Skip to contentSkip to conversation

Tracing/ftrace: Don't consume entries unhandled by mmiotrace

PATCH Applied Frédéric Weisbecker 4 participants 8 messages Sep 25, 2008 – Sep 27, 2008

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 3/3 Tracing/ftrace: Don't consume entries unhandled by mmiotrace

When mmiotrace can't handle an entry output, it returns 1. It should return 0 to relay on other output functions.

---

kernel/trace/trace_mmiotrace.c +1 -1
index a108c32..6fe1315 100644
@@ -277,7 +277,8 @@ static int mmio_print_line(struct trace_iterator *iter)
 	case TRACE_PRINT:
 		return mmio_print_mark(iter);
 	default:
-		return 1; /* ignore unknown entries */
+		return 0; /* ignore unknown entries whithout consuming */
 	}
 }
 
Applied
#1 lore raw
Quoted 4 lines · Frédéric Weisbecker

When mmiotrace can't handle an entry output, it returns 1. It should return 0 to relay on other output functions.

applied to tip/tracing/mmiotrace, thanks Frédéric!

Ingo

#2 lore raw
In reply to Frédéric Weisbecker
Quoted 10 lines · On Thu, 25 Sep 2008 13:28:56 +0100 Frédéric Weisbecker <[email protected]>…

When mmiotrace can't handle an entry output, it returns 1. It should return 0 to relay on other output functions.

---

kernel/trace/trace_mmiotrace.c +0 -0
index a108c32..6fe1315 100644

kernel/trace/trace_mmiotrace.c static int mmio_print_line(struct trace_iterator *iter) · line 282

Quoted 7 lines
kernel/trace/trace_mmiotrace.c +1 -1
@@ -277,7 +277,8 @@ static int mmio_print_line(struct trace_iterator *iter)
 	case TRACE_PRINT:
 		return mmio_print_mark(iter);
 	default:
-		return 1; /* ignore unknown entries */
+		return 0; /* ignore unknown entries whithout consuming */
 	}
 }

NACK.

mmiotrace's log output is stricly specified, the standard/default printing functions may NOT be used. The output is supposed to be machine readable in addition to human readable.

The ftrace infrastructure assumes there is only one active tracer at a time, therefore destroying unhandled entries is not a problem.

Signature · 2 lines
Pekka Paalanen
http://www.iki.fi/pq/
#3 lore raw

2008/9/25 Pekka Paalanen <[email protected]>:

Quoted 6 lines

NACK.

mmiotrace's log output is stricly specified, the standard/default printing functions may NOT be used. The output is supposed to be machine readable in addition to human readable.

Quoted 7 lines

The ftrace infrastructure assumes there is only one active tracer at a time, therefore destroying unhandled entries is not a problem.

--
Pekka Paalanen
http://www.iki.fi/pq/

Hi Pekka.

It's up to you.
I just guessed that one would trace IO and stack for example. Or why not IO and functions: This sounds to me very useful (who performed this io..?)

As you want....

Regards...

#4 lore raw
Quoted 10 lines · On Thu, 25 Sep 2008 16:44:08 +0200 "Frédéric Weisbecker" <[email protected]>…

2008/9/25 Pekka Paalanen <[email protected]>:

Quoted 8 lines

NACK.

mmiotrace's log output is stricly specified, the standard/default printing functions may NOT be used. The output is supposed to be machine readable in addition to human readable.

The ftrace infrastructure assumes there is only one active tracer at a time, therefore destroying unhandled entries is not a problem.

Quoted 6 lines

Hi Pekka.

It's up to you.
I just guessed that one would trace IO and stack for example. Or why not IO and functions: This sounds to me very useful (who performed this io..?)

We can (must) return to this if/when multiple tracers are allowed to be active simultaneously. But in the current situation, I do not see a way to trace more than one thing, but I haven't really looked into the other tracers.

OTOH, mmiotrace log format has a pid field, which is pretty much unused currently. It could be used to record the current process, if one exists. That can be added, should the need arise, but so far the field has been reserved for tracing accesses done in user space - but those are not caught yet.

