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/backing-dev.h>
19#include <linux/uio.h>
20
21#define DEBUG 0
22
23#include <linux/sched.h>
24#include <linux/fs.h>
25#include <linux/file.h>
26#include <linux/mm.h>
27#include <linux/mman.h>
28#include <linux/mmu_context.h>
29#include <linux/slab.h>
30#include <linux/timer.h>
31#include <linux/aio.h>
32#include <linux/highmem.h>
33#include <linux/workqueue.h>
34#include <linux/security.h>
35#include <linux/eventfd.h>
36#include <linux/blkdev.h>
37#include <linux/mempool.h>
38#include <linux/hash.h>
39
40#include <asm/kmap_types.h>
41#include <asm/uaccess.h>
42
43#if DEBUG > 1
44#define dprintk printk
45#else
46#define dprintk(x...) do { ; } while (0)
47#endif
48
49/*------ sysctl variables----*/
50static DEFINE_SPINLOCK(aio_nr_lock);
51unsigned long aio_nr; /* current system wide number of aio requests */
52unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
53/*----end sysctl variables---*/
54
55static struct kmem_cache *kiocb_cachep;
56static struct kmem_cache *kioctx_cachep;
57
58static struct workqueue_struct *aio_wq;
59
60/* Used for rare fput completion. */
61static void aio_fput_routine(struct work_struct *);
62static DECLARE_WORK(fput_work, aio_fput_routine);
63
64static DEFINE_SPINLOCK(fput_lock);
65static LIST_HEAD(fput_head);
66
67#define AIO_BATCH_HASH_BITS 3 /* allocated on-stack, so don't go crazy */
68#define AIO_BATCH_HASH_SIZE (1 << AIO_BATCH_HASH_BITS)
69struct aio_batch_entry {
70    struct hlist_node list;
71    struct address_space *mapping;
72};
73mempool_t *abe_pool;
74
75static void aio_kick_handler(struct work_struct *);
76static void aio_queue_work(struct kioctx *);
77
78/* aio_setup
79 * Creates the slab caches used by the aio routines, panic on
80 * failure as this is done early during the boot sequence.
81 */
82static int __init aio_setup(void)
83{
84    kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
85    kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
86
87    aio_wq = create_workqueue("aio");
88    abe_pool = mempool_create_kmalloc_pool(1, sizeof(struct aio_batch_entry));
89    BUG_ON(!abe_pool);
90
91    pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
92
93    return 0;
94}
95__initcall(aio_setup);
96
97static void aio_free_ring(struct kioctx *ctx)
98{
99    struct aio_ring_info *info = &ctx->ring_info;
100    long i;
101
102    for (i=0; i<info->nr_pages; i++)
103        put_page(info->ring_pages[i]);
104
105    if (info->mmap_size) {
106        down_write(&ctx->mm->mmap_sem);
107        do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
108        up_write(&ctx->mm->mmap_sem);
109    }
110
111    if (info->ring_pages && info->ring_pages != info->internal_pages)
112        kfree(info->ring_pages);
113    info->ring_pages = NULL;
114    info->nr = 0;
115}
116
117static int aio_setup_ring(struct kioctx *ctx)
118{
119    struct aio_ring *ring;
120    struct aio_ring_info *info = &ctx->ring_info;
121    unsigned nr_events = ctx->max_reqs;
122    unsigned long size;
123    int nr_pages;
124
125    /* Compensate for the ring buffer's head/tail overlap entry */
126    nr_events += 2; /* 1 is required, 2 for good luck */
127
128    size = sizeof(struct aio_ring);
129    size += sizeof(struct io_event) * nr_events;
130    nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
131
132    if (nr_pages < 0)
133        return -EINVAL;
134
135    nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
136
137    info->nr = 0;
138    info->ring_pages = info->internal_pages;
139    if (nr_pages > AIO_RING_PAGES) {
140        info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
141        if (!info->ring_pages)
142            return -ENOMEM;
143    }
144
145    info->mmap_size = nr_pages * PAGE_SIZE;
146    dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
147    down_write(&ctx->mm->mmap_sem);
148    info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
149                  PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
150                  0);
151    if (IS_ERR((void *)info->mmap_base)) {
152        up_write(&ctx->mm->mmap_sem);
153        info->mmap_size = 0;
154        aio_free_ring(ctx);
155        return -EAGAIN;
156    }
157
158    dprintk("mmap address: 0x%08lx\n", info->mmap_base);
159    info->nr_pages = get_user_pages(current, ctx->mm,
160                    info->mmap_base, nr_pages,
161                    1, 0, info->ring_pages, NULL);
162    up_write(&ctx->mm->mmap_sem);
163
164    if (unlikely(info->nr_pages != nr_pages)) {
165        aio_free_ring(ctx);
166        return -EAGAIN;
167    }
168
169    ctx->user_id = info->mmap_base;
170
171    info->nr = nr_events; /* trusted copy */
172
173    ring = kmap_atomic(info->ring_pages[0], KM_USER0);
174    ring->nr = nr_events; /* user copy */
175    ring->id = ctx->user_id;
176    ring->head = ring->tail = 0;
177    ring->magic = AIO_RING_MAGIC;
178    ring->compat_features = AIO_RING_COMPAT_FEATURES;
179    ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
180    ring->header_length = sizeof(struct aio_ring);
181    kunmap_atomic(ring, KM_USER0);
182
183    return 0;
184}
185
186
187/* aio_ring_event: returns a pointer to the event at the given index from
188 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
189 */
190#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
191#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
192#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
193
194#define aio_ring_event(info, nr, km) ({ \
195    unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
196    struct io_event *__event; \
197    __event = kmap_atomic( \
198            (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
199    __event += pos % AIO_EVENTS_PER_PAGE; \
200    __event; \
201})
202
203#define put_aio_ring_event(event, km) do { \
204    struct io_event *__event = (event); \
205    (void)__event; \
206    kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
207} while(0)
208
209static void ctx_rcu_free(struct rcu_head *head)
210{
211    struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
212    unsigned nr_events = ctx->max_reqs;
213
214    kmem_cache_free(kioctx_cachep, ctx);
215
216    if (nr_events) {
217        spin_lock(&aio_nr_lock);
218        BUG_ON(aio_nr - nr_events > aio_nr);
219        aio_nr -= nr_events;
220        spin_unlock(&aio_nr_lock);
221    }
222}
223
224/* __put_ioctx
225 * Called when the last user of an aio context has gone away,
226 * and the struct needs to be freed.
