blktrace: fix possible memleak in '__blk_trace_remove'
authorYe Bin <yebin10@huawei.com>
Wed, 19 Oct 2022 03:36:01 +0000 (11:36 +0800)
committerJens Axboe <axboe@kernel.dk>
Thu, 20 Oct 2022 13:02:52 +0000 (06:02 -0700)
When test as follows:
step1: ioctl(sda, BLKTRACESETUP, &arg)
step2: ioctl(sda, BLKTRACESTART, NULL)
step3: ioctl(sda, BLKTRACETEARDOWN, NULL)
step4: ioctl(sda, BLKTRACESETUP, &arg)
Got issue as follows:
debugfs: File 'dropped' in directory 'sda' already present!
debugfs: File 'msg' in directory 'sda' already present!
debugfs: File 'trace0' in directory 'sda' already present!

And also find syzkaller report issue like "KASAN: use-after-free Read in relay_switch_subbuf"
"https://syzkaller.appspot.com/bug?id=13849f0d9b1b818b087341691be6cc3ac6a6bfb7"

If remove block trace without stop(BLKTRACESTOP) block trace, '__blk_trace_remove'
will just set 'q->blk_trace' with NULL. However, debugfs file isn't removed, so
will report file already present when call BLKTRACESETUP.
static int __blk_trace_remove(struct request_queue *q)
{
        struct blk_trace *bt;

        bt = rcu_replace_pointer(q->blk_trace, NULL,
                                 lockdep_is_held(&q->debugfs_mutex));
        if (!bt)
                return -EINVAL;

if (bt->trace_state != Blktrace_running)
         blk_trace_cleanup(q, bt);

        return 0;
}

If do test as follows:
step1: ioctl(sda, BLKTRACESETUP, &arg)
step2: ioctl(sda, BLKTRACESTART, NULL)
step3: ioctl(sda, BLKTRACETEARDOWN, NULL)
step4: remove sda

There will remove debugfs directory which will remove recursively all file
under directory.
>> blk_release_queue
>> debugfs_remove_recursive(q->debugfs_dir)
So all files which created in 'do_blk_trace_setup' are removed, and
'dentry->d_inode' is NULL. But 'q->blk_trace' is still in 'running_trace_lock',
'trace_note_tsk' will traverse 'running_trace_lock' all nodes.
>>trace_note_tsk
>>  trace_note
>>    relay_reserve
>>       relay_switch_subbuf
>>        d_inode(buf->dentry)->i_size

To solve above issues, reference commit '5afedf670caf', call 'blk_trace_cleanup'
unconditionally in '__blk_trace_remove' and first stop block trace in
'blk_trace_cleanup'.

Signed-off-by: Ye Bin <yebin10@huawei.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221019033602.752383-3-yebin@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
kernel/trace/blktrace.c

index 50b6f241b5f778b7ee39fca48d56ff065e8fcb33..e17bba027a2c6edd4d45569bffb792b321d4deb5 100644 (file)
@@ -379,6 +379,7 @@ static int blk_trace_stop(struct blk_trace *bt)
 
 static void blk_trace_cleanup(struct request_queue *q, struct blk_trace *bt)
 {
+       blk_trace_stop(bt);
        synchronize_rcu();
        blk_trace_free(q, bt);
        put_probe_ref();
@@ -393,8 +394,7 @@ static int __blk_trace_remove(struct request_queue *q)
        if (!bt)
                return -EINVAL;
 
-       if (bt->trace_state != Blktrace_running)
-               blk_trace_cleanup(q, bt);
+       blk_trace_cleanup(q, bt);
 
        return 0;
 }