[PATCH V3 10/18] coresight: tmc: getting the right read_count on tmc_open()

Mathieu Poirier mathieu.poirier at linaro.org
Mon Apr 25 07:25:12 PDT 2016


On 25 April 2016 at 04:47, Suzuki K Poulose <Suzuki.Poulose at arm.com> wrote:
> On 22/04/16 18:14, Mathieu Poirier wrote:
>>
>> In function tmc_open(), if tmc_read_prepare() fails variable
>> drvdata->read_count is not decremented, causing unwanted
>> access to drvdata->buf and very likely, a crash dump.
>>
>> By moving the incrementation to a place where we know things
>> are stable this kind of situation is avoided.
>>
>> Signed-off-by: Mathieu Poirier <mathieu.poirier at linaro.org>
>> Reviewed-by: Suzuki K Poulose <Suzuki.Poulose at arm.com>
>> ---
>>   drivers/hwtracing/coresight/coresight-tmc.c | 3 ++-
>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/hwtracing/coresight/coresight-tmc.c
>> b/drivers/hwtracing/coresight/coresight-tmc.c
>> index e8e12a9b917a..55806352b1f1 100644
>> --- a/drivers/hwtracing/coresight/coresight-tmc.c
>> +++ b/drivers/hwtracing/coresight/coresight-tmc.c
>> @@ -121,13 +121,14 @@ static int tmc_open(struct inode *inode, struct file
>> *file)
>>                                                    struct tmc_drvdata,
>> miscdev);
>>         int ret = 0;
>>
>
> On a second thought, I think there could be a race here.
>
>
>> -       if (drvdata->read_count++)
>> +       if (drvdata->read_count)
>>                 goto out;
>>
>>         ret = tmc_read_prepare(drvdata);
>>         if (ret)
>>                 return ret;
>>   out:
>
>
> What prevents someone else doing a release() on the file when we get here,
> without
> incrementing the read_count ? Also, read_count accesses are not protected.
> Either should
> be covered by the drvdata->spinlock or convert it to atomic.

I agree - I'll move it to an atomic type.

Thanks,
Mathieu

>
>
>
>> +       drvdata->read_count++;
>>         nonseekable_open(inode, file);
>
>
>
> Cheers
> Suzuki



More information about the linux-arm-kernel mailing list