Jan Kiszka wrote:
Philippe Gerum wrote:

Jan Kiszka wrote:

Philippe Gerum wrote:

I've just rolled out two patches, the first issue of the 1.1 series for
x86, and the accompanying tracer patch contributed by Jan Kiszka and
Luotao Fu. With the latter patch, the I-pipe shall trace the longest
stalled path of the domain with the highest priority. Apply them in that



Two remarks: First, the tracer patch claims more in its config option
than it actually provides - mea culpa. The patch itself does not contain
any instrumentation of ipipe. This has to be fixed. Meanwhile, please
have a look at this posting for instrumentation options:

Philippe, do you remember the issues I described about my original
ipipe_trace.instr? How can we avoid too short worst-case traces due to
domain unstalling followed by re-stalling inside the same IRQ context?
Do you see further issues with this approach? I think it would be best
if we can provide a clean CONFIG_IPIPE_TRACE_STALLS for the highest (or
later maybe even for an arbitrary) domain together with the tracer.

+static inline void ipipe_trace_stall(struct ipipe_domain *ipd, int code)
+    if (__ipipe_pipeline_head_p(ipd) && (ipd != ipipe_root_domain))
+        ipipe_trace_begin(code);
+static inline void ipipe_trace_unstall(struct ipipe_domain *ipd, int code)
+    if (__ipipe_pipeline_head_p(ipd) && (ipd != ipipe_root_domain))
+        ipipe_trace_end(code);

The test is wrong in both case. You need to check that ipd is above or
equal to ipipe_current_domain in the pipeline. To determine that quickly
while tracing, you will probably need to insert an integer giving the
position of each domain into the ipipe_domain struct.

So the orderning in __ipipe_pipeline does say nothing about the priority
of the domain? Then this seems to have worked only by chance so far for me.

Of course it does. The thing is that your test must reflect the fact that stalling above and up to the current domain actually blocks the interrupts for the latter, and unstalling at least from your current domain unblocks them. The position value is just a suggestion to quickly compare the effective priority of two domains given their descriptor, without being stuck with the uncertainty of ipd->priority which might be the same for multiple domains, and without having to scan the pipeline top-down.

Anyway, fixing this does not seem to address the other issue I
mentioned. I once also traced the evaluation of those two conditions and
found no indication that this triggers the preliminary end-of-stall
reports I'm facing.

And second, the separation between both patches is not clean. There are
tracer related fragments in the 1.1-00 base patch, intentionally? What's
the idea of the separated patches? I mean, doesn't this increase the
maintenance effort?

It's intentional, those (very few) bits always evaluate to false when
the tracer is not in, and become conditional depending on the value of
CONFIG_IPIPE_TRACE when the support available. IOW, they should be seen
as sleeping hooks serving the purpose of allowing a further optional
extension of the I-pipe.

I see. Then these hooks are intended to keep the effort of breaking up
the patches low.

Yes. Actually, the latency tracer is merged into the Adeos CVS tree on top of the core system; it's just my patch release script that splits them since they are well separated. The remaining hooks do the necessary glue between them.

The key issue here is not about ease of maintenance for us, but rather
about ease of use for the people who don't necessarily want to drag
what's fundamentally a debug infrastructure into the codebase of
production systems, even if it's passive and can be compiled out. Adeos
for x86 is about 151k without tracing, and goes beyond 189k with the
tracer, which is nearly a 20% increase. Add to this that since a latency
tracer is now available for vanilla Linux as an independent patch, it's
likely wiser to allow people to keep the I-pipe tracing facility as a
patch option too, so that you won't create conflicts (e.g. mcount).

Actually, both traces should not collide as long as only one is active
at the same time.

Unfortunately, we can't bet on this for the vanilla kernel part, who knows what's going to happen to this support in the future?

Anyway, I already assumed that this more or less psychological aspect of
patch size makes a difference. I don't have a problem with this separation!




Reply via email to