Root/fs/aio.c

1/*
2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
4 *
5 * Implements an efficient asynchronous io interface.
6 *
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 *
9 * See ../COPYING for licensing terms.
10 */
11#include <linux/kernel.h>
12#include <linux/init.h>
13#include <linux/errno.h>
14#include <linux/time.h>
15#include <linux/aio_abi.h>
16#include <linux/module.h>
17#include <linux/syscalls.h>
18#include <linux/uio.h>
19
20#define DEBUG 0
21
22#include <linux/sched.h>
23#include <linux/fs.h>
24#include <linux/file.h>
25#include <linux/mm.h>
26#include <linux/mman.h>
27#include <linux/slab.h>
28#include <linux/timer.h>
29#include <linux/aio.h>
30#include <linux/highmem.h>
31#include <linux/workqueue.h>
32#include <linux/security.h>
33#include <linux/eventfd.h>
34
35#include <asm/kmap_types.h>
36#include <asm/uaccess.h>
37#include <asm/mmu_context.h>
38
39#if DEBUG > 1
40#define dprintk printk
41#else
42#define dprintk(x...) do { ; } while (0)
43#endif
44
45/*------ sysctl variables----*/
46static DEFINE_SPINLOCK(aio_nr_lock);
47unsigned long aio_nr; /* current system wide number of aio requests */
48unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
49/*----end sysctl variables---*/
50
51static struct kmem_cache *kiocb_cachep;
52static struct kmem_cache *kioctx_cachep;
53
54static struct workqueue_struct *aio_wq;
55
56/* Used for rare fput completion. */
57static void aio_fput_routine(struct work_struct *);
58static DECLARE_WORK(fput_work, aio_fput_routine);
59
60static DEFINE_SPINLOCK(fput_lock);
61static LIST_HEAD(fput_head);
62
63static void aio_kick_handler(struct work_struct *);
64static void aio_queue_work(struct kioctx *);
65
66/* aio_setup
67 * Creates the slab caches used by the aio routines, panic on
68 * failure as this is done early during the boot sequence.
69 */
70static int __init aio_setup(void)
71{
72    kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
73    kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
74
75    aio_wq = create_workqueue("aio");
76
77    pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
78
79    return 0;
80}
81
82static 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
102static int aio_setup_ring(struct kioctx *ctx)
103{
104    struct aio_ring *ring;
105    struct aio_ring_info *info = &ctx->ring_info;
106    unsigned nr_events = ctx->max_reqs;
107    unsigned long size;
108    int nr_pages;
109
110    /* Compensate for the ring buffer's head/tail overlap entry */
111    nr_events += 2; /* 1 is required, 2 for good luck */
112
113    size = sizeof(struct aio_ring);
114    size += sizeof(struct io_event) * nr_events;
115    nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
116
117    if (nr_pages < 0)
118        return -EINVAL;
119
120    nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
121
122    info->nr = 0;
123    info->ring_pages = info->internal_pages;
124    if (nr_pages > AIO_RING_PAGES) {
125        info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
126        if (!info->ring_pages)
127            return -ENOMEM;
128    }
129
130    info->mmap_size = nr_pages * PAGE_SIZE;
131    dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
132    down_write(&ctx->mm->mmap_sem);
133    info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
134                  PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
135                  0);
136    if (IS_ERR((void *)info->mmap_base)) {
137        up_write(&ctx->mm->mmap_sem);
138        info->mmap_size = 0;
139        aio_free_ring(ctx);
140        return -EAGAIN;
141    }
142
143    dprintk("mmap address: 0x%08lx\n", info->mmap_base);
144    info->nr_pages = get_user_pages(current, ctx->mm,
145                    info->mmap_base, nr_pages,
146                    1, 0, info->ring_pages, NULL);
147    up_write(&ctx->mm->mmap_sem);
148
149    if (unlikely(info->nr_pages != nr_pages)) {
150        aio_free_ring(ctx);
151        return -EAGAIN;
152    }
153
154    ctx->user_id = info->mmap_base;
155
156    info->nr = nr_events; /* trusted copy */
157
158    ring = kmap_atomic(info->ring_pages[0], KM_USER0);
159    ring->nr = nr_events; /* user copy */
160    ring->id = ctx->user_id;
161    ring->head = ring->tail = 0;
162    ring->magic = AIO_RING_MAGIC;
163    ring->compat_features = AIO_RING_COMPAT_FEATURES;
164    ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
165    ring->header_length = sizeof(struct aio_ring);
166    kunmap_atomic(ring, KM_USER0);
167
168    return 0;
169}
170
171
172/* aio_ring_event: returns a pointer to the event at the given index from
173 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
174 */
175#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
176#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
177#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
178
179#define aio_ring_event(info, nr, km) ({ \
180    unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
181    struct io_event *__event; \
182    __event = kmap_atomic( \
183            (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
184    __event += pos % AIO_EVENTS_PER_PAGE; \
185    __event; \
186})
187
188#define put_aio_ring_event(event, km) do { \
189    struct io_event *__event = (event); \
190    (void)__event; \
191    kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
192} while(0)
193
194static void ctx_rcu_free(struct rcu_head *head)
195{
196    struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
197    unsigned nr_events = ctx->max_reqs;
198
199    kmem_cache_free(kioctx_cachep, ctx);
200
201    if (nr_events) {
202        spin_lock(&aio_nr_lock);
203        BUG_ON(aio_nr - nr_events > aio_nr);
204        aio_nr -= nr_events;
205        spin_unlock(&aio_nr_lock);
206    }
207}
208
209/* __put_ioctx
210 * Called when the last user of an aio context has gone away,
211 * and the struct needs to be freed.
