Re: [PATCH 1/6] usb: xhci: add xhci_log_cmd trace events

[Lu Baolu]

Hi,

On 02/15/2017 04:56 PM, Felipe Balbi wrote:
> Hi,
>
> Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx> writes:
>>> Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx> writes:
>>>> diff --git a/drivers/usb/host/xhci-trace.h b/drivers/usb/host/xhci-trace.h
>>>> index 1ac2cdf..c31eeaf 100644
>>>> --- a/drivers/usb/host/xhci-trace.h
>>>> +++ b/drivers/usb/host/xhci-trace.h
>>>> @@ -285,6 +285,96 @@ DEFINE_EVENT(xhci_log_urb, xhci_urb_dequeue,
>>>>  	TP_ARGS(urb)
>>>>  );
>>>>  
>>>> +DECLARE_EVENT_CLASS(xhci_log_cmd,
>>>> +	TP_PROTO(struct xhci_command *cmd),
>>>> +	TP_ARGS(cmd),
>>>> +	TP_STRUCT__entry(
>>>> +		__field(struct xhci_command *, cmd)
>>>> +		__field(struct xhci_container_ctx *, in_ctx)
>>>> +		__field(union xhci_trb *, cmd_trb)
>>>> +		__field(int, slot_id)
>>>> +		__field(int, status)
>>>> +		__field(int, type)
>>>> +	),
>>>> +	TP_fast_assign(
>>>> +		__entry->cmd = cmd;
>>>> +		__entry->in_ctx = cmd->in_ctx;
>>>> +		__entry->cmd_trb = cmd->command_trb;
>>>> +		__entry->slot_id = cmd->slot_id;
>>>> +		__entry->status = cmd->status;
>>>> +		__entry->type = TRB_FIELD_TO_TYPE(le32_to_cpu(cmd->command_trb->generic.field[3]))
>>>> +	),
>>>> +	TP_printk("cmd @%p: %s: in_ctx=@%p, slot_id=%d, cmd_trb=@%p, status=%d",
>>>> +		__entry->cmd, xhci_trb_type_string(__entry->type),
>>>> +		__entry->in_ctx, __entry->slot_id, __entry->cmd_trb,
>>>> +		__entry->status
>>>> +	)
>>>> +);
>>> we already have a generic TRB tracer that decodes every single TRB. What
>>> is this bringing that the previous doesn't provide?
>> This tracer traces the life cycle of a command. It gives,
>>
>> 1) Which function started an xhci command?
>> 2) What was name of that command?
>> 3) Did hardware respond to it, or timed out?
>> 4) If hardware responded, what was the execution result?
>> 5) If timed out, did 'abort command ring operation' abort it successfully?
>> 6) Was the command structure freed at last?
> We already have all that, AFAICT. Command is enqueued, then an event
> triggers for command completion, then we look at results.

The TRB tracer can directly tell us above 2 and 4 by tracing the enqueue
and dequeue of command trbs. It doesn't tell us 1 and 6. For 3 and 5, we
might be able to get to know, but it needs deep understand of the driver
code and many efforts to look through the trace log.

This command tracer will make the life easier when we are debugging
command related issues in xhci driver.


>  The only thing
> missing for completeness is slot/EP context tracers (which I've pointed
> you to) so we can see what changes each command cause to the different
> contexts.
>
> Frankly, I don't think printing out context pointers brings
> anything. What can you do with that address? :-p Same goes for cmd
> pointer, it brings nothing; gives no insight into the problem
> whatsoever.
>
> We certainly need to know which command was enqueued, the slot, etc. But
> addresses... not so sure.

My thought was 'if we find a command failed, we might want to
know what parameters did we feed hardware'. The trb and ctx
pointers will let us go through the trb and context trace log.

For the command pointer, it can help us to grep the trace log
of a single command, so we are able to know how did the
command go.

Best regards,
Lu Baolu

>
>>> BTW, I also have
>>> ready Slot and EP context tracers, I didn't send before because I
>>> already had quite a large series pending for Mathias :-p
>> Sorry for the duplication.
> no need to apologize, you didn't know :-)
>