} else {
1529 		list_add_tail(&req->ki_run_list, &ctx->run_list);
1530 		/* drain the run list */
1531 		while (__aio_run_iocbs(ctx))
1532 			;
1533 	}
1534 	spin_unlock_irq(&ctx->ctx_lock);
1535 	aio_put_req(req);	/* drop extra ref to req */
1536 	return 0;
1537 
1538 out_put_req:
1539 	aio_put_req(req);	/* drop extra ref to req */
1540 	aio_put_req(req);	/* drop i/o ref to req */
1541 	return ret;
1542 }
1543 
1544 /* sys_io_submit:
1545  *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1546  *	the number of iocbs queued.  May return -EINVAL if the aio_context
1547  *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
1548  *	*iocbpp[0] is not properly initialized, if the operation specified
1549  *	is invalid for the file descriptor in the iocb.  May fail with
1550  *	-EFAULT if any of the data structures point to invalid data.  May
1551  *	fail with -EBADF if the file descriptor specified in the first
1552  *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
1553  *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1554  *	fail with -ENOSYS if not implemented.
1555  */
1556 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1557 			      struct iocb __user * __user *iocbpp)
1558 {
1559 	struct kioctx *ctx;
1560 	long ret = 0;
1561 	int i;
1562 
1563 	if (unlikely(nr < 0))
1564 		return -EINVAL;
1565 
1566 	if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1567 		return -EFAULT;
1568 
1569 	ctx = lookup_ioctx(ctx_id);
1570 	if (unlikely(!ctx)) {
1571 		pr_debug("EINVAL: io_submit: invalid context id\n");
1572 		return -EINVAL;
1573 	}
1574 
1575 	/*
1576 	 * AKPM: should this return a partial result if some of the IOs were
1577 	 * successfully submitted?
1578 	 */
1579 	for (i=0; i<nr; i++) {
1580 		struct iocb __user *user_iocb;
1581 		struct iocb tmp;
1582 
1583 		if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1584 			ret = -EFAULT;
1585 			break;
1586 		}
1587 
1588 		if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1589 			ret = -EFAULT;
1590 			break;
1591 		}
1592 
1593 		ret = io_submit_one(ctx, user_iocb, &tmp);
1594 		if (ret)
1595 			break;
1596 	}
1597 
1598 	put_ioctx(ctx);
1599 	return i ? i : ret;
1600 }
1601 
1602 /* lookup_kiocb
1603  *	Finds a given iocb for cancellation.
1604  *	MUST be called with ctx->ctx_lock held.
1605  */
1606 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1607 				  u32 key)
1608 {
1609 	struct list_head *pos;
1610 	/* TODO: use a hash or array, this sucks. */
1611 	list_for_each(pos, &ctx->active_reqs) {
1612 		struct kiocb *kiocb = list_kiocb(pos);
1613 		if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1614 			return kiocb;
1615 	}
1616 	return NULL;
1617 }
1618 
1619 /* sys_io_cancel:
1620  *	Attempts to cancel an iocb previously passed to io_submit.  If
1621  *	the operation is successfully cancelled, the resulting event is
1622  *	copied into the memory pointed to by result without being placed
1623  *	into the completion queue and 0 is returned.  May fail with
1624  *	-EFAULT if any of the data structures pointed to are invalid.
1625  *	May fail with -EINVAL if aio_context specified by ctx_id is
1626  *	invalid.  May fail with -EAGAIN if the iocb specified was not
1627  *	cancelled.  Will fail with -ENOSYS if not implemented.
1628  */
1629 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1630 			      struct io_event __user *result)
1631 {
1632 	int (*cancel)(struct kiocb *iocb, struct io_event *res);
1633 	struct kioctx *ctx;
1634 	struct kiocb *kiocb;
1635 	u32 key;
1636 	int ret;
1637 
1638 	ret = get_user(key, &iocb->aio_key);
1639 	if (unlikely(ret))
1640 		return -EFAULT;
1641 
1642 	ctx = lookup_ioctx(ctx_id);
1643 	if (unlikely(!ctx))
1644 		return -EINVAL;
1645 
1646 	spin_lock_irq(&ctx->ctx_lock);
1647 	ret = -EAGAIN;
1648 	kiocb = lookup_kiocb(ctx, iocb, key);
1649 	if (kiocb && kiocb->ki_cancel) {
1650 		cancel = kiocb->ki_cancel;
1651 		kiocb->ki_users ++;
1652 		kiocbSetCancelled(kiocb);
1653 	} else
1654 		cancel = NULL;
1655 	spin_unlock_irq(&ctx->ctx_lock);
1656 
1657 	if (NULL != cancel) {
1658 		struct io_event tmp;
1659 		pr_debug("calling cancel\n");
1660 		memset(&tmp, 0, sizeof(tmp));
1661 		tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1662 		tmp.data = kiocb->ki_user_data;
1663 		ret = cancel(kiocb, &tmp);
1664 		if (!ret) {
1665 			/* Cancellation succeeded -- copy the result
1666 			 * into the user's buffer.
1667 			 */
1668 			if (copy_to_user(result, &tmp, sizeof(tmp)))
1669 				ret = -EFAULT;
1670 		}
1671 	} else
1672 		printk(KERN_DEBUG "iocb has no cancel operation\n");
1673 
1674 	put_ioctx(ctx);
1675 
1676 	return ret;
1677 }
1678 
1679 /* io_getevents:
1680  *	Attempts to read at least min_nr events and up to nr events from
1681  *	the completion queue for the aio_context specified by ctx_id.  May
1682  *	fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1683  *	if nr is out of range, if when is out of range.  May fail with
1684  *	-EFAULT if any of the memory specified to is invalid.  May return
1685  *	0 or < min_nr if no events are available and the timeout specified
1686  *	by when	has elapsed, where when == NULL specifies an infinite
1687  *	timeout.  Note that the timeout pointed to by when is relative and
1688  *	will be updated if not NULL and the operation blocks.  Will fail
1689  *	with -ENOSYS if not implemented.
1690  */
1691 asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1692 				 long min_nr,
1693 				 long nr,
1694 				 struct io_event __user *events,
1695 				 struct timespec __user *timeout)
1696 {
1697 	struct kioctx *ioctx = lookup_ioctx(ctx_id);
1698 	long ret = -EINVAL;
1699 
1700 	if (likely(ioctx)) {
1701 		if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1702 			ret = read_events(ioctx, min_nr, nr, events, timeout);
1703 		put_ioctx(ioctx);
1704 	}
1705 
1706 	return ret;
1707 }
1708 
1709 __initcall(aio_setup);
1710 
1711 EXPORT_SYMBOL(aio_complete);
1712 EXPORT_SYMBOL(aio_put_req);
1713 EXPORT_SYMBOL(wait_on_sync_kiocb);
1714