Thanks for the patch Corey.
I am afraid that the system does not have interrupts enabled. It uses polling
mode.
When the error is seen, I know for a fact that in function ipmi_thread()
smi_result is SI_SM_CALL_WITH_DELAY,
I have some logs where in busy_wait always reads as 1. Not sure if it was ever
set to 0. (ill check this again).
Ill anyway run the test using the patch that you have shared.
b/w would it harm if we were to do to something like this ?
Signed-off-by: Srinivas Gowda <srinivas_g_go...@dell.com>
---
drivers/char/ipmi/ipmi_si_intf.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/drivers/char/ipmi/ipmi_si_intf.c b/drivers/char/ipmi/ipmi_si_intf.c
index 15e4a60..e23484f 100644
--- a/drivers/char/ipmi/ipmi_si_intf.c
+++ b/drivers/char/ipmi/ipmi_si_intf.c
@@ -1008,6 +1008,7 @@ static int ipmi_thread(void *data)
spin_unlock_irqrestore(&(smi_info->si_lock), flags);
busy_wait = ipmi_thread_busy_wait(smi_result, smi_info,
&busy_until);
+ ipmi_start_timer_if_necessary(smi_info);
if (smi_result == SI_SM_CALL_WITHOUT_DELAY)
; /* do nothing */
else if (smi_result == SI_SM_CALL_WITH_DELAY && busy_wait)
--
1.8.1.2
Thanks,
G
On 11/30/2013 05:49 AM, Corey Minyard wrote:
> On 11/27/2013 04:34 AM, srinivas_g_go...@dell.com wrote:
>>
>> *Dell - Internal Use - Confidential *
>>
>> I hit a bug during one of our stress tests, Here is the issue that I
>> am looking at.
>>
>> We have IPMI_READ_EVENT_MSG_BUFFER_CMD getting invoked from
>> smi_event_handler.
>>
>> In case we hit error scenario, say "OBF not ready in time" we do not
>> have smi_timeout driving the interface.
>>
>> Seems like the timer is not armed when we invoke
>> IPMI_READ_EVENT_MSG_BUFFER_CMD from smi_event_handler.
>>
>> For the proposed patch I checked the return value of mod_timer just
>> before smi_info->handlers->start_transaction, that returns 0 !!!
>>
>> gWithout smi_timeout handler getting called periodically, if the BMC
>> fails to set OBF flag during the msg transaction of
>> IPMI_READ_EVENT_MSG_BUFFER_CMD,
>>
>> the driver just keeps looping until the flag is set. Ideally we would
>> want BMC to set the flag, but in case it doesn’t we do not want the
>> driver to loop indefinitely rather hit KCS_ERROR states.
>>
>> To summarize, we do not have timer set to invoke smi_timeout() when we
>> call IPMI_READ_EVENT_MSG_BUFFER_CMD from smi_event_handler.
>>
>> Do you feel there is a better way to fix it or a bug elsewhere…!
>>
>
> Ok, I think I know what is happening, and I think I have a fix. I'm
> betting that you have interrupts on this, and
> I found a situation where if an interrupt came in at a certain time, it
> wouldn't start the timer. The attached patch should fix the problem.
>
> Can you try this out?
>
> Thanks for the detailed description.
>
> -corey
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