212 */
213static void __put_ioctx(struct kioctx *ctx)
214{
215    BUG_ON(ctx->reqs_active);
216
217    cancel_delayed_work(&ctx->wq);
218    cancel_work_sync(&ctx->wq.work);
219    aio_free_ring(ctx);
220    mmdrop(ctx->mm);
221    ctx->mm = NULL;
222    pr_debug("__put_ioctx: freeing %p\n", ctx);
223    call_rcu(&ctx->rcu_head, ctx_rcu_free);
224}
225
226#define get_ioctx(kioctx) do { \
227    BUG_ON(atomic_read(&(kioctx)->users) <= 0); \
228    atomic_inc(&(kioctx)->users); \
229} while (0)
230#define put_ioctx(kioctx) do { \
231    BUG_ON(atomic_read(&(kioctx)->users) <= 0); \
232    if (unlikely(atomic_dec_and_test(&(kioctx)->users))) \
233        __put_ioctx(kioctx); \
234} while (0)
235
236/* ioctx_alloc
237 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
238 */
239static struct kioctx *ioctx_alloc(unsigned nr_events)
240{
241    struct mm_struct *mm;
242    struct kioctx *ctx;
243    int did_sync = 0;
244
245    /* Prevent overflows */
246    if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
247        (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
248        pr_debug("ENOMEM: nr_events too high\n");
249        return ERR_PTR(-EINVAL);
250    }
251
252    if ((unsigned long)nr_events > aio_max_nr)
253        return ERR_PTR(-EAGAIN);
254
255    ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
256    if (!ctx)
257        return ERR_PTR(-ENOMEM);
258
259    ctx->max_reqs = nr_events;
260    mm = ctx->mm = current->mm;
261    atomic_inc(&mm->mm_count);
262
263    atomic_set(&ctx->users, 1);
264    spin_lock_init(&ctx->ctx_lock);
265    spin_lock_init(&ctx->ring_info.ring_lock);
266    init_waitqueue_head(&ctx->wait);
267
268    INIT_LIST_HEAD(&ctx->active_reqs);
269    INIT_LIST_HEAD(&ctx->run_list);
270    INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
271
272    if (aio_setup_ring(ctx) < 0)
273        goto out_freectx;
274
275    /* limit the number of system wide aios */
276    do {
277        spin_lock_bh(&aio_nr_lock);
278        if (aio_nr + nr_events > aio_max_nr ||
279            aio_nr + nr_events < aio_nr)
280            ctx->max_reqs = 0;
281        else
282            aio_nr += ctx->max_reqs;
283        spin_unlock_bh(&aio_nr_lock);
284        if (ctx->max_reqs || did_sync)
285            break;
286
287        /* wait for rcu callbacks to have completed before giving up */
288        synchronize_rcu();
289        did_sync = 1;
290        ctx->max_reqs = nr_events;
291    } while (1);
292
293    if (ctx->max_reqs == 0)
294        goto out_cleanup;
295
296    /* now link into global list. */
297    spin_lock(&mm->ioctx_lock);
298    hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
299    spin_unlock(&mm->ioctx_lock);
300
301    dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
302        ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
303    return ctx;
304
305out_cleanup:
306    __put_ioctx(ctx);
307    return ERR_PTR(-EAGAIN);
308
309out_freectx:
310    mmdrop(mm);
311    kmem_cache_free(kioctx_cachep, ctx);
312    ctx = ERR_PTR(-ENOMEM);
313
314    dprintk("aio: error allocating ioctx %p\n", ctx);
315    return ctx;
316}
317
318/* aio_cancel_all
319 * Cancels all outstanding aio requests on an aio context. Used
320 * when the processes owning a context have all exited to encourage
321 * the rapid destruction of the kioctx.
322 */
323static void aio_cancel_all(struct kioctx *ctx)
324{
325    int (*cancel)(struct kiocb *, struct io_event *);
326    struct io_event res;
327    spin_lock_irq(&ctx->ctx_lock);
328    ctx->dead = 1;
329    while (!list_empty(&ctx->active_reqs)) {
330        struct list_head *pos = ctx->active_reqs.next;
331        struct kiocb *iocb = list_kiocb(pos);
332        list_del_init(&iocb->ki_list);
333        cancel = iocb->ki_cancel;
334        kiocbSetCancelled(iocb);
335        if (cancel) {
336            iocb->ki_users++;
337            spin_unlock_irq(&ctx->ctx_lock);
338            cancel(iocb, &res);
339            spin_lock_irq(&ctx->ctx_lock);
340        }
341    }
342    spin_unlock_irq(&ctx->ctx_lock);
343}
344
345static void wait_for_all_aios(struct kioctx *ctx)
346{
347    struct task_struct *tsk = current;
348    DECLARE_WAITQUEUE(wait, tsk);
349
350    spin_lock_irq(&ctx->ctx_lock);
351    if (!ctx->reqs_active)
352        goto out;
353
354    add_wait_queue(&ctx->wait, &wait);
355    set_task_state(tsk, TASK_UNINTERRUPTIBLE);
356    while (ctx->reqs_active) {
357        spin_unlock_irq(&ctx->ctx_lock);
358        io_schedule();
359        set_task_state(tsk, TASK_UNINTERRUPTIBLE);
360        spin_lock_irq(&ctx->ctx_lock);
361    }
362    __set_task_state(tsk, TASK_RUNNING);
363    remove_wait_queue(&ctx->wait, &wait);
364
365out:
366    spin_unlock_irq(&ctx->ctx_lock);
367}
368
369/* wait_on_sync_kiocb:
370 * Waits on the given sync kiocb to complete.
371 */
372ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
373{
374    while (iocb->ki_users) {
375        set_current_state(TASK_UNINTERRUPTIBLE);
376        if (!iocb->ki_users)
377            break;
378        io_schedule();
379    }
380    __set_current_state(TASK_RUNNING);
381    return iocb->ki_user_data;
382}
383
384/* exit_aio: called when the last user of mm goes away. At this point,
385 * there is no way for any new requests to be submited or any of the
386 * io_* syscalls to be called on the context. However, there may be
387 * outstanding requests which hold references to the context; as they
388 * go away, they will call put_ioctx and release any pinned memory
389 * associated with the request (held via struct page * references).
390 */
391void exit_aio(struct mm_struct *mm)
392{
393    struct kioctx *ctx;
394
395    while (!hlist_empty(&mm->ioctx_list)) {
396        ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
397        hlist_del_rcu(&ctx->list);
398
399        aio_cancel_all(ctx);
400
401        wait_for_all_aios(ctx);
402        /*
403         * Ensure we don't leave the ctx on the aio_wq
404         */
405        cancel_work_sync(&ctx->wq.work);
406
407        if (1 != atomic_read(&ctx->users))
408            printk(KERN_DEBUG
409                "exit_aio:ioctx still alive: %d %d %d\n",
410                atomic_read(&ctx->users), ctx->dead,
411                ctx->reqs_active);
412        put_ioctx(ctx);
413    }
414}
415
416/* aio_get_req
417 * Allocate a slot for an aio request. Increments the users count
418 * of the kioctx so that the kioctx stays around until all requests are
419 * complete. Returns NULL if no requests are free.
420 *
421 * Returns with kiocb->users set to 2. The io submit code path holds
422 * an extra reference while submitting the i/o.
423 * This prevents races between the aio code path referencing the
424 * req (after submitting it) and aio_complete() freeing the req.
425 */
426static struct kiocb *__aio_get_req(struct kioctx *ctx)
427{
428    struct kiocb *req = NULL;
429    struct aio_ring *ring;
430    int okay = 0;
431
432    req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
433    if (unlikely(!req))
434        return NULL;
435
436    req->ki_flags = 0;
437    req->ki_users = 2;
438    req->ki_key = 0;
439    req->ki_ctx = ctx;
440    req->ki_cancel = NULL;
441    req->ki_retry = NULL;
442    req->ki_dtor = NULL;
443    req->private = NULL;
444    req->ki_iovec = NULL;
445    INIT_LIST_HEAD(&req->ki_run_list);
446    req->ki_eventfd = NULL;
447
448    /* Check if the completion queue has enough free space to
449     * accept an event from this io.
