lkml.org 
[lkml]   [2021]   [Oct]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH -next] block: fix UAF from race of ioc_release_fn() and __ioc_clear_queue()
From
Date
On 10/12/21 2:30 AM, Laibin Qiu wrote:
> KASAN reports a use-after-free report when doing block test:
>
> [293762.535116]
> ==================================================================
> [293762.535129] BUG: KASAN: use-after-free in
> queued_spin_lock_slowpath+0x78/0x4c8
> [293762.535133] Write of size 2 at addr ffff8000d5f12bc8 by task sh/9148
> [293762.535135]
> [293762.535145] CPU: 1 PID: 9148 Comm: sh Kdump: loaded Tainted: G W
> 4.19.90-vhulk2108.6.0.h824.kasan.eulerosv2r10.aarch64 #1
> [293762.535148] Hardware name: QEMU KVM Virtual Machine, BIOS 0.0.0
> 02/06/2015
> [293762.535150] Call trace:
> [293762.535154] dump_backtrace+0x0/0x310
> [293762.535158] show_stack+0x28/0x38
> [293762.535165] dump_stack+0xec/0x15c
> [293762.535172] print_address_description+0x68/0x2d0
> [293762.535177] kasan_report+0x130/0x2f0
> [293762.535182] __asan_store2+0x80/0xa8
> [293762.535189] queued_spin_lock_slowpath+0x78/0x4c8
> [293762.535194] __ioc_clear_queue+0x158/0x160
> [293762.535198] ioc_clear_queue+0x194/0x258
> [293762.535202] elevator_switch_mq+0x64/0x170
> [293762.535206] elevator_switch+0x140/0x270
> [293762.535211] elv_iosched_store+0x1a4/0x2a0
> [293762.535215] queue_attr_store+0x90/0xe0
> [293762.535219] sysfs_kf_write+0xa8/0xe8
> [293762.535222] kernfs_fop_write+0x1f8/0x378
> [293762.535227] __vfs_write+0xe0/0x360
> [293762.535233] vfs_write+0xf0/0x270
> [293762.535237] ksys_write+0xdc/0x1b8
> [293762.535241] __arm64_sys_write+0x50/0x60
> [293762.535245] el0_svc_common+0xc8/0x320
> [293762.535250] el0_svc_handler+0xf8/0x160
> [293762.535253] el0_svc+0x10/0x218
> [293762.535254]
> [293762.535258] Allocated by task 3466763:
> [293762.535264] kasan_kmalloc+0xe0/0x190
> [293762.535269] kasan_slab_alloc+0x14/0x20
> [293762.535276] kmem_cache_alloc_node+0x1b4/0x420
> [293762.535280] create_task_io_context+0x40/0x210
> [293762.535284] generic_make_request_checks+0xc78/0xe38
> [293762.535288] generic_make_request+0xf8/0x640
> [293762.535394] generic_file_direct_write+0x100/0x268
> [293762.535401] __generic_file_write_iter+0x128/0x370
> [293762.535467] vfs_iter_write+0x64/0x90
> [293762.535489] ovl_write_iter+0x2f8/0x458 [overlay]
> [293762.535493] __vfs_write+0x264/0x360
> [293762.535497] vfs_write+0xf0/0x270
> [293762.535501] ksys_write+0xdc/0x1b8
> [293762.535505] __arm64_sys_write+0x50/0x60
> [293762.535509] el0_svc_common+0xc8/0x320
> [293762.535387] ext4_direct_IO+0x3c8/0xe80 [ext4]
> [293762.535394] generic_file_direct_write+0x100/0x268
> [293762.535401] __generic_file_write_iter+0x128/0x370
> [293762.535452] ext4_file_write_iter+0x610/0xa80 [ext4]
> [293762.535457] do_iter_readv_writev+0x28c/0x390
> [293762.535463] do_iter_write+0xfc/0x360
> [293762.535467] vfs_iter_write+0x64/0x90
> [293762.535489] ovl_write_iter+0x2f8/0x458 [overlay]
> [293762.535493] __vfs_write+0x264/0x360
> [293762.535497] vfs_write+0xf0/0x270
> [293762.535501] ksys_write+0xdc/0x1b8
> [293762.535505] __arm64_sys_write+0x50/0x60
> [293762.535509] el0_svc_common+0xc8/0x320
> [293762.535513] el0_svc_handler+0xf8/0x160
> [293762.535517] el0_svc+0x10/0x218
> [293762.535521]
> [293762.535523] Freed by task 3466763:
> [293762.535528] __kasan_slab_free+0x120/0x228
> [293762.535532] kasan_slab_free+0x10/0x18
