Re: [dpdk-dev] [PATCH] vfio: fix close unchecked file descriptor

[Patrick MacArthur <patrick@patrickmacarthur.net>]

On 09/20/2017 10:39 AM, Burakov, Anatoly wrote:
> On 20-Sep-17 3:34 PM, Patrick MacArthur wrote:
>> On 09/20/2017 05:59 AM, Kuba Kozak wrote:
>>> Add file descriptor value check before calling close() function.
>>>
>>> Coverity issue: 141297
>>> Fixes: 811b6b25060f ("vfio: fix file descriptor leak in multi-process")
>>> Cc: patrick@patrickmacarthur.net
>>> Cc: stable@dpdk.org
>>>
>>> Signed-off-by: Kuba Kozak <kubax.kozak@intel.com>
>>> ---
>>>   lib/librte_eal/linuxapp/eal/eal_vfio_mp_sync.c | 3 ++-
>>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/lib/librte_eal/linuxapp/eal/eal_vfio_mp_sync.c 
>>> b/lib/librte_eal/linuxapp/eal/eal_vfio_mp_sync.c
>>> index 7e8095c..c04f548 100644
>>> --- a/lib/librte_eal/linuxapp/eal/eal_vfio_mp_sync.c
>>> +++ b/lib/librte_eal/linuxapp/eal/eal_vfio_mp_sync.c
>>> @@ -301,7 +301,8 @@ vfio_mp_sync_thread(void __rte_unused * arg)
>>>                   vfio_mp_sync_send_request(conn_sock, SOCKET_ERR);
>>>               else
>>>                   vfio_mp_sync_send_fd(conn_sock, fd);
>>> -            close(fd);
>>> +            if (fd != -1)
>>> +                close(fd);
>>
>> IMHO this should be:
>>
>>          if (fd >= 0)
>>
>> What specifically is Coverity complaining about here? Is there a 
>> specific code path that leads to fd being -1 here?
>>
> Hi Patrick,
> 
> There's no way the fd will be 0 - the function we get the value from 
> returns a valid fd, or a -1 in case of error. In this particular case, 
> the "specific code path that leads to fd being -1" is when we can't get 
> a container fd for some reason. I believe this is a very remote 
> possibility as by the time we're spinning up the socket listening thread 
> we're pretty sure we have a working VFIO container, but this is a valid 
> fix nevertheless. Maybe having it >= 0 (or > 0, to be precise) would be 
> cleaner, but it really makes no difference here.

The point of my suggestion is that it would catch *any* negative value 
for fd as opposed to just -1.

I agree 0 should never happen since it is stdin but it is technically a 
valid fd that could occur if the user program did close(STDIN_FILENO) 
for some reason.

I don't feel too strongly about it but feel like if we are going to fix 
what amounts to close() possibly returning EBADF we might as well fix it 
for all cases.

Thanks,
Patrick