Re: [PATCH resend ] tty/n_gsm.c: use gsm->num to remove mux itself from gsm_mux[]

[Pan Xinhui <xinhui.pan@linux.vnet.ibm.com>]

Hi, Alan
	thanks for your reply :)

On 2015/12/18 21:17, xinhui wrote:
> hi, Alan
> this is xinhui. My eyes got badly hurt, and i am ooo this whole week and next coming week. sorry for late responce.
> I just review the codes in my mind. gsm ioctl callback might change gsm->num, so you are right.
> i still have many confusion. but tears came out several times:( when i am back, i will reply you again.
>
> thx
> xinhui
>
>
>
> On 2015-12-14 23:40 , One Thousand Gnomes Wrote:
>
> On Mon, 14 Dec 2015 15:08:03 +0800
> Pan Xinhui <xinhui.pan@linux.vnet.ibm.com> wrote:
>
>> From: Pan Xinhui <xinhui.pan@linux.vnet.ibm.com>
>>
>> There is one filed gsm->num to store mux's index of gsm_mux[]. So use
>> gsm->num to remove itself from gsm_mux[] instead of the for-loop
>> traverse in gsm_cleanup_mux().
>>
>> Signed-off-by: Pan Xinhui <xinhui.pan@linux.vnet.ibm.com>
>> ---
>>   drivers/tty/n_gsm.c |   14 +++++---------
>>   1 files changed, 5 insertions(+), 9 deletions(-)
>>
>> diff --git a/drivers/tty/n_gsm.c b/drivers/tty/n_gsm.c
>> index 9aff371..cf28054 100644
>> --- a/drivers/tty/n_gsm.c
>> +++ b/drivers/tty/n_gsm.c
>> @@ -2037,18 +2037,14 @@ static void gsm_cleanup_mux(struct gsm_mux *gsm)
>>
>>        gsm->dead = 1;
>>
>> -     spin_lock(&gsm_mux_lock);
>> -     for (i = 0; i < MAX_MUX; i++) {
>> -          if (gsm_mux[i] == gsm) {
>> -               gsm_mux[i] = NULL;
>> -               break;
>> -          }
>> -     }
>> -     spin_unlock(&gsm_mux_lock);
>>        /* open failed before registering => nothing to do */
>> -     if (i == MAX_MUX)
>> +     if (gsm_mux[gsm->num] != gsm)
>>             return;
>>
>> +     spin_lock(&gsm_mux_lock);
>> +     gsm_mux[gsm->num] = NULL;
>> +     spin_unlock(&gsm_mux_lock);
>
> Its a highly theoretical and probably impossible corner case but I can't
> help thinking the lock should be held for the if () as well as NULLing
> this out.
>
yes, gsm_mux[] must be touched with gsm_mux_lock held.

I am still wondering if it's possible that two gsm_cleanup_mux() run on the same mux.
seems gsmld_config() -> gsm_cleanup_mux() might have race with gsmld_detach_gsm() -> gsm_cleanup_mux().
what's more, we need make sure gsm_mux[gsm->num] == gsm, as if there is a new mux put into gsm_mux[], we might NULL this new mux out.

here is one possible race.
CPUA				  CPUB					CPUC
in cleanup()                     in cleanup()                         in activate()
if (gsm_mux[gsm->num] != gsm)    if (gsm_mux[gsm->num] != gsm)
..					...
spin_lock(&gsm_mux_lock);
gsm_mux[gsm->num] = NULL;
spin_unlock(&gsm_mux_lock);
									spin_lock(&gsm_mux_lock);
									...
										gsm->num = i;
										gsm_mux[i] = gsm;
									...
									spin_unlock(&gsm_mux_lock);
				spin_lock(&gsm_mux_lock);
				gsm_mux[gsm->num] = NULL;//this NULLing might cause BUGS!!
				spin_unlock(&gsm_mux_lock);

I will send out patch V2 to avoid any possible race.
thanks for pointing it out.

thanks
xinhui

> Alan
>