> [293762.535536] kmem_cache_free+0x68/0x248
> [293762.535540] put_io_context+0x104/0x190
> [293762.535545] put_io_context_active+0x204/0x2c8
> [293762.535549] exit_io_context+0x74/0xa0
> [293762.535553] do_exit+0x658/0xae0
> [293762.535557] do_group_exit+0x74/0x1a8
> [293762.535561] get_signal+0x21c/0xf38
> [293762.535564] do_signal+0x10c/0x450
> [293762.535568] do_notify_resume+0x224/0x4b0
> [293762.535573] work_pending+0x8/0x10
> [293762.535574]
> [293762.535578] The buggy address belongs to the object at
> ffff8000d5f12bb8
> which belongs to the cache blkdev_ioc of size 136
> [293762.535582] The buggy address is located 16 bytes inside of
> 136-byte region [ffff8000d5f12bb8, ffff8000d5f12c40)
> [293762.535583] The buggy address belongs to the page:
> [293762.535588] page:ffff7e000357c480 count:1 mapcount:0
> mapping:ffff8000d8563c00 index:0x0
> [293762.536201] flags: 0x7ffff0000000100(slab)
> [293762.536540] raw: 07ffff0000000100 ffff7e0003118588 ffff8000d8adb530
> ffff8000d8563c00
> [293762.536546] raw: 0000000000000000 0000000000140014 00000001ffffffff
> 0000000000000000
> [293762.536551] page dumped because: kasan: bad access detected
> [293762.536552]
> [293762.536554] Memory state around the buggy address:
> [293762.536558] ffff8000d5f12a80: 00 00 00 00 00 00 fc fc fc fc fc fc fc
> fc fb fb
> [293762.536562] ffff8000d5f12b00: fb fb fb fb fb fb fb fb fb fb fb fb fb
> fb fb fc
> [293762.536566] >ffff8000d5f12b80: fc fc fc fc fc fc fc fb fb fb fb fb
> fb fb fb fb
> [293762.536568] ^
> [293762.536572] ffff8000d5f12c00: fb fb fb fb fb fb fb fb fc fc fc fc fc
> fc fc fc
> [293762.536576] ffff8000d5f12c80: 00 00 00 00 00 00 00 00 00 00 00 00 00
> 00 00 00
>
> [293762.536577]
> ==================================================================
>
> ioc_release_fn() will destroy icq from ioc->icq_list and
> __ioc_clear_queue() will destroy icq from request_queue->icq_list.
> However, the ioc_release_fn() will hold ioc_lock firstly, and
> free ioc finally. Then __ioc_clear_queue() will get ioc from icq
> and hold ioc_lock, but ioc has been released, which will result
> in a use-after-free.
>
> CPU0 CPU1
> put_io_context elevator_switch_mq
> queue_work &ioc->release_work ioc_clear_queue
> ^^^ splice q->icq_list
> __ioc_clear_queue
> ^^^get icq from icq_list
> get ioc from icq->ioc
> ioc_release_fn
> spin_lock(ioc->lock)
> ioc_destroy_icq(icq)
> spin_unlock(ioc->lock)
> free(ioc)
> spin_lock(ioc->lock) <= UAF
>
> Fix by grabbing the request_queue->queue_lock in ioc_clear_queue() to
> avoid this race scene.
>
> Signed-off-by: Laibin Qiu <qiulaibin@huawei.com>
> ---
> block/blk-ioc.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/block/blk-ioc.c b/block/blk-ioc.c
> index 57299f860d41..1d6ba8ff5a66 100644
> --- a/block/blk-ioc.c
> +++ b/block/blk-ioc.c
> @@ -242,9 +242,9 @@ void ioc_clear_queue(struct request_queue *q)
>
> spin_lock_irq(&q->queue_lock);
> list_splice_init(&q->icq_list, &icq_list);
> - spin_unlock_irq(&q->queue_lock);
>
> __ioc_clear_queue(&icq_list);
> + spin_unlock_irq(&q->queue_lock);
> }

This looks good to me - but I'd change __ioc_clear_queue() to use
spin_lock() instead of spin_lock_irqsave() as well at the same time, as
we just nest inside this call.


--
Jens Axboe

\
 
 \ /
  Last update: 2021-10-12 21:55    [W:1.208 / U:0.368 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site