227 */
228static void __put_ioctx(struct kioctx *ctx)
229{
230    BUG_ON(ctx->reqs_active);
231
232    cancel_delayed_work(&ctx->wq);
233    cancel_work_sync(&ctx->wq.work);
234    aio_free_ring(ctx);
235    mmdrop(ctx->mm);
236    ctx->mm = NULL;
237    pr_debug("__put_ioctx: freeing %p\n", ctx);
238    call_rcu(&ctx->rcu_head, ctx_rcu_free);
239}
240
241#define get_ioctx(kioctx) do { \
242    BUG_ON(atomic_read(&(kioctx)->users) <= 0); \
243    atomic_inc(&(kioctx)->users); \
244} while (0)
245#define put_ioctx(kioctx) do { \
246    BUG_ON(atomic_read(&(kioctx)->users) <= 0); \
247    if (unlikely(atomic_dec_and_test(&(kioctx)->users))) \
248        __put_ioctx(kioctx); \
249} while (0)
250
251/* ioctx_alloc
252 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
253 */
254static struct kioctx *ioctx_alloc(unsigned nr_events)
255{
256    struct mm_struct *mm;
257    struct kioctx *ctx;
258    int did_sync = 0;
259
260    /* Prevent overflows */
261    if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
262        (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
263        pr_debug("ENOMEM: nr_events too high\n");
264        return ERR_PTR(-EINVAL);
265    }
266
267    if ((unsigned long)nr_events > aio_max_nr)
268        return ERR_PTR(-EAGAIN);
269
270    ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
271    if (!ctx)
272        return ERR_PTR(-ENOMEM);
273
274    ctx->max_reqs = nr_events;
275    mm = ctx->mm = current->mm;
276    atomic_inc(&mm->mm_count);
277
278    atomic_set(&ctx->users, 1);
279    spin_lock_init(&ctx->ctx_lock);
280    spin_lock_init(&ctx->ring_info.ring_lock);
281    init_waitqueue_head(&ctx->wait);
282
283    INIT_LIST_HEAD(&ctx->active_reqs);
284    INIT_LIST_HEAD(&ctx->run_list);
285    INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
286
287    if (aio_setup_ring(ctx) < 0)
288        goto out_freectx;
289
290    /* limit the number of system wide aios */
291    do {
292        spin_lock_bh(&aio_nr_lock);
293        if (aio_nr + nr_events > aio_max_nr ||
294            aio_nr + nr_events < aio_nr)
295            ctx->max_reqs = 0;
296        else
297            aio_nr += ctx->max_reqs;
298        spin_unlock_bh(&aio_nr_lock);
299        if (ctx->max_reqs || did_sync)
300            break;
301
302        /* wait for rcu callbacks to have completed before giving up */
303        synchronize_rcu();
304        did_sync = 1;
305        ctx->max_reqs = nr_events;
306    } while (1);
307
308    if (ctx->max_reqs == 0)
309        goto out_cleanup;
310
311    /* now link into global list. */
312    spin_lock(&mm->ioctx_lock);
313    hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
314    spin_unlock(&mm->ioctx_lock);
315
316    dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
317        ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
318    return ctx;
319
320out_cleanup:
321    __put_ioctx(ctx);
322    return ERR_PTR(-EAGAIN);
323
324out_freectx:
325    mmdrop(mm);
326    kmem_cache_free(kioctx_cachep, ctx);
327    ctx = ERR_PTR(-ENOMEM);
328
329    dprintk("aio: error allocating ioctx %p\n", ctx);
330    return ctx;
331}
332
333/* aio_cancel_all
334 * Cancels all outstanding aio requests on an aio context. Used
335 * when the processes owning a context have all exited to encourage
336 * the rapid destruction of the kioctx.
337 */
338static void aio_cancel_all(struct kioctx *ctx)
339{
340    int (*cancel)(struct kiocb *, struct io_event *);
341    struct io_event res;
342    spin_lock_irq(&ctx->ctx_lock);
343    ctx->dead = 1;
344    while (!list_empty(&ctx->active_reqs)) {
345        struct list_head *pos = ctx->active_reqs.next;
346        struct kiocb *iocb = list_kiocb(pos);
347        list_del_init(&iocb->ki_list);
348        cancel = iocb->ki_cancel;
349        kiocbSetCancelled(iocb);
350        if (cancel) {
351            iocb->ki_users++;
352            spin_unlock_irq(&ctx->ctx_lock);
353            cancel(iocb, &res);
354            spin_lock_irq(&ctx->ctx_lock);
355        }
356    }
357    spin_unlock_irq(&ctx->ctx_lock);
358}
359
360static void wait_for_all_aios(struct kioctx *ctx)
361{
362    struct task_struct *tsk = current;
363    DECLARE_WAITQUEUE(wait, tsk);
364
365    spin_lock_irq(&ctx->ctx_lock);
366    if (!ctx->reqs_active)
367        goto out;
368
369    add_wait_queue(&ctx->wait, &wait);
370    set_task_state(tsk, TASK_UNINTERRUPTIBLE);
371    while (ctx->reqs_active) {
372        spin_unlock_irq(&ctx->ctx_lock);
373        io_schedule();
374        set_task_state(tsk, TASK_UNINTERRUPTIBLE);
375        spin_lock_irq(&ctx->ctx_lock);
376    }
377    __set_task_state(tsk, TASK_RUNNING);
378    remove_wait_queue(&ctx->wait, &wait);
379
380out:
381    spin_unlock_irq(&ctx->ctx_lock);
382}
383
384/* wait_on_sync_kiocb:
385 * Waits on the given sync kiocb to complete.
386 */
387ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
388{
389    while (iocb->ki_users) {
390        set_current_state(TASK_UNINTERRUPTIBLE);
391        if (!iocb->ki_users)
392            break;
393        io_schedule();
394    }
395    __set_current_state(TASK_RUNNING);
396    return iocb->ki_user_data;
397}
398EXPORT_SYMBOL(wait_on_sync_kiocb);
399
400/* exit_aio: called when the last user of mm goes away. At this point,
401 * there is no way for any new requests to be submited or any of the
402 * io_* syscalls to be called on the context. However, there may be
403 * outstanding requests which hold references to the context; as they
404 * go away, they will call put_ioctx and release any pinned memory
405 * associated with the request (held via struct page * references).
406 */
407void exit_aio(struct mm_struct *mm)
408{
409    struct kioctx *ctx;
410
411    while (!hlist_empty(&mm->ioctx_list)) {
412        ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
413        hlist_del_rcu(&ctx->list);
414
415        aio_cancel_all(ctx);
416
417        wait_for_all_aios(ctx);
418        /*
419         * Ensure we don't leave the ctx on the aio_wq
420         */
421        cancel_work_sync(&ctx->wq.work);
422
423        if (1 != atomic_read(&ctx->users))
424            printk(KERN_DEBUG
425                "exit_aio:ioctx still alive: %d %d %d\n",
426                atomic_read(&ctx->users), ctx->dead,
427                ctx->reqs_active);
428        put_ioctx(ctx);
429    }
430}
431
432/* aio_get_req
433 * Allocate a slot for an aio request. Increments the users count
434 * of the kioctx so that the kioctx stays around until all requests are
435 * complete. Returns NULL if no requests are free.
436 *
437 * Returns with kiocb->users set to 2. The io submit code path holds
438 * an extra reference while submitting the i/o.
439 * This prevents races between the aio code path referencing the
440 * req (after submitting it) and aio_complete() freeing the req.
441 */
442static struct kiocb *__aio_get_req(struct kioctx *ctx)
443{
444    struct kiocb *req = NULL;
445    struct aio_ring *ring;
446    int okay = 0;
447
448    req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
449    if (unlikely(!req))
450        return NULL;
451
452    req->ki_flags = 0;
453    req->ki_users = 2;
454    req->ki_key = 0;
455    req->ki_ctx = ctx;
456    req->ki_cancel = NULL;
457    req->ki_retry = NULL;
458    req->ki_dtor = NULL;
459    req->private = NULL;
460    req->ki_iovec = NULL;
461    INIT_LIST_HEAD(&req->ki_run_list);
462    req->ki_eventfd = NULL;
463
464    /* Check if the completion queue has enough free space to
465     * accept an event from this io.
