Re: [PATCH] x86/tracing: introduce enter/exit tracepoint pairs for page faults

[Junxuan Liao <ljx@cs.wisc.edu>]

Sorry I forgot to reply to the list. :( It's my first time doing this.

On 4/14/25 4:56 PM, Dave Hansen wrote:
> On 4/14/25 13:52, Junxuan Liao wrote:
>> On 4/14/25 3:42 PM, Dave Hansen wrote:
>>> Is there a reason kprobes don't work for this?
>>
>> In this code path it's either noinstr or inline functions, so I believe
>> explicit tracepoints are necessary?
> 
> It'd be great to turn this "??" into some more certainty. ;)
> 
>> As Steven has mentioned, there are similar tracepoints for irq and it's
>> nice to have them for page faults too.
> 
> So, that code is:
> 
>>                 trace_irq_handler_entry(irq, action);
>>                 res = action->handler(irq, action->dev_id);
>>                 trace_irq_handler_exit(irq, action, res);
> 
> ... I think.
> 
> That's a heck of a lot simpler than a couple static keys and new
> conditionals.
> 
> Also, I honestly don't think we need separate user and kernel fault
> trace points. Maybe we should just zap that in the process.
> 
> Is there a reason we couldn't get this down to something dirt simple like:
> 
>  DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault)
>  {
>         instrumentation_begin();
> +	trace_page_fault_entry(...);
>         handle_page_fault(regs, error_code, address);
> +	trace_page_fault_exit(...);
>         instrumentation_end();
> 
> ??

If we don't need to separate user and kernel tracepoints, we won't need
those static keys anyway, and it's indeed much simpler.

Do people find separate user/kernel tracepoints useful? For me, I can
check regs in eBPF tracing code instead.

Junxuan