450     */
451    spin_lock_irq(&ctx->ctx_lock);
452    ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
453    if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
454        list_add(&req->ki_list, &ctx->active_reqs);
455        ctx->reqs_active++;
456        okay = 1;
457    }
458    kunmap_atomic(ring, KM_USER0);
459    spin_unlock_irq(&ctx->ctx_lock);
460
461    if (!okay) {
462        kmem_cache_free(kiocb_cachep, req);
463        req = NULL;
464    }
465
466    return req;
467}
468
469static inline struct kiocb *aio_get_req(struct kioctx *ctx)
470{
471    struct kiocb *req;
472    /* Handle a potential starvation case -- should be exceedingly rare as
473     * requests will be stuck on fput_head only if the aio_fput_routine is
474     * delayed and the requests were the last user of the struct file.
475     */
476    req = __aio_get_req(ctx);
477    if (unlikely(NULL == req)) {
478        aio_fput_routine(NULL);
479        req = __aio_get_req(ctx);
480    }
481    return req;
482}
483
484static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
485{
486    assert_spin_locked(&ctx->ctx_lock);
487
488    if (req->ki_eventfd != NULL)
489        eventfd_ctx_put(req->ki_eventfd);
490    if (req->ki_dtor)
491        req->ki_dtor(req);
492    if (req->ki_iovec != &req->ki_inline_vec)
493        kfree(req->ki_iovec);
494    kmem_cache_free(kiocb_cachep, req);
495    ctx->reqs_active--;
496
497    if (unlikely(!ctx->reqs_active && ctx->dead))
498        wake_up(&ctx->wait);
499}
500
501static void aio_fput_routine(struct work_struct *data)
502{
503    spin_lock_irq(&fput_lock);
504    while (likely(!list_empty(&fput_head))) {
505        struct kiocb *req = list_kiocb(fput_head.next);
506        struct kioctx *ctx = req->ki_ctx;
507
508        list_del(&req->ki_list);
509        spin_unlock_irq(&fput_lock);
510
511        /* Complete the fput(s) */
512        if (req->ki_filp != NULL)
513            __fput(req->ki_filp);
514
515        /* Link the iocb into the context's free list */
516        spin_lock_irq(&ctx->ctx_lock);
517        really_put_req(ctx, req);
518        spin_unlock_irq(&ctx->ctx_lock);
519
520        put_ioctx(ctx);
521        spin_lock_irq(&fput_lock);
522    }
523    spin_unlock_irq(&fput_lock);
524}
525
526/* __aio_put_req
527 * Returns true if this put was the last user of the request.
528 */
529static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
530{
531    dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
532        req, atomic_long_read(&req->ki_filp->f_count));
533
534    assert_spin_locked(&ctx->ctx_lock);
535
536    req->ki_users--;
537    BUG_ON(req->ki_users < 0);
538    if (likely(req->ki_users))
539        return 0;
540    list_del(&req->ki_list); /* remove from active_reqs */
541    req->ki_cancel = NULL;
542    req->ki_retry = NULL;
543
544    /*
545     * Try to optimize the aio and eventfd file* puts, by avoiding to
546     * schedule work in case it is not __fput() time. In normal cases,
547     * we would not be holding the last reference to the file*, so
548     * this function will be executed w/out any aio kthread wakeup.
549     */
550    if (unlikely(atomic_long_dec_and_test(&req->ki_filp->f_count))) {
551        get_ioctx(ctx);
552        spin_lock(&fput_lock);
553        list_add(&req->ki_list, &fput_head);
554        spin_unlock(&fput_lock);
555        queue_work(aio_wq, &fput_work);
556    } else {
557        req->ki_filp = NULL;
558        really_put_req(ctx, req);
559    }
560    return 1;
561}
562
563/* aio_put_req
564 * Returns true if this put was the last user of the kiocb,
565 * false if the request is still in use.
566 */
567int aio_put_req(struct kiocb *req)
568{
569    struct kioctx *ctx = req->ki_ctx;
570    int ret;
571    spin_lock_irq(&ctx->ctx_lock);
572    ret = __aio_put_req(ctx, req);
573    spin_unlock_irq(&ctx->ctx_lock);
574    return ret;
575}
576
577static struct kioctx *lookup_ioctx(unsigned long ctx_id)
578{
579    struct mm_struct *mm = current->mm;
580    struct kioctx *ctx, *ret = NULL;
581    struct hlist_node *n;
582
583    rcu_read_lock();
584
585    hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
586        if (ctx->user_id == ctx_id && !ctx->dead) {
587            get_ioctx(ctx);
588            ret = ctx;
589            break;
590        }
591    }
592
593    rcu_read_unlock();
594    return ret;
595}
596
597/*
598 * use_mm
599 * Makes the calling kernel thread take on the specified
600 * mm context.
601 * Called by the retry thread execute retries within the
602 * iocb issuer's mm context, so that copy_from/to_user
603 * operations work seamlessly for aio.
604 * (Note: this routine is intended to be called only
605 * from a kernel thread context)
606 */
607static void use_mm(struct mm_struct *mm)
608{
609    struct mm_struct *active_mm;
610    struct task_struct *tsk = current;
611
612    task_lock(tsk);
613    active_mm = tsk->active_mm;
614    atomic_inc(&mm->mm_count);
615    tsk->mm = mm;
616    tsk->active_mm = mm;
617    switch_mm(active_mm, mm, tsk);
618    task_unlock(tsk);
619
620    mmdrop(active_mm);
621}
622
623/*
624 * unuse_mm
625 * Reverses the effect of use_mm, i.e. releases the
626 * specified mm context which was earlier taken on
627 * by the calling kernel thread
628 * (Note: this routine is intended to be called only
629 * from a kernel thread context)
630 */
631static void unuse_mm(struct mm_struct *mm)
632{
633    struct task_struct *tsk = current;
634
635    task_lock(tsk);
636    tsk->mm = NULL;
637    /* active_mm is still 'mm' */
638    enter_lazy_tlb(mm, tsk);
639    task_unlock(tsk);
640}
641
642/*
643 * Queue up a kiocb to be retried. Assumes that the kiocb
644 * has already been marked as kicked, and places it on
645 * the retry run list for the corresponding ioctx, if it
646 * isn't already queued. Returns 1 if it actually queued
647 * the kiocb (to tell the caller to activate the work
648 * queue to process it), or 0, if it found that it was
649 * already queued.
650 */
651static inline int __queue_kicked_iocb(struct kiocb *iocb)
652{
653    struct kioctx *ctx = iocb->ki_ctx;
654
655    assert_spin_locked(&ctx->ctx_lock);
656
657    if (list_empty(&iocb->ki_run_list)) {
658        list_add_tail(&iocb->ki_run_list,
659            &ctx->run_list);
660        return 1;
661    }
662    return 0;
663}
664
665/* aio_run_iocb
666 * This is the core aio execution routine. It is
667 * invoked both for initial i/o submission and
668 * subsequent retries via the aio_kick_handler.
669 * Expects to be invoked with iocb->ki_ctx->lock
670 * already held. The lock is released and reacquired
671 * as needed during processing.
672 *
673 * Calls the iocb retry method (already setup for the
674 * iocb on initial submission) for operation specific
675 * handling, but takes care of most of common retry
676 * execution details for a given iocb. The retry method
677 * needs to be non-blocking as far as possible, to avoid
678 * holding up other iocbs waiting to be serviced by the
679 * retry kernel thread.