466     */
467    spin_lock_irq(&ctx->ctx_lock);
468    ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
469    if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
470        list_add(&req->ki_list, &ctx->active_reqs);
471        ctx->reqs_active++;
472        okay = 1;
473    }
474    kunmap_atomic(ring, KM_USER0);
475    spin_unlock_irq(&ctx->ctx_lock);
476
477    if (!okay) {
478        kmem_cache_free(kiocb_cachep, req);
479        req = NULL;
480    }
481
482    return req;
483}
484
485static inline struct kiocb *aio_get_req(struct kioctx *ctx)
486{
487    struct kiocb *req;
488    /* Handle a potential starvation case -- should be exceedingly rare as
489     * requests will be stuck on fput_head only if the aio_fput_routine is
490     * delayed and the requests were the last user of the struct file.
491     */
492    req = __aio_get_req(ctx);
493    if (unlikely(NULL == req)) {
494        aio_fput_routine(NULL);
495        req = __aio_get_req(ctx);
496    }
497    return req;
498}
499
500static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
501{
502    assert_spin_locked(&ctx->ctx_lock);
503
504    if (req->ki_eventfd != NULL)
505        eventfd_ctx_put(req->ki_eventfd);
506    if (req->ki_dtor)
507        req->ki_dtor(req);
508    if (req->ki_iovec != &req->ki_inline_vec)
509        kfree(req->ki_iovec);
510    kmem_cache_free(kiocb_cachep, req);
511    ctx->reqs_active--;
512
513    if (unlikely(!ctx->reqs_active && ctx->dead))
514        wake_up(&ctx->wait);
515}
516
517static void aio_fput_routine(struct work_struct *data)
518{
519    spin_lock_irq(&fput_lock);
520    while (likely(!list_empty(&fput_head))) {
521        struct kiocb *req = list_kiocb(fput_head.next);
522        struct kioctx *ctx = req->ki_ctx;
523
524        list_del(&req->ki_list);
525        spin_unlock_irq(&fput_lock);
526
527        /* Complete the fput(s) */
528        if (req->ki_filp != NULL)
529            __fput(req->ki_filp);
530
531        /* Link the iocb into the context's free list */
532        spin_lock_irq(&ctx->ctx_lock);
533        really_put_req(ctx, req);
534        spin_unlock_irq(&ctx->ctx_lock);
535
536        put_ioctx(ctx);
537        spin_lock_irq(&fput_lock);
538    }
539    spin_unlock_irq(&fput_lock);
540}
541
542/* __aio_put_req
543 * Returns true if this put was the last user of the request.
544 */
545static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
546{
547    dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
548        req, atomic_long_read(&req->ki_filp->f_count));
549
550    assert_spin_locked(&ctx->ctx_lock);
551
552    req->ki_users--;
553    BUG_ON(req->ki_users < 0);
554    if (likely(req->ki_users))
555        return 0;
556    list_del(&req->ki_list); /* remove from active_reqs */
557    req->ki_cancel = NULL;
558    req->ki_retry = NULL;
559
560    /*
561     * Try to optimize the aio and eventfd file* puts, by avoiding to
562     * schedule work in case it is not __fput() time. In normal cases,
563     * we would not be holding the last reference to the file*, so
564     * this function will be executed w/out any aio kthread wakeup.
565     */
566    if (unlikely(atomic_long_dec_and_test(&req->ki_filp->f_count))) {
567        get_ioctx(ctx);
568        spin_lock(&fput_lock);
569        list_add(&req->ki_list, &fput_head);
570        spin_unlock(&fput_lock);
571        queue_work(aio_wq, &fput_work);
572    } else {
573        req->ki_filp = NULL;
574        really_put_req(ctx, req);
575    }
576    return 1;
577}
578
579/* aio_put_req
580 * Returns true if this put was the last user of the kiocb,
581 * false if the request is still in use.
582 */
583int aio_put_req(struct kiocb *req)
584{
585    struct kioctx *ctx = req->ki_ctx;
586    int ret;
587    spin_lock_irq(&ctx->ctx_lock);
588    ret = __aio_put_req(ctx, req);
589    spin_unlock_irq(&ctx->ctx_lock);
590    return ret;
591}
592EXPORT_SYMBOL(aio_put_req);
593
594static struct kioctx *lookup_ioctx(unsigned long ctx_id)
595{
596    struct mm_struct *mm = current->mm;
597    struct kioctx *ctx, *ret = NULL;
598    struct hlist_node *n;
599
600    rcu_read_lock();
601
602    hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
603        if (ctx->user_id == ctx_id && !ctx->dead) {
604            get_ioctx(ctx);
605            ret = ctx;
606            break;
607        }
608    }
609
610    rcu_read_unlock();
611    return ret;
612}
613
614/*
615 * Queue up a kiocb to be retried. Assumes that the kiocb
616 * has already been marked as kicked, and places it on
617 * the retry run list for the corresponding ioctx, if it
618 * isn't already queued. Returns 1 if it actually queued
619 * the kiocb (to tell the caller to activate the work
620 * queue to process it), or 0, if it found that it was
621 * already queued.
622 */
623static inline int __queue_kicked_iocb(struct kiocb *iocb)
624{
625    struct kioctx *ctx = iocb->ki_ctx;
626
627    assert_spin_locked(&ctx->ctx_lock);
628
629    if (list_empty(&iocb->ki_run_list)) {
630        list_add_tail(&iocb->ki_run_list,
631            &ctx->run_list);
632        return 1;
633    }
634    return 0;
635}
636
637/* aio_run_iocb
638 * This is the core aio execution routine. It is
639 * invoked both for initial i/o submission and
640 * subsequent retries via the aio_kick_handler.
641 * Expects to be invoked with iocb->ki_ctx->lock
642 * already held. The lock is released and reacquired
643 * as needed during processing.
644 *
645 * Calls the iocb retry method (already setup for the
646 * iocb on initial submission) for operation specific
647 * handling, but takes care of most of common retry
648 * execution details for a given iocb. The retry method
649 * needs to be non-blocking as far as possible, to avoid
650 * holding up other iocbs waiting to be serviced by the
651 * retry kernel thread.
652 *
653 * The trickier parts in this code have to do with
654 * ensuring that only one retry instance is in progress
655 * for a given iocb at any time. Providing that guarantee
656 * simplifies the coding of individual aio operations as
657 * it avoids various potential races.
658 */
659static ssize_t aio_run_iocb(struct kiocb *iocb)
660{
661    struct kioctx *ctx = iocb->ki_ctx;
662    ssize_t (*retry)(struct kiocb *);
663    ssize_t ret;
664
665    if (!(retry = iocb->ki_retry)) {
666        printk("aio_run_iocb: iocb->ki_retry = NULL\n");
667        return 0;
668    }
669
670    /*
671     * We don't want the next retry iteration for this
672     * operation to start until this one has returned and
673     * updated the iocb state. However, wait_queue functions
674     * can trigger a kick_iocb from interrupt context in the
675     * meantime, indicating that data is available for the next
676     * iteration. We want to remember that and enable the
677     * next retry iteration _after_ we are through with
678     * this one.
