[lvm-devel] [PATCHv2] libdm: mark control fd as close-on-exec

[Zdenek Kabelac]

Dne 19.8.2015 v 15:42 Mathias Krause napsal(a):
> On 15.07.2015 09:34, Mathias Krause wrote:
>> The control fd should be marked as close-on-exec to avoid file
>> descriptor leaks in forking applications executing other programs.
>>
>> Signed-off-by: Mathias Krause <mathias.krause at secunet.com>
>>
>> ---
>> v2: fix return value mix-up (should be 1 on success)
>> ---
>>   libdm/ioctl/libdm-iface.c | 13 +++++++++++++
>>   1 file changed, 13 insertions(+)
>>
>> diff --git a/libdm/ioctl/libdm-iface.c b/libdm/ioctl/libdm-iface.c
>> index e3b33b805e93..0f9e98a4a061 100644
>> --- a/libdm/ioctl/libdm-iface.c
>> +++ b/libdm/ioctl/libdm-iface.c
>> @@ -385,11 +385,24 @@ static void _close_control_fd(void)
>>   #ifdef DM_IOCTLS
>>   static int _open_and_assign_control_fd(const char *control)
>>   {
>> +#ifdef O_CLOEXEC
>> +	/*
>> +	 * O_CLOEXEC is supported since v2.6.23, so this may fail on old
>> +	 * kernels. Nonetheless favour it to a two staged approach as it's
>> +	 * atomic.
>> +	 */
>> +	if ((_control_fd = open(control, O_RDWR | O_CLOEXEC)) >= 0)
>> +		return 1;
>> +#endif
>> +
>>   	if ((_control_fd = open(control, O_RDWR)) < 0) {
>>   		log_sys_error("open", control);
>>   		return 0;
>>   	}
>>
>> +	if (fcntl(_control_fd, F_SETFD, FD_CLOEXEC))
>> +		log_sys_error("fcntl", "setting FD_CLOEXEC");
>> +
>>   	return 1;
>>   }
>>   #endif
>
> Ping? Any objections to the patch?
>


Hi

_control_fd  is then 'set' in forked code - means 'forked' libdm user might 
live with impression it has opened control handler.

So you would also need to ensure _control_fd is zeroed.

Zdenek