680 *
681 * The trickier parts in this code have to do with
682 * ensuring that only one retry instance is in progress
683 * for a given iocb at any time. Providing that guarantee
684 * simplifies the coding of individual aio operations as
685 * it avoids various potential races.
686 */
687static ssize_t aio_run_iocb(struct kiocb *iocb)
688{
689    struct kioctx *ctx = iocb->ki_ctx;
690    ssize_t (*retry)(struct kiocb *);
691    ssize_t ret;
692
693    if (!(retry = iocb->ki_retry)) {
694        printk("aio_run_iocb: iocb->ki_retry = NULL\n");
695        return 0;
696    }
697
698    /*
699     * We don't want the next retry iteration for this
700     * operation to start until this one has returned and
701     * updated the iocb state. However, wait_queue functions
702     * can trigger a kick_iocb from interrupt context in the
703     * meantime, indicating that data is available for the next
704     * iteration. We want to remember that and enable the
705     * next retry iteration _after_ we are through with
706     * this one.
707     *
708     * So, in order to be able to register a "kick", but
709     * prevent it from being queued now, we clear the kick
710     * flag, but make the kick code *think* that the iocb is
711     * still on the run list until we are actually done.
712     * When we are done with this iteration, we check if
713     * the iocb was kicked in the meantime and if so, queue
714     * it up afresh.
715     */
716
717    kiocbClearKicked(iocb);
718
719    /*
720     * This is so that aio_complete knows it doesn't need to
721     * pull the iocb off the run list (We can't just call
722     * INIT_LIST_HEAD because we don't want a kick_iocb to
723     * queue this on the run list yet)
724     */
725    iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
726    spin_unlock_irq(&ctx->ctx_lock);
727
728    /* Quit retrying if the i/o has been cancelled */
729    if (kiocbIsCancelled(iocb)) {
730        ret = -EINTR;
731        aio_complete(iocb, ret, 0);
732        /* must not access the iocb after this */
733        goto out;
734    }
735
736    /*
737     * Now we are all set to call the retry method in async
738     * context.
739     */
740    ret = retry(iocb);
741
742    if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
743        BUG_ON(!list_empty(&iocb->ki_wait.task_list));
744        aio_complete(iocb, ret, 0);
745    }
746out:
747    spin_lock_irq(&ctx->ctx_lock);
748
749    if (-EIOCBRETRY == ret) {
750        /*
751         * OK, now that we are done with this iteration
752         * and know that there is more left to go,
753         * this is where we let go so that a subsequent
754         * "kick" can start the next iteration
755         */
756
757        /* will make __queue_kicked_iocb succeed from here on */
758        INIT_LIST_HEAD(&iocb->ki_run_list);
759        /* we must queue the next iteration ourselves, if it
760         * has already been kicked */
761        if (kiocbIsKicked(iocb)) {
762            __queue_kicked_iocb(iocb);
763
764            /*
765             * __queue_kicked_iocb will always return 1 here, because
766             * iocb->ki_run_list is empty at this point so it should
767             * be safe to unconditionally queue the context into the
768             * work queue.
769             */
770            aio_queue_work(ctx);
771        }
772    }
773    return ret;
774}
775
776/*
777 * __aio_run_iocbs:
778 * Process all pending retries queued on the ioctx
779 * run list.
780 * Assumes it is operating within the aio issuer's mm
781 * context.
782 */
783static int __aio_run_iocbs(struct kioctx *ctx)
784{
785    struct kiocb *iocb;
786    struct list_head run_list;
787
788    assert_spin_locked(&ctx->ctx_lock);
789
790    list_replace_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        aio_run_iocb(iocb);
800        __aio_put_req(ctx, iocb);
801     }
802    if (!list_empty(&ctx->run_list))
803        return 1;
804    return 0;
805}
806
807static void aio_queue_work(struct kioctx * ctx)
808{
809    unsigned long timeout;
810    /*
811     * if someone is waiting, get the work started right
812     * away, otherwise, use a longer delay
813     */
814    smp_mb();
815    if (waitqueue_active(&ctx->wait))
816        timeout = 1;
817    else
818        timeout = HZ/10;
819    queue_delayed_work(aio_wq, &ctx->wq, timeout);
820}
821
822
823/*
824 * aio_run_iocbs:
825 * Process all pending retries queued on the ioctx
826 * run list.
827 * Assumes it is operating within the aio issuer's mm
828 * context.
829 */
830static inline void aio_run_iocbs(struct kioctx *ctx)
831{
832    int requeue;
833
834    spin_lock_irq(&ctx->ctx_lock);
835
836    requeue = __aio_run_iocbs(ctx);
837    spin_unlock_irq(&ctx->ctx_lock);
838    if (requeue)
839        aio_queue_work(ctx);
840}
841
842/*
843 * just like aio_run_iocbs, but keeps running them until
844 * the list stays empty
845 */
846static inline void aio_run_all_iocbs(struct kioctx *ctx)
847{
848    spin_lock_irq(&ctx->ctx_lock);
849    while (__aio_run_iocbs(ctx))
850        ;
851    spin_unlock_irq(&ctx->ctx_lock);
852}
853
854/*
855 * aio_kick_handler:
856 * Work queue handler triggered to process pending
857 * retries on an ioctx. Takes on the aio issuer's
858 * mm context before running the iocbs, so that
859 * copy_xxx_user operates on the issuer's address
860 * space.
861 * Run on aiod's context.
862 */
863static void aio_kick_handler(struct work_struct *work)
864{
865    struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
866    mm_segment_t oldfs = get_fs();
867    struct mm_struct *mm;
868    int requeue;
869
870    set_fs(USER_DS);
871    use_mm(ctx->mm);
872    spin_lock_irq(&ctx->ctx_lock);
873    requeue =__aio_run_iocbs(ctx);
874    mm = ctx->mm;
875    spin_unlock_irq(&ctx->ctx_lock);
876     unuse_mm(mm);
877    set_fs(oldfs);
878    /*
879     * we're in a worker thread already, don't use queue_delayed_work,
880     */
881    if (requeue)
882        queue_delayed_work(aio_wq, &ctx->wq, 0);
883}
884
885
886/*
887 * Called by kick_iocb to queue the kiocb for retry
888 * and if required activate the aio work queue to process
889 * it
890 */
891static void try_queue_kicked_iocb(struct kiocb *iocb)
892{
893     struct kioctx *ctx = iocb->ki_ctx;
894    unsigned long flags;
895    int run = 0;
896
897    /* We're supposed to be the only path putting the iocb back on the run
898     * list. If we find that the iocb is *back* on a wait queue already
899     * than retry has happened before we could queue the iocb. This also
900     * means that the retry could have completed and freed our iocb, no
901     * good. */
902    BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
903
904    spin_lock_irqsave(&ctx->ctx_lock, flags);
905    /* set this inside the lock so that we can't race with aio_run_iocb()
906     * testing it and putting the iocb on the run list under the lock */
907    if (!kiocbTryKick(iocb))
908        run = __queue_kicked_iocb(iocb);
909    spin_unlock_irqrestore(&ctx->ctx_lock, flags);
910    if (run)
911        aio_queue_work(ctx);
912}
913
914/*
915 * kick_iocb:
916 * Called typically from a wait queue callback context
917 * (aio_wake_function) to trigger a retry of the iocb.