679     *
680     * So, in order to be able to register a "kick", but
681     * prevent it from being queued now, we clear the kick
682     * flag, but make the kick code *think* that the iocb is
683     * still on the run list until we are actually done.
684     * When we are done with this iteration, we check if
685     * the iocb was kicked in the meantime and if so, queue
686     * it up afresh.
687     */
688
689    kiocbClearKicked(iocb);
690
691    /*
692     * This is so that aio_complete knows it doesn't need to
693     * pull the iocb off the run list (We can't just call
694     * INIT_LIST_HEAD because we don't want a kick_iocb to
695     * queue this on the run list yet)
696     */
697    iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
698    spin_unlock_irq(&ctx->ctx_lock);
699
700    /* Quit retrying if the i/o has been cancelled */
701    if (kiocbIsCancelled(iocb)) {
702        ret = -EINTR;
703        aio_complete(iocb, ret, 0);
704        /* must not access the iocb after this */
705        goto out;
706    }
707
708    /*
709     * Now we are all set to call the retry method in async
710     * context.
711     */
712    ret = retry(iocb);
713
714    if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED)
715        aio_complete(iocb, ret, 0);
716out:
717    spin_lock_irq(&ctx->ctx_lock);
718
719    if (-EIOCBRETRY == ret) {
720        /*
721         * OK, now that we are done with this iteration
722         * and know that there is more left to go,
723         * this is where we let go so that a subsequent
724         * "kick" can start the next iteration
725         */
726
727        /* will make __queue_kicked_iocb succeed from here on */
728        INIT_LIST_HEAD(&iocb->ki_run_list);
729        /* we must queue the next iteration ourselves, if it
730         * has already been kicked */
731        if (kiocbIsKicked(iocb)) {
732            __queue_kicked_iocb(iocb);
733
734            /*
735             * __queue_kicked_iocb will always return 1 here, because
736             * iocb->ki_run_list is empty at this point so it should
737             * be safe to unconditionally queue the context into the
738             * work queue.
739             */
740            aio_queue_work(ctx);
741        }
742    }
743    return ret;
744}
745
746/*
747 * __aio_run_iocbs:
748 * Process all pending retries queued on the ioctx
749 * run list.
750 * Assumes it is operating within the aio issuer's mm
751 * context.
752 */
753static int __aio_run_iocbs(struct kioctx *ctx)
754{
755    struct kiocb *iocb;
756    struct list_head run_list;
757
758    assert_spin_locked(&ctx->ctx_lock);
759
760    list_replace_init(&ctx->run_list, &run_list);
761    while (!list_empty(&run_list)) {
762        iocb = list_entry(run_list.next, struct kiocb,
763            ki_run_list);
764        list_del(&iocb->ki_run_list);
765        /*
766         * Hold an extra reference while retrying i/o.
767         */
768        iocb->ki_users++; /* grab extra reference */
769        aio_run_iocb(iocb);
770        __aio_put_req(ctx, iocb);
771     }
772    if (!list_empty(&ctx->run_list))
773        return 1;
774    return 0;
775}
776
777static void aio_queue_work(struct kioctx * ctx)
778{
779    unsigned long timeout;
780    /*
781     * if someone is waiting, get the work started right
782     * away, otherwise, use a longer delay
783     */
784    smp_mb();
785    if (waitqueue_active(&ctx->wait))
786        timeout = 1;
787    else
788        timeout = HZ/10;
789    queue_delayed_work(aio_wq, &ctx->wq, timeout);
790}
791
792
793/*
794 * aio_run_iocbs:
795 * Process all pending retries queued on the ioctx
796 * run list.
797 * Assumes it is operating within the aio issuer's mm
798 * context.
799 */
800static inline void aio_run_iocbs(struct kioctx *ctx)
801{
802    int requeue;
803
804    spin_lock_irq(&ctx->ctx_lock);
805
806    requeue = __aio_run_iocbs(ctx);
807    spin_unlock_irq(&ctx->ctx_lock);
808    if (requeue)
809        aio_queue_work(ctx);
810}
811
812/*
813 * just like aio_run_iocbs, but keeps running them until
814 * the list stays empty
815 */
816static inline void aio_run_all_iocbs(struct kioctx *ctx)
817{
818    spin_lock_irq(&ctx->ctx_lock);
819    while (__aio_run_iocbs(ctx))
820        ;
821    spin_unlock_irq(&ctx->ctx_lock);
822}
823
824/*
825 * aio_kick_handler:
826 * Work queue handler triggered to process pending
827 * retries on an ioctx. Takes on the aio issuer's
828 * mm context before running the iocbs, so that
829 * copy_xxx_user operates on the issuer's address
830 * space.
831 * Run on aiod's context.
832 */
833static void aio_kick_handler(struct work_struct *work)
834{
835    struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
836    mm_segment_t oldfs = get_fs();
837    struct mm_struct *mm;
838    int requeue;
839
840    set_fs(USER_DS);
841    use_mm(ctx->mm);
842    spin_lock_irq(&ctx->ctx_lock);
843    requeue =__aio_run_iocbs(ctx);
844    mm = ctx->mm;
845    spin_unlock_irq(&ctx->ctx_lock);
846     unuse_mm(mm);
847    set_fs(oldfs);
848    /*
849     * we're in a worker thread already, don't use queue_delayed_work,
850     */
851    if (requeue)
852        queue_delayed_work(aio_wq, &ctx->wq, 0);
853}
854
855
856/*
857 * Called by kick_iocb to queue the kiocb for retry
858 * and if required activate the aio work queue to process
859 * it
860 */
861static void try_queue_kicked_iocb(struct kiocb *iocb)
862{
863     struct kioctx *ctx = iocb->ki_ctx;
864    unsigned long flags;
865    int run = 0;
866
867    spin_lock_irqsave(&ctx->ctx_lock, flags);
868    /* set this inside the lock so that we can't race with aio_run_iocb()
869     * testing it and putting the iocb on the run list under the lock */
870    if (!kiocbTryKick(iocb))
871        run = __queue_kicked_iocb(iocb);
872    spin_unlock_irqrestore(&ctx->ctx_lock, flags);
873    if (run)
874        aio_queue_work(ctx);
875}
876
877/*
878 * kick_iocb:
879 * Called typically from a wait queue callback context
880 * to trigger a retry of the iocb.
881 * The retry is usually executed by aio workqueue
882 * threads (See aio_kick_handler).
883 */
884void kick_iocb(struct kiocb *iocb)
885{
886    /* sync iocbs are easy: they can only ever be executing from a
887     * single context. */
888    if (is_sync_kiocb(iocb)) {
889        kiocbSetKicked(iocb);
890            wake_up_process(iocb->ki_obj.tsk);
891        return;
892    }
893
894    try_queue_kicked_iocb(iocb);
895}
896EXPORT_SYMBOL(kick_iocb);
897
898/* aio_complete
899 * Called when the io request on the given iocb is complete.
900 * Returns true if this is the last user of the request. The
901 * only other user of the request can be the cancellation code.
