Re: [PATCH RFC 0/2] Add basic tracing support for m68k

[Jean-Michel Hautbois <jeanmichel.hautbois@yoseli.org>]

Hi Steve,

On 19/11/2024 19:10, Steven Rostedt wrote:
> On Tue, 19 Nov 2024 19:06:45 +0100
> Jean-Michel Hautbois <jeanmichel.hautbois@yoseli.org> wrote:
>>>
>>> It shouldn't crash, but it also found a bug in your code ;-)
>>
>> In my code is a really big assumption :-).
> 
> Well, not your personally, but I meant "your" as in m68k code.
> 
>>
>>> You reference two variables that are not part of the event:
>>>
>>>    "mem_map" and "m68k_memory[0].addr"
>>>
>>> Do these variables ever change? Because the TP_printk() part of the
>>> TRACE_EVENT() macro is called a long time after the event is recorded. It
>>> could be seconds, minutes, days or even months (and unlikely possibly
>>> years) later.
>>
>> I am really not the best placed to answer.
>> AFAIK, it sounds like those are never changing.
> 
> That would mean they are OK and will not corrupt the trace, but it will be
> meaningless for tools like perf and trace-cmd.
> 
>>
>>>
>>> The event takes place and runs the TP_fast_assign() to record the event in
>>> the ring buffer. Then some time later, when you read the "trace" file, the
>>> TP_printk() portion gets run. If you wait months before reading that, it is
>>> executed months later.
>>>
>>> Now you have "mem_map" and "m68k_memory[0].addr" in that output that gets
>>> run months after the fact. Are they constant throughout the boot?
>>
>> I don't know.
>>
>>> Now another issue is that user space has no idea what those values are. Now
>>> user space can not print the values. Currently the code crashes because you
>>> are the first one to reference a global value from a trace event print fmt.
>>> That should probably be fixed to simply fail to parse the event and ignore
>>> the print format logic (which defaults to just printing the raw fields).
>>
>> The patch you sent works...
>> But, it fails a bit later:
>> Dispatching timerlat u procs
>> starting loop
>> User-space timerlat pid 230 on cpu 0
>> Segmentation fault
>>
> 
> More printk? ;-)

Indeed, but the result is not straightforward this time :-(.

Long story short: it fails at kbuffer_load_subbuffer() call in
read_cpu_pages().

I added printf in the kbuffer helpers in libevent, and it finishes at:
__read_long_4: call read_4 at 0x600230c2
__read_4_sw: ptr=0x8044e2ac

static unsigned int __read_4_sw(void *ptr)
{
	printf("%s: ptr=%p, value: %08x\n", __func__, ptr, *(unsigned int *)ptr);
	unsigned int data = *(unsigned int *)ptr;
	printf("%s: data=%08x\n", __func__, data);

	return swap_4(data);
}

As soon as ptr is dereferenced, the segfault appears.
ptr should be ok though, as the address is valid afaik...

I must say that now I am stuck :-(.

Thanks,
JM