918 * The retry is usually executed by aio workqueue
919 * threads (See aio_kick_handler).
920 */
921void kick_iocb(struct kiocb *iocb)
922{
923    /* sync iocbs are easy: they can only ever be executing from a
924     * single context. */
925    if (is_sync_kiocb(iocb)) {
926        kiocbSetKicked(iocb);
927            wake_up_process(iocb->ki_obj.tsk);
928        return;
929    }
930
931    try_queue_kicked_iocb(iocb);
932}
933EXPORT_SYMBOL(kick_iocb);
934
935/* aio_complete
936 * Called when the io request on the given iocb is complete.
937 * Returns true if this is the last user of the request. The
938 * only other user of the request can be the cancellation code.
939 */
940int aio_complete(struct kiocb *iocb, long res, long res2)
941{
942    struct kioctx *ctx = iocb->ki_ctx;
943    struct aio_ring_info *info;
944    struct aio_ring *ring;
945    struct io_event *event;
946    unsigned long flags;
947    unsigned long tail;
948    int ret;
949
950    /*
951     * Special case handling for sync iocbs:
952     * - events go directly into the iocb for fast handling
953     * - the sync task with the iocb in its stack holds the single iocb
954     * ref, no other paths have a way to get another ref
955     * - the sync task helpfully left a reference to itself in the iocb
956     */
957    if (is_sync_kiocb(iocb)) {
958        BUG_ON(iocb->ki_users != 1);
959        iocb->ki_user_data = res;
960        iocb->ki_users = 0;
961        wake_up_process(iocb->ki_obj.tsk);
962        return 1;
963    }
964
965    info = &ctx->ring_info;
966
967    /* add a completion event to the ring buffer.
968     * must be done holding ctx->ctx_lock to prevent
969     * other code from messing with the tail
970     * pointer since we might be called from irq
971     * context.
972     */
973    spin_lock_irqsave(&ctx->ctx_lock, flags);
974
975    if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
976        list_del_init(&iocb->ki_run_list);
977
978    /*
979     * cancelled requests don't get events, userland was given one
980     * when the event got cancelled.
981     */
982    if (kiocbIsCancelled(iocb))
983        goto put_rq;
984
985    ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
986
987    tail = info->tail;
988    event = aio_ring_event(info, tail, KM_IRQ0);
989    if (++tail >= info->nr)
990        tail = 0;
991
992    event->obj = (u64)(unsigned long)iocb->ki_obj.user;
993    event->data = iocb->ki_user_data;
994    event->res = res;
995    event->res2 = res2;
996
997    dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
998        ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
999        res, res2);
1000
1001    /* after flagging the request as done, we
1002     * must never even look at it again
1003     */
1004    smp_wmb(); /* make event visible before updating tail */
1005
1006    info->tail = tail;
1007    ring->tail = tail;
1008
1009    put_aio_ring_event(event, KM_IRQ0);
1010    kunmap_atomic(ring, KM_IRQ1);
1011
1012    pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1013
1014    /*
1015     * Check if the user asked us to deliver the result through an
1016     * eventfd. The eventfd_signal() function is safe to be called
1017     * from IRQ context.
1018     */
1019    if (iocb->ki_eventfd != NULL)
1020        eventfd_signal(iocb->ki_eventfd, 1);
1021
1022put_rq:
1023    /* everything turned out well, dispose of the aiocb. */
1024    ret = __aio_put_req(ctx, iocb);
1025
1026    /*
1027     * We have to order our ring_info tail store above and test
1028     * of the wait list below outside the wait lock. This is
1029     * like in wake_up_bit() where clearing a bit has to be
1030     * ordered with the unlocked test.
1031     */
1032    smp_mb();
1033
1034    if (waitqueue_active(&ctx->wait))
1035        wake_up(&ctx->wait);
1036
1037    spin_unlock_irqrestore(&ctx->ctx_lock, flags);
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 */
1047static 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
1076out:
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
1083struct aio_timeout {
1084    struct timer_list timer;
1085    int timed_out;
1086    struct task_struct *p;
1087};
1088
1089static 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
1097static inline void init_timeout(struct aio_timeout *to)
1098{
1099    setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1100    to->timed_out = 0;
1101    to->p = current;
1102}
1103
1104static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1105                   const struct timespec *ts)
1106{
1107    to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1108    if (time_after(to->timer.expires, jiffies))
1109        add_timer(&to->timer);
1110    else
1111        to->timed_out = 1;
1112}
1113
1114static inline void clear_timeout(struct aio_timeout *to)
1115{
1116    del_singleshot_timer_sync(&to->timer);
1117}
1118
1119static int read_events(struct kioctx *ctx,
1120            long min_nr, long nr,
1121            struct io_event __user *event,
1122            struct timespec __user *timeout)
1123{
1124    long start_jiffies = jiffies;
1125    struct task_struct *tsk = current;
1126    DECLARE_WAITQUEUE(wait, tsk);
1127    int ret;
1128    int i = 0;
1129    struct io_event ent;
1130    struct aio_timeout to;
1131    int retry = 0;
1132
1133    /* needed to zero any padding within an entry (there shouldn't be
1134     * any, but C is fun!
1135     */
1136    memset(&ent, 0, sizeof(ent));
1137retry:
1138    ret = 0;
1139    while (likely(i < nr)) {
1140        ret = aio_read_evt(ctx, &ent);
1141        if (unlikely(ret <= 0))
1142            break;
1143
1144        dprintk("read event: %Lx %Lx %Lx %Lx\n",
1145            ent.data, ent.obj, ent.res, ent.res2);
1146
1147        /* Could we split the check in two? */
1148        ret = -EFAULT;
1149        if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1150            dprintk("aio: lost an event due to EFAULT.\n");
1151            break;
1152        }
1153        ret = 0;
1154
1155        /* Good, event copied to userland, update counts. */
1156        event ++;
1157        i ++;
1158    }
1159
1160    if (min_nr <= i)
1161        return i;
1162    if (ret)
1163        return ret;
1164
1165    /* End fast path */
1166
1167    /* racey check, but it gets redone */
1168    if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1169        retry = 1;
1170        aio_run_all_iocbs(ctx);
1171        goto retry;
1172    }
1173
1174    init_timeout(&to);
1175    if (timeout) {
1176        struct timespec ts;
1177        ret = -EFAULT;
1178        if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1179            goto out;
1180
1181        set_timeout(start_jiffies, &to, &ts);
1182    }
1183
1184    while (likely(i < nr)) {
1185        add_wait_queue_exclusive(&ctx->wait, &wait);
1186        do {
1187            set_task_state(tsk, TASK_INTERRUPTIBLE);
1188            ret = aio_read_evt(ctx, &ent);
1189            if (ret)
1190                break;
1191            if (min_nr <= i)
1192                break;
1193            if (unlikely(ctx->dead)) {
1194                ret = -EINVAL;
1195                break;
1196            }
1197            if (to.timed_out) /* Only check after read evt */
1198                break;
1199            /* Try to only show up in io wait if there are ops
1200             * in flight */
1201            if (ctx->reqs_active)
1202                io_schedule();
1203            else
1204                schedule();
1205            if (signal_pending(tsk)) {
1206                ret = -EINTR;
1207                break;
1208            }
1209            /*ret = aio_read_evt(ctx, &ent);*/
1210        } while (1) ;
1211
1212        set_task_state(tsk, TASK_RUNNING);
1213        remove_wait_queue(&ctx->wait, &wait);
1214
1215        if (unlikely(ret <= 0))
1216            break;
1217
1218        ret = -EFAULT;
1219        if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1220            dprintk("aio: lost an event due to EFAULT.\n");
1221            break;
1222        }
1223
1224        /* Good, event copied to userland, update counts. */
1225        event ++;
1226        i ++;
1227    }
1228
1229    if (timeout)
1230        clear_timeout(&to);
1231out:
1232    destroy_timer_on_stack(&to.timer);
1233    return i ? i : ret;
1234}
1235
1236/* Take an ioctx and remove it from the list of ioctx's. Protects
1237 * against races with itself via ->dead.