902 */
903int aio_complete(struct kiocb *iocb, long res, long res2)
904{
905    struct kioctx *ctx = iocb->ki_ctx;
906    struct aio_ring_info *info;
907    struct aio_ring *ring;
908    struct io_event *event;
909    unsigned long flags;
910    unsigned long tail;
911    int ret;
912
913    /*
914     * Special case handling for sync iocbs:
915     * - events go directly into the iocb for fast handling
916     * - the sync task with the iocb in its stack holds the single iocb
917     * ref, no other paths have a way to get another ref
918     * - the sync task helpfully left a reference to itself in the iocb
919     */
920    if (is_sync_kiocb(iocb)) {
921        BUG_ON(iocb->ki_users != 1);
922        iocb->ki_user_data = res;
923        iocb->ki_users = 0;
924        wake_up_process(iocb->ki_obj.tsk);
925        return 1;
926    }
927
928    info = &ctx->ring_info;
929
930    /* add a completion event to the ring buffer.
931     * must be done holding ctx->ctx_lock to prevent
932     * other code from messing with the tail
933     * pointer since we might be called from irq
934     * context.
935     */
936    spin_lock_irqsave(&ctx->ctx_lock, flags);
937
938    if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
939        list_del_init(&iocb->ki_run_list);
940
941    /*
942     * cancelled requests don't get events, userland was given one
943     * when the event got cancelled.
944     */
945    if (kiocbIsCancelled(iocb))
946        goto put_rq;
947
948    ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
949
950    tail = info->tail;
951    event = aio_ring_event(info, tail, KM_IRQ0);
952    if (++tail >= info->nr)
953        tail = 0;
954
955    event->obj = (u64)(unsigned long)iocb->ki_obj.user;
956    event->data = iocb->ki_user_data;
957    event->res = res;
958    event->res2 = res2;
959
960    dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
961        ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
962        res, res2);
963
964    /* after flagging the request as done, we
965     * must never even look at it again
966     */
967    smp_wmb(); /* make event visible before updating tail */
968
969    info->tail = tail;
970    ring->tail = tail;
971
972    put_aio_ring_event(event, KM_IRQ0);
973    kunmap_atomic(ring, KM_IRQ1);
974
975    pr_debug("added to ring %p at [%lu]\n", iocb, tail);
976
977    /*
978     * Check if the user asked us to deliver the result through an
979     * eventfd. The eventfd_signal() function is safe to be called
980     * from IRQ context.
981     */
982    if (iocb->ki_eventfd != NULL)
983        eventfd_signal(iocb->ki_eventfd, 1);
984
985put_rq:
986    /* everything turned out well, dispose of the aiocb. */
987    ret = __aio_put_req(ctx, iocb);
988
989    /*
990     * We have to order our ring_info tail store above and test
991     * of the wait list below outside the wait lock. This is
992     * like in wake_up_bit() where clearing a bit has to be
993     * ordered with the unlocked test.
994     */
995    smp_mb();
996
997    if (waitqueue_active(&ctx->wait))
998        wake_up(&ctx->wait);
999
1000    spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1001    return ret;
1002}
1003EXPORT_SYMBOL(aio_complete);
1004
1005/* aio_read_evt
1006 * Pull an event off of the ioctx's event ring. Returns the number of
1007 * events fetched (0 or 1 ;-)
1008 * FIXME: make this use cmpxchg.
1009 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1010 */
1011static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1012{
1013    struct aio_ring_info *info = &ioctx->ring_info;
1014    struct aio_ring *ring;
1015    unsigned long head;
1016    int ret = 0;
1017
1018    ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1019    dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1020         (unsigned long)ring->head, (unsigned long)ring->tail,
1021         (unsigned long)ring->nr);
1022
1023    if (ring->head == ring->tail)
1024        goto out;
1025
1026    spin_lock(&info->ring_lock);
1027
1028    head = ring->head % info->nr;
1029    if (head != ring->tail) {
1030        struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1031        *ent = *evp;
1032        head = (head + 1) % info->nr;
1033        smp_mb(); /* finish reading the event before updatng the head */
1034        ring->head = head;
1035        ret = 1;
1036        put_aio_ring_event(evp, KM_USER1);
1037    }
1038    spin_unlock(&info->ring_lock);
1039
1040out:
1041    kunmap_atomic(ring, KM_USER0);
1042    dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1043         (unsigned long)ring->head, (unsigned long)ring->tail);
1044    return ret;
1045}
1046
1047struct aio_timeout {
1048    struct timer_list timer;
1049    int timed_out;
1050    struct task_struct *p;
1051};
1052
1053static void timeout_func(unsigned long data)
1054{
1055    struct aio_timeout *to = (struct aio_timeout *)data;
1056
1057    to->timed_out = 1;
1058    wake_up_process(to->p);
1059}
1060
1061static inline void init_timeout(struct aio_timeout *to)
1062{
1063    setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1064    to->timed_out = 0;
1065    to->p = current;
1066}
1067
1068static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1069                   const struct timespec *ts)
1070{
1071    to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1072    if (time_after(to->timer.expires, jiffies))
1073        add_timer(&to->timer);
1074    else
1075        to->timed_out = 1;
1076}
1077
1078static inline void clear_timeout(struct aio_timeout *to)
1079{
1080    del_singleshot_timer_sync(&to->timer);
1081}
1082
1083static int read_events(struct kioctx *ctx,
1084            long min_nr, long nr,
1085            struct io_event __user *event,
1086            struct timespec __user *timeout)
1087{
1088    long start_jiffies = jiffies;
1089    struct task_struct *tsk = current;
1090    DECLARE_WAITQUEUE(wait, tsk);
1091    int ret;
1092    int i = 0;
1093    struct io_event ent;
1094    struct aio_timeout to;
1095    int retry = 0;
1096
1097    /* needed to zero any padding within an entry (there shouldn't be
1098     * any, but C is fun!
1099     */
1100    memset(&ent, 0, sizeof(ent));
1101retry:
1102    ret = 0;
1103    while (likely(i < nr)) {
1104        ret = aio_read_evt(ctx, &ent);
1105        if (unlikely(ret <= 0))
1106            break;
1107
1108        dprintk("read event: %Lx %Lx %Lx %Lx\n",
1109            ent.data, ent.obj, ent.res, ent.res2);
1110
1111        /* Could we split the check in two? */
1112        ret = -EFAULT;
1113        if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1114            dprintk("aio: lost an event due to EFAULT.\n");
1115            break;
1116        }
1117        ret = 0;
1118
1119        /* Good, event copied to userland, update counts. */
1120        event ++;
1121        i ++;
1122    }
1123
1124    if (min_nr <= i)
1125        return i;
1126    if (ret)
1127        return ret;
1128
1129    /* End fast path */
1130
1131    /* racey check, but it gets redone */
1132    if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1133        retry = 1;
1134        aio_run_all_iocbs(ctx);
1135        goto retry;
1136    }
1137
1138    init_timeout(&to);
1139    if (timeout) {
1140        struct timespec ts;
1141        ret = -EFAULT;
1142        if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1143            goto out;
1144
1145        set_timeout(start_jiffies, &to, &ts);
1146    }
1147
1148    while (likely(i < nr)) {
1149        add_wait_queue_exclusive(&ctx->wait, &wait);
1150        do {
1151            set_task_state(tsk, TASK_INTERRUPTIBLE);
1152            ret = aio_read_evt(ctx, &ent);
1153            if (ret)
1154                break;
1155            if (min_nr <= i)
1156                break;
1157            if (unlikely(ctx->dead)) {
1158                ret = -EINVAL;
1159                break;
1160            }
1161            if (to.timed_out) /* Only check after read evt */
1162                break;
1163            /* Try to only show up in io wait if there are ops
1164             * in flight */
1165            if (ctx->reqs_active)
1166                io_schedule();
1167            else
1168                schedule();
1169            if (signal_pending(tsk)) {
1170                ret = -EINTR;
1171                break;
1172            }
1173            /*ret = aio_read_evt(ctx, &ent);*/
1174        } while (1) ;
1175
1176        set_task_state(tsk, TASK_RUNNING);
1177        remove_wait_queue(&ctx->wait, &wait);
1178
1179        if (unlikely(ret <= 0))
1180            break;
1181
1182        ret = -EFAULT;
1183        if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1184            dprintk("aio: lost an event due to EFAULT.\n");
1185            break;
1186        }
1187
1188        /* Good, event copied to userland, update counts. */
1189        event ++;
1190        i ++;
1191    }
1192
1193    if (timeout)
1194        clear_timeout(&to);
1195out:
1196    destroy_timer_on_stack(&to.timer);
1197    return i ? i : ret;
1198}
1199
1200/* Take an ioctx and remove it from the list of ioctx's. Protects
1201 * against races with itself via ->dead.