I'm not sure what stack trace collects, but tracing functions is useless or harmful in the usual use case of mmiotrace, i.e. tracing a binary-only proprietary driver. The functions will be anonymous anyway, and we can't disassemble the driver due to possible legal reasons. That's why mmiotrace does not even collect instruction pointers by default, although it supports them (or did, can't recall if it works now, since it hasn't been needed). I assume the goal is to create a free open driver without any legal issues by watching what the proprietary driver does.

Even if the log would be cluttered with MMIO from other, uninteresting drivers, you can use the physical address to filter the entries by device (and therefore driver) afterwards. This is one reason why an mmiotrace log starts with the contents of /proc/bus/pci/devices.

Was there some other scenario you were thinking about?

Signature · 2 lines
Pekka Paalanen
http://www.iki.fi/pq/
#5 lore raw

2008/9/25 Pekka Paalanen <[email protected]>:

Quoted 4 lines

We can (must) return to this if/when multiple tracers are allowed to be active simultaneously. But in the current situation, I do not see a way to trace more than one thing, but I haven't really looked into the other tracers.

You can enable another tracer simultaneously. I never tried it but you can launch one from debugfs and another from kernel code for example. The new boot tracer will have to launch the sched_switch tracer. And since the functions which display the sched_switch entries are already inside
trace.c I wanted to relay the display to the "default display functions". That's one of
the reason of my patch (+ the bug).
But I could proceed in a another way than default relaying. Why not using a function exported
by the tracing engine to display sched_switches.

Quoted 9 lines

but tracing functions is useless
or harmful in the usual use case of mmiotrace, i.e. tracing a binary-only proprietary driver. The functions will be anonymous anyway, and we can't disassemble the driver due to possible legal reasons. That's why mmiotrace does not even collect instruction pointers by default, although it supports them (or did, can't recall if it works now, since it hasn't been needed). I assume the goal is to create a free open driver without any legal issues by watching what the proprietary driver does.

Quoted 6 lines

Even if the log would be cluttered with MMIO from other, uninteresting drivers, you can use the physical address to filter the entries by device (and therefore driver) afterwards. This is one reason why an mmiotrace log starts with the contents of /proc/bus/pci/devices.

Was there some other scenario you were thinking about?

When I wanted to help on the ath5k project, I used to compare madwifi traces with the ath5k traces. The goal was to find where was the trouble when support for ar2425 cards didn't worked.
And I wanted to see the name of the functions where the code entered to see where I was in the traces. That's why I wanted to have an mmiotrace_printk function primarily :)
But I think one could use it for debugging too. Just to know where we are in the trace.

#6 lore raw
Quoted 9 lines · Frédéric Weisbecker

2008/9/25 Pekka Paalanen <[email protected]>:

Quoted 4 lines

We can (must) return to this if/when multiple tracers are allowed to be active simultaneously. But in the current situation, I do not see a way to trace more than one thing, but I haven't really looked into the other tracers.

You can enable another tracer simultaneously. I never tried it but you can launch one from debugfs and another from kernel code for example. The new boot tracer will have to launch the sched_switch tracer.

hm, right now i can see a lot of problems with that approach - the buffers are partly per tracer and partly global.

the new trace-ringbuffer design Steve is working on would abstract away this complication - then we could make tracers truly independent entities.

Ingo

maintainer
#7 lore raw
Quoted 4 lines · Ingo Molnar

the new trace-ringbuffer design Steve is working on would abstract away this complication - then we could make tracers truly independent entities.

Yep, I even started implementing this. What is nice about this too, is that we could even mix the tracer buffers together in the merge sort. Since all buffers use the same timestamping, we can keep the separate buffers out, turn on different tracers, and have the output show them all together.

This would mean that we can keep the function tracer buffer large, and the other tracer buffers smaller. And merge them together when needed.

I'll have to recreate it, since the buffer code changed quite a bit since I first started making ftrace do this. Also, I want to make the changes to ftrace as incremental as possible. That way we could easily diagnose problems that arise.

-- Steve