1238 */
1239static void io_destroy(struct kioctx *ioctx)
1240{
1241    struct mm_struct *mm = current->mm;
1242    int was_dead;
1243
1244    /* delete the entry from the list is someone else hasn't already */
1245    spin_lock(&mm->ioctx_lock);
1246    was_dead = ioctx->dead;
1247    ioctx->dead = 1;
1248    hlist_del_rcu(&ioctx->list);
1249    spin_unlock(&mm->ioctx_lock);
1250
1251    dprintk("aio_release(%p)\n", ioctx);
1252    if (likely(!was_dead))
1253        put_ioctx(ioctx); /* twice for the list */
1254
1255    aio_cancel_all(ioctx);
1256    wait_for_all_aios(ioctx);
1257
1258    /*
1259     * Wake up any waiters. The setting of ctx->dead must be seen
1260     * by other CPUs at this point. Right now, we rely on the
1261     * locking done by the above calls to ensure this consistency.
1262     */
1263    wake_up(&ioctx->wait);
1264    put_ioctx(ioctx); /* once for the lookup */
1265}
1266
1267/* sys_io_setup:
1268 * Create an aio_context capable of receiving at least nr_events.
1269 * ctxp must not point to an aio_context that already exists, and
1270 * must be initialized to 0 prior to the call. On successful
1271 * creation of the aio_context, *ctxp is filled in with the resulting
1272 * handle. May fail with -EINVAL if *ctxp is not initialized,
1273 * if the specified nr_events exceeds internal limits. May fail
1274 * with -EAGAIN if the specified nr_events exceeds the user's limit
1275 * of available events. May fail with -ENOMEM if insufficient kernel
1276 * resources are available. May fail with -EFAULT if an invalid
1277 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1278 * implemented.
1279 */
1280SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1281{
1282    struct kioctx *ioctx = NULL;
1283    unsigned long ctx;
1284    long ret;
1285
1286    ret = get_user(ctx, ctxp);
1287    if (unlikely(ret))
1288        goto out;
1289
1290    ret = -EINVAL;
1291    if (unlikely(ctx || nr_events == 0)) {
1292        pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1293                 ctx, nr_events);
1294        goto out;
1295    }
1296
1297    ioctx = ioctx_alloc(nr_events);
1298    ret = PTR_ERR(ioctx);
1299    if (!IS_ERR(ioctx)) {
1300        ret = put_user(ioctx->user_id, ctxp);
1301        if (!ret)
1302            return 0;
1303
1304        get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1305        io_destroy(ioctx);
1306    }
1307
1308out:
1309    return ret;
1310}
1311
1312/* sys_io_destroy:
1313 * Destroy the aio_context specified. May cancel any outstanding
1314 * AIOs and block on completion. Will fail with -ENOSYS if not
1315 * implemented. May fail with -EFAULT if the context pointed to
1316 * is invalid.
1317 */
1318SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1319{
1320    struct kioctx *ioctx = lookup_ioctx(ctx);
1321    if (likely(NULL != ioctx)) {
1322        io_destroy(ioctx);
1323        return 0;
1324    }
1325    pr_debug("EINVAL: io_destroy: invalid context id\n");
1326    return -EINVAL;
1327}
1328
1329static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1330{
1331    struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1332
1333    BUG_ON(ret <= 0);
1334
1335    while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1336        ssize_t this = min((ssize_t)iov->iov_len, ret);
1337        iov->iov_base += this;
1338        iov->iov_len -= this;
1339        iocb->ki_left -= this;
1340        ret -= this;
1341        if (iov->iov_len == 0) {
1342            iocb->ki_cur_seg++;
1343            iov++;
1344        }
1345    }
1346
1347    /* the caller should not have done more io than what fit in
1348     * the remaining iovecs */
1349    BUG_ON(ret > 0 && iocb->ki_left == 0);
1350}
1351
1352static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1353{
1354    struct file *file = iocb->ki_filp;
1355    struct address_space *mapping = file->f_mapping;
1356    struct inode *inode = mapping->host;
1357    ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1358             unsigned long, loff_t);
1359    ssize_t ret = 0;
1360    unsigned short opcode;
1361
1362    if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1363        (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1364        rw_op = file->f_op->aio_read;
1365        opcode = IOCB_CMD_PREADV;
1366    } else {
1367        rw_op = file->f_op->aio_write;
1368        opcode = IOCB_CMD_PWRITEV;
1369    }
1370
1371    /* This matches the pread()/pwrite() logic */
1372    if (iocb->ki_pos < 0)
1373        return -EINVAL;
1374
1375    do {
1376        ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1377                iocb->ki_nr_segs - iocb->ki_cur_seg,
1378                iocb->ki_pos);
1379        if (ret > 0)
1380            aio_advance_iovec(iocb, ret);
1381
1382    /* retry all partial writes. retry partial reads as long as its a
1383     * regular file. */
1384    } while (ret > 0 && iocb->ki_left > 0 &&
1385         (opcode == IOCB_CMD_PWRITEV ||
1386          (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1387
1388    /* This means we must have transferred all that we could */
1389    /* No need to retry anymore */
1390    if ((ret == 0) || (iocb->ki_left == 0))
1391        ret = iocb->ki_nbytes - iocb->ki_left;
1392
1393    /* If we managed to write some out we return that, rather than
1394     * the eventual error. */
1395    if (opcode == IOCB_CMD_PWRITEV
1396        && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1397        && iocb->ki_nbytes - iocb->ki_left)
1398        ret = iocb->ki_nbytes - iocb->ki_left;
1399
1400    return ret;
1401}
1402
1403static ssize_t aio_fdsync(struct kiocb *iocb)
1404{
1405    struct file *file = iocb->ki_filp;
1406    ssize_t ret = -EINVAL;
1407
1408    if (file->f_op->aio_fsync)
1409        ret = file->f_op->aio_fsync(iocb, 1);
1410    return ret;
1411}
1412
1413static ssize_t aio_fsync(struct kiocb *iocb)
1414{
1415    struct file *file = iocb->ki_filp;
1416    ssize_t ret = -EINVAL;
1417
1418    if (file->f_op->aio_fsync)
1419        ret = file->f_op->aio_fsync(iocb, 0);
1420    return ret;
1421}
1422
1423static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb)
1424{
1425    ssize_t ret;
1426
1427    ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf,
1428                    kiocb->ki_nbytes, 1,
1429                    &kiocb->ki_inline_vec, &kiocb->ki_iovec);
1430    if (ret < 0)
1431        goto out;
1432
1433    kiocb->ki_nr_segs = kiocb->ki_nbytes;
1434    kiocb->ki_cur_seg = 0;
1435    /* ki_nbytes/left now reflect bytes instead of segs */
1436    kiocb->ki_nbytes = ret;
1437    kiocb->ki_left = ret;
1438
1439    ret = 0;
1440out:
1441    return ret;
1442}
1443
1444static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1445{
1446    kiocb->ki_iovec = &kiocb->ki_inline_vec;
1447    kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1448    kiocb->ki_iovec->iov_len = kiocb->ki_left;
1449    kiocb->ki_nr_segs = 1;
1450    kiocb->ki_cur_seg = 0;
1451    return 0;
1452}
1453
1454/*
1455 * aio_setup_iocb:
1456 * Performs the initial checks and aio retry method
1457 * setup for the kiocb at the time of io submission.