1202 */
1203static void io_destroy(struct kioctx *ioctx)
1204{
1205    struct mm_struct *mm = current->mm;
1206    int was_dead;
1207
1208    /* delete the entry from the list is someone else hasn't already */
1209    spin_lock(&mm->ioctx_lock);
1210    was_dead = ioctx->dead;
1211    ioctx->dead = 1;
1212    hlist_del_rcu(&ioctx->list);
1213    spin_unlock(&mm->ioctx_lock);
1214
1215    dprintk("aio_release(%p)\n", ioctx);
1216    if (likely(!was_dead))
1217        put_ioctx(ioctx); /* twice for the list */
1218
1219    aio_cancel_all(ioctx);
1220    wait_for_all_aios(ioctx);
1221
1222    /*
1223     * Wake up any waiters. The setting of ctx->dead must be seen
1224     * by other CPUs at this point. Right now, we rely on the
1225     * locking done by the above calls to ensure this consistency.
1226     */
1227    wake_up(&ioctx->wait);
1228    put_ioctx(ioctx); /* once for the lookup */
1229}
1230
1231/* sys_io_setup:
1232 * Create an aio_context capable of receiving at least nr_events.
1233 * ctxp must not point to an aio_context that already exists, and
1234 * must be initialized to 0 prior to the call. On successful
1235 * creation of the aio_context, *ctxp is filled in with the resulting
1236 * handle. May fail with -EINVAL if *ctxp is not initialized,
1237 * if the specified nr_events exceeds internal limits. May fail
1238 * with -EAGAIN if the specified nr_events exceeds the user's limit
1239 * of available events. May fail with -ENOMEM if insufficient kernel
1240 * resources are available. May fail with -EFAULT if an invalid
1241 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1242 * implemented.
1243 */
1244SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1245{
1246    struct kioctx *ioctx = NULL;
1247    unsigned long ctx;
1248    long ret;
1249
1250    ret = get_user(ctx, ctxp);
1251    if (unlikely(ret))
1252        goto out;
1253
1254    ret = -EINVAL;
1255    if (unlikely(ctx || nr_events == 0)) {
1256        pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1257                 ctx, nr_events);
1258        goto out;
1259    }
1260
1261    ioctx = ioctx_alloc(nr_events);
1262    ret = PTR_ERR(ioctx);
1263    if (!IS_ERR(ioctx)) {
1264        ret = put_user(ioctx->user_id, ctxp);
1265        if (!ret)
1266            return 0;
1267
1268        get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1269        io_destroy(ioctx);
1270    }
1271
1272out:
1273    return ret;
1274}
1275
1276/* sys_io_destroy:
1277 * Destroy the aio_context specified. May cancel any outstanding
1278 * AIOs and block on completion. Will fail with -ENOSYS if not
1279 * implemented. May fail with -EFAULT if the context pointed to
1280 * is invalid.
1281 */
1282SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1283{
1284    struct kioctx *ioctx = lookup_ioctx(ctx);
1285    if (likely(NULL != ioctx)) {
1286        io_destroy(ioctx);
1287        return 0;
1288    }
1289    pr_debug("EINVAL: io_destroy: invalid context id\n");
1290    return -EINVAL;
1291}
1292
1293static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1294{
1295    struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1296
1297    BUG_ON(ret <= 0);
1298
1299    while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1300        ssize_t this = min((ssize_t)iov->iov_len, ret);
1301        iov->iov_base += this;
1302        iov->iov_len -= this;
1303        iocb->ki_left -= this;
1304        ret -= this;
1305        if (iov->iov_len == 0) {
1306            iocb->ki_cur_seg++;
1307            iov++;
1308        }
1309    }
1310
1311    /* the caller should not have done more io than what fit in
1312     * the remaining iovecs */
1313    BUG_ON(ret > 0 && iocb->ki_left == 0);
1314}
1315
1316static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1317{
1318    struct file *file = iocb->ki_filp;
1319    struct address_space *mapping = file->f_mapping;
1320    struct inode *inode = mapping->host;
1321    ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1322             unsigned long, loff_t);
1323    ssize_t ret = 0;
1324    unsigned short opcode;
1325
1326    if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1327        (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1328        rw_op = file->f_op->aio_read;
1329        opcode = IOCB_CMD_PREADV;
1330    } else {
1331        rw_op = file->f_op->aio_write;
1332        opcode = IOCB_CMD_PWRITEV;
1333    }
1334
1335    /* This matches the pread()/pwrite() logic */
1336    if (iocb->ki_pos < 0)
1337        return -EINVAL;
1338
1339    do {
1340        ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1341                iocb->ki_nr_segs - iocb->ki_cur_seg,
1342                iocb->ki_pos);
1343        if (ret > 0)
1344            aio_advance_iovec(iocb, ret);
1345
1346    /* retry all partial writes. retry partial reads as long as its a
1347     * regular file. */
1348    } while (ret > 0 && iocb->ki_left > 0 &&
1349         (opcode == IOCB_CMD_PWRITEV ||
1350          (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1351
1352    /* This means we must have transferred all that we could */
1353    /* No need to retry anymore */
1354    if ((ret == 0) || (iocb->ki_left == 0))
1355        ret = iocb->ki_nbytes - iocb->ki_left;
1356
1357    /* If we managed to write some out we return that, rather than
1358     * the eventual error. */
1359    if (opcode == IOCB_CMD_PWRITEV
1360        && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1361        && iocb->ki_nbytes - iocb->ki_left)
1362        ret = iocb->ki_nbytes - iocb->ki_left;
1363
1364    return ret;
1365}
1366
1367static ssize_t aio_fdsync(struct kiocb *iocb)
1368{
1369    struct file *file = iocb->ki_filp;
1370    ssize_t ret = -EINVAL;
1371
1372    if (file->f_op->aio_fsync)
1373        ret = file->f_op->aio_fsync(iocb, 1);
1374    return ret;
1375}
1376
1377static ssize_t aio_fsync(struct kiocb *iocb)
1378{
1379    struct file *file = iocb->ki_filp;
1380    ssize_t ret = -EINVAL;
1381
1382    if (file->f_op->aio_fsync)
1383        ret = file->f_op->aio_fsync(iocb, 0);
1384    return ret;
1385}
1386
1387static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb)
1388{
1389    ssize_t ret;
1390
1391    ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf,
1392                    kiocb->ki_nbytes, 1,
1393                    &kiocb->ki_inline_vec, &kiocb->ki_iovec);
1394    if (ret < 0)
1395        goto out;
1396
1397    kiocb->ki_nr_segs = kiocb->ki_nbytes;
1398    kiocb->ki_cur_seg = 0;
1399    /* ki_nbytes/left now reflect bytes instead of segs */
1400    kiocb->ki_nbytes = ret;
1401    kiocb->ki_left = ret;
1402
1403    ret = 0;
1404out:
1405    return ret;
1406}
1407
1408static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1409{
1410    kiocb->ki_iovec = &kiocb->ki_inline_vec;
1411    kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1412    kiocb->ki_iovec->iov_len = kiocb->ki_left;
1413    kiocb->ki_nr_segs = 1;
1414    kiocb->ki_cur_seg = 0;
1415    return 0;
1416}
1417
1418/*
1419 * aio_setup_iocb:
1420 * Performs the initial checks and aio retry method
1421 * setup for the kiocb at the time of io submission.
