speedfreq: epia + longhaul + speedfreq + copying large files = freeze

Dear Jeremy,

thanks for your tool! I like it best of all cpufreq daemons,

Anyway, the Epia board normally runs fine with speedfreq. Frequency gets adjusted between 400Mhz and 600MHz. But when I start cutting advertisements out of the DVB recordings with the VDR software (Video Disk Recorder from Klaus Schmiedinger, www.cadsoft.de/vdr) after a short while it freezes. The cutting process is simply like copying large files from one folder to another folder on the same partition. Mostly 1.7 to 2.7GB. I tried the following:

1. Cutting a recording, cpufreq enabled, speedfreq daemon running:

System freeze.

2. Copying large files manually, cpufreq enabled, speedfreq daemon running:

System freeze.

3. Cutting a recording, cpufreq enabled, speedfreq daemon disabled:

Works.

4. Copying large files manually, cpufreq enabled, speedfreq daemon disabled:

Works.

I'd like to add that sometimes I get messages in syslog telling me that speedfreq thought the actual frequency was xxx and discovered that the real frequency was something else. Once that was the last thing in syslog before a freeze. I don't know if this is relevant, because 99% of the freezes the syslog stays like it is, nothing gets added.

The kernel is a vanilla 2.6.9. The Linux Distribution is Gentoo stable. Flags are "-O2 -march=c3 -fomit-frame-pointer -mmmx -m3dnow -pipe". I have no 4k stacks, no preemption, no nptl, no perl ithreads.

Maybe speedfreq changing frequencies too fast for the little Epia is the reason for the freezes? What can I do?

Quoted 3 lines · Sebastian

Dear Jeremy,

thanks for your tool! I like it best of all cpufreq daemons,

Thanks!

This problem sounds like a bug in the longhaul driver though, which I don't really know anything about. (The system should never crash, regardless of what speedfreqd does.)

I've cc:d the cpufreq list to see if someone there can help. Dave?

Yeah, longhaul is badly broken right now, and tbh, I've no idea what on earth is wrong. Speed transitions happen, and work correctly. Then a few minutes later, system locks up.

I've been trying all sorts of things, from disabling PCI mastering, disabling CPU caches, etc, but nothing seems to help. I'm truly puzzled.

Quoted 5 lines · Jeremy Fitzhardinge

Quoted 3 lines · Sebastian

Dear Jeremy,

thanks for your tool! I like it best of all cpufreq daemons,

Quoted 7 lines

Thanks!

This problem sounds like a bug in the longhaul driver though, which I don't really know anything about. (The system should never crash, regardless of what speedfreqd does.)

I've cc:d the cpufreq list to see if someone there can help. Dave?

Yeah, longhaul is badly broken right now, and tbh, I've no idea what on earth is wrong. Speed transitions happen, and work correctly. Then a few minutes later, system locks up.

I've been trying all sorts of things, from disabling PCI mastering, disabling CPU caches, etc, but nothing seems to help. I'm truly puzzled.

Dave

Quoted 2 lines

I've been trying all sorts of things, from disabling PCI mastering, disabling CPU caches, etc, but nothing seems to help. I'm truly puzzled.

I remember old versions of the C3 used to forget to snoop DMA traffic while in the hlt instruction, so it wouldn't surprise me if there's some other bug like that related to longhaul.

Do you mean that it will crash some minutes after a single transition, or is it less deterministic than that?

The reason I ask is that I have the ear of Via engineering support at the moment (we're using embedded C3's in a product at work), so I could ask a few well-placed questions.

Usually takes a few. I've noticed if I run cpuspeed, it steps it down from 1.2GHz to 500MHz through a half dozen or so steps. Whilst it does this, its stable. If I then kill cpuspeed, its stable.
But, if I leave cpuspeed run, and then do something CPU intensive, it tries to jump up to fullspeed quickly (through less steps than it took to scale down), when it does this, it hangs shortly afterwards.

We're not doing voltage scaling in longhaul yet, but that shouldn't matter, as we should be booting at max voltage (as we boot at max speed)

Sebastian is reporting the problem with speedfreqd, which will use the same number of steps both ways (and, depending on the CPU load curve, at the same rate).

Does rate-limiting the transitions help stability?

Quoted 2 lines

We're not doing voltage scaling in longhaul yet, but that shouldn't matter, as we should be booting at max voltage (as we boot at max speed)

Is voltage switching complex? It might be worth trying - its not something which normally improves stability, but maybe there's some regulator issue under load changes, or something...