1458 */
1459static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1460{
1461    struct file *file = kiocb->ki_filp;
1462    ssize_t ret = 0;
1463
1464    switch (kiocb->ki_opcode) {
1465    case IOCB_CMD_PREAD:
1466        ret = -EBADF;
1467        if (unlikely(!(file->f_mode & FMODE_READ)))
1468            break;
1469        ret = -EFAULT;
1470        if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1471            kiocb->ki_left)))
1472            break;
1473        ret = security_file_permission(file, MAY_READ);
1474        if (unlikely(ret))
1475            break;
1476        ret = aio_setup_single_vector(kiocb);
1477        if (ret)
1478            break;
1479        ret = -EINVAL;
1480        if (file->f_op->aio_read)
1481            kiocb->ki_retry = aio_rw_vect_retry;
1482        break;
1483    case IOCB_CMD_PWRITE:
1484        ret = -EBADF;
1485        if (unlikely(!(file->f_mode & FMODE_WRITE)))
1486            break;
1487        ret = -EFAULT;
1488        if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1489            kiocb->ki_left)))
1490            break;
1491        ret = security_file_permission(file, MAY_WRITE);
1492        if (unlikely(ret))
1493            break;
1494        ret = aio_setup_single_vector(kiocb);
1495        if (ret)
1496            break;
1497        ret = -EINVAL;
1498        if (file->f_op->aio_write)
1499            kiocb->ki_retry = aio_rw_vect_retry;
1500        break;
1501    case IOCB_CMD_PREADV:
1502        ret = -EBADF;
1503        if (unlikely(!(file->f_mode & FMODE_READ)))
1504            break;
1505        ret = security_file_permission(file, MAY_READ);
1506        if (unlikely(ret))
1507            break;
1508        ret = aio_setup_vectored_rw(READ, kiocb);
1509        if (ret)
1510            break;
1511        ret = -EINVAL;
1512        if (file->f_op->aio_read)
1513            kiocb->ki_retry = aio_rw_vect_retry;
1514        break;
1515    case IOCB_CMD_PWRITEV:
1516        ret = -EBADF;
1517        if (unlikely(!(file->f_mode & FMODE_WRITE)))
1518            break;
1519        ret = security_file_permission(file, MAY_WRITE);
1520        if (unlikely(ret))
1521            break;
1522        ret = aio_setup_vectored_rw(WRITE, kiocb);
1523        if (ret)
1524            break;
1525        ret = -EINVAL;
1526        if (file->f_op->aio_write)
1527            kiocb->ki_retry = aio_rw_vect_retry;
1528        break;
1529    case IOCB_CMD_FDSYNC:
1530        ret = -EINVAL;
1531        if (file->f_op->aio_fsync)
1532            kiocb->ki_retry = aio_fdsync;
1533        break;
1534    case IOCB_CMD_FSYNC:
1535        ret = -EINVAL;
1536        if (file->f_op->aio_fsync)
1537            kiocb->ki_retry = aio_fsync;
1538        break;
1539    default:
1540        dprintk("EINVAL: io_submit: no operation provided\n");
1541        ret = -EINVAL;
1542    }
1543
1544    if (!kiocb->ki_retry)
1545        return ret;
1546
1547    return 0;
1548}
1549
1550/*
1551 * aio_wake_function:
1552 * wait queue callback function for aio notification,
1553 * Simply triggers a retry of the operation via kick_iocb.
1554 *
1555 * This callback is specified in the wait queue entry in
1556 * a kiocb.
1557 *
1558 * Note:
1559 * This routine is executed with the wait queue lock held.
1560 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1561 * the ioctx lock inside the wait queue lock. This is safe
1562 * because this callback isn't used for wait queues which
1563 * are nested inside ioctx lock (i.e. ctx->wait)
1564 */
1565static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1566                 int sync, void *key)
1567{
1568    struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1569
1570    list_del_init(&wait->task_list);
1571    kick_iocb(iocb);
1572    return 1;
1573}
1574
1575static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1576             struct iocb *iocb)
1577{
1578    struct kiocb *req;
1579    struct file *file;
1580    ssize_t ret;
1581
1582    /* enforce forwards compatibility on users */
1583    if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1584        pr_debug("EINVAL: io_submit: reserve field set\n");
1585        return -EINVAL;
1586    }
1587
1588    /* prevent overflows */
1589    if (unlikely(
1590        (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1591        (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1592        ((ssize_t)iocb->aio_nbytes < 0)
1593       )) {
1594        pr_debug("EINVAL: io_submit: overflow check\n");
1595        return -EINVAL;
1596    }
1597
1598    file = fget(iocb->aio_fildes);
1599    if (unlikely(!file))
1600        return -EBADF;
1601
1602    req = aio_get_req(ctx); /* returns with 2 references to req */
1603    if (unlikely(!req)) {
1604        fput(file);
1605        return -EAGAIN;
1606    }
1607    req->ki_filp = file;
1608    if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1609        /*
1610         * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1611         * instance of the file* now. The file descriptor must be
1612         * an eventfd() fd, and will be signaled for each completed
1613         * event using the eventfd_signal() function.