1422 */
1423static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1424{
1425    struct file *file = kiocb->ki_filp;
1426    ssize_t ret = 0;
1427
1428    switch (kiocb->ki_opcode) {
1429    case IOCB_CMD_PREAD:
1430        ret = -EBADF;
1431        if (unlikely(!(file->f_mode & FMODE_READ)))
1432            break;
1433        ret = -EFAULT;
1434        if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1435            kiocb->ki_left)))
1436            break;
1437        ret = security_file_permission(file, MAY_READ);
1438        if (unlikely(ret))
1439            break;
1440        ret = aio_setup_single_vector(kiocb);
1441        if (ret)
1442            break;
1443        ret = -EINVAL;
1444        if (file->f_op->aio_read)
1445            kiocb->ki_retry = aio_rw_vect_retry;
1446        break;
1447    case IOCB_CMD_PWRITE:
1448        ret = -EBADF;
1449        if (unlikely(!(file->f_mode & FMODE_WRITE)))
1450            break;
1451        ret = -EFAULT;
1452        if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1453            kiocb->ki_left)))
1454            break;
1455        ret = security_file_permission(file, MAY_WRITE);
1456        if (unlikely(ret))
1457            break;
1458        ret = aio_setup_single_vector(kiocb);
1459        if (ret)
1460            break;
1461        ret = -EINVAL;
1462        if (file->f_op->aio_write)
1463            kiocb->ki_retry = aio_rw_vect_retry;
1464        break;
1465    case IOCB_CMD_PREADV:
1466        ret = -EBADF;
1467        if (unlikely(!(file->f_mode & FMODE_READ)))
1468            break;
1469        ret = security_file_permission(file, MAY_READ);
1470        if (unlikely(ret))
1471            break;
1472        ret = aio_setup_vectored_rw(READ, kiocb);
1473        if (ret)
1474            break;
1475        ret = -EINVAL;
1476        if (file->f_op->aio_read)
1477            kiocb->ki_retry = aio_rw_vect_retry;
1478        break;
1479    case IOCB_CMD_PWRITEV:
1480        ret = -EBADF;
1481        if (unlikely(!(file->f_mode & FMODE_WRITE)))
1482            break;
1483        ret = security_file_permission(file, MAY_WRITE);
1484        if (unlikely(ret))
1485            break;
1486        ret = aio_setup_vectored_rw(WRITE, kiocb);
1487        if (ret)
1488            break;
1489        ret = -EINVAL;
1490        if (file->f_op->aio_write)
1491            kiocb->ki_retry = aio_rw_vect_retry;
1492        break;
1493    case IOCB_CMD_FDSYNC:
1494        ret = -EINVAL;
1495        if (file->f_op->aio_fsync)
1496            kiocb->ki_retry = aio_fdsync;
1497        break;
1498    case IOCB_CMD_FSYNC:
1499        ret = -EINVAL;
1500        if (file->f_op->aio_fsync)
1501            kiocb->ki_retry = aio_fsync;
1502        break;
1503    default:
1504        dprintk("EINVAL: io_submit: no operation provided\n");
1505        ret = -EINVAL;
1506    }
1507
1508    if (!kiocb->ki_retry)
1509        return ret;
1510
1511    return 0;
1512}
1513
1514static void aio_batch_add(struct address_space *mapping,
1515              struct hlist_head *batch_hash)
1516{
1517    struct aio_batch_entry *abe;
1518    struct hlist_node *pos;
1519    unsigned bucket;
1520
1521    bucket = hash_ptr(mapping, AIO_BATCH_HASH_BITS);
1522    hlist_for_each_entry(abe, pos, &batch_hash[bucket], list) {
1523        if (abe->mapping == mapping)
1524            return;
1525    }
1526
1527    abe = mempool_alloc(abe_pool, GFP_KERNEL);
1528    BUG_ON(!igrab(mapping->host));
1529    abe->mapping = mapping;
1530    hlist_add_head(&abe->list, &batch_hash[bucket]);
1531    return;
1532}
1533
1534static void aio_batch_free(struct hlist_head *batch_hash)
1535{
1536    struct aio_batch_entry *abe;
1537    struct hlist_node *pos, *n;
1538    int i;
1539
1540    for (i = 0; i < AIO_BATCH_HASH_SIZE; i++) {
1541        hlist_for_each_entry_safe(abe, pos, n, &batch_hash[i], list) {
1542            blk_run_address_space(abe->mapping);
1543            iput(abe->mapping->host);
1544            hlist_del(&abe->list);
1545            mempool_free(abe, abe_pool);
1546        }
1547    }
1548}
1549
1550static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1551             struct iocb *iocb, struct hlist_head *batch_hash)
1552{
1553    struct kiocb *req;
1554    struct file *file;
1555    ssize_t ret;
1556
1557    /* enforce forwards compatibility on users */
1558    if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1559        pr_debug("EINVAL: io_submit: reserve field set\n");
1560        return -EINVAL;
1561    }
1562
1563    /* prevent overflows */
1564    if (unlikely(
1565        (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1566        (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1567        ((ssize_t)iocb->aio_nbytes < 0)
1568       )) {
1569        pr_debug("EINVAL: io_submit: overflow check\n");
1570        return -EINVAL;
1571    }
1572
1573    file = fget(iocb->aio_fildes);
1574    if (unlikely(!file))
1575        return -EBADF;
1576
1577    req = aio_get_req(ctx); /* returns with 2 references to req */
1578    if (unlikely(!req)) {
1579        fput(file);
1580        return -EAGAIN;
1581    }
1582    req->ki_filp = file;
1583    if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1584        /*
1585         * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1586         * instance of the file* now. The file descriptor must be
1587         * an eventfd() fd, and will be signaled for each completed
1588         * event using the eventfd_signal() function.