1614         */
1615        req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1616        if (IS_ERR(req->ki_eventfd)) {
1617            ret = PTR_ERR(req->ki_eventfd);
1618            req->ki_eventfd = NULL;
1619            goto out_put_req;
1620        }
1621    }
1622
1623    ret = put_user(req->ki_key, &user_iocb->aio_key);
1624    if (unlikely(ret)) {
1625        dprintk("EFAULT: aio_key\n");
1626        goto out_put_req;
1627    }
1628
1629    req->ki_obj.user = user_iocb;
1630    req->ki_user_data = iocb->aio_data;
1631    req->ki_pos = iocb->aio_offset;
1632
1633    req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1634    req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1635    req->ki_opcode = iocb->aio_lio_opcode;
1636    init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1637    INIT_LIST_HEAD(&req->ki_wait.task_list);
1638
1639    ret = aio_setup_iocb(req);
1640
1641    if (ret)
1642        goto out_put_req;
1643
1644    spin_lock_irq(&ctx->ctx_lock);
1645    aio_run_iocb(req);
1646    if (!list_empty(&ctx->run_list)) {
1647        /* drain the run list */
1648        while (__aio_run_iocbs(ctx))
1649            ;
1650    }
1651    spin_unlock_irq(&ctx->ctx_lock);
1652    aio_put_req(req); /* drop extra ref to req */
1653    return 0;
1654
1655out_put_req:
1656    aio_put_req(req); /* drop extra ref to req */
1657    aio_put_req(req); /* drop i/o ref to req */
1658    return ret;
1659}
1660
1661/* sys_io_submit:
1662 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1663 * the number of iocbs queued. May return -EINVAL if the aio_context
1664 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1665 * *iocbpp[0] is not properly initialized, if the operation specified
1666 * is invalid for the file descriptor in the iocb. May fail with
1667 * -EFAULT if any of the data structures point to invalid data. May
1668 * fail with -EBADF if the file descriptor specified in the first
1669 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1670 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1671 * fail with -ENOSYS if not implemented.
1672 */
1673SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1674        struct iocb __user * __user *, iocbpp)
1675{
1676    struct kioctx *ctx;
1677    long ret = 0;
1678    int i;
1679
1680    if (unlikely(nr < 0))
1681        return -EINVAL;
1682
1683    if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1684        return -EFAULT;
1685
1686    ctx = lookup_ioctx(ctx_id);
1687    if (unlikely(!ctx)) {
1688        pr_debug("EINVAL: io_submit: invalid context id\n");
1689        return -EINVAL;
1690    }
1691
1692    /*
1693     * AKPM: should this return a partial result if some of the IOs were
1694     * successfully submitted?
1695     */
1696    for (i=0; i<nr; i++) {
1697        struct iocb __user *user_iocb;
1698        struct iocb tmp;
1699
1700        if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1701            ret = -EFAULT;
1702            break;
1703        }
1704
1705        if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1706            ret = -EFAULT;
1707            break;
1708        }
1709
1710        ret = io_submit_one(ctx, user_iocb, &tmp);
1711        if (ret)
1712            break;
1713    }
1714
1715    put_ioctx(ctx);
1716    return i ? i : ret;
1717}
1718
1719/* lookup_kiocb
1720 * Finds a given iocb for cancellation.
1721 */
1722static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1723                  u32 key)
1724{
1725    struct list_head *pos;
1726
1727    assert_spin_locked(&ctx->ctx_lock);
1728
1729    /* TODO: use a hash or array, this sucks. */
1730    list_for_each(pos, &ctx->active_reqs) {
1731        struct kiocb *kiocb = list_kiocb(pos);
1732        if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1733            return kiocb;
1734    }
1735    return NULL;
1736}
1737
1738/* sys_io_cancel:
1739 * Attempts to cancel an iocb previously passed to io_submit. If
1740 * the operation is successfully cancelled, the resulting event is
1741 * copied into the memory pointed to by result without being placed
1742 * into the completion queue and 0 is returned. May fail with
1743 * -EFAULT if any of the data structures pointed to are invalid.
1744 * May fail with -EINVAL if aio_context specified by ctx_id is
1745 * invalid. May fail with -EAGAIN if the iocb specified was not
1746 * cancelled. Will fail with -ENOSYS if not implemented.
1747 */
1748SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1749        struct io_event __user *, result)
1750{
1751    int (*cancel)(struct kiocb *iocb, struct io_event *res);
1752    struct kioctx *ctx;
1753    struct kiocb *kiocb;
1754    u32 key;
1755    int ret;
1756
1757    ret = get_user(key, &iocb->aio_key);
1758    if (unlikely(ret))
1759        return -EFAULT;
1760
1761    ctx = lookup_ioctx(ctx_id);
1762    if (unlikely(!ctx))
1763        return -EINVAL;
1764
1765    spin_lock_irq(&ctx->ctx_lock);
1766    ret = -EAGAIN;
1767    kiocb = lookup_kiocb(ctx, iocb, key);
1768    if (kiocb && kiocb->ki_cancel) {
1769        cancel = kiocb->ki_cancel;
1770        kiocb->ki_users ++;
1771        kiocbSetCancelled(kiocb);
1772    } else
1773        cancel = NULL;
1774    spin_unlock_irq(&ctx->ctx_lock);
1775
1776    if (NULL != cancel) {
1777        struct io_event tmp;
1778        pr_debug("calling cancel\n");
1779        memset(&tmp, 0, sizeof(tmp));
1780        tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1781        tmp.data = kiocb->ki_user_data;
1782        ret = cancel(kiocb, &tmp);
1783        if (!ret) {
1784            /* Cancellation succeeded -- copy the result
1785             * into the user's buffer.
1786             */
1787            if (copy_to_user(result, &tmp, sizeof(tmp)))
1788                ret = -EFAULT;
1789        }
1790    } else
1791        ret = -EINVAL;
1792
1793    put_ioctx(ctx);
1794
1795    return ret;
1796}
1797
1798/* io_getevents:
1799 * Attempts to read at least min_nr events and up to nr events from
1800 * the completion queue for the aio_context specified by ctx_id. May
1801 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1802 * if nr is out of range, if when is out of range. May fail with
1803 * -EFAULT if any of the memory specified to is invalid. May return
1804 * 0 or < min_nr if no events are available and the timeout specified
1805 * by when has elapsed, where when == NULL specifies an infinite
1806 * timeout. Note that the timeout pointed to by when is relative and
1807 * will be updated if not NULL and the operation blocks. Will fail
1808 * with -ENOSYS if not implemented.
1809 */
1810SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1811        long, min_nr,
1812        long, nr,
1813        struct io_event __user *, events,
1814        struct timespec __user *, timeout)
1815{
1816    struct kioctx *ioctx = lookup_ioctx(ctx_id);
1817    long ret = -EINVAL;
1818
1819    if (likely(ioctx)) {
1820        if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1821            ret = read_events(ioctx, min_nr, nr, events, timeout);
1822        put_ioctx(ioctx);
1823    }
1824
1825    asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1826    return ret;
1827}
1828
1829__initcall(aio_setup);
1830
1831EXPORT_SYMBOL(aio_complete);
1832EXPORT_SYMBOL(aio_put_req);
1833EXPORT_SYMBOL(wait_on_sync_kiocb);
1834

Archive Download this file



interactive