1589         */
1590        req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1591        if (IS_ERR(req->ki_eventfd)) {
1592            ret = PTR_ERR(req->ki_eventfd);
1593            req->ki_eventfd = NULL;
1594            goto out_put_req;
1595        }
1596    }
1597
1598    ret = put_user(req->ki_key, &user_iocb->aio_key);
1599    if (unlikely(ret)) {
1600        dprintk("EFAULT: aio_key\n");
1601        goto out_put_req;
1602    }
1603
1604    req->ki_obj.user = user_iocb;
1605    req->ki_user_data = iocb->aio_data;
1606    req->ki_pos = iocb->aio_offset;
1607
1608    req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1609    req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1610    req->ki_opcode = iocb->aio_lio_opcode;
1611
1612    ret = aio_setup_iocb(req);
1613
1614    if (ret)
1615        goto out_put_req;
1616
1617    spin_lock_irq(&ctx->ctx_lock);
1618    aio_run_iocb(req);
1619    if (!list_empty(&ctx->run_list)) {
1620        /* drain the run list */
1621        while (__aio_run_iocbs(ctx))
1622            ;
1623    }
1624    spin_unlock_irq(&ctx->ctx_lock);
1625    if (req->ki_opcode == IOCB_CMD_PREAD ||
1626        req->ki_opcode == IOCB_CMD_PREADV ||
1627        req->ki_opcode == IOCB_CMD_PWRITE ||
1628        req->ki_opcode == IOCB_CMD_PWRITEV)
1629        aio_batch_add(file->f_mapping, batch_hash);
1630
1631    aio_put_req(req); /* drop extra ref to req */
1632    return 0;
1633
1634out_put_req:
1635    aio_put_req(req); /* drop extra ref to req */
1636    aio_put_req(req); /* drop i/o ref to req */
1637    return ret;
1638}
1639
1640/* sys_io_submit:
1641 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1642 * the number of iocbs queued. May return -EINVAL if the aio_context
1643 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1644 * *iocbpp[0] is not properly initialized, if the operation specified
1645 * is invalid for the file descriptor in the iocb. May fail with
1646 * -EFAULT if any of the data structures point to invalid data. May
1647 * fail with -EBADF if the file descriptor specified in the first
1648 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1649 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1650 * fail with -ENOSYS if not implemented.
1651 */
1652SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1653        struct iocb __user * __user *, iocbpp)
1654{
1655    struct kioctx *ctx;
1656    long ret = 0;
1657    int i;
1658    struct hlist_head batch_hash[AIO_BATCH_HASH_SIZE] = { { 0, }, };
1659
1660    if (unlikely(nr < 0))
1661        return -EINVAL;
1662
1663    if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1664        return -EFAULT;
1665
1666    ctx = lookup_ioctx(ctx_id);
1667    if (unlikely(!ctx)) {
1668        pr_debug("EINVAL: io_submit: invalid context id\n");
1669        return -EINVAL;
1670    }
1671
1672    /*
1673     * AKPM: should this return a partial result if some of the IOs were
1674     * successfully submitted?
1675     */
1676    for (i=0; i<nr; i++) {
1677        struct iocb __user *user_iocb;
1678        struct iocb tmp;
1679
1680        if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1681            ret = -EFAULT;
1682            break;
1683        }
1684
1685        if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1686            ret = -EFAULT;
1687            break;
1688        }
1689
1690        ret = io_submit_one(ctx, user_iocb, &tmp, batch_hash);
1691        if (ret)
1692            break;
1693    }
1694    aio_batch_free(batch_hash);
1695
1696    put_ioctx(ctx);
1697    return i ? i : ret;
1698}
1699
1700/* lookup_kiocb
1701 * Finds a given iocb for cancellation.
1702 */
1703static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1704                  u32 key)
1705{
1706    struct list_head *pos;
1707
1708    assert_spin_locked(&ctx->ctx_lock);
1709
1710    /* TODO: use a hash or array, this sucks. */
1711    list_for_each(pos, &ctx->active_reqs) {
1712        struct kiocb *kiocb = list_kiocb(pos);
1713        if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1714            return kiocb;
1715    }
1716    return NULL;
1717}
1718
1719/* sys_io_cancel:
1720 * Attempts to cancel an iocb previously passed to io_submit. If
1721 * the operation is successfully cancelled, the resulting event is
1722 * copied into the memory pointed to by result without being placed
1723 * into the completion queue and 0 is returned. May fail with
1724 * -EFAULT if any of the data structures pointed to are invalid.
1725 * May fail with -EINVAL if aio_context specified by ctx_id is
1726 * invalid. May fail with -EAGAIN if the iocb specified was not
1727 * cancelled. Will fail with -ENOSYS if not implemented.
1728 */
1729SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1730        struct io_event __user *, result)
1731{
1732    int (*cancel)(struct kiocb *iocb, struct io_event *res);
1733    struct kioctx *ctx;
1734    struct kiocb *kiocb;
1735    u32 key;
1736    int ret;
1737
1738    ret = get_user(key, &iocb->aio_key);
1739    if (unlikely(ret))
1740        return -EFAULT;
1741
1742    ctx = lookup_ioctx(ctx_id);
1743    if (unlikely(!ctx))
1744        return -EINVAL;
1745
1746    spin_lock_irq(&ctx->ctx_lock);
1747    ret = -EAGAIN;
1748    kiocb = lookup_kiocb(ctx, iocb, key);
1749    if (kiocb && kiocb->ki_cancel) {
1750        cancel = kiocb->ki_cancel;
1751        kiocb->ki_users ++;
1752        kiocbSetCancelled(kiocb);
1753    } else
1754        cancel = NULL;
1755    spin_unlock_irq(&ctx->ctx_lock);
1756
1757    if (NULL != cancel) {
1758        struct io_event tmp;
1759        pr_debug("calling cancel\n");
1760        memset(&tmp, 0, sizeof(tmp));
1761        tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1762        tmp.data = kiocb->ki_user_data;
1763        ret = cancel(kiocb, &tmp);
1764        if (!ret) {
1765            /* Cancellation succeeded -- copy the result
1766             * into the user's buffer.
1767             */
1768            if (copy_to_user(result, &tmp, sizeof(tmp)))
1769                ret = -EFAULT;
1770        }
1771    } else
1772        ret = -EINVAL;
1773
1774    put_ioctx(ctx);
1775
1776    return ret;
1777}
1778
1779/* io_getevents:
1780 * Attempts to read at least min_nr events and up to nr events from
1781 * the completion queue for the aio_context specified by ctx_id. May
1782 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1783 * if nr is out of range, if when is out of range. May fail with
1784 * -EFAULT if any of the memory specified to is invalid. May return
1785 * 0 or < min_nr if no events are available and the timeout specified
1786 * by when has elapsed, where when == NULL specifies an infinite
1787 * timeout. Note that the timeout pointed to by when is relative and
1788 * will be updated if not NULL and the operation blocks. Will fail
1789 * with -ENOSYS if not implemented.
1790 */
1791SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1792        long, min_nr,
1793        long, nr,
1794        struct io_event __user *, events,
1795        struct timespec __user *, timeout)
1796{
1797    struct kioctx *ioctx = lookup_ioctx(ctx_id);
1798    long ret = -EINVAL;
1799
1800    if (likely(ioctx)) {
1801        if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1802            ret = read_events(ioctx, min_nr, nr, events, timeout);
1803        put_ioctx(ioctx);
1804    }
1805
1806    asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1807    return ret;
1808}
1809

Archive